Gaussian Mixture Model Using Semisupervised Learning for Probabilistic Fault Diagnosis Under New Data Categories

Abstract

Fault diagnosis has played a vital role in industry to prevent operation hazards and failures. To overcome the limitation of conventional diagnosis approaches, which misclassify new types of faults into existing categories from training, a novel probabilistic diagnosis framework will be proposed in this paper for effective detection on new data categories. Gaussian mixture model (GMM) is applied for the pattern recognition, while its training procedure is improved from conventional unsupervised learning to novel semisupervised learning. Even with unlabeled training data, component number in our GMM can be autoselected instead of predetermined. For online testing, the probabilistic classification results from GMM’s soft assignment assist to improve overall diagnosis framework, which is able to first detect whether new types of faults occur and further categorize them in detail via the GMM update. The effectiveness of our fault diagnosis framework is testified on an industrial fault simulator of rotary machine and the partial discharge measurement of various high-voltage electronic equipment components. Compared with existing approaches, our probabilistic diagnosis framework is able to achieve an average diagnosis accuracy of 97.9% without new data categories and it can also classify new data categories with diagnosis accuracy of at least 86.3% if occurred.