Multi-Sensor Fusion Techniques for Anomaly Detection in Safety-Critical Automation Systems

Abstract

The reliability and intelligence of anomaly detection in safety automation systems can be improved through multi-sensor fusion technology. First, this paper aims to address the issues of heterogeneous sensor data, high dimensionality, and different operating environments in next-generation industrial systems through the use of deep learning fusion technologies. One-dimensional convolutional neural networks are used for local feature extraction; bidirectional recurrent neural networks combined with attention mechanisms are used to simultaneously process spatial and temporal information from multiple sensors. The industrial dataset used for experimental evaluation contains over 18,000 synchronized multi-channel sequences and is labeled with three types of events. The average detection accuracy and recall rate of the fusion model are 0.92 and 0.89, respectively, both surpassing traditional statistical methods and single-sensor neural networks. Through visual analysis, feature learning has achieved its goals. The saliency map improved interpretability, and the robustness to sensor loss and data noise was enhanced. Ablation studies have found that the fusion and attention modules in the system are indeed beneficial. Therefore, through detailed multi-sensor fusion, a stable and all-weather complex automation fault detection system can be created. This method can help monitor and digitize the maintenance and digitalization of factories.