Mechanical Fault Diagnosis Method Based on Multi-Source Information Fusion to Construct a Deep Tensor Projection Network

Abstract

The invention discloses a mechanical fault diagnosis method for constructing a deep tensor projection network by multi-source information fusion. The collected vibration signals are divided into multiple signal samples of equal length, and the time-frequency diagrams of different signal features of the same type of faults are feature-fused to construct a third-order tensor dataset of time × frequency × number of source information, which is divided into a training sample set and a test sample set. A deep tensor projection network fault diagnosis model is established, network parameters are set, the deep tensor projection network fault diagnosis model is trained by using the training sample set, and the model is verified by the test set to form a final diagnosis model. The invention can not only extract feature information more completely, but also effectively retain fault feature information and identify faults, and the model is more accurate, having broad application prospects in intelligent diagnosis of mechanical faults.

Description

Technical Field

[0001] The present invention relates to a mechanical fault diagnosis technology based on artificial intelligence, and particularly to a mechanical fault diagnosis method for constructing a deep tensor projection network by multi-source information fusion. Background Art

[0002] Rotating machinery usually operates in complex, harsh and unstable environments, and the collected signals have strong non-linearity and coupling. However, directly using this vibration signal will cause the model to be unable to effectively extract the complete fault feature information of the rotating machinery, thus affecting the model performance and even resulting in missed diagnosis or misjudgment of the model. Therefore, in order to avoid the occurrence of this problem, it is very necessary to use richer and more comprehensive fault information data as the input of the model to improve the performance of the fault diagnosis model and the recognition accuracy. Multi-source information fusion is a method that can integrate and analyze information from different angles and aspects, which not only enhances the comprehensiveness and accuracy of the information, but also improves the quality and reliability of the information.

[0003] At present, the fault diagnosis method based on multi-source information fusion can stably and widely extract fault features, but there are still some problems. After information fusion, the single-dimensional data information is expanded into multi-dimensional data information, and there are cross-correlation and redundant information in these data. Directly using them as the input of the model is likely to cause the curse of dimensionality, further leading to a decline in diagnostic performance and computational efficiency. Therefore, it is necessary to consider converting the high-dimensional data after multi-signal fusion into a new expression form and effectively reducing the dimension of these data to avoid destroying the data structure and causing the loss of feature information, so as not to affect the recognition accuracy of the model and the fault recognition precision.

[0004] The Tensor Projection Layer (TPL) is a variable-dimensional dimensionality reduction layer, and its implementation method is similar to Multiway Principal Component Analysis (MPCA for short). When reducing the dimension of the high-order tensor data structure, it will not expand the tensor into a vector, retains the original data structure, avoids destroying the internal information of the tensor, and retains the original feature information. Based on the advantages of the tensor projection layer, the tensor projection layer is used to replace the pooling layer in the deep convolutional neural network to build a deep tensor projection network.

[0005] Therefore, it is necessary to use multi-source information fusion to construct a fault diagnosis model of a deep tensor projection network and use the tensor projection layer to replace the pooling layer in the deep convolutional neural network. Summary of the Invention

[0006] To solve the problem that directly using multi-source information fusion of multi-dimensional data as the input of the model is likely to cause the curse of dimensionality and lead to a decline in diagnostic performance, the present invention proposes a mechanical fault diagnosis method for constructing a deep tensor projection network by multi-source information fusion. By using the method of multi-information fusion to construct a third-order tensor and combining it with a deep tensor projection network to build a mechanical intelligent fault diagnosis model, experimental verification is carried out to achieve high-precision and high-efficiency identification of the state of mechanical equipment.

[0007] The present invention adopts the following technical solutions to achieve the above object. A mechanical fault diagnosis method for constructing a deep tensor projection network by multi-source information fusion, the specific steps are as follows:

[0008] 1) Signal acquisition and sample division: Divide the collected vibration signals into multiple equally long signal samples;

[0009] 2) Construct a tensor data set by multi-source information fusion: Use synchronous extraction transformation for the divided signal samples to obtain the time-frequency diagrams of each sample, then convert the time-frequency diagrams into grayscale images, and then perform feature fusion on the time-frequency diagrams of different signal features of the same type of fault. Finally, construct a third-order tensor data set of time × frequency × source information number;

[0010] 3) Divide the constructed tensor data set into a training sample set and a test sample set according to a ratio of 8:2;

