Intelligent control of spunlace production line using classification of current production state of real-time production line data

Abstract

Disclosed is an intelligent control system of spunlace production line, which includes a data acquiring module, which is used for acquiring and storing real-time production line data; the production line data includes cotton feeding roller value, real-time moisture value, real-time speed value and real-time gram weight value; the data process module is used for classify and controlling that production line data, and giving the adjustment opinions of the cotton feeding roller parameters; the parameter control module is used for verifying the parameter adjustment opinions and applying the opinions to the control system; the data acquiring module, the data processing module and the parameter control module are connected in sequence.

Claims

1. An intelligent control system of a spunlace production line implemented in a computer system using a set of computer-executable instructions, comprising: a data acquiring module used for acquiring and storing real-time production line data; wherein the real-time production line data comprises cotton feeding roller values, real-time moisture values, real-time speed values and real-time gram weight values; a data process module used for classifying and controlling the real-time production line data, and giving adjustment opinions of cotton feeding roller parameters; and a parameter control module used for verifying parameter adjustment opinions and applying the parameter adjustment opinions to the intelligent control system, wherein the data acquiring module, the data processing module and the parameter control module are connected in sequence; wherein the data processing module comprises: a classifying unit used for predicting a probability of exceeding a gram weight threshold in future through a classifying model; a control unit used for setting up an automatic control closed loop and giving the adjustment opinions of cotton feeding roller parameters; the classifying model is used for classifying according to the real-time production line data to obtain current production states and classification results; defining label types for the classification results, and judging whether an adjustment operation is needed or not according to the classification results; wherein if the adjustment operation is needed, the classification results are input into the control unit for a further processing, and if the adjustment operation is not needed, new data is continuously re-input.

2. The intelligent control system of a spunlace production line according to claim 1, wherein the data acquiring module acquires the real-time production line data through sensors on the process production line, and saves the real-time production line data into a time series database through a data integration software program for subsequent modules to process and analyze.

3. The intelligent control system of a spunlace production line according to claim 2, wherein the data acquiring module further comprises a preprocessing unit; the preprocessing unit is used for cleaning, segmenting and extracting the real-time production line data, performing a data enhancement processing by adopting an up-sampling method or a down-sampling method to obtain preprocessed data, and storing the preprocessed data into the time series database.

4. The intelligent control system of a spunlace production line according to claim 1, wherein the classifying model adopts a double-layer classifier; the double-layer classifier comprises a first classifier and a second classifier; and an output quantity of the first classifier is taken as an input quantity of the second classifier.

5. The intelligent control system of a spunlace production line according to claim 4, wherein the first classifier adopts a random forest model, extracts features of a historical gram weight window and a historical cotton feeding roller window, inputs the features into the random forest model, and outputs a predicted percentage of each label; the second classifier inputs a historical window of the predicted percentage of the output label of the first classifier by adopting a long short-term memory (LSTM) model, and outputs a predicted label of a production state; the LSTM model uses a small batch gradient descent method to train historical data, adjusts real-time data through random gradient descent, and obtains a probability of the label of the production state through a full connection layer activated by softmax.

6. The intelligent control system of a spunlace production line according to claim 5, wherein the control unit comprises: a prediction model used to predict the cotton feeding roller value output by the system in future according to the historical production line data and the predicted percentage of labels; an optimization sub-unit used for adjusting the cotton feeding roller value through the prediction model between a variable quantity of gram weight and a variable quantity of cotton feeding roller, using the prediction percentage of the labels output by the classifying model as a membership degree of the classification of the production state, and outputting an optimal adjustment amount of the cotton feeding roller based on a model conversion.

7. The intelligent control system of a spunlace production line according to claim 1, wherein the parameter control module comprises a feedback correction unit; and the feedback correction unit is used for re-predicting at each new sampling moment, and correcting a prediction result using real-time information, and then performing a new optimization.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

(1) In order to more clearly explain the embodiments of the present application or the technical solutions in the prior art, the following will briefly introduce the drawings to be used in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, other drawings may be obtained according to these drawings without any creative labor.

(2) FIG. 1 is a flow diagram of the prior art in the present application.

(3) FIG. 2 is a system design diagram of an embodiment of the present application.

(4) FIG. 3 is an example diagram of the normal trend of gram weight according to the embodiment of the present application.

(5) FIG. 4 is an example diagram of abnormal trend of gram weight according to the embodiment of the present application.

(6) FIG. 5 is an example of cleaning of cotton feeding roller data according to the embodiment of the present application.

(7) FIG. 6 is a brief flow chart of down-sampling and up-sampling methods according to an embodiment of the present application.

(8) FIG. 7 is a model design diagram of an embodiment of the present application.

