CN-121997169-A - Multi-working condition industrial process soft measurement method based on multi-task learning and probability modeling

CN121997169ACN 121997169 ACN121997169 ACN 121997169ACN-121997169-A

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 accuracy caused by heterogeneous mixing of multi-working-condition process samples. The method comprises three core modules of feature decoupling coding, hierarchical feature fusion and probability information aggregation. Firstly, a space-time feature extractor is designed to explicitly decouple multi-working condition data into working condition sharing features and specific features, so that hybrid characteristic expression and working condition identification are realized. And then, constructing a hierarchical feature fusion module, realizing deep fusion of information among working conditions through a hierarchical special control network, and modeling complex interaction relations among the working conditions. And finally, providing a probability information aggregation strategy, inputting the fusion characteristics into a corresponding predictor, and weighting each prediction result by using the working condition identification probability to generate a final prediction output. The method brings classification tasks and regression tasks into a unified multi-task learning framework, and ensures good performance of soft measurement of performance indexes in a multi-working-condition process.

Inventors

YANG CHUNJIE
Yan Duojin
GAO DALI
XIAO HANG

Assignees

浙江大学
湖州工业控制技术研究院

Dates

Publication Date: 20260508
Application Date: 20260117

Claims (10)

1. The multi-working condition industrial process soft measurement method based on multi-task learning and probability modeling is characterized by comprising the following steps of: (1) Reading historical data from a DCS database of the industrial process, dividing the data into a plurality of subsets according to working condition labels, constructing a multi-working condition process variable time sequence data set, carrying out serialization processing on each sample by adopting a preset time window T to form an input sample The corresponding quality variable is used as a label y i , and the working condition label is z i epsilon { 1.. The M }; (2) The method comprises the steps of explicitly decoupling an input sample x i into a working condition sharing characteristic h c , i and a working condition specific characteristic h s , i through two parallel space-time characteristic extractors, wherein the sharing characteristic is used for modeling a common dynamic law of a cross working condition, and the specific characteristic is used for distinguishing different working conditions; (3) Constructing a hierarchical working condition perception feature fusion module, wherein the module comprises an L-layer special expert control network, each layer comprises a shared expert group and M special expert groups, each expert group consists of N e expert networks, the shared features and the special features are dynamically weighted and fused through the gate control network, and the fusion features corresponding to each working condition M are iteratively output layer by layer ; (4) Constructing an independent regression predictor R m for each working condition m, and receiving corresponding fusion characteristics And outputs the predicted value Meanwhile, the recognition probability of each working condition is output by using the working condition specific characteristic h s , i through the classifier phi And finally, generating a comprehensive prediction result through probability weighted aggregation: , The method jointly optimizes the following three tasks in the training process: A prediction task, namely minimizing the mean square error loss L p of quality variable prediction; A classification task of minimizing cross entropy loss L c of chemical identification; sharing feature alignment task-minimizing the Maximum Mean Difference (MMD) loss in the shared feature space for different conditions L m , Feature decoupling, working condition identification and soft measurement end-to-end collaborative optimization are realized through a multi-task learning framework.
2. The method of claim 1, wherein the length of the predetermined time window T is determined by analyzing the time auto-correlation and cross-correlation of the process variable to ensure a primary dynamic response period of the overlay system and to satisfy T≥max (τ i ), where τ i is the dominant time constant of the ith key variable.
3. The method of claim 1, wherein the spatio-temporal feature extractor is TSMixer constructed from a sequential hybrid network and a feature hybrid network in series; The time sequence mixing network carries out full-connection cross projection on different time steps on each characteristic dimension, and captures long-range time sequence dependence; the characteristic mixing network carries out channel-level fusion on all process variables in the same time step, so as to enhance inter-variable interaction; And stable training is realized between the two networks through residual connection and layer normalization.
4. The method of claim 1, wherein in the hierarchical condition-aware feature fusion module, the first layer, L e { 1..once., L }, the fusion of the condition m is calculated as: , ; Wherein, the For the normalized weight coefficients output by the gating network, Representing the splice input of the nth e expert pair sharing and specific features, Outputting for the upper layer; The first layer inputs are h c , i and h s , i , and the last layer outputs As input to predictor R m .
5. The method of claim 1, wherein the classifier phi is a single-layer or multi-layer fully connected network, the output of which is an M-dimensional probability distribution vector, normalized by a softmax function; The working condition identification probability Satisfy the following requirements And is optimized in combination with the classification loss L c during training.
6. The method of claim 1, wherein the overall loss function of the multi-tasking learning framework is: L = w p L p + w c L c + w m L m, Wherein w p 、w c 、w m is a learnable or preset weight coefficient and satisfies w p + w c + w m =1; The MMD loss L m is defined as the average maximum mean difference between all pairs of conditions: 。
7. the method of any one of claims 1 to 6, wherein the method is deployed in an industrial process real-time database system, wherein soft measurements are performed on the online collected process variable data according to steps (1) to (4), and the prediction results are fed back and written into a DCS database to form a "perception-prediction-optimization" closed-loop control circuit.
8. The method according to claim 1 to 7, wherein the industrial process is blast furnace iron making, chemical reactor or rotary cement kiln, and the quality variable comprises at least one of silicon content, sulfur content, temperature or component concentration.
9. The method of any one of claims 1 to 8, wherein when the number of conditions M > 10, the condition clustering method is used to combine similar conditions into subsets and share the same proprietary regression model for each subset to reduce model size and inference delay.
10. The method of any one of claims 1 to 9, wherein the proprietary regression model R m is a three-layer fully-connected network, the hidden layer widths are 128, 64, 32, respectively, the output layer is a linear layer, and an Adam optimizer is used in the training process, and the initial learning rate is 1e -3 .

