Pump monitoring system and method

Abstract

A system for monitoring the operation of a surface pump such as a hand-operated water pump or oil pump, which uses an accelerometer mounted to a component of the pump to monitor movement of a pump component, for example the handle, and transmitted via a data connection such as a mobile data communications network to a server. The accelerometer measurements are processed by using a trained model such as a support vector machine to output an indication of the condition of the pump or the level of liquid in the well or borehole served by the pump. The model may be trained using a training data set of sensor measurements associated with liquid level in the well and condition of the pump.

Claims

1. A system for monitoring a surface pump for raising liquid from a well, the monitoring system comprising: a surface pump comprising a handle for operating the pump; a sensor mounted on or associated with the handle, wherein the sensor is configured to sense movement of the handle and provide an output signal representative thereof; and a signal processor for receiving the sensor output signal and processing it to derive therefrom an estimate of the level of liquid in the well or an estimate of the condition of the pump or both.

2. The system according to claim 1, wherein the signal processor comprises a trained model for deriving the estimate of the level of liquid or the condition of the pump, or both, from the output signal from the sensor.

3. The system according to claim 2, wherein the trained model comprises a classifier.

4. The system according to claim 2, wherein the trained model comprises a support vector machine, artificial neural network, kernel-based machine or Gaussian process classifier.

5. The system according to claim 1, wherein the signal processor is adapted to extract a plurality of features from successive periods of the output signal and form them into a feature vectors respectively representing the output signal in the successive periods.

6. The system according to claim 5, wherein the signal processor is adapted to form a feature vector for each cycle of a periodic output signal from the sensor.

7. The system according to claim 6, wherein the signal processor is adapted to divide each cycle into a plurality of segments and to form as the feature vector for that cycle the values of the underlying waveform of the output signal at a predetermined position in each segment together with an estimate of the noise in each segment.

8. The system according to claim 7, wherein the signal processor is adapted to divide each cycle into the same number of segments whereby the duration of each segment may vary from cycle to cycle.

9. The system according to claim 1, wherein the system comprises a surface pump and a monitoring system, wherein the monitoring system comprises the sensor and the signal processor.

10. The system according to claim 1, wherein the surface pump is one of: a water pump, an oil pump, a hand pump, or a reciprocating pump.

11. The system according to claim 1, wherein the sensor is an accelerometer, pressure sensor or vibration transducer.

12. The system claim 2, wherein the sensor is an accelerometer, pressure sensor or vibration transducer.

13. The system according to claim 1, wherein the signal processor processes the output signal to classify at least one of: the condition of pump, or the user of the pump.

14. The system according to claim 1, further comprising a data transmitter for sending data from at least one of the sensor and the signal processor via a data communications link.

15. A method of monitoring a surface pump for raising liquid from a well, the method comprising: measuring movement of a handle of the pump and providing an output signal representative thereof; receiving the output signal and processing it to derive therefrom an estimate of the level of liquid in the well, or an estimate of the condition of the pump or both.

16. The method according to claim 15, wherein the output signal is processed to derive the estimate of the level of liquid in the well or the condition of the pump by means of a trained model.

17. The method according to claim 16, wherein the trained model comprises a classifier, support vector machine, artificial neural network, kernel-based machine or Gaussian process classifier.

18. The method according to claim 15, comprising the steps of extracting a plurality of features from successive periods of the output signal and forming them into a feature vectors respectively representing the output signal in the successive periods.

19. The method according to claim 18, wherein each successive period is divided into a plurality of segments and the feature vector for that period is formed by the values of the underlying waveform of the output signal at a predetermined position in each segment together with an estimate of the noise in each segment.

20. The system according to claim 1, wherein the sensor is an accelerometer or vibration transducer.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

(1) The present disclosure will be further described by way of example with reference to the accompanying drawings in which:

(2) FIGS. 1A and 1B schematically illustrate the two most common designs of water hand pump;

(3) FIGS. 2A and 2B illustrate example sensor outputs for the two designs of hand pump illustrated in FIGS. 1A and 1B for an embodiment of the present disclosure utilising an accelerometer on the hand pump handle;

(4) FIG. 3 illustrates an example feature extraction method according to one embodiment of the present disclosure;

(5) FIGS. 4A and 4B are visualisations of feature vectors corresponding to recordings of a variety of hand pumps;

(6) FIGS. 5A and 5B compare estimates of water level in a well obtained by an embodiment of the present disclosure with water levels measured directly;

(7) FIGS. 6A and 6B illustrate heteroscedastic Gaussian processes fitted to sensor data from a water hand pump;

(8) FIG. 7 is a flow diagram of the method according to one embodiment of the present disclosure;

(9) FIG. 8 is a flow diagram of a method of training a model according to one embodiment of the present disclosure; and

(10) FIG. 9 is a schematic diagram of a monitoring system according to one embodiment of the present disclosure.

