Gearbox credible fault intelligent diagnosis method based on multi-channel information fusion

Abstract

The invention belongs to the technical field of rotating machinery fault diagnosis, and relates to a gearbox credible fault intelligent diagnosis method based on multi-channel information fusion, which comprises the steps of preprocessing multi-channel information, establishing a corrected graph convolutional network and constructing a credible multi-channel information fusion framework. According to the method, Fourier transform is adopted to obtain features related to fault information, and a k-nearest neighbor algorithm is adopted to realize graph data construction of multi-channel information; a batch normalization layer and a Leakey ReLU layer are added between graph convolution layers, a full connection layer is added behind the two graph convolution network layers, a Softplus layer is used for replacing a Softmax layer, and a corrected graph convolution network frame is obtained; according to the method, subjective logic is introduced, the uncertainty of information of each channel is calculated, diagnosis results of multi-modal information are fused through a reduced evidence fusion theory, and whether the diagnosis results are credible or not is verified by combining the fused uncertainty, so that a gearbox credible fault diagnosis result is obtained.

Description

Technical Field

[0001] The present invention belongs to the technical field of rotating machinery fault diagnosis, and in particular relates to an intelligent diagnosis method for credible gearbox faults based on multi-channel information fusion. Background Art

[0002] The gear transmission system is the core of various heavy equipment and has been widely used in heavy-duty and high-power mechanical equipment such as automobile gearboxes, axles, wind turbines, coal mining machines, and trains. However, such equipment often works under harsh conditions, which can easily lead to failure of key components of the mechanical system (gears, bearings, etc.). Therefore, real-time online monitoring of the health status of the gear transmission system and accurate diagnosis and identification of faults are of great significance to ensure the normal operation of mechanical equipment and equipment fault diagnosis.

[0003] With the continuous development of sensor and network technology, data-driven fault diagnosis methods have been widely used in rotating machinery and equipment. As an advanced machine learning method, deep learning provides a new solution for feature mining and fault state identification of rotating machinery and equipment. In recent years, with the rapid development of graph representation learning and graph machine learning, graph neural network models have brought revolutionary breakthroughs in the field of fault diagnosis of rotating machinery and equipment. In graph data, the relationships between nodes are represented by edges, which provides the possibility of mining potential relationships between samples. The process of graph convolution operation can be understood as the central node aggregating the information of neighboring nodes to form a new node, and the new node features formed participate in the subsequent label prediction task. However, due to the particularity of the graph data structure, how to use the existing multi-channel signals to construct input suitable for graph neural networks is an urgent problem to be solved.

[0004] At present, the graph data constructed by existing research institutes only rely on single vibration sensor data, and the graph deep learning model built cannot process multi-channel graph data at the same time. Although single-channel vibration signals can reflect some key characteristic information of rotating mechanical equipment, fault diagnosis research based on vibration signals also has certain defects: (a) Due to the complexity of the planetary gearbox structure, it is necessary to consider the influence of the transmission path on it; (b) Mechanical equipment is easily affected by external excitation, resulting in serious noise interference in the collected vibration signal. Therefore, using a single-channel vibration sensor to identify equipment faults is prone to missing key characteristic information. In addition, for the results of fault classification that integrates multi-channel information, the common problems are "Is this classification reliable?" and "Why is the classification reliability too high or too low?" Based on this, quantifying the uncertainty of each channel information is particularly important for enhancing the reliability of the classification results.

[0005] Patent CN 115329906 A discloses a fault diagnosis method and device based on a graph regularization network. The trained fault diagnosis model based on the graph regularization network is used to perform fault diagnosis on the weakly supervised test sample set, and the fault patterns of the weakly supervised test sample set are identified. The associated graph construction process and the model construction only consider single-channel signals, which are prone to missing key feature information.

[0006] Patent CN 109540520 A discloses a rolling bearing fault fusion diagnosis method based on improved D-S evidence theory. First, various fault diagnosis methods are used to classify and identify the fault state data to obtain various primary diagnosis results. Then, the correlation matrix between each evidence is constructed using the conflict factor in the DS evidence theory, and the reliability of each evidence is calculated. Finally, the DS combination rule is used to fuse the similar evidence and the modified conflict evidence to obtain the final result. This method fuses the diagnosis results of multiple fault diagnosis methods of a single signal source, and it remains to be considered whether the diagnosis results of multi-channel signals can be fused, and the reliability of each diagnosis result is also a potential problem.

