Industrial monitoring method, system, device and storage medium based on soft measurement model

Abstract

The present invention relates to an industrial monitoring method, system, device and storage medium based on a soft measurement model, and belongs to the field of soft measurement technology. The present invention monitors quality variables in industrial processes by constructing a soft measurement model based on a multi-scale gated Transformer, and realizes the effective fusion of global and local features by designing a hierarchical feature fusion structure. First, the convolution operation is embedded in the attention mechanism to extract the multi-scale features of the data; secondly, the multi-scale feature weights are adaptively adjusted using the gated unit, and low-correlation or redundant features are eliminated through the sparse attention mechanism, thereby enhancing the feature extraction capability of the model in the case of data loss. The method is applied to the debutanizer industrial process to predict butane concentration. The results show that the present invention has higher prediction accuracy than the existing soft measurement industrial monitoring method. The present invention also provides an industrial monitoring system, device and storage medium based on the soft measurement model, which improves the convenience of industrial monitoring.

Description

Technical Field

[0001] The invention relates to an industrial monitoring method, system, device and storage medium based on a soft measurement model, and belongs to the technical field of soft measurement. Background Art

[0002] Complex industrial processes contain many difficult-to-measure variables, which often directly impact product quality and production efficiency and are therefore crucial to the process. Therefore, real-time monitoring of key process parameters is essential. However, traditional instrumentation and offline testing methods are time-consuming and labor-intensive, and often result in significant delays in measurement results. To better monitor these industrial variables in real time, soft sensing methods have been proposed. These methods use easily measurable variables (i.e., auxiliary variables) as input and difficult-to-measure variables (i.e., dominant variables) as output.

[0003] Soft sensor modeling primarily involves four steps: selecting auxiliary variables, data acquisition and preprocessing, establishing a soft sensor model, and online calibration. Establishing a soft sensor model is the core of soft sensor modeling. Soft sensor models are primarily categorized into two types: mechanism models and data-driven models. Soft sensor methods based on mechanism modeling require an analysis of the underlying principles of chemical processes. However, as industrial processes become increasingly complex, the study of mechanisms is also becoming increasingly complex. Relying on extensive experience to comprehensively analyze industrial processes is unable to meet actual industrial needs. Compared to mechanism modeling, data-driven soft sensor modeling relies solely on process data, effectively improving modeling efficiency and thus attracting widespread attention. Numerous data-driven soft sensor methods exist, primarily including statistical learning, machine learning, and deep learning.

[0004] Deep learning, with its ability to automatically extract features from multi-layered nonlinear structures, has demonstrated significant advantages in processing complex industrial data in the field of soft sensing. Typical deep learning methods include convolutional neural networks (CNNs), recurrent neural networks (RNNs), long short-term memory (LSTM) networks, and transformer networks. As a popular technology, the transformer, with its self-attention mechanism, can process sequential data in parallel, overcoming the limitations of traditional RNN and LSTM networks in processing long sequences. It effectively captures global data dependencies, exhibiting higher computational efficiency and performance advantages, and has become a hot topic and mainstream research direction in the field of deep learning soft sensing.

[0005] Deep learning methods have become a better solution for the soft measurement modeling problems of nonlinearity, dynamics and time series correlation of actual industrial process data. However, in actual industrial processes, sensor data collection is periodic, making it difficult to obtain all variables at the same time. This sampling method with periodic sampling characteristics leads to data missing and hysteresis, which brings huge challenges to soft measurement prediction. This application mainly addresses the data missing problem caused by the inconsistent sampling period of auxiliary variables and quality variables, that is, the dual sampling rate problem. For soft measurement data with different sampling rates, Chang et al. proposed an improved RNN network that divides and maps data with different sampling rates into a common hidden space, updates the hidden space at each moment, and thus integrates and unifies the information to predict the quality variables (see Chang S, Chen X, Zhao C. Flexible Clockwork Recurrent Neural Network for multirate industrial softsensor[J]. Journal of Process Control, 2022, 119: 86-100.). Song et al. proposed a deep learning model based on a multi-temporal channel CNN. This model uses a temporal feature sharing network and a parallel prediction network for variable prediction. By improving the backpropagation algorithm, blank values ​​in the data are excluded from backpropagation (see Song B, Zhou Y, Shi H, et al. A soft sensor for multirate quality variables based on MC-CNN[J]. IEEE Transactions on Neural Networks and Learning Systems, 2024. DOI: 10.1109 / tnnls.2024.3360030.). However, both CNNs and RNNs have shortcomings when processing data with long-term dependencies. CNNs are limited by the size of the convolution kernel, and RNNs may suffer from the vanishing gradient problem, which further affects the accuracy of industrial data monitoring and the generalization ability of the model. Summary of the Invention

