ONLINE CENTRALIZED MONITORING AND ANALYSIS METHOD FOR MULTI-POINT MALODOROUS GASES USING ELECTRONIC NOSE INSTRUMENT

Abstract

Provided is an online centralized monitoring and analysis system based on an electronic nose instrument for multi-point malodorous gases, and the system includes an electronic nose instrument, which connects with multiple monitoring points through pipes. On-site malodorous gases in the maximum range of 2.5 km are drawn into odor electronic nose instrument within 1.0 min by the external vacuum pump, and forced to flow through the annular gas sensor array for 30 s by the internal vacuum pump periodically. The modular convolution neural networks online learn the recent time-series responses of the gas sensor array and predict their coming responses, and the modular deep neural networks offline set up the relationship between the responses and multiple concentration items according to odor big data. The odor electronic nose instrument monitors up to 10 pollution sites cyclically and uses the cascade machine learning model to online predict one dimensionless unit and 10 specified-component concentration index values of malodorous gases.

Claims

1. An online centralized monitoring and analysis system based on an electronic nose instrument for multi-point malodorous gases, comprising an odor electronic nose instrument (I), a plurality of gas sampling heads (II), an external vacuum pump (III), an ambient air purification device (IV), a clean air cylinder (V), gas pipelines, an electronic hygrometer (VI), a central control room (VII) and a plurality of stationary/mobile terminals (VIII), for long-term cyclical monitoring of 10 malodorous pollution sites, and online estimation and prediction of multiple concentration control index values of malodorous gases; wherein the odor electronic nose instrument (I) comprises a gas sensor array (I-1) and a thermostatic working room (I(a)) of the gas sensor array, a multi-point centralized auto-sampling system (I(b)), and a computer control and data analyzing system (1(c)); the thermostatic working room (I(a)) of the gas sensor array is comprised of the gas sensor array and an annular working chamber of the gas sensor array, a thermal insulation layer (I-2), a resistance heating wire (I-3) and a fan (I-4); the gas sensor array (I-1) is comprised of 16 gas sensors, which are uniformly distributed in a sealed chamber having a middle diameter of 140 mm and a section size of 21 mm17 mm to form the annular working chamber of the gas sensor array; the thermostatic working room (I(a)) is with a constant temperature of 550.1 C. and located at a top right of the odor electronic nose instrument (I); wherein the multi-point centralized auto-sampling system (I(b)) comprises an internal miniature vacuum pump (I-14), a first two-position two-port electromagnetic valve (I-5), a second two-position two-port electromagnetic valve (I-6-1), a third two-position two-port electromagnetic valve (I-6-2), a fourth two-position two-port electromagnetic valve (I-6-3), a fifth two-position two-port electromagnetic valve (I-6-4), a sixth two-position two-port electromagnetic valve (I-6-5), a seventh two-position two-port electromagnetic valve (I-6-6), an eighth two-position two-port electromagnetic valve (I-6-7), a ninth two-position two-port electromagnetic valve (I-6-8), a tenth two-position two-port electromagnetic valve (I-6-9), an eleventh two-position two-port electromagnetic valve (I-6-10), a twelfth two-position two-port electromagnetic valve (I-8), a thirteenth two-position two-port electromagnetic valve (I-10), and a fourteenth two-position two-port electromagnetic valve (I-13), a throttle valve (I-11), a flowmeter (I-12), a vacuum pressure gauge (I-7), a gas buffer cavity (I-9); the multi-point centralized auto-sampling system (I(b)) is located at a lower right of the odor electronic nose instrument (I); wherein the computer control and data analyzing system (I(c)) comprises a computer mainboard (I-15), a data acquisition card (I-16), a monitor (I-17), a drive and control circuit module (I-18), a precision linear and switching power module (I-19), a hard disk, a network card, a video card; the computer control and data analyzing system (I(c)) is located on a left side of the odor electronic nose instrument (I); wherein the multi-point centralized auto-sampling system (I(b)) has a gas sampling period of T.sub.0=180-300 s for a malodorous gas at a single monitoring point, with a default value T.sub.0=240 s, so the gas sensor array (I-1) generates a 16-dimensional response vector for the single monitoring point; according to the 16-dimensional response vector, using, by the computer control and data analyzing system (I(c)) is configured to use a cascade machine learning model to perform real-time analysis and prediction of an olfactory concentration value of the malodorous gas at the single monitoring point, concentrations of eight compounds specified in a China national standard GB14554: NH.sub.3, H.sub.2S, CS.sub.2, C.sub.3H.sub.9N, CH.sub.4S, C.sub.2H.sub.6S, C.sub.2H.sub.6S.sub.2, C.sub.8H.sub.8, and concentrations of SO.sub.2 and the total volatile organic compound (TVOC) specified in a China national standard GB/T18883, totaling 10+1 items of the malodorous gas at the single monitoring point; and finally transmit, through wireless Internet, monitoring data and prediction results to the central control room (VII) and designated ones of the plurality of stationary/mobile terminals (VIII); the odor electronic nose instrument (I) is configured to obtain; the 16-dimensional response vector every single gas sampling period T.sub.0, which is stored in a data file of a computer hard disk, use 10 two-position two-port electromagnetic valves from the second two-position two-port electromagnetic valve (I-6-1) to the eleventh two-position two-port electromagnetic valve (I-6-10) to sequentially control on-and-off of the malodorous gases at 10 monitoring points within a 4 km.sup.2 area, online monitor the malodorous gases at the 10 monitoring points by using the cyclic sampling period of T=10T.sub.0 to obtain 10 monitoring data results, and sequentially store the 10 monitoring data results in 10 data files, wherein the 10 monitoring data results are a numerical basis for the odor electronic nose instrument (I) to realize cyclically online quantitative prediction of 10+1 concentration control index values of the malodorous gases; wherein the gas sampling period T.sub.0 comprises following five stages: an initial recovery stage of the gas sensor array lasting 95-215 s, an accurate calibration stage of clean air lasting 30 s, a balance stage lasting 5 s, a stage of a headspace sampling of the malodorous gases lasting 30 s and a flushing stage of purified ambient air lasting 20 s; in the gas sampling period T.sub.0, under a control of a computer, a two-position two-port electromagnetic valve (I-6-k, k=1, 2, . . . , 10) corresponding to one of the 10 monitoring points is connected, and another nine two-position two-port electromagnetic valves corresponding to another nine of the 10 monitoring points are disconnected; the internal miniature vacuum pump (I-14) draws a malodorous gas into the gas buffer cavity (I-9) with a flow rate of 1,000 ml/min, thereby enabling the malodorous gas to flow through the annular working chamber of the gas sensor array and skim over surfaces of the sensitive films of the gas sensor array, so that the gas sensor array generates sensitive responses lasting 30 s; from a beginning of the balance stage, the computer control and data analyzing system (1(c)) continuously records sensitive response data, wherein the sensitive response data comprises response data of the gas sensor array in following three stages, which last 45 s: the balance stage lasting 5 s, the stage of the headspace sampling of the malodorous gases lasting 30 s, and the flushing stage of the purified ambient air lasting first 10 s, which are temporarily stored in a text file; response data of other time slots in the gas sampling period T.sub.0 are not recorded; in the response data of 45 s, a difference between a steady-state maximum value and a minimum value of a response curve of a single gas sensor is extracted as a response component, so that the gas sensor array generates the 16-dimensional response vector; in 10 s after the end of the data recording, that is, a later 10 s of the purified ambient air flushing stage, the computer control and data analyzing system (I(c)) predicts the 10+1 concentration control index values of the malodorous gases based on the 16-dimensional response vector; wherein the odor electronic nose instrument (I) is configured to perform long-term online monitoring of multiple monitoring points and online prediction of various concentration control index values of the malodorous gases in the pollution sites by executing the following operations: (1), a power-on operation: when the odor electronic nose instrument (I) is preheated for 30 minutes and an Air purifier on option on a screen menu is clicked, triggering the ambient air purification device (IV) to start purifying an ambient air where the odor electronic nose instrument (I) is located, and keeping the ambient air purification device (IV) working until an Air purifier off option is clicked; under a drawing action of the internal miniature vacuum pump (I-14) inside the ambient air purification device (IV), making the purified ambient air sequentially flow through the first two-position two-port electromagnetic valve (I-5), the annular working chamber of the gas sensor array (I-1) and the thirteenth two-position two-port electromagnetic valve (I-10) with a flow rate of 6,500 ml/min, and then be discharged to outdoor; wherein a temperature in the annular working chamber of the gas sensor array (I-1) reaches a constant 550.1 C. from a room temperature; when an External vacuum pump on option on the screen menu is clicked; drawing, by the external vacuum pump (III) with a drawing flow rate of 250-280 l/min and a limit vacuum degree of 100-120 mbar, a malodorous gas at a certain monitoring point to the odor electronic nose instrument (I) with a linear distance of up to 2.5 km to the odor electronic nose instrument (I) within 1 minute through a stainless steel pipe with an inner diameter of 10 mm, and then making the malodorous gas at the certain monitoring point flow through a corresponding two-position two-port electromagnetic valve (I-6-k) k=1, 2, . . . , 10, the vacuum pressure gauge (I-7) and the gas buffer cavity (I-9) and be directly discharged to the outdoor; keeping, by the external vacuum pump (III) draw the malodorous gas at the certain monitoring point until an external vacuum pump off option on the screen menu