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

Abstract

The invention relates to an industrial monitoring method, system and device based on a soft measurement model, and a storage medium, and belongs to the technical field of soft measurement. According to the method, the quality variables in the industrial process are monitored by constructing the soft measurement model based on the multi-scale gating Transform, and effective fusion of global and local features is realized by designing a hierarchical feature fusion structure. Firstly, convolution operation is embedded into an attention mechanism to extract multi-scale features of data; secondly, the multi-scale feature weight is adjusted in a self-adaptive mode through a gating unit, low-correlation or redundant features are removed through a sparse attention mechanism, and therefore the feature extraction capacity of the model under the data missing condition is enhanced; the method is applied to butane concentration prediction in the debutanizer industrial process, and the result shows that compared with an existing soft measurement industrial monitoring method, the method has higher prediction precision. The invention further provides an industrial monitoring system and device based on the soft measurement model and a storage medium, and the convenience of industrial monitoring is improved.

Description

Technical Field

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

[0002] There are many difficult-to-measure variables in complex industrial processes. These variables usually directly affect the product quality and production efficiency of industrial production and are crucial for industrial processes. Therefore, it is necessary to monitor the important parameters of industrial processes in real time. However, the methods of using traditional instruments for measurement and off-line chemical analysis are time-consuming and laborious, and there is a large delay in the measurement results. To better monitor industrial variables in real time, a soft sensor method with easily measurable variables (i.e., auxiliary variables) as inputs and difficult-to-measure variables (i.e., dominant variables) as outputs has been proposed.

[0003] Soft sensor modeling mainly includes four steps: selecting auxiliary variables, data acquisition and preprocessing, establishing a soft sensor model, and online correction. Among them, establishing a soft sensor model is the core of the soft sensor modeling method. Soft sensor models are mainly divided into two categories: mechanism models and data-driven models. The soft sensor method based on mechanism modeling requires an analysis of the underlying principles of chemical processes. However, as the complexity of industrial processes increases, the research on mechanisms becomes more and more complex, and the method of comprehensively analyzing industrial processes relying on rich experience is difficult to meet the actual industrial needs. Compared with mechanism modeling, the data-driven soft sensor modeling method only relies on process data, effectively improving the modeling efficiency, and thus has received extensive attention. There are many data-driven soft sensor methods, mainly including: statistical learning methods, machine learning methods, and deep learning methods.

[0004] Deep learning, with its automatic feature extraction ability of multi-layer non-linear structures, shows significant advantages in processing complex industrial data in the field of soft sensors. Typical deep learning methods include Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), Long Short-Term Memory (LSTM) networks, and Transformer networks, etc. As a current popular technology, Transformer, with its self-attention mechanism, can process sequence data in parallel, breaking the limitation of the loop structures of traditional RNN and LSTM networks for processing long sequences, effectively obtaining the global dependence relationship of data, and showing higher computational efficiency and performance advantages. It has become a research hotspot and mainstream direction in the current field of deep learning soft sensors.

