A soft measurement method for variable working condition chemical process based on adaptive migration DeCPLS model

Abstract

The application discloses a variable working condition chemical process soft measurement method based on an adaptive migration DeCPLS model, belongs to the technical field of intelligent modeling of industrial processes, and is used for solving the problem that quality variables are difficult to be accurately measured in real time when working conditions frequently change and new working condition labels are scarce in chemical production. The method comprises the following steps: collecting historical multi-working condition data to construct a source domain and dividing the source domain into sub-source domains; adopting an MMD index to screen the sub-source domains similar to the distribution of a target domain; constructing a multi-constraint feature alignment model containing MMD, maximum variance and manifold regularization, and mapping heterogeneous data to a common feature space; training DeCPLS sub-models for each similar sub-source domain and performing weighted integration. The application suppresses negative migration through similar domain selection, realizes deep feature matching through multi-constraint alignment, improves the generalization ability through weighted integration, and realizes continuous adaptive learning through fine-tuning of the DeCPLS sub-models; and solves the problems of label scarcity, negative migration and working condition drift under variable working conditions, thereby providing a reliable technical scheme for soft measurement of process industries.

Description

Technical Field

[0001] This invention belongs to the field of intelligent monitoring and soft measurement technology of industrial processes, specifically involving a soft measurement method for chemical processes under varying operating conditions based on an adaptive migration DeCPLS model. Background Technology

[0002] Modern chemical production is becoming increasingly complex and large-scale, placing higher demands on the control of product quality indicators. However, due to high hardware costs, harsh measurement environments, and poor sensor reliability, many key quality indicators are difficult to measure online in real time, such as component concentrations in reactors, tray efficiency in distillation columns, and SO2 concentrations in sulfur recovery units. These key variables can usually only be obtained through offline laboratory analysis, resulting in significant measurement lag and severely impacting process optimization and closed-loop quality control.

[0003] In the context of industrial big data, data-driven soft measurement technology has become an important means of online quality estimation. In recent years, artificial intelligence technologies such as deep learning have been widely used in soft measurement modeling of industrial processes. In 2025, researchers proposed a novel Deep Convolutional Partial Least Squares (DeCPLS) model, which combines the nonlinear feature extraction capability of convolutional neural networks (CNNs) with the dimensionality reduction and regression characteristics of partial least squares (PLS), demonstrating significant advantages in high-dimensional spatiotemporal industrial data modeling.

[0004] However, existing technologies still have the following drawbacks;

[0005] 1. Lack of cross-condition transfer capability: Changes in raw material composition, load adjustments, catalyst aging, etc. in industrial production can lead to frequent changes in operating conditions. When new operating conditions occur, there is very little labeled modeling data, while historical multi-condition data is abundant but cannot be effectively utilized. Traditional models cannot achieve knowledge transfer, resulting in a sharp decline in prediction accuracy under new operating conditions.

[0006] 2. There is a risk of negative transfer; if all historical data is simply mixed with the target domain data, unrelated source domain data will introduce negative transfer, which will reduce model performance.

[0007] 3. Lack of online adaptive capability; after a new operating condition model is established, the industrial process database will continuously record new online data. Most existing models are static models, which cannot use this new data to continuously update their own parameters and are difficult to track the time-varying drift of the process.

[0008] 4. Difficulty in aligning heterogeneous data: The distribution of process data under different operating conditions varies significantly. Traditional feature alignment methods either only consider the distribution differences or only consider the data variance, ignoring the local geometric structure, resulting in the loss of feature information after alignment. Summary of the Invention

[0009] To address the aforementioned technical problems in existing technologies, a soft measurement method for chemical processes under varying operating conditions based on an adaptive migration DeCPLS model is provided. This method solves the problems of low prediction accuracy across operating conditions, high negative migration risk, and lack of online adaptive capability in existing technologies.

