Fault Diagnosis Strategy Based on Complex Network Analysis

Fault diagnosis, whose essence is pattern recognition of object’s operation state, can be accomplished through clustering methods. The network model is used to represent the fault data structure and thus the clustering problem is converted into the detection task for sub-network structures. Thereby, a fault diagnosis strategy based on complex network structure analysis is proposed. Corresponding to the two central issues for sub-network partition:similarity measure between samples and partition criteria, the modularity concept used broadly in the analysis of community structures in a complex network is introduced into the design of a states differentiating criterion function. To optimize this criterion and accordingly classify fault states, an agglomerative hierarchical clustering algorithm is developed. In applications such as benchmark data classification and four-stage piston compressor diagnosis problem, the effect of similarity measure on algorithm is discussed and the algorithm performance is testified. The comparative results with several artificial intelligent diagnosis algorithms show that the new algorithm can achieve higher diagnosis accuracy, and it is more straightforward, able to extract critical features of the data samples more accurately and therefore accomplishes data clustering with less computational cost.