An adaptive imbalance modified online broad learning system-based fault diagnosis for imbalanced chemical process data stream

Modern chemical process industry is becoming larger and more complicated to achieve a higher level of technical functionality. There is less tolerance for functional degeneration, productivity retrogression, and safety hazards, which significantly leads to an ever-increasing demand on detecting any potential faults as early as possible. In reality, chemical process data are continuous generated with imbalanced fault patterns, which leads to the fault diagnosis models failing to assign the same attention to minority fault patterns as the majority and further leads to the lack of generalization ability. In the present study, a novel adaptive imbalance modified online broad learning system (AIM-OBLS) is developed to promote fault diagnosis in the contexts of imbalanced chemical process data streams. The proposed AIM-OBLS is developed on a flat linear network, which can excavate potential information efficiently in an incremental manner. An adaptive imbalance modified method combined with the Niche technique, oversampling technique, and manifold regularization is presented for imbalanced data streams modification. The advantages of the proposed AIM-OBLS are demonstrated on two widely used industrial datasets. Experimental results indicate that AIM-OBLS can effectively deal with the imbalance chemical process data streams. The performance of AIM-OBLS is competitive in terms of both diagnosis accuracy and time consumption.

Automated summary

With the increasing complexity and size of the modern chemical process industry, there is a growing demand for detecting potential faults as early as possible. However, the imbalanced fault patterns in continuous chemical process data streams hinder the generalization ability of fault diagnosis models. In this study, a novel adaptive imbalance modified online broad learning system (AIM-OBLS) is proposed to address this issue and has shown competitive performance on industrial datasets.