Industrial process soft measurement method based on hidden layer target space-time correlation feature regression
By employing the hidden-layer target spatiotemporal correlation feature regression method, and utilizing orthogonal canonical correlation analysis and time-weighted spatial dilatation network, the problem of online measurement of key variables in industrial processes was solved, achieving high-precision quality prediction and production monitoring, and ensuring the stability and efficiency of industrial production.
Patent Information
- Application Number
- CN202411113709.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-14
- Publication Date
- 2024-11-26
AI Technical Summary
Existing technologies struggle to effectively measure key variables in industrial processes online in harsh measurement environments, and traditional models fail to fully uncover the deep quality-related characteristics of the data, resulting in inadequate quality predictions.
A method based on the spatiotemporal correlation feature regression of the hidden layer target is adopted. Orthogonal canonical correlation analysis is used to calculate the correlation between the hidden layer features and the target variable. Potential spatiotemporal features are extracted by gated neuron weighting and time-weighted spatial dilation network, and a deep network is constructed for quality prediction.
It enables real-time monitoring of product quality in industrial processes, improves prediction accuracy and production stability, reduces operational difficulty, and ensures normal production operation.
Smart Images

Figure CN119026471A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to an industrial process soft measurement method, and in particular relates to an industrial process soft measurement method based on latent layer target spatiotemporal correlation feature regression. Background Art
[0002] Process industry has become the foundation and pillar industry of my country's national economy. In order to ensure production safety and improve production efficiency, the monitoring of process performance and the extraction of key variables are crucial to the output of high-quality products. However, due to the harsh measurement environment, large time delays, expensive measurement costs, and large analyzer volume, most key variables are difficult to measure online. To solve these problems, soft measurement technology has become a research hotspot in the field of industrial process control. By selecting a set of measurable variables related to the target variable, a mathematical model with auxiliary variables as input and target variables as output is constructed to estimate those variables that cannot be directly measured.
[0003] Industrial data often has strong nonlinearity, dynamic characteristics and slow time-varying characteristics. There are multi-level and strongly coupled relationships among many parameter variables. Autoencoders have a place in soft measurement modeling due to their excellent data feature extraction and dimensionality reduction capabilities. The nonlinearity and dynamic characteristics of data will affect the measurement results of the experiment. Traditional models rarely consider extracting deep quality-related features in data under nonlinear and time-varying environments at the same time. Summary of the invention
[0004] The purpose of the present invention is to provide an industrial process soft measurement method based on the regression of hidden layer target spatiotemporal correlation features, which uses orthogonal canonical correlation analysis to calculate the correlation between hidden layer features and target variables, gated neurons to weight hidden layer features according to relevant information, and time-weighted spatial expansion networks to focus on spatiotemporal features. This solves the problem of incomplete quality prediction caused by the accumulation of irrelevant information and insufficient mining of information.
[0005] The objective of the present invention is achieved through the following technical solutions:
[0006] A method for soft sensing of industrial processes based on latent target spatiotemporal correlation feature regression, the method comprising the following steps:
[0007] S1: In-depth understanding of industrial production processes and the relationship between process variables and product quality variables;
[0008] S2: Normal samples X and product samples Y obtained based on relevant process data collected by industrial field sensors or industrial system simulation platforms;
[0009] S3: Divide the collected data into training data set and test data set;
[0010] S4: preprocess the training data and test data;
[0011] S5: The training data is passed through the stacked hidden layer target related autoencoder, and the correlation calculation module is used to calculate the correlation between the hidden layer features and the target variables and assign weights to obtain target related information. The time attention mechanism dilated convolution is introduced to build a time-weighted space dilation network to learn the potential time-space related features;
[0012] S6: Train the model by minimizing the loss function and optimize the model parameters;
[0013] S7: Input the test data into the trained model to directly obtain the predicted value of product quality variables, adjust the production process in real time according to the product quality, and ensure the normal operation of industrial production.
[0014] The industrial process soft measurement method based on latent target spatiotemporal correlation feature regression is described, and the overall steps of the prediction method are divided into a training part and a testing part; and the stacked target correlation autoencoder and the time-weighted space expansion network are used to remove irrelevant information in the data to solve the problem of insufficient quality prediction.
