Multi-Modal Fusion for Multi-Task Fuzzy Detection of Rail Anomalies

Due to prolonged exposure to heavy train loads, various anomalies can emerge on the surface of railway tracks, posing a direct threat to safe train operation. The accurate and timely detection of these anomalies is important to ensure safe transportation and advancing intelligent maintenance. However, in the domain of anomaly detection, several challenges have arisen owing to the variability in illumination conditions, imaging blur inherent to the capture devices, and the introduction of noise from environmental factors such as dust particles. These interferences have significantly undermined the accuracy of rail anomaly detection based on target detection techniques. In response to these challenges, this study introduced a novel approach for rail anomaly detection in the presence of image artifacts by utilizing a multi-modal multi-task framework. The objective of this study is to enhance the performance of rail anomaly detection under interference-prone conditions. This study integrated color moment features, HU invariant moment features, and Haralick features to construct a fuzzy detection model for rail anomalies using a multi-task learning (MTL) strategy. The model prioritized the primary task of classifying rail anomalies, with interference level classification and fuzzy logic interference judgment serving as auxiliary tasks within the network. Finally, based on the results of fuzzy logic and interference level detection, a fuzzy judgment was made distinguishing between “with interference” and “without interference.” Experimental findings consistently demonstrated that the integration of multi-modal features and multi-task learning methodologies significantly improves the accuracy of rail anomaly recognition in the presence of interference, thus establishing an effective approach for rail anomaly identification in challenging scenarios.