Li Tian's research while affiliated with Shaoxing University and other places

Modern industrial processes are increasingly complex, where multiple characteristics usually coexist in process data. Therefore, traditional monitoring methods based on a single model may ignore other data characteristics and obtain poor monitoring performance. Aiming at this problem, a novel monitoring method based on multi-model information extraction and fusion is proposed in this paper. Firstly, several methods are used to extract different characteristics from process data. For example, principal component analysis, independent component analysis and slow features analysis can be used to extract Gaussian, non-Gaussian and dynamic characteristics respectively. Secondly, features extracted from multiple models are combined into new potential features. Then, Lasso regression models between potential features and process variables are established. In this way, not only are multiple characteristics in process data considered during the reconstruction, but key potential features (KPFs) can be selected for each process variable. The KPFs for each process variable can form a monitoring subspace to enhance the sensitivity for fault detection. Furthermore, cluster analysis is used to reduce the redundancy of monitoring subspaces based on the similarity of each subspace. Process monitoring can be achieved by fusing the monitoring results of finally determined multiple subspaces and residual space. Case studies on three simulation processes and a real industrial process demonstrate the effectiveness and better performance.

Traditional deep neural network (DNN) based process monitoring methods only use the deep features of the last layer and residuals to achieve fault detection. However, the features in different hidden layers are different representations of the input data, which may be beneficial to process monitoring. Only using the deepest features for process monitoring will cause the problems of information loss and low monitoring performance. To obtain more useful information for fault detection, this paper considers the features in all hidden layers and proposed an ensemble monitoring model based on multi-subspace partition of deep features. Firstly, a DNN model is established based on the collected faultless data to obtain the features in all hidden layers and residuals. Secondly, a new feature matrix is constructed based on the retained deep features and residuals. Then, the multi-subspace partition of the new feature matrix is realized by combining correlation analysis and cluster analysis. Finally, the monitoring statistics that are established based on the features in each subspace are fused to realize process monitoring. The proposed method can not only avoid information loss but also enrich the fault-related information. The monitoring performance is verified through two benchmark processes and one actual industrial process.

Industrial process data shows the coexistence of multiple characteristics, such as linear, nonlinear, Gaussian, non-Gaussian, and dynamic. Various multivariate statistical analysis methods were applied for different process characteristics. However, using only one method may not have the ability to capture complex characteristics and the relationship between the variables. Therefore, this study designs an ensemble monitoring framework that can automatically determine the local models and the optimal monitoring variables. Firstly, multiple models that can extract different data characteristics are selected as candidate models. Secondly, combining the fault information and intelligent optimization algorithm, the monitoring performance differences are compared when different local models are selected for ensemble, so as to realize the elimination of monitoring redundant models. Finally, the process monitoring is implemented by integrating the determined local models. Under this framework, multiple models that describe the complex characteristics of process data from different aspects can be automatically determined to establish an ensemble monitoring model based on the various characteristics of the process data. Tennessee Eastman process and wastewater treatment process are used to verify the monitoring performance of the proposed framework.

In actual industrial processes, data are usually time series. Each process variable may have strong autocorrelation and cross-correlation with other variables with different delays. In addition, there are usually complex nonlinearities among variables. To further improve the monitoring performance for dynamic nonlinear processes, establishing a nonlinear filtering model for each variable is necessary. Therefore, a novel dynamic nonlinear process monitoring method based on dynamic nonlinear feature selection and kernel principal component regression (KPCR) is proposed in this study. First, dynamic nonlinear related variables are selected for each variable through mutual information by considering variables with different time delays. Second, process variables are divided into response and independent variable sets. Third, corresponding KPCR models are established to describe the dynamic relationships with the selected dynamic related variables as input variables and with response variables as output variables. To monitor the dynamic processes, kernel principal component analysis model is constructed on the basis of the residuals, where the residuals are obtained by comparing measured values of instruments with the predicted values of KPCR models. A support vector data description model is established to monitor the independent variables. Three cases are used to verify the performance of the novel approach. Results show that the proposed method is superior to and more effective than other advanced dynamic process monitoring methods.

Fault diagnostic methods based on deep learning for industrial processes are becoming a research hotspot. Most existing methods focus on algorithmic improvements and attempt to establish a single model to extract effective features of faults. However, effective information related to different faults is diverse. Therefore, instead of using a single model to extract features and build a model to correctly diagnose all types of faults, we propose a novel fault diagnostic method based on deep learning and multimodel feature fusion. First, the minimum redundancy-maximum relevance method is used to select the variables that are the most relevant to each fault. Next, the features of each fault are extracted using a stack autoencoder, and the corresponding residual matrices are obtained. The features and residuals obtained using each model are then spliced as new inputs to establish a classifier for fault diagnosis. Finally, we apply the proposed method to the Tennessee Eastman benchmark process to demonstrate its performance and efficiency.

Determining whether a fault occurs locally or globally is highly important for large-scale industrial processes involving multiple operating units. Moreover, the complex non-linearity among process variables is a prominent feature of modern industries. This paper proposes a distributed-ensemble stacked autoencoder (DE-SAE) model based on deep learning technology for monitoring non-linear, large-scale, multi-unit processes. First, the deep features of the variables involved in each operating unit are extracted with the stacked autoencoder (SAE) to represent the essential structure of the unit. Two statistics are separately constructed using the deep features and the reconstruction error for detecting the faults in local units. Subsequently, the deep representations of the variables from each operating unit are modeled with the SAE to extract the global information for global monitoring. The proposed DE-SAE model uses deep learning techniques to solve the complex non-linear relationships in industrial processes, while considering their local and global information. Therefore, the method can explain the monitoring results better. Experimental results obtained from the numerical simulation and Tennessee-Eastman process confirm the feasibility and superiority of this method.

In the differential evolution (DE) algorithm, many adaptive methods have been studied in terms of fitness values. However, few studies exist on the information from individuals with potential, which presents a large difference in fitness values from that of previous individuals and contains much evolution information. This study proposes a high-performance DE (PDE) algorithm guided by information from individuals with potential. In PDE, all individuals are divided into individuals with potential and individuals without potential according to their improvement in fitness values. The experience learned from the generation of individuals with potential is used to guide future individuals. At each generation, the selection probability of each strategy in the strategy pool is determined by the strategy’s contribution to the improvement in fitness values when generating individuals with potential. The parameters are randomly generated with two distributions, and the location parameters of the two distributions are adjusted on the basis of the improvement in fitness values of individuals with potential. Different individuals (with or without potential) may have different characteristics and evolution methods. Therefore, the generation process of individuals with potential is separated into two cases according to whether they are from previous individuals with or without potential. The study results of the two cases are applied to guide the evolution of current individuals with and without potential. The proposed algorithm is evaluated by comparing it with five advanced DE variants on CEC2005 and seven up-to-date evolutionary algorithms on CEC2014. Comparison results demonstrate the competitive performance of the proposed algorithm. The PDE is also applied to estimate the parameters of a kinetic model of p-xylene oxidation process.