is clicked; modifying a setting of the gas sampling period T.sub.0 on the screen menu as a default value T.sub.0=4 minutes and modifying a cyclic sampling period of the malodorous gases at the 10 monitoring points as T=10T.sub.0; (2), an operation for starting a cyclic sampling period of the malodorous gases: when a start detection button on the screen menu is clicked, sequentially monitoring, by the odor electronic nose instrument (I), the 10 monitoring points, and automatically generating, by the computer control and data analyzing system (I(c)), 10 text files in a designated folder to store the response data of the gas sensor array (I-1) to the malodorous gases at the 10 monitoring points; (3), an operation for starting a single sampling period of a malodorous gas at a monitoring point k; taking T.sub.0=240 seconds as an example, wherein k=1, 2, . . . , 10: (3.1), a preliminary recovery operation of the gas sensor array (I-1): in the 0-155 s of the gas sampling period T.sub.0, under the drawing action of the internal miniature vacuum pump (I-14) inside the ambient air purification device (IV), making the purified ambient air sequentially flow through the first two-position two-port electromagnetic valve (I-5), the annular working chamber of the gas sensor array (I-1), the thirteenth two-position two-port electromagnetic valve (I-10) with a flow rate of 6,500 ml/min, and then be discharged to the outdoor; so that the accumulated heat in the annular working chamber of the gas sensor array (I-1) is taken away under an action of the 6,500 ml/min purified ambient air, the malodorous gas molecules adhered to the sensitive membrane surfaces of the gas sensor array (I-1) and the inner walls of pipelines are preliminarily washed away, and the gas sensor array (I-1) preliminarily returns to a reference state which lasts 155 s; wherein among 10 two-position two-port electromagnetic valves from the second two-position two-port electromagnetic valve (I-6-1) to the eleventh two-position two-port electromagnetic valve (I-6-10), only one two-position two-port electromagnetic valve (I-6-k) is on, another nine two-position two-port electromagnetic valves are off, and the external vacuum pump (III) draws the malodorous gas at the monitoring point k into the odor electronic nose instrument (I), wherein k=1, 2, . . . , 10; (3.2), an accurate calibration operation of the gas sensor array by clean air: in 156-185 s range of the gas sampling period T.sub.0, the fourteenth two-position two-port electromagnetic valve (I-13) is on, the first two-position two-port electromagnetic valve (I-5), the twelfth two-position two-port electromagnetic valve (I-8), and the thirteenth two-position two-port electromagnetic valve (I-10) are off, and among the 10 two-position two-port electromagnetic valves from the second two-position two-port electromagnetic valve (I-6-1) to the eleventh two-position two-port electromagnetic valve (I-6-10), only the one two-position two-port electromagnetic valve (I-6-k) is on, the other nine two-position two-port electromagnetic valves are off; under the drawing action of the internal miniature vacuum pump (I-14) inside the ambient air purification device (IV), making the clean air sequentially flow through the fourteenth two-position two-port electromagnetic valve (I-13), the gas pipelines, the annular working chamber of the gas sensor array (I-1), the throttle valve (I-11), the flowmeter (I-12) and the internal miniature vacuum pump (I-14) with the flow rate of 1,000 ml/min, and then be discharged to the outdoor; wherein the clean air makes the gas sensor array (I-1) accurately return to the reference state which lasts 30 s; and the external vacuum pump (III) keeps drawing for 30 seconds; (3.3), an operation of a balance stage: in 186-190 s of the gas sampling period T.sub.0, the first two-position two-port electromagnetic valve (I-5), the twelfth two-position two-port electromagnetic valve (I-8), the thirteenth two-position two-port electromagnetic valve (I-10) and the fourteenth two-position two-port electromagnetic valve (I-13) are disconnected, among the 10 two-position two-port electromagnetic valves from the second two-position two-port electromagnetic valve (I-6-1) to the eleventh two-position two-port electromagnetic valve (I-6-10), only the one two-position two-port electromagnetic valve (I-6-k) is on, the other nine two-position two-port electromagnetic valves are off; and there is no gas flow in the annular working chamber of the gas sensor array (I-1); from 186.sup.th second of the gas sampling period T.sub.0, recording and storing, by the computer control and data analyzing system (I(c)), real-time response data of the gas sensor array (I-1) in a designated temporary text file temp.txt; wherein the external vacuum pump (III) keeps drawing the malodorous gas for 5 seconds, (3.4), the headspace sampling operation of the malodorous gas at the monitoring point k: in 190-220 seconds of the gas sampling period T.sub.0, the twelfth two-position two-port electromagnetic valve (I-8) is on, the first two-position two-port electromagnetic valve (I-5), the fourteenth two-position two-port electromagnetic valve (I-13) and the thirteenth two-position two-port electromagnetic valve (I-10) are off, and among the 10 two-position two-port electromagnetic valves from the second two-position two-port electromagnetic valve (I-6-1) to the eleventh two-position two-port electromagnetic valve (I-6-10), only the one two-position two-port electromagnetic valve (I-6-k) is on, the other nine two-position two-port electromagnetic valves are off; under the drawing action of the internal miniature vacuum pump (I-14) inside the ambient air purification device (IV), making the malodorous gas in the gas buffer cavity (I-9) sequentially flow through the annular working chamber of the gas sensor array (I-1), the throttle valve (I-11), the flowmeter (I-12), the internal miniature vacuum pump (I-14) with a flow rate of 1,000 ml/min, and be finally discharged to the outdoor; wherein the sensitive responses of the gas sensor array (I-1) are recorded in the temporary file temp.txt, and the external vacuum pump (III) keeps drawing for 30 seconds; (3.5), a flushing operation of the gas sensor array: in the 221-230 seconds of the gas sampling period T.sub.0, the first two-position two-port electromagnetic valve (I-5) and the thirteenth two-position two-port electromagnetic valve (I-10) are connected, and the twelfth two-position two-port electromagnetic valve (I-8) and the fourteenth two-position two-port electromagnetic valve (I-13) are disconnected; under the drawing action of the internal miniature vacuum pump (I-14) inside the ambient air purification device (IV), making the purified ambient air sequentially flow through the first two position two-port solenoid valve (I-5), the annular working chamber of the gas sensor array (I-1) and the thirteenth two-position two-port electromagnetic valve (I-10) with a flow rate of 6,500 ml/min, and then be discharged to the outdoor; at the same time, among the 10 two-position two-port electromagnetic valves from the second two-position two-port electromagnetic valve (I-6-1) to the eleventh two-position two-port electromagnetic valve (I-6-10), only one two-position two-port electromagnetic valve (I-6-(k+1)) is on, another nine two-position two-port electromagnetic valves are off, and drawing, by the external vacuum pump (III), a malodorous gas at a monitoring point (k+1); due to the role of the purified ambient air, the accumulated heat in the annular working chamber of the gas sensor array (I-1) is taken away, the malodorous gas molecules adhered to the sensitive film surfaces of the gas sensor array (I-1) and the inner walls of the pipelines are preliminarily washed away, and the gas sensor array (I-1) gradually returns to the reference state, which takes 20 seconds, wherein: (a), in the 221-230 seconds of the gas sampling period T.sub.0, continuing to record the response data of the gas sensor array in the temporary file temp.txt which lasts 10 s; at the end of 230.sup.th second, stopping, by the computer control and data analyzing system (I(c)), recording the response data of the gas sensor array; (b), in the 231-240 seconds of the gas sampling period T.sub.0, performing, by the computer control and data analyzing system (I(c)), following three operations: (b1), a feature extraction: from the 231.sup.st second, extracting the maximum and the minimum steady-state response values of each gas sensor with a time duration of 45 seconds from the temporary file temp.txt, and taking and recording a difference between the maximum and the minimum response value as a characteristic response component x.sub.i(t) of each gas sensor to the malodorous gas at the monitoring point k at the current time t in a corresponding data file, wherein i=1, 2, . . . , 16; (b2), a response prediction of the gas sensor array: realizing, by the first level of the cascade machine learning model, an online self-learning action according to the time-series response vectors of the gas sensor array within a period that has occurred before the current time t, and predicting responses of the gas sensor array (I-1) in three future time points of T.sub.0, 2T.sub.0 and 3T.sub.0; wherein the first level of the cascade machine learning model is 16*3 convolutional neural networks and the period comprises three time segments of [t18, t], [t19, t1] and [t20, t2]; (b3), a prediction of concentration control index values of the malodorous gases: continuing to predict, by the second level of the cascade machine learning model, the 10+1 concentration control index values of the malodorous gas at the monitoring point k according to the response values of the gas sensor array predicted by the 16*3 convolution neural networks in the first level of the cascade machine learning model, showing the 10+1 concentration control index values on the monitor, and transmitting monitoring and prediction results to a central control room (VII) and the plurality of stationary/mobile terminals (VIII) through the Internet network wherein the second level of the cascade machine learning model is 10+1 deep neural networks; (3.6), an ending operation of the gas sampling period T.sub.0 at the monitoring point k: kk+1, returning to the step (3.1), and starting the gas sampling period T.sub.0 at the monitoring point k+1; if k+1>10, then starting to detect a malodorous gas at a monitoring point k=1 of a next gas sampling period; (4), repeating the steps (3.1)(3.6); so that the odor electronic nose instrument (I) realizes cyclically online measurement, identification and prediction of 10+1 concentration control index values of the malodorous gases at the 10 monitoring points.

