MARPP: Memory-Augmented Reconstruction for Pseudo-Periodic Time Series Anomaly Detection

With the boom in automated manufacturing spurred by the rapid development of Industry 4.0, the accurate and efficient identification of anomalies in the manufacturing process has become a critical issue. However, most reconstruction-based time series anomaly detection algorithms perform poorly on manufacturing monitoring data due to their insensitive to trend anomalies caused by over-generalization. The strong pseudo-periodicity of monitoring data, especially from assembly processes, indicates most monitoring anomalies are trend anomalies. To overcome this problem, we proposed MARPP, a reconstruction-based anomaly detection method tailored to pseudo-periodic time series, which uses an Encoder-Decoder architecture with a trendcapture module to enhance the encoder's representations and a shrinkage memory module to constrain the decoder's reconstruction, effectively reducing over-generalization. Experiments on the force-displacement curves of self-piercing riveting process originating from the real-world assembly scenario of NIO's body-in-white manufacturing demonstrate that MARPP is superior than existing methods on standard anomaly detection metrics and gain an improvement of 3.7 % in terms of f1-score.