A deep domain-adaptive anomaly detection method for scroll compressor speed transfer using domain-adversarial training and attention mechanisms

Anomaly detection during the production process of electric scroll compressors is a critical step in ensuring product quality. However, under variable-speed operating conditions, data distribution shifts and sparse labeled samples in the target domain cause existing diagnostic models to significantly decline in generalization capability. This paper proposes a novel deep domain-adaptive anomaly detection method to address this. Its core contributions are: First, by integrating an improved residual network with a channel attention mechanism, a feature-selection-capable network adaptively focuses on critical frequency-domain features related to faults, effectively suppressing interference from speed variations. Second, a dynamic adversarial training strategy is designed. Periodically adjusting domain loss weights achieves excellent feature distribution alignment while ensuring training stability, simultaneously addressing the challenges of insufficient feature discriminability and training instability in cross-speed transfer. Experiments on real scroll compressor datasets demonstrate that this method achieves average accuracies of 97.67% and 94.93% in low-to-high-speed bidirectional transfer tasks, significantly outperforming mainstream domain adaptation methods and highlighting its outstanding engineering application value. • A novel deep domain-adaptive anomaly detection method for scroll compressors. • Dynamic adversarial training with cosine annealing enhances training stability. • Achieved over 97% accuracy in cross-speed anomaly detection tasks. • Model validated on industrial data with 0.032 ms inference latency.