[0006] In order to improve the prediction accuracy of soft measurement models and the convenience of industrial monitoring, the present invention provides an industrial monitoring method, system, device and storage medium based on soft measurement models. The technical solution is as follows:

[0007] A first object of the present invention is to provide an industrial monitoring method based on a soft sensor model, the method comprising:

[0008] Step 1: Periodically collect easily measurable auxiliary variables and historically difficult-to-measurable leading variables in the industrial process;

[0009] Step 2: Use the MGT model to extract features from the easily measurable auxiliary variables and the historically unmeasurable dominant variables, and perform regression prediction on the current unmeasurable dominant variables;

[0010] The MGT model divides the input easy-to-measure auxiliary variables into four attention subspaces, extracts local features in each subspace using the convolutional layer within the multi-scale attention mechanism, and uses the attention mechanism to extract global features; secondly, a trainable gating unit is used to adaptively assign weights to fuse multi-scale features; then, a sparse attention mechanism is used to perform a sparse operation on the fused features; finally, a multi-scale gated attention output is obtained, and after fusing the features of each subspace, a fully connected layer decoding is used for regression prediction.

[0011] Optionally, the calculation process of the multi-scale gated attention output is:

[0012]

[0013]

[0014]

[0015]

[0016] in, is the fused feature matrix, is the attention score after filtering by the TopK operation. The TopK operation indicates the selection of Before Attention score; Representation matrix No. i Rank j Elements of the column, Represents a sparse matrix of features No. i Rank j Elements of the column, express No. i row vector; represents the sparse attention score matrix after normalization operation, Represents a small constant to avoid the denominator being 0. represents the multi-scale gated attention output.

[0017] Optionally, the fused feature matrix Expressed as:

[0018]

[0019]

[0020] in, Represents the global feature, the value is the standard attention score, represents the gate control unit, Represents the dot product operation between vectors, Express Q Perform convolution operation, is the query vector Q and key vector K dimension.

[0021] Optionally, the gate control unit is represented as:

[0022]

[0023] in, and is the weight matrix and bias vector of the gate unit, is the Sigmoid function.

[0024] Optionally, the construction process of the soft sensor model includes:

[0025] Step 1: Determine the easy-to-measure auxiliary variables and the difficult-to-measure dominant variables, which serve as the input samples and output samples of the soft sensor model respectively. Construct a data set and divide it into a training set, a validation set, and a test set.

[0026] Step 2: Construct the MGT model;

[0027] Step 3: Initialize the parameters of the MGT model, train the model using the training set, perform forward propagation and calculate the error; calculate the gradient of the loss function with respect to the model parameters using the backpropagation algorithm; and update the model parameters using the Adam optimization algorithm.

[0028] Step 4: Evaluate the MGT model on the validation set, use grid search to find model parameters and use the validation set to adjust the hyperparameters, and use the trained model for test set prediction;

[0029] Step 5: Input the test set into the trained model to evaluate the predictive ability.

[0030] Optionally, the method further includes position encoding the input data, which is expressed as:

[0031]

[0032]

[0033] in, i represents the index of the dimension, POSIndicates the position where position encoding is performed, d m is the embedding dimension of the data, PE Represents the output after position encoding.

