Multi-channel convolutional neural network and decision-level fusion for multi-part cover fault diagnosis

Multi-part covers (MPCs) are a special type of urban road cover, which plays an important role in underground facilities maintenance and road safety. MPC failure results in catastrophic safety consequences for pedestrians and vehicles. At present, methods for road cover health detection mainly rely on manual inspection, which consumes a large amount of manpower and resources, and is also inefficient. Additionally, the accuracy is unacceptable in most situations due to the internal failures being easily overlooked. Therefore, an accurate and intelligent diagnosis approach for MPC is urgently needed. This paper proposes an integrated approach based on a multi-channel convolutional neural network (CNN) and decision-level fusion for MPC diagnosis. Using a multi-channel CNN model to adaptively extract features from vibration in three directions and realize preliminary classification. Based on the preliminary classification results, a decision-level fusion approach based on improved Dempster–Shafer (DS) evidence theory is proposed to resolve the complex fault diagnosis results. Three directions of the vibration signal in real road conditions are collected to verify the proposed methodology. The results demonstrated that the proposed method shows superior performance compared to alternative diagnosis approaches. • Proposed a convolutional neural network model for multi-channel fault diagnosis. • The decision-level fusion strategy is optimal for multi-part covers fault diagnosis. • The diagnostic accuracy is enhanced by an improved Dempster–Shafer evidence theory.