Industrial process soft measurement method for dynamically scaling long short-term memory network
By applying dynamic scaling long and short-term memory networks and graph attention networks in industrial processes, combining error factors and threshold mechanisms, the problem of difficulty in monitoring quality variables in industrial processes is solved, and high accuracy and stability quality prediction is achieved.
Patent Information
- Application Number
- CN202510007221.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-03
- Publication Date
- 2025-05-13
AI Technical Summary
The complexity of modern industrial processes increases, and traditional hard equipment is difficult to effectively monitor key quality variables, resulting in inaccurate quality prediction and poor robustness.
Dynamic scaling long short-term memory network (LSTM) and graph attention network (GAT) are introduced, and the LSTM hidden state is updated through error factors, temporal features are extracted, and spatial features are extracted using GAT. At the same time, the validity of unlabeled data is judged by the threshold calculation of labeled data, and a pseudo-label data update model is generated.
Accurate online prediction of key quality variables in industrial processes is achieved, product quality is improved, prediction stability and accuracy are enhanced, and real-time monitoring and control needs are met.
Smart Images
Figure CN119987326A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to an industrial process soft measurement method, in particular to an industrial process soft measurement method of a dynamic scaling long short-term memory network. Background Art
[0002] Faced with large-scale and highly complex modern industrial processes, monitoring, evaluating and optimizing the production process are of great significance for ensuring safe production and improving product quality. In the past, the scale of the industry was relatively simple, and the key variables of the industrial process could be directly obtained through online hard equipment or offline analytical instruments. However, with the development of the economy and the advancement of science and technology, the complexity of industrial production processes has continued to increase, and the harsh measurement environment conditions and imperfect measurement instruments have made the monitoring of quality variables extremely difficult. The traditional strategy of using online hard equipment is slowly being abandoned and replaced by the popular soft measurement technology. Soft measurement technology is widely used to predict key quality variables that are difficult to obtain from easy-to-measure process variables. This technology first collects a set of process variables related to the quality variables to be predicted, and then establishes a mathematical model using these process variables as input and quality variables as output, allowing quality variables that cannot be directly measured to be predicted.
[0003] Industrial data often have strong nonlinearity, dynamic characteristics and redundant information. Using these data to directly establish a regression model for quality variables may lead to inaccurate model prediction and poor robustness. Therefore, it is necessary to represent the input data and extract useful information for soft measurement modeling. Long short-term memory networks are widely used in the field of soft measurement due to their excellent ability to extract time features. Traditional long short-term memory networks do not consider the relationship between the current input variables and the historical state, and do not consider the spatiality of process variables. Therefore, a scaling factor is introduced into the hidden state of LSTM to extract time features, and GAT extracts spatial features. At the same time, the quality variable calculation threshold obtained by labeled data is used to judge the validity of unlabeled data, and pseudo-labeled data is generated from the valid unlabeled data to update the model. Summary of the invention
[0004] The purpose of the present invention is to provide an industrial process soft measurement method for dynamically scaling long short-term memory networks. The present invention introduces a scaling factor into the long short-term memory network and uses the GAT network to construct a dynamic spatiotemporal module to extract features. A threshold mechanism is established to calculate the threshold using the quality variable obtained from the labeled data to judge the validity of the unlabeled data, and the valid unlabeled data is used to generate pseudo-labeled data to update the model. Accurate online prediction of product quality is achieved to improve product quality.
[0005] The purpose of the present invention is achieved through the following technical solutions:
[0006] A method for industrial process soft sensing using a dynamically scaled long short-term memory network, the method comprising the following steps:
[0007] S1: Master the industrial process theory and basic operation procedures, and understand the relationship between process variables and quality variables;
[0008] S2: Collect and organize process data collected by industrial field instruments or samples obtained by industrial system simulation platforms;
[0009] S3: Divide the data into training set and test set according to 8:2;
[0010] S4: preprocessing the divided data sets;
[0011] S5: Introduce a scaling factor to update the hidden state in LSTM to extract temporal features, and use the GAT network to extract spatial features. Finally, calculate the likelihood and threshold of the evaluation index calculated based on the labeled data, and compare the likelihood and threshold calculated under unlabeled data to determine its effectiveness. Generate pseudo-label data from valid unlabeled data to update the model;
[0012] S6: Minimize the pre-training loss to train the model and fine-tune the model parameters;
[0013] S7: Calculate RMSE, MAE and R 2 The evaluation indicators are used to input the test data into the trained model for prediction to obtain the predicted values of quality variables and improve product quality.