2. (canceled)

3. The system according to claim 1, wherein the gas sensor array (I-1) is comprised of 11 metal oxide semiconducting (MOS) elements, 4 electrochemical (EC) elements and a photo ionization detector (PM); wherein the 11 MOS elements are configured to detect a plurality of organic/inorganic compounds; the 4 EC elements are configured to detect 4 inorganic compounds: NH.sub.3, H.sub.2S, CS.sub.2 and SO.sub.2; the PID is configured to detect the TVOC.

4. The system according to claim 1, wherein the online multi-point centralized monitoring and analysis system is operative to realize online monitoring and analysis of multi-point malodorous gases in a certain specific area; and 10 monitoring points are set in a maximum area of 2 km*2 km=4 km.sup.2, including 9 stationary monitoring points and 1 mobile monitoring point; the odor electronic nose instrument (I) is located indoor, which connects with each of the 10 monitoring points through a stainless steel pipe with an inner diameter of 10 mm; each gas sampling head is in a form of a water tap, is connected to a commercial dedusting, dehumidification and purification part, and is installed or moved to a designated position; when a monitoring point is changed, the stainless steel pipe is relayed and the gas sampling head is re-installed and re-moved to the designated position.

5. The system according to claim 1, wherein eight or more of the 10 monitoring points are arranged around a boundary of a specified area, and a target is to make a total length of stainless steel pipelines between the odor electronic nose instrument (I) and the 10 monitoring points be a shortest value; for an area with accessible paths, the odor electronic nose instrument (I) is configured to be arranged indoor in a center of the area, wherein the area with accessible paths comprises a chemical industrial park, a residential area and other areas with accessible paths; for an area without accessible paths, the odor electronic nose instrument (I) is configured to be arranged indoor at a boundary of the area, wherein the area without accessible paths comprises a landfill, a sewage treatment plant and other areas without accessible paths.

6. The system according to claim 1, wherein the external vacuum pump (III) has a drawing rate of 250-280 l/min, a limit vacuum degree of 100-120 mbar, and is operative to work continuously for a long period of time; the external vacuum pump (III) is configured to draw a malodorous gas at one of the 10 monitoring points with a linear distance of 2.5 km into the odor electronic nose instrument (I) through a stainless steel pipe of a 10 mm inner diameter within less than 1 min; and in the gas sampling period T.sub.0, except for a 30 s headspace sampling stage, the external vacuum pump (III) is configured to make the malodorous gas flow into the odor electronic nose instrument (I) and be discharged to the outdoor directly, but not flow through the annular working chamber of the gas sensor array (I-1).

7. The system according to claim 1, wherein the gas buffer cavity (I-9) has a size of 40 mm*5 mm and is set inside the odor electronic nose instrument (I); a flow rate of the malodorous gas measured in the gas buffer cavity (I-9) is 16 times lower than a flow rate in the stainless steel pipe with an inner diameter of 10 mm; only at a 30 s headspace sampling stage, the internal miniature vacuum pump (I-14) is configured to draws the malodorous gas in the gas buffer cavity (I-9) into the annular working chamber of the gas sensor array (I-1), such that the gas sensor array (I-1) generates a sensitive response; wherein the malodorous gases drawn by the internal miniature vacuum pump (I-14) are always fresh.

8. The method system according to claim 1, wherein before a headspace sampling stage of the malodorous gas, an accurate calibration stage of clean air, lasting 30 s with a flow rate of 1,000 ml/min, makes multiple perceptions of the gas sensor array (I-1) for the malodorous gases on a same baseline; a standard volume of a 12-15 MPa clean air cylinder (V) is 40 L, and the clean air is 6 m.sup.3 when the standard volume is converted to standard temperature and pressure; when the gas sampling period T.sub.0=3, 4 and 5 minutes, a bottle of 40L compressed clean air is respectively used for 25, 33 and 41 days; and an outdoor ambient air where the odor electronic nose instrument (I) is located is first purified by the ambient air purification device (IV), and then is used to flush the gas sensor array (I-1), so as to primarily restore the gas sensor array (I-1) to a reference state and reduce an operation cost.

9. The system according to claim 1, wherein a set of big data of the malodorous gases comprises: (1), online detection data monitored by the gas sensor array (I-1) for a large number of malodorous pollutants in chemical industrial parks including fragrance and flavor factories, pharmaceutical factories, landfill sites, sewage treatment plants, farm and neighboring residential areas; (2), off-line laboratory test data monitored by the gas sensor array (I-1) for a large number of headspace volatile gases of standard malodorous samples, including 5 standard odorants specified in a China national standard GB/T14675, the standard malodorous samples made up of nine single-component malodorous pollutants with different concentrations designated by GB14554: C.sub.3H.sub.9N, C.sub.8H.sub.8, H.sub.2S, CH.sub.4S, C.sub.2H.sub.6S, C.sub.2H.sub.6S.sub.2, NH.sub.3, CS.sub.2, and SO.sub.2 by GB/T18883, and standard malodorous samples of mixed components prepared with different concentrations of multiple single compounds, wherein the 5 standard odorants are -phenylethanol, isovaleric acid, methylcyclopentanone, peach aldehyde and -methylindole; (3), off-line panel evaluation data of olfactory concentration values specified in GB/T14675 and a China industry standard HJ 905-2017 for the malodorous gases sampled by vacuum bottles or bags at a large number of malodorous sites and immediately transported back to laboratories; (4), off-line TVOC data by gas chromatography, and off-line SO.sub.2 data obtained by spectrophotometry, depending upon the malodorous pollutants in Tenax GC/TA adsorption tubes sampled on sites according to GB/T18883; (5), off-line laboratory test data of 8 malodorous components specified in China national standards from GB/T14676 to GB/T14680 by gas chromatography, mass spectrometry and spectrophotometry for the on-site sampling malodorous pollutants; and (6), residents' complaint data in vicinities of malodorous pollution sources.

10. The system according to claim 1, wherein the odor electronic nose instrument (I) is configured to use the cascade machine learning model to predict olfactory concentration values of the malodorous gases and several specified concentration control index values of malodorous gases in time points of t+1, t+2 and t+3 in a near future; wherein a first level of the cascade machine learning model, is responsible for predicting responses of the gas sensor array (I-1) to the malodorous gases in the time points of t+1, t+2 and t+3, based on occurred time-series responses of the gas sensor array (I-1) at a current time t and a recent past, wherein the first level of the cascade machine learning model is a layer of convolutional neural networks; wherein a second level of the cascade machine learning model, further predicts the olfactory concentration values of the malodorous gases and multiple specified concentration control index values of various malodorous gases in the time points of t+1, t+2 and t+3, based on long-term accumulation of big data of the malodorous gases and the prediction values of the first-level, wherein the second level of the cascade machine learning model is a layer of deep neural networks.

