Search

During the process of monitoring conveyor belt faults in underground coal mines, issues such as false alarms and low recognition rates arise. This was attributed to the extended monitoring distance of the belt conveyors, strong electromagnetic interference, and high environmental noise during the monitoring. Simultaneously, interference from other events can occur when a specific event is being monitored. To address this issue, this study proposes a Siamese recognition network based on a distributed acoustic sensor (DAS) system, featuring a multichannel hybrid parallel classification network (Si-MCHPCN). This network can accurately identify fault events even when they occur at different locations of the conveyor system. The experimental results show that an average accuracy of 94.8% can be achieved for the five event types. Compared to using the same feature extraction network, Si-MCHPCN improved the recognition accuracy by 9.5% and 11.6% compared to using a single-signal feature extraction network and by 36.7% compared to using a multichannel parallel classification network alone. Finally, after applying the Siamese network to the multichannel hybrid parallel classification network, not only did it improve recognition accuracy but it also resolved the issue of inaccurate data labeling, further enhancing the data reliability.

Perimeter security intrusion monitoring, relying on distributed optical fiber vibration sensing (DVS) systems, is prevalent, yet effectively identifying vibration signals remains a challenge. This article introduces a method for intrusion event pattern recognition based on deep learning (DL) and multimodal feature fusion, facilitating automatic feature extraction and fusion. The approach integrates 1-D raw time-domain signals (1D-RTD-signal) and 2-D time–frequency spectrum (2D-TF-spectrum) as training data. While the 1D-RTD-signal preserves comprehensive raw data information, the 2D-TF-spectrum unveils time–frequency characteristics of the vibration signal. Specifically, the 1DCNN-DN multimodal feature fusion model is tailored for this purpose. Initially, the branch model independently extracts features, followed by fusion for recognition. Leveraging original time-domain signals streamlines data preprocessing, and the multimodal fusion model enhances model generalization while averting overfitting. Experimental outcomes showcase the scheme’s prowess in recognizing background noise and three types of intrusion events in outdoor environments, achieving an average recognition rate of 99.36%. This presents an optimized solution for DVS perimeter security technology, facilitating the classification and recognition of vibration signals from various intrusion events.