Gearbox fault diagnosis method based on WATD-MTF residual network

Abstract

The invention provides a gearbox fault diagnosis method based on a WATD-MTF residual network, and relates to the technical field of gearbox testing. Wavelet transform adaptive denoising is carried out on the vibration signals of the gearbox to be detected, noise in the vibration signals is removed, the influence of the noise on a fault diagnosis result is reduced, two-dimensional MTF transformation is further carried out, a two-dimensional MTF graph is generated, time sequence signals between fault features in the vibration signals are reserved, and the fault diagnosis accuracy is improved. According to the method, the WATD-MTF residual network can conveniently carry out fault identification on the vibration signals, the influence of useless information in the vibration signals on a fault diagnosis result is reduced, effective diagnosis of gearbox faults is realized, and the accuracy of gearbox fault diagnosis is improved.

Description

Technical Field

[0001] The invention relates to the technical field of gearbox testing, and in particular to a gearbox fault diagnosis method based on a WATD-MTF residual network. Background Art

[0002] With the rapid development of science and technology and modern industry, the mechanical equipment in the transportation, aerospace, defense and other industries are becoming more and more automated, integrated, high-speed and large-scale. Gearboxes have played a vital role in the development of various industries. Due to the advantages of small size, small mass, fixed transmission ratio, large transmission torque, compact structure, large load capacity, high efficiency, stable operation and strong adaptability, gearboxes have become the most common transmission equipment for transmitting power and changing speed in the machinery industry, and are widely used in various types of mechanical equipment. Therefore, the performance of gearboxes seriously affects the performance of mechanical equipment in actual production and life applications.

[0003] As a key component of the mechanical transmission chain, the working state of the gearbox plays an important role in the transmission system. As the transmission mechanism of the entire system, the gearbox has a compact internal structure and strong coupling between components, which can easily cause gear pitting, broken teeth and other faults. At the same time, the gearbox failure repair process is relatively complicated and the repair time is long, resulting in great economic losses.

[0004] However, the working environment of the gearbox is complex and changeable, and the signal contains a lot of noise when the gearbox is performing fault analysis, resulting in too much useless information and interference information in the gearbox fault diagnosis process, making the gearbox fault diagnosis result inaccurate. Summary of the invention

[0005] The present invention provides a gearbox fault diagnosis method based on a WATD-MTF residual network, which can realize effective diagnosis of gearbox faults and improve the accuracy of gearbox fault diagnosis.

[0006] In a first aspect, the present invention provides a gearbox fault diagnosis method based on a WATD-MTF residual network, the method comprising: obtaining a vibration signal of a gearbox to be tested when it is working; using wavelet transform to perform adaptive denoising on the vibration signal to obtain a denoised signal; based on the denoised signal, performing a two-dimensional MTF transform to generate an MTF graph, the MTF graph is used to characterize time series information of fault characteristics in the vibration signal; based on the MTF graph and a preset WATD-MTF residual network, predicting a fault diagnosis result of the gearbox to be tested, the fault diagnosis result including a fault type and a fault probability.

[0007] In a possible implementation, a vibration signal is subjected to adaptive noise reduction using wavelet transform to obtain a denoised signal, including: performing wavelet decomposition on the vibration signal to obtain a low-frequency signal and a high-frequency signal; calculating the noise standard deviation of each layer in the low-frequency signal and the high-frequency signal based on the low-frequency signal and the high-frequency signal; determining an adaptive threshold of each layer based on the noise standard deviation of each layer and the signal length of the vibration signal; and denoising and recombining the low-frequency signal and the high-frequency signal based on the adaptive threshold of each layer to obtain a denoised signal.

[0008] In a possible implementation, determining the adaptive threshold of each layer based on the noise standard deviation of each layer and the signal length of the vibration signal includes: determining the adaptive threshold of each layer based on the following formula;

[0009]

[0010] Where N is the signal length of the vibration signal, l j is the adaptive threshold of the i-th layer, s j is the noise standard deviation of the i-th layer.

[0011] In a possible implementation, a two-dimensional MTF transformation is performed based on the denoised signal to generate an MTF graph, including: discretizing the signal value at each moment in the denoised signal to obtain a value range of the signal value; equally dividing the value range to obtain a plurality of dividers; dividing the denoised signal based on a preset time window to obtain denoised signals of a plurality of time windows; performing probability calculation based on the denoised signals of the plurality of time windows and the plurality of dividers to construct an MTF graph for each time window; wherein the element in the i-th row and j-th column of the MTF graph of any time window represents the signal value at each moment in the time window, and the probability that the current moment belongs to the i-th divider under the condition that the previous moment belonged to the j-th divider.

[0012] In a possible implementation, based on the MTF diagram and the preset WATD-MTF residual network, before predicting the fault diagnosis result of the gearbox to be tested, it also includes: obtaining the vibration signal of the gearbox in the historical period, the fault time and fault type of each fault in the historical period; based on the fault time of each fault, splitting the vibration signal of the gearbox in the historical period to obtain the vibration signal of the time period where each fault time is located; performing wavelet transform and two-dimensional MTF transform on the vibration signal of the time period where each fault time is located to obtain the MTF diagram at each fault time; using the MTF diagram at each fault time as input and the fault type at each fault time as output to generate multiple training samples; based on the multiple training samples, performing neural network training to obtain the WATD-MTF residual network.