11. The system according to claim 1, wherein according to a divide-and-conquer strategy, a first level of the cascade machine learning model is configured to use 16*3 groups of single-output single-hidden-layer convolution neural networks to predict responses of each gas sensor in the time points of t+1, t+2 and t+3; for a single gas sampling period of T.sub.0=4 minutes, predict responses at 40, 80 and 120 minutes in a coming future from a current time t; when three single-output single-hidden-layer convolution neural network modules with the gas sampling period of T.sub.0=4 minutes are used to respectively predict the responses of a gas sensor i in the time points of t+1, t+2 and t+3: (a), a single-output single-hidden-layer convolution neural network CNN.sub.i1 is configured to predict, a response of a gas sensor i in the time point of t+1: if the convolutional neural network CNN.sub.i1 is used to learn 18 time-series response data of the gas sensor i that have occurred before the current time t, a delay length t=9, then a number of input nodes is m.sub.i=9, a number of hidden nodes is h.sub.i=5, and a number of output nodes is n.sub.i=1; a preprocessed time-series response data set X.sub.i1 of the gas sensor i learned online by the convolutional neural network CNN.sub.i1 is: $X_{i.Math..Math.1}=\left(\begin{array}{ccccc}{x}_i\left(t-9\right)& x_i\left(t-8\right)& x_i\left(t-7\right)& x_i\left(t-6\right)& x_i\left(t-5\right)\\ x_i\left(t-4\right)& x_i\left(t-3\right)& x_i\left(t-2\right)& x_i\left(t-1\right).Math.& \\ x_i\left(t-10\right)& x_i\left(t-9\right)& x_i\left(t-8\right)& x_i\left(t-7\right)& x_i\left(t-6\right)\\ x_i\left(t-5\right)& x_i\left(t-4\right)& x_i\left(t-3\right)& x_i\left(t-2\right)& \\ & & .Math.\mathrm{.Math.}& & \\ x_i\left(t-17\right)& x_i\left(t-16\right)& x_i\left(t-15\right)& x_i\left(t-14\right)& x_i\left(t-13\right)\\ x_i\left(t-12\right)& x_i\left(t-11\right)& x_i\left(t-10\right)& x_i\left(t-9\right)& \\ x_i\left(t-18\right)& x_i\left(t-17\right)& x_i\left(t-16\right)& x_i\left(t-15\right)& x_i\left(t-14\right)\\ x_i\left(t-13\right)& x_i\left(t-12\right)& x_i\left(t-11\right)& x_i\left(t-10\right)& \end{array}\right)R^{109}$ a target output d.sub.i1 is:
d.sub.i1=(x.sub.i(t)x.sub.i(t1)x.sub.i(t2)x.sub.i(t3)x.sub.i(t4)x.sub.i(t5)x.sub.i(t6)x.sub.i(t7)x.sub.i(t8)x.sub.i(t9).sup.TR.sup.10 the convolutional neural network CNN.sub.i1 is configured to learn a 18 dimensional time-series responses of the gas sensor i that has occurred in last 12 hours, generate ten 9-dimensional time-series response samples, wherein the ten 9-dimensional time-series response samples represent that a number of samples is N.sub.i1=10; when the activation functions of hidden and output layers in the CNN.sub.i1 are a modified Sigmoid function ()=3/(1+exp(/3)), and an error back propagation algorithm is adopted, wherein a learning factor is .sub.i=5/N.sub.i1=0.2; the data set X.sub.i1 and the target output d.sub.i1 are transformed to a range of [0, 3]; by undergoing an online learning process of 10 seconds, the convolutional neural network CNN.sub.i1 is configured to predict, a response x.sub.i(t+1) of the gas sensor i in the time point of t+1, based on a following 9-dimensional time-series response in a latest time period:
x.sub.i1=(x.sub.i(t8)x.sub.i(t7)x.sub.i(t6)x.sub.i(t5)x.sub.i(t4)x.sub.i(t3)x.sub.i(t2)x.sub.i(t1)x.sub.i(t).sup.TR.sup.9 when T.sub.0=4 minutes, the convolutional neural network CNN.sub.i1 is configured to predict a response of the gas sensor i in a coming 40 minutes; (b), two single-output single-hidden-layer convolution neural networks CNN.sub.i2 and CNN.sub.i3 are configured to predict, responses of the gas sensor i in the time points of t+2 and t+3: structures of the convolutional neural networks CNN.sub.i2 and CNN.sub.i3 are m.sub.i=9, h.sub.i=5, and n.sub.i=1; and pre-processed data sets X.sub.i2 and X.sub.i3 for online learning are respectively: $X_{i.Math..Math.2}=\left(\begin{array}{ccccc}{x}_i\left(t-10\right)& x_i\left(t-9\right)& x_i\left(t-8\right)& x_i\left(t-7\right)& x_i\left(t-6\right)\\ x_i\left(t-5\right)& x_i\left(t-4\right)& x_i\left(t-3\right)& x_i\left(t-2\right).Math.& \\ x_i\left(t-11\right)& x_i\left(t-10\right)& x_i\left(t-9\right)& x_i\left(t-8\right)& x_i\left(t-7\right)\\ x_i\left(t-6\right)& x_i\left(t-5\right)& x_i\left(t-4\right)& x_i\left(t-3\right)& \\ & & .Math.\mathrm{.Math.}& & \\ x_i\left(t-18\right)& x_i\left(t-17\right)& x_i\left(t-16\right)& x_i\left(t-15\right)& x_i\left(t-14\right)\\ x_i\left(t-13\right)& x_i\left(t-12\right)& x_i\left(t-11\right)& x_i\left(t-10\right)& \\ x_i\left(t-19\right)& x_i\left(t-18\right)& x_i\left(t-17\right)& x_i\left(t-16\right)& x_i\left(t-15\right)\\ x_i\left(t-14\right)& x_i\left(t-13\right)& x_i\left(t-12\right)& x_i\left(t-11\right)& \end{array}\right)R^{109}$ and $X_{i.Math..Math.3}=\left(\begin{array}{ccccc}{x}_i\left(t-11\right)& x_i\left(t-10\right)& x_i\left(t-9\right)& x_i\left(t-8\right)& x_i\left(t-7\right)\\ x_i\left(t-6\right)& x_i\left(t-5\right)& x_i\left(t-4\right)& x_i\left(t-3\right).Math.& \\ x_i\left(t-12\right)& x_i\left(t-11\right)& x_i\left(t-10\right)& x_i\left(t-9\right)& x_i\left(t-8\right)\\ x_i\left(t-7\right)& x_i\left(t-6\right)& x_i\left(t-5\right)& x_i\left(t-4\right)& \\ & & .Math.\mathrm{.Math.}& & \\ x_i\left(t-19\right)& x_i\left(t-18\right)& x_i\left(t-17\right)& x_i\left(t-16\right)& x_i\left(t-15\right)\\ x_i\left(t-14\right)& x_i\left(t-13\right)& x_i\left(t-12\right)& x_i\left(t-11\right)& \\ x_i\left(t-20\right)& x_i\left(t-19\right)& x_i\left(t-18\right)& x_i\left(t-17\right)& x_i\left(t-16\right)\\ x_i\left(t-15\right)& x_i\left(t-14\right)& x_i\left(t-13\right)& x_i\left(t-12\right)& \end{array}\right)R^{109}$ wherein X.sub.i2 and X.sub.i3 have ten 9-dimensional time-series response samples respectively, and have a same number of samples of N.sub.i2=N.sub.i3=N.sub.i1=10; target outputs of the CNN.sub.i2 and CNN.sub.i3 in a learning phase and depended time-series responses of the CNN.sub.i2 and CNN.sub.i3 in a prediction phase are similarly to the target output of the CNN.sub.i1 in a learning phase and depended time-series response of the CNN.sub.i1 in a prediction phase; when T.sub.0=4 minutes, the two single-output single-hidden-layer convolution neural networks CNN.sub.i2 and CNN.sub.i3 are configured to learn the responses of the gas sensor i during 12 hours before 40 and 80 minutes, and respectively predict the responses x.sub.i(t+2) and x.sub.i(t+3) of the gas sensor i in the time points of t+2 and t+3, wherein predicting the responses x.sub.i(t+2) and x.sub.i(t+3) of the gas sensor i in the time points of t+2 and t+3 represents that predicting the responses of the gas sensor i in next 80 and 120 minutes.

12. The system according to claim 1, wherein according to a divide-and-conquer strategy, an overall prediction problem of 10+1 concentration control index values of the malodorous gases, including NH.sub.3, H.sub.2S, CS.sub.2, C.sub.3H.sub.9N, CH.sub.4S, C.sub.2H.sub.6S, C.sub.2H.sub.6S.sub.2, C.sub.8H.sub.8, SO.sub.2, TVOC and the olfactory concentration values of the malodorous gases is divided into 11 single concentration prediction problems; a second level of the cascade machine learning model is configured to use 10+1 single-output three-hidden-layer deep neural network modules to predict a 10+1 malodorous pollution control index values; wherein a training set of a single-output deep neural network is big data online detected by the gas sensor array (I-1) of the odor electronic nose instrument (I) for standard malodorous liquid/gas samples and a large number of malodorous pollutants; wherein target outputs are the olfactory evaluation values, off-line measurement values of conventional instruments such as a gas chromatographer, mass spectrometer and spectrophotometer, and data of residents' complaints; each single-output three-hidden-layer deep neural network DNN.sub.j is configured to adopt a bottom-up off-line learning manner; wherein parameters of a first hidden layer and a second hidden layer are determined by a single-hidden-layer peer-to-peer neural network, wherein the single-hidden-layer peer-to-peer neural network represents that weights of a hidden-to-output layer are directly equal to weights of an input-to-hidden layer and the target outputs are directly equal to input values; wherein input and output components are proportionally transformed to a range of [0, 3]; wherein an activation function of hidden units of each single-hidden-layer peer-to-peer neural network are modified sigmoid functions ()=3/(1+exp(/3)), an error back-propagation algorithm is adopted, a learning factor is .sub.j=1/N.sub.j, and the hidden-to-output layer is discarded after ending a learning operation, wherein N.sub.j is a number of samples in the big data of the malodorous gases; a j.sup.th single-output deep neural network DNN.sub.j is configured to, based on the predicted responses of 16 convolutional neural networks to the gas sensor array (I-1) in the time point of t+1, {x.sub.1(t+1), x.sub.2(t+1), . . . , x.sub.16(t+1)}, predict a j.sup.th concentration index value y.sub.j(t+1) of the malodorous gas in the time point of t+1; the DNN.sub.j is configured to, according to the predicted responses of 16 convolutional neural networks, {x.sub.1(t+2), x.sub.2(t+2), . . . , x.sub.16(t+2)} and {x.sub.1(t+3), x.sub.2(t+3), . . . , x.sub.16(t+3)}, respectively predict j.sup.th concentration index values y.sub.j(t+2) and y.sub.j(t+3) of the malodorous gas in the time points of t+2 and t+3; if an actual input is a current response vector of the gas sensor array, x(t)=(x.sub.1(t), x.sub.2(t), . . . , x.sub.16(t)).sup.T, temperature and humidity values at the current time t are added if necessary, wherein an actual output of DNN.sub.j is an estimation of a current concentration value y.sub.j(t) of a component j of the malodorous gas.