[0007] There is no relevant patent disclosure for the intelligent diagnosis of credible faults that can simultaneously fuse multi-channel signals, consider the uncertainty of the diagnosis results of each channel signal, and form an end-to-end intelligent diagnosis framework. Summary of the Invention

[0008] In view of the above problem analysis, the purpose of the present invention is to provide a credible fault intelligent diagnosis method for a gearbox based on multi-channel information fusion, including multi-channel information preprocessing, establishing a modified graph convolutional network, and constructing a credible multi-channel information fusion framework. This method aims at problems such as the weak early fault characteristics of the mechatronic coupling system under strong background noise being difficult to extract, the low reliability of fault classification results, and the single-channel vibration signal being difficult to accurately reflect some key feature information of rotating mechanical equipment.

[0009] The technical solution adopted by the present invention to solve its technical problems is:

[0010] A credible fault intelligent diagnosis method for a gearbox based on multi-channel information fusion, the specific steps are as follows:

[0011] Step 1: Install multiple acceleration sensors on the surface of the planetary gearbox housing in a magnetic adsorption manner, and collect multi-channel vibration signals of the planetary gearbox under different fault states respectively.

[0012] Step 2: Preprocess the vibration signals collected in Step 1 to construct graph data suitable for the credible fusion diagnosis framework, specifically including the following steps:

[0013] 2.1: Perform unified standardization processing on the multi-channel vibration signals using the maximum-minimum standardization. For the m-th channel vibration signal xm , and its standardized result X m is as follows:

[0014]

[0015] where max(·) is the maximum function and min(·) is the minimum function.

[0016] 2.2: Based on the data standardized in Step 2.1, divide X m into n non - overlapping subsets, each with a length of d. Let:

[0017]

[0018] where L is the total length of the signal; d is the length of the divided signal; floor(·) is the floor function. The subsets of the m - channel signal can be expressed as:

[0019] H m = [X m,1 , X m,2 ,..., X m,n

[0020] 2.3: To identify the sensitive fault features contained in the signal, perform a fast Fourier transform on each data in the subset. The new samples after the transformation are the nodes in the graph data:

[0021]

[0022] where i = 1,..., n is the sample index within the subset. Subsequently, assign the corresponding label y to each sample m,i to obtain the labeled dataset D, where the labels include correct labels and incorrect (fault) classification labels.

[0023]

[0024] 2.4: After obtaining the graph data nodes , use the k - nearest neighbor method to construct edges. Calculate the Euclidean distance between nodes, select the k nodes closest to the central node as the neighbors of the central node, and the close relationship between every two nodes can form an edge. The weight of the edge is calculated according to the Gaussian kernel function.

[0025]

[0026] where w i,j represents the weight of the edge formed between the i - th sample node x i and its neighbor sample node x j ; ζ represents the bandwidth of the Gaussian kernel; Ne(·) represents the node x i ​Neighbors.

[0027] 2.5: According to the method in step 2.4, establish the edges between all samples and domain nodes to obtain the KNN graph. Repeat steps 2.1 to 2.4 to obtain the KNN graph results of the vibration acceleration signals of all channels, and then the multi-channel fusion diagnosis framework dataset applicable to G = {G1,..., G m ,..., G N} can be obtained, where N is the total number of channels.

[0028] Step 3: Improve the graph convolutional network model (GCN). Sequentially add a batch normalization layer and a Leaky ReLU layer between the graph convolutional layers, add a fully connected layer after two graph convolutional layers, and replace the Softmax layer with a Softplus layer to obtain a modified graph convolutional network model.

[0029] Step 4: Input the multi-channel fusion diagnosis framework dataset obtained in step 2 into the modified graph convolutional network model. By aggregating the information of each node and its neighbors, generate the non-linear node representation e m = [e1 m , e2 m ,..., e k m for each channel, where k represents the output feature dimension. To further enhance the interpretability and robustness of the node representation, introduce the Dirichlet distribution to model and optimize the node representation e m . Specifically, define the probability density function of the Dirichlet distribution as:

[0030]

[0031] where is the K-dimensional simplex, and K takes the same value as the output feature dimension k. μ is the distribution parameter. When , Dir(μ|α) = 0; α is the concentration parameter; map the node representation e m to the parameter α m = [α1 m , α2 m ,..., α k m such that the relationship between and satisfies Therefore, the optimized node and the distribution parameter μ m can be obtained by the following formula:

[0032]

[0033]

[0034] Step 5: Fuse multi-channel information through the reduced evidence theory to generate a unified opinion distribution. An "opinion" is a quantitative representation of the beliefs of different-channel information, manifested as belief mass and uncertainty mass μ m . Suppose the opinion obtained from channel signal 1 is and the opinion obtained from channel signal 2 is The fusion rule can be defined as:

[0035]

[0036] Furthermore, the specific calculation process of the fusion rule is:

[0037]

[0038] where is the conflict factor, and 0 ≤ C ≤ 1. The magnitude of the conflict factor value reflects the degree of conflict between the evidences. The larger C is, the greater the conflict between the evidences; conversely, the smaller it is.

[0039] Step 6: According to the transitivity of the D-S evidence theory combination rule, the fusion method of N-channel information can be obtained:

[0040]

[0041] where the opinion obtained from channel signal N is

[0042] Step 7: By calculating the loss between the correct label y corresponding to the sample m,i in each channel information and other misclassifications, the overall loss is obtained. Using the obtained overall loss for backpropagation, the modified graph convolutional network model with updated parameters is obtained. Since the output of the modified graph convolutional network model is a Dirichlet distribution, the traditional cross-entropy loss function needs to be improved. The improved cross-entropy loss function The calculation formula is:

[0043]

[0044] where p ij is the predicted probability distribution, ψ(·) is the Digamma function, which is the logarithmic derivative of the gamma function, expressed as After calculating the loss for each modal signal separately, the overall loss of the reliable fault intelligent diagnosis of the electromechanical coupling system based on multi-source information fusion includes:

[0045]

[0046] Step 8: Use the test data set to perform performance testing on the trained and corrected graph convolutional network model obtained in Steps 1 to 7, realize the fault diagnosis of the planetary gearbox based on multi-channel information fusion, and identify the fault modes of the gearbox fault test sample set.

[0047] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0048] (1) In the multi-channel vibration signal graph data preprocessing algorithm of the present invention, Fourier transform is used to obtain the features related to the fault information, and the k-nearest neighbor algorithm is used to realize the graph data construction of multi-channel information. The graph data preprocessing process can effectively enhance the weak fault features and is very suitable for the identification and diagnosis of weak fault features of the mechatronic coupling system under strong noise background;

[0049] (2) In the corrected graph convolutional network algorithm of the present invention, a batch normalization layer and a Leaky ReLU layer are added between the graph convolutional layers, a fully connected layer is added after two graph convolutional network layers, and the Softmax layer is replaced by a Softplus layer. The obtained output result is a non-linear vector, and the related improvements are more conducive to integrating with the Dirichlet distribution parameters, thereby forming a more unified fault diagnosis framework.

[0050] (3) In the credible fusion diagnosis framework of the present invention, by introducing subjective logic, the uncertainty of each channel information is calculated respectively, and then the diagnosis results of multi-modal information are fused through the reduced evidence fusion theory, and the credibility of the diagnosis results is verified by combining the fused uncertainty. Compared with the traditional intelligent diagnosis algorithm, it has the advantages of good diagnosis effect, high diagnosis efficiency, and credible diagnosis results.

[0051] (4) The fusion diagnosis mechanism in the present invention uses a method based on the graph network model to parallelly process the multi-channel vibration signals of the gearbox, deeply mines the hidden features in the multi-channel information, and compared with the traditional method of using a single-channel vibration sensor to judge the faults of the equipment, which is prone to problems such as missing key feature information, this method greatly improves the fault classification accuracy and the model operation efficiency.

[0052] The beneficial effects of the present invention: Through the collaborative analysis of the multi-modal signal correlation graph construction and the corrected graph convolutional network, the present invention can directly identify the fault state based on the real-time monitoring data, and can quantify and reduce the multi-source uncertainty in the diagnosis results through the subjective logic and Dirichlet distribution fusion mechanism, which can effectively improve the accuracy and conclusion credibility of the gearbox fault diagnosis. Description of the Drawings

[0053] Figure 1 It is the overall flowchart of the embodiment of the present invention.

[0054] Figure 2It is the industrial-level back-to-back planetary gear fault simulation test bench of the embodiment of the present invention; in the figure, 1 is a frequency converter, 2 is a driving motor, 3 is a coupling, 4 is a rotational speed and torque sensor, 5 is a signal acquisition card, 6 is a vibration acceleration sensor, 7 is a test gearbox, 8 is a speed increasing gearbox, 9 is a computer, 10 is a starting motor, and 11 is an earthquake-resistant base.