[0013] In a possible implementation, a neural network is trained based on multiple training samples to obtain a WATD-MTF residual network, including: constructing a WATD-MTF residual network based on a convolutional layer, an SKNet module, a residual module, an average pooling layer, and a fully connected layer; based on multiple training samples, the constructed WATD-MTF residual network is trained to obtain a preset WATD-MTF residual network.

[0014] In one possible implementation, based on the MTF graph and a preset WATD-MTF residual network, a fault diagnosis result of the gearbox to be tested is predicted, including: based on the MTF graph, attention weight fusion is performed to obtain a fused feature graph; based on the fused feature graph and a preset residual module, a fault diagnosis result is obtained.

[0015] In a possible implementation, attention weight fusion is performed based on the MTF map to obtain a fused feature map, including: based on preset convolution kernels of different sizes, feature extraction is performed on the MTF feature map to obtain multiple convolution features; feature fusion is performed on multiple convolution features to obtain fused features; based on the fused features, a global average pooling operation is performed to generate channel statistics; based on the channel statistics of multiple channels, a compact feature vector is generated; based on the compact feature vector, the attention weight of each convolution kernel is calculated; based on multiple convolution features and the attention weight of each convolution kernel, a fused feature map is obtained.

[0016] In the second aspect, an embodiment of the present invention provides a gearbox fault diagnosis device based on a WATD-MTF residual network, the device including a communication module and a processing module, the communication module is used to obtain a vibration signal of the gearbox to be tested when it is working; the processing module is used to perform adaptive denoising on the vibration signal using wavelet transform to obtain a denoised signal; based on the denoised signal, a two-dimensional MTF transform is performed to generate an MTF graph, and the MTF graph is used to characterize the time series information of the fault characteristics in the vibration signal; based on the MTF graph and a preset WATD-MTF residual network, the fault diagnosis result of the gearbox to be tested is predicted, and the fault diagnosis result includes the fault type and the fault probability.

[0017] In a third aspect, an embodiment of the present invention provides an electronic device, comprising a memory and a processor, the memory storing a computer program, the processor being used to call and run the computer program stored in the memory to execute the steps of the method described in the first aspect and any possible implementation method of the first aspect.

[0018] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and wherein when the computer program is executed by a processor, the steps of the method described in the first aspect and any possible implementation method of the first aspect are implemented.

[0019] The present invention provides a gearbox fault diagnosis method based on a WATD-MTF residual network. The present invention performs wavelet transform adaptive denoising on a vibration signal of a gearbox to be tested, thereby removing noise in the vibration signal and reducing the influence of the noise on the fault diagnosis result. Further, a two-dimensional MTF transform is performed to generate a two-dimensional MTF graph, and a time series signal between fault features in the vibration signal is retained, so that the WATD-MTF residual network is used to perform fault identification on the vibration signal, and the influence of useless information in the vibration signal on the fault diagnosis result is reduced, thereby achieving effective diagnosis of the gearbox fault and improving the accuracy of the gearbox fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0021] Figure 1 It is a flow chart of a gearbox fault diagnosis method based on a WATD-MTF residual network provided by an embodiment of the present invention;

[0022] Figure 2 It is a schematic diagram of a wavelet threshold denoising algorithm flow provided by an embodiment of the present invention;

[0023] Figure 3 It is a schematic diagram of a process of MTF sliding coding provided by an embodiment of the present invention;

[0024] Figure 4 It is a schematic diagram of the structure of an SKNet network provided by an embodiment of the present invention;

[0025] Figure 5 is a schematic structural diagram of a residual module provided by an embodiment of the present invention;

[0026] Figure 6 1 is a schematic diagram of the structure of a WATD-MTF residual network provided by an embodiment of the present invention;

[0027] Figure 7 is a schematic diagram of a gear box vibration signal provided by an embodiment of the present invention;

[0028] Figure 8 is a schematic diagram of recognition accuracy and loss rate before and after noise reduction provided by an embodiment of the present invention;

[0029] Fig. 9 1 is a schematic diagram of the structure of a gearbox fault diagnosis device based on a WATD-MTF residual network provided by an embodiment of the present invention;

[0030] Fig.10 It is a schematic diagram of the structure of another electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0031] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present invention. However, it should be clear to those skilled in the art that the present invention may be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present invention.

[0032] In the description of the present invention, unless otherwise specified, " / " means "or", for example, A / B can mean A or B. "And / or" in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. In addition, "at least one" and "plurality" refer to two or more. The words "first", "second", etc. do not limit the quantity and execution order, and the words "first", "second", etc. do not limit them to be different.

[0033] In the embodiments of the present application, words such as "exemplary" or "for example" are used to indicate examples, illustrations or descriptions. Any embodiment or design described as "exemplary" or "for example" in the embodiments of the present application should not be interpreted as being more preferred or more advantageous than other embodiments or designs. Specifically, the use of words such as "exemplary" or "for example" is intended to present related concepts in a concrete way for easy understanding.

[0034] In addition, the terms "including" and "having" and any variations thereof mentioned in the description of this application are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or modules is not limited to the listed steps or modules, but may optionally include other steps or modules that are not listed, or may optionally include other steps or modules that are inherent to these processes, methods, products or devices.