13. (canceled)

Description

BRIEF DESCRIPTION OF THE DRAWINGS

[0114] FIG. 1 illustrates a block diagram of the relationship among odor electronic nose instrument development, cascade machine learning model and algorithm, and online detection and prediction of malodorous pollutants according to the present disclosure named an online centralized monitoring and analysis system based on an electronic nose instrument for multi-point malodorous gases.

[0115] FIG. 2 illustrates a schematic diagram illustrating the working principle of the odor electronic nose instrument and multi-point centralized monitoring and analysis system for malodorous pollution sites according to the present disclosure named an online centralized monitoring and analysis system based on an electronic nose instrument for multi-point malodorous gases.

[0116] FIG. 3 illustrates a schematic diagram illustrating the working principle of the odor electronic nose instrument (in headspace sampling state) according to the present disclosure named an online centralized monitoring and analysis system based on an electronic nose instrument for multi-point malodorous gases.

[0117] FIG. 4a illustrates a schematic diagram illustrating the mutual position relationship of the odor electronic nose instrument and multiple monitoring points in an area with road access according to the present disclosure named an online centralized monitoring and analysis system based on an electronic nose instrument for multi-point malodorous gases.

[0118] FIG. 4b illustrates a schematic diagram illustrating the mutual position relationship of the odor electronic nose instrument and multiple monitoring points in an area without road access according to the present disclosure named an online centralized monitoring and analysis system based on an electronic nose instrument for multi-point malodorous gases.

[0119] FIG. 5a illustrates a schematic diagram illustrating composition units of a gas sensor array according to the present disclosure named an online centralized monitoring and analysis system based on an electronic nose instrument for multi-point malodorous gases.

[0120] FIG. 5b illustrates a schematic diagram illustrating an annular working chamber cover according to the present disclosure named an online centralized monitoring and analysis system based on an electronic nose instrument for multi-point malodorous gases.

[0121] FIG. 5c illustrates a schematic diagram illustrating a sectional drawing of an annular working chamber according to the present disclosure named an online centralized monitoring and analysis system based on an electronic nose instrument for multi-point malodorous gases using an odor electronic nose instrument.

[0122] FIG. 6 illustrates a schematic diagram illustrating an odor buffer cavity according to the present disclosure named an online centralized monitoring and analysis system based on an electronic nose instrument for multi-point malodorous gases.

[0123] FIG. 7 illustrates an on-off change diagram of 14 two-position two-port electromagnetic valves (in seconds) with a single gas sampling period of T.sub.0=240s and a cyclic gas sampling period T=10T.sub.0 according to the present disclosure named an online centralized monitoring and analysis system based on an electronic nose instrument for multi-point malodorous gases.

[0124] FIG. 8 illustrates a schematic diagram illustrating the three-dimensional appearance of the odor electronic nose instrument according to the present disclosure named an online centralized monitoring and analysis system based on an electronic nose instrument for multi-point malodorous gases.

[0125] FIG. 9 illustrates a schematic diagram illustrating a back of the odor electronic nose instrument according to the present disclosure named an online centralized monitoring and analysis system based on an electronic nose instrument for multi-point malodorous gases.

[0126] FIG. 10 illustrates a schematic diagram illustrating a convolutional neural network CNN.sub.i1 to predict the response x.sub.i(t+1) of a gas sensor i at the time point of t+1 (i.e. the 40.sup.th minute in the future) according to the present disclosure named an online centralized monitoring and analysis system based on an electronic nose instrument for multi-point malodorous gases.

[0127] FIG. 11a illustrates a learning process diagram of a k.sup.th layer in a deep neural network DNN.sub.j, where a structure and parameters of a single-hidden-layer peer-to-peer neural network, according to the present disclosure named an online centralized monitoring and analysis system based on an electronic nose instrument for multi-point malodorous gases.

[0128] FIG. 11b illustrates a learning process diagram of a k.sup.th layer in a deep neural network DNN.sub.j, where a reserved structure after a learning of a peer-to-peer neural network, according to the present disclosure named an online centralized monitoring and analysis system based on an electronic nose instrument for multi-point malodorous gases.

[0129] FIG. 12a illustrates a schematic diagram illustrating a concentrations of multiple malodorous pollutants at the time point of t+1 (say the 40.sup.th minute in the future) with a first level of a cascade machine learning model, i.e. a convolution neural network layer, according to the present disclosure named an online centralized monitoring and analysis system based on an electronic nose instrument for multi-point malodorous gases.

[0130] FIG. 12b illustrates a schematic diagram illustrating a concentrations of multiple malodorous pollutants at the time point of t+1 (say the 40.sup.th minute in the future) with a second level of a cascade machine learning model, i.e. the deep neural network (DNN) layer, according to the present disclosure named an online centralized monitoring and analysis system based on an electronic nose instrument for multi-point malodorous gases.

DETAILED DESCRIPTION

[0131] The detailed description of the present disclosure is further given below in conjunction with the above drawings.

[0132] FIG. 1 is a block diagram illustrating the relationship of odor electronic nose instrument development, cascade machine learning model and algorithm, and online detection and prediction of malodorous pollutants.

[0133] The disclosure first analyzes the characteristics of malodorous pollutants and gas sensors from the chemical and physical perspectives. The components of malodorous gases are numerous and complex, often containing dozens or even hundreds of components, including both organic and inorganic. Some malodorous components contribute a lot to the dimensionless malodorous concentrations, but the real concentrations of the components may be very low, so the responses of a gas sensor are very small; some malodorous components contribute a little to the dimensionless malodorous concentrations, but the real concentrations may be very high, so the response of the gas sensor are very large; and vice versa. Considering the factors of sensitivity, selectivity, response speed, stability, commercialization, miniaturization, service life, cost and so on, the present disclosure selects MOS-type, EC-type and PID-type gas sensors to form a small-sized gas sensor array module. In order to avoid the wind, sun, and rain outside the monitoring areas, this disclosure proposes a centralized monitoring mode of multi-point malodorous gases with the key parts located indoors and develops an odor electronic nose instrument accordingly. In consideration of the complex composition of malodorous pollutants and the variable environment of the monitoring sites, this disclosure further proposes to establish the odor big data, and presents a new cascade machine learning model to realize online monitoring and prediction of various malodorous pollutants.

[0134] According to FIG. 1, the big data of malodorous gases include: (1) The off-line laboratory test data of the gas sensor array (I-1) of the odor electronic nose instrument for a large number of standard malodorous samples; including 5 kinds of standard malodorous liquids, i.e., -phenylethanol, isovaleric acid, methylcyclopentanone, -undecanolide, -methylindole; and 9 kinds of single standard malodorous compounds, i.e., C.sub.3H.sub.9N, C.sub.8H.sub.8, H.sub.2S, CH.sub.4S, C.sub.2H.sub.6S, C.sub.2H.sub.6S.sub.2, NH.sub.3, CS.sub.2, SO.sub.2, are prepared with different concentrations; in addition, also including the standard malodorous samples of mixed components prepared by a variety of the above compounds with different concentrations. (2) On-line detection data of a large number of malodorous pollutants by the gas sensor array (I-1). (3) Off-line olfactory data of a large number of malodorous pollutants in the laboratory for determining the dimensionless concentration values. (4) Off-line detection data of TVOCs and 9 kinds of malodorous components obtained by gas chromatographs, mass spectrometers and spectrophotometers for a large number of malodorous pollutants. (5) Complaint data of residents in the vicinities of malodorous pollution sources.

[0135] FIG. 2 is a schematic diagram illustrating the working principle of the odor electronic nose instrument and the centralized monitoring and analysis system for multiple malodorous pollution areas. The online centralized monitoring and analysis system based on an electronic nose instrument for multiple malodorous pollution areas includes an odor electronic nose instrument (I), 10 sampling heads (II-1)(II-10) for 10 outdoor monitoring points, an external vacuum pump (III), an ambient air purification device (IV), a clean air cylinder (V), an electronic hygrometer (VI), a central control room (VII), gas pipelines, as well as multiple stationary/mobile terminals (VIII), to realize the long-term online monitoring and the online estimation and prediction of important concentration control index values of malodorous gases for 10 points in some specified pollution areas. At this time, the gas paths and the electromagnetic valves are located at the working state by drawing the malodorous gas of the sampling head (II-1) from the first outdoor monitoring point to the odor electronic nose instrument (I), and the gas sensor array (I-1) thus produces its sensitive responses.

