The work of this paper belongs to the application of deep learning in industry , Learn from the solution of computer vision , For the scene of machine fault detection, an adaptive APReLU, The accuracy of fault detection has been greatly improved . The whole idea of the paper should also be applied to computer vision , The code is also open source , You can try it
source ： Xiaofei's algorithm Engineering Notes official account
The paper : Deep Residual Networks with Adaptively Parametric Rectifier Linear Units for Fault Diagnosis
- Address of thesis ：https://ieeexplore.ieee.org/document/8998530/metrics#metrics
- Code address ：https://github.com/zhao62/Adaptively-Parametric-ReLU
The scenario discussed in this paper is the error detection of electronic devices , Due to long-term operation in harsh environments , Electronic equipment often inevitably fails , And then cause accidents and losses . And the vibration signal (vibration signal) It usually includes pulses and fluctuations due to machine failure , Can be used to detect equipment failure . In the near future , Deep learning is also used in error detection of electronic devices , Take the vibration as the input signal , Output whether the current device is normal .
Mainstream classification neural networks use an identical set of nonlinear transformations to deal with different inputs , Pictured a Shown ,F、G and H Represents a nonlinear change ,$=$ Represents whether the nonlinear transformation is the same . For vibration signal scenarios , Machines in the same state of health , Because the current operation is different , The difference of the feedback vibration signals may be large , It is difficult to classify different waveforms into the same health state . Contrary , Machines in different health states occasionally produce the same vibration signal , Neural networks map them to similar regions , It's hard to distinguish . Sum up , The fixed nonlinear transformation may have a negative impact on the feature learning ability in the vibration signal scene , It is very meaningful to be able to learn automatically and use different nonlinear transformations according to the input signal .
This paper is based on ResNet An improved version of ResNet-APReLU, Pictured b Shown , Different nonlinear transformations are assigned according to the input signal , Specifically by inserting a similar SE(squeeze-and-excitation) The module's subnet to adjust the slope of the activation function , It can greatly improve the accuracy of fault detection . Because of the special scene of the thesis , So I mainly study the methods proposed in the paper , As for the application scenario related part and experimental part , Just take it easy .
Fundamentals of classical ResNets
The paper ResNet Based on ,ResNet Its core structure is shown in the figure 2a Shown , I'm sure you all know , No more introduction . take ResNet Applied to machine error recognition , Pictured 2b Shown , Input vibration signal , After the feature extraction of the network, the state recognition is carried out , Determine whether the machine is healthy or in some other wrong state . The core of the paper is through improvement ReLU Adaptive nonlinear transformation , original edition ReLU It can be formulated as ：
Design of the developed ResNet-APReLU
Design of the fundamental architecture for APReLU
APReLU Integrated with a specially designed subnet , It's kind of like SE modular , The multiplicative factor used for nonlinear transformation is predicted adaptively according to the input , Structure is shown in figure 3a Shown , Output channel-wise Of ReLU Parameters , The following steps are included ：
- use ReLU and GAP Mapping input features to 1D vector , Get positive features (positive feature) The overall information of . use min(x, 0) and GAP Map the input feature to another 1D vector , Get negative features (negative feature) The overall information of , Negative information may contain some useful fault information .GAP It can deal with the problem of signal offset , The input feature map information is compressed into two 1D vector , They represent positive and negative information respectively .
- Put two 1D vector Concate together , Conduct FC-BN-ReLU-FC-BN-Sigmoid Calculation , Two FC The dimension of output and input characteristics is consistent , Last sigmoid The output is used for the formula 10 Of $\\alpha \\in (0, 1)$ factor ：
Architecture of the developed ResNet-APReLU for vibration-based gearbox fault diagnosis
be based on APEeLU Building new ResBlock, Pictured b Shown , With the original ResBlock Almost the same , Just to ReLU Replace with APReLU Adaptive nonlinear activation .APReLU The output size is the same as the input size , It can be simply embedded in a variety of networks . The complete network structure is shown in the figure c Shown , Finally, output the prediction of multiple machine states , Calculate the cross entropy loss , Do gradient descent learning .
From the results , For machine failure scenarios , The method proposed in this paper is very effective .
The work of this paper belongs to the application of deep learning in industry , Learn from the solution of computer vision , For the scene of machine fault detection, an adaptive APReLU, The accuracy of fault detection has been greatly improved . The whole idea of the paper should also be applied to computer vision , The code is also open source , You can try it .
If this article helps you , Please give me a compliment or watch it ～
More on this WeChat official account 【 Xiaofei's algorithm Engineering Notes 】