A bearing composite fault diagnosis method using multi-label deep learning

Abstract

The application discloses a bearing composite fault diagnosis method adopting multi-label deep learning, aiming at the problem that fault decoupling is difficult in composite fault diagnosis, and combining a multi-label classification thought, 01 coding mode is used for multi-label processing on training samples. A deep neural network is introduced as a feature extractor, a sigmoid function is used as a neural network output end, a multi-label deep learning composite fault diagnosis model is obtained, and verification is carried out on a bearing composite fault data set. The results show that the accuracy of the model can reach 90%. The proposed multi-label deep learning composite fault diagnosis model can effectively solve the decoupling problem of composite faults.

Description

Technical Field

[0001] This invention belongs to the field of composite fault diagnosis and aims to improve the accuracy of composite fault diagnosis. Background Technology

[0002] Deep neural networks excel in feature extraction and classification tasks thanks to their powerful fitting capabilities. Deep learning also holds great potential in handling multi-label problems, effectively addressing the correlation between different labels. In the field of fault diagnosis, this translates to a strong ability to decouple complex faults. However, a unified architecture for deep neural networks used in fault diagnosis remains lacking. Summary of the Invention

[0003] This invention addresses the challenge of decoupling faults in bearing composite fault diagnosis. It combines multi-label classification with 0-1 encoding for multi-label processing of training samples. A deep neural network is introduced as a feature extractor, and the sigmoid function is used as the neural network output to obtain a multi-label deep learning composite fault diagnosis model, which is then validated on a bearing composite fault dataset.

[0004] The technical solution of this invention is a bearing composite fault diagnosis method using multi-label deep learning, the method comprising:

[0005] Step 1: Vibration time-series signals of bearing rotation are collected using vibration sensors. First, the collected vibration time-series signals are labeled according to their different fault characteristics. Then, using the multi-label approach, the labeled data is processed by 0-1 encoding. 0-1 encoding can extend the values ​​of discrete features to Euclidean space, mapping discrete features to points in Euclidean space, which makes the similarity calculation or distance calculation between different features more reasonable. Finally, the entire dataset is divided into two parts: a training set and a test set, each containing vibration time-series signals and their corresponding labels.

[0006] Step 2: Fault diagnosis phase. First, a deep neural network model is constructed. The model's input and output, as well as the complete propagation process of the deep neural network, are as follows:

[0007] Input: Total number of layers L = 6; number of neurons in each hidden layer m = 512; number of neurons in the output layer 5 (for 5-complex fault diagnosis); loss function is Binary cross-entropy; activation function for both input and intermediate layers is ReLU; activation function for output layer is sigmoid; iteration step size α = 0.001, maximum number of iterations epochs = 30; stopping iteration threshold ε = 0.001; input sample X is input complex fault data.

[0008] Output: The linear relationship coefficient matrix W between each hidden layer and the output layer, and the bias vector B;

[0009] The propagation process is as follows:

[0010] S1: Initialize the linear relationship coefficient matrix W and bias vector B between each hidden layer and the output layer;

[0011] S2: Forward and backward propagation process;

[0012] S2-1: For iter = 1 to epochs, the number of iterations increases from 1 to the maximum number of iterations;

[0013] S2-1-1: Input A to the deep neural network 1 Set to X;

[0014] S2-1-2: For l = 2 to L, perform forward propagation algorithm A from the second layer to the last layer. l =W l X l +B l , where the superscript l indicates the number of layers in the deep neural network;

[0015] S2-1-3: Calculate the loss value using the loss function and take the derivative with respect to W and B;

[0016] S2-1-4: For l = L to 2, starting from the last layer and working backwards to the second layer, perform the backpropagation algorithm and calculate the step size. Where J represents the cross-entropy loss function:

[0017] J(W,B,A,Y)=-[Ylog(A L )+(1-Y)·log(1-A L )]

[0018] In the formula, W and B are the weights and biases; A is the N-dimensional output vector; and Y is the N-dimensional actual label vector.

[0019] This represents the derivative of the cross-entropy loss function with respect to W and B:

[0020]

[0021]

[0022] In the formula, The superscript T indicates an intermediate function; the superscript T indicates transpose.

