Multi-working-condition industrial process soft measurement method based on multi-task learning and probability modeling

Abstract

The invention discloses a multi-working-condition industrial process soft measurement method based on multi-task learning and probability modeling, and aims to solve the problem of insufficient measurement precision caused by heterogeneous mixing of multi-working-condition process samples. The method comprises three core modules, namely a feature decoupling coding module, a hierarchical feature fusion module and a probability information aggregation module. Firstly, a spatial-temporal feature extractor is designed to explicitly decouple multi-working-condition data into working condition shared features and specific features, and hybrid feature expression and working condition recognition are achieved. Then, a hierarchical feature fusion module is constructed, deep fusion of information between working conditions is realized through a hierarchical expert gating network, and a complex interaction relationship between the working conditions is modeled; and finally, proposing a probability information aggregation strategy, inputting the fusion features into corresponding predictors, and weighting prediction results by using the working condition identification probability to generate final prediction output. According to the method, a classification task and a regression task are incorporated into a unified multi-task learning framework, and the good performance of multi-working-condition process performance index soft measurement is ensured.

Description

Technical Field

[0001] This invention belongs to the field of soft measurement modeling in process industries, and specifically relates to a multi-condition industrial process soft measurement method based on multi-task learning and probabilistic modeling, which solves the problem of insufficient measurement accuracy caused by heterogeneous mixing of multi-condition process samples. Background Technology

[0002] In modern process industries, accurately sensing key quality variables is crucial for ensuring stable production, improving energy efficiency, and achieving operational safety. Soft measurement technology, by utilizing easily measurable process variables to predict quality variables, effectively overcomes the inherent delays and high costs of direct measurement methods. With the development of industrial digitalization and intelligence, data-driven soft measurement has become the mainstream paradigm for achieving real-time quality prediction. In recent years, deep learning-based methods have been extensively studied, such as autoencoders (AEs), long short-term memory networks (LSTMs), and convolutional neural networks (CNNs).

[0003] Most existing deep learning methods assume that training and testing data originate from the same operating condition and satisfy the condition of independence and identical distribution. However, this assumption is often inconsistent with actual industrial processes involving multiple operating conditions caused by parameter variations and raw material fluctuations. This leads to models that perform well on training data failing to maintain performance on testing data. Taking blast furnace ironmaking as an example, differences in iron ore grades from different batches and changes in operating parameters set by operators can cause significant changes in operating conditions. In this case, the data distribution exhibits heterogeneity across multiple operating conditions. Traditional models lack the ability to accurately characterize multiple patterns, resulting in decreased prediction accuracy. Current solutions for quality prediction in multi-operating-condition processes can be broadly categorized into three types: transfer learning methods, single global methods, and hybrid ensemble methods. Specifically, the core of transfer learning methods is learning transferable knowledge to adapt from source conditions to target conditions. However, in transfer learning, each operating condition is treated as an independent domain and projected onto a unified shared feature space, leading to the loss of discriminative information between operating conditions. Single global methods train directly on the entire multi-operating-condition dataset to learn multiple patterns holistically. While single global methods preserve some variation between load conditions, they lack explicit load condition identification and tend to learn the average effects of load conditions, making it difficult to capture complex interactions. In contrast, hybrid ensemble methods first identify load conditions and then build separate sub-models for each load condition, thus providing fine-grained predictions.

[0004] While the aforementioned methods can address the soft measurement needs of multi-condition processes, they still face challenges in multi-condition representation and information aggregation. Due to the complexity of production processes, multi-condition data contains two types of characteristics: nonlinear dynamics within each condition and intricate shared and specific information between different conditions. If these mixed characteristics are not explicitly decoupled during feature extraction, the model will inevitably be disturbed, leading to a decline in predictive performance. Furthermore, the dependence of each prediction target in multiple outputs on different conditions varies. Therefore, the final prediction should selectively integrate information from each condition while maintaining industrial interpretability, necessitating the exploration of new solutions to overcome the shortcomings of traditional deep learning-based soft measurement. Summary of the Invention

[0005] In view of the above background, this invention proposes a multi-condition industrial process soft measurement method based on multi-task learning and probabilistic modeling, which can be used to realize real-time high-precision soft measurement in multi-condition scenarios.

