How to diagnose a wind turbine planetary gearbox failure? Is a diagnostic method based on enhanced convolutional neural networks feasible?
1 Preface
As a key component of wind turbines, planetary gearboxes often have various failures, in order to solve this problem, this paper proposes a new planetary gearbox intelligent fault diagnosis method.
2 IDCNN's intelligent diagnostic method
Convolutional neural network is one of the important algorithms in deep learning methods, and it is outstanding in pattern recognition applications due to its powerful feature extraction ability.
In this paper, the method of converting one-dimensional time domain signals into two-dimensional matrices is used to preprocess the test data and use it as the input of the CNN network.
Assuming that the size of the two-dimensional matrix to be generated is N×N, N2 sampling points are selected in the vibration time domain signal, and N2 sampling points are divided into N parts, the ordinate of each part is used as the data point, and the data of the data point is taken as a row in the two-dimensional matrix.
In this paper, 4096 sampling points are selected in the one-dimensional vibration signal, and each 64 sampling points are used as a row in the two-dimensional matrix to generate a two-dimensional matrix with a size of 64×64, which can make the sampled sample contain more feature information while adapting to the computer ability.
Due to the complex structure and harsh working conditions, the weakly coupled fault features in the original vibration signal are often overwhelmed by background noise or other interference, and it is difficult for CNNs to extract discriminative fault features directly from the original vibration signal.
In order to solve this problem, IDCNN is proposed for the fault diagnosis of planetary gearboxes with expanded receptive fields.
Compared with the traditional CNN, the IDCNN model enlarges the receptive field, enhances the effectiveness and robustness, and improves the feature learning ability of the CNN model.
The developed IDCNN model takes the preprocessed two-dimensional matrix as the input of the network, and extracts the fault characteristics through the stacks of the initial expansion convolutional layer, pooling layer and fully connected layer.
Then, all features are fed into the softmax classifier during the classification stage, the cross-entropy of the true probability and the estimated probability is calculated, the weights of the IDCNN are updated using gradient descent, and the optimal parameters are obtained through multiple backpropagations.
3. Experimental verification
In order to verify the superiority of the proposed algorithm, this paper collects data from the planetary gearbox on the wind turbine test bench built by a company.
In this experiment, the vibration fault signals of planetary gearboxes and bearings in the transmission system under eight different working conditions were collected, and nine fault data types were obtained.
There are two types of planetary gear faults, namely tooth root crack fault and tooth surface pitting failure, which are represented by PRC and PPS, respectively;
Planetary bearing faults are divided into bearing inner ring failures, bearing outer ring faults and rolling element faults on bearings, which are represented by IF, OF and BF respectively, and all faults are manufactured in the form of wire cutting and mechanical failure.
According to the wheel train schematic diagram of the planetary gearbox of the test bench, the planetary gearbox of the test bench is mainly composed of ring gear, planetary carrier, sun gear, planetary gear, planetary bearing and pin shaft.
The measurement point position of sensor number 11 is placed horizontally and vertically, and the acceleration sensor is used for the acquisition of vibration signals in the experiment, and the sampling frequency of the sensor is 16384Hz.
In the obtained vibration test data, each type of data is divided into 500 groups, each group of data length is 4096, a total of 4500 data samples, and the IDCNN method proposed in this paper is applied to train and test the vibration data.
In this paper, the hyperparameter selection details of the proposed network are as follows: IDCNN consists of one input layer, five initial expansion convolutional layers, five maximum pooling layers, one fully connected layer and one output layer.
To ensure that the model has converged, the number of iterations is set to 200, and the Adam optimizer is selected during the training process, and the learning rate is set to 0.002, which allows the model to converge quickly, and the batch_size is set to 100.
The number of convolutional layers plays an important role in the deep learning method, and the initial expansion convolutional layer of the proposed model is set to 5 layers, in order to provide more convincing and comprehensive results, 3-layer, 4-layer and 5-layer initial expansion convolutional layers are selected for comparison.
According to the comparison results, the average test accuracy is 98.89% when 3 initial expansion convolutional layers are used in the test, and the average test accuracy is increased to 98.91% compared with the 3-layer initial expansion convolutional layer when 4 initial expansion convolutional layers are used in the second scheme.
In another scheme, when 5 initial expansion convolutional layers are used, the average test accuracy is improved to 99.1%, because with the stacking of the initial expansion convolution, the disadvantage of losing a few features due to the convolution kernel of the expansion convolution has cavities is corrected and the network is deeper.
Therefore, it is more efficient to use an initial expansion convolution of 5 layers.
4 Conclusion
In order to enhance the ability of CNN to learn features, this paper proposes a research method for IDCNN planetary gearbox fault diagnosis, which increases the width and depth of the network by branching, which can better improve the network performance and avoid overfitting.
The effectiveness of the proposed algorithm is verified on the experimental vibration data of the planetary gearbox test bench, and the average test accuracy of the proposed IDCNN reaches 99.1% compared with CNN, OCRNN, MDCNN and ICNN 10 times, which is more effective for the fault diagnosis of planetary gearbox.
bibliography
[1] LI Yuheng, JIANG Zhanglei, LIANG Hao, et al. Research on fault diagnosis method of planetary gearbox based on HEI quantitative fault information[J].Mechanical and Electrical Engineering,2021,38:836-842.)
[2] HU Ruijie, PANG Xuebo, SHE Caiqing, et al. Fault diagnosis of planetary gearbox under variable working condition based on optimal window function Gabor transform[J].Fan Technology,2021,63(2):79-90.)
[3] LI Dongdong, LIU Yuhang, ZHAO Yang, et al. Fault diagnosis method of wind turbine planetary gearbox based on improved generative countermeasure network[J].Proceedings of the CSEE,2021,41(21):7496-7506.)