[0055] Figure 3 It is a schematic diagram of the figure data construction process of the embodiment of the present invention.

[0056] Figure 4 It is a schematic diagram of the modified graph convolutional neural network of the embodiment of the present invention.

[0057] Figure 5 It is a comparison of the diagnostic confusion matrices of seven signal combination methods of the model built in the embodiment of the present invention. Specific implementation manner

[0058] The following further describes the specific implementation process of the present invention in conjunction with the drawings and embodiments.

[0059] The present invention provides a reliable fault intelligent diagnosis method for a gearbox based on multi-channel information fusion, and the diagnosis process is as Figure 1 shown, including a multi-channel information preprocessing method, a modified graph convolutional network algorithm, and a reliable multi-channel information fusion framework construction method. The main steps are as follows:

[0060] Step 1: Install multiple acceleration sensors on the surface of the planetary gearbox in a magnetic adsorption manner, and collect multi-channel vibration signals of the planetary gearbox in different fault states respectively.

[0061] Step 2: Preprocess the vibration signals collected in Step 1 to construct graph data applicable to the reliable fusion diagnosis framework.

[0062] Step 3: Improve the basic graph convolutional network model, sequentially add a batch normalization layer and a Leaky ReLU layer between the graph convolutional layers, add a fully connected layer after two graph convolutional layers, and replace the Softmax layer with a Softplus layer to obtain a modified graph convolutional network model.

[0063] Step 4: Input the multi-channel fusion diagnosis framework data set obtained in Step 2 into the modified graph convolutional network model, generate a non-linear node representation for each channel by aggregating the information of each node and its neighbors, and optimize it.

[0064] Step 5: Introduce subjective logic to calculate the uncertainty of each Dirichlet distribution. Combine the evidence of the multi-channel signals with the parameters of the Dirichlet distribution, and The relationship between them satisfies

[0065] Step 6: After obtaining the uncertainty of multi-channel information through Step 5, the belief quality and uncertainty quality obtained from different channel information are fused using the reduced evidence theory.

[0066] Step 7: According to the transitivity of the combination rule of D-S evidence theory, the fusion method of N-channel information can be obtained.

[0067] Step 8: Calculate the loss of the correct label corresponding to the sample in each channel information relative to other misclassifications to obtain the overall loss of the intelligent diagnosis model based on multi-channel information fusion. Use the obtained overall loss for backpropagation to obtain the corrected graph convolutional network model with updated parameters.

[0068] Step 9: Use the test data set to perform a performance test on the corrected graph convolutional network model trained through Steps 1 to 7, realize the fault diagnosis of the planetary gearbox based on multi-channel information fusion, and identify the fault modes of the gearbox fault test sample set.

[0069] Embodiment

[0070] Build as Figure 2The industrial-grade back-to-back planetary gear fault simulation test bench shown. This test bench mainly consists of a frequency converter 1, a drive motor 2, a coupling 3, a rotational speed and torque sensor 4, a signal acquisition card 5, a vibration acceleration sensor 6, a test gearbox 7, a speed increasing gearbox 8, a computer 9, a starting motor 10, an earthquake-resistant base 11, etc. The drive motor drives the test gearbox to rotate, and the power source drives the load motor to operate through a speed increasing and reducing gearbox. The load motor is a load for the test gearbox. By adjusting the magnitude of the excitation current of the load motor, the magnitude of the load torque can be changed, thereby realizing the loading control of the test gearbox. The signal acquisition system consists of a B&K vibration acceleration sensor, an NI-9231 data acquisition card, a laptop computer, and test software. Three vibration acceleration sensors are respectively arranged at the positions of the input shaft, output shaft, and upper end of the gear ring of the test gearbox. At the start of the test, the three channels simultaneously acquire the vibration signals during the operation of the gearbox. The sampling rate of all channels is set to 12,800 Hz, the sampling time is 28 s, the rotational speed of the drive motor is set to 1,200 r / min, and the load is 300 Nm. In order to verify the accuracy of the constructed credible fusion framework in weak fault identification, a total of 15 kinds of gear health states are set in this experiment, mainly including the healthy state, 6 kinds of sun gear fault parts, 6 kinds of planet gear fault parts, and 2 kinds of compound faults of planet gears and sun gears. It should be noted that during the processing of the fault parts, three types of faults with tooth root cracks of 1 mm, 1.8 mm, and 2.5 mm are processed according to different fault damage degrees. The detailed health states of the planetary gearbox in this test bench and the corresponding labels are shown in Table 1. Under the above experimental conditions, the parameters and characteristic frequencies of the reduction planetary gearbox 4 are shown in Table 2.

