Multi-Branch Transformer-Based Anomaly Detection Method for Multi-Source Sensor Monitoring Signals of Milling Machine Tools

Abstract In the milling process, tool wear is inevitable, and the effective prediction of tool wear status plays a significant role in improving machining quality and production efficiency. However, during the milling monitoring process, abnormal signal data often occurs due to factors such as sensor installation, operator actions, and environmental interference. This reduces the quality of the monitoring signal data, which in turn affects the accuracy and stability of the tool wear monitoring system. To address this issue, a multi-branch Transformer-based method is proposed in this paper for anomaly detection in milling tool monitoring signals from multi-source sensors. The primary objective is to identify abnormal signal data, which can then be discarded or corrected, ultimately improving the quality of the monitoring data. Specifically, the proposed method extracts local and global features in the feature dimension through multi-scale convolution and fully connected networks. It optimizes the weights of multi-sensor monitoring signals using an attention mechanism and employs parallel Transformer encoders for temporal feature extraction and fusion, generating reconstructed data. Subsequently, an anomaly determination strategy based on reconstruction loss and Euclidean distance is used to detect the monitoring signals of multi-source sensors, mark the abnormal data, and provide the corresponding reconstructed data as correction values. Experimental results demonstrate the effectiveness and reliability of the proposed method. This approach can provide data support for further signal data cleaning and tool wear state monitoring, showing promising application prospects.