Few-shot anomaly detection in engine cold testing : A synthetic fault augmentation approach with self-attention memory and dynamic thresholding

Engine cold test anomaly detection is a critical component of automotive manufacturing quality control. However, practical applications face significant challenges due to the low probability of fault occurrences and the high cost of acquiring fault data. Consequently, while health data is relatively abundant, fault data remains scarce, resulting in a few-shot phenomenon that limits detection accuracy and model generalization. To address this issue, this paper focuses on a Self-Attention Memory Conditional Generation Model (SAGM) for feature generation to synthesize fault samples, while utilizing a Long Short-Term Memory Autoencoder (LSTM-AE) for anomaly detection. The SAGM learns the distribution of key fault-related parameters from source-domain fault data and generates virtual fault samples by integrating these parameters with normal samples from the target domain, effectively mitigating the scarcity of fault data. The LSTM-AE then models the probability distribution of reconstruction errors and dynamically updates detection thresholds using a sliding window mechanism. To validate the effectiveness of the proposed method, experiments are conducted on a 1.5L four-cylinder engine to distinguish normal operating conditions from rocker arm bearing outer race wear fault. The results demonstrate that the virtual fault samples generated by the proposed approach closely resemble actual fault data in terms of feature distribution and statistical properties. Furthermore, the model achieves a detection accuracy of 99.23 %, significantly enhancing the ability and robustness of anomaly detection in faulty engines, outperforming traditional methods.