[0006] The technical solution for achieving the technical objective of this invention is as follows: A multi-condition soft measurement method for industrial processes based on multi-task learning and probabilistic modeling includes the following steps: (1) Historical data are read from the DCS database of the industrial process, divided into multiple subsets according to the operating condition label, and a multi-operating condition process variable time series dataset is constructed; a preset time window is applied to each sample. T Perform serialization processing to form input samples The corresponding quality variable is used as a label. y i The operating condition label is z i ∈{1, ..., M}; (2) The input samples are processed by two parallel spatiotemporal feature extractors. x i Explicit decoupling is a feature of shared operating conditions. h c , i and special characteristics of working conditions h s , i The shared features are used to model common dynamic patterns across working conditions, while the specific features are used to distinguish different working conditions. (3) Construct a hierarchical working condition perception feature fusion module, which includes L Layered expert gating network; each layer contains a shared expert group and M Each of the special expert groups consists of [number] members. N eThe system consists of an expert network; through a gated network, shared features and unique features are dynamically weighted and fused, and each working condition is output iteratively layer by layer. m Corresponding fusion features ; (4) For each working condition m Build an independent regression predictor R m Receive corresponding fusion features And output the predicted value At the same time, utilizing the unique characteristics of the operating conditions h s , i The classifier φ outputs the recognition probability of each working condition. Finally, a comprehensive prediction result is generated through probability-weighted aggregation: , The method jointly optimizes the following three tasks during training: Prediction task: Minimize the mean squared error loss in the prediction of quality variables. L p ; Classification task: Minimize the cross-entropy loss for work condition identification L c ; Shared Feature Alignment Task: Minimize the Maximum Mean Difference (MMD) loss among different operating conditions in the shared feature space. L m An end-to-end collaborative optimization of feature decoupling, operating condition identification, and soft measurement is achieved through a multi-task learning framework.

[0007] The preset time window T The length is determined by analyzing the time autocorrelation and cross-correlation of process variables to ensure coverage of the main dynamic response cycle of the system and to satisfy the following conditions. T ≥ max( τ i ),in τ i For the first i The dominant time constants of key variables.

[0008] The spatiotemporal feature extractor is a TSMixer structure, which consists of a cascaded temporal mixing network and a feature mixing network. The temporal hybrid network performs fully connected cross-projection on different time steps in each feature dimension to capture long-range temporal dependencies. The feature mixing network performs channel-level fusion of all process variables within the same time step, enhancing cross-variable interaction; Stable training is achieved between the two networks through residual connections and layer normalization.

[0009] In the hierarchical working condition perception feature fusion module, the first l layer, l ∈ {1,...,L}, working condition m The fusion calculation is as follows: , ; in, These are the normalized weight coefficients output by the gating network. Indicates the first n e An expert's input combining shared and unique features. This is the output of the previous layer; The first layer input is h c , i and h s , i Final layer output As a predictor R m Input.

[0010] The classifier φ is a single-layer or multi-layer fully connected network, and its output is M A probability distribution vector is normalized using the softmax function; The probability of identifying the working condition satisfy And during training, it is related to classification loss. L c Joint optimization.

[0011] The overall loss function of the multi-task learning framework is: L = w p L p + w c L c + w m L m in, w p , w c , w mThe weight coefficients are learnable or preset, and satisfy the following conditions: w p + w c + w m = 1; The MMD loss L m Defined as the average maximum mean difference among all operating condition pairs: .

[0012] The method is deployed in an industrial process real-time database system. The process variable data collected online is subjected to soft measurement in steps (1) to (4), and the prediction results are fed back into the DCS database to form a "sensing-prediction-optimization" closed-loop control loop.

[0013] The industrial process is blast furnace ironmaking, chemical reactor, or cement rotary kiln, and the quality variables include at least one of silicon content, sulfur content, temperature, or component concentration.

[0014] The method described above, when the number of working conditions M > 10. The working condition clustering method is used to merge similar working conditions into subsets and share the same dedicated regression model for each subset in order to reduce model size and inference latency.

[0015] The specific regression model R m It is a three-layer fully connected network with hidden layer widths of 128, 64, and 32 respectively. The output layer is a linear layer, and the Adam optimizer is used during training with an initial learning rate of 1e. -3 .

[0016] The beneficial effects of this invention are: 1. The designed spatiotemporal feature extractor and dual constraint mechanism realize explicit decoupling of multi-condition data, while retaining shared and specific features of the conditions. This captures the common patterns across conditions while preserving the specific information within each condition.

[0017] 2. Hierarchical expert gating networks achieve fine-grained feature fusion through dynamic weight adjustment, effectively modeling complex interaction relationships between working conditions and improving the model's ability to utilize information related to working conditions.

[0018] 3. A probabilistic information aggregation strategy is proposed, which uses the probability of working condition identification to quantify classification uncertainty. Through weighted integration, the impact of working condition identification error on prediction results is reduced, and the robustness of the model is improved.