[0010] To achieve the above objectives, the present invention adopts the following technical solution;

[0011] A soft sensor method for chemical processes under varying operating conditions based on an adaptive migration DeCPLS model includes the following steps;

[0012] Step 1: Historical operating condition data collection and source domain construction; collect historical operating condition data, construct the source domain, and use clustering algorithms to divide the source domain into several sub-source domains;

[0013] Step 2: Measure the distribution difference between the target domain and the sub-source domains; collect process variable data under new operating conditions as the target domain, and use the maximum mean difference index to calculate the distribution difference between the target domain and each sub-source domain;

[0014] Step 3: Similarity domain selection; Based on the distribution difference degree and the transferability threshold, similar sub-source domains are selected from the sub-source domains;

[0015] Step 4: Align the target domain with similar domain features; For all similar sub-source domains and the target domain, construct a multi-constraint feature alignment optimization model including maximum mean difference, maximum variance and manifold regularization, solve the feature transformation matrix, and align the similar sub-source domain data to the common feature space;

[0016] Step 5: DeCPLS sub-model training; For each aligned similar sub-source domain, train a deep convolutional partial least squares sub-model respectively.

[0017] Step 6: Weighted ensemble modeling; Based on the distribution differences between each similar source domain and the target domain, design ensemble weights, and perform weighted ensemble of all deep convolutional partial least squares sub-models to obtain an ensemble model for multi-source domain fusion.

[0018] Step 7: Online adaptive correction model; When a new online data stream arrives, the feature transformation matrix is ​​used to align the features of the new samples, and then the data is input into the ensemble model for quality variable prediction; When the true quality variable labels are obtained, the prediction error is calculated, and the regression parameters of the sub-model corresponding to the most similar sub-source domain are fine-tuned according to the prediction error to achieve online adaptive correction.

[0019] Preferably, in step 1, for chemical processes involving multiple operating conditions, historical data is collected as source domain data, including process variable data. With quality variable data ,use Mean clustering algorithm for process variable data Perform analysis on process variable data. With quality variable data Divided into K sub-source domains .

[0020] Preferably, in step 2, the process variable data of the new operating condition are collected as the target domain data. The target domain data is calculated using the maximum mean difference index. Data from each sub-source domain Distributional variability The distributions between quantization domains are inconsistent, among which ;

[0021] The maximum mean difference index is calculated according to formula (1);

[0022] (1);

[0023] in, These represent the number of samples in the source domain and the target domain, respectively. Representing the sub-source domain datasets respectively With the target domain dataset Data in the middle; Indicates the first The maximum mean difference between the source and target domains is used to quantify the degree of distribution difference between the two domains. Indicates the first The first sub-source domain One sample; Represents the first in the target domain One sample; Represents the Gaussian kernel function; ∥⋅∥ 2 Let represent the square norm of the difference between two mean vectors in Hilbert space, and let represent the distance between the means of two domain distributions.

[0024] Preferably, in step 3, the data from each sub-source domain are... According to the target domain data Sort by difference in ascending order, migration threshold The sub-source domain data is determined by cross-validation of labeled samples in the target domain. With target domain data Difference Less than or equal to the migration threshold When the sub-source domain is determined to be a similar sub-source domain, it is discarded otherwise to avoid introducing negative migration from irrelevant source domains; selection A set of similar sub-source domains ;

[0025] The ranking of differences is shown in formula (2);

[0026] (2);

[0027] in, The transferability threshold is defined as the threshold value when the difference is less than 1%. The source domain is determined to be mobile at that time.

[0028] Preferably, in step 4, all similar sub-source domains are... With target domain data United Construct a multi-constraint feature alignment optimization model including maximum mean difference, maximum variance, and manifold regularization; obtain each similar sub-source domain based on the model optimization results. Aligned data The corresponding quality variable data is , in .

[0029] Preferably, the objective function of the multi-constraint feature alignment optimization model in step 4 is shown in formula (3);

[0030] (3);

[0031] in, This represents the term with the largest variance. Denotes the manifold regularization term. This represents the term with the largest mean difference. The characteristic transformation matrix, The graph Laplacian matrix is ​​used for manifold regularization constraints. This is the matrix of maximum mean difference. This represents the joint data of all similar sub-source domains and the target domain. To balance manifold regularization with the regularization weights that account for the difference between the maximum mean and the regularization weights;

[0032] The eigenvalue transformation matrix is ​​obtained by solving this optimization problem using the generalized eigenvalue decomposition method. And align the similar sub-source domain data according to formula (4);

[0033] (4).

