Nonlinear Activation Functions are Not Necessary: A Lightweight Nonlinear Activation Free Network Based on Multiscale Large Kernel Attention Mechanism for Fault Diagnosis

In recent years, there has been a notable advancement in the field of fault diagnosis, with the introduction of Transformers and convolutional neural networks (CNNs). These two approaches have demonstrated distinct advantages in global and local feature extraction, respectively. Nevertheless, the introduction of the multihead self-attention (SA) mechanism and the cross-channel convolutional mechanism has led to a notable increase in the system complexity of the networks, resulting in higher hardware requirements, limited industry applications, and the potential for comparative analyses between methods to be hindered. Accordingly, this article proposes a novel lightweight fault diagnosis framework, designated MSLKA-NAFNet, to address the limitations identified in current research. First, a plainNet is constructed using the most common components (e.g., convolutional blocks, ReLU, and shot-cut); second, a multiscale large-kernel attention (MSLKA) mechanism is designed to extract the global information of the vibration signals through a large receptive field, which is added to the plainNet along with other state-of-the-art methodological components, to propose a simple and efficient baseline. To further simplify the network complexity, basic multiplication operations are used instead of nonlinear activation functions (e.g., Sigmoid, ReLU, and GELU) to transform the baseline into a nonlinear activation-free network. The results of three challenging rotating machinery benchmark tests show that the proposed framework is not only more advantageous in terms of lightweight and robustness compared to state-of-the-art algorithms, but also proves that the nonlinear activation functions are nonessential for fault diagnosis tasks.