[0005] Regarding the problem of soft sensor modeling for the nonlinearity, dynamics, and temporal correlation of actual industrial process data, deep learning methods have become a better solution. However, in actual industrial processes, sensor-collected data is periodic, making it difficult to obtain all variables simultaneously. This sampling method with periodic sampling characteristics leads to data missing and lag, posing a huge challenge to soft sensor prediction. This application mainly focuses on the data missing problem caused by inconsistent sampling periods between auxiliary variables and quality variables, namely the dual sampling rate problem. For soft sensor data with different sampling rates, Chang et al. proposed an improved RNN network, which divides and maps data with different sampling rates to a common hidden space and updates the hidden space at each moment, so as to integrate and unify the information for predicting 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 multi-time-channel CNN, which uses a time feature sharing network and a parallel prediction network for variable prediction, and makes the blank values in the data not participate in the backpropagation by improving the backpropagation algorithm (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 CNN and RNN have deficiencies in dealing with data with long-term dependence relationships. CNN is limited by the size of the convolutional kernel, and RNN may have the problem of gradient disappearance, which will further affect 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 the soft sensor model and the convenience of industrial monitoring, the present invention provides an industrial monitoring method, system, device, and storage medium based on the soft sensor model. The technical solutions are as follows: The first object of the present invention is to provide an industrial monitoring method based on a soft sensor model, and the method includes: Step 1: Periodically collect easily measurable auxiliary variables and historical difficult-to-measure dominant variables in the industrial process; Step 2: Use the MGT model to extract features from the easy-to-measure auxiliary variables and historical difficult-to-measure dominant variables, and perform regression prediction on the current difficult-to-measure dominant variables; The MGT model divides the input easy-to-measure auxiliary variables into 4 attention subspaces, extracts local features using the convolutional layer inside the multi-scale attention mechanism in each subspace, and uses the attention mechanism to extract global features; secondly, uses a trainable gating unit to adaptively allocate weights to fuse multi-scale features; then, uses a sparse attention mechanism to sparsify the fused features; finally, obtains the multi-scale gated attention output. After fusing the features of each subspace, a fully connected layer is used for decoding for regression prediction.

[0007] Optionally, the calculation process of the multi-scale gated attention output is as follows:

[0008]

[0009]

[0010]

[0011] Among them, is the fused feature matrix, is the attention score after being screened by the TopK operation, and the TopK operation means selecting the first attention scores; represents the element in the matrix at the i row and j column, represents the element in the feature sparse matrix at the i row and j column, represents the i row vector of; represents the sparse attention score matrix after normalization operation, represents a very small constant used to avoid the denominator being 0, represents the multi-scale gated attention output.

[0012] Optionally, the fused feature matrix is expressed as:

[0013]

[0014] Among them, Represents the global feature, and the value is the standard attention score. Represents the gating unit. Represents the dot product operation between vectors. Represents Q Performs a convolution operation on Is the query vector Q And the key vector K Of the dimension.

[0015] Optionally, the gating unit is represented as:

[0016] Where And Are the weight matrix and bias vector of the gating unit, Is the Sigmoid function.

[0017] Optionally, the construction process of the soft measurement model includes: Step 1: Determine the easily measurable auxiliary variable and the difficult-to-measure dominant variable, and use them as the input sample and output sample of the soft measurement model respectively. Construct a data set and divide it into a training set, a validation set, and a test set; Step 2: Construct the MGT model; Step 3: Initialize the parameters of the MGT model, train the model through the training set, perform forward propagation and calculate the error; calculate the gradient of the loss function with respect to the model parameters through the backpropagation algorithm; update the model parameters through the Adam optimization algorithm; Step 4: Evaluate the MGT model on the validation set, use the grid search method to find the model parameters and adjust the hyperparameters using the validation set, and use the trained model for test set prediction; Step 5: Input the test set into the trained model and evaluate the prediction ability.

[0018] Optionally, the method further includes performing positional encoding on the input data, which is represented as:

[0019]

[0020] Where i Represents the index of the dimension, pos Represents the position for performing positional encoding, d m Is the embedding dimension of the data, PE Represents the output after positional encoding.

[0021] Optionally, the MSE loss function is used to calculate the error during the training process.

[0022] The second object of the present invention is to provide an industrial monitoring system based on a soft measurement model, the system comprising: A data acquisition module configured to periodically acquire easily measurable auxiliary variables and historical difficult-to-measure dominant variables in an industrial process; A prediction module configured to perform feature extraction on the easily measurable auxiliary variables and historical difficult-to-measure dominant variables using an MGT model, and perform regression prediction on the current difficult-to-measure dominant variables; The MGT model divides the input easily measurable auxiliary variables into 4 attention subspaces, extracts local features using the convolutional layer inside the multi-scale attention mechanism in each subspace, and extracts global features using the attention mechanism; secondly, a trainable gating unit is used to adaptively allocate weights to fuse the multi-scale features; then, a sparse attention mechanism is used to sparsify the fused features; finally, a multi-scale gated attention output is obtained. After fusing the features of each subspace, a fully connected layer is used for decoding and regression prediction.