[0019] 4. By adopting a multi-task learning framework to collaboratively optimize classification and regression tasks, end-to-end joint modeling of feature decoupling, working condition identification and quality prediction is achieved, which significantly improves the accuracy of soft measurement in multi-working condition scenarios. Attached Figure Description

[0020] Figure 1 The flowchart shown is for the soft measurement method.

[0021] Figure 2 The diagram shows the network structure of the spatiotemporal feature extractor.

[0022] Figure 3 The diagram shows the network structure of the hybrid expert model.

[0023] Figure 4 The image shows a comparison between the method's prediction results and the actual data. Detailed Implementation

[0024] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0025] The purpose of this invention is to provide a multi-condition soft measurement method for industrial processes based on multi-task learning and probabilistic modeling, as shown in the flowchart below. Figure 1 As shown, the specific steps are as follows: (1): Historical data were read from the DCS database of the industrial process, divided into multiple subsets according to the operating condition label, and a multi-operating condition process variable time series dataset was constructed; a preset time window was applied to each sample. T Perform serialization processing to form input samples The corresponding quality variable is used as a label. y i The operating condition label is z i ∈{1, ..., M}; (2): The input samples are processed by two parallel spatiotemporal feature extractors. x i Explicit decoupling is a feature of shared operating conditions. h c , i and special characteristics of working conditions h s , i The shared features are used to model common dynamic patterns across working conditions, while the specific features are used to distinguish different working conditions. (3): Construct a hierarchical working condition perception feature fusion module, which includes L Layered expert gating network; each layer contains a shared expert group and M Each of the special expert groups consists of [number] members. Ne The system consists of an expert network; through a gated network, shared features and unique features are dynamically weighted and fused, and each working condition is output iteratively layer by layer. m Corresponding fusion features ; (4): For each working condition m Build an independent regression predictor R m Receive corresponding fusion features And output the predicted value At the same time, utilizing the unique characteristics of the operating conditions h s , i The classifier φ outputs the recognition probability of each working condition. Finally, a comprehensive prediction result is generated through probability-weighted aggregation. The method jointly optimizes the following three tasks during training: Prediction task: minimizing the mean squared error loss of the quality variable prediction. L p Classification task: Minimize the cross-entropy loss for work condition identification. L c Shared Feature Alignment Task: Minimize the maximum mean difference (MMD) loss among different operating conditions in the shared feature space. L m The end-to-end collaborative optimization of feature decoupling, working condition identification, and soft measurement is achieved through a multi-task learning framework.

[0026] In step (1), the preset time window T The length is determined by analyzing the time autocorrelation and cross-correlation of process variables to ensure coverage of the main dynamic response cycle of the system and to satisfy the following conditions. T ≥ max( τ i ),in τ i For the first i The dominant time constants of key variables.

[0027] In step (2), the spatiotemporal feature extractor is a TSMixer structure, consisting of a cascaded temporal mixing network and a feature mixing network. The temporal mixing network performs fully connected cross-projection on different time steps at each feature dimension to capture long-range temporal dependencies. The feature mixing network performs channel-level fusion on all process variables within the same time step to enhance cross-variable interactions. Stable training is achieved between the two networks through residual connections and layer normalization. Its network structure is as follows: Figure 2 As shown.

[0028] In multi-condition industrial processes, differences and correlations coexist between operating conditions. For example, differences in operating settings and raw material batches can alter the distribution of process variables, resulting in condition-specificity. Simultaneously, different operating conditions, based on the same equipment and internal mechanisms, generate shared common characteristics. Traditional methods struggle to learn complex mixed features and fail to capture subtle differences between patterns. To overcome this limitation, the input multi-condition data needs to be explicitly decomposed into shared and specific features. Shared features aim to capture the inherent information of each operating condition, while specific features strive to retain the discriminative information for condition identification. The decomposed features also lay the foundation for subsequent hierarchical fusion of condition-aware features. Shared Features h c , i and special characteristics of working conditions h s , i The features are obtained through two parallel spatiotemporal feature extractors, and their forward calculation formula is as follows: , in, FET c and FET s TSMixer represents the extraction of shared features and specific features of operating conditions, respectively. θ c and θ s These represent the parameters that the network needs to learn. During the backpropagation process of network training, the shared features of the operating conditions are minimized by minimizing the maximum mean difference (MMD) loss. L m Extraction is performed to reduce the distribution differences of data under different operating conditions within the shared feature space. L m The calculation formula is: , Its meaning is operating condition. m With working conditions n The MMD distance between features. Condition-specific features are minimized by the cross-entropy classification loss. L c Extracting data enhances the differentiation of operating condition categories, enabling operating condition identification. L c The calculation formula is: , in, For the sample x i The actual working conditions to which it belongs For the sample x iThe features are then analyzed by a classifier to predict which operating condition it belongs to. m The probability of this is high. Both condition sharing and specific feature branching are crucial for quality prediction; relying on only one branch will result in the loss of information from the other. To fully utilize this complementary information and form more comprehensive pattern-aware features, it is necessary to effectively fuse shared and specific features.

