Lightweight multiscale feature soft-sensing method for multisampling rates in industrial processes

Abstract

The present application relates to a light multi-scale feature soft measurement method for multi-sampling rate in industrial processing, comprising S1, determining the history length of the process variable; S2, grouping the process variable according to the difference of the sampling frequency; S3, designing a feature extractor, the feature extractor is used for feature extraction of each group of process variables, obtaining a plurality of groups of feature data groups of process variables, the frequency of all feature data groups is the same; S4, training the soft measurement model; S5, verifying whether the soft measurement model meets the requirements, if the soft measurement model meets the requirements, then the feature extractor is used for feature extraction of each group of process variables, and the obtained feature data group is input into the soft measurement model for prediction; if the soft measurement model does not meet the requirements, then steps S2 to S5 are repeated. The present application solves the problem of different sampling frequencies of multiple process variables, reduces the data volume, reduces the input of redundant and disturbance information, significantly reduces the parameter quantity of the soft measurement model, and improves the data processing.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of industrial prediction and process control, and particularly relates to a light multi-scale feature soft measurement method for multiple sampling rates in industrial processing. BACKGROUND

[0002] In modern industrial systems, due to technical or cost limitations, quality variables such as yield and composition are often analyzed offline, which will cause serious time delay, and the quality variables often have large hysteresis, so as to fail to meet real-time requirements such as real-time processing, control, monitoring and alarm.

[0003] In order to solve the problem, there are currently two technical solutions, that is, using advanced hardware equipment to measure quality variables or using soft measurement technology.

[0004] The advanced hardware equipment solution can effectively help enterprises realize online measurement, but has two main disadvantages. First, this type of hardware equipment is often high in cost, which is not conducive to unit cost control. Second, this type of hardware equipment can only provide online measurement, that is, it solves the problem of offline analysis, but cannot solve the hysteresis problem caused by the inertia link or hysteresis link of the industrial system. Therefore, the advanced hardware equipment solution cannot fundamentally solve the problem.

[0005] The soft measurement technology solution establishes a mathematical model between process variables and quality variables, and predicts the quality variables through process variables such as pressure, temperature and flow that can be measured online in real time. This solution does not need to increase additional hardware equipment, and solves the problem through a software solution. It is the focus of most enterprises and scholars. The common soft measurement technology solutions at present mainly include model-based methods and data-driven methods. For complex industrial processes, it is usually difficult to obtain accurate process mechanism models, but it is relatively easy to collect industrial production data, and these data are the manifestations of variables. Therefore, the data-driven soft measurement method has high value and strong universality. In the industrial field, the data-based soft measurement method is a kind of algorithm based on data analysis and independent of accurate mathematical models. This method can establish a mathematical model of the industrial system through process data.

[0006] Currently, the data-driven soft measurement techniques such as the multi-layer perception (MLP) technique or the PLS technique do not consider the characteristics of the industrial process variables with multiple sampling frequencies. That is, the current data-driven soft measurement scheme adopts the scheme of filling the low sampling frequency data to make it consistent with the high sampling frequency data in scale, and then establishing a "process variable-quality variable" soft measurement model for the data with different sampling frequencies. This will result in a large amount of repeated redundant information in the model input, and these useless information may introduce additional noise and reduce the model accuracy. In addition, this kind of scheme will result in a too large "process variable-quality variable" soft measurement model, which reduces the "timeliness" index in the soft measurement technology. SUMMARY

[0007] The technical problem to be solved by the present application is to provide a lightweight multi-scale feature soft measurement method for multiple sampling rates in industrial processing, which is designed for lightweight soft measurement technology suitable for multiple sampling frequency industrial processes without the need for the same scale filling processing of the original sampling data, reduces the input of redundant and disturbance information, significantly reduces the number of soft measurement model parameters under the premise of ensuring the accuracy of the soft measurement model.

[0008] To solve the above problems, the technical scheme adopted by the present application is as follows: a lightweight multi-scale feature soft measurement method for multiple sampling rates in industrial processing, comprising

[0009] S1, determining the history length of the process variables used;

[0010] S2, grouping the process variables according to the different sampling frequencies, the sampling frequencies of the process variables in each group being the same, and the sampling frequencies of the process variables in different groups being different;

[0011] S3, designing a feature extractor, the feature extractor being used for feature extraction of each group of process variables to obtain a plurality of groups of feature data of the process variables, so that the frequencies of all the groups of feature data are the same;

[0012] S4, training the soft measurement model;

[0013] S5, verifying whether the soft measurement model meets the requirements, if yes, using the feature extractor to extract features of each group of process variables, and inputting the obtained groups of feature data into the soft measurement model for prediction; if not, repeating steps S2 to S5.

