Electric Motor Bearing Fault Noise Detection via Mel-Spectrum-Based Contrastive Self-Supervised Transformer Model

Bearings are vital components of motor drive systems and are widely used in various industrial applications. Bearing failures can lead to system collapse and pose a risk to human safety. Therefore, real-time monitoring and diagnosis of multi-fault bearings are crucial. This paper proposes a Mel-spectrum-based contrastive self-supervised Transformer (Mel-CSST) model to efficiently detect multiple bearing faults in electric motors through vibration noise signals. Among them, the contrastive self-supervised Transformer model (CSST) can be pre-trained without the need for labeled data, significantly improving the fault detection accuracy of the target bearing after transfer learning using the parameter-frozen domain-adversarial (PFDA) method. Mel-spectrums are converted from a mass of sub-signals generated by the random-masked sliding window (RMSW) method, providing training data sample pairs for the CSST model. Mel-spectrums can analyze significant vibration noise signals at lower frequencies in more detail, revealing the fault features missed by the standard fast Fourier transform. Furthermore, the encoder part of Mel-CSST uses a modified Transformer network to ensure the feature extraction effectiveness of CSST. The proposed method can be easily transferred to be used on target bearings without expensive labelling data in practical applications. Experiments using two real bearing datasets measured from two test rigs, along with comparison experiments with other existing methods, validate the effectiveness of the proposed method.