[0023] The third object of the present invention is to provide an industrial monitoring device based on a soft measurement model, comprising a memory and a processor; The memory is used to store a computer program; The processor is used to implement the industrial monitoring method based on the soft measurement model as described in any one of the above when executing the computer program.

[0024] The 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 is implemented.

[0025] The beneficial effects of the present invention are as follows: The present invention monitors the quality variables in an industrial process 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, a convolutional operation is embedded in the attention mechanism to extract multi-scale features of the data; secondly, a gating unit is used to adaptively adjust the weights of the multi-scale features, and low-correlation or redundant features are removed through a sparse attention mechanism, thereby enhancing the feature extraction ability of the model in the case of data loss. Applying this method to industrial processes such as debutanizer towers and sulfur recovery can effectively improve the accuracy of monitoring.

[0026] Experimental results show that compared with other existing soft measurement model industrial monitoring methods, the present invention exhibits higher prediction accuracy and better generalization ability.

[0027] The present invention provides an industrial monitoring system and device based on a soft-sensing model and a computer-readable storage medium, improving the convenience of industrial monitoring. Description of the Drawings

[0028] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0029] Figure 1 It is a diagram for describing double-sampling-rate data.

[0030] Figure 2 It is the structure diagram of the proposed MGT model.

[0031] Figure 3 It is the structure diagram of multi-scale gated attention.

[0032] Figure 4 It is the process flow diagram of the debutanizer tower.

[0033] Figure 5A It is when the sampling period is 3 T d The butane concentration prediction curve diagrams of each model.

[0034] Figure 5B It is when the sampling period is 4 T d The butane concentration prediction curve diagrams of each model.

[0035] Figure 5C It is when the sampling period is 5 T d The butane concentration prediction curve diagrams of each model.

[0036] Figure 5D It is when the sampling period is 6 T d The butane concentration prediction curve diagrams of each model.

[0037] Figure 6A It is when the sampling period is 3 T d The scatter diagrams of the prediction results of each network model.

[0038] Figure 6B It is when the sampling period is 4 T d The scatter diagrams of the prediction results of each network model.

[0039] Figure 6C It is when the sampling period is 5 Td Scatter plot of the prediction results of each network model at this time.

[0040] Figure 6D The sampling period is 6 T d Scatter plot of the prediction results of each network model at this time.

[0041] Figure 7A It is a graph showing the change of the RMSE value of the prediction results of each network model in the debutanizer process at different sampling rates.

[0042] Figure 7B It is the prediction result R of each network model in the debutanizer process at different sampling rates 2 Value change graph. Specific implementation manners

[0043] To make the objectives, technical solutions and advantages of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0044] First, the double sampling rate problem is introduced as follows: In actual industrial processes, due to different measurement principles of sensors, there are differences in the measurement frequencies of auxiliary variables and quality variables. Auxiliary variables are usually sampled at a higher frequency, while quality variables are sampled at a lower frequency due to higher measurement costs. This problem of inconsistent data acquisition is the double sampling rate problem. Inconsistent sampling rates can make it difficult to align data in the time dimension and there are many missing values, posing great challenges to soft measurement prediction.

[0045] Figure 1 A diagram illustrating double sampling rate data is given. Since the sampling rate of the auxiliary variable is higher, let the sampling period of the auxiliary variable be T , and the left-to-right in the figure represents the time sequence. As can be seen from the figure, all auxiliary variables are sampled every period, while the acquisition interval of the dominant variable is 2 T , and there are two periods of data missing between every two acquisitions of the dominant variable.

