Industrial process multi-sampling-rate data-oriented index soft measurement method

Abstract

The invention provides an index soft measurement method for industrial process multi-sampling-rate data, and the method comprises the steps: collecting process variables and quality index operation data in an industrial process, and carrying out the grouping and sorting of the process variables according to sampling rates, and obtaining the multi-sampling-rate data, the method comprises the following steps: dividing multi-sampling-rate data and quality index operation data into a training set and a test set, preprocessing the training set and the test set, constructing a block diagonal attention shielding module, carrying out self-supervised training based on the training set to obtain potential features of process variables, constructing a convolutional neural network module, and obtaining process variables; and performing supervised training based on the potential features of the process variables and the quality index operation data in the training set, constructing a soft measurement model, and inputting the test set into the constructed soft measurement model to obtain an estimated value of a related quality index. According to the method, the defects in the prior art can be overcome, and the soft measurement accuracy of the multi-sampling-rate industrial process quality index is improved.

Description

Technical Field

[0001] This invention relates to the field of soft measurement technology for industrial processes, and in particular to a soft measurement method for indicators of industrial process multi-sampling rate data. Background Technology

[0002] In modern industrial production, ensuring the controllability and stability of product quality is crucial not only for the smooth operation of the production process but also for the cost control and economic benefits of enterprises. However, many key quality indicators are difficult to obtain through direct measurement, often requiring time-consuming offline testing or complex measurement methods, which poses a significant challenge to real-time monitoring. Therefore, soft measurement technology has become an important means to solve this problem. At the same time, because the sampling frequency of quality indicators is much lower than that of process variables, this difference in sampling rate leads to the scarcity of data labels, further increasing the difficulty of real-time monitoring and modeling.

[0003] Due to limitations in measurement technology and hardware costs, many process variables in industrial processes cannot typically be sampled at a uniform sampling rate. For example, process variables such as temperature are usually sampled on a minute-by-minute basis, while process variables such as concentration may be sampled on an hourly basis. Therefore, the problem of process variables with multiple sampling rates is a pressing issue in industrial processes. Traditional upsampling interpolation methods lead to the gradual accumulation of errors, ultimately affecting the accuracy of soft measurement model predictions; while using simple downsampling methods results in a significant waste of data, and the reduced data volume may lead to underfitting of the soft measurement model. Therefore, it is essential to design an index soft measurement method based on block-diagonal occlusion attention-convolutional neural networks for multi-sampling rate data in industrial processes. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a soft measurement method for indicators of multi-sampling rate data in industrial processes.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] This invention provides a soft measurement method for indicators based on multi-sampling rate data in industrial processes, comprising:

[0007] Step 100: Collect process variables and quality index operation data in the industrial process, and group and sort the process variables according to the sampling rate to obtain multi-sampling rate data;

[0008] Step 200: Divide the multi-sampling rate data and quality index running data into training set and test set, and preprocess them;

[0009] Step 300: Construct a block-based diagonal occlusion attention module, perform self-supervised training based on the training set, and obtain the latent features of the process variables;

[0010] Step 400: Construct a convolutional neural network module, perform supervised training based on the latent features of process variables and the quality indicators in the training set, and build a soft measurement model;

[0011] Step 500: Input the test set into the constructed soft measurement model to obtain the estimated values ​​of the relevant quality indicators.

[0012] Preferably, in step 100, the process variables are grouped and sorted according to the sampling rate to obtain multi-sampling rate data, specifically as follows:

[0013] Process variables with the same sampling rate are grouped together, and the process variable groups are sorted from high to low sampling rate to obtain multi-sampling rate data.

[0014] Preferably, in step 200, the preprocessing specifically includes:

[0015] Step 201: Standardize the process variables and quality indicators in the training and test sets respectively. The standardization formula is as follows:

[0016]

[0017] In the formula, X and Z represent all sample values ​​of the variable before and after standardization, μ and σ represent the mean and standard deviation of variable X, respectively, and the process variable of the training set obtained after standardization is X. norm ;

[0018] Step 202: Select appropriate review window length L and prediction window length P based on the scale of the original industrial process data and actual requirements, and divide the process variables and quality index operation data into time windows of corresponding lengths with a step size of 1.

