An Improved Genetic Clustering-Based Fault Detection for Reciprocation Compressor Valve Train

Abstract

There is an irregular data distribution when using multi-sensor to monitor reciprocating compressor conditions. It is difficult to deal with conventional approaches. In this paper an improved genetic algorithm based clustering method is used to solve the problem. First a prototype-based genetic representation is utilized, where each chromosome is a set of positive integer numbers that represent a specific sequence number of the k-medoids. A geodesic distance based proximity measures is adopted to measure the similarity among data points. To improve the algorithm searching performance, the probabilities of crossover and mutation can be dynamically adjusted based on the information entropy of population fitness distribution. We apply the algorithm to detect fault conditions of inlet valve train leakage in a two-stage reciprocating compressor. Experimental results demonstrate the effectiveness of the algorithm as a clustering technique. Compared with generic K-means algorithm for clustering task, the presented algorithm has the ability to identify complicated non-convex clusters and its clustering performance is clearly better than that of the K-means algorithm for complex manifold structures.