[0046] Example 1: This embodiment provides an industrial monitoring method based on a soft measurement model. This method constructs an MGT model and then uses this model to achieve the prediction of butane concentration in the debutanizer process.

[0047] First, the MGT model is introduced as follows: When constructing the MGT model, first embed the convolution operation into the attention mechanism to extract data features at multiple scales. Secondly, use the gating unit to adaptively adjust the weights of the multi-scale features. Then, utilize the sparse attention mechanism to eliminate redundant features, effectively enhancing the model's ability to extract features of data missing objects and accurately predicting the butane concentration in the de-butane tower process.

[0048] The modeling process of the MGT soft sensor model mainly includes two stages: the dual-sampling rate data processing and model training stage, and the multi-scale gated Transformer prediction stage for temporal feature extraction. In the data processing and training stage, first fill in the missing data with neighboring values, and then train the model and save the trained parameters. In the prediction stage, dynamically obtain the multi-scale features of the data through the multi-head multi-scale gated attention mechanism, perform sparsification operations on the extracted features to eliminate redundant information and extract key features, and finally perform full connection layer decoding to predict the true value. The MGT network model structure is as Figure 2 shown.

[0049] For unlabeled samples with missing quality variables, first perform a filling operation. Fill in the missing items with the quality variable data collected at the nearest time to reduce the impact of data missing on model training. Secondly, train in combination with labeled data.

[0050] For dual-sampling rate data, only analyzing the observed values at the current moment may lead to the loss of historical information and temporal features, thereby affecting the prediction accuracy of quality variables. To alleviate this problem, introduce the sliding window technique in the data preprocessing stage to divide the time series data and expand the input information at the current moment by introducing historical observations. To enable the model to effectively perceive the sequential information of the input data, add positional encoding before the model input to add positional information to each element in the input sequence. The calculation formula for positional encoding is: (1) (2) where i represents the index of the dimension, and its value range is i ∈[0, d m / 2], pos represents the position for positional encoding, d m is the embedding dimension of the data, PE represents the output after positional encoding.

[0051] The Transformer can effectively extract the self-correlation and global features of the input data by virtue of its internal self-attention mechanism. However, its local feature extraction ability is weaker than that of the CNN. Therefore, to make up for the deficiency of the self-attention mechanism, the convolutional operation of the 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 the Transformer. The improved model simultaneously performs convolution and attention mechanism weight calculation on the query vector Q and can obtain features in parallel, retaining the global feature extraction ability while enhancing the local feature extraction ability. In addition, a gating structure is constructed between the convolutional output and the self-attention score matrix, and the weight distribution of each feature is dynamically adjusted through model training. Finally, feature fusion is performed to output multi-scale features. The design of the gating unit is more conducive to the reasonable distribution of local and global weights of information. Finally, a selection mechanism of sparse attention 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 Figure 3 shown. The calculation formula of multi-scale gated attention is as follows: (3) (4) (5) (6) (7) (8) (9) The value of in formula (3) is the standard attention score, is the query vector Q and the dimension of the key vector K , is the scaling factor, which is used to prevent the dot product from being too large and causing the gradient to disappear.

[0052] In formula (4), represents the convolution operation on Q , and are the weight matrix and bias vector of the gating unit, is the Sigmoid function, represents the gating unit; is a weight coefficient learned through model training, usually ranging from 0 to 1, indicating the preference degree for the current feature.

[0053] In formula (5), is the output of the hidden layer for feature fusion, represents the dot product operation between vectors.

[0054] In formula (6), is the attention score after being screened by the TopK operation, and the TopK operation means selecting the top larger attention scores.

[0055] Formula (7) represents updating the feature matrix to a feature sparse matrix, where represents the i row vector of

[0056] Formula (8) is a normalization operation, where represents the sparse attention score matrix after the normalization operation, and

[0057] in formula (9), represents the output of the proposed multi-scale gated attention.