[0034] Preferably, in step 5, the aligned similar sub-source domain data The corresponding quality variable data serves as the input data for the DeCPLS sub-model. As output data, each similar sub-source domain Train one Sub-model ;

[0035] The depthwise convolutional partial least squares sub-model is represented as shown in Equation (5);

[0036] (5);

[0037] in, This is a depthwise convolutional partial least squares model. The model's mapping function, These are the convolution parameters. This is the partial least squares regression coefficient matrix.

[0038] Preferably, in step 6, the sub-model ensemble weights are designed based on the maximum mean difference distance between the aligned similar sub-source domains and the target domain. All training completed Sub-model Weighted integration is performed according to the aforementioned weights to obtain multi-source domain fusion. ensemble model ;

[0039] The integration weight is calculated according to formula (6);

[0040] (6);

[0041] in, It is a very small positive number. The maximum mean difference distance between the similar sub-source domain and the target domain after alignment;

[0042] The weighted integration model is represented as shown in formula (7);

[0043] (7).

[0044] Preferably, in step 7, when the new chemical process online data stream... Upon arrival, the feature transformation matrix obtained in step 4 is first used. Perform feature alignment on the new samples, and then input them into the ensemble model obtained in step 6. Perform quality variable prediction; based on the true quality variable labels obtained from the online data stream. Calculate the prediction error of the model The model established based on the prediction error for the most similar sub-source domain regression parameters Fine-tuning is performed to achieve real-time adaptation to drift in industrial operating conditions.

[0045] Preferably, the prediction error in step 7 is calculated according to formula (8);

[0046] (8);

[0047] in, For a moment New data stream true quality variable labels; For a moment The collected online process variable data samples (i.e., model input vectors); For the ensemble model to input The predicted output values ​​of the quality variables; Let be the prediction error function, representing the predicted value. Compared with the true value The squared error between them;

[0048] The regression parameters are fine-tuned according to the gradient descent formula (9);

[0049] (9);

[0050] in, For the first At time 1, similar sub-source domain sub-models The regression parameters, For the first At time 1, similar sub-source domain sub-models The regression parameters; For learning rate, For prediction error Regression parameters at the previous time step The gradient.

[0051] The beneficial technical effects of this invention are as follows: This invention systematically solves the technical bottlenecks of traditional soft measurement models from the perspectives of data, features, models, and online applications through four core steps: adaptive screening of similar source domains based on maximum mean difference; multi-constraint feature alignment integrating distribution difference measures with maximum variance and manifold regularization; weighted ensemble modeling of DeCPLS sub-models; and online lightweight adaptive correction based on prediction error gradients. Specific beneficial effects are as follows:

[0052] 1. Suppress the risk of negative transfer and make full use of historical knowledge; through adaptive screening of similar source domains, irrelevant sub-source domains with distributions that differ too much from the target domain are eliminated from the source, effectively avoiding the damage of negative transfer to model performance.

[0053] 2. Achieve accurate cross-domain matching of deep features; the multi-constraint feature alignment criteria of maximum mean difference, maximum variance and manifold regularization map heterogeneous data to a common feature space that preserves structure, thereby reducing the distribution differences between domains while retaining the main information and local geometric structure of the target domain.

[0054] 3. Significantly improves the generalization prediction capability under multiple working conditions; by adopting feature alignment and weighted ensemble strategies, the weighted model has stronger generalization performance under various working conditions.

[0055] 4. Lightweight online adaptive, real-time tracking of operating condition drift; the parameter fine-tuning mechanism based on the prediction error gradient only updates the DeCPLS regression coefficients, without retraining the entire deep model, achieving continuous learning on the fly while avoiding the forgetting of historical knowledge, and significantly reducing online computational overhead.