[0034] Optionally, the MSE loss function is used to calculate the error during training.

[0035] A second object of the present invention is to provide an industrial monitoring system based on a soft sensor model, the system comprising:

[0036] A data acquisition module is configured to periodically collect easily measurable auxiliary variables and historically difficult-to-measurable leading variables in industrial processes;

[0037] A prediction module is configured to use the MGT model to perform feature extraction on the easily measurable auxiliary variables and historically unmeasurable leading variables, and to perform regression prediction on the current unmeasurable leading variables;

[0038] The MGT model divides the input easy-to-measure auxiliary variables into four attention subspaces, extracts local features in each subspace using the convolutional layer within the multi-scale attention mechanism, and uses the attention mechanism to extract global features; secondly, a trainable gating unit is used to adaptively assign weights to fuse multi-scale features; then, a sparse attention mechanism is used to perform a sparse operation on the fused features; finally, a multi-scale gated attention output is obtained, and after fusing the features of each subspace, a fully connected layer decoding is used for regression prediction.

[0039] A third object of the present invention is to provide an industrial monitoring device based on a soft sensor model, comprising a memory and a processor;

[0040] The memory is used to store computer programs;

[0041] The processor is configured to implement the industrial monitoring method based on the soft measurement model as described in any one of the above items when executing the computer program.

[0042] A fourth object of the present invention is to provide a computer-readable storage medium, characterized in that a computer program is stored on the storage medium, and when the computer program is executed by a processor, the industrial monitoring method based on the soft measurement model as described in any one of the above items is implemented.

[0043] The beneficial effects of the present invention are:

[0044] This paper monitors quality variables in industrial processes by constructing a soft-sensing model based on a multi-scale gated Transformer. By designing a hierarchical feature fusion structure, it effectively integrates global and local features. First, convolution operations are embedded within an attention mechanism to extract multi-scale features from the data. Second, gating units are used to adaptively adjust the weights of multi-scale features, and a sparse attention mechanism is used to remove low-correlation or redundant features, thereby enhancing the model's feature extraction capabilities in the absence of data. This approach, when applied to industrial processes such as debutanizers and sulfur recovery, can effectively improve monitoring accuracy.

[0045] Experimental results show that compared with other existing soft sensor model industrial monitoring methods, the present invention exhibits higher prediction accuracy and better generalization ability.

[0046] The present invention provides an industrial monitoring system and device based on a soft measurement model and a computer-readable storage medium, thereby improving the convenience of industrial monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of 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 creative work.

[0048] Figure 1 It is a description diagram of dual sampling rate data.

[0049] Figure 2 This is the structural diagram of the proposed MGT model.

[0050] Figure 3 It is a multi-scale gated attention structure diagram.

[0051] Figure 4 This is the process flow diagram of the debutanizer tower.

[0052] Figure 5A The sampling period is 3 T d Butane concentration prediction curve of each model at .

[0053] Figure 5B The sampling period is 4 T d Butane concentration prediction curve of each model at .

[0054] Figure 5C The sampling period is 5 T d Butane concentration prediction curve of each model at .

[0055] Figure 5D The sampling period is 6 T d Butane concentration prediction curve of each model at .

[0056] Figure 6A The sampling period is 3 T d Scatter plot of the prediction results of each network model at .

[0057] Figure 6B The sampling period is 4 T d Scatter plot of the prediction results of each network model at .

[0058] Figure 6C The sampling period is 5 T d Scatter plot of the prediction results of each network model at .

[0059] Figure 6D The sampling period is 6 T d Scatter plot of the prediction results of each network model at .

[0060] Figure 7A This is a graph showing the changes in the RMSE values ​​of the prediction results of each network model in the debutanizer process at different sampling rates.