[0035] In order to make the purpose, technical solutions and advantages of the present invention more clear, the following will be described through specific embodiments in conjunction with the accompanying drawings of the present invention.

[0036] As described in the background technology, the gearbox is one of the commonly used components in rotating machinery. Its internal gears and bearings and other parts will inevitably fail during service, causing unnecessary losses and even casualties in serious cases. In order to ensure the stable and safe operation of mechanical equipment, it is of great significance to study intelligent fault diagnosis methods that can avoid the shortcomings of manual feature extraction methods and have strong adaptability.

[0037] Traditional fault diagnosis methods require manual feature extraction. Fault signal features are difficult to extract due to the complexity of actual working conditions, which is not only time-consuming but also inefficient. With further research by scholars, algorithms for signal analysis have gradually developed. Currently, there is a nonparametric complementary ensemble empirical mode decomposition (NPCEEMD) method for identifying bearing defects using weak features. There is also a method that uses the local mean method to decompose the vibration signal into multiple product function components, and uses the permutation entropy algorithm on the decomposed components to calculate the components containing fault information, thereby determining the fault type.

[0038] At present, there is a method that uses variational mode decomposition (VMD) to achieve adaptive decomposition of signals, and uses the gray wolf algorithm to optimize the decomposition parameters, and extracts fault features by calculating the time domain parameters, energy and permutation entropy of the fault signal. However, in the actual signal acquisition process, there will be a lot of noise, and the main features of the signal will be covered by a lot of noise, resulting in poor effect of extracting features by signal analysis methods. In order to solve the above problems, some diagnostic methods select the components of Hilbert vibration decomposition modally, and construct their feature vectors with wavelet packets to complete the denoising of the signal, and obtain high recognition accuracy in the bearing fault diagnosis task. There is also an adaptive wavelet denoising method. Experiments show that the fault recognition accuracy of the signal after denoising is 6% higher than that of the signal without denoising. The proposed wavelet threshold denoising can also be compared with the other three algorithms, and the effectiveness of the proposed method is verified by experiments.

[0039] From the above analysis, it can be seen that effectively reducing the noise of the actual collected signal and improving the fault identification accuracy are the key to the entire fault diagnosis. In recent years, the data-driven intelligent fault diagnosis method has gradually become the mainstream algorithm in the field of fault diagnosis. By combining with the noise reduction algorithm, it has shown a high advantage in the field of fault diagnosis. There is also a method that combines feature fusion and support vector machine. First, the original signal is preprocessed using wavelet packets, variational mode decomposition, etc., and then the features are linearly fused in multiple groups, and finally the gear fault is classified using SVM. VMD can also be combined with singular spectrum analysis (SSA) to effectively suppress the noise, and finally SVM is used for fault diagnosis. However, the intelligent diagnosis algorithm has certain shortcomings in fault diagnosis. When processing complex and irregular data, the model will have overfitting and underfitting problems. In response to this, the concept of deep learning is proposed. Deep learning has made great progress in the fields of computer vision, speech recognition and image recognition. With the in-depth research of many scholars, deep learning has been widely used in the field of fault diagnosis. There is also an adaptive neural network model that performs fault diagnosis on gearboxes and verifies the superiority of the network. There is also a multi-scale convolutional network, which has been verified experimentally to have strong feature extraction capabilities in bearing fault diagnosis tasks. There is also a dynamic wide convolution residual network bearing fault diagnosis method, which achieves effective fault identification by extracting feature information of different scales. Although the above studies have achieved fruitful research results, they are all based on one-dimensional signal verification and cannot give full play to the advantages of deep learning image processing. The one-dimensional signal can also be converted into a Gram Angular Difference Field (GADF) and input into CNN to complete the extraction and classification of rolling bearing fault features. The Markov transfer field (MTF) can also be used to encode the collected vibration signal and input it into the CNN model for fault classification. Through experimental verification, a bearing fault identification accuracy rate of 99.8% has been achieved.

[0040] Although the above-mentioned methods have achieved good results in fault diagnosis tasks, fault diagnosis in actual industry still faces serious problems: (1) Modern mechanical equipment is complex and has uncertain working conditions. In addition to being noisy, the signals collected in the working environment also show nonlinear characteristics and variable characteristic parameters, which reduces the sensitivity of the model to key features and makes it impossible to effectively distinguish the types of faults. (2) When processing complex signals, the number of diagnostic model layers needs to be deepened to capture more complex features, but when the model structure is deepened, problems such as gradient vanishing and explosion may occur.

[0041] Based on the above analysis, the present invention proposes a gearbox fault diagnosis method that combines wavelet adaptive threshold denoising with Markov transfer field and improved residual network. In view of the fact that the collected signal contains a lot of noise that affects the accuracy of the model, the signal is denoised by wavelet adaptive threshold denoising. At the same time, relying on the advantage of the residual network in processing two-dimensional signals, the denoised signal is converted into an MTF graph. In order to solve the problems of gradient explosion and disappearance when the number of model layers is deepened, the ResNet18 model is used for analysis, and the SKNet attention network is integrated into the ResNet18 model to enhance the model's sensitivity to key features and improve the ability to capture weak features. A large number of experiments have fully verified that the proposed method has higher fault identification accuracy and generalization performance than other intelligent algorithms.

[0042] like Figure 1 As shown, an embodiment of the present invention provides a gearbox fault diagnosis method based on WATD-MTF residual network. The method includes steps S101-S104.