[0056] 5. Experimental verification of superior performance: In the soft measurement task of SO2 concentration in sulfur recovery unit (SRU), the RMSE, MAE and R² indices of the present invention are all superior to the four comparative methods of single source domain DeCPLS, single target domain DeCPLS, mixed domain DeCPLS and migration DeCPLS. The prediction effect diagram shows that the method of the present invention has the best fitting effect in both the stationary and non-stationary stages.

[0057] This invention provides a novel technical approach for cross-condition soft measurement modeling in the process industry, fundamentally overcoming the technical obstacles of low prediction accuracy and weak anti-drift performance of traditional static prediction models when dealing with heterogeneous data. It effectively solves core technical problems such as scarce target domain labels, negative migration risk, and adaptive operation condition drift. Attached Figure Description

[0058] Figure 1 The flowchart of the present invention illustrates the overall process of the method and the interrelationships of the four core modules: similarity domain selection, feature alignment, integrated modeling, and online adaptive correction.

[0059] Figure 2 The image shows the prediction results of a single-source domain DeCPLS model.

[0060] Figure 3 The image shows the prediction results of a single-objective domain DeCPLS model.

[0061] Figure 4 The image shows the prediction results of the hybrid domain DeCPLS model.

[0062] Figure 5 The prediction results of the DeCPLS model are shown in the figure.

[0063] Figure 6 This is a diagram showing the predicted effect of the method of the present invention. Detailed Implementation

[0064] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0065] Example 1: Soft measurement of SO2 concentration in sulfur recovery unit.

[0066] 1. Data acquisition and preprocessing.

[0067] The sulfur recovery unit (SRU) is a crucial component in oil refining processes. Its exhaust gas contains H2S and SO2, requiring timely measurement of SO2 concentration to monitor conversion process performance and control the air feed ratio. Due to the difficulty of online hardware sensor measurement, this invention employs a soft measurement model to establish SO2 concentration.

[0068] Five easily measurable process variables were selected as input features, as shown in Table 1.

[0069] Table 1: Description of Process Variables;

[0070] .

[0071] We collected 1000 fully labeled historical data points as the source domain and divided them into 5 sub-source domains using K-means clustering (K=5). We collected 100 data points under the new operating conditions as the target domain, of which only 25 were labeled. We also collected 300 new data points as the test set.

[0072] 2. Model training.

[0073] Follow steps 1 through 6;

[0074] Calculate the distribution difference between each sub-source domain and the target domain;

[0075] The transferability threshold was determined using cross-validation. Three similar sub-source domains were selected.

[0076] Construct a multi-constraint feature alignment optimization model and set... , Solve for the eigentransformation matrix ;

[0077] For each similar sub-source domain, a DeCPLS sub-model is trained with a kernel size of 3×3, 2 convolutional layers, and 10 partial least squares principal components.

[0078] The integration weights are calculated according to equation (5) to obtain the weighted integration model.

[0079] 3. Online adaptive correction.

[0080] During the testing phase, whenever a new online data stream and its true labels are obtained (obtained once every 10 samples), proceed to step 7: calculate the prediction error and set the learning rate. The partial least squares regression coefficients of the corresponding sub-model of the most similar sub-source domain are updated according to Equation (8).

[0081] 4. Comparative experiment.

[0082] Four comparison methods are set up;

[0083] Single-source domain DeCPLS is trained using only all data from the source domain.

[0084] Single-target domain DeCPLS is trained using only 25 labeled datasets from the target domain.

[0085] Hybrid Domain DeCPLS is trained by simply concatenating labeled data from the source and target domains.

[0086] By migrating DeCPLS, the online adaptive phase is removed from the method of this invention.

[0087] The root mean square error (RMSE), maximum absolute error (MAE), and coefficient of determination (R²) were used as evaluation indicators.

[0088] 5. Experimental results.

[0089] Table 2: Performance of five models in predicting SO2 concentration in the SRU process;

[0090] .

[0091] The results show that;

[0092] Single-objective domain models have a negative R² due to the scarcity of labels, and their prediction performance is even worse than that of mean prediction.

