Machinery Early Fault Detection Based on Dirichlet Process Mixture Model

The most commonly used single feature-based anomaly detection method for the complex machinery, such as large wind power equipment, steam turbine generator sets, and reciprocating compressors, exhibits a defect of low-alarm accuracy due to the non-stationary characteristic of the vibration signals. In order to improve the accuracy of fault detection, a novel method based on the Dirichlet process mixture model (DPMM) is proposed. First, the features of the mechanical vibration signals are used to construct the feature space of the equipment. The DPMM modeling method is then applied to self-learn the probabilistic mixture model of the feature space. The normal working condition model is used as the benchmark model. The early fault detection is realized by using a precise difference measurement method based on Kullback-Leibler divergence to calculate the difference between the real-time model and the benchmark model accurately, and by comparing the calculation result with a self-learned alarm threshold. The effectiveness and the adaptability of this novel early fault detection method are verified by comparing it to the single feature-based anomaly detection method and the Gaussian mixture model (GMM)-based early fault detection method.