[0136] FIG. 3 is a schematic diagram illustrating the working principle of the odor electronic nose instrument (I). The components of the instrument (I) include:

[0137] (a), The thermostatic gas sensor array working room (I(a)) is composed of a gas sensor array and an annular working chamber of the gas sensor array (I-1), a thermal insulation layer (I-2), a resistance heating wire (I-3) and a fan (I-4), located at the top right of the odor electronic nose instrument (I).

[0138] (b), The multi-point centralized malodorous auto-sampling system (I(b)) includes a first two-position two-port electromagnetic valves (I-5) to control the on-off of purified ambient air, 10 two-position two-port electromagnetic valves from the second two-position two-port electromagnetic valve (I-6-1) to the eleventh two-position two-port electromagnetic valve (I-6-10) to control the on-off of malodorous gases at 10 monitoring points, a vacuum pressure gauge (I-7) to show the working state of the external vacuum pump (III), the twelfth two-position two-port electromagnetic valve (I-8) to control when the malodorous gases flow into the annular working chamber of the gas sensor array (I-1), a gas buffer cavity (I-9), the thirteenth two-position two-port electromagnetic valve (I-10) to control the flow conversion between 6,500 ml/min of malodorous gases and 1,000 ml/min of clean air in the annular working chamber of the gas sensor array (I-1), a throttle valve (I-11), a flowmeter (I-12), a vacuum pressure gauge (I-7), and located at the lower right of the odor electronic nose instrument (I).

[0139] (c), The computer control and data analysis system (I(c)) includes a computer mainboard (I-15), a data acquisition card (I-16), a monitor (I-17), a drive and control circuit module (I-18), a precision linear and switching power module (I-19), a hard disk, a network card, a video card, located on the left side of the odor electronic nose instrument (I).

[0140] FIG. 4 is a schematic diagram illustrating the mutual positions of the odor electronic nose instrument (I) and 10 sampling heads (II-1)(II-10) located at 10 monitoring points. (a) An area with road access; and (b) an area without road access. For the areas with access paths shown in FIG. 4(a), such as a chemical industry park or a residential area, the odor electronic nose instrument (I) ought to be arranged within a room in the center of the monitoring area; and for the areas without access roads shown in FIG. 4(b), the odor electronic nose instrument (I) ought to be arranged within a room at the boundary of the monitoring area. The position determination of the odor electronic nose instrument (I) is based on the rule of the shortest straight-line distance between it and each monitoring point. The external vacuum pump (III), the ambient air purification device (IV), the clean air cylinder (V) and the electronic hygrometer (VI) are arranged near the odor electronic nose instrument (I).

[0141] Suppose that the maximum monitoring area is 2 km*2 km=4 km.sup.2 and the inner diameter of stainless steel pipe connecting odor electronic nose instrument (I) and each monitoring point is d=10 mm. let us consider the most disadvantageous situation: The longest gas pipeline appears in the area without road access as shown in FIG. 4(b), and the maximum straight-line length is

[00004]$l_{\max }=\sqrt{2^2+1}=2.2.Math.4.Math..Math.\mathrm{km}.$

Let us still assume that the maximum flow rate of the external vacuum pump (III) is Q=280 L/min and the maximum vacuum pressure is P=100 mbar, the flow speed of the gas is =4Q/(d.sup.2)=59.42 m/s=3.57 km/min, and the malodorous gas is drawn from the farthest sampling point (II-3) or (II-9) to the odor electronic nose instrument (I) only needs t.sub.max=l.sub.max/=37.65 s. It is noted that the maximum volume of the pipeline is about U.sub.max=l.sub.max*d.sup.2/4=176 L, when l.sub.max=2.24 km and d=10 mm, which is smaller than the maximum gas volume that can be drawn by the external vacuum pump (III) for 1.0 min, Q=250-280 L. Considering the leakage factor, it means that the time duration for the tested gas at a certain monitoring point to be drawn to the odor electronic nose instrument (I) is only about 1.0 min. In such a short time, there is no chance for the odor to undergo deterioration or adsorption effect.

[0142] The most industrial parks, waste and sewage treatment zones, breeding farms, adjacent residential areas and other pollution areas to be monitored are within the area of 1 km.sup.2. Assuming that the maximum monitoring area is 1 km*1 km=1 km.sup.2, and the gas pipeline connecting the odor the odor electronic nose instrument (I) and each monitoring point is arranged around the boundary, and still considering the disadvantageous situation that the longest gas pipeline is l.sub.max=0.5+1+0.5=2 km, the external vacuum pump (III) can draw the gas from the monitoring point into the odor electronic nose instrument (I) within 1.0 min. In this disclosure, the odor electronic nose instrument and the multi-point centralized monitoring and analysis system for the malodorous pollution areas are particularly suitable for the places such as production workshops, sewage pools, breeding farms, etc., and can realize the online monitoring and analysis of some specific areas as large as several km.sup.2, as small as a production workshop or a building, or even a point.

[0143] FIG. 5 is a schematic diagram illustrating the arrangement of the gas sensor array (I-1) and the annular working chamber. FIG. 5(a) shows such a specific example: The gas sensor array consists of 16 sensitive with three types, including 11 MOS-type, i.e., four TGS2000 series (I-1-1), three TGS800 series with plastic shell (I-1-2), four TGS800 series (I-1-3) with stainless steel shell; four EC-type (I-1-4) and one PID-type (I-1-5). The MOS-type gas sensors have high sensitivity, long life and are sensitive to both organic and inorganic components; the EC-type gas sensors have good selectivity and are mainly used to detect the inorganic gases; and the PID gas sensor is sensitive to the VOCs between n-hexane and n-cetane. The cascade machine learning model determines the concentrations of H.sub.2S, NH.sub.3, SO.sub.2, CS.sub.2 and other inorganic components depending upon the responses of eleven MOS-type and four EC-type gas sensors; estimates the concentration values of TVOCs and such organic components as C.sub.3H.sub.9N, C.sub.8H.sub.8, CH.sub.4S, C.sub.2H.sub.6S, C.sub.2H.sub.6S.sub.2 depending upon the responses of eleven MOS-type gas sensors and one PM-type, and quantizes the dimensionless concentration OU values depending upon the responses of all the 16 gas sensors.

[0144] According to FIGS. 5(a), 5(b) and 5(c), the working chamber of gas sensor array (I-1) is composed of stainless steel base (I-1-6), sealing ring (I-1-7), stainless steel cover (I-1-8), partition (I-1-9), gas sensor socket (I-1-10), sealing material (I-1-11) and screw (I-1-12), forming a sealed annular chamber. During the headspace sampling, the malodorous gas is inhaled from the gas inlet, then passes through four TGS2000-series (I-1-1), three TGS800-series (I-1-2) with plastic shell, four TGS800-series (I-1-3) with stainless steel shell, four EC-type (I-1-4) and one PID-type gas sensor (I-1-5) around the working chamber, in order, and finally flows out from the air outlet. As a result, the gas sensor array generates the sensitive responses.

[0145] FIG. 6 is a schematic diagram illustrating the gas buffer cavity (I-9). The buffer cavity is located within the odor electronic nose instrument (I), with an inner diameter of 40 mm and a clear depth of 5-10 mm. Because the ratio between the inner diameter of the buffer cavity and the inner diameter of the gas pipelines connecting the odor electronic nose instrument (I) and 10 sampling heads (II-1)(II-10) located at 10 monitoring points is 4:1, the gas flow speed in this buffer cavity drops 16 times than in the pipelines, and the internal miniature vacuum pump (I-14) can draw enough malodorous gas from here.

[0146] FIG. 7 shows the on-off changes and mutual relationship of 14 two-position two-port electromagnetic valves in the centralized automatic sampling system of multi-point malodorous gases when the single sampling period is T.sub.0 and the cycle period is T=10T.sub.0. In the cycle sampling period T, the 10 two-position two-port electromagnetic valves from the second two-position two-port electromagnetic valve (I-6-1) to the eleventh two-position two-port electromagnetic valve (I-6-10) that control the on-off states of malodorous gases at 10 monitoring points are only on and off once. At any time in any single gas sampling period T.sub.0, among 10 two-position two-port electromagnetic valves, only one is on, and the other nine are off.

[0147] In the cyclic sampling period T, the first two-position two-port electromagnetic valve (I-5) that controls the on-off of purified ambient air, the twelfth two-position two-port electromagnetic valve (I-8) that controls the malodorous gas flowing into the annular working chamber of gas sensor array and the fourteenth two-position two-port electromagnetic valve (I-13) that controls the on-off of clean air are on-off for 10 times, and the thirteenth two-position two-port electromagnetic valve (I-10) that controls the flow conversion is on-off for 20 times.