Description

Multi-working condition industrial process soft measurement method based on multi-task learning and probability modeling Technical Field The invention belongs to the field of soft measurement modeling of flow industry, in particular to a multi-working-condition industrial process soft measurement method based on multi-task learning and probability modeling, which solves the problem of insufficient measurement precision caused by heterogeneous mixing of samples of a multi-working-condition process. Background In modern process industry, accurate perception of key quality variables is critical to ensuring production stability, improving energy utilization efficiency and realizing operation safety. Soft measurement techniques can effectively overcome the inherent delay and high cost problems of direct measurement by using easily measured process variables to predict quality variables. With the development of industrial digitization and intelligence, data-driven soft measurement has become a dominant paradigm for achieving real-time quality prediction. In recent years, methods based on deep learning, such as a self encoder (AE), a long-short-term memory network (LSTM), a Convolutional Neural Network (CNN), and the like, have been widely studied. Most existing deep learning methods assume that the training data and the test data originate from the same working condition and satisfy independent identical distribution conditions. However, this assumption is generally not consistent with actual industrial multi-regime processes caused by parameter set variations and material fluctuations. This results in models that perform well on training data that fail to maintain performance on test data. Taking blast furnace ironmaking as an example, the difference of the grades of iron ores in different batches and the switching of the operating parameters set by operators can cause the obvious change of working conditions. In this case, the data distribution exhibits multi-operating heterogeneity. The conventional model lacks the ability to accurately describe multiple modes, resulting in a decrease in prediction accuracy. Current solutions for multi-condition process quality prediction can be broadly divided into three categories, a transfer learning method, a single global method, and a hybrid integration method. Specifically, the core of the migration learning method is to learn migratable knowledge and realize the adaptation from the source working condition to the target working condition. However, in the migration learning, each condition is regarded as an independent domain and projected to a unified shared feature space, which results in loss of discrimination information between conditions. The single global approach trains directly on all multi-working condition datasets to learn the overall multiple patterns. While single global methods preserve some variation between conditions, they lack explicit condition identification and tend to learn the averaging effect of the conditions, making it difficult to capture complex interactions. In contrast, the hybrid integration method first identifies the operating conditions and then builds a separate sub-model for each operating condition, thereby providing fine-grained predictions. Although the above methods can cope with the soft measurement requirements of the multi-working-condition process, they still face the problems of insufficient multi-working-condition characterization and information aggregation. Due to the complex production flow, the multi-working-condition data comprises two types of characteristics, namely nonlinear dynamic characteristics inside each working condition and intricate sharing and specific information among different working conditions. If these mixing characteristics are not explicitly decoupled during feature extraction, the model will inevitably be disturbed, resulting in a reduced predictive performance. In addition, the degree of dependence of each prediction target in the multiple outputs on different working conditions is different. Thus, the final prediction should selectively integrate information from various operating conditions and preserve industrial interpretability, and it is necessary to explore new solutions to make up for the deficiencies of traditional deep learning-based soft measurements. Disclosure of Invention Aiming at the background, the invention provides a multi-task learning and probability modeling-based multi-task industrial process soft measurement method which is used for realizing real-time high-precision soft measurement under a multi-working condition scene. The technical scheme for realizing the technical purpose of the invention is as follows: A multi-working condition industrial process soft measurement method based on multi-task learning and probability modeling comprises the following steps: (1) Reading historical data from a DCS database of the industrial process, dividing the data into a plurality of s