[0014] Further, in step S3, the feature extraction process is as follows:

[0015] Let the measurable space be A probability P is defined on this set, and the probability space is Space of random variables with finite second moment At sampling time t, the quality variable Y(t) and the process variables X1(t), X2(t),..., then Y(t) = ψ([X(t)]), where X(t) = [X1(t), X2(t),...] T , ψ(·) is the true mapping function from Y(t) to X(t), i.e. the mapping to be estimated, there is an optimal Y(t k+1 ) approximation function So that:

[0016]

[0017] Where Y = [Y(t k ), Y(t k-1 ),..., Y(t1)] T , is the Borel function from Y to Y(t k+1 ), Φ (·) is the optimal Borel function, t k is the sampling time, E is the mathematical expectation;

[0018] Using higher sampling frequency, the process variable X containing more information to fit Y, formula (1) is transformed into:

[0019] E[Y(t k+1 )-Φ * (X)] 2 = inf Φ E[Y(t k+1 )-Φ(X)] 2 Formula (2)

[0020] Where X = [X(t x ), X(t k-1 ),..., X(t1)] T , Φ(·) is the Borel function from X to Y(t k+1 ), Φ * (·) is the optimal Borel function;

[0021] Formula (2) shows that there is Φ * (x) = E(Y(t k+1 )|X = x), so the process variable-quality variable transformation model Φ * (x) = E(Y(t k+1 )|x = x) can be obtained, according to the universal approximation theorem, there is an infinite neural network that approximates Φ * (·);

[0022] For the same sampling frequency T iProcess variable design feature extractor, feature is:

[0023]

[0024] wherein X i (t k ), X i-1 (t k ), … satisfy t k -t k-1 =1 / T i , is the optimal fitting function from X i (t k ), X i-1 (t k ), … to Y(t k+1 ), therefore, Further, in step S4, the soft measurement model adopts a multi-layer perception model.

[0025] The beneficial effects of the present application are: by grouping the process variables, the sampling frequency of each group of process variables is the same, and feature extraction is performed on each group of process variable data, so that the number of data of each group of process variables is the same within a certain time period, i.e. part of the data of the high-frequency sampling process variable is discarded, so that the data amount of the high-frequency sampling process variable is the same as that of the low-frequency sampling process variable, solving the problem of different sampling frequencies of multiple process variables, while reducing the data amount, reducing the input of redundant and disturbance information, significantly reducing the parameter amount of the soft measurement model, and improving data processing. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 is an industrial data schematic diagram;

[0027] Figure 2 is a soft measurement model structure schematic diagram of the present application based on a multi-scale feature extractor;

[0028] Figure 3 is a prediction error schematic diagram of the present application and the conventional technology;

[0029] Figure 4 is a prediction effect schematic diagram of the present application and the conventional technology;

[0030] Figure 5 is an effect schematic diagram of the present application and the conventional technology applied to wind power plant power prediction. DETAILED DESCRIPTION

[0031] The present application will be further described below in conjunction with the drawings and examples.

[0032] The light multi-scale feature soft measurement method for multi-sampling rates in industrial processing of the application comprises

[0033] S1, determine the history length of the process variable adopted.

[0034] Specifically, the grid search strategy is adopted to determine the optimal length of the history data for prediction.If the length of the history data is too long, additional noise is introduced, and the performance is reduced.If the length of the history data is too short, the information carried is insufficient for establishing an accurate soft measurement model.

[0035] S2, according to the different sampling frequencies, the process variables are grouped, the sampling frequencies of the process variables in each group are the same, and the sampling frequencies of the process variables in different groups are different.

[0036] S3, a feature extractor is designed, which is used for feature extraction of each group of process variables to obtain a plurality of groups of feature data of the process variables, so that the frequencies of all the feature data groups are the same.

[0037] Due to the design limitation of the sampling frequency of modern industrial equipment, industrial data can be roughly divided into equipment level data, process level data, quality level data and user data, which will lead to different sampling frequencies of industrial data. The sampling period of the equipment level data can be seconds or even milliseconds. However, the sampling period of the quality data or user data can be hours, resulting in different data quantities or data scales, as shown in Figure 1

[0038] Figure 1 As can be seen, the quantities of the collected equipment data, process data and quality data are significantly different in the same time period, and there are also differences in the quantities of the data collected by different equipment. Therefore, it is necessary to design a feature extractor for multi-sampling frequencies according to this feature, and then complete the data-driven soft measurement technology.