[0023] This indicates that the iteration rate is a function of the gradient of the loss function, and includes the calculation of the first and second moments.

[0024] S2-2: For l = 2 to L, starting from the second layer and continuing to the last layer, update W in the l-th layer. l and B l In the formula, m represents the number of neurons in each layer:

[0025]

[0026]

[0027] S2-3: If all the changes in W and B are less than the corresponding threshold for stopping the iteration, then exit the iteration and output the linear relationship coefficient matrix W and the bias vector B between each hidden layer and the output layer.

[0028] Step 3: Import the training set into the fully connected layer of Step 2 for feature extraction. The dimension of the input layer of the deep neural network is the dimension of the fault data. Compared with traditional multi-label learning, DNN has good decoupling performance in extracting fault features. It can effectively distinguish different features under mutual influence in composite fault data. After continuous iteration, a well-trained model is obtained.

[0029] Step 4: Use the trained neural network to perform actual fault diagnosis on the bearing based on the bearing vibration signal.

[0030] The beneficial effects of this invention are as follows:

[0031] 1) In the data processing stage, this invention uses 01 encoding to generate multi-label data for different fault types, which is more conducive to feature extraction and classification compared with traditional neural network input.

[0032] 2) This invention uses deep neural networks for composite fault diagnosis, which effectively solves the problem of difficult decoupling of composite faults. Combined with the multi-label processing in 1), it can be broadly classified as a type of multi-label learning. Attached Figure Description

[0033] Figure 1 It is a composite fault diagnosis model based on multi-label deep learning;

[0034] Figure 2 This is the complete propagation process of a deep neural network;

[0035] Figure 3 This is a schematic diagram of the data experiment platform used;

[0036] Figure 4 This is a schematic diagram of bearing failure modes;

[0037] Figure 5 This is the result obtained by using publicly available datasets in this invention. Detailed Implementation

[0038] The method described in this invention is validated using a publicly available bearing dataset from Paderborn University, USA. This dataset contains four types of data: normal, single fault, recurring fault, and compound fault. The compound fault portion consists of real test data generated during accelerated life reliability testing. The accelerated life test bench comprises a motor, bearing housing protective mesh, bearing dock, spring assembly, adjusting nuts, and other structures. The motor provides power to the shafts of four 6203-type test bearings within the bearing dock. The bearings receive radial loads applied by a spring-screw mechanism, and vibration signals from the bearing rotational motion are collected using vibration sensors.

[0039] Step 1: For the composite fault dataset, process the data by assigning different labels based on its different fault characteristics. Then, using a multi-label approach, encode the labeled data using 0-1 encoding. Finally, divide the entire dataset into training and testing sets with a ratio of 0.8:0.2, each containing both data and labels.

[0040] Step 2: The input and output of the neural network model are:

[0041] Input: Total number of layers L = 6, number of neurons in each hidden layer is 512, number of neurons in the output layer is 5 (for 5-complex fault diagnosis), loss function is Binary cross-entropy, activation function of input and intermediate layers is ReLU function, activation function of output layer is sigmoid, iteration step size α = 0.001, maximum number of iterations epochs = 30, stopping iteration threshold ε = 0.001, input sample X is input complex fault data.

[0042] Output: The probability of each fault category output by the sigmoid function.

[0043] The deep neural network model used in this invention comprises six layers: an input layer, four intermediate layers (also called hidden layers), and an output layer. The input layer contains 500 neurons, and the intermediate layers contain 512 neurons. To improve training efficiency, the ReLU function is used for activation in both the input and intermediate layers. The output layer uses a sigmoid function to activate each node once, independently outputting a probability value between [0,1]. The model is iteratively trained during training, using a binary cross-entropy loss function and an Adam optimizer, with the goal of minimizing the cross-entropy between the output value and the label.

[0044] Step 3: Import the training set into the deep neural network model from Step 2 for feature extraction, iterate continuously, and save the model with the highest accuracy.

[0045] Step 4: Input the data from the test set into the best-performing model saved in Step 3 for testing, and provide the test accuracy.