[0058] 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 convolutional operations and attention mechanisms to features at different sampling rates. The MGT method in this embodiment effectively makes up for the deficiencies of a single mechanism: the attention mechanism performs well in obtaining the global dependency relationship of data but is not sensitive to local features; convolutional operations are good at local feature extraction but difficult to handle long-term dependency relationships. Combining convolution and the attention mechanism can handle both local and global features simultaneously and is suitable for multi-scale modeling.

[0059] To verify the prediction effect of the proposed MGT model in the dual-sampling rate problem, experiments are conducted using the industrial process of the debutanizer shown in Figure 4 as an example.

[0060] The debutanizer is an important device in the petroleum processing process, mainly used for desulfurization and naphtha cracking. As shown in Figure 4 , it mainly includes a condenser, a heat exchanger, a reflux pump, a reboiler, and a reflux accumulator. In the actual industrial production process, the butane remaining at the bottom of the debutanizer will greatly affect the industrial reaction, and its content should be detected in real time to minimize it. However, since the concentration of butane is difficult to directly measure, a soft sensor model needs to be established for modeling and prediction, and the easily measurable auxiliary variables selected for the model are shown in Table 1.

[0061] Table 1 Selection of auxiliary variables

[0062] A total of 2,394 groups of samples were collected for the entire debutanizer industrial process. Each group of samples contains 7 input variables and 1 output variable. The first 60% of the total samples were selected as the training set of the model in this embodiment, and half of the remaining data was used as the validation set and half as the test set. Let the sampling period of the auxiliary variable be T d , to verify the effective performance of the method proposed in this embodiment, the corresponding quality variables were sampled at a sampling period of 3 T d , 4 T d , 5 T d and 6 T d to select data. Due to the industrial characteristics of the debutanizer, t the quality variables collected at the t moment cannot be measured immediately. t At the t moment, only the butane concentration value at (10) -4 obtained by off-line chemical analysis can be obtained. Therefore, the data set at the 2 moment when the quality variables are collected is composed of: (11) (12) where, is the true value of the i th sample, is the predicted value of the model, is the average value of the true values, n is the number of samples.

[0063] First, ablation experiments were carried out, and the experimental results are shown in Table 2.

[0064] Table 2 Prediction results of butane concentration by each network model

[0065] 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.

[0066] Figure 5A 、 Figure 5B 、 Figure 5C and Figure 5D The 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.

[0067] 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.

[0068] 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 7BIt can be seen that when the sampling period is relatively low, the prediction effects of all 6 models are optimal, and the prediction effects of the sparse Transformer method and the CNN-sparse Transformer method are relatively good. As the sampling period increases, the prediction effects of the SSLSTM and SSSAE methods gradually surpass those of other methods. The RMSE and R 2 of the method proposed in the present invention both have relatively small change ranges and are optimal under four sampling period conditions.

[0069] Embodiment 2: This embodiment provides an industrial variable monitoring system based on a soft measurement model, including: A data acquisition module configured to periodically acquire easily measurable auxiliary variables and historical difficult-to-measure dominant variables in an industrial process; A prediction module configured to use the MGT model to extract features from the easily measurable auxiliary variables and historical difficult-to-measure dominant variables, and perform regression prediction on the to-be-measured industrial variable (current difficult-to-measure dominant variable); The MGT model divides the input easily measurable auxiliary variables into 4 attention subspaces, extracts local features using the convolutional layer inside the multi-scale attention mechanism in each subspace, and extracts global features using the attention mechanism; secondly, an adaptable gating unit is used to adaptively allocate weights to fuse the multi-scale features; then, a sparse attention mechanism is used to perform a sparsification operation on the fused features; finally, a multi-scale gated attention output is obtained. After fusing the features of each subspace, a fully connected layer is used for decoding for regression prediction.

[0070] Some steps in the embodiments of the present invention can be implemented by software, and the corresponding software program can be stored in a readable storage medium, such as an optical disc or a hard disk, etc.