[0029] In step (3), the hierarchical working condition perception feature fusion module is composed of... L The expert gating network consists of layers, and the structure of the expert gating network is as follows: Figure 3 As shown, it employs a layered architecture, progressively fusing feature streams to achieve fine-grained integration of shared and specific information. For each layer of the network... l ∈ {1,...,L}, which contains 1 shared expert group and M A special expert group Each expert group contains N e Each expert network processes local correlation information of shared features and specific features related to the operating conditions, respectively. The structure of the expert network is identical to that of the spatiotemporal feature extractor network. The output features of the shared expert group are... For working conditions m The output characteristics of the special expert group are Each expert group corresponds to a gated network. Weighting coefficients are generated based on the current input features and used to weight and fuse the output results of the expert group, thereby enabling the interaction between shared and specific features.

[0030] For working conditions m , its in l The formula for calculating the working condition-specific fusion characteristics of a layer is:

[0031] , in, These are the output weight coefficients of the gating network. These are the specific features after fusion. Since fusion is performed after the first layer, the input to the first layer is the decoupled feature. h c , i and h s , i In the hierarchical fusion process, the fusion features output by the previous layer are used as the input to the next layer of the expert gating network. After a preset number of fusion layers, the output fusion feature vector corresponding to the number of working conditions is used as the input to the predictor.

[0032] In step (4), the predictor R m Representatives address working conditions m The independently constructed regression model takes as input the corresponding working condition fusion features output from step (3). The output is the predicted value of the quality variable under this working condition. The specific regression model R m This is a three-layer fully connected network with hidden layer widths of 128, 64, and 32 respectively. The output layer is a linear layer, and the Adam optimizer is used during training with an initial learning rate of 1e-3. According to the law of total probability, the quality variable... y i The probability distribution is as follows: , The above formula means joint distribution It can be decomposed into and The product of Θ, where Θ is the set of model parameters, can be integrated along the z-axis into Θ. M The finite sum of each working condition. Among them... yes x i This is a working condition. m The likelihood is obtained through classification probability; For working conditions m The prediction results under the given conditions. Therefore, the final prediction result can be calculated as a weighted sum of the working condition classification probability and the working condition prediction. The probability weighted aggregation will combine the working condition specific features in step (2). Operating condition identification probability output by classifier φ As weights, the predicted values ​​output by each predictor are weighted and summed to obtain the final soft measurement result. The calculation formula is: .

[0033] During the backpropagation process of network training, the loss function of the predictor is the mean squared error (MSE) loss. L p The calculation is as follows: .

[0034] The overall model contains multiple optimization objectives, and the total loss is: L = w p L p + w c L c + wm L m in, w p , w c , w m These are the weighting coefficients for prediction loss, classification loss, and maximum mean difference loss, respectively.

[0035] This method was applied to predict the silicon content of molten iron in blast furnace ironmaking. Sixteen process variables were selected as inputs, including oxygen enrichment rate, permeability index, oxygen enrichment flow rate, cold blast flow rate, furnace gas volume, furnace gas index, blast kinetic energy, top pressure, total pressure difference, hot blast pressure, actual wind speed, hot blast temperature, furnace top temperature, drag coefficient, blast temperature, and actual pulverized coal injection rate. The dataset contains three operating conditions, and the prediction results are as follows: Figure 4 As shown, the root mean square error (RMSE) of the model's prediction of silicon content in molten iron is 7.061, the mean absolute error (MAE) is 5.442, and the coefficient of determination is... R 2 The result reached 0.724, indicating that the model has good predictive performance and good fit. After the model training is completed, the optimized network parameters are solidified and stored, and deployed to the industrial process real-time database system. The newly collected process variable data are processed according to the process of steps (1)-(4), and the real-time quality variable soft measurement results are output to provide data support for online monitoring and optimization control of industrial processes.

[0036] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the scope of protection of the present invention. The scope of protection of the present invention is defined by the appended claims and any equivalent technical solutions.