[0015] The industrial process soft sensing method based on latent target spatiotemporal correlation feature regression, the training part is:
[0016] (1) Understand the industrial process and preprocess the industrial process data while dividing it into training and test sets;
[0017] (2) Augmenting the preprocessed training data with historical data;
[0018] (3) Construct a latent target spatiotemporal correlation feature regression model, set model parameters, input data into the model to extract deep quality-related features, and train network parameters;
[0019] (4) Minimizing the loss function to obtain the optimal parameters and obtain the soft measurement model; the industrial process soft measurement method based on the regression of the temporal and spatial correlation features of the hidden layer target is described.
[0020] The test parts are:
[0021] (1) The preprocessed test data is augmented and fused with historical data;
[0022] (2) The augmented test data is substituted into the quality prediction model to predict quality variables, obtain the predicted values of product quality variables, and effectively adjust the abnormalities that occur in the production process to ensure the normal operation of industrial production.
[0023] The present invention considers the problem of layer-by-layer accumulation of target irrelevant information and insufficient mining of temporal and spatial relationships between process variables in deep network hierarchical feature learning, and proposes a soft measurement modeling method for regression of hidden target spatiotemporal correlation features. First, a hidden target correlation autoencoder (HTAE) is proposed. The orthogonal canonical correlation analysis method (OCCA) is used to calculate the correlation coefficient between each layer of hidden features and the target variable, and the hidden features weighted by gated neurons are used to replace the original hidden features to improve the dynamic expression ability of target-related features and avoid ignoring high-level features related to the target. Subsequently, multiple HTAEs are stacked to construct a deep network (SHTAE). Secondly, a time-weighted spatial expansion network (TWSDN) is built, and the time attention mechanism and expansion convolution are used to extract spatiotemporal correlation information for the time dimension and the space dimension, respectively, to mine deep spatiotemporal relationships and enhance the model's ability to capture dynamic features. Finally, a prediction regression model is constructed for the deep target spatiotemporal correlation features using a fully connected layer.
[0024] The advantages and effects of the present invention are:
[0025] 1. Based on the autoencoder network, the present invention proposes a hidden target spatiotemporal correlation soft measurement modeling method. The correlation between the hidden features and the target and time and space is considered at the same time. First, a hidden target correlation autoencoder (HTAE) is proposed, and then a deep network is constructed by stacking. In HTAE, the orthogonal canonical correlation information between each hidden feature and the target variable is calculated. The gate value is used to dynamically adjust the information, weight the hidden features and replace the original hidden features, so as to obtain deep features that are fully correlated with the target. After that, a time-weighted space expansion network (TWSDN) is built. After integrating the attention mechanism and dilated convolution to obtain the target spatiotemporal correlation features, a feature regression soft measurement model is established using a fully connected layer.
[0026] 2. The operation status of the industrial production process is monitored in real time, which improves the stability and accuracy of production. The hidden layer target spatiotemporal correlation regression model can make full use of historical time data and current data to predict product quality. Compared with other soft measurement models, this model can effectively solve the problem of layer-by-layer accumulation of irrelevant target information during feature extraction and insufficient mining of the temporal and spatial relationships between hidden layer features, improve the accuracy of quality prediction in the industrial process, and meet the requirements of real-time and accurate monitoring of industrial product quality.
[0027] 3. The predicted data of industrial process product quality is more intuitive and reliable, which reduces the difficulty of operators to supervise the system during operation and improves the monitoring efficiency. When used online, the present invention can directly obtain the predicted value of product quality variables through the soft measurement model, effectively adjust the abnormalities in the production process, ensure the normal operation of industrial production, and avoid the loss of resources and property. BRIEF DESCRIPTION OF THE DRAWINGS
[0028] Figure 1 It is a schematic diagram of the technical route of the present invention;
[0029] Figure 2 It is a prediction result diagram of quality variables in the debutanizer industrial process of the present invention;
[0030] Figure 3 The figure is a prediction result diagram of quality variables in the sulfur recovery industrial process of the present invention;
[0031] Figure 4 A scatter plot of the prediction results in the sulfur recovery industrial process of the present invention;
[0032] Figure 5 It is a schematic diagram of the overall structure of the present invention. DETAILED DESCRIPTION
[0033] The present invention will be described in detail below with reference to the embodiments shown in the accompanying drawings.