[0148] Referring now to FIG. 7. Taking as an example the first single sampling period T.sub.0=240 seconds within the cyclic sampling period T=10T.sub.0, there are the several following situations:

[0149] (a), The whole single sampling period of T.sub.0=240 s. Among the 10 two-position two-port electromagnetic valves, the second two-position two-port electromagnetic valve (I-6-1) is always on, and the other nine two-position two-port electromagnetic valves are disconnected. Under the suction action of the external vacuum pump (III), the malodorous gas sequentially flows through the sampling head (II-1) located at the first monitoring point, pipelines, the second two-position two-port electromagnetic valve (I-6-1), the gas buffer cavity (I-9) and the external vacuum pump (III) with a flow rate of 250-280 l/min, and is finally discharged to outdoor.

[0150] (b), The 0-175 s segment in the single sampling period T.sub.0. Although the external vacuum pump (III) draws the measured malodorous gas of the sampling head (II-1) from the first monitoring point to the odor electronic nose instrument (I), because the twelfth two-position two-port electromagnetic valve (I-8) is disconnected, the malodorous gas at this time does not flow through the annular working chamber of the gas sensor array (I-1) of the gas sensor array, but is directly discharged from the external vacuum pump (III) to outdoor. The period of 175 s can be further divided into two sub-stages: (b1), the first 155 s for the preliminary recovery stage of the gas sensor array; and (b2), the last 20 s for the flushing stage of the gas sensor array. In the two time sub-stages, the fourteenth two-position two-port electromagnetic valve (I-13) is disconnected, the first two-position two-port electromagnetic valve (I-5) and the thirteenth two-position two-port electromagnetic valve (I-10) are on, and under the suction action of the internal miniature vacuum pump (I-14), the purified ambient air by the ambient air purification device (IV) sequentially flows through the first two-position two-port electromagnetic valve (I-5), the gas pipelines, the annular working chamber of the gas sensor array (I-1), the thirteenth two-position two-port electromagnetic valve (I-10) and the internal miniature vacuum pump (I-14) with a flow rate of 6,500 ml/min, and then is discharged to outdoor.

[0151] (c), The 176-205 s in the single sampling period T.sub.0=240 seconds. The first two-position two-port electromagnetic valve (I-5), the twelfth two-position two-port electromagnetic valve (I-8) and the thirteenth two-position two-port electromagnetic valve (I-10) are disconnected, and the fourteenth two-position two-port electromagnetic valve (I-13) is on. The clean air in the clean air cylinder (V) sequentially flows by the own pressure through the fourteenth two-position two-port electromagnetic valve (I-13), gas pipelines, the annular working chamber of the gas sensor array (I-1), the flowmeter (I-12) and the internal miniature vacuum pump (I-14) with a flow rate of 1,000 ml/min, and then is discharged to outdoor.

[0152] (d), The 111-240 s in the single sampling period T.sub.0=240 seconds. The first two-position two-port electromagnetic valve (I-5), the twelfth two-position two-port electromagnetic valve (I-8) and the thirteenth two-position two-port electromagnetic valve (I-10) are disconnected, and the twelfth two-position two-port electromagnetic valve (I-8) is on. Under the suction effect of the internal miniature vacuum pump (I-14), the malodorous gas sequentially flows through the fourteenth two-position two-port electromagnetic valve (I-13), gas pipelines, the annular working chamber of the gas sensor array (I-1), the flowmeter (I-12) and the internal miniature vacuum pump (I-14) with a flow rate of 1,000 ml/min, and then is discharged to outdoor.

[0153] FIG. 8 is a schematic diagram illustrating the three-dimensional appearance of the odor electronic nose instrument (I). The gas sensor array (I-1) is located in the upper right part of the odor electronic nose instrument (I). The monitor (I-17), the vacuum pressure gauge (I-7) and the flowmeter (I-12) can be seen from the front view.

[0154] FIG. 9 is a schematic diagram illustrating the back of the odor electronic nose instrument (I). The odor electronic nose instrument (I) is set with an interface for the external monitor, 2 USB interfaces, a mouse interface, a keyboard interface, an Internet interface, a clean and a purified air inlet, 10 inlets for the malodorous gases of 10 monitoring points, an outlet for the external vacuum pump (III) and an exhaust gas outlet.

[0155] In the 45 s response data recorded in a single sampling period T.sub.0, the difference between the steady-state maximum value U.sub.imax(t) and the minimum value U.sub.imin(t) of the response curve of a single gas sensor i is extracted as the characteristic component x.sub.i(t)=U.sub.max(t)U.sub.imin(t). Therefore, the gas sensor array generates a 16-dimensional response vector x(t)=(x.sub.1(t), . . . , x.sub.i(t), . . . , x.sub.16(t)).sup.TR.sup.16. Within 10 seconds after the end of data recording, i.e., 10 seconds after the ambient air flushing stage, the cascade machine learning model of the computer control and data analysis system (I(c)) predicts 10+1 concentration control index values of malodorous pollutants based on the 16-dimensional response vector x(t).

[0156] According to the divide-and-conquer strategy, the first level of the cascade machine learning model, namely, the convolution neural network layer, uses multiple single-output single-hidden-layer convolution neural networks to predict the response of each gas sensor. FIG. 10 is the schematic diagram illustrating the structure of the convolutional neural network CNN.sub.i1 for predicting the response x.sub.i(t+1) of the gas sensor i at the time point of t+1 (say the 40.sup.th minute in the near future). Table 2(a) shows the time-series response of the training set X.sub.i1R.sup.109 of the CNN.sub.i1, with ten 9-dimensional samples in total. The time-series span of the training set X.sub.i1 is [t18, t1]. When T.sub.0=240s and T=10T.sub.0, it is equivalent to the CNN.sub.i1 to learning the responses of the gas sensor i what have happened from the 12 hours ago to the current time. According to Table 2(a), a learning sample of CNN.sub.i1 is equivalent to a time-series response of the gas sensor i with the time length t=9. Table 2(b) shows such a response sample x.sub.1=(x.sub.i(t8), . . . , x.sub.i(t)).sup.TR.sup.9.

TABLE-US-00002 TABLE 2(a) The time-series training set X.sub.i1 of the convolutional neural network CNN.sub.i1. Input node Desired 1 2 3 4 5 6 7 8 9 output #Sample x.sub.i(t 9) x.sub.i(t 8) x.sub.i(t 7) x.sub.i(t 6) x.sub.i(t 5) x.sub.i(t 4) x.sub.i(t 3) x.sub.i(t 2) x.sub.i(t 1) x.sub.i(t) 1 x.sub.i(t 10) x.sub.i(t 9) x.sub.i(t 8) x.sub.i(t 7) x.sub.i(t 6) x.sub.i(t 5) x.sub.i(t 4) x.sub.i(t 3) x.sub.i(t 2) x.sub.i(t 1) 2 . . . . . . x.sub.i(t 17) x.sub.i(t 16) x.sub.i(t 15) x.sub.i(t 14) x.sub.i(t 13) x.sub.i(t 12) x.sub.i(t 11) x.sub.i(t 10) x.sub.i(t 9) x.sub.i(t 8) 9 x.sub.i(t 18) x.sub.i(t 17) x.sub.i(t 16) x.sub.i(t 15) x.sub.i(t 14) x.sub.i(t 13) x.sub.i(t 12) x.sub.i(t 11) x.sub.i(t 10) x.sub.i(t 9) 10

TABLE-US-00003 TABLE 2(b) The time-series response x.sub.i(t) of gas sensor array learned by the CNN.sub.i1 to predict x.sub.i(t + 1) at the time point of t + 1 Input node Predicted 1 2 3 4 5 6 7 8 9 output #Sample x.sub.i(t 8) x.sub.i(t 7) x.sub.i(t 6) x.sub.i(t 5) x.sub.i(t 4) x.sub.i(t 3) x.sub.i(t 2) x.sub.i(t 1) x.sub.i(t) x.sub.i(t + 1) 1

[0157] The convolutional neural network CNN.sub.i1 learns the time-series responses of the gas sensor i at 18 times that have occurred before the time t. If the delay length t=9, then the number of input nodes is set to be m.sub.i=9. The number of hidden nodes is set to be h.sub.i=5, and the number of output nodes is set to be n.sub.i=1. The convolutional neural network CNN.sub.i1 online learns the preprocessed time-series responses of the gas sensor ithe data set X.sub.i1, as shown in Table 2(a). The hidden and output activation functions of the CNN.sub.i1 are the modified Sigmoid function ()=3/(1+exp(/3)), and the error back-propagation algorithm is used for learning. The learning factor is n.sub.i1=(5/N.sub.i1=0.5, and the maximum number of iterations is 10,000. The input and output components in Tables 2(a) and 2(b) are scaled to the range of [0, 3].

[0158] the convolutional neural network CNN.sub.i1 completes online learning within 10 seconds after the ambient air flushing phase of the gas sensor array, and thus predicts the response x.sub.i(t+1) of the gas sensor i at the time point of t+1 according to the time-series response sample x.sub.i(t)=(x.sub.i(t8), . . . , x.sub.i(t)).sup.T given in Table 2(b).