[0011] 4) Establish a deep tensor projection network fault diagnosis model and set network parameters. The process is as follows:

[0012] (1) Replace the pooling layer in the deep convolutional neural network with a tensor projection layer, and alternately stack the convolutional layer and the tensor projection layer to build a deep tensor projection network;

[0013] (2) Set the number of iterations N, the size (m, n) of the convolutional kernel in the convolutional layer, the stride, and the output tensor dimension of the tensor projection layer. Put the training sample set in the data set into the deep tensor projection network fault diagnosis model for training. Finally, test the model effect through the test set;

[0014] (3) In the deep tensor projection network, the input n tensors After passing through the convolutional layer, the output is of dimension p1×p2×p3 Then the output After passing through the tensor projection layer, the output is of dimension q1×q2×q3 And q ≤ p, that is:

[0015]

[0016] In the formula: × k, where \(k = 1, 2, 3\) represents the \(k\)-mode product; \(U_1, U_2, U_3\) are orthogonal matrices of size \(p\) k × \(q\) k and there is

[0017] In the process of finding the optimal parameters of the network model, it is necessary to obtain the minimum loss function by taking the partial derivatives of the loss function \(L\) with respect to the weights \(w\), that is:

[0018]

[0019] By continuously updating the loss function, the optimal parameters of the model are obtained;

[0020] 5) Use the training sample set described in step 3) to train the deep tensor projection network fault diagnosis model constructed in step 4), and verify the model with the test set to form the final diagnosis model.

[0021] The present invention constructs a mechanical intelligent fault diagnosis model by combining multi-source information fusion and a deep tensor projection network. By synchronously extracting the time-frequency features of the original signal through transformation, and then using the multi-source information fusion technology to fuse different fault features to construct a third-order tensor. Finally, the deep tensor projection network is used to reduce the dimension of the high-order tensor samples and extract features to achieve fault classification. The proposed method can not only extract feature information more completely, but also effectively retain fault feature information and identify faults. The model is more accurate and has broad prospects in mechanical fault diagnosis. It can overcome the deficiency of the dimensionality disaster caused by the multi-dimensional data formed by multi-source information fusion and can effectively and comprehensively extract fault feature information, and has broad application prospects in mechanical fault intelligent diagnosis. Description of the Drawings

[0022] Figure 1 is the model flowchart of the present invention;

[0023] In the figure: 101. Signal sampling and sample division, 102. Sample Ⅰ, 103. Sample Ⅱ, 104. Sample n, 105. Synchronous extraction transformation and multi-source information feature fusion, 106. Fault diagnosis, 107. Deep tensor projection network;

[0024] Figure 2 is the flowchart of the construction of the third-order tensor in the present invention;

[0025] In the figure: 201. Source signal Ⅰ, 202. Source signal Ⅱ, 203. Synchronous extraction transformation, 204. Time-frequency feature fusion, 205. Third-order tensor X;

[0026] Figure 3 is the deep convolutional neural network structure;

[0027] In the figure: 301. Input layer, 302. Convolution layer I, 303. Pooling layer I, 304. Convolution layer II, 305. Pooling layer II, 306. Fully connected layer, 307. Output layer, 308. Classifier layer;

[0028] Figure 4 is the deep tensor network diagram in the present invention;

[0029] In the figure: 401. Input layer, 402. First convolution layer, 403. First tensor projection layer, 404. Second convolution layer, 405. Second tensor projection layer, 406. Flattening layer, 407. First fully connected layer, 408. Second fully connected layer, 409. Output layer;

[0030] Figure 5a is the accuracy curve diagram obtained by the present invention;

[0031] In the figure: 1. Training set accuracy curve; 2. Test set accuracy curve;

[0032] Figure 5b is the loss function curve diagram of the multi-source information fusion-based deep tensor projection network fault diagnosis model obtained by the present invention;

[0033] In the figure: 1. Training set loss function curve; 2. Test set loss function curve;

[0034] Figure 6a is the fault feature map extracted by the first convolution layer in the present invention;

[0035] Figure 6b is the fault feature map extracted by the first tensor projection layer in the present invention;

[0036] Figure 6c is the fault feature map extracted by the second convolution layer in the present invention;

[0037] Figure 6d is the fault feature map extracted by the second tensor projection layer in the present invention. Detailed implementation manners