[0034] Related term explanation: Orthogonal canonical correlation calculation: a correlation calculation strategy; HTAE: hidden layer target correlation autoencoder; SHTAE: stacked hidden layer target correlation autoencoder; TWSDN: time-weighted spatial dilation network; HTSTC-FR: hidden layer target spatiotemporal correlation feature regression model.
[0035] The present invention designs a hidden layer target related autoencoder and a time weighted space expansion network to solve the problem of insufficient quality prediction caused by the accumulation of irrelevant information layer by layer and the lack of spatiotemporal relationship in the process of hidden layer feature extraction. A soft measurement method for predicting the quality of industrial process products. It includes the following steps:
[0036] S1: Master the theoretical knowledge and operation procedures of industrial processes, and determine relevant process variables and quality variables based on theoretical analysis and experience.
[0037] S2: Collect and organize process data collected by industrial field instruments or process samples X and product samples Y obtained by industrial system simulation platforms;
[0038] S3: Divide the collected data into training data set and test data set;
[0039] S4: preprocess the training data and test data;
[0040] S5: The training data is passed through the stacked hidden layer target related autoencoder, and the correlation calculation module is used to calculate the correlation between the hidden layer features and the target variables and assign weights to obtain target related information. The time attention mechanism dilated convolution is introduced to build a time-weighted space dilation network to learn the potential time-space related features;
[0041] S6: By minimizing the improved loss function and calculating RMSE, MAE and R 2 Three model performance evaluation indicators are used to train the model and optimize the model parameters to improve the model prediction accuracy;
[0042] S7: Input the preprocessed test data into the trained model to directly predict the quality variables, obtain the predicted values of product quality variables, realize real-time monitoring of the production process, and ensure the normal operation of product quality and industrial production.
[0043] As a further improvement, the present invention uses the hidden layer feature target correlation calculation module described in step S5, and the orthogonal canonical correlation analysis method helps calculate the correlation between the hidden layer features and the target variable, and uses these related information to use gated neurons to quantify the contribution of related information to the hidden layer features. The specific processing is as follows:
[0044] S51: Assume that the hidden layer feature H∈R N×l and the target variable Y∈R N×M :
[0045]
[0046] S52: After H and Y are standardized to zero mean and unit variance, their covariance and cross-covariance matrices are estimated as shown in formula (2):
[0047]
[0048] S53: The objective function of a typical target correlation analysis is:
[0049]
[0050] S54: There are two key constraints in the OCCA algorithm, namely orthogonality requirements, which ensure that the feature spaces of H and Y after transformation by J and L are orthogonal and normalized, thus maintaining the independence and stability between variables in multivariate analysis.
[0051] The formula requirements are as follows:
[0052]
[0053] S55: Then define the formula as follows:
[0054]
[0055] S56: Perform SVD decomposition on Y to obtain the following formula:, where Γ and M are both unitary matrices, Γ∈R l×l ; Λ=diag(λ1,λ2,…,λ r )Λ=diag(λ1,λ2,…,λ r),r=rank(S).
[0056]
[0057] S57: Extract the coefficient matrix representing the maximum correlation direction from the covariance matrix of the original data and the matrix obtained by singular value decomposition. The correlation coefficient matrix is as follows:
[0058]
[0059] S58: Take the square root of the singular value to obtain the correlation coefficient, the formula is as follows:
[0060]
[0061] S59: Use the correlation coefficient ρ i To calculate the gating weight of each hidden unit, W gi and b gi are the corresponding weight matrix and bias vector respectively. The formula is as follows:
[0062] g i =σ(W gi ·ρ i +b gi ) (9)
[0063] S60: The hidden layer feature h weighted by the gated neuron t 'Replace the original hidden layer feature h t , and add correlation error to the loss function in the pre-training phase. HTAE, a calculation and replacement method, dynamically adjusts weights to strengthen the correlation between hidden features and target variables, thereby improving the prediction performance of the model.