Claims

1. A multi-condition soft measurement method for industrial processes based on multi-task learning and probabilistic modeling, characterized in that, Includes the following steps: (1) Historical data are read from the DCS database of the industrial process, divided into multiple subsets according to the operating condition label, and a multi-operating condition process variable time series dataset is constructed; a preset time window is applied to each sample. T Perform serialization processing to form input samples The corresponding quality variable is used as a label. y i The operating condition label is z i ∈{1, ..., M}; (2) The input samples are processed by two parallel spatiotemporal feature extractors. x i Explicit decoupling is a feature of shared operating conditions. h c , i and special characteristics of working conditions h s , i The shared features are used to model common dynamic patterns across working conditions, while the specific features are used to distinguish different working conditions. (3) Construct a hierarchical working condition perception feature fusion module, which includes L Layered expert gating network; each layer contains a shared expert group and M Each of the special expert groups consists of [number] members. N e The system consists of an expert network; through a gated network, shared features and unique features are dynamically weighted and fused, and each working condition is output iteratively layer by layer. m Corresponding fusion features ; (4) For each working condition m Build an independent regression predictor R m Receive corresponding fusion features And output the predicted value At the same time, utilizing the unique characteristics of the operating conditions h s , i The classifier φ outputs the recognition probability of each working condition. Finally, a comprehensive prediction result is generated through probability-weighted aggregation: ， The method jointly optimizes the following three tasks during training: Prediction task: Minimize the mean squared error loss in the prediction of quality variables. L p ; Classification task: Minimize the cross-entropy loss for work condition identification L c ; Shared Feature Alignment Task: Minimize the Maximum Mean Difference (MMD) Loss for Different Operating Conditions in the Shared Feature Space L m , An end-to-end collaborative optimization of feature decoupling, operating condition identification, and soft measurement is achieved through a multi-task learning framework.

2. The method according to claim 1, characterized in that: The preset time window T The length is determined by analyzing the time autocorrelation and cross-correlation of process variables to ensure coverage of the main dynamic response cycle of the system and to satisfy the following conditions. T ≥ max( τ i ),in τ i For the first i The dominant time constants of the key variables.

3. The method according to claim 1, characterized in that: The spatiotemporal feature extractor is a TSMixer structure, which consists of a cascaded temporal mixing network and a feature mixing network. The temporal hybrid network performs fully connected cross-projection on different time steps in each feature dimension to capture long-range temporal dependencies. The feature mixing network performs channel-level fusion of all process variables within the same time step, enhancing cross-variable interaction; Stable training is achieved between the two networks through residual connections and layer normalization.

4. The method according to claim 1, characterized in that: In the hierarchical working condition perception feature fusion module, the first l layer, l ∈ {1,...,L}, working condition m The fusion calculation is as follows: ， ； in, These are the normalized weight coefficients output by the gating network. Indicates the first n e An expert's input combining shared and unique features. This is the output of the previous layer; The first layer input is h c , i and h s , i Final layer output As a predictor R m Input.

5. The method according to claim 1, characterized in that: The classifier φ is a single-layer or multi-layer fully connected network, and its output is M A probability distribution vector is normalized using the softmax function; The probability of identifying the working condition satisfy And during training, it is related to classification loss. L c Joint optimization.

6. The method according to claim 1, characterized in that: The overall loss function of the multi-task learning framework is: L = w p L p + w c L c + w m L m， in, w p , w c , w m The weights are learnable or preset weights, and satisfy the following conditions: w p + w c + w m = 1; The MMD loss L m Defined as the average maximum mean difference among all operating condition pairs: 。 7. The method according to any one of claims 1 to 6, characterized in that: The method is deployed in an industrial process real-time database system. The process variable data collected online is subjected to soft measurement in steps (1) to (4), and the prediction results are fed back into the DCS database to form a "sensing-prediction-optimization" closed-loop control loop.

8. The method according to any one of claims 1 to 7, characterized in that: The industrial process is blast furnace ironmaking, chemical reactor, or cement rotary kiln, and the quality variables include at least one of silicon content, sulfur content, temperature, or component concentration.

9. The method according to any one of claims 1 to 8, characterized in that: When the number of working conditions M > 10. The working condition clustering method is used to merge similar working conditions into subsets and share the same dedicated regression model for each subset in order to reduce model size and inference latency.

10. The method according to any one of claims 1 to 9, characterized in that: The specific regression model R m It is a three-layer fully connected network with hidden layer widths of 128, 64, and 32 respectively. The output layer is a linear layer, and the Adam optimizer is used during training with an initial learning rate of 1e. -3 .

Images