[0019] Preferably, in step 300, a segmented diagonal occlusion attention module is constructed, and self-supervised training is performed based on the training set to obtain the latent features of the process variables, specifically including:

[0020] Step 301: Perform artificial data masking on the high-sampling-rate process variable group to make it have the same sampling rate as the low-sampling-rate process variable group, thus obtaining the process variables. According to the current Obtaining the masking matrix in the case of missing data and indicator matrix I, where:

[0021]

[0022]

[0023] Step 302: Determine the block length L1 and step size P1 based on the review window length L, perform time window block division, and perform dimensional projection and time encoding embedding on the variable samples in each block;

[0024] Step 303: Feed the processed data into the diagonal occlusion attention block for self-supervised training to obtain the estimated process variables. Given the latent features h, the loss function for self-supervised training is:

[0025]

[0026] Where z M z I Represent the occlusion matrix respectively The number of 1 values ​​in the indicator matrix I, where α represents the manually set weight.

[0027] Preferably, the diagonal occlusion attention block consists of a diagonal occlusion multi-head self-attention layer, a residual connection and normalization layer, a feedforward neural network layer, and another residual connection and normalization layer. In self-supervised training, the diagonal occlusion attention block is stacked and used N times to generate latent features h of the process variable. Finally, the estimated process variable is obtained through a linear layer.

[0028] Preferably, the convolutional neural network module consists of Q convolutional layers, dropout layers, and R fully connected layers. The size of Q and the hyperparameters of each convolutional layer are determined by the latent feature dimension of the process variable and the prediction length P of the quality index. The size of R is the same as the number of quality indices.

[0029] Preferably, the fully connected layer module consists of two fully connected layers, and the dimension of the intermediate layer is determined based on the potential feature dimension of the process variable and the prediction length P of the quality index.

[0030] Preferably, in step 400, the loss function for supervised training is RMSE, which is:

[0031]

[0032] In the formula, y represents the true value of the quality indicator. denoted as the predicted value of the quality indicator, and z represents the number of observable data points in the quality indicator.

[0033] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0034] This invention provides a soft measurement method for indicators based on multi-sampling rate data in industrial processes. The method includes: collecting operational data of process variables and quality indicators in industrial processes; grouping and sorting the process variables according to the sampling rate to obtain multi-sampling rate data; dividing the multi-sampling rate data and the operational data of quality indicators into training sets and test sets, and preprocessing them; constructing a block-diagonal occlusion attention module; performing self-supervised training based on the training set to obtain the latent features of the process variables; constructing a convolutional neural network module; performing supervised training based on the latent features of the process variables and the operational data of quality indicators in the training set to construct a soft measurement model; and inputting the test set into the constructed soft measurement model to obtain the estimated values ​​of the relevant quality indicators.

[0035] The present invention has the following advantages:

[0036] 1. By combining the self-supervised training of the block-diagonal occlusion attention module with the supervised training of the convolutional neural network module, the potential features in the process variables can be fully explored, reducing the dependence on data labels, i.e., quality indicator data. Even when data labels are scarce, the modeling ability and prediction accuracy of the soft measurement model can still be effectively improved.

[0037] 2. By masking high-sampling-rate variables to maintain consistency with low-sampling-rate variables in the time dimension, and combining a block-diagonal masking attention mechanism to uniformly model multi-sampling-rate data, the shortcomings of traditional upsampling interpolation methods, such as error accumulation and data waste caused by simple downsampling, are overcome. This significantly improves the utilization efficiency of multi-sampling-rate industrial process data and the prediction accuracy of soft measurement models. Attached Figure Description

[0038] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0039] Figure 1 A flowchart of the method provided in an embodiment of the present invention;

[0040] Figure 2 This is a schematic diagram illustrating the construction of the segmented diagonal occlusion attention module of the present invention;

[0041] Figure 3 This is a schematic diagram of the diagonal attention block of the present invention. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] The purpose of this invention is to provide a soft measurement method for indicators based on multi-sampling rate data of industrial processes, which can overcome the shortcomings of existing technologies and improve the accuracy of soft measurement of quality indicators of multi-sampling rate industrial processes.

[0044] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0045] Figure 1 The method flowchart provided in the embodiments of the present invention is as follows: Figure 1 As shown, this invention provides a soft measurement method for indicators based on multi-sampling rate data in industrial processes, comprising:

[0046] Step 100: Collect process variables and quality index operation data in the industrial process, and group and sort the process variables according to the sampling rate to obtain multi-sampling rate data;

[0047] Step 200: Divide the multi-sampling rate data and quality index running data into training set and test set, and preprocess them;

[0048] Step 300: Construct a block-based diagonal occlusion attention module, perform self-supervised training based on the training set, and obtain the latent features of the process variables;

[0049] Step 400: Construct a convolutional neural network module, perform supervised training based on the latent features of process variables and the quality indicators in the training set, and build a soft measurement model;

[0050] Step 500: Input the test set into the constructed soft measurement model to obtain the estimated values ​​of the relevant quality indicators.

