Contrastive Learning-Based Dual Autoencoder for Anomaly Detection in Loader Gearboxes

Anomaly detection (AD) of gearboxes is essential for ensuring the operational safety and reliability of the loader. However, identifying anomalies in non-stationary signals remains challenging as anomalies often emerge within the normal fluctuation, especially when normal and abnormal samples exhibit high similarity. This brief proposes a contrastive learning-based dual autoencoder (AE) AD method for loader gearboxes. Specifically, the continuous wavelet transform is employed to capture dynamic characteristics of non-stationary signals. A compound scaling network is then designed into the unified encoder to extract complex features while maintaining a lightweight architecture. Subsequently, a sparse representation channel is integrated into the second AE framework, complementing the basis for contrastive mechanisms and promoting the learning of consistency across normal samples with the reconstruction channel. By minimizing the contrastive loss between two samples from different channels, the model learns the inherent consistency of normal samples. Finally, the contrastive loss of the second AE and the reconstruction error of the first AE serve as indicators for detecting abnormalities. Experimental results on real-world loader gearbox data demonstrate that the proposed method achieves a high fault detection rate, a low false alarm rate, and robust reliability, validating its effectiveness.