A Hierarchical Coarse-to-Fine Fault Diagnosis Method for Industrial Processes Based on Decision Fusion of Class-Specific Stacked Autoencoders

Fault diagnosis (FD) is crucial for ensuring the safety and stability of industrial processes. In real industrial processes, fault features in measurement data are prone to be misidentified due to the feature overlap, which hinders the effective FD. Besides, the large-scale characteristic of industrial processes also poses challenges for FD. To solve the above problems, a hierarchical coarse-to-fine FD method based on decision fusion of class-specific stacked autoencoders (DFCSSAEs) is proposed to achieve accurate FD by identifying overlapping features. The entire FD process is divided into three stages: sub-block division, coarse diagnosis, and fine diagnosis. In the sub-block division stage, for the large-scale characteristic, a sub-block division method based on prior knowledge and mutual information (MI) is proposed to divide the whole process into several new sub-blocks. On this basis, the faults with overlapping features in each sub-block are identified by leveraging the feature visualization technique and summarized into a new composite class. In the coarse diagnosis stage, several SAE-Softmax-based coarse FD models are established to achieve the targeted diagnosis of individual faults and the composite class fault. At the fine diagnosis stage, a weighted Dempster-Shafer (D-S) evidence theory is proposed to solve decision conflicts among different coarse FD models by assigning reasonable weights. Moreover, a new SAE-Softmax is established to diagnose the unclassifiable faults after the decision fusion and ultimately improve diagnostic accuracy. Finally, the effectiveness and advantages of our method are validated by the Tennessee Eastman process (TEP) and the real dataset of the PRONTO process. Experimental results demonstrate that our method achieves high FD rates of 0.962 and 0.997.

I. Introduction

With the increasing complexity of industrial processes, it is increasingly challenging to ensure the safe operation of industrial processes [1], [2]. Fault detection and diagnosis (FDD) plays a pivotal role in enhancing the reliability and security of industrial processes. Fault diagnosis (FD) is an important part of FDD and its function is to identify fault classes of samples [3]. FD methods mainly include model-based methods, knowledge-based methods, and data-driven methods [4]. Among them, model-based methods achieve FD by analyzing the operating principle of the industrial process to construct a suitable mechanism model, while knowledge-based methods rely on expert knowledge and experience to achieve FD [5]. Yet, the scale and complexity of industrial processes are increasing day by day, making it difficult to build accurate mechanism models and obtain sufficient expert knowledge, which is not conducive to the application of the above two methods in industrial process FD. Recently, with the rapid development of computer technology and sensor technology, a large amount of historical data on industrial processes has been collected and stored [6]. Thus, data-driven-based FD methods have attracted extensive attention [7], [8], [9].