[0046] This invention was validated using experimental data from the publicly available bearing dataset from Paderborn University, USA. This dataset includes four types of data: normal, single fault, recurring fault, and compound fault. The compound fault portion consists of real experimental data generated during accelerated life reliability testing. The accelerated life test bench is shown in... Figure 3 As shown, it consists of a motor, bearing housing, protective net, bearing dock, spring assembly, and measuring module. All structures are tightly fixed to the bearing housing and arranged along the motor's axis. The motor provides power to the shafts of four 6203-type test bearings within the bearing dock. Because it is an accelerated life test, the bearings are subjected to radial loads applied by a spring-screw mechanism, ranging from 0N to 3000N. Finally, vibration sensors are used to sample the vibration signals from the bearing's rotational motion.

[0047] As attached Figure 4 As shown, bearing failure locations are categorized into inner ring failures and outer ring failures. The main failure types are fatigue and plastic deformation. Fatigue primarily manifests as pitting, while plastic deformation manifests as indentations caused by particles. There are four failure types based on the combination of failure location and failure mode. Multi-label processing was performed on different failure data based on actual files from the publicly available dataset. Each composite failure type in the data used contains 500 samples, with 512 sampling points per sample.

[0048] Figure 5 The results obtained using the PU public dataset are presented, showing that the model achieves an accuracy of 90.125%.

Claims

1. A bearing composite fault diagnosis method employing multi-label deep learning, the method comprising: Step 1: Vibration time-series signals of bearing rotation are collected using vibration sensors. First, the collected vibration time-series signals are labeled according to their different fault characteristics. Then, using the multi-label approach, the labeled data is processed by 0-1 encoding. 0-1 encoding can extend the values ​​of discrete features to Euclidean space, mapping discrete features to points in Euclidean space, which makes the similarity calculation or distance calculation between different features more reasonable. Finally, the entire dataset is divided into two parts: a training set and a test set, each containing vibration time-series signals and their corresponding labels. Step 2: Fault diagnosis phase. First, a deep neural network model is constructed. The model's input and output, as well as the complete propagation process of the deep neural network, are as follows: Input: Total number of floors =6; the number of neurons in each hidden layer is The number of neurons in the output layer is 5; the loss function is Binary cross-entropy; the activation functions for both the input and intermediate layers are ReLU; the activation function for the output layer is sigmoid; the iteration step size is... =0.001, maximum number of iterations =30; Stop iteration threshold =0.001; Input sample To input composite fault data; Output: Linear relationship coefficient matrix between each hidden layer and the output layer and bias vector ; The propagation process is as follows: S1: Initialize the linear relationship coefficient matrix between each hidden layer and the output layer. and bias vector ; S2: Forward and backward propagation process; S2-1: For to The number of iterations increases from 1 to the maximum number of iterations; S2-1-1: Input the deep neural network Set as ; S2-1-2: For to The forward propagation algorithm is performed from the second layer to the last layer. , where superscript Indicates the number of layers in a deep neural network; S2-1-3: Calculate the loss value using the loss function and then... and Find the derivative; S2-1-4: For to Starting from the last layer and working backwards to the second layer, perform the backpropagation algorithm and calculate the step size. ,in, The cross-entropy loss function is represented as follows: ； In the formula, and For weights and biases; for 3D output vector; for Dimensional actual label vector; The cross-entropy loss function represents the... and Differentiate: ； ； In the formula, Indicates an intermediate function; The superscript indicates transpose; This indicates that the iteration rate is a function of the gradient of the loss function, including the calculation of the first and second moments; S2-2: For to From the second level to the last level, update the... Layer and In the formula This indicates the number of neurons in each layer: ； ； S2-3: If all and If the changes in all values ​​are less than the corresponding stopping threshold, the iteration stops, and the linear relationship coefficient matrix between each hidden layer and the output layer is output. and bias vector ; Step 3: Import the training set into the fully connected layer of Step 2 for feature extraction. The dimension of the input layer of the deep neural network is the dimension of the fault data. Compared with traditional multi-label learning, DNN has good decoupling performance in extracting fault features. It can effectively distinguish different features under mutual influence in composite fault data. After continuous iteration, a well-trained model is obtained. Step 4: Use the trained neural network to perform actual fault diagnosis on the bearing based on the bearing vibration signal.

Images