A self-adaptive DRSN-GPReLU for bearing fault diagnosis under variable working conditions

Abstract

Recently, deep learning has become an effective fault diagnosis method due to its no-mankind feature extraction capability. However, in real applications, rolling bearings are often used in the strong noise and variable working conditions, which lead to the degradation of fault diagnosis ability of neural network model. In order to solve the problem, a self-adaptive deep residual shrinkage network with global parametric rectifier linear unit (DRSN-GPReLU) is presented in this article for intelligent fault diagnosis under variable working conditions, which adopts the DRSN as the basic network architecture. We develop a novel activation function named global parametric rectifier linear unit (GPReLU), which can achieve better intra-class compactness for vibration signals, and the inter-class samples are better mapped into remote areas. Furthermore, a novel network based on the attention mechanism is designed to automatically infer the slope of the GPReLU. Various experimental results demonstrate that the DRSN-GPReLU can realize better performance compared with the traditional methods under variable working conditions, and has better robustness under noise interferences.

Deep residual shrinkage networks with adaptively convex global parametric rectifier linear units for fault diagnosis

“…(3) We propose a novel evaluation metric designed to measure the collective influence of diagnostic accuracy and computational complexity. In addition, it is experimentally demonstrated that the developed DRSN-CGPReLU can maintain a better diagnostic performance while greatly improving the diagnostic efficiency under variable operating conditions compared to our previous work [34].…”

“…This innovative sub-network facilitates the autonomous creation of a suite of nonlinear transformations that take into account global features across diverse input signals. A visual representation of the sub-network's overall structure is provided in figure 1, while comprehensive details regarding its specific design are expounded in [34]. Nonetheless, in contrast to conventional ReLU and APReLU, the incorporation of an attention sub-network within GPReLU amplifies its operational efficacy, albeit at the expense of escalated computational intricacy.…”

“…Given that PReLUs exclusively attend to the nonlinear transformation of the negative feature domain, there exists the potential for an adverse impact on the neural network's discernment capacity in relation to patterns. Consequently, a novel activation function, denoted as GPReLUs, has been introduced in our earlier research [34]. In contrast to PReLUs, this function adds an additional slope parameter within the positive feature region, implementing nonlinear transformations of global region features.…”

“…This bias could potentially undermine the efficacy of feature extraction. In response to this limitation, we introduced the global PReLU(GPReLU) in a prior study [34]. GPReLU excels in facilitating the adaptive nonlinear transformation of global features across diverse signals, yielding superior diagnostic outcomes under operational conditions.…”

SO-IMCKD processed signal improving MSCNN model’s fault diagnosis accuracy for drilling pump fluid end

“…Song and Jiang [18] transformed the time series signals in the chemical process into matrix graphs and input them into multi-scale convolutional neural networks (MCNNs), realizing various fault diagnosis in the chemical production process, with a recognition accuracy of 88.54%. Some scholars have also improved CNN to improve the diagnostic effect of bearings [19,20]. Research has also found that CNN can effectively extract images generated by continuous wavelet transform (CWT) [21][22][23][24], and realize the diagnosis of different conditions in different fields.…”