A Fast Fault Diagnosis Method for The Unlabeled Signal Based on Improved PSO-DBSCAN Algorithm

The fault diagnosis of different components with supervised learning method usual requires a large number of training samples. In practical engineering applications, the diagnosis efficiency is low and the failure rate is high due to the small amount of training samples. In order to solve these problems, a step-by-step fast fault diagnosis method based on improved Particle Swarm Optimization (PSO)- Density Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm and Least Squares Support Vector Machine (LSSVM) is proposed. Firstly, the original signal is pre-processed by normalization and wavelet threshold de-noising. Then, the dimensionality reduction by Principal Component Analysis (PCA) is used as the input of the improved PSO-DBSCAN algorithm to cluster the data, and the train samples are formed after the data categories. Secondly, the train samples are used as the input of LSSVM to train the fault classifier. Finally, by using the trained classifier to classify other data, the working state of the component can be obtained. In this paper, by simulating a certain type of engine oil monitoring data, the accuracy of the classification result is 96.67%, which verifies the feasibility and effectiveness of the method, and realizes the fast fault diagnosis of unlabelled signals.