[0093] Single-source domain models struggle to adapt to target domain distribution shifts, limiting their generalization capabilities.

[0094] Although hybrid domain models integrate data from multiple sources, they are still constrained by distribution differences.

[0095] Transfer DeCPLS (without adaptation) significantly improves performance through multi-stage feature optimization;

[0096] Based on this, the method of the present invention further introduces an online adaptive mechanism, which achieves optimal performance on all indicators, with R² reaching 0.842.

[0097] 6. Prediction effect analysis.

[0098] Figures 2 to 6 The prediction curves of five models on the test set are shown. The single-source domain and single-target domain models deviate significantly in the non-stationary phase (e.g., sample 250-300); the mixed domain model shows some improvement but still has a large bias; the transfer DeCPLS model has a better fit; the prediction curve of the method in this invention is the closest to the true value curve, and it can maintain high-precision tracking even when the operating conditions fluctuate drastically.

[0099] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A soft sensing method for chemical processes under varying operating conditions based on an adaptive transfer DeCPLS model, characterized in that, Includes the following steps; Step 1: Historical operating condition data collection and source domain construction; collect historical operating condition data, construct the source domain, and use clustering algorithms to divide the source domain into several sub-source domains; Step 2: Measure the distribution difference between the target domain and the sub-source domains; collect process variable data under new operating conditions as the target domain, and use the maximum mean difference index to calculate the distribution difference between the target domain and each sub-source domain; Step 3: Similarity domain selection; Based on the distribution difference degree and the transferability threshold, similar sub-source domains are selected from the sub-source domains; Step 4: Align the target domain with similar domain features; For all similar sub-source domains and the target domain, construct a multi-constraint feature alignment optimization model including maximum mean difference, maximum variance and manifold regularization, solve the feature transformation matrix, and align the similar sub-source domain data to the common feature space; Step 5: DeCPLS sub-model training; For each aligned similar sub-source domain, train a deep convolutional partial least squares sub-model respectively. Step 6: Weighted ensemble modeling; Based on the distribution differences between each similar source domain and the target domain, design ensemble weights, and perform weighted ensemble of all deep convolutional partial least squares sub-models to obtain an ensemble model for multi-source domain fusion. Step 7: Online adaptive correction model; When a new online data stream arrives, the feature transformation matrix is ​​used to align the features of the new samples, and then the data is input into the ensemble model for quality variable prediction. Once the true quality variable labels are obtained, the prediction error is calculated, and the regression parameters of the sub-model corresponding to the most similar sub-source domain are fine-tuned based on the prediction error to achieve online adaptive correction.

2. The soft sensing method for variable operating conditions chemical processes based on the adaptive migration DeCPLS model according to claim 1, characterized in that, In step 1, for chemical processes involving multiple operating conditions, historical data is collected as source domain data, including process variable data. With quality variable data ,use Mean clustering algorithm for process variable data Perform analysis on process variable data. With quality variable data Divided into K sub-source domains .

3. The soft sensing method for variable operating conditions chemical processes based on the adaptive migration DeCPLS model according to claim 2, characterized in that, In step 2, process variable data under the new operating condition are collected as target domain data. The target domain data is calculated using the maximum mean difference index. Data from each sub-source domain Distributional variability The distributions between quantization domains are inconsistent, among which ; The maximum mean difference index is calculated according to formula (1); （1）； in, These represent the number of samples in the source domain and the target domain, respectively. Representing the sub-source domain datasets respectively With the target domain dataset Data in the middle; Indicates the first The maximum mean difference between the source and target domains is used to quantify the degree of distributional difference between the two domains. Indicates the first The first sub-source domain One sample; Represents the first in the target domain One sample; This represents the Gaussian kernel function.