[0051] In step 100, the process variables are grouped and sorted according to the sampling rate to obtain multi-sampling rate data, specifically as follows:

[0052] Process variables with the same sampling rate are grouped together, and the process variable groups are sorted from high to low sampling rate to obtain multi-sampling rate data.

[0053] In step 200, the multi-sampling rate data and the quality index running data are randomly divided in a ratio of 8:2 to obtain the training set and the test set.

[0054] In step 200, the preprocessing specifically includes:

[0055] Step 201: Standardize the process variables and quality indicators in the training and test sets respectively. The standardization formula is as follows:

[0056]

[0057] In the formula, X and Z represent all sample values ​​of the variable before and after standardization, μ and σ represent the mean and standard deviation of variable X, respectively, and the process variable of the training set obtained after standardization is X. norm ;

[0058] Step 202: Select appropriate review window length L and prediction window length P based on the scale of the original industrial process data and actual requirements. Divide the process variables and quality index operation data into time windows of corresponding lengths with a step size of 1. In order for supervised training to operate normally, P should contain at least one observable value of a quality index.

[0059] like Figure 2 As shown, in step 300, a block-based diagonal occlusion attention module is constructed, and self-supervised training is performed based on the training set to obtain the latent features of the process variables, specifically including:

[0060] Step 301: Perform artificial data masking on the high-sampling-rate process variable group to make it have the same sampling rate as the low-sampling-rate process variable group, thus obtaining the process variables. According to the current Obtaining the masking matrix in the case of missing data and indicator matrix I, where:

[0061]

[0062] By masking the high-sampling-rate process variable group to the same sampling rate as the low-sampling-rate variable group, the model can discover the potential patterns of the high-sampling-rate data during the process of reconstructing the masked data in self-supervised training, adapt to the characteristics of sparse data, and thus enhance its understanding and completion ability of the low-sampling-rate data.

[0063] Step 302: Determine the block length L1 and step size P1 based on the review window length L, perform time window block division, and perform dimensional projection and time encoding embedding on the variable samples in each block;

[0064] The time window segmentation performed here can extract local information from the process variable data, allowing the model to focus on the characteristics of different regions of the data, thereby better learning the local patterns of the process variables; in addition, using time blocks as the smallest unit of model operation can significantly reduce computational complexity.

[0065] Step 303: Feed the processed data into the diagonal occlusion attention block for self-supervised training to obtain the estimated process variables. Given the latent features h, the loss function for self-supervised training is:

[0066]

[0067] In the formula, z M 、z I Represent the occlusion matrix respectively The number of 1 values ​​in the indicator matrix I, where α represents the manually set weight;

[0068] In the loss function of self-supervised training Some of these are observable reconstruction terms, which guide the model to reconstruct observables, thereby learning the overall distribution of the data. The loss function includes a component for reconstructing obscured values, guiding the model to reconstruct these values ​​and recover missing data. If the loss function only includes observable value reconstruction terms, the model will tend to minimize the reconstruction error of the obscured values ​​while ignoring the prediction of the obscured values, making it difficult to effectively complete the missing data. The weights in the loss function can be dynamically adjusted based on the soft measurement effect of the quality metrics.

[0069] like Figure 3 As shown, the diagonal occlusion attention block consists of a diagonal occlusion multi-head self-attention layer, a residual connection and normalization layer, a feedforward neural network layer, and another residual connection and normalization layer. In self-supervised training, the diagonal occlusion attention block is stacked and used N times to generate latent features h of the process variable. Finally, the estimated process variable is obtained through a linear layer.

[0070] because There are a large number of masked values ​​in the data. In order to effectively handle these missing information, a diagonal masking self-attention mechanism is adopted. The diagonal masking self-attention mechanism constructs a main diagonal masking matrix and sets the attention weights on the main diagonal to 0, thereby forcing the model to infer the value of the current time step from the context information. This design not only effectively enhances the model's ability to complete masked data, but also can more fully explore the local and global dependencies in the time series data, thereby improving the accuracy and robustness of prediction.

[0071] In step 400, a convolutional neural network module is constructed. Supervised training is performed based on the latent features of the process variables and the quality indicators in the training set to build a soft measurement model. Specifically:

[0072] The convolutional neural network module consists of Q convolutional layers, dropout layers, and R fully connected layers. The size of Q and the hyperparameters of each convolutional layer are determined by the latent feature dimension of the process variable and the prediction length P of the quality index. The size of R is the same as the number of quality indices.

[0073] The fully connected layer module consists of two fully connected layers, and the dimension of the intermediate layer is determined based on the potential feature dimension of the process variable and the prediction length P of the quality index.

[0074] The loss function for supervised training is RMSE, which is:

[0075]

[0076] In the formula, y represents the true value of the quality indicator. denoted as the predicted value of the quality indicator, and z represents the number of observable data points in the quality indicator.

