A fault diagnosis method for complex chemical process integrating shallow learning and deep learning

The accurate identification and diagnosis of chemical process faults are crucial for ensuring the safe and stable operation of production plants. The current hot topic in industrial process fault diagnosis research is data-driven methods. Most of the existing fault diagnosis methods focus on a single shallow or deep learning model. This paper proposes a novel hybrid fault diagnosis method to fully utilize various features to improve the accuracy of fault diagnosis. Furthermore, the method addresses the issue of incomplete data, which has been largely overlooked in the majority of existing research. Firstly, the variable data is effectively fitted using orthogonal non-negative matrix tri-factorization, and the missing data in the matrix is solved to construct a complete production condition relationship. Next, the support vector machine model and the deep residual contraction network model are trained in parallel to pre-diagnose process faults by mining linear and non-linear interaction features. Finally, a novel mapping relationship is established between the result and model levels using the multi-layer perceptron algorithm to complete the final diagnosis and evaluation of the fault. To demonstrate the effectiveness of the proposed method, we conducted extensive comparative experiments on the Tennessee Eastman dataset and the ethylene plant cracking unit dataset. The experimental results show that the method has advantages in different evaluation metrics.