[0043] S101, obtaining a vibration signal of the gearbox to be tested when it is working.

[0044] S102, using wavelet transform to perform adaptive noise reduction on the vibration signal to obtain a denoised signal.

[0045] As a possible implementation manner, step S102 can be specifically implemented as steps S1021-S1024.

[0046] S1021. Perform wavelet decomposition on the vibration signal to obtain a low-frequency signal and a high-frequency signal.

[0047] S1022. Calculate the noise standard deviation of each layer in the low-frequency signal and the high-frequency signal based on the low-frequency signal and the high-frequency signal.

[0048] S1023. Determine an adaptive threshold of each layer based on the noise standard deviation of each layer and the signal length of the vibration signal.

[0049] Exemplarily, the embodiment of the present invention may determine the adaptive threshold of each layer based on the following formula.

[0050]

[0051] Where N is the signal length of the vibration signal, l j is the adaptive threshold of the i-th layer, s j is the noise standard deviation of the i-th layer.

[0052] S1024. Based on the adaptive threshold of each layer, denoise and reorganize the low-frequency signal and the high-frequency signal to obtain a denoised signal.

[0053] It should be noted that the main principle of wavelet threshold denoising is to decompose the noisy signal layer by layer, and decompose it into approximate components and detail components at each different scale of the signal. The approximate component represents the low-frequency part of the signal, that is, the noise-free part, and the detail component represents the high-frequency part, that is, the noisy part. The noise component is mainly concentrated in the detail component of wavelet decomposition. The basic process of wavelet threshold denoising is to select a threshold according to the signal characteristics, and stipulate that the signal part greater than this value is low-frequency information and should be retained; the part less than this value is a high-frequency signal and should be filtered out. The low-frequency signal after decomposition is reorganized by using the inverse wavelet transform to obtain the denoised signal. The process of the wavelet threshold denoising algorithm is as follows: Figure 2 shown.

[0054] The key to wavelet threshold denoising lies in the selection of threshold. Common denoising methods include hard threshold denoising and soft threshold denoising. However, these two methods have discontinuities and different degrees of deviation. Due to the lack of adaptability to signal characteristics, they are not applicable to all working conditions, which affects the final denoising effect. Therefore, this paper uses adaptive threshold to denoise the gearbox fault signal. The specific formula is as follows.

[0055]

[0056] Where N is the signal length of the vibration signal, l j is the adaptive threshold of the i-th layer, s j is the noise standard deviation of the i-th layer.

[0057]

[0058] The signal after layer-by-layer decomposition and reconstruction by adaptive threshold denoising is:

[0059]

[0060] Among them, CA j is the coefficient of the j-th level approximate component, j j (t) is the scale function, Represents the coefficient of the detail component at level j, y j (tk) represents the j-th level wavelet function, k is the translation parameter of the wavelet function, which controls the position of the wavelet function.

[0061] According to the complex characteristics of the gearbox fault signal, the db4 wavelet basis function is selected and the decomposition layer number is set to 4 to fully filter out the noise influence. The optimal threshold is automatically selected using the noise level and the statistical characteristics of the signal to achieve a balance between noise suppression and signal retention.

[0062] S103: Perform a two-dimensional MTF transformation based on the denoised signal to generate an MTF graph.

[0063] In the embodiments of the present application, the MTF graph is used to characterize the time series information of the fault features in the vibration signal.

[0064] As a possible implementation manner, step S103 can be specifically implemented as steps S1031 - S1034.

[0065] S1031. Based on the signal values at each moment in the denoised signal, perform discretization processing to obtain the value range of the signal values.

[0066] S1032. Divide the value range equally to obtain a plurality of quantizers.

[0067] S1033. Based on a preset time window, divide the denoised signal to obtain the denoised signals of multiple time windows.

[0068] S1034. Based on the denoised signals of multiple time windows and a plurality of quantizers, perform probability calculation to construct the MTF graph of each time window.

[0069] Among them, the element in the i-th row and j-th column of the MTF graph of any time window represents the probability that the signal value at each moment in this time window belongs to the i-th quantizer under the condition that the previous moment belongs to the j-th quantizer.

[0070] It should be noted that the Markov Transition Field (MTF) refers to a method of transforming a time series signal into a two-dimensional graph. The MTF method effectively retains the time series information of the signal by constructing a Markov transition field matrix and introducing time position information. The specific steps of the encoding process are as follows:

[0071] Step 1. Given a time series X = {x1, x2, …, x n}, discretize the sampling points in the time series into Q equal parts (Q < N) to ensure that each data point x i is mapped to the corresponding q i quantizer.

[0072] Step 2. Construct a Markov transition matrix W of size Q × Q through the state probabilities between every two quantizers, and its expression is as follows:

[0073]

[0074] Among them, w ij is p{x t ∈q i |x t+1 ∈q i+1}, p is the probability, and x tis the value of the signal at time t, and q is the quantizer. It can be further expressed as the following formula along the time axis through a first-order Markov chain.

[0075]

[0076] Step 3: Since the Markov transfer matrix W ignores the influence of time on state probability, in order to retain the time correlation of the gear fault signal, the Markov transfer field M is constructed by adding time bit information to the Markov matrix W, as shown in the following formula:

[0077]

[0078] Among them, m ij is the quantile q i Transfer to quantile q j The transition probability, that is, m ij =P(q i →q j ).

