Pioneering Industrial Anomaly Detection with a Hierarchical LSTM-Rola Framework

Dingyu Chen, Shaohua Liu, Le Yuan

Abstract: In the manufacturing sector, establishing a system for fault diagnosis and analysis on production lines is of paramount importance. This research presents a novel hierarchical anomaly management method, addressing issues such as data category imbalance, challenges in detecting abnormal signs. The study utilizes the LSTM-Rola (rolling accumulation) approach, specifically designed for time series forecasting, to effectively identify anomalies. This method employs a stacked LSTM structure in an encoder-decoder framework, combining single-step and multi-step predictions to enhance both short-term and long-term forecasting capabilities. The anomaly detection aspect of the method incorporates an accumulation of abnormal occurrences and categorizes anomalies into different levels. This strategy not only improves detection accuracy but also resonates with traditional fault mechanism theories, facilitating easier interpretation of the model. Additionally, the paper includes comparative studies on various normalization methods and early warning accumulation tactics, demonstrating the model's effectiveness. The model shows remarkable performance in time series anomaly detection, achieving an F_0.5 score of 0.8259, a high precision of 91.8%, and a recall rate of approximately 60%.

External IDs:dblp:conf/ijcnn/ChenLY24