LSC-VAE: a lightweight separable convolutional variational autoencoder for unsupervised and generalisable gearbox fault detection

Artificial intelligence can help detect gearbox faults, which is critical to timely intervention in the operation to reduce downtime and costs. However, current methods are limited because they require labelled fault data that is expensive and impractical to obtain in real-world scenarios. Additionally, the real-time performance of these models is still concerning. Based on this motivation, this study employed an unsupervised, Lightweight Separable Convolution-Based Variational Autoencoder (LSC-VAE) approach for fault detection in gearboxes. The proposed approach processes the raw accelerometer signals in periods to monitor the health of the mechanism directly. The performance of the proposed approach was demonstrated considering a two-stage industrial gearbox and a benchmark dataset to ensure the generalisability. The real-time suitability of the model was determined by establishing a local network using streaming data, and predictions were obtained in 2-s intervals. It was found that the LSC-VAE elicited an overall accuracy of 99.22% for the industrial gearbox and 99.39% for the benchmark dataset. Moreover, the developed model successfully perceived early-fault occurrence with an accuracy of 99.13%. Due to its high performance, small size, and computational complexity, the proposed approach can be deployed as an easy-to-tune, generalisable, and effective model for real-time fault detection in gearboxes.