[0038] The present invention will be further described below in conjunction with the accompanying drawings and embodiments. See Figures 1 to 6dSince the multi-source information fusion method can stably and more extensively extract fault features, it is widely used in fault diagnosis. However, there are still some problems. After information fusion, the single-dimensional data information is expanded into multi-dimensional data information, and there are cross-correlation and redundant information in these data. Directly using them as the input of the model is likely to cause the curse of dimensionality, further leading to a decline in diagnostic performance and computational efficiency. Therefore, it is necessary to consider converting the high-dimensional data after multi-signal fusion into a new expression form and effectively reducing the dimension of these data to avoid destroying the data structure and causing the loss of feature information, so as not to affect the recognition accuracy of the model and the fault recognition accuracy. Therefore, consider converting the data of multi-source information fusion into a high-order tensor, and then using the constructed deep tensor projection network to effectively reduce the dimension of the high-order tensor and extract features to achieve fault classification.

[0039] 1. Mechanical fault diagnosis model for constructing a deep tensor projection network by multi-source information fusion:

[0040] On this basis, a mechanical fault diagnosis model for constructing a deep tensor projection network by multi-source information fusion is built, and its process is as Figure 1 shown.

[0041] 1) First, collect vibration signals, and then divide the collected vibration signals into multiple equally long signal samples 101.

[0042] 2) Tensor acquisition based on synchroextracting transform (SET) - multi-source information fusion:

[0043] The original data information divided into samples (sample Ⅰ 102, sample Ⅱ 103... sample n 104) is transformed into a time-frequency feature map through synchroextracting transform, and then these feature maps are fused 105 and constructed into a high-order tensor for use in the fault diagnosis 106 model. The tensor construction process is as Figure 2 shown. Two source signals (source signal Ⅰ 201 and source signal Ⅱ 202.) are transformed 203 into time-frequency feature maps through synchroextracting transform, and then the obtained time-frequency feature maps are fused 204, and finally a third-order tensor X I ×J×K is obtained, where I represents time, J represents frequency, and K represents the number of source information 205.

[0044] 3) Divide the constructed tensor sample data set into training samples and test samples according to a ratio of 8:2. The training samples are used to train the fault diagnosis model, and the test samples are used to test the performance of the model.

[0045] 4) Establish a deep tensor projection network fault diagnosis model and set network parameters. The process is as follows:

[0046] (1) Replace the pooling layer Ⅰ 303 and pooling layer Ⅱ 305 in the deep convolutional neural network with a tensor projection layer, and alternately stack the convolutional layer and the tensor projection layer to build a deep tensor projection network, as Figure 3 shown. Its structure is "input layer 401 - first convolutional layer 402 - first tensor projection layer 403 - second convolutional layer 404 - second tensor projection layer 405 - flattening layer 406 - first fully connected layer 407 - second fully connected layer 408 - output layer 409". Figure 4

[0047] (2) Set the number of iterations N of the network model to 150 times and the learning rate to 0.001, and input the data set into the network for training and testing, as Figure 5a shown, to form the training set accuracy curve 1 and the test set accuracy curve 2.

[0048] (3) Forward propagation and backward propagation are required in the deep tensor projection network. The specific process is as follows:

[0049] Forward propagation: Assume that the input of the deep tensor projection network is n tensors After passing through the convolutional layer, the output size is a tensor with dimensions p1×p2×p3 As shown in Equation (3):

[0050]

[0051] Among them: w l represents the weight, b l represents the bias, i represents the number of tensors, l represents the network layer (l = 1, 2,..., l0), and conv is the convolution operation function. The output of the convolutional layer passes through the tensor projection layer and the output size is a tensor with dimensions q1×q2×q3 And q ≤ p, as shown in Equation (4) namely:

[0052]

[0053] The neural network trains the parameters through gradient descent. However, if the algorithm is simply used to train and update the parameters in Equation (4) here, then will not satisfy orthogonality. Therefore, introduce a matrix k with dimensions p k × q U k and the relationship between S k is as shown in Equation (5):

[0054]

[0055] So, the training parameters in the tensor projection layer become​ However, in Equation (5) is not necessarily full rank (or invertible). For this, make some small changes to Equation (5), such as Equation (6):

[0056]

[0057] where: is a small fixed positive constant.

[0058] Then, the output after the tensor projection layer is Enter the flattening layer to obtain the output F i , as shown in Equation (7), that is:

[0059] F i = Flatten(Z i ) (7)

[0060] where Flatten represents the flattening function.

