Prediction of abnormal gearbox conditions using an improved variational autoencoder

This paper presents an enhanced Convolutional Variational Autoencoder (ConvVAE) model for gearbox vibration signal fault detection and diagnosis. Combining convolutional neural networks and variational autoencoders, the model automatically extracts health features from time-series signals, performs nonlinear compression and reconstruction, and generates high-quality latent space representations. The ConvVAE uses a multi-layer 1D convolutional network as the encoder to extract and compress features, employing a reparameterization technique for optimized latent vector sampling. The decoder applies deconvolution to reconstruct the signal's structure. A fusion of time-domain and frequency-domain signals improves pattern recognition, while an adaptive thresholding method dynamically sets fault detection thresholds, reducing false positives. Experimental results show that the proposed method achieves high accuracy and robustness in detecting abnormal signals from gearboxes, providing efficient fault detection and early warning across various operating conditions.