[0071] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. An industrial monitoring method based on a soft sensor model, characterized in that The method includes: Step 1: Periodically collect easily measurable auxiliary variables and historical difficult-to-measure dominant variables in the industrial process; Step 2: Use the MGT model to extract features from the easily measurable auxiliary variables and historical difficult-to-measure dominant variables, and perform regression prediction on the current difficult-to-measure dominant variables; The MGT model divides the input easily measurable auxiliary variables into 4 attention subspaces, extracts local features using the convolutional layer inside the multi-scale attention mechanism in each subspace, and extracts global features using the attention mechanism; Secondly, a trainable gating unit is used to adaptively allocate weights to fuse the multi-scale features; Then, a sparse attention mechanism is used to sparsify the fused features; Finally, a multi-scale gated attention output is obtained. After fusing the features of each subspace, a fully connected layer is used for decoding and regression prediction.

2. The industrial monitoring method based on the soft sensor model according to claim 1, wherein 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.

3. The industrial monitoring method based on the soft sensor model according to claim 2, wherein The fused feature matrix is expressed as: Among them, represents the global feature, and the value is the standard attention score, represents the gating unit, represents the dot product operation between vectors, represents the operation on Q for convolution operation, is the query vector Q and the key vector K is the dimension of.

4. The industrial monitoring method based on a soft sensor model according to claim 3, characterized in that, The gating unit is expressed as: Among them, and are the weight matrix and bias vector of the gating unit, is the Sigmoid function.

5. The industrial monitoring method based on a soft sensor model according to claim 1, characterized in that, The construction process of the soft sensor model includes: Step 1: Determine the easily measurable auxiliary variables and difficult-to-measure dominant variables, which are used 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: Construct the MGT model; Step 3: Initialize the parameters of the MGT model, train the model through the training set, perform forward propagation and calculate the error; Calculate the gradient of the loss function with respect to the model parameters through the backpropagation algorithm; Update the model parameters through the Adam optimization algorithm; Step 4: Evaluate the MGT model on the validation set, use the grid search method to find the model parameters and adjust the hyperparameters using the validation set, and use the trained model for test set prediction; Step 5: Input the test set into the trained model to evaluate the prediction ability.

6. The industrial monitoring method based on a soft sensor model according to claim 1, characterized in that The method further includes performing positional encoding on the input data, expressed as: Among them, i represents the index of the dimension, pos represents the position for performing positional encoding, d m is the embedding dimension of the data, PE represents the output after positional encoding.

7. The industrial monitoring method based on a soft sensor model according to claim 5, characterized in that, During the training process, the MSE loss function is used to calculate the error.

8. An industrial monitoring system based on a soft sensor model, characterized in that, The system includes: A data acquisition module configured to periodically collect easily measurable auxiliary variables and historical difficult-to-measure dominant variables in the industrial process; A prediction module configured to use the MGT model to extract features from the easily measurable auxiliary variables and historical difficult-to-measure dominant variables, and perform regression prediction on the current difficult-to-measure dominant variables; The MGT model divides the input easily measurable auxiliary variables into 4 attention subspaces, extracts local features using the convolutional layer inside the multi-scale attention mechanism in each subspace, and extracts global features using the attention mechanism; Secondly, a trainable gating unit is used to adaptively allocate weights to fuse the multi-scale features; Then, a sparse attention mechanism is used to sparsify the fused features; Finally, a multi-scale gated attention output is obtained. After fusing the features of each subspace, a fully connected layer is used for decoding and regression prediction.

9. An industrial variable monitoring device based on a soft sensor model, characterized in that, It includes a memory and a processor; The memory is used to store computer programs; The processor is used to, when executing the computer program, implement the industrial monitoring method based on the soft sensor model according to any one of claims 1 to 7.

10. 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 according to any one of claims 1 to 7 is implemented.

Images