(9) FIG. 8 is a schematic diagram of the classifying model design of the embodiment of the present application.

(10) FIG. 9 is a schematic diagram of LSTM model of an embodiment of the present application.

(11) FIG. 10 is a structural diagram of a conventional control system according to an embodiment of the present application.

(12) FIG. 11 is a schematic diagram of optimal control according to an embodiment of the present application.

(13) FIG. 12 is a structural diagram of a control model of an embodiment of the present application.

(14) FIG. 13 is a design flow chart of an embodiment of the present application.

DETAILED DESCRIPTION OF THE EMBODIMENTS

(15) The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, but not all of them. Based on the embodiment of the present application, all other embodiments obtained by ordinary technicians in the field without creative labor are within the scope of the present application.

(16) In order to make the above objects, features and advantages of the present application more obvious and understandable, the present application will be explained in further detail below with reference to the drawings and detailed description.

(17) This embodiment provides an intelligent control system of the spunlace production line (as shown in FIG. 2), including: a data acquiring module used for acquiring and storing real-time production line data; and the production line data includes cotton feeding roller values, real-time moisture values, real-time speed values and real-time gram weight values; a data process module used for classifying and controlling the production line data, and giving adjustment opinions of cotton feeding roller parameters; a parameter control module used for verifying parameter adjustment opinions and applying the parameter adjustment opinions to the intelligent control system; and the data acquiring module, the data processing module and the parameter control module are connected in sequence.

(18) The data acquiring module acquires the real-time production line data through sensors on the process production line, and saves the production line data into an influxDB time series database through KEPServer for subsequent modules to process and analyze.

(19) The data acquiring module further includes a preprocessing unit; the preprocessing unit is used for cleaning, segmenting and extracting the production line data, performing a data enhancement processing by adopting an up-sampling method or a down-sampling method to obtain preprocessed data, and storing the preprocessed data into the influxDB time series database.

(20) At present, the data obtained from spunlace production line include online detection data, spunlace data and carding machine data. The gram weight data will fluctuate up and down normally due to various factors. When the gram weight value exceeds the target value, the staff will keep the gram weight value within a stable range by adjusting the cotton feeding roller data in the carding machine data. The gram weight data and cotton feeding roller data are shown in Table 1 and Table 2 below.

(21) TABLE-US-00001 TABLE 1 Average Target gram weight gram weight Time 35.2 35 2022 Feb. 27 23:10:00 35 35 2022 Feb. 27 23:10:00 35 35 2022 Feb. 27 23:10:00 36 35 2022 Feb. 27 23:10:00 36.6 35 2022 Feb. 27 23:10:00 36.7 35 2022 Feb. 27 23:10:00 36.2 35 2022 Feb. 27 23:10:00 34.7 35 2022 Feb. 27 23:10:00 34.9 35 2022 Feb. 27 23:10:00 35.3 35 2022 Feb. 27 23:10:00 35.2 35 2022 Feb. 27 23:10:00

(22) TABLE-US-00002 TABLE 2 Low-frame Elevated cotton feed- cotton feed- Time ing roller ing roller 3.443 3.49417 2022 Feb. 27 23:10:00 3.47022 3.50506 2022 Feb. 27 23:10:00 3.43429 3.50506 2022 Feb. 27 23:10:00 3.46695 3.51051 2022 Feb. 27 23:10:00 3.46695 3.50289 2022 Feb. 27 23:10:00 3.45389 3.49744 2022 Feb. 27 23:10:00 3.46151 3.50833 2022 Feb. 27 23:10:00 3.46478 3.51051 2022 Feb. 27 23:10:00 3.45389 3.50289 2022 Feb. 27 23:10:00 3.42884 3.49417 2022 Feb. 27 23:10:00 3.46695 3.50833 2022 Feb. 27 23:10:00

(23) In order to show the trend of gram weight data changing with time more intuitively, FIG. 3 and FIG. 4 respectively select a time series of normal gram weight and a time series of abnormal gram weight to draw images. It can be seen that if the abnormal threshold is set at 2, the gram weight curve of FIG. 3 always fluctuates randomly within the target gram weight range of 352, while the gram weight curve of FIG. 4 obviously keeps increasing in the second half, and it has exceeded the threshold value of 37 (that is, the target gram weight of 35+ threshold value of 2) many times. At this time, it is necessary to adjust the cotton feeding roller parameters to make the gram weight curve return to about 35. The target is to identify the abnormal trend as soon as possible during the real-time monitoring of the production line, and give the adjustment opinions of the cotton feeding roller.

