Bearing Fault Diagnosis Using Domain Adaptation Approach for Acoustic Data

Bearing failures are one of the most occurring problems in industrial machines. Thus, bearings require improved fault detection methods. In this direction, data-driven approaches for machine fault diagnosis have proven to be more effective than the model-based approaches. However, conventional data-driven methods in domain shift conditions are unable to yield optimal performance. Bearing faults usually occur under different operational conditions. Related to this, acoustic emissions as a non-invasive can capture valuable information about machine health conditions and it is considered as an effective alternative to vibration and current-based methods. Moreover, the application of acoustic data in the domain-shift scenario has not been much explored. In this research, we implement a transfer learning approach for bearing fault diagnosis using machine acoustic data while considering the domain shift problem. Three deep learning models including 1DCNN, 1DCNN-LSTM, and a Residual network are developed and investigated in this research. The pre-trained models are implemented based on the DCASE dataset. The pre-trained models are established using the Air Compressor dataset. By using transfer learning, the feature parameters obtained during model development on the Air compressor dataset are utilized to fine-tune the model on the DCASE dataset. The results demonstrate that the model accuracy through the proposed approach is improved to 89.7% for the target domain. The hybrid model 1DCNN-LSTM demonstrated the best results than the other two algorithms.