[0061] Figure 7B is the prediction result R of each network model in the debutanizer process at different sampling rates 2 Value change graph. DETAILED DESCRIPTION

[0062] To make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0063] First, the dual sampling rate problem is introduced as follows:

[0064] In actual industrial processes, due to different sensor measurement principles, the measurement frequencies of auxiliary variables and quality variables differ. Auxiliary variables are typically sampled at a higher frequency, while quality variables, due to their higher measurement costs, are often sampled at a lower frequency. This inconsistent data acquisition is known as the dual sampling rate problem. Inconsistent sampling rates make it difficult to align data in the temporal dimension and result in many missing values, posing significant challenges to soft sensor prediction.

[0065] Figure 1 A diagram illustrating dual sampling rate data is given. Since the auxiliary variable sampling rate is higher, let the auxiliary variable sampling period be T, the time sequence is from left to right in the figure. As can be seen from the figure, all auxiliary variables are sampled every cycle, while the sampling interval of the dominant variable is 2 T , there are two cycles of missing data between each two collections of the dominant variable.

[0066] Example 1:

[0067] This embodiment provides an industrial monitoring method based on a soft sensor model. The method constructs an MGT model and then uses the model to predict the butane concentration in a debutanizer process.

[0068] First, the MGT model is introduced as follows:

[0069] When constructing the MGT model, the convolution operation is first embedded in the attention mechanism to extract data features at multiple scales. Secondly, the gating unit is used to adaptively adjust the multi-scale feature weights. Then, the sparse attention mechanism is used to eliminate redundant features, effectively enhancing the model's feature extraction ability for objects with missing data and accurately predicting the butane concentration in the debutanizer process.

[0070] The modeling process of the MGT soft sensor model mainly includes two stages: dual sampling rate data processing and model training stage and multi-scale gated Transformer prediction stage of time series feature extraction. In the data processing and training stage, the missing data is first filled with neighboring values, and then the model is trained and the trained parameters are saved. In the prediction stage, the multi-scale features of the data are dynamically obtained through the multi-head multi-scale gated attention mechanism, and the extracted features are sparsely operated to eliminate redundant information and extract key features. Finally, the full connection layer is decoded to predict the true value. The MGT network model structure is as follows Figure 2 shown.

[0071] For unlabeled samples with missing quality variables, we first perform a fill operation. We use the most recently collected quality variable data to fill the missing items to reduce the impact of missing data on model training. Then, we combine them with labeled data for training.

[0072] For dual-sampling rate data, using only the current observations for analysis may result in the loss of historical information and time series features, which in turn affects the prediction accuracy of quality variables. To alleviate this problem, a sliding window technique is introduced in the data preprocessing stage to partition the time series data and expand the current input information by introducing historical observations. To enable the model to effectively perceive the sequential information of the input data, position encoding is added before the model input, adding position information to each element in the input sequence. The calculation formula for position encoding is:

[0073] (1)

[0074] (2)

[0075] in, i Indicates the index of the dimension, its value range is i ∈[0, d m / 2], POS Indicates the position where position encoding is performed, d m is the embedding dimension of the data, PE Represents the output after position encoding.

[0076] Transformer can effectively extract the autocorrelation and global features of input data with its internal self-attention mechanism. However, its local feature extraction ability is weaker than CNN. Therefore, to make up for the shortcomings of the self-attention mechanism, the convolution operation of CNN is introduced into the internal structure of the Transformer self-attention mechanism to improve the attention mechanism structure, thereby enhancing the local feature extraction ability of Transformer. The improved model also improves the query vector Q By performing convolution and attention mechanism weight calculation, features can be obtained in parallel, which can enhance the local feature extraction capability while retaining the global feature extraction capability. In addition, a gating structure is constructed between the convolution output and the self-attention score matrix, and the weight of each feature is dynamically adjusted through model training, and finally the features are fused into multi-scale feature output. The design of the gating unit is more conducive to the reasonable distribution of local and global weights of information. Finally, the sparse attention selection mechanism is adopted to eliminate irrelevant and redundant features with low weights, so as to perform soft measurement modeling and prediction. The core of the proposed MGT method is multi-scale gated attention, and its internal structure is as follows: Figure 3 As shown in Figure 2, the calculation formula for multi-scale gated attention is as follows:

[0077] (3)

[0078] (4)

[0079] (5)

[0080] (6)

[0081] (7)

[0082] (8)

[0083] (9)

[0084] In formula (3) The value of is the standard attention score, is the query vector Q and key vector K Dimensions, is a scaling factor used to prevent the gradient from disappearing due to the dot product being too large.