[0079] The one-dimensional signal is converted to MTF using a sliding window. In order to ensure that the MTF graph can fully include the fault characteristics, the number of sliding window data points is set to 512, and the data is enhanced by 50% resampling to improve the characterization ability of the model. Each MTF graph is divided into many grid-like cells, and the probabilities of grids with different pixel intensity values ​​(i.e., different colors) are transformed and stored through the Markov chain. There is a temporal correlation between different cells in the MTF, i.e., different probabilities, which retains important information to a certain extent. The process of MTF sliding coding is as follows Figure 3 shown.

[0080] S104, based on the MTF diagram and the preset WATD-MTF residual network, predict and obtain the fault diagnosis result of the gearbox to be tested.

[0081] In the embodiment of the present application, the fault diagnosis result includes the fault type and the fault probability.

[0082] As a possible implementation manner, step S104 may be specifically implemented as steps S1041 - S1042 .

[0083] S1041. Based on the MTF graph, attention weight fusion is performed to obtain a fused feature graph.

[0084] Exemplarily, step S104 may be specifically implemented as steps A1-A6.

[0085] A1. Based on the preset convolution kernels of different sizes, the MTF feature map is extracted to obtain multiple convolution features.

[0086] A2. Perform feature fusion on multiple convolutional features to obtain fused features.

[0087] A3. Based on the fused features, a global average pooling operation is performed to generate channel statistics.

[0088] A4. Generate a compact feature vector based on channel statistics of multiple channels.

[0089] A5. Calculate the attention weight of each convolution kernel based on the compact feature vector.

[0090] A6. Based on multiple convolutional features and the attention weight of each convolution kernel, a fused feature map is obtained.

[0091] S1042. Obtain a fault diagnosis result based on the fused feature map and the preset residual module.

[0092] The present invention provides a gearbox fault diagnosis method based on a WATD-MTF residual network. The method performs wavelet transform adaptive denoising on a vibration signal of a gearbox to be tested, thereby removing noise in the vibration signal and reducing the influence of the noise on the fault diagnosis result. Furthermore, a two-dimensional MTF map is generated by a two-dimensional MTF transform, and a time series signal between fault features in the vibration signal is retained, so that the WATD-MTF residual network is used to perform fault identification on the vibration signal, and the influence of useless information in the vibration signal on the fault diagnosis result is reduced, thereby achieving effective diagnosis of the gearbox fault and improving the accuracy of the gearbox fault diagnosis.

[0093] Optionally, the gearbox fault diagnosis method based on WATD-MTF residual network provided by the embodiment of the present invention further includes steps S201-S205 before step S104.

[0094] S201. Obtain a vibration signal of a gearbox in a historical period, and a fault time and fault type of each fault in the historical period.

[0095] S202: based on the fault time of each fault, the vibration signal of the gearbox in the historical period is split to obtain the vibration signal of the time period where each fault time is located.

[0096] S203 , performing wavelet transform and two-dimensional MTF transform based on the vibration signal of the time period at each fault moment to obtain an MTF graph at each fault moment.

[0097] S204 , taking the MTF graph at each fault moment as input and the fault type at each fault moment as output, to generate a plurality of training samples.

[0098] S205 , performing neural network training based on multiple training samples to obtain a WATD-MTF residual network.

[0099] Exemplarily, step S205 can be specifically implemented as steps S2051 - S2052.

[0100] S2051. Based on a convolutional layer, an SKNet module, a residual module, an average pooling layer, and a fully connected layer, construct a WATD - MTF residual network.

[0101] It should be noted that SKNet is an attention - enhanced network that can adaptively select convolutional kernels of different scales for feature extraction. It not only expands the receptive field but also enables the model to obtain multi - scale feature information. The network structure is as Figure 4 shown. The SKNet network mainly includes three modules, namely Split, Fuse, and Select.

[0102] (1) Split operation: Extract features from the input MTF map using convolutional kernels of different sizes. To reduce the occupation of computing resources, in this paper, two convolutional kernels with K = 1 and K = 3 are selected to expand the receptive field. Add and fuse the two parts of features to obtain feature U, as shown in the following formula.

[0103] U = U1 + U2;

[0104] (2) Fuse operation: Perform global average pooling operation Fgp on the fused feature U to generate channel statistical information to obtain S, whose dimension is C×1×1. Sc represents the elements of the C - th channel, as shown in the following formula.

[0105]

[0106] After Fgp, generate a compact feature Z through a fully connected layer. The dimension of vector Z is Z×1×1 (Z < C), as shown in the following formula.

[0107] Z = F fc (s)=δ(β(W×s));

[0108] where F fc is the fully connected layer, δ is the Relu activation function, β represents the BN batch normalization processing layer, and W represents the operation matrix of the fully connected layer.

[0109] (3) Select operation: Expand Z into two vectors with the same dimension as S through a fully connected layer. Then, calculate the weights of each convolutional kernel through Softmax.

[0110]

[0111] where, A, B ∈ R C×d , A c and B c respectively represent the data of the c - th row, Ac , B c ∈R 1×d .a c and b c is the weight value of two different convolution kernels. Since this paper selects two convolution kernels for weighting, a c +b c =1.

