A Distributed Detection Method for Quality-related Faults in Complex Non-stationary Industrial Processes
The statistical characteristics of process data time series in the production process typically vary over time due to changes in production conditions, unknown external interference, load fluctuations, and other factors. Hence, complex industrial processes often exhibit non-stationary characteristic...
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.202400079 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | The statistical characteristics of process data time series in the production process typically vary over time due to changes in production conditions, unknown external interference, load fluctuations, and other factors. Hence, complex industrial processes often exhibit non-stationary characteristics. These non-stationary characteristics can obscure faults in industrial processes. Traditional quality-related fault detection methods cannot detect abnormalities on time, leading to the transfer and evolution of faults between different operational units of industrial processes. This failure threatens the quality of products and the stable operation of the production process and can also damage the economic benefits of enterprises, presenting significant challenges to the quality-related fault detection of complex industrial processes. It is essential to investigate quality-related fault detection methods for non-stationary processes, detect abnormal product quality in time, and take appropriate measures to reduce the adverse impact on enterprises’ production processes and minimize their economic losses. Therefore, this study proposes a new distributed quality-related fault detection method for complex non-stationary industrial processes based on dynamic and static feature fusion. This distributed method considers the local neighborhood relationships and information interaction between sub-blocks, providing an effective process monitoring approach for complex industrial systems. Initially, the minimal redundancy maximal relevance (mRMR) algorithm is employed to reveal and quantify the linear and nonlinear relationships between process and quality variables. This classical method identifies a feature subset that maximizes the correlation between features and targets and minimizes redundancy between features based on mutual information. Process variables highly correlated with quality variables are selected using the mRMR algorithm. Simultaneously, the algorithm eliminates redundancy among the selected process variables to simplify the subsequent construction of the fault detection model, enhancing its accuracy. Then, the selected process variables are categorized into stationary and non-stationary variables using the augmented Dickey–Fuller (ADF) test method, a type of unit root test. It hypothesizes that if a time series can be expressed as an autoregressive model containing a unit root, it is considered non-stationary. In addition, mechanistic knowledge of the industrial process is applied to divide the complex industrial process into multiple physically meaningful sub-blocks. Common variables facilitate the information interaction between sub-blocks. When constructing models for local sub-blocks, information from neighboring sub-blocks is incorporated, enabling the transmission of information in a physically meaningful manner. In contrast, traditional distributed monitoring methods only utilize the information of variables within each sub-block without considering the correlation between different sub-blocks. The static and dynamic features of stationary and non-stationary variables within the sub-blocks are then extracted using partial least squares (PLS) and long short-term memory (LSTM) network methods, respectively, and the features are fused. The dynamic and static information of the industrial process is fully utilized to reduce the dimensionality and feature extraction. The canonical variable analysis (CVA) algorithm maximizes the correlation between the fused features and quality variables. Specifically, the fused features and quality variables are employed to construct past and future data vectors in the CVA algorithm, further extracting the canonical variables and conducting statistical monitoring on the canonical variables. In each sub-block, quality anomaly detection statistics in the state subspace and the residual subspace are constructed using <inline-formula><tex-math id="M1">$ {T^2} $</tex-math><alternatives><graphic specific-use="online" xmlns:xlink="http://www.w3.org/1999/xlink" xlink:href="202400079_M1.jpg"/><graphic specific-use="print" xmlns:xlink="http://www.w3.org/1999/xlink" xlink:href="202400079_M1.png"/></alternatives></inline-formula> and SPE statistics, respectively, and the corresponding control limits are further calculated. Then, the local detection results of each sub-block are fused through Bayesian inference. Bayesian reasoning is effective for decision fusion, and the global quality-related distributed PLS–LSTM–CVA fault detection model is constructed to realize global quality-related fault detection. Finally, the proposed distributed quality anomaly monitoring method is applied to a company’s float glass production line in Hebei Province, and the actual data from the furnace process in the float glass manufacturing process are simulated. The detection performance is evaluated from the perspectives of fault detection rate (FDR) and false alarm rate (FAR). The detection results demonstrated that the proposed distributed quality anomaly monitoring method can quickly detect quality-related faults. In addition, it is compared to the traditional CVA and centralized PLS–LSTM–CVA methods to verify the effectiveness and superiority of the proposed method. The comparison results showed that the proposed method has a better fault detection effect, significantly reducing the FAR. Specifically, the FDR of the proposed method in the state subspace is 100%, which is higher than that of the traditional CVA and centralized PLS–LSTM–CVA methods. The FAR of the proposed method is 4%, significantly lower than that of the traditional CVA and centralized PLS–LSTM–CVA methods. The effectiveness and superiority of the distributed quality anomaly detection method for non-stationary characteristics proposed in this study are further proved. Accordingly, the quality-related distributed detection method proposed in this study employs the mRMR algorithm for variable selection and then decomposes the process using mechanistic knowledge. LSTM and PLS are utilized respectively to extract dynamic and static features in each sub-block, integrating dynamic and static features. The local quality-related fault detection model is constructed based on the CVA method by fully using the process information and characteristics reflecting the abnormal quality, and global fault detection is realized by fusing the local detection results through Bayesian reasoning. The performance of quality-related fault detection is improved. It can effectively reduce the interference of operators, improve work efficiency, and avoid the waste of resources caused by false alarms. This plays a role in ensuring safety and guiding production in the float glass manufacturing process. It has excellent prospects for application in industrial processes. |
|---|---|
| ISSN: | 2096-3246 |