[0085] In formula (4) Express Q Perform convolution operation, and is the weight matrix and bias vector of the gate unit, is the Sigmoid function, represents the gate control unit; It is a weight coefficient learned through model training, usually ranging from 0 to 1, indicating the preference for the current feature.

[0086] In formula (5) is the hidden layer output of feature fusion, Represents the dot product operation between vectors.

[0087] In formula (6) is the attention score after filtering by the TopK operation, which means selecting the query position Before A larger attention score.

[0088] Formula (7) means The feature matrix is ​​updated as Feature sparse matrix, express No. i Row vector.

[0089] Formula (8) is the normalization operation, represents the sparse attention score matrix after normalization operation, Represents a small constant used to avoid the denominator being 0.

[0090] Formula (9) Represents the output of the proposed multi-scale gated attention.

[0091] When processing multi-sampling rate data, feature extraction at different scales is very important. The multi-scale feature fusion method can jointly extract multi-scale information by applying convolution operations and attention mechanisms on features at different sampling rates. The MGT method of this embodiment effectively makes up for the shortcomings of a single mechanism: the attention mechanism performs well in obtaining global data dependencies, but is insensitive to local features; the convolution operation is good at extracting local features, but has difficulty processing long-term dependencies. Combining convolution and attention mechanisms can process local and global features simultaneously, making it suitable for multi-scale modeling.

[0092] In order to verify the prediction effect of the MGT model proposed in this invention in the dual sampling rate problem, Figure 4 The experiment was conducted using the debutanizer industrial process shown as an example.

[0093] The debutanizer is an important equipment in the oil processing process, mainly used for desulfurization and naphtha cracking. Figure 4 As shown in Figure 1, it primarily consists of a condenser, heat exchanger, reflux pump, reboiler, and reflux accumulator. In actual industrial production, butane remaining at the bottom of the debutanizer significantly impacts industrial reactions and should be monitored in real time to minimize its concentration. However, since butane concentration is difficult to measure directly, a soft sensor model is required for modeling and prediction. The easily measurable auxiliary variables selected for this model are shown in Table 1.

[0094] Table 1 Auxiliary variable selection

[0095]

[0096] A total of 2394 sets of samples were collected from the entire debutanizer industrial process, each set of samples contained 7 input variables and 1 output variable; the first 60% of the total samples were selected as the training set for the model in this embodiment, and half of the remaining data were used as the validation set and the other half as the test set. Assume that the auxiliary variable sampling period is T d In order to verify the effectiveness of the method proposed in this embodiment, the corresponding quality variables are sampled at a period of 3 T d , 4 T d , 5 T d and 6 T d Select data. Due to the industrial characteristics of debutanizer, t The quality variables collected at all times cannot be measured immediately. t Only chemical off-line analysis can be obtained at any time t -4 time butane concentration value. Therefore, the mass variable is collected t The moment data consists of:

[0097] (10)

[0098] To verify the effectiveness of the proposed MGT method, an ablation experiment was conducted on the MGT method and compared with the mainstream methods SSLSTM and SSSAE. The parameters of the proposed method are as follows: the number of hidden layer neurons is [128-128-256-1], the number of encoder layers is 2, the number of attention mechanism heads is 4, the time step is set to 20, the number of batches is 512, and the learning rate of the model is calculated by a custom function through the number of iterations. The evaluation indicators of the model prediction are RMSE and R 2 , the calculation formula is as follows:

[0099] (11)

[0100] (12)

[0101] in, For the i The true value of the sample, is the model prediction value, is the mean of the true values, n is the number of samples.

[0102] First, an ablation experiment is conducted, and the experimental results are shown in Table 2.

