A Novel Feature Extraction Approach for Mechanical Fault Diagnosis Based on ESAX and BoW Model

Condition monitoring and fault diagnosis are of great significance to the development of modern industry, for they enable enterprises to avoid unexpected interruptions or severe accidents, and extracting the fault-related features from vibration signals is a critical step to achieve accurate diagnosis. Among diverse of feature extraction approaches, symbolic aggregate approximation (SAX) is a promising one that has been introduced into fault diagnosis recently. Nevertheless, when dealing with the sampled vibration signals, the SAX ignores the change of signal frequency characteristics, which eventually leads to information aliasing and cannot ensure the information validity. In this work, the information aliasing is analyzed from the perspective of signal processing, and the extremum symbolic aggregate approximation (ESAX) is developed as a substitution on the premise of maintaining the validity of the information. Subsequently, to convert the symbol strings generated by the ESAX into usable digital feature vectors, the bag-of-words (BoW) model in natural language processing (NLP) is employed to perform the counting statistics of the fault-related words, and the Laplacian score (LS) algorithm is then utilized to rerank the statistical results, thereby realizing the extraction of mechanical fault feature. The superiority of the developed method is verified by experiments.