[0071] Table 1 Detailed health states of the planetary gearbox and the corresponding labels

[0072]

[0073] Table 2 Parameters and characteristic frequencies of the reduction planetary gearbox

[0074]

[0075] Specifically as follows:

[0076] Step 1: Install the three acceleration sensors on the surface of the planetary gearbox body in a magnetic adsorption manner, and respectively acquire the multi-channel vibration signals of the planetary gearbox in different fault states. Figure 3 The time-domain and frequency-domain waveforms of the three channels when the gear is faulty. Due to the weak fault characteristics, in Figure 3 the time-domain waveform, there is no obvious periodic impact information, the frequency spectrum is chaotic and disordered, and the fault characteristic frequency is completely submerged by noise. Therefore, it is impossible to judge the current health state of the gearbox only from the time-domain and frequency-domain analysis;

[0077] Step 2: Preprocess the vibration signals collected in Step 1 to construct graph data suitable for the trusted fusion diagnosis framework, which specifically includes the following steps:

[0078] (1) Perform unified standardization processing on multi-channel vibration signals using maximum-minimum standardization. For the vibration signal x of the m-th channel m , its standardized result is as follows:

[0079]

[0080] (2) After the standardization processing, divide the data into n non-overlapping subsets, each group with a length of d. Let:

[0081]

[0082] where L is the total length of the signal; d is the length of the divided signal; floor(·) is the floor function. The subset of the m-channel signal can be expressed as:

[0083] H m = [X m,1 , X m,2 ,..., X m,n

[0084] (3) To identify the sensitive fault features contained in the signal, perform a fast Fourier transform on each data in the subset. The new samples after the transformation are the nodes in the graph data:

[0085]

[0086] where i = 1,..., n is the sample index within the subset. Subsequently, assign the corresponding label y to each sample m,i to obtain the labeled dataset D, where the labels include correct labels and wrong (fault) classification labels.

[0087]

[0088] (4) After obtaining the graph data nodes , use the k-nearest neighbor method to construct edges. Calculate the Euclidean distance between nodes, select the k nearest nodes to the central node as the neighbors of the central node, and the close relationship between every two nodes can form an edge. The weight of the edge is calculated according to the Gaussian kernel function.

[0089]

[0090] where w i,j represents the i-th sample node x i and its neighbor sample node x j ​The weight of the edge formed between; ζ represents the bandwidth of the Gaussian kernel; Ne(·) represents the neighbor of node x i of.

[0091] (5) According to the method in (1-4), establish the edges between all samples and the domain nodes to obtain the KNN graph. Repeat steps 2.1 to 2.4 to obtain the KNN graph results of the vibration acceleration signals of all channels, and then the applicable G = {G1,..., G m ,..., G N} can be obtained, where N is the total number of channels.

[0092] Step 3: Improve the basic graph convolutional network model by sequentially adding a batch normalization layer and a Leaky ReLU layer between the graph convolutional layers, adding a fully connected layer after two graph convolutional layers, and replacing the Softmax layer with a Softplus layer to obtain the modified graph convolutional network model. The modified graph convolutional network model is as Figure 4 shown.

[0093] Step 4: Input the multi-channel fusion diagnosis framework dataset obtained in step 2 into the modified graph convolutional network model. By aggregating the information of each node and its neighbors, generate the non-linear node representation e m = [e1 m , e2 m ,..., e k m , where k represents the output feature dimension. To further enhance the interpretability and robustness of the node representation, introduce and optimize the node representation e m . Specifically, define the probability density function of the Dirichlet distribution as:

[0094]

[0095] where is the K-dimensional simplex, which is consistent with the value of the output feature dimension k here. u is the distribution parameter. When , Dir(μ|α) = 0. α is the concentration parameter. Map the node representation e m to the parameter α m = [α1 m , α2 m ,..., α k m such that the relationship between and satisfies Therefore, the optimized node and the distribution parameter μ m can be obtained by the following formula:

[0096]

[0097] Step 5: Fuse multi-channel information through the reduced evidence theory to generate a unified opinion distribution. An "opinion" is a quantitative representation of the beliefs of different channel information, manifested as belief mass and uncertainty mass μ m . Suppose the opinion obtained from channel signal 1 is The opinion obtained from channel signal 2 is The fusion rule is:

[0098]

[0099] Furthermore, the specific calculation process of the fusion rule is:

[0100]

[0101] where is the conflict factor, and 0 ≤ C ≤ 1. The value of the conflict factor reflects the degree of conflict between the evidences. The larger the value of C, the greater the conflict between the evidences, and vice versa.