[0064] S61: The formula for improving the loss function to guide feature learning is as follows:
[0065]
[0066] S62: As attached to the instruction manual Figure 5 The soft sensing method of industrial process based on the regression of hidden layer target spatiotemporal correlation features is shown in FIG. The method realizes real-time monitoring of industrial process and effective prediction of product quality based on orthogonal canonical correlation analysis, gated neurons and time-weighted space expansion network. In order to intuitively evaluate the performance of the model, the present invention uses RMSE, MAE and R 2 Three indicators are used to quantify the prediction effect of the model. Formulas (11), (12) and (13) define RMSE, MAE and R 2 :
[0067]
[0068] Where N is the number of samples, yi is the actual output value of the ith sample, is the predicted output value of the ith sample, and is the average of the actual output values of all samples. RMSE represents the prediction error, R 2 Represents the square correlation between the actual value and the predicted value. Generally, the smaller the RMSE, the higher the R 2 The closer it is to 1, the better the prediction performance of the model.
[0069] The process of predicting product quality variables of the present invention is as follows:
[0070] Training phase:
[0071] (1) Normalization preprocessing.
[0072] (2) Set the time window length to k, and set X = {x1, x2, ..., x m} The time dependency of process data is captured by sliding window to obtain time series data X t ={x t-k+1 ,x t-k+2 ,...,x t}.
[0073] (3) After the first layer encoding process of the HTSTC-FR model, its encoding network parameters {W e1 ,b e1} and the first hidden layer features The first hidden layer feature h1 obtained by the OCCA algorithm is combined with the target variable Y={y1,y2,...,y n} Calculate the correlation and use the gated neurons to weight h1 according to the target correlation contribution weight to obtain the weighted first hidden layer feature By minimizing the pre-training loss function, the encoding network parameters {W e1 ,b e1}, complete the pre-training of the first layer network of the HTSTC-FR model.
[0074] (4) For the second layer of HTAE, the weighted first hidden layer features are used Replace the original first hidden layer features and use h1' as the input of the second layer of the HTSTC-FR model. After the second layer encoding process of the model, the encoding network parameters {W e2 ,b e2} and the second hidden layer features h2 reflects the deeper characteristics of h1 after being weighted and adjusted by target relevance. After training, update the network parameters {W e2 ,b e2}.
[0075] (5) Perform supervised pre-training layer by layer in a similar way, and use the correlation between the hidden layer and the target variable to calculate the weights. The weighted hidden layer features obtained replace the original hidden layer features, making the HTSTC-FR model more sensitive to target-related features. t '={h1',h'2,...,h t '} and all pre-trained weights {W ek ,b ek} k=1,2...L .
[0076] (6) The last layer of weighted hidden features As the input of TWSDN, on the basis of hidden layer target correlation, the time-related information and space-related information are finally extracted from the hidden layer features. By integrating the two-dimensional spatiotemporal information, the HTSTC-FR model is pre-trained layer by layer to extract features.
[0077] (7) By applying the back-propagation algorithm and fine-tuning the network with the minimum error loss, we can obtain improved weights {W e ' k ,b e ' k} k=1,2...L . Save all network parameters for predicting the target variable y.
[0078] Testing phase:
[0079] The preprocessed test samples are input into the input layer of the HTSTC-FR model, and the corresponding target variable prediction value can be obtained through network forward propagation. The root mean square error RMSE, mean absolute error MAE and determination coefficient R are also calculated. 2 Three performance evaluation indicators are used to evaluate the model prediction performance.
[0080] The implementation of the present invention is specifically described below in conjunction with examples:
[0081] [Example 1]
[0082] 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.
[0083] Table 1 Performance evaluation indicators of different models of debutanizer process
[0084] Model RMSE MAE <![CDATA[R 2 ]]> SAE 0.04341 0.02293 0.93476 VW-SAE 0.04275 0.02071 0.93732 STAN 0.03102 0.01948 0.96392 HTSTC-FR 0.02694 0.01845 0.97899
[0085] From Table 1 and the instruction manual Figure 1 It can be seen that the evaluation performance indicators of the HTSTC-FR model are better than those of other models, and its predicted values can track the changes in the true values well. This is because HTSTC-FR plays the role of quality variables, has stronger performance and deep quality feature extraction capabilities, and good nonlinear learning ability, which greatly improves the prediction performance.