4. The soft sensing method for variable operating conditions chemical processes based on the adaptive migration DeCPLS model according to claim 3, characterized in that, In step 3, the data from each sub-source domain are... According to the target domain data Sort by difference in ascending order, migration threshold The sub-source domain data is determined by cross-validation of labeled samples in the target domain. With target domain data Difference Less than or equal to the migration threshold When the sub-source domain is determined to be a similar sub-source domain, it is discarded otherwise to avoid introducing negative migration from irrelevant source domains; selection A set of similar sub-source domains ; The ranking of differences is shown in formula (2); （2）； in, The transferability threshold is defined as the threshold value when the difference is less than 1%. The source domain is determined to be mobile at that time.

5. The soft sensing method for variable operating conditions chemical processes based on the adaptive migration DeCPLS model according to claim 4, characterized in that, In step 4, all similar sub-source domains are... With target domain data United Construct a multi-constraint feature alignment optimization model including maximum mean difference, maximum variance, and manifold regularization; obtain each similar sub-source domain based on the model optimization results. Aligned data The corresponding quality variable data is , in .

6. The soft sensor method for variable operating conditions chemical processes based on the adaptive migration DeCPLS model according to claim 5, characterized in that, The objective function of the multi-constraint feature alignment optimization model described in step 4 is shown in formula (3); （3）； in, This represents the term with the largest variance. Denotes the manifold regularization term. This represents the term with the largest mean difference. The characteristic transformation matrix, The graph Laplacian matrix is ​​used for manifold regularization constraints. This is the matrix of maximum mean difference. This represents the joint data of all similar sub-source domains and the target domain. To balance manifold regularization with the regularization weights that account for the difference between the maximum mean and the regularization weights; The eigenvalue transformation matrix is ​​obtained by solving this optimization problem using the generalized eigenvalue decomposition method. And align the similar sub-source domain data according to formula (4); （4）。 7. The soft sensing method for variable operating conditions chemical processes based on the adaptive migration DeCPLS model according to claim 6, characterized in that, In step 5, the aligned similar sub-source domain data The corresponding quality variable data serves as the input data for the DeCPLS sub-model. As output data, each similar sub-source domain Train one Sub-model ; The depthwise convolutional partial least squares sub-model is represented as shown in Equation (5); （5）； in, This is a depthwise convolutional partial least squares model. The model's mapping function, These are the convolution parameters. This is the partial least squares regression coefficient matrix.

8. The soft sensing method for variable operating conditions chemical processes based on the adaptive migration DeCPLS model according to claim 7, characterized in that, In step 6, the sub-model ensemble weights are designed based on the maximum mean difference distance between the aligned similar sub-source domains and the target domain. All training completed Sub-model Weighted integration is performed according to the aforementioned weights to obtain multi-source domain fusion. ensemble model ; The integration weight is calculated according to formula (6); （6）； in, It is a very small positive number. The maximum mean difference distance between the similar sub-source domain and the target domain after alignment; The weighted integration model is represented as shown in formula (7); （7）。 9. The soft sensing method for variable operating conditions chemical processes based on the adaptive migration DeCPLS model according to claim 8, characterized in that, In step 7, when the new chemical process online data stream... Upon arrival, the feature transformation matrix obtained in step 4 is first used. Perform feature alignment on the new samples, and then input them into the ensemble model obtained in step 6. Perform quality variable prediction; based on the true quality variable labels obtained from the online data stream. Calculate the prediction error of the model The model established based on the prediction error for the most similar sub-source domain regression parameters Fine-tuning is performed to achieve real-time adaptation to industrial operating condition drift.

10. The soft sensing method for chemical processes under varying operating conditions based on the adaptive migration DeCPLS model according to claim 1, characterized in that, The prediction error mentioned in step 7 is calculated according to formula (8); （8）； in, For a moment New data stream true quality variable labels; For a moment The collected online process variable data samples are the model input vectors; For the ensemble model to input The predicted output values ​​of the quality variables; Let be the prediction error function, representing the predicted value. Compared with the true value The squared error between them; The regression parameters are fine-tuned according to the gradient descent formula (9); （9）； in, For the first At time 1, similar sub-source domain sub-models The regression parameters, For the first At time 1, similar sub-source domain sub-models The regression parameters; For learning rate, For prediction error Regression parameters at the previous time step The gradient.

Images