A novel method based on wavelet transform and prototypical network for gearbox detection in few-shot learning

Abstract

Fault diagnosis is crucial for industrial systems, with traditional methods such as CNN heavily reliant on large training datasets to achieve high accuracy. However, such datasets are often-times inaccessible in the real world. Even in few-shot learning models, such as Model-Agnostic Meta-Learning (MAML), the quantity of training data significantly impacts the stability and accuracy of the models, posing challenges for reliable fault diagnosis under limited data conditions. To address these issues, the Wavelet Transform Prototypical Network (WTPN) is proposed, which integrates discrete wavelet transform with prototypical networks for limited training dataset. There are two main structures in WTPN. Firstly, this method transforms one-dimensional vibration signals into two-dimensional distance matrices, enhancing feature extraction and classification accuracy. Secondly, a confidence weighting mechanism assigns weights to decomposed signals based on their classification reliability, thereby improving consistency and reducing performance variability. Then, results from both experimental and publicly available datasets validate that WTPN consistently outperforms existing few-shot learning models in terms of accuracy and stability. Furthermore, the contributions include enhanced feature extraction through DWT, improved stability via confidence weighting, and robust performance in scenarios with limited training data. In conclusion, WTPN represents a significant advancement in fault diagnosis, offering reliable outcomes with minimal training data, making it particularly suitable for applications where data availability is constrained.