A Quality Soft Sensing Method Designed for Complex Multi-process Manufacturing Procedures
Objective Accurately perceiving key quality variables in complex manufacturing processes is essential for achieving system optimization control and ensuring safe and stable operation. Given the numerous production procedures, interconnected loops, quality inheritance across procedures, and the spati...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Editorial Department of Journal of Sichuan University (Engineering Science Edition)
2024-11-01
|
| Series: | 工程科学与技术 |
| Subjects: | |
| Online Access: | http://jsuese.scu.edu.cn/thesisDetails#10.12454/j.jsuese.202400080 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Objective Accurately perceiving key quality variables in complex manufacturing processes is essential for achieving system optimization control and ensuring safe and stable operation. Given the numerous production procedures, interconnected loops, quality inheritance across procedures, and the spatiotemporal distribution of data, precisely perceiving process quality presents many challenges. In this context, this study proposes a soft measurement modeling method for quality in multi-procedure complex manufacturing processes, considering process time delays, based on mRMR-GA-ResNet.Methods First, in complex manufacturing processes with continuous multi-procedure production, different process variables and quality variables exhibit varying time delays, complicating the selection of auxiliary variables. This study proposes a multi-sensor process variable selection method that considers the time delay between process and quality variables based on minimum redundancy maximum relevance (mRMR) and genetic algorithm (GA). Multi-sensor process variables are segmented by the procedure according to the process knowledge. For each procedure, the time delay effect between different procedure segments and quality variables is considered, and data alignment is performed. Using mRMR as the relevance measure, the initial feature subset for each procedure is obtained. GA is employed as the optimal feature search strategy, where the initial feature subset is binary encoded (1 for selected feature, 0 for unselected feature), and an initial population is randomly generated. The fitness function is constructed by combining the prediction accuracy of the extreme gradient boosting (XGBoost) model and the number of selected features, to measure the fitness of each individual in the population. Additionally, an iterative optimization search is performed to determine the final optimal feature subset, thus deriving the auxiliary variable set for modeling. Second, considering the spatiotemporal correlation coupling and quality inheritance characteristics of complex manufacturing processes, to better utilize the optimal feature subset and fully explore the deep information within and between different procedures, this study proposes a three-dimensional (feature-time-procedure) sample space representation method. Temporal-spatial feature extraction is performed using a residual network, and the final quality prediction value is obtained through local-global feature fusion. Since the number of features in the optimal feature subset of each procedure may vary, the procedure with the fewest features is taken as the reference, and the optimal feature subsets of other procedures are dimensionally reduced using principal component analysis (PCA) to ensure that all procedures have the same number of features. Based on the optimal feature subsets of different procedures after PCA dimensionality reduction, three-dimensional samples are constructed. In traditional soft measurement modeling methods, samples are generally in a two-dimensional form. In this study, within each procedure, samples are represented in a two-dimensional (feature-time) form, and procedures are constructed as channels to build three-dimensional (feature-time-procedure) samples. Upon extracting local (within-procedure) spatiotemporal features, global (inter-procedure) feature fusion is performed through a combination of three-dimensional samples and convolution operations.Results and Discussions Experimental validation is conducted using an actual manufacturing process, specifically the float glass production process. In order to verify the effectiveness of the proposed time-delay handling method, the following comparative experiments are set up: In Experiment 1, the time delay between process variables and quality variables is considered, and data alignment is performed according to the proposed method. In Experiment 2, the time delay is not considered. Four correlation-based feature selection algorithms, Pearson correlation coefficient (PCC), Spearman correlation coefficient (SCC), mutual information (MI), and minimum redundancy maximum relevance (mRMR), are used for feature selection, and the same number of features are selected for both experiments. XGBoost is used as the prediction model to predict the quality variables. The results indicate that the proposed time-delay handling method improve the prediction accuracy of the model, validating the effectiveness of the proposed method. To further verify the effectiveness of the proposed multi-sensor process variable selection method, comparative tests are conducted using the four correlation-based feature selection algorithms (PCC, SCC, MI, and MIC) with the same number of selected features for each algorithm. Both XGBoost and a residual network are used as prediction models to predict waviness.Conclusions The results indicate that, under the same number of selected features, the proposed multi-sensor process variable selection method outperforms the other four correlation-based feature selection methods (PCC, SCC, MI, MIC) in terms of predictive performance, achieving a balance between feature selection time complexity and prediction accuracy. Moreover, compared to the traditional two-dimensional (feature-time) sample construction method, the proposed three-dimensional (feature-time-procedure) sample space representation method, utilizing ResNet for temporal-spatial feature extraction and local-global feature fusion, further improves the model’s prediction accuracy. The <italic>E</italic><sub>RMS</sub>, <italic>E</italic><sub>MA</sub>, and <italic>E</italic><sub>SMAP</sub> are improved by 9.2%, 10.8%, and 9.8%, respectively, verifying the effectiveness of the proposed method. |
|---|---|
| ISSN: | 2096-3246 |