[0014] The industrial process soft measurement method of a dynamically scaled long short-term memory network takes into account the spatiotemporal correlation of process variables, introduces error factors into the hidden state of LSTM, and combines the GAT network. Considering the validity of the data, the validity of the unlabeled data is judged by the threshold determined under the labeled data, and the valid unlabeled data is used to generate pseudo labels to update the model and expand the data.
[0015] The industrial process soft sensing method of a dynamic scaling long short-term memory network, the training process is:
[0016] (1) Preprocess the data;
[0017] (2) Divide the training set and test set;
[0018] (3) Divide the training set into labeled data and unlabeled data;
[0019] (4) First, input the labeled data, build the hybrid model, add the error scaling factor to the hidden state of LSTM to extract the temporal features, and use GAT to extract the spatial features;
[0020] (5) Calculating the likelihood value and threshold value based on the obtained predicted value of the quality variable;
[0021] (6) Calculate the likelihood value of the unlabeled data according to the above steps, and generate pseudo-labeled data from the unlabeled data that meets the threshold range to update the model;
[0022] (7) Minimize the pre-training loss to obtain the optimal parameters and obtain the soft measurement model;
[0023] The industrial process soft measurement method of a dynamic scaling long short-term memory network, the test process is:
[0024] (1) Preprocess the test data;
[0025] (2) The processed data is input into the trained model to obtain the predicted values of the key quality variables of industrial products and calculate the evaluation indicators RMSE, MAE and R 2 To evaluate the prediction effect of the model, finally draw conclusions and output the model.
[0026] The advantages and effects of the present invention are:
[0027] 1. Based on the long short-term memory network, the present invention proposes an industrial process soft measurement method of a semi-supervised dynamic scaling long short-term memory network (SDSTS-LSTM). First, the error factor is introduced into the hidden state of the LSTM, and it is updated to extract the time characteristics of the process variables. At the same time, considering the spatiality between process variables, the GAT network is combined to extract spatial features. Considering the validity of the data, the likelihood value and threshold are calculated by using the evaluation index representative of the quality variable predicted by labeled data. The validity of the current unlabeled data is judged by comparing the likelihood value and the threshold under the unlabeled data. And the effective unlabeled data is generated into a pseudo-label to update the model. Real-time prediction of product quality is achieved while improving the stability of the prediction.
[0028] 2. Semi-supervised dynamic scaling long short-term memory network deeply extracts spatiotemporal related features and ensures the validity of data. Compared with other models, semi-supervised dynamic scaling long short-term memory network has better prediction performance under nonlinearity and dynamics.
[0029] 3. The present invention meets the needs of real-time monitoring of actual industrial production operation and control of industrial product quality, and improves the stability and accuracy of online measurement. Real-time monitoring of process status to adjust abnormal conditions in the production process reduces production costs and ensures production safety and efficiency.
[0030] 4. In terms of quality prediction, the present invention can accurately predict the key quality variables of the actual industrial production process; in terms of online detection, it can monitor the operating status of the production process in real time, and realize the evaluation, control and detection of the process equipment; in terms of product output, it plays an important role in the output of high-quality products, and is of great significance in improving production efficiency and ensuring production safety, and ultimately achieving the goal of green production. BRIEF DESCRIPTION OF THE DRAWINGS
[0031] Figure 1 It is a schematic diagram of the technical route of the present invention;
[0032] Figure 2 is a prediction result diagram of quality variables of the debutanizer process of the present invention;
[0033] Figure 3 is a prediction error result diagram of quality variables of the debutanizer process of the present invention;
[0034] Figure 4 is a prediction result diagram of quality variables of the sulfur recovery process of the present invention;
[0035] Figure 5 It is a schematic diagram of the overall structure of the present invention. DETAILED DESCRIPTION
[0036] The present invention will be described in detail below with reference to the embodiments shown in the accompanying drawings.
[0037] Glossary:
[0038] LSTM: Long Short-Term Memory Network;
[0039] GAT: Graph Attention Network;
[0040] SDSTS-LSTM: Semi-supervised Dynamically Scaling Long Short-Term Memory Network;
[0041] The present invention proposes a semi-supervised dynamic scaling long short-term memory network (SDSTS-LSTM) industrial process soft sensing method, the process of which is as follows:
[0042] S1: Master the industrial process theory and basic operation procedures, and understand the relationship between process variables and quality variables;
[0043] S2: Collect and organize process data collected by industrial field instruments or samples obtained by industrial system simulation platforms;
[0044] S3: Divide the data into training set and test set according to 8:2;
[0045] S4: preprocessing the divided data sets;
[0046] S5: Introduce a scaling factor to update the hidden state in LSTM to extract temporal features, and use the GAT network to extract spatial features. Finally, calculate the likelihood and threshold of the evaluation index calculated based on the labeled data, and compare the likelihood and threshold calculated under unlabeled data to determine its effectiveness. Generate pseudo-label data from valid unlabeled data to update the model;
[0047] S6: Minimize the pre-training loss to train the model and fine-tune the model parameters;
[0048] S7: Calculate RMSE, MAE and R 2 The evaluation indicators are used to input the test data into the trained model for prediction to obtain the predicted values of quality variables and improve product quality.
