A novel cross-fault-condition transfer strategy for remaining useful life prediction of gear

Abstract Accurate prediction of the remaining useful life (RUL) of gears is essential for ensuring the safe operation and efficient maintenance of rotating machinery. However, in practical industrial settings, the gear failures exhibit diverse and complex patterns, and life cycle data, especially failure samples, are often extremely scarce. Consequently, the RUL prediction of gears under different fault conditions with limited samples faces significant challenges. In this study, a cross-fault-condition health assessment system (CFC-HAS) is proposed, which is designed to overcome the limitations of the gear RUL prediction under varying fault conditions with a limited sample size. The proposed method consists of two key components. First, a novel cross-fault health index (CFHI) transfer framework, named TQSDCAE, is proposed. This framework integrates a dynamic convolutional autoencoder (DYCAE) with quadratic function shape constraints and self-attention mechanisms, further enhanced by multi-kernel maximum mean discrepancy. The DYCAE adaptively adjusts kernel weights for flexible cross-fault feature extraction, while self-attention mechanisms enhance focus on temporal correlations, optimizing the capture of fault feature evolution and improving fault identification accuracy. Second, an enhanced gated temporal convolutional framework (EGTCF) is employed to predict the RUL of gears based on the generated CFHI. The EGTCF incorporates gated convolutional units, dilated convolutions, and gated activation units to enhance the model’s ability to understand time-series data, improving its predictive accuracy for long-term dependencies and short-term fluctuations. According to the experiments, the effectiveness of the proposed CFC-HAS method is demonstrated through mutual transfer prediction between tooth breakage and pitting faults. The CFC-HAS method performs exceptionally well at various prediction points, and it outperforms both traditional methods and the current state-of-the-art RUL prediction techniques in terms of prediction accuracy and stability.