[0077] The effectiveness of this invention will be verified using Tennessee-Eastman (TE) process data. Components G and H were selected as the key quality indicators to be predicted for the TE process, with a sampling rate of 12 minutes. Eleven manipulated variables and 22 process variables were used as input variables, with the manipulated variables and the first 11 process variables sampled at a rate of 3 minutes, and the latter 11 process variables sampled at a rate of 6 minutes.

[0078] The present invention was compared with the soft measurement effects of upsampled convolutional neural network U-CNN, downsampled convolutional neural network D-CNN, upsampled multilayer neural network U-MLP, downsampled multilayer neural network D-MLP, upsampled long short-term memory network U-LSTM, and downsampled long short-term memory network D-LSTM. The experimental results are shown in Table 1.

[0079] Table 1 Comparison of Experimental Results for Each Method on the Test Set

[0080]

[0081] The bolded data in the table represents the optimal values. As shown in Table 1, the present invention has the highest accuracy in predicting components G and H in most cases.

[0082] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0083] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A soft-sensing method for industrial process multi-sampling rate data, characterized in that: include: Step 100: Collecting process variables and quality indicator operation data in the industrial process, and grouping and sorting the process variables according to the sampling rate to obtain multi-sampling rate data; Step 200: Divide the multi-sampling rate data and the quality indicator operation data into a training set and a test set, and pre-process them; Step 300: construct a block diagonal masking attention module, perform self-supervised training based on the training set, and obtain the latent features of the process variables; Step 400: construct a convolutional neural network module, perform supervised training based on the potential characteristics of process variables and the quality indicator operation data in the training set, and construct a soft measurement model; Step 500: Input the test set into the constructed soft measurement model to obtain the estimated value of the relevant quality index.

2. The method according to claim 1, characterized in that In step 100, the process variables are grouped and sorted according to the sampling rate to obtain multi-sampling rate data, specifically: The process variables with the same sampling rate are grouped into one group, and the process variable groups are sorted from high to low according to the sampling rate to obtain multi-sampling rate data.

3. The method according to claim 1, characterized in that In step 200, the preprocessing process specifically includes: Step 201: Standardize the process variables and quality indicator operation data in the training set and the test set respectively. The standardization formula is: In the formula, X and Z represent all sample values ​​of the variable before and after standardization, μ and σ represent the mean and standard deviation of variable X respectively. The training set process variable obtained after standardization is X norm ; Step 202: According to the scale of the original data of the industrial process and the actual requirements, a suitable review window length L and prediction window length P are selected, and the process variables and quality indicator operation data are divided into time windows of corresponding lengths and a step length of 1.

4. The method according to claim 1, characterized in that In step 300, a block diagonal masking attention module is constructed, and self-supervised training is performed based on the training set to obtain the potential features of the process variables, specifically including: Step 301: Perform artificial data masking on the high sampling rate process variable group to make it have the same sampling rate as the low sampling rate process variable group, and obtain the process variable According to the current The masking matrix is ​​obtained in the case of missing data. and indicator matrix I, where: Step 302: Determine the block length L1 and step length P1 according to the review window length L, perform time window block division, and perform dimension projection and time coding embedding on the variable samples in each block; Step 303: Send the processed data to the diagonal masked attention block for self-supervised training to obtain the estimated process variables And the latent feature h, the loss function of self-supervised training is: In the formula, z M 、z I Represent the masking matrix Indicates the number of 1 values ​​in the matrix I, and α represents the artificially set weight.

5. The method according to claim 4, characterized in that The diagonal masked attention block consists of a diagonal masked multi-head self-attention layer, a residual connection and normalization layer, a feedforward neural network layer, and another residual connection and normalization layer. In self-supervised training, the diagonal masked attention block is stacked N times to generate the process variable potential feature h, and finally the estimated process variable is obtained through a linear layer.

6. The method according to claim 1, characterized in that The convolutional neural network module consists of Q convolutional layers, a dropout layer and R fully connected layer modules. The size of Q and the hyperparameters of each convolutional layer are determined by the potential feature dimension of the process variable and the prediction length P of the quality indicator. The size of R is the same as the number of quality indicators.

7. The method according to claim 6, characterized in that The fully connected layer module consists of two fully connected layers, and the dimension of the middle layer is determined according to the potential feature dimension of the process variable and the prediction length P of the quality indicator.

8. The method according to claim 7, characterized in that In step 400, the loss function of supervised training is RMSE, which is: In the formula, y is the true value of the quality indicator, is the predicted value of the quality indicator, and z is the number of data that can be observed in the quality indicator.

Images