[0086] [Example 2]
[0087] The sulfur recovery process is an important link in controlling sulfur emissions, and usually treats environmental pollutants in acidic gas streams that are harmful to the atmosphere and human body. Its input is MEA gas rich in hydrogen sulfide (H2S) and SWS gas rich in H2S and ammonia (NH3). During the sulfur recovery process, H2S is converted into pure sulfur by partial oxidation with air and forms sulfur dioxide (SO2). The tail gas from the process contains residual H2S and SO2, and its concentration needs to be monitored before being released into the atmosphere. The present invention uses a prediction experiment of SO2 concentration in the sulfur recovery process to verify the effectiveness of the proposed method. The five process variables selected are MEA_GAS gas flow, AIR_MEA_1 air flow, AIR_MEA_2 air flow, SWS regional gas flow, and SWS regional air flow. A total of 10075 samples were collected, the first 7600 samples were used as training data, and the remaining 2475 samples were used as test data.
[0088] Table 2 Performance evaluation indicators of different models of industrial polyethylene process
[0089] Prediction methods. RMSE MAE <![CDATA[R 2 ]]> SAE 0.01662 0.01085 0.89077 VW-SAE 0.01443 0.01411 0.89780 STAN 0.01322 0.00813 0.93279 HTSTC-FR 0.01268 0.00554 0.95274
[0090] From Table 2 and the instruction manual Figure 2-3 It can be seen that the prediction results of the HTSTC-FR model track the trend of the true value very well and the performance evaluation index is better than other methods, indicating that the performance of the HTSTC-FR model is the best; the prediction results of SAE, VW-SAE and STAN cannot track the true value changes of the test data set very well, while the prediction results of HTSTC-FR are closest to the true value. In particular, it can be more clearly seen in the local magnification diagram that the test samples of HTSTC-FR perform better, have a high degree of fit with the true value, have a good ability to learn features, attach importance to key related information when extracting features, and greatly improve the prediction accuracy.
[0091] Although the embodiments of the present invention have been disclosed as above, they are not limited to the applications listed in the specification and the embodiments. They can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily implemented. Therefore, without departing from the general concept defined by the claims and the scope of equivalents, the present invention is not limited to the specific details and the illustrations shown and described herein.
Claims
1. A soft sensing method for industrial processes based on latent target spatiotemporal correlation feature regression, 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 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, and preprocess the data; S4: Use orthogonal canonical correlation analysis to calculate the correlation between hidden layer features and target variables, use gated neurons to weight hidden layer features, and build a hidden layer target correlation model to extract deep target correlation features; S5: Add a time-weighted spatial expansion network to extract the spatiotemporal features in the hidden layer features; S6: Train the model by minimizing the loss function 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 based on latent layer target spatiotemporal correlation feature regression as claimed in claim 1, characterized in that: This method is familiar with the operating procedures of industrial processes and the processing of data. The overall steps of the prediction method include training and testing. It uses orthogonal canonical correlation analysis and gated neurons to weight features. It uses time attention mechanism and dilated convolution to extract spatiotemporal features and remove irrelevant information in the data to solve incomplete quality prediction.
3. The industrial process soft sensing method based on latent layer target spatiotemporal correlation feature regression according to claim 1 is characterized in that: The training part is: (1) Understand the industrial process and preprocess the industrial process data while dividing it into training and test sets; (2) Augmenting the preprocessed training data with historical data; (3) Construct a latent target spatiotemporal correlation feature regression model, set model parameters, input data into the model to extract deep quality-related features, and train network parameters; (4) Minimize the loss function to obtain the optimal parameters and obtain the soft sensor model.
4. The industrial process soft sensing method based on latent layer target spatiotemporal correlation feature regression according to claim 3 is characterized in that: The test parts are: (1) The preprocessed test data is augmented and fused with historical data; (2) The augmented test data is substituted into the quality prediction model to predict quality variables, obtain the predicted values of product quality variables, and effectively adjust the abnormalities that occur in the production process to ensure the normal operation of industrial production.
Citation Information
Cited By
Quality monitoring method and system based on related component regression
CN120632812A
