Multi-grained mode partition and robust fault diagnosis for multimode industrial processes

Practical industrial processes usually operate under multiple conditions to meet the requirements of manufacturing strategies. A general choice is to partition data according to the number of operating modes and then develop learning models to suit each mode. However, data from a same mode still exhibit multi-grained patterns due to the non-Gaussian processes, occurrence of faults, etc. Only discovering coarse between-mode correlations may fail to achieve precise mode partition, resulting in unsatisfied diagnosis performance. To this end, this paper presents a novel method for multimode industrial processes fault diagnosis. Firstly, the hierarchical clustering strategy exploits the multi-grained information of process data, modeling both the between-mode (different operation conditions) and within-mode correlations (patterns in each mode). Then, a feature learning algorithm based on nonnegative matrix factorization (NMF) is proposed to learn data features, allowing a sample to be represented by the discovered multi-grained structural information. A weighted metric is also designed to reasonably measure the feature similarities learned by the NMF. Particularly in our framework, a l(p)-norm (0 < p <= 1) based minimization strategy deals with noise and outliers in practical processes. The effectiveness of the proposed framework is validated on a numerical example and a multiple-phase flow process.

Automated summary

This paper proposes a novel method for multimode industrial processes fault diagnosis, which utilizes hierarchical clustering strategy to analyze the multi-grained information of process data and model the correlations between operating modes and patterns within each mode. A feature learning algorithm based on nonnegative matrix factorization (NMF) is then proposed to learn data features and represent samples by discovered multi-grained structural information. Moreover, a weighted metric is designed to measure the feature similarities learned by NMF. The effectiveness of the proposed framework is validated on a numerical example and a multiple-phase flow process.