[0049] As a further improvement, the present invention uses the error factor described in step S5 to update the hidden state in LSTM to extract time features, and combines GAT to extract spatial features. Finally, the validity of the unlabeled data is judged by the threshold determined under the labeled data, and the valid unlabeled data is used to generate pseudo labels to update the model and expand the labeled data. The specific steps are as follows:
[0050] S51: Introduce an error factor to update the hidden state in LSTM to reduce the redundancy of current input variables and hidden features;
[0051] S52: The formula for introducing the error factor is as follows:
[0052]
[0053] f t =layer_norm(error) (2)
[0054] f t =layer_norm(error) (3)
[0055]
[0056] S53: The formula for extracting spatial features from the GAT network is as follows:
[0057]
[0058] S54: The threshold formula for setting labeled data is as follows:
[0059] PA i =R i +M i (7)
[0060] Among them, R irepresents RMSE, M i MAE
[0061]
[0062] S55: The formula for calculating the likelihood value of unlabeled data is as follows:
[0063]
[0064] If lh i 'Greater than or equal to lh t , then the unlabeled sequence data is retained. After all unlabeled data training is completed, the retained unlabeled data is used to generate pseudo-labeled data, and the model is retrained with the labeled data as input; if lh i 'Small than lh t , it means that the unlabeled sequence data is abnormal data and will be discarded. Figure 5 A semi-supervised dynamic scaling long short-term memory network industrial process soft measurement method is shown. The method realizes real-time monitoring of industrial processes and effective prediction of product quality based on a semi-supervised dynamic scaling long short-term memory network and a GAT network. The process of predicting quality variables of the present invention is as follows:
[0065] Training phase:
[0066] 1. Normalize and preprocess the collected data and divide it into training set and test set.
[0067] 2. First, the labeled data sequence X in the training set is passed through a sliding window s ={x s1 ,x s2 ,…,x sl} as input, and perform dynamic spatiotemporal feature extraction to obtain the hidden feature h s1 , so that the pre-training error is minimized.
[0068] 3. Obtain the current prediction value through the fully connected layer, and calculate the RMSE and MAE of the model at the current moment. Calculate the prediction accuracy and its likelihood value.
[0069] 4. Extract the dynamic spatiotemporal features of the labeled data sequence in turn to obtain the hidden features h s2 ,h s3 ,…,h sl , calculate the prediction accuracy and its likelihood. Calculate the threshold value under supervised learning.
[0070] 5. Use the sliding window to convert the unlabeled data sequence X in the training set u ={x u1 ,x u2 ,…,xuL} as input, dynamic feature extraction is performed to obtain hidden features h u1 ,h u2 ,…,h uL , and calculate its prediction accuracy and likelihood value.
[0071] 6. Compare the likelihood value under unlabeled data with the threshold value determined under labeled data. If the threshold condition is met, the current sequence data and quality variables are retained and pseudo-label data X is generated. ss ; Otherwise, discard the current input data sequence.
[0072] 7. Retrain the model using all the retained pseudo-labeled data and labeled data as input;
[0073] 8. Fine-tune the overall parameters by minimizing the prediction error loss, preserving the network structure and parameters.
[0074] Testing phase:
[0075] The same method is used for data preprocessing in the testing phase and the training phase, and the predicted value is obtained using the trained SDSTS-LSTM model. The root mean square error (RMSE), mean absolute error (MAE) and coefficient of determination (R 2 ) evaluation index to verify the prediction effect of the model. As shown in equations (6), (7) and (8):
[0076]
[0077]
[0078] Where N is the number of test samples; is the i-th predicted value, y i is the i-th true value. is the average value of the output value in the test data set. The closer RMSE and MAE are to 0, the higher the R 2 The closer to 1, the better the effect.
[0079] Example 1
[0080] The debutanizer is an important part of the oil refining process and is used for desulfurization and naphtha cracking. In the industrial process, in order to improve the control quality of the debutanizer, it is necessary to estimate the butane content in real time. The present invention collects 7 process variables as model inputs, including tower top temperature, tower top pressure, tower top reflux flow, tower top effluent flow, tower temperature, tower bottom temperature 036 and tower bottom temperature 037. 2390 samples are selected, of which the first 1600 samples are used as training sets to train model parameters, and the remaining samples are used as test sets to evaluate the model prediction ability.