[0103] Table 2 Prediction results of butane concentration by each network model

[0104]

[0105] It can be seen from the experimental results in Table 2 that the prediction effect of the traditional Transformer method is weaker than that of the sparse Transformer. Sparse Transformer relies on sparse attention, which can only focus on key features and eliminate invalid features, thereby improving the modeling prediction accuracy. The sparse Transformer combined with CNN uses a one-dimensional weighted convolution module to extract local features of the input data, and then uses the sparse Transformer to extract key global features, which can effectively improve the prediction accuracy, and its prediction effect is better than the sparse Transformer. The MGT method of the present invention does not rely on a serial structure, but adopts fused convolution operations and gating units to enhance the local feature extraction capability while adaptively fusing features, effectively improving the prediction accuracy, and is better than the CNN-Transformer method. Under the conditions of the same sampling period, it can be seen from the experimental results in Table 2 that, compared with the SSLSTM and SSSAE methods, the MGT method of the present invention has the best prediction indicators.

[0106] Figure 5A 、 Figure 5B 、 Figure 5C and Figure 5DThe prediction fitting curves of different models under different sampling periods are shown respectively. For a more intuitive comparison, 100 sets of prediction results are magnified and displayed. Figure 5A 、 Figure 5B 、 Figure 5C and Figure 5D It can be seen that as the sampling period increases, the number of missing values ​​in the data increases, and the prediction fitting curve of the model deviates from the true value curve. However, the method proposed in this invention is closer to the true value curve than other methods. Figure 5A 、 Figure 5B 、 Figure 5C and Figure 5D It can be seen that when the sampling period is 3 T d When the sampling period is 1, the prediction result curves of each model are not much different, but with the increase of the sampling period, the prediction result curves of other models except the present invention have large deviations.

[0107] Figure 6A 、 Figure 6B 、 Figure 6C and Figure 6D The scatter plot in FIG2 also shows that, in the four sampling period cases, the prediction results of the model proposed in the present invention are all closer to the 45-degree line.

[0108] Figure 7A and Figure 7B The evaluation indicators RMSE and R of each model are shown respectively. 2 The changes in different sampling rates. Figure 7A and Figure 7B It can be seen that when the sampling period is low, the prediction effects of the six models are the best, and the sparse Transformer method and the CNN-sparse Transformer method have better prediction effects. As the sampling period increases, the prediction effects of the SSLSTM and SSSAE methods gradually surpass the other methods. The RMSE and R 2 The variation range is small, and the performance is optimal under the four sampling periods.

[0109] Example 2:

[0110] This embodiment provides an industrial variable monitoring system based on a soft sensor model, including:

[0111] The data acquisition module is configured to periodically collect easily measurable auxiliary variables and historically difficult-to-measurable leading variables in the industrial process;

[0112] The prediction module is configured to use the MGT model to extract features from the easily measurable auxiliary variables and the historically difficult-to-measurable leading variables, and to perform regression prediction on the industrial variables to be measured (currently difficult-to-measurable leading variables);

[0113] The MGT model divides the input easy-to-measure auxiliary variables into four attention subspaces, extracts local features using the convolutional layer within the multi-scale attention mechanism in each subspace, and uses the attention mechanism to extract global features; secondly, a trainable gating unit is used to adaptively assign weights to fuse multi-scale features; then, a sparse attention mechanism is used to perform a sparse operation on the fused features; finally, a multi-scale gated attention output is obtained, and after fusing the features of each subspace, a fully connected layer decoding is used for regression prediction.

[0114] Some steps in the embodiments of the present invention may be implemented using software, and the corresponding software program may be stored in a readable storage medium, such as a CD or a hard disk.