[0061] Next, the output F i obtained by the flattening layer then enters the fully connected layer to obtain FC i , as shown in Equation (8), that is:

[0062] FC i = f(wF i + b) (8)

[0063] where: f(.) is the function of the connection layer, w is the weight parameter, and b is the bias.

[0064] Finally, FC i enters the softmax classifier to predict the model output, as shown in Equation (9), that is:

[0065]

[0066] In summary, the forward propagation process is completed.

[0067] Backpropagation: After the above forward propagation is completed, the model parameters need to be adjusted through backpropagation. The cross-entropy loss function loss is used in the backpropagation of the model, as shown in Equation (10):

[0068]

[0069] where: y is the output, is the expected output.

[0070] Let all the weight parameters of the model be w, then the expression of the total loss function L(w, b) of the model is as shown in Equation (11):

[0071]

[0072] To achieve the optimal accuracy of model recognition and find the optimal parameters, as shown in Equation (12):

[0073]

[0074] where: η is the learning rate, and then it is necessary to calculate to find the minimum loss function L, then The calculation formula of is shown in Equation (13):

[0075]

[0076] 5) Use the training set in step 3) to train the model constructed in step 4), and verify the model through the test set to form the final diagnosis model.

[0077] 2. Experimental verification:

[0078] To verify the effectiveness of the mechanical fault diagnosis model based on multi-source information fusion to construct a deep tensor projection network, the present invention uses the experimental data collected by the planetary gearbox test bench of Xi'an Jiaotong University, with a sampling frequency of 12,800 Hz. A total of four types of gearbox states are collected, namely normal state, first-stage sun gear spalling fault, first-stage planetary gear spalling fault, and first-stage planetary gear root crack fault. In this experiment, the data information collected in the horizontal direction from two different position sensors is applied. Each working condition has 300 samples, and there are a total of 1,200 samples. The dimension of each sample is 256×512×2. At the same time, the training set and the test set are divided according to the ratio in step 1). Then the sample set is input into the 107 network model for training and testing. The parameters of the model are shown in Table 1. After passing through the first convolutional layer, the output dimension is 256×512×16. After passing through the first tensor projection layer, the output dimension is 32×64×16. After passing through the second convolutional layer, the output dimension is 32×64×32. After passing through the second tensor projection layer, the output dimension is 8×16×32. Then, through the flattening layer, the tensor is flattened, which is a transition to the fully connected layer. For better classification, all features are flattened through the first fully connected layer; the second fully connected layer is used for a transition between the first fully connected layer and the output; the output layer uses softmax as the multi-classifier function and outputs in a probabilistic manner.

[0079]

[0080] The results obtained after the model is trained and tested are as Figures 5a to 5b shown, obtaining Figure 5b the training set loss function curve 1 and the test set loss function curve 2 in Figure 5aIn it, the two curves of the model (the training set change curve 1a and the test set change curve 2a) both quickly reach an accuracy rate of 100%. At the same time, in Figure 5b In it, the two loss function curves (the training set loss function curve 1b and the test set loss function curve 2b) are continuously decreasing and gradually converging, and the final loss value decreases to 0. During this process, there is no overfitting phenomenon. It can be seen that fusing the information data collected by different sensors enriches the fault feature information, and the deep tensor projection network effectively reduces the dimension and extracts features from high-order tensors. Therefore, the model can identify various fault types more effectively and quickly.

[0081] To further demonstrate the effectiveness of the model, t-distributed stochastic neighbor embedding (t-SNE for short) is used to show the feature extraction of the model for each gear state. Therefore, t-SNE is used to visualize the feature distribution of each convolutional layer and tensor projection layer of the proposed model, and its feature scatter distribution is as Figures 6a to 6d shown. Figure 6a is the fault feature map extracted by the first convolutional layer in the present invention; Figure 6b is the fault feature map extracted by the first tensor projection layer in the present invention; Figure 6c is the fault feature map extracted by the second convolutional layer in the present invention; Figure 6d is the fault feature map extracted by the second tensor projection layer in the present invention. Figures 6a to 6d In it: ● represents the spalling fault of the planet gear, ▲ represents the crack fault of the planet gear, ★ represents the spalling fault of the sun gear, and ⅹ represents normal. From Figure 6a It can be seen that after synchronous extraction transformation and fusing the time-frequency feature maps of the two sensors, the fault information is complementary, making the fault features more perfect. Therefore, after the first convolutional layer, the four states of the planetary gearbox have been separated from each other. Looking at Figure 6b and Figure 6c , although the various fault states have been separated after the first convolutional layer, the various feature scatter points are relatively scattered. After the first tensor projection layer and the second convolutional layer, the various feature scatter points begin to gather, and finally after the second tensor projection layer, it presents the Figure 6d distribution situation, where the various feature scatter points are separated and gathered respectively. Therefore, it can be seen that the fault feature information is extracted more accurately after synchronous extraction transformation. At the same time, fusing multi-source information makes the fault feature information more perfect, thereby improving the performance of the fault diagnosis model.