[0102] Step 7: According to the transitivity of the D-S evidence theory combination rule, the fusion method of N-channel information can be obtained:

[0103]

[0104] where the opinion obtained from channel signal N is

[0105] Step 8: By calculating the loss between the correct label y corresponding to the sample m,i in each channel information and other misclassifications, the overall loss of the intelligent diagnosis model based on multi-channel information fusion is obtained. Using the obtained overall loss for backpropagation, the modified graph convolutional network model with updated parameters is obtained. Since the output of the model is a Dirichlet distribution, the traditional cross-entropy loss function needs to be improved, and the calculation formula is:

[0106]

[0107] where p ij is the predicted probability distribution, ψ(·) is the Digamma function, which is the logarithmic derivative of the gamma function, expressed as After calculating the loss for each modal signal separately, the overall loss of the credible fault intelligent diagnosis of the electromechanical coupling system based on multi-source information fusion includes:

[0108]

[0109] Step 9: Use the test data set to perform performance testing on the trained and corrected graph convolutional network model obtained in Steps 1 to 7, realize the fault diagnosis of the planetary gearbox based on multi-channel information fusion, and identify the fault modes of the gearbox fault test sample set.

[0110] To verify the superiority of the established fusion diagnosis framework in dealing with multi-channel information fusion tasks, a multi-channel combination method was set for fault classification, and the accuracy of fusion diagnosis was compared with the accuracy of single-channel information fault diagnosis. During the model training process, the epoch was set to 100, the initial learning rate was set to 0.001, and the optimizer was set to Adam. To exclude the randomness and instability during the testing process, each signal combination method was tested 10 times, and the classification accuracy was mainly used as the evaluation criterion for the signal combination method. The evaluation criteria adopted mainly include the maximum accuracy (max-acc), the minimum accuracy (min-acc), the average accuracy (avg-acc), and the standard deviation (std). The diagnostic results obtained by using various signal combination methods and the proposed fusion diagnosis framework are shown in Table 3.

[0111] Table 3 Credible Fusion Diagnosis Results of Multi-channel Signals

[0112]

[0113] As can be seen from the table, the diagnostic accuracy obtained by using multi-channel vibration signals based on the proposed method reached as high as 100%, which was 6.63%, 4.14%, and 2.09% higher than the diagnostic accuracy of single-channel signals respectively. In addition, the confusion matrices of the diagnostic results of the seven signal combination methods in the above table are as Figure 5 shown. The experimental results show that in the multi-modal data test set constructed by seven different signal combinations, the diagnostic performance of the proposed corrected graph convolutional network model for 14 types of typical faults is at a relatively high level. It should be particularly noted that when using dual-channel vibration signals as input, due to the information complementarity between different sensing channels, the comprehensive recognition accuracy of the model is higher than that of single-channel input. Moreover, under the condition of triple-channel vibration signal input, the model achieves a 100% fault recognition accuracy, and the main diagonal elements between classes of its confusion matrix reach the ideal state value of 1, and all samples are correctly classified. The experimental results not only verify the effectiveness of multi-source heterogeneous data fusion, but also fully verify the accuracy and superiority of the method proposed in the present invention in intelligent diagnosis.