[0115] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An industrial monitoring method based on a soft measurement model, characterized in that: The method comprises: Step 1: Periodically collect easily measurable auxiliary variables and historically difficult-to-measurable leading variables in the industrial process; Step 2: Use the MGT model to extract features from the easily measurable auxiliary variables and the historically unmeasurable dominant variables, and perform regression prediction on the current unmeasurable dominant variables; The MGT model divides the input easy-to-measure auxiliary variables into four attention subspaces. In each subspace, the convolutional layer within the multi-scale attention mechanism is used to extract local features, and the attention mechanism is used to extract global features. Secondly, a trainable gated unit is used to adaptively assign weights to fuse the multi-scale features. Then, the fused features are sparsified using the sparse attention mechanism. Finally, the multi-scale gated attention output is obtained. After fusing the features of each subspace, a fully connected layer decoding is used for regression prediction. The calculation process of the multi-scale gated attention output is: in, is the fused feature matrix, is the attention score after filtering by the TopK operation. The TopK operation indicates the selection of Before Attention score; Representation matrix No. i Rank j Elements of the column, Represents a sparse matrix of features No. i Rank j Elements of the column, express No. i row vector; represents the sparse attention score matrix after normalization operation, Represents a small constant to avoid the denominator being 0. represents the multi-scale gated attention output; The fused feature matrix Expressed as: in, Represents the global feature, the value is the standard attention score, represents the gate control unit, Represents the dot product operation between vectors, Express Q Perform convolution operation, is the query vector Q and key vector K Dimensions; The gate control unit is represented as: in, and is the weight matrix and bias vector of the gate unit, is the Sigmoid function.

2. The industrial monitoring method based on the soft measurement model according to claim 1, characterized in that: The construction process of the soft sensor model includes: Step 1: Determine the easy-to-measure auxiliary variables and the difficult-to-measure dominant variables, which serve as the input samples and output samples of the soft sensor model respectively. Construct a data set and divide it into a training set, a validation set, and a test set. Step 2: Build the MGT model; Step 3: Initialize the parameters of the MGT model, train the model using the training set, perform forward propagation and calculate the error; calculate the gradient of the loss function with respect to the model parameters using the backpropagation algorithm; and update the model parameters using the Adam optimization algorithm. Step 4: Evaluate the MGT model on the validation set, use grid search to find model parameters and use the validation set to adjust the hyperparameters, and use the trained model for test set prediction; Step 5: Input the test set into the trained model to evaluate the predictive ability.

3. The industrial monitoring method based on the soft measurement model according to claim 1, characterized in that: The method further includes position encoding the input data, which is represented as: in, i represents the index of the dimension, POS Indicates the location where position encoding is performed, d m is the embedding dimension of the data, PE Represents the output after position encoding.

4. The industrial monitoring method based on the soft measurement model according to claim 2, characterized in that: The MSE loss function is used to calculate the error during training.

5. An industrial monitoring system based on a soft measurement model, characterized in that: The system is used to implement the industrial monitoring method based on the soft measurement model according to claim 1, comprising: The data acquisition module is configured to periodically collect easily measurable auxiliary variables and historically difficult-to-measurable leading variables in the industrial process; A prediction module is configured to use the MGT model to extract features from the easily measurable auxiliary variables and the historically unmeasurable leading variables, and to perform regression prediction on the current unmeasurable leading variables; The MGT model divides the input easy-to-measure auxiliary variables into four attention subspaces, extracts local features in each subspace using the convolutional layer within the multi-scale attention mechanism, and uses the attention mechanism to extract global features; secondly, a trainable gating unit is used to adaptively assign weights to fuse multi-scale features; then, a sparse attention mechanism is used to perform a sparse operation on the fused features; finally, a multi-scale gated attention output is obtained, and after fusing the features of each subspace, a fully connected layer decoding is used for regression prediction.

6. An industrial variable monitoring device based on a soft measurement model, characterized in that: including memory and processor; The memory is used to store computer programs; The processor is configured to implement the industrial monitoring method based on the soft measurement model according to any one of claims 1 to 4 when executing the computer program.

7. A computer-readable storage medium, characterized in that The storage medium stores a computer program, and when the computer program is executed by a processor, the industrial monitoring method based on the soft measurement model according to any one of claims 1 to 4 is implemented.

Images