[0112] The weight value is weighted with U1 and U2 to obtain the fused feature map A, as shown in the following formula.

[0113] A=A1+A2, A1=a·U1, A2=b·U2;

[0114] Among them, a, b are the weights generated by fusion of different convolution kernels, A = [A1, A2, ...A c ],A∈R H×W .

[0115] It should be noted that the residual network is a variant of the convolutional neural network. It uses the skip connection structure to build a deeper network structure, solves the problem of gradient explosion and vanishing in the traditional convolutional neural network, and reduces the risk of overfitting to a certain extent. The residual module is mainly divided into two categories, as follows: Figure 5 As shown. The residual block structure of the ResNet18 model is Figure 5 As shown in (a), the residual network module with more layers is Figure 5 The (b) figure in the figure has fewer parameters.

[0116] Based on the above analysis, this paper improves the ResNet18 model and integrates the SKNet module into the ResNet-18 structure. After the MTF graph is input, it is first processed by a convolution layer, batch normalization (BN) and ReLU activation function, and then the SKNet module is introduced to enhance the feature extraction process. In the backbone network part, in order not to introduce more parameters and cause model structure redundancy, the original residual structure is retained. After being processed by 4 residual blocks, the SKNet module is embedded for secondary enhancement. By adaptively adjusting the selection of convolution kernels, it ensures that features of different scales can be effectively expressed and utilized. Finally, the model is subjected to global average pooling and a fully connected layer to achieve fault classification.

[0117] Compared with the traditional ResNet18, the improved model shows higher recognition accuracy in fault diagnosis tasks and stronger generalization performance in different working conditions. Figure 6 shown.

[0118] S2052: Based on multiple training samples, the constructed WATD-MTF residual network is trained to obtain a preset WATD-MTF residual network.

[0119] In this way, the embodiment of the present invention can construct a WATD-MTF residual network through the attention module and the improved residual module, and perform training to obtain a preset WATD-MTF residual network, which is convenient for real-time diagnosis of the gearbox and improves the accuracy of gearbox fault diagnosis.

[0120] Exemplarily, the present invention uses an adaptive wavelet threshold denoising method combined with a Markov transfer field method to denoise a one-dimensional signal and generate a two-dimensional feature map as a model input, and combines SKNet with ResNet18 to form an SK-ResNet18 enhanced model. This paper proposes a WATD-MTF-SK-ResNet18 gearbox fault diagnosis method, and the specific steps of the method are as follows:

[0121] Step 1: Build a test bench to collect gear fault data for subsequent signal processing.

[0122] Step 2: Use the wavelet adaptive threshold denoising method and Markov transition field to construct the processed data set for training the improved residual network model.

[0123] Step 3: Divide the Markov two-dimensional image into a training set and a test set in a ratio of 4:1.

[0124] Step 4: Build the SK-ResNet18 network model and initialize the parameters.

[0125] Step 5: Input the MTF graph into the model for training, continuously optimize and update the parameters, and if the optimal value is reached, proceed to step 6; otherwise, continuously update the parameters until convergence, and save the optimal model parameters.

[0126] Step 6: Input the test set into the trained model for testing, and finally obtain the gear fault classification results and classification accuracy.

[0127] According to the analysis of the impact of training parameters on model classification effect, the training Batch Size is set to 32, the test Batch Size is 16, the optimizer selects the Adam adaptive optimizer, the learning rate is 0.0001, and all experiments are trained for 50 rounds. The experimental environment is Windows 11 64-bit operating system, the running memory is 16GB, the network model is built using the Pytorch framework in the software Pycharm, and the hardware is the RTX 3050Laptop GPU graphics card.

[0128] The present invention uses a gearbox failure data set from a university to verify the improved model. The data set includes two sub-datasets, bearings and gears, and both are collected from a transmission system dynamics simulator. The gear failure data set has two working conditions, 20Hz-0v (speed-load) and 30Hz-2v. There are 8 channels of signals in each file, where channels 2, 3, and 4 represent the vibration signals of the planetary gearbox in the x, y, and z directions, respectively. This paper uses the signal in the x direction for analysis. The gear failure data collects a total of 5 gear operating status signals, namely health, chipping, missing, root, and surface.

[0129] The original one-dimensional vibration signal is Figure 7 As shown in Table 1. For different faults, 2000 images are randomly sampled and divided into training set and test set in the ratio of 8:2. The experimental samples of gears with 5 types of faults under different working conditions are shown in Table 1.

[0130] Table 1

[0131]

[0132] Based on the advantages of Markov transfer field method, such as fully retaining time correlation and effectively extracting internal features, the one-dimensional signal of data set A is sampled in two dimensions by 50% resampling. The total number of images generated for each type of fault is 2000, and the image size is 224*224. After being divided proportionally, they are input into the model for training. Different fault MTF images generated by one-dimensional signals before and after denoising.

[0133] In order to fully verify the superiority of the present invention on the gear fault dataset, the accuracy, precision, recall, F1-score and AUC value are used to comprehensively analyze and evaluate the model, where the formulas for accuracy, precision, recall and F1-score are as follows.

[0134]

[0135] Among them, TP is the number of samples correctly predicted as positive, TN is the number of samples correctly predicted as negative, FP is the number of negative samples incorrectly predicted as positive, and FN is the number of positive samples incorrectly predicted as negative.

