Multilevel feature encoder for transfer learning-based fault detection on acoustic signal

Abstract

The intelligent diagnosis of faults in industrial assets is crucial for preventing unexpected disruptions to critical services. Although numerous deep learning methods based on acoustic data have been developed to enhance fault detection accuracy, these methods often prove suboptimal in transfer learning due to two key challenges: (1) insufficient generalization capability that causes overfitting to source device characteristics, and (2) failure to capture domain-invariant patterns essential for cross-device fault detection. This work seeks to alleviate these limitations by proposing a multilevel features encoder (MLFE) for transfer learning-based fault detection on acoustic signal. The acoustic data are initially preprocessed with a frequency mask to filter out high-frequency noise. Subsequently, feature engineering techniques are employed to extract several statistical features, such as the mean, standard deviation, and median absolute deviation, etc. with an emphasis on frequency characteristics. Moreover, unsupervised method is then applied to extract additional essential features. These multilevel features are then combined and fed into MLFE to differentiate between faulty and non-faulty signals. After being trained on several source devices, the pre-trained MLFE is transferred to a new target device to evaluate its transfer learning capability. MLFE is evaluated using the pump and fan datasets in MIMII, where it outperforms existing methods and offers a novel solution for transfer learning-based fault detection using acoustic signals.