(24) As for the data of the spunlace production line, each batch of materials has different categories and different working conditions (for example, the adjustment times of workers are less and the gram weight is more unstable during handing-off), which will affect the final actual gram weight curve result. Meanwhile, due to the complicated situation on the production line, the on-off operation of the machine and the reading values of sensors in other abnormal situations are not directly marked in the acquired data, and the data obtained in these situations are often invalid abnormal values.

(25) In order to use effective and reliable data when training the model, the acquired data is cleaned in the early stage. The main contents include: (1) performing data segmentation and extraction according to the stable working conditions provided by the production line; (2) performing data cleaning for some abnormal values which obviously deviate from the average value, or data with a very large jump in a period of time.

(26) FIG. 5 shows an example of data cleaning for cotton feeding roller data. It can be seen that there are some meaningless jumps before cleaning. Skip the jumps during cleaning, and smooth the average.

(27) The abnormal data samples on the actual production line account for a small proportion, and most of the data still fluctuate within the normal range, so there is a problem of data imbalance, which will lead to the trained model biased towards the majority of results. Meanwhile, the lack of abnormal samples may also affect the characteristics of abnormal trends in model learning. Therefore, some data enhancement methods are adopted in this embodiment to solve this problem.

(28) Specifically, in the process of training, the method of down-sampling or up-sampling is used to solve the problem of data imbalance, so that the data volume of different labels can reach a basically balanced state, which is convenient for the subsequent actual model analysis. FIG. 6 shows a brief flow chart of down-sampling and up-sampling methods.

(29) The data processing module includes: a classifying unit used for predicting the probability of exceeding the gram weight threshold in the future through the classifying model; a control unit used to set up automatic control closed loop and give the adjustment opinions of cotton feeding roller parameters.

(30) In order to predict the weight trend and give corresponding adjustment suggestions, a two-stage model is proposed to complete these tasks, including a classifying model and a control model. The specific model design diagram is shown in FIG. 7. The classifying model defines a multi-classification problem, aiming at predicting the possibility of exceeding the gram weight threshold in the future according to the current real-time window data. Firstly, the model uses the historical data of Yongxin process production line for training, and through the historical data, the characteristic data related to gram weight and cotton feeding roller are extracted, and a time series classifying model is built for prediction. The control model uses the historical manual adjustment of cotton feeding roller and the corresponding historical data of gram weight change to build a model, build an automatic control closed loop, and give the adjustment opinions of cotton feeding roller parameters.

(31) The classifying model is used for classifying according to real-time production line data, obtaining the current data state and classification result, defining the label type of the classification result, and judging whether the adjustment operation is needed according to the classification result; if the adjustment operation is needed, the classification result is input into the control unit for further processing; if the adjustment operation is not needed, new data is continuously input.

(32) The goal of the classifying model is to classify the status of the current data according to the real-time data, and then make the decision whether to adjust the operation according to the classification results.

(33) Because the obtained data has not been manually marked, it is necessary to mark the data through certain judgment standards. According to the existing experience of spunlace production line, the following empirical rules are summarized for consideration: (1) small amplitude fluctuations can be ignored, and medium and small amplitude fluctuation in gram weight can also be ignored; (2) medium and small amplitude fluctuation with order of minutes continuously higher than the target value needs attention, and two or three consecutive volumes small amplitude fluctuation slightly higher than the target value need attention; (3) for cost reasons, staying above the target value requires more attention than staying below the target value.

(34) According to these rules, the judgment of whether or not to make adjustment depends largely on whether there are cases that are higher than the target value in a certain period of time, and whether there are cases that are higher than the target value in a small period of time. According to the above experience and actual production situation, in this embodiment, the threshold of abnormal samples exceeding the target value is set to 2, and the label is defined as shown in Table 3 (the label definition is only an example, and will be changed according to the actual production situation).

(35) TABLE-US-00003 TABLE 3 Label Meaning Judgement standard VL Strong Whether the average value of the upper sample is greater than (target supersample value +2) in the next 20 time points L Upper Whether the average value of the supersample samplei s greater than (target value +1) in the next 20 time points N Normal Whether the sample will always sample fluctuate within the target range in the next 20 time points S Lower Whether the average value of the supersample sample is less than (target value 1) in the next 20 time points VS Strong Whether the average value of the lower sample is less than (target value 2) supersample in the next 20 time points

(36) Because it is hoped that the label can predict the change of the future gram weight trend through the historical window value, the classifying model design as shown in FIG. 8 is adopted. The production line itself has a high demand for model sensitivity, and the abnormal trend needs to be adjusted as soon as possible. Therefore, the accuracy of classification may be well improved by adopting a double-layer classifier.