[0136] The AUC value is defined as the probability that the model predicts that the output is a positive sample, which is greater than the probability of the predicted negative sample. The two most important evaluation indicators are FRP and TPR. TPR represents the probability of correctly classifying a positive sample, and FRP represents the probability of incorrectly classifying a negative sample. These two parameters constitute the horizontal and vertical coordinates of the ROC curve, and the AUC value is equal to the area under the ROC curve. The ideal state is that TPR tends to 1 and FRP tends to 0. At this time, the area of ​​the ROC curve is the largest. Therefore, the larger the AUC value, the better the classification performance of the model.

[0137] Using dataset A, we conducted a comparative analysis of the fault diagnosis of the ResNet18 model before and after denoising. The MTF graph after denoising obviously contains clearer features than before denoising. Figure 8 shown.

[0138] Figure 8 (a) The accuracy curve converges around the 15th step, with the highest accuracy reaching 99%. However, the average test accuracy fails to reach 80%. This is because the data contains a lot of noise, which causes the model to fail to fully learn the fault features, resulting in overfitting. The test loss rate curve oscillates too much, and the loss does not show a decreasing trend. Fig. 9 (c)(d) It can be seen that the model performance has been significantly improved after denoising. The training converges around the 10th step, and the highest accuracy reaches 100%. The accuracy and loss rate curves are basically consistent, which conforms to the gradient descent law. However, there are tip values ​​in the curve during the iteration process. This is because the ResNet model does not pay enough attention to some key features, resulting in insufficient feature extraction.

[0139] The previous section used dataset A to verify the ResNet18 model. Although this model showed a high diagnostic accuracy after denoising, a few cutting-edge values ​​appeared, affecting the overall diagnostic effect. Therefore, this section used the denoised dataset B to verify the robustness of the SK-ResNet18 model under load conditions. By comparing the AlexNet, LeNet-5, ConvNext and ResNet-18 models, the effectiveness of the proposed method was verified.

[0140] The diagnostic effect of the ResNet-18 model is slightly lower than that of the model of the present invention, indicating that the introduction of SKNet can enhance the model's ability to capture important features. The AlexNet and LeNet-5 architectures are relatively simple and it is difficult to effectively capture complex features. The indicators of the ConvNeXt model are relatively low. Due to the complex structure of the model and the slow gradient descent, the fault cannot be effectively identified. The improved model of the present invention has excellent fault identification ability in gear fault diagnosis.

[0141] It should be understood that the order of execution of the steps in the above embodiment does not necessarily mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiment of the present invention.

[0142] The following is an embodiment of the device of the present invention. For details not described in detail therein, reference may be made to the corresponding method embodiment described above.

[0143] Fig. 9 The schematic diagram of the structure of a gearbox fault diagnosis device based on WATD-MTF residual network provided by an embodiment of the present invention is shown. The fault diagnosis device 300 includes a communication module 301 and a processing module 302 .

[0144] The communication module 301 is used to obtain the vibration signal of the gearbox to be tested when it is working;

[0145] The processing module 302 is used to perform adaptive noise reduction on the vibration signal by wavelet transform to obtain a denoised signal; based on the denoised signal, perform a two-dimensional MTF transform to generate an MTF graph, and the MTF graph is used to characterize the time series information of the fault characteristics in the vibration signal; based on the MTF graph and the preset WATD-MTF residual network, predict and obtain the fault diagnosis result of the gearbox to be tested, and the fault diagnosis result includes the fault type and the fault probability.

[0146] Fig.10 Schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. Fig.10 As shown, the electronic device 400 of this embodiment includes: a processor 401, a memory 402, and a computer program 403 stored in the memory 402 and executable on the processor 401. When the processor 401 executes the computer program 403, the steps in the above-mentioned method embodiments are implemented, for example Figure 1 Alternatively, when the processor 401 executes the computer program 403, the functions of each module / unit in the above-mentioned device embodiments are realized, for example, Fig. 9 The functions of the communication module 301 and the processing module 302 are shown.

[0147] Exemplarily, the computer program 403 may be divided into one or more modules / units, which are stored in the memory 402 and executed by the processor 401 to implement the present invention. The one or more modules / units may be a series of computer program instruction segments that can implement specific functions, and the instruction segments are used to describe the execution process of the computer program 403 in the electronic device 400. For example, the computer program 403 may be divided into Fig. 9 The communication module 301 and the processing module 302 are shown.

[0148] The processor 401 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.

[0149] The memory 402 may be an internal storage unit of the electronic device 400, such as a hard disk or memory of the electronic device 400. The memory 402 may also be an external storage device of the electronic device 400, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. equipped on the electronic device 400. Further, the memory 402 may also include both an internal storage unit of the electronic device 400 and an external storage device. The memory 402 is used to store the computer program and other programs and data required by the terminal. The memory 402 may also be used to temporarily store data that has been output or is to be output.

[0150] The embodiments described above are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included in the protection scope of the present invention.

Claims

1. A gearbox fault diagnosis method based on WATD-MTF residual network, characterized in that: include: Obtain the vibration signal of the gearbox to be tested when it is working; Adopting wavelet transform to perform adaptive noise reduction on the vibration signal to obtain a denoised signal; Based on the denoised signal, a two-dimensional MTF transformation is performed to generate an MTF graph, wherein the MTF graph is used to characterize time series information of fault characteristics in the vibration signal; Based on the MTF diagram and the preset WATD-MTF residual network, a fault diagnosis result of the gearbox to be tested is predicted and obtained, and the fault diagnosis result includes a fault type and a fault probability.