DETAILED DESCRIPTION

(11) FIG. 1A and FIG. 1B of the accompanying drawings schematically illustrate respectively the Afridev and India Mark II types of water hand pump. These two pumps form the majority of hand pumps in use in the developing world. Both are positive displacement, piston type pumps in which a pivotably mounted handle 1 is connected either directly or with a connecting chain to a pump rod 2 which slides a non-return piston valve (not visible) in a vertical cylindrical pipe 4 fitted at its base with a foot valve. In use the vertical cylindrical pipe 4 is disposed within a well or borehole 6. Pumping the handle 1 causes vertical reciprocation of the pumping rod 2 and piston valve with upwards movement of the piston valve drawing water from the well or borehole into the cylindrical pipe 4 through the foot valve and also moving water above the piston valve (from the previous stroke) up through the pump head 8 to be dispensed. The subsequent downward movement of the piston rod 2 forces the piston valve through the water which has just been drawn in, to start the cycle again. As illustrated the pump handle 1 is fitted with a sensor fitted within a package 10 known as a waterpoint data transmitter 10 which, in this embodiment, includes a consumergrade, low-cost IC-based accelerometer such as that found in a mobile phone handset or games controller, such as an Analogue Devices ADXL335. FIG. 9 schematically illustrates the waterpoint data transmitter 10 fitted to a pump. The accelerometer 11 senses movement in the X, Y and Z directions and produces three analogue output signals proportional to the acceleration sensed along each axis. The analogue output can be filtered by a simple RC filter 12 to remove any high-frequency noise, and passed via an analogue to digital converter 13 to a data processor 15 for processing the data. Alternatively the accelerometer 11 can be a digital accelerometer obviating the need for the RC filter 12 and separate A/D converter 13. The output of the data processor 15 is passed to a modem 17 for dispatch via a data link 19, for example provided by a communications network such as the cellular mobile telephone network, and another modem 18 to a server 21.

(12) The estimation of liquid level in the well and pump condition based on the sensor data can be carried out by the data processor 15 or at the server 21. Thus the data processor 15 can be adapted only to compress and package the sensor data to be sent via a data connection provided, for example, by mobile telephone or other communication network e.g. via an SMS text message on the GSM network, or can obtain the liquid level and pump condition data and compress and package that for transmission to the server 21 via the data connection. The explanation below applies to processing either at the server 21 or at the pump. The water point data transmitter 10 can be retrofitted to water pumps or can be fitted on manufacture.

(13) FIGS. 2A and 2B show approximately six seconds of recorded accelerometer data for each of an Afridev pump (FIG. 2A) and an India Mark II pump (FIG. 2B) for each of the three, X, and Z directions. For these recordings the Z direction corresponds to the main up and down direction of the handle, the Y direction to the longitudinal axis of the handle and the X direction transverse of the handle. There is a marked difference in the X direction recordings, this is thought to be because the India Mark II pump has a different connection between the handle and pumping rod resulting in a slightly elliptical motion of the handle. Thus the X direction for the India Mark II pump shows a greater periodicity. It can also be seen that in the Y and Z traces for both pumps the amount of noise differs between the upstroke and downstroke. This being because during the downstroke the handle is under load whereas the upstroke is a lower load return stroke.

(14) In a first embodiment the sensor outputs shown in FIGS. 2A and 2B are processed to reduce their dimensionality and put them in a form which is suitable for analysis by a conventional machine learning algorithm such as a support vector machine. A preferred type of feature extraction is illustrated in FIG. 3.

(15) The accelerometer 11 used in this embodiment has a sampling rate of 96 Hz, meaning that it provides 96 acceleration measurements per second (per axis). FIG. 3 shows magnified one period from one axis of the recording of FIG. 2B with the individual acceleration samples shown as dots. The aim is to express the recording as a combination of a smooth underlying waveform together with noise, i.e.:
x=ƒ(t)+ε
where ƒ(t) is the function describing the underlying waveform and ε is the noise. For the function ƒ(t) a smoothing spline can be selected which minimises the weighted sum of the function fluctuation and the corresponding mean square error as shown in the equation below:
Σ.sub.i=1.sup.n(x.sub.i−s(t.sub.i)).sup.2+(λ−1)∫.sub.t[s″(t.sub.i)].sup.2dt
where s is the point on the smoothing spline that minimises the function.
The smoothing parameter λ controls the complexity of the spline that is fitted to the data. For this embodiment a value of λ=0.002 was selected, though the results are relatively insensitive to changes in λ.

(16) Having fitted the spline to the data, the noise can be taken as the distance between each original data point and the spline.

