Mechanical Fault Diagnosis Using Audio Signal Feature Fusion and Fuzzy Inference

Fault diagnosis is one of the key technologies for achieving intelligent manufacturing. However, implementing fault diagnosis in special environments, such as high-pressure, high-altitude, and toxic conditions, presents a significant challenge. The method of mechanical fault diagnosis based on audio signal analysis (MFDA) offers unique advantages in handling fault diagnosis in special environments, as it enables non-invasive and non-contact measurements. This paper proposes a feature called “spectral similarity” for MFDA and validates the effectiveness of the proposed feature through experiments. Additionally, this paper combines feature fusion with a fuzzy inference system to propose an MFDA method suitable for high-noise and small-sample scenarios. Firstly, the audio signals are preprocessed and then segmented into frames. Next, the spectral similarity feature, sound quality features, and time-domain features are extracted. Subsequently, feature fusion is performed based on the results of correlation analysis and kernel principal component analysis (KPCA). Then, a fuzzy inference system is established by determining the membership functions and linguistic terms, and defining fuzzy rules. Finally, the fused features are input into the fuzzy inference system to obtain the diagnostic results. The experimental results demonstrate that the proposed method achieves an accuracy of 98.16% for motor fault diagnosis in low-noise scenarios and an accuracy of 97.64% for tool wear diagnosis in high-noise scenarios. The proposed method outperforms the one-class support vector machine (OCSVM) method, the K-Means clustering method, etc.