Online Fault Diagnosis in Industrial Processes Using Multimodel Exponential Discriminant Analysis Algorithm

Fault diagnosis is used to identify the fault cause of the online abnormality, which is crucial for efficient and optimal operation of industrial processes. Due to the time-varying characteristics of fault process, the historical fault data may consist of multiple patterns and cannot be described accurately using a single model, which may result in poor performance of conventional multivariable statistical methods. In this brief, a multimodel exponential discriminant analysis (MEDA) algorithm is proposed for solving the aforementioned fault diagnosis problem. First, the samples of each fault class are clustered into different subclasses using the fast search and find of density peaks algorithm and the proposed cluster index. Subsequently, the between-and within-subclass exponential covariance matrices are calculated based on the subclasses and then the multimodel exponential discriminant model can be developed. Finally, recursively update the cluster center and the multimodel exponential discriminant model until the developed subclasses can be well separated and better discriminant performance can be obtained. Besides, a probabilistic MEDA algorithm and its corresponding online probabilistic diagnosis are also described to develop fuzzy model, so that the fault classes whose subclasses have fuzzy boundaries can also be effectively diagnosed. The Tennessee Eastman process is used to validate the diagnosis performance of the proposed MEDA algorithm, and the experimental results illustrate that the proposed algorithm can efficiently diagnose different classes of faults.