[0081] Table 1 Performance evaluation indicators of different models of debutanizer process
[0082]
[0083] From Table 1 and the instruction manual Figure 2 And the instruction manual Figure 3 It can be seen that the evaluation performance indicators of the SDSTS-LSTM model are better than those of other models, and its predicted values can track the changes of the true values well. This is because SDSTS-LSTM ensures the integrity of feature information extraction, as well as good nonlinear and dynamic learning capabilities, which greatly improves the prediction performance.
[0084] Example 2
[0085] Sulfur recovery unit is an important chemical production process, mainly used to treat the acidic gas stream generated in the production process and remove it before discharging it into the atmosphere. The two acidic gases MEA and SWS from the previous stage undergo oxidation reaction in the reactor to generate elemental sulfur. The resulting tail gas steam contains hydrogen sulfide and sulfur dioxide gas residues. The reaction process is controlled by real-time detection of sulfur dioxide concentration in the tail gas.
[0086] The present invention collects five conventional measurements as input, selects 10071 samples, of which the first 8000 samples are used as training sets and the rest are used as test sets.
[0087] Table 2 Performance evaluation indicators of different models of sulfur recovery process
[0088]
[0089] Table 2 and the instruction manual Figure 4 The results show that the trained SDSTS-LSTM model has the smallest RMSE and MAE values compared with the other four models. 2 The value is the largest, and the predicted result matches the actual curve better in terms of value and trend, and the predicted result is closest to the true value. This is mainly because the present invention can extract the spatiotemporal characteristics of process variables under nonlinearity and dynamics, while ensuring the validity of input data. Therefore, the SDSTS-LSTM soft sensor model has better online prediction ability for sulfur recovery than the traditional model, and is a soft sensor model more suitable for the sulfur recovery process.
[0090] The specific embodiments of the present invention have been disclosed above. The present invention is applicable to various fields mentioned in the specification and other suitable fields. Without departing from the general concept defined by the claims and their equivalent scope, the present invention is not limited to the description illustrations and specific related details given in the specification.
Claims
1. A method for industrial process soft sensing using a dynamically scaled long short-term memory network, characterized in that: The method steps are as follows: S1: Familiar with the specific operation process of actual industrial production and understand the relationship between process variables and quality variables; S2: Acquire samples based on on-site collection of process data from industrial devices or industrial system simulation platforms; S3: Divide the data into training set and test set according to the ratio of 8:2, and preprocess the data to divide the training set into labeled data and unlabeled data according to the ratio of 4:6; S4: Extract spatiotemporal features of labeled data using error-scaled long short-term memory network and GAT, and calculate the likelihood value and threshold based on the obtained quality variable prediction value; S5: Calculate the likelihood value of the unlabeled data according to the above steps, and generate pseudo-labeled data from the unlabeled data that meets the threshold range to update the model; S6: Minimize the pre-training loss to train the model and optimize the model parameters; S7: Input the test data into the trained soft measurement model to obtain the predicted values of the key quality variables of the product, monitor the production process in real time, and adjust the operation plan to ensure normal production operation.
2. The industrial process soft sensing method of a dynamically scaled long short-term memory network according to claim 1 is characterized in that Be familiar with the operation procedures and data processing of industrial processes.
3. The industrial process soft sensing method of a dynamic scaling long short-term memory network according to claim 3 is characterized in that Construct a hybrid model to extract spatiotemporal features using a dynamically scaled spatiotemporal long short-term memory network; The validity of the current unlabeled data is judged by the threshold determined under the labeled data; the valid unlabeled data is used to generate pseudo labels to update the model.
4. The industrial process soft sensing method of a dynamic scaling long short-term memory network according to claim 1 is characterized in that: The training part is: (1) Preprocess the data; (2) Divide the training set and test set; (3) Divide the training set into labeled data and unlabeled data (4) First, input the labeled data, build the hybrid model, add the error scaling factor to the hidden state of LSTM to extract the temporal features, and use GAT to extract the spatial features; (5) Calculating the likelihood value and threshold value based on the obtained predicted value of the quality variable; (6) Calculate the likelihood value of the unlabeled data according to the above steps, and generate pseudo-labeled data from the unlabeled data that meets the threshold range to update the model; (7) Minimize the pre-training loss to obtain the optimal parameters and obtain the soft measurement model.
5. The industrial process soft sensing method of a dynamic scaling long short-term memory network according to claim 1 is characterized in that: The test parts are: (1) Preprocess the test data; (2) The processed data is input into the trained model to obtain the predicted values of the key quality variables of industrial products and calculate the evaluation indicators RMSE, MAE and R 2 To evaluate the prediction effect of the model, finally draw conclusions and output the model.