[0039] The application divides the process variables with the same frequency into a group, then performs feature extraction, extracts representative data, and discards the rest of the data. Therefore, the data quantity can be reduced, and in a time period, the quantities of the feature data extracted in each group of process variables are the same, that is, the frequencies of the feature data of all the process variables are the same, thereby solving the problem of different sampling frequencies of the process variables and reducing the data quantity. Compared with the same scale padding processing of the original sampling data, the input of redundant and disturbance information is reduced, the parameter quantity of the soft measurement model is significantly reduced under the premise of ensuring the accuracy of the soft measurement model.

[0040] For example, the sampling frequency of process variable 1 is 3 minutes, the sampling frequency of process variable 2 is 5 minutes, the sampling frequency of process variable 3 and process variable 4 is 2 minutes, process variable 1 is taken as the first group, process variable 2 is taken as the second group, and process variable 3 and process variable 4 are taken as the third group, feature extraction is performed on the collected data of the first group, the second group and the third group respectively, one feature data is extracted from every 10 data in the first group, one data is extracted from every 6 data in the second group, and one feature data is extracted from every 15 data in the third group, so that the data frequency of the first group, the second group and the third group of process variables is 30 minutes, solving the problem of different sampling frequencies and different data quantities of original process variables 1, 2, 3 and 4.

[0041] Feature extraction technology principle:

[0042] Let the measurable space be On this set, the probability P can be defined, and the probability space is Random variable space with finite second moment At the sampling time t, the quality variable Y(t) and the process variables X1(t), X2(t),..., then Y(t) = ψ([X(t)]), where X(t) = [X1(t), X2(t),...] T , ψ(·) is the true mapping function of Y(t) to X(t), that is, the mapping to be estimated. There is an optimal Y(t k+1 ) approximation function So that:

[0043]

[0044] Where Y = [Y(t k ), Y(t k-1 ),..., Y(t1)] T , is the Borel function from Y to Y(t k+1 ), is the optimal Borel function, t k is the sampling time, and E is the mathematical expectation;

[0045] Because the sampling frequency of the quality variable is often much lower than that of the process variable, the data quantity of the quality variable is small and the information quantity is sparse. Therefore, the process variable X with higher sampling frequency and richer information can be used to fit Y. Based on the above formula (1) can be transformed into:

[0046] E[Y(t k+1 )-Φ * (X)] 2 =inf Φ E[Y(tk+1 )-Φ(X)] 2 Formula (2)

[0047] Where, X = [X(t)] k ), X(t) k-1 ), ..., X(t1)] T Φ(·) is from X to Y(t) k+1 The Borel function of Φ * (·) is the optimal Borel function (i.e., the optimal estimation mapping of ψ(·));

[0048] Formula (2) shows that there exists Φ * (x)=E(Y(t k+1 Therefore, we can obtain the transformation model of process variable to quality variable: Φ |X=x). * (x)=E(Y(t k+1 According to the universal approximation theorem, there exists an infinite approximation Φ(x)|X=x). * A multilayer neural network of (x);

[0049] For the same sampling frequency T i Process variables (i.e., those satisfying {X) i (t k )|t k -t k-1 =1 / T i Design a feature extractor, features It can be written as:

[0050]

[0051] Among them, X i (t k ), X i-1 (t k ), ... satisfy t k -t k-1 =1 / T i , For from X i (t k ), X i-1 (t k ), ... to Y(t) k+1 The optimal fitting function for ). Therefore,

[0052] The feature extractor described above is used to extract the feature data of each group of process variables. According to equation (3), the features at the extraction points of process data of different frequencies are denoted as follows: The features obtained by this way greatly reduce the redundant information introduced by filling data in traditional data-driven based soft-sensing technology, thus the parameter amount of soft-sensing model can be significantly reduced.

[0053] S4, training the soft-sensing model.

[0054] The soft-sensing model adopts a multi-layer perception (MLP), and the multi-layer perception (MLP model fitting The MLP structure is (i+1)*[w / T i ] / 2° (i+1)*[w / T i ] / 2 1 (i+1)*[w / T i ] / '2 2 ……, wherein w is a time window required for predicting Y(t k+1 ) and is the optimal length of historical data for prediction.