[0159] The present disclosure uses the convolution neural networks CNN.sub.i2 and CNN.sub.i3 to predict the responses x.sub.i(t+2) and x.sub.i(t+3) of the gas sensor i at the time point of t+2 (for example, the 80.sup.th minute in the future) and the time point of t+3 (for example, the 120.sup.th minute in the future), respectively. The structure and learning parameters of CNN.sub.i2 and CNN.sub.i3 are the same as those of CNN.sub.i1. Tables 3 and 4 show the time-series response training sets X.sub.i2R.sup.109 and X.sub.i3R.sup.109 for the two convolutional neural networks. The CNN.sub.i2 and CNN.sub.i3 still use the same time-series response samples x.sub.i(t) as the CNN.sub.i1 shown in Table 2(b) to predict the responses x.sub.i(t+2) and x.sub.i(t+3) of the gas sensors i at the time points of t+2 and t+3, respectively. Compared with the time span of [t18, t1] in the dataset X.sub.i1, the time spans of X.sub.i2 and X.sub.i3 are [t19, t2] and [t20, t3], respectively, which are a little far away from the current time t. Therefore, the reliability of the predicted values of the CNN.sub.i2 and CNN.sub.i3 is relatively low, compared with that of the CNN.sub.i1.

TABLE-US-00004 TABLE 3 The time-series training set X.sub.i2 of the convolutional neural network CNN.sub.i2. Input node Desired 1 2 3 4 5 6 7 8 9 output #Sample x.sub.i(t 10) x.sub.i(t 9) x.sub.i(t 8) x.sub.i(t 7) x.sub.i(t 6) x.sub.i(t 5) x.sub.i(t 4) x.sub.i(t 3) x.sub.i(t 2) x.sub.i(t) 1 x.sub.i(t 11) x.sub.i(t 10) x.sub.i(t 9) x.sub.i(t 8) x.sub.i(t 7) x.sub.i(t 6) x.sub.i(t 5) x.sub.i(t 4) x.sub.i(t 3) x.sub.i(t 1) 2 x.sub.i(t 18) x.sub.i(t 17) x.sub.i(t 16) x.sub.i(t 15) x.sub.i(t 14) x.sub.i(t 13) x.sub.i(t 12) x.sub.i(t 11) x.sub.i(t 10) x.sub.i(t 8) 9 x.sub.i(t 19) x.sub.i(t 18) x.sub.i(t 17) x.sub.i(t 16) x.sub.i(t 15) x.sub.i(t 14) x.sub.i(t 13) x.sub.i(t 12) x.sub.i(t 11) x.sub.i(t 9) 10

TABLE-US-00005 TABLE 4 The time-series training set X.sub.i3 of the convolutional neural network CNN.sub.i3. Input node Desired 1 2 3 4 5 6 7 8 9 output #Sample x.sub.i(t 11) x.sub.i(t 10) x.sub.i(t 9) x.sub.i(t 8) x.sub.i(t 7) x.sub.i(t 6) x.sub.i(t 5) x.sub.i(t 4) x.sub.i(t 3) x.sub.i(t) 1 x.sub.i(t 12) x.sub.i(t 11) x.sub.i(t 10) x.sub.i(t 9) x.sub.i(t 8) x.sub.i(t 7) x.sub.i(t 6) x.sub.i(t 5) x.sub.i(t 4) x.sub.i(t 1) 2 x.sub.i(t 19) x.sub.i(t 18) x.sub.i(t 17) x.sub.i(t 16) x.sub.i(t 15) x.sub.i(t 14) x.sub.i(t 13) x.sub.i(t 12) x.sub.i(t 11) x.sub.i(t 8) 9 x.sub.i(t 20) x.sub.i(t 19) x.sub.i(t 18) x.sub.i(t 17) x.sub.i(t 16) x.sub.i(t 15) x.sub.i(t 14) x.sub.i(t 13) x.sub.i(t 12) x.sub.i(t 9) 10

[0160] The three networks, CNN.sub.i1, CNN.sub.i2 and CNN.sub.i3, all completed their online learnings and predictions within 10 seconds after the ambient air flushing phase of the gas sensor array. Therefore, 3*16 convolution neural networks are adopted in the present disclosure to predict the responses at the time points of t+1, t+2 and t+3 for all 16 response curves of the gas sensor array. If only the response at the time point of t+1 is predicted, only 16 single-output convolution neural networks are needed.

[0161] According to the divide-and-conquer strategy, the present disclosure decomposes the overall prediction problem of multiple concentration values of malodorous gas into multiple individual concentration value prediction problems, and uses the second level of the cascade machine learning model, i.e. multiple single-output depth neural networks, to predict multiple individual concentration values one by one, thereby effectively reducing the complexity of machine learning models and algorithms. The number of the single-output DNNs are equal to the number of concentration control indicators of malodorous gases to be predicted, namely they are in one-to-one correspondence. For example, 10+1 single-output DNNs are needed to predict the dimensionless concentration OU value, the concentration values of 9 specified compounds including NH.sub.3, H.sub.2S, CS.sub.2, C.sub.3H.sub.9N, CH.sub.4S, C.sub.2H.sub.6S, C.sub.2H.sub.6S.sub.2, C.sub.8H.sub.8, SO.sub.2, as well as the TVOC concentration value, present in malodorous pollutants. A single-output DNN learns the big data of malodorous gases. The input values are the detection data of the gas sensor array as well as the temperature and humidity data on the site of the odor electronic nose instrument, and the target output is the off-line measurement values of odor olfactory identification, the conventional instruments including gas chromatography, mass spectrometry, and others, as well as the data of residents' complaints. Some samples in the big data of malodorous gases that only have the responses of the gas sensor array but without the off-line measurement values such as dimensionless olfactory discrimination, gas chromatography, mass spectrometry and residents' complaint data, will not participate in the study.

[0162] A single-output DNN.sub.j has three hidden layers, and the hidden-layer and the output-layer nodes use the modified Sigmoid activation function ()=3/(1+exp(/3)); the response data of the gas sensor arrays and the target output components are respectively and proportionally transformed to the ranges of [0, 3]. The first and second hidden layers are the feature transformation (coding) layers. The structure and weight parameters are determined by the single-hidden-layer and peer-to-peer neural networks. FIG. 11 shows the learning process of a peer-to-peer neural network to determine the weights and thresholds of between the k.sup.th and the k+1.sup.th hidden layer in the DNN. FIG. 11(a) shows that the number of the output and the input nodes in a peer-to-peer neural network are equal, both of which are linear activation functions; the weights and thresholds of the hidden-to-output layer are directly equal to those of its input-to-hidden layer, and the target outputs are directly equal to their actual inputs. FIG. 11(b) shows that after the learning of the peer-to-peer neural network, the number of hidden nodes in the (k+1).sup.th layer of DNN, is equal to the number of hidden nodes in the peer-to-peer neural network, and the weights and thresholds of between the k.sup.th and the (k+1).sup.th layer are equal to those in the input-to-hidden layer of the peer-to-peer neural network. If the number of the learning samples of the DNN.sub.j is N, then the learning factor of the peer-to-peer neural network is =2/N, and the maximum number of iterative steps is .sub.max=10,000. The third hidden layer of the DNN.sub.j is a nonlinear mapping layer, which is used to fit the j.sup.th concentration control index value of malodorous gases with the single output unit j.

[0163] FIG. 12 is a schematic diagram illustrating the cascade machine learning model to predict the concentrations of various malodorous pollutants at the time point of t+1 (say the 40.sup.th minute in the coming future). According to FIG. 12(a), the first level of the cascade machine learning model uses 16*3 groups of single-output single-hidden-layer convolution neural networks to learn the time-series responses generated by 16 gas sensors, and then to predict the responses of each gas sensor at the time points of t+1, t+2 and t+3, respectively according to the time-series response sample x.sub.i(t)=(x.sub.i(t8), . . . , x.sub.i(t)).sup.T. In the second level of the cascade machine learning model, 10+1 single-output three-hidden-layer depth neural network modules are used to predict the above 10+1 malodorous concentration control indicator values.

[0164] Assuming that the DNN.sub.j predicts the concentration value y.sub.j(t+1) of a malodorous gas at the time point of t+1, it is based on the predicted response vector (x.sub.1(t+1), x.sub.2(t+1), . . . , x.sub.16(t+1)).sup.T by the previous 16 CNN.sub.i1 (i=1, 2, . . . , 16) and the current temperature and humidity values; the DNN.sub.j forecasts y.sub.j(t+1) based on the predicted response vector (x.sub.1(t+2), x.sub.2(t+2), . . . , x.sub.16(t+2)).sup.T by the previous 16 CNN.sub.i2 and current temperature and humidity values, and so on.

[0165] If the actual input is the current response vector of the gas sensor array (x.sub.1(t), x.sub.2(t), . . . , x.sub.16(t)).sup.T, and the temperature and the humidity values at the current time t are added if necessary, the actual output of DNN.sub.j is the estimating value of the current concentration y.sub.j(t) of the malodorous component j.