A Production Process Anomaly Detection Model Combining XLSTM and Negative Sample-Aware VAE
Ni Tang, Tao Zhou, Zhi Jun Ding, Yang Ni, Jie Shi, Yang Wang · 2025
Anomaly detection algorithms can leverage time-series data collected from multi-source sensors to locate production anomalies in real-time. However, existing generative model-based methods often suffer from overfitting to sparse anomaly features, inefficiency in modeling long sequences with low information density, and difficulty in dynamically updating memory, leading to low detection accuracy and severe anomaly region offset issues. To address these challenges, a novel anomaly detection model combining XLSTM and Negative Sample-Aware VAE is proposed. This model introduces negative samples based on shared model weights to enhance the VAE’s perception of anomaly pattern distributions, increasing the difference between negative samples and reconstructed positive samples. Using XLSTM, the model performs sequence modeling across multiple consecutive detection windows, addressing the efficiency bottleneck in long-sequence modeling and the inefficiency of computing correlations in sparse data. Additionally, an anomaly transition region shifting enhancement strategy is applied to improve the model’s recognition at sequence transition points, alleviating the problem of severe anomaly region offset. Experiments were conducted on four publicly available labeled datasets and one cigarette production process dataset. Compared with the AE-LSTM model, the proposed method achieved an average precision improvement of 22.9 percentage points, an average recall improvement of 18.7 percentage points, and an average F1-score improvement of 18.3 percentage points on the public datasets. Compared with the VAE-LSTM model, it improved average precision by 9.1 percentage points, average recall by 6.3 percentage points, and average F1-score by 8.8 percentage points. On the cigarette production process dataset, the optimized model showed better fitting performance, achieving a lower false detection rate, which verified the effectiveness and generalizability of the optimized model in practical production scenarios.