(17) A feature vector for this cycle (cycles can easily be recognised by detecting the maxima or minima) can then be formed by dividing the cycle into, for example, p=16 intervals as shown by the short vertical lines, taking the value of the spline at each of the interval boundaries and taking an estimate of the noise in each interval as the sum of the distances between the original data points and the spline in that interval. Thus in this example each feature vector consists of 16 spline values and 16 noise values. Each of the feature vectors for a complete recording thus represents a point in a 32 dimensional “feature vector space”. It should be appreciated that by dividing the sensor output into cycles and dividing each cycle into an equal number of intervals, the method is effectively distorting the time base to allow for different timing of pump operation by different users or in different circumstances.

(18) Typically water pump handles are operated at about 1 Hz, thus each axis of the sensor output provides typically one feature vector per second, each feature vector having 32 components, though the method works for any cycle length. It should also be noted that other aspects of the signal can be added to the feature vector. For example the cycle length can be informative, it goes up if the aquifer is low because of the extra effort required, and can go down if the pump is leaky, and so period length can be added as a component of the feature vector.

(19) The feature vectors thus provide a representation of the sensor output recording which can be analysed by a machine learning algorithm.

(20) FIGS. 4A and 4B illustrate respectively the feature vectors from the recordings of FIGS. 2A and 2B but with their dimensionality reduced to two dimensions for easy visualisation by means of Sammon's mapping. Sammon's mapping is a visualisation technique which tries to preserve the relative spacings of high dimensional feature vectors in a low dimensional display (in this case two dimensional). It can be seen from FIGS. 4A and 4B that the feature vectors from different recordings group together demonstrating that the feature vector representation of the recordings preserves the useful information in the recordings. It should be noted that the full 32 dimensions are used in the training and monitoring discussed below—the two dimensional plots of FIGS. 4 A and 4B are just for visualisation.

(21) With this embodiment the estimates of liquid level in the well and condition of the pump are obtained from the feature vectors by use of a trained model, in this case a support vector machine. A support vector machine is one type of machine learning algorithm, but other types can be used. The model must first be trained on a training set of data for which the desired output (i.e. the liquid level or pump condition) is known. Once the support vector machine has been trained on a training data set, it can be presented with new feature vectors and it will output an estimate of the liquid level or pump condition.

(22) Rather than a support vector machine, other machine learning algorithms can be used such as neural networks or kernel-based machines. Techniques for training these on a training data set, validating them and using them to classify further data are well known.

(23) FIGS. 5A and 5B illustrates the results of ground water level predictions (crosses) and measured level (lighter dots) with FIG. 5A being the individual spline estimations (i.e. one for each cycle) and FIG. 5B showing the average estimation from each recording. It can be seen that there is good agreement between the estimations and the directly measured level.

(24) FIGS. 7 and 8 summarise the test and training aspects of this embodiment. As illustrated in FIG. 8 in step 90 a training set of accelerometer data is taken together with measured water level data. For each of the cycles of accelerometer data a feature vector is created in step 91 as explained above and these feature vectors together with the measured water levels are used in step 92 to train the model (such as the SVM above).

(25) For monitoring performance, as shown in FIG. 7, accelerometer data is taken in step 80 and in step 81 is formed into feature vectors for each cycle as before. These feature vectors are input in step 82 to the trained model, which in step 83 outputs the water level and any other aspects which it has been trained to distinguish, such as the pump condition or the user. Training for other aspects corresponds to the training process of FIG. 8. Training to detect the condition of the pump can either utilise data from pumps which are known to be faulty (for example by fitting them with faulty components), or can follow a novelty detection approach in which the distance of a feature vector from a predefined region of normality in the multi-dimensional feature vector space (32 dimensions in the embodiment above) is calculated and, if it is greater than a preset threshold, a malfunction alarm is generated. The region of novelty may be defined by using a training data set of recordings of pumps known to be in normal operation. The distance of an input feature vector from the region of normality can be calculated as the distance from the centroid of the normal feature vectors or the distance from a certain number of nearby feature vectors. Similarly training to distinguish users can be based on a training data set consisting of recordings from different users such as male adult, female adult, child etc.

(26) The embodiment above uses feature extraction to reduce the dimensionality of the input data and a machine learning algorithm such as a support vector machine. However, alternative approaches are possible, for example the entire waveform can be described using a Gaussian process model thus obviating the need for feature extraction. Gaussian process models are trained using a training data set and thus can classify input data to output estimations of liquid level, pump condition, user as before. FIGS. 6A and 6B illustrate Gaussian processes fitted to a 15 second interval of data from an Afridev pump for the Y axis (top) and Z axis (bottom). FIG. 6A is data from a deep well and FIG. 6B from a shallow well. In each case the data points from the accelerometer are shown as dots and the fitted Gaussian process shown as a line.