(37) The first classifier adopts a random forest model. Random forest is an algorithm that integrates multiple trees through the idea of integrated learning. Its basic unit is decision trees, and each decision tree is a classifier. For an input sample, N trees will have N classification results. The random forest integrates all the classified voting results, and designates the category with the most voting times as the final output. Random forests can effectively train on large data sets, and have the advantages of fast training speed and high accuracy.

(38) Specifically, in the feature extraction stage, the features of the historical gram weight window and the cotton feeding roller historical window are extracted and input into the model. Finally, the predicted percentage of each label is output by the random forest algorithm.

(39) The second classifier adopts the recurrent neural network model LSTM shown in FIG. 9. The input is the output label confidence history window of classifier 1, and the output is the defined label.

(40) LSTM adopts cross entropy loss function, historical data is trained by a small batch gradient descent, and real-time data is fine-tuned by a random gradient descent. The hidden state output in the last time step aggregates the information of all time steps, and hidden state is input to the full connection layer activated by softmax to get the probability of each set label.

(41) LSTM can effectively aggregate information of multiple time steps, and automatically extract important features by using the feature screening function of deep network. LSTM uses a forget gate and an output gate to control the circulation and loss of features in each time step, which can effectively deal with the problem of long-term dependence in long time steps, and make the model learn the influence of early time steps on future time points.

(42) The control unit includes: a prediction model used to predict the cotton feeding roller value output by the system in the future according to the historical production line data and the predicted percentage of labels; an optimization sub-unit used for adjusting the cotton feeding roller value through the prediction model between a variable quantity of gram weight and the variable quantity of cotton feeding roller, using the prediction percentage of labels output by the classifying model as a membership degree of the classification of the production state, and outputting an optimal adjustment amount of cotton feeding roller based on a model conversion.

(43) After inputting the real-time data, the adjustment opinions will be output for the on-site operators to verify. The final adjustment opinions will be directly applied to the control system of the process line, and the console will change the specific parameters of the process line in real time according to the adjustment opinions.

(44) The goal of the control model is to predict the future change state of the gram weight based on the current prediction, and to adjust and feedback the parameters of the cotton feeding roller. Here the related concepts of automatic control theory are introduced to help adjust the feedback.

(45) In the field of industrial process control, for a long time, industrial controllers mainly adopt the control method based on feedback regulation. By feeding back the system response measured by sensors to the input of the controller, the function of tracking the set value of the system response is realized. FIG. 10 shows the conventional control system structure. PID control is such a control method based on feedback regulation. Although the conventional control scheme represented by PID can realize non-differential regulation, it is difficult to guarantee the rapid control and the timely response in the control of complex working conditions such as large delay and inertia.

(46) Referring to the basic ideas of model predictive control and fuzzy control, the control model proposed in this embodiment uses the following steps to control.

(47) Prediction model: the prediction model is the basis of model prediction control. The main function is to predict the future output of the system according to the historical information and future input of the object. Since the future output state has been predicted by the classifying model in the preceding step of the control model, the classifying model may be reused as the result of our prediction model. The prediction model also saves the running time of the model in online detection.

(48) Optimization control: the model predictive control determines the control function through the optimization of a certain performance indicator. In off-line training, the mathematical model between the change of gram weight and the change of cotton feeding roller will be established through the historical data of gram weight and cotton feeding roller, and the mathematical model is used as the basis for adjusting the value of the cotton feeding roller. Then, the label prediction possibility output by the classifying model is used as the membership degree of each classification, and the optimal adjustment amount of the cotton feeding roller is found through model conversion as the output of this optimization step. FIG. 11 is a schematic diagram of optimal control.

(49) The parameter control module includes a feedback correction unit. In order to prevent the control from deviating from the ideal state due to model mismatch or environmental interference, the feedback correction unit will make a new prediction at each new sampling moment, and use this real-time information to correct the prediction result based on the model, and then make a new optimization.

(50) The overall structure of the control model is shown in FIG. 12. The relations between the change of gram weight and the change of cotton feeding roller parameters can be better established, the opinions of parameter adjustment can be output to workers in an intelligent way, and the automatic adjustment can be completed, which saves a lot of manual work on the basis of ensuring accurate adjustment.

(51) FIG. 13 shows the process flow of applying the intelligent control system, including using the classifying model to predict the change trend in the future, and using the feedback model to predict the parameter change. This process optimizes the original manual intervention parameter adjustment process in FIG. 1, and completes an automatic closed loop of monitoring-adjusting-changing through intelligent control.

(52) The above-mentioned embodiments only describe the preferred mode of the application, but do not limit the scope of the application. On the premise of not departing from the design spirit of the application, all kinds of modifications and improvements made by ordinary technicians in the field to the technical scheme of the application shall fall within the scope of protection determined by the claims of the application.