A Lightweight Dilated Convolution- Channel Attention Network for Fault Detection

With rapid advancement of industrial and intelligent manufacturing, Prognostics and Health Management (PHM) has emerged as a pivotal technology for ensuring industrial production safety and reducing maintenance costs. Among its critical components, fault detection serves as the first line of defense in preventing catastrophic accidents in industrial operations. Addressing the challenges of large parameter sizes and high computational costs associated with deep learning models in industrial fault detection scenarios, this paper proposes a lightweight one-dimensional dilated convolution-channel attention network. The proposed method constructs a deep feature extractor by stacking one-dimensional dilated convolutional layers, leveraging exponentially increasing dilation rates to model longterm temporal dependencies. The dilated convolutional layers, while maintaining a small parameter size, extend the effective receptive field to the length of the signal period, significantly outperforming the local convolution of WDCNN (Deep Convolutional Neural Networks with Wide First-layer Kernels). Additionally, a convolutional channel weighting module is integrated to dynamically enhance fault-sensitive features. Compared to the classical WDCNN, the proposed network reduces the parameter size to <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$1 / 15$</tex> of the original and achieves a 1.77 -fold improvement in computational speed. On the CWRU (Case Western Reserve University) bearing dataset, it attains an average detection accuracy of 99.98 %. Furthermore, the model demonstrates robust adaptability to noise, maintaining strong performance on noise-augmented datasets. Under high-intensity noise conditions, the accuracy only decreases by 1.83 %, underscoring the network's ability to achieve superior performance and enhanced noise resistance with fewer parameters.