The work of this paper belongs to the application of deep learning in the industrial field. Referring to the solution of computer vision, the adaptive aprelu is proposed for the scene of machine fault detection, which greatly improves the accuracy of fault detection. The overall idea of the paper should also be applied to computer vision, the code is also open source, we can try
Source: Xiaofei’s algorithm Engineering Notes official account
Paper: deep residual networks with adaptive parametric rectifier linear units for fault diagnosis
- Thesis address:https://ieeexplore.ieee.org/document/8998530/metrics#metrics
- Code address:https://github.com/zhao62/Adaptively-Parametric-ReLU
The scene discussed in this paper is the error detection of electronic equipment. Due to long-term operation in harsh environment, electronic equipment will inevitably fail, which will cause accidents and losses. Vibration signal usually contains pulse and fluctuation caused by machine fault, which can be used to detect equipment failure. Recently, deep learning method has also been used in the error detection of electronic equipment. The vibration signal is taken as the input to output whether the current equipment is normal.
The mainstream classification neural networks use a set of identical nonlinear transformations to deal with different inputs. As shown in figure a, F, G and H represent nonlinear changes, and $= $represents whether the nonlinear transformations are the same. For the vibration signal scene, due to the different current operation of the machine in the same health state, the vibration signal feedback may vary greatly, so it is difficult to classify different waveforms into the same health state. On the contrary, machines with different health status will occasionally produce the same vibration signal, which will be mapped to similar regions by neural network, which is difficult to distinguish. In conclusion, the fixed nonlinear transformation may have a negative impact on the feature learning ability in the vibration signal scene, so it is very meaningful to be able to learn automatically and use different nonlinear transformation according to the input signal.
In this paper, an improved version of RESNET aprelu is proposed based on RESNET. As shown in Figure B, different nonlinear transformations are given according to the input signal. Specifically, the slope of activation function is adjusted by inserting a subnet similar to se (squeeze and excitation) module, which can greatly improve the accuracy of fault detection. Because the scene of the paper is special, so we mainly study the methods proposed in the paper. As for the application scenario related part and the experimental part, it is good to simply take it.
Fundamentals of classical ResNets
This paper is based on RESNET. The core structure of RESNET is shown in Figure 2A. I believe you are very clear, so we will not introduce it any more. RESNET is applied to the machine error recognition, as shown in Fig. 2B. The vibration signal is input and the state recognition is carried out after the feature extraction of the network to determine whether the machine is in health or in other error states. The core of this paper is adaptive nonlinear transformation by improving relu
Design of the developed ResNet-APReLU
Design of the fundamental architecture for APReLU
Aprelu integrates a specially designed subnet, which is somewhat similar to the se module. It adaptively predicts the multiplication factor for nonlinear transformation according to the input. The structure is shown in Fig. 3a. The output of relu parameters of channel wise includes the following steps:
- The global information of positive features is obtained by mapping input features to 1D vectors with relu and gap. Min (x, 0) and gap are used to map the input features to another 1D vector to obtain the global information of negative features, which may contain some useful fault information. Gap can deal with the problem of signal offset and compress the input feature map information into two 1D vectors, representing positive and negative information respectively.
- Two 1D vectors are concatenated together to calculate fc-bn-relu-fc-bn-sigmoid. The output of the two FC is consistent with the dimension of the input feature. Finally, the sigmoid output is used for the $\ \ alpha \ \ in (0,1) $factor of formula 10
Architecture of the developed ResNet-APReLU for vibration-based gearbox fault diagnosis
A new resblock is constructed based on apelu, as shown in Figure B, which is basically the same as the original resblock, except that relu is replaced by aprelu for adaptive nonlinear activation. The output size of aprelu is the same as the input size, and can be easily embedded into various networks. The complete network structure is shown in Figure C. finally, the prediction of multiple machine states is output, the cross entropy loss is calculated, and the gradient descent learning is carried out.
The results show that the proposed method is very effective for the scenario of machine failure.
The work of this paper belongs to the application of deep learning in the industrial field. Referring to the solution of computer vision, the adaptive aprelu is proposed for the scene of machine fault detection, which greatly improves the accuracy of fault detection. The overall idea of the paper should also be applied to computer vision, the code is also open source, we can try.
If this article is helpful to you, please give me a like or read it
More content, please pay attention to WeChat official account.