Claims

1. A gearbox credible fault intelligent diagnosis method based on multi-channel information fusion, characterized in that: The specific steps are as follows: Step 1: Install multiple acceleration sensors on the surface of the planetary gearbox by magnetic attraction, and collect multi-channel vibration signals of the planetary gearbox under different fault states; Step 2: Preprocess the vibration signal collected in step 1 to construct graph data suitable for the trusted fusion diagnosis framework, which specifically includes the following steps: 2.1: Use maximum and minimum standardization to uniformly standardize multi-channel vibration signals. For the vibration signal x of the mth channel m , the standardized result X m As follows: Among them, max(·) is the maximum value function, min(·) is the minimum value function; 2.2: Based on the data standardized in step 2.1, m Divide into n non-overlapping subsets, each with a length of d, let: Where L is the total length of the signal; d is the length of the divided signal; floor(·) is the floor function; the subset of the m-channel signal can be expressed as: H m =[X m,1 ,X m,2 ,...,X m,n ] 2.3: In order to identify the sensitive fault features contained in the signal, each data in the subset is subjected to fast Fourier transform, and the transformed new samples are the nodes in the graph data: Where i=1,...,n is the sample index in the subset; then, for each sample Assign the corresponding label y m,i Obtain a labeled data set D, where the labels include correct labels and error (fault) classification labels; 2.4: Getting graph data nodes After that, the k-nearest neighbor method is used to construct edges, calculate the Euclidean distance between nodes, and select the k nodes closest to the central node as the neighbors of the central node. The close relationship between every two nodes can form an edge, and the weight of the edge is calculated according to the Gaussian kernel function; in, w i,j Represents the i-th sample node x i and its neighboring sample node x j The weight of the edge formed between them; ζ represents the bandwidth of the Gaussian kernel; Ne(·) represents the node x i Neighbors; 2.5: According to the method in step 2.4, establish the edges between all samples and domain nodes to obtain the KNN graph; repeat steps 2.1 to 2.4 to obtain the KNN graph results of the vibration acceleration signals of all channels, and you can get the KNN graph applicable to G={G1,...,G m ,...,G N }, N is the total number of channels; Step 3: Improve the graph convolutional network model GCN by adding batch normalization layers and Leakey ReLU layers between the graph convolutional layers, adding a fully connected layer after the two graph convolutional layers, and replacing the Softmax layer with the Softplus layer to obtain a modified graph convolutional network model; Step 4: Input the multi-channel fusion diagnosis framework dataset obtained in step 2 into the modified graph convolutional network model, and generate a nonlinear node representation of each channel by aggregating the information of each node and its neighbors. m =[e1 m ,e2 m ,...,e k m ], where k represents the feature dimension of the output; in order to further enhance the interpretability and robustness of the node representation, the Dirichlet distribution model is introduced and the node representation e is optimized. m ; Specifically, the probability density function of the Dirichlet distribution is defined as: in, is a K-dimensional simplex, K is consistent with the output feature dimension k; μ is the distribution parameter, when When Dir(μ|α)=0; α is the concentration parameter; the node represents e m Parameter α mapped to Dirichlet distribution m =[α1 m ,α2 m ,...,α k m ],make and The relationship between Therefore, the optimized node and distribution parameter μ m It can be obtained by the following formula: Step 5: Fusion of multi-channel information through evidence theory of reduction to generate a unified opinion distribution. "Opinion" is a quantitative representation of beliefs in different channel information, which is reflected in the quality of belief. and the uncertainty mass μ m ; Assume that the opinion obtained from channel signal 1 is The opinion obtained from channel signal 2 is The fusion rule can be defined as: Furthermore, the specific calculation process of the fusion rule is as follows: in, is the conflict factor, and 0≤C≤1. The value of the conflict factor reflects the conflict between the evidences. The larger the C, the greater the conflict between the evidences, and vice versa. Step 6: According to the transferability of the combination rule of DS evidence theory, the fusion method of N channel information is obtained: Among them, the opinion obtained from the channel signal N is Step 7: By calculating the samples in each channel information The corresponding correct label y m,i Compared with the losses between other misclassifications, the overall loss is obtained; the overall loss is used for back propagation to obtain a modified graph convolutional network model with updated parameters; Step 8: Use the test data set to perform a performance test on the modified graph convolutional network model trained in steps 1 to 7, implement planetary gearbox fault diagnosis based on multi-channel information fusion, and identify the fault mode of the gearbox fault test sample set.

2. The gearbox reliable fault intelligent diagnosis method based on multi-channel information fusion according to claim 1 is characterized in that: The cross entropy loss function needs to be improved. The improved cross entropy loss function The calculation formula is: Among them, p ij is the predicted probability distribution, ψ(·) is the Digamma function, which is the logarithmic derivative of the gamma function and is expressed as After calculating the loss of each modal signal, the overall loss of intelligent diagnosis of trusted faults of electromechanical coupling system based on multi-source information fusion is obtained, which includes:

Images