[0055] The soft-sensing model structure based on the multi-scale feature extractor obtained by the application is shown in Figure 2 The feature extraction is performed first, and then the prediction is performed.

[0056] The order of the above steps S2, S3 and S4 can be changed.

[0057] S5, verifying whether the soft-sensing model meets the requirements, if yes, the feature extractor is used to extract features from each group of process variables, and the obtained feature data group is input into the soft-sensing model for prediction; if not, steps S2 to S5 are repeated.

[0058] The Tennessee Eastman Process is used to test the performance of the application, 22 equipment variables with a sampling period of 3 minutes and 14 process variables with a sampling period of 6 minutes are sampled, and the soft-sensing model of the application, the traditional MLP and the traditional PLS are respectively used to perform soft-sensing on the quality variables, the prediction errors of the three are shown in Figure 3 , and the prediction effect is shown in Figure 4 From the figure, it can be seen that the technical solution proposed by the application can significantly improve the prediction accuracy. In addition, the parameter amount required by the soft-sensing model based on the multi-scale feature extractor proposed by the application is reduced by 60.9% compared with the MLP algorithm.

[0059] Example One

[0060] The soft-sensing method of the application is used for soft-sensing of the power of a certain wind power plant, and the prediction effect is shown in Figure 5 ​As shown, it can be seen that the prediction result of the application is closest to the true value, and the prediction effect is best. In addition, the amount of parameters required by the soft measurement method of the application is reduced by 71.5%.

[0061] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. The present application can be variously changed and modified by those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A lightweight, multi-scale feature soft measurement method for multiple sampling rates in industrial processing, characterized in that, include S1. Determine the historical length of the process variables used; S2. Based on the different sampling frequencies, the process variables are grouped, with each group of process variables having the same sampling frequency and different sampling frequencies for different groups of process variables. S3. Design a feature extractor. The feature extractor is used to extract features from each set of process variables to obtain feature data sets of multiple sets of process variables, so that all feature data sets have the same frequency. S4. Train the soft measurement model; S5. Verify whether the soft sensor model meets the requirements. If it does, use a feature extractor to extract features from each group of process variables and input the obtained feature data into the soft sensor model for prediction. If it does not meet the requirements, repeat steps S2 to S5.

2. The lightweight multi-scale feature soft measurement method for multiple sampling rates in industrial processing as described in claim 1, characterized in that, In step S3, the feature extractor design process is as follows: Assume measurable space Define a probability P on this set, and its probability space. Space of random variables with finite second moments At sampling time t, given the quality variable Y(t) and process variables X1(t), X2(t), ..., then Y(t) = ψ([X(t)]), where X(t) = [X1(t), X2(t), ...] T Let ψ(·) be the true mapping function from Y(t) to X(t), i.e., the mapping to be estimated, and let there exist optimal Y(t). k+1 Approximate function Make: Where, Y = [Y(t)] k ), Y(t) k-1 ), ..., Y(t1)] T , From Y to Y(t) k+1 Borel function, For the optimal Borel function, T k Let E be the sampling time, and E be the expected value. By using the process variable X, which has a higher sampling frequency and contains richer information, to fit Y, formula (1) is transformed into: E[Y(t k+1 ) - Φ * (X)] 2 = inf Φ E[Y(t k+1 ) - Φ(X)] 2 Equation (2) Where, X = [X(t)] k ), X(t) k-1 ), ..., X(t1)] T Φ(·) is from X to Y(t) k+1 The Borel function of Φ * (·) represents the optimal Borel function; Formula (2) shows that there exists Φ * (x)=E(Y(t k+1 Therefore, we can obtain the transformation model of process variable to quality variable: Φ |X=x). * (x)=E(Y(t k+1 According to the universal approximation theorem, there exists an infinite approximation Φ(x)|X=x). * A multilayer neural network of (x); For the same sampling frequency T i Design a feature extractor for process variables, features for: Among them, X i (t k ), X i-1 (t k ), ... satisfy t k -t k-1 =1 / T i , For from X i (t k ), X i-1 (t k ), ... to Y(t) k+1 The optimal fitting function of ), therefore, 3. The lightweight multi-scale feature soft measurement method for multiple sampling rates in industrial processing as described in claim 1 or 2, characterized in that, In step S4, the soft measurement model adopts a multilayer perceptron model.

Images