[0082] In the present invention, a multi-source information fusion is used to construct a third-order tensor sample set, which is divided into a training set and a test set. A deep tensor projection network (Deep Tensor Projection Networks, DTPN) fault diagnosis model is established by replacing the pooling layer of the convolutional neural network with a tensor projection layer. The network model parameters are set, the model is trained using the training set, and the model is verified using the test set to obtain the final fault diagnosis model. A third-order tensor of time × frequency × number of source information is constructed through the feature fusion of multi-source information, and a mechanical fault diagnosis model is established using the deep tensor projection network. This not only overcomes the deficiency of the dimensionality disaster caused by directly using the multi-dimensional data fused from multi-source information, but also can effectively and comprehensively extract fault feature information. At the same time, it avoids the loss of feature information caused during the process of reducing the data dimension. The performance of the final model is superior to that of the traditional deep convolutional neural network.

[0083] From the above results, it can be seen that the third-order tensor constructed by the present invention through time, frequency, and the number of source information can effectively represent the data fused from multi-source information. It not only avoids the problem of affecting the model performance by directly using the high-dimensional data fused from multi-source data, but also realizes the effective dimensionality reduction and feature extraction of the high-order tensor data formed after multi-source information fusion through the constructed deep tensor projection network (DTPN). The experimental results of the multi-source information feature fusion fault diagnosis of the planetary gearbox and the bearing show that the proposed method can not only effectively retain the fault feature information and identify the faults, demonstrating the unique advantages of the present invention and having broad application prospects in the intelligent diagnosis of mechanical faults.

Claims

1. A mechanical fault diagnosis method for constructing a depth tensor projection network by multi-source information fusion, characterized in that, The specific steps are as follows: 1) Signal acquisition and sample division: Divide the collected vibration signals into multiple signal samples of equal length; 2) Multi-source information fusion to construct a tensor dataset: Use synchronous extraction transformation on the divided signal samples to obtain the time-frequency diagrams of each sample, then convert the time-frequency diagrams into grayscale images, and then perform feature fusion on the time-frequency diagrams of different signal features of the same type of fault. Finally, construct a third-order tensor dataset of time × frequency × number of source information; 3) Divide the constructed tensor dataset into a training sample set and a test sample set according to a ratio of 8:2; 4) Establish a deep tensor projection network fault diagnosis model and set network parameters. The process is as follows: (1) Replace the pooling layer in the deep convolutional neural network with a tensor projection layer, and alternately stack the convolutional layer and the tensor projection layer to build a deep tensor projection network; (2) Set the number of iterations N, the size (m, n) of the convolutional kernels in the convolutional layer, the stride, and the output tensor dimension of the tensor projection layer. Put the training sample set in the dataset into the deep tensor projection network fault diagnosis model for training. Finally, test the model effect through the test set; (3) In the deep tensor projection network, n input tensors l = 1, 2, 3, …, n; after passing through the convolutional layer, the output has a dimension of p1×p2×p3 Then the output After passing through the tensor projection layer, the output has a dimension of q1×q2×q3 And q ≤ p, that is: Where: × k , k = 1, 2, 3 represents the k - modulus product; U1, U2, U3 are orthogonal matrices of size p k ×q k , and there are i = 1, 2, 3; In the process of finding the optimal parameters of the network model, it is necessary to obtain the minimum loss function by taking the partial derivatives of the loss function L and the weights w, that is: In the formula, vec(·) represents vector operation. By continuously updating the loss function, the optimal parameters of the model are obtained; 5) Use the training sample set in step 3) to train the deep tensor projection network fault diagnosis model constructed in step 4), and verify the model by the test set to form a final diagnosis model.

Images