2. The gearbox fault diagnosis method based on WATD-MTF residual network according to claim 1 is characterized in that: The method of using wavelet transform to adaptively reduce noise on the vibration signal to obtain a denoised signal comprises: Performing wavelet decomposition on the vibration signal to obtain a low-frequency signal and a high-frequency signal; Based on the low-frequency signal and the high-frequency signal, calculate the noise standard deviation of each layer in the low-frequency signal and the high-frequency signal; Determining an adaptive threshold of each layer based on the noise standard deviation of each layer and the signal length of the vibration signal; Based on the adaptive threshold of each layer, the low-frequency signal and the high-frequency signal are denoised and reorganized to obtain the denoised signal.

3. The gearbox fault diagnosis method based on WATD-MTF residual network according to claim 1 is characterized in that: The step of determining the adaptive threshold of each layer based on the noise standard deviation of each layer and the signal length of the vibration signal comprises: Based on the following formula, the adaptive threshold of each layer is determined; Wherein, N is the signal length of the vibration signal, l j is the adaptive threshold of the i-th layer, s j is the noise standard deviation of the i-th layer.

4. The gearbox fault diagnosis method based on WATD-MTF residual network according to claim 1 is characterized in that: The performing of two-dimensional MTF transformation based on the denoised signal to generate an MTF graph includes: Based on the signal value at each moment in the denoised signal, discretization processing is performed to obtain a value range of the signal value; Dividing the value range into equal parts to obtain a plurality of position dividers; Based on a preset time window, the denoised signal is divided to obtain denoised signals of multiple time windows; Based on the denoised signals of the multiple time windows and the multiple quantizers, probability calculation is performed to construct an MTF diagram for each time window; wherein, the element in the i-th row and j-th column in the MTF diagram of any time window represents the signal value at each moment in the time window, and the probability that the current moment belongs to the i-th quantizer under the condition that the previous moment belonged to the j-th quantizer.

5. The gearbox fault diagnosis method based on WATD-MTF residual network according to claim 1 is characterized in that: Before predicting the fault diagnosis result of the gearbox to be tested based on the MTF diagram and the preset WATD-MTF residual network, the method further includes: Obtain the vibration signal of the gearbox in the historical period, the fault time and fault type of each fault in the historical period; Based on the fault time of each fault, the vibration signal of the gearbox in the historical period is split to obtain the vibration signal of the time period of each fault moment; Based on the vibration signal of the time period at each fault moment, wavelet transform and two-dimensional MTF transform are performed to obtain an MTF graph at each fault moment; Taking the MTF graph at each fault moment as input and the fault type at each fault moment as output, a plurality of training samples are generated; Based on the multiple training samples, neural network training is performed to obtain the WATD-MTF residual network.

6. The gearbox fault diagnosis method based on WATD-MTF residual network according to claim 5 is characterized in that: The performing of neural network training based on the multiple training samples to obtain the WATD-MTF residual network includes: Based on the convolutional layer, SKNet module, residual module, average pooling layer and fully connected layer, the WATD-MTF residual network is constructed; Based on the multiple training samples, the constructed WATD-MTF residual network is trained to obtain the preset WATD-MTF residual network.

7. The gearbox fault diagnosis method based on WATD-MTF residual network according to claim 1 is characterized in that: The method of predicting the fault diagnosis result of the gearbox to be tested based on the MTF diagram and the preset WATD-MTF residual network includes: Based on the MTF graph, attention weight fusion is performed to obtain a fused feature graph; Based on the fused feature map and the preset residual module, the fault diagnosis result is obtained.

8. The gearbox fault diagnosis method based on WATD-MTF residual network according to claim 7 is characterized in that: The attention weight fusion is performed based on the MTF graph to obtain a fused feature graph, including: Based on preset convolution kernels of different sizes, feature extraction is performed on the MTF feature map to obtain multiple convolution features; Perform feature fusion on multiple convolutional features to obtain fused features; Based on the fusion features, a global average pooling operation is performed to generate channel statistical information; generating a compact feature vector based on channel statistics of the plurality of channels; Based on the compact feature vector, calculating the attention weight of each convolution kernel; Based on the multiple convolution features and the attention weight of each convolution kernel, a fused feature map is obtained.

9. A gearbox fault diagnosis device based on WATD-MTF residual network, characterized in that: include: A communication module is used to obtain the vibration signal of the gearbox to be tested when it is working; A processing module, used for performing adaptive noise reduction on the vibration signal by using wavelet transform to obtain a denoised signal; Based on the denoised signal, a two-dimensional MTF transformation is performed to generate an MTF graph, wherein the MTF graph is used to characterize time series information of fault characteristics in the vibration signal; Based on the MTF diagram and the preset WATD-MTF residual network, a fault diagnosis result of the gearbox to be tested is predicted and obtained, and the fault diagnosis result includes a fault type and a fault probability.

10. An electronic device, characterized in that: The electronic device comprises a memory and a processor, the memory stores a computer program, and the processor is configured to execute the steps of the method according to any one of claims 1 to 8 when calling and running the computer program stored in the memory.

Images