Gearbox fault diagnosis method based on domain adaptation and deep residual network

Abstract

The invention discloses a gearbox fault diagnosis method based on domain adaptation and a deep residual network, and the method comprises the steps: collecting vibration signals of a gear under the fault conditions of different working conditions through an acceleration sensor, and building a gear fault data set; converting a one-dimensional vibration signal into a two-dimensional image of which the features are more easily extracted by a deep residual network by adopting a Grubm angle difference field, and dividing the two-dimensional image into a training set and a test set according to a certain proportion; establishing a fault diagnosis model which comprises a feature extractor, a label predictor and a domain classifier, aligning feature distribution differences between a source domain and a target domain by combining the maximum average difference and the domain classifier, and enhancing the data information integration capability of the two domains; inputting the training sets with the labels in the source domain and the target domain into the fault diagnosis model for training to obtain a trained fault diagnosis model; and inputting the label-free test set of the target domain into the trained fault diagnosis model for verification, thereby realizing accurate classification of gear fault types of different target domains.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of gear fault diagnosis, and particularly relates to a gearbox fault diagnosis method based on domain adaptation and a deep residual network. BACKGROUND

[0002] As a core mechanical transmission element, gears are widely used in key industrial fields such as aerospace, wind power, ships, rail transportation and mining machinery, and bear the heavy responsibility of motion and power transmission. However, due to their long-term work in harsh environments such as high temperature, high pressure, impact and overload, fatigue, cracks, wear and broken teeth and other faults inevitably occur. These faults not only seriously damage the performance of mechanical systems and greatly shorten the service life of equipment, but also may cause serious production interruptions or safety accidents. Therefore, developing and deploying an intelligent gear fault diagnosis model suitable for actual industrial needs to achieve early and accurate fault identification is of great strategic significance for reducing unplanned downtime, reducing maintenance costs and improving system operation reliability and safety. This has become a core research direction in the field of mechanical condition monitoring.

[0003] Currently, deep learning has made significant progress in intelligent fault diagnosis, but its wide application is limited by excessive dependence on a large amount of high-quality labeled data. In addition, existing models can usually only process data under a single working condition. In complex and variable actual industrial environments, changes in equipment operating conditions (such as load and speed) will directly cause significant shifts in the feature distribution of collected signals, which will seriously weaken the generalization ability of the model. In the face of new working conditions, traditional methods often require time-consuming retraining of the model, and must perform costly and tedious manual labeling for new target domain data, which undoubtedly causes a huge waste of time and economic cost. Therefore, there is an urgent need to develop an intelligent fault diagnosis model that can adapt to multiple complex working conditions and does not require a large amount of labeled data, in order to greatly improve the practicality and deployment efficiency of the diagnosis system and meet the actual needs of dynamic industrial scenarios. SUMMARY

[0004] In order to solve the problems of low fault diagnosis rate and efficiency of gears under multiple complex working conditions and a small amount of labeled data, a gearbox fault diagnosis method based on domain adaptation and a deep residual network is provided. The deep residual network and attention mechanism are used to extract features from fault data, and JMMD and domain discriminators are used to reduce the feature distribution between the source domain and the target domain, thereby reducing the dependence on labeled data of the target domain, improving the diagnosis efficiency, and achieving accurate classification of gear faults under different working conditions.

[0005] The present application provides a gearbox fault diagnosis method based on domain adaptation and a deep residual network, comprising:

[0006] Step 1: collect the vibration signals of gears in different working conditions by using acceleration sensors, and establish a gear fault data set;

[0007] Step 2: convert the one-dimensional vibration signal into a two-dimensional image that is easier to extract features by a deep residual network by using the GADF, and divide it into a training set and a test set according to a certain proportion;

[0008] Step 3: establish a fault diagnosis model, including a feature extractor, a label predictor and a domain classifier, and align the feature distribution difference between the source domain and the target domain by using the joint maximum mean difference and the domain classifier, and enhance the integration ability of the data information of the two domains;

[0009] Step 4: input the labeled training set in the source domain and the target domain into the fault diagnosis model for training, and obtain the trained fault diagnosis model;

[0010] Step 5: input the unlabeled test set in the target domain into the trained fault diagnosis model for verification, so as to realize accurate classification of the fault types of gears in different target domains.

[0011] The gear box fault diagnosis method based on domain adaptation and deep residual network has the following beneficial effects:

[0012] 1. By converting the one-dimensional vibration signal into a two-dimensional image by using the GADF, the powerful feature learning ability of the two-dimensional convolutional neural network can be fully utilized, the time dependence and interaction inside the signal can be effectively captured and enhanced, and the visualization and interpretability of the features can be enhanced.

[0013] 2. The feature extractor of the model adopts an attention residual network and an attention feature fusion module, which can fully extract the position information of the feature information and fuse the features extracted at each stage.

[0014] 3. By using the JMMD and the domain discriminator in the model, the flexibility of the adversarial learning and the explicitness based on distance measurement can be cooperatively utilized, the limitations of a single method can be overcome, so that more stable, more robust and deeper cross-domain feature alignment can be realized, and the diagnosis performance and generalization ability of the model on different working condition target domains can be significantly improved. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 It is a flow chart of the gear box fault diagnosis method based on domain adaptation and deep residual network;

[0016] Figure 2 It is a process chart for converting one-dimensional signal into two-dimensional image by using GADF;

[0017] Figure 3 Fig. 1 is a schematic diagram of a fault diagnosis model of the present application;

[0018] Figure 4 Fig. 2 is a schematic diagram of a multi-scale feature extraction module;

[0019] Figure 5 Fig. 3 is a schematic diagram of an attention residual module;

[0020] Figure 6 Fig. 4 is a schematic diagram of an improved attention feature fusion module;

[0021] Fig. 7(a) is a confusion matrix diagram of DANN under C-B transfer task;

[0022] Fig. 7(b) is a confusion matrix diagram of Resnet18 under C-B transfer task;

[0023] Fig. 7(c) is a confusion matrix diagram of MoblieNetV3 under C-B transfer task;

[0024] Fig. 7(d) is a confusion matrix diagram of the method of the present application under C-B transfer task. DETAILED DESCRIPTION

[0025] As shown in Figure 1 Fig. 1, a gear box fault diagnosis method based on domain adaptation and deep residual network of the present application comprises:

[0026] Step 1: Use an acceleration sensor to collect vibration signals of gears under different working conditions, and establish a gear fault data set.

[0027] In the implementation, a fault gear box experimental platform is used, which is composed of a CL-100 type gear tester, a dynamic signal testing system, a gear box, and an acceleration sensor. The experimental platform can collect vibration signals of gears under different working conditions, thereby establishing a gear fault data set. In the specific operation, the motor speed is set to 600 r / min, 900 r / min, and 1200 r / min, and the sampling frequency is set to 25.6 Hz. In order to more comprehensively simulate the faults of the gear box in actual industrial production, the healthy gears are pretreated, and the most common gear surface wear and gear surface pitting are obtained.

[0028] Step 2: Use GADF to convert one-dimensional vibration signals into two-dimensional images that are easier to extract features by deep residual network, and divide them into training set and test set according to the ratio of 8:2. Figure 2 The process of converting one-dimensional signals into GADF is shown. Specifically:

[0029] Step 2.1: The original one-dimensional vibration signal data X=(x iX(i) = {x(i) | i = 1, 2,..., N} is a time series with N samples, the original one-dimensional vibration signal data X is normalized to scale the data range to the interval [-1, 1]:

[0030]

[0031] Step 2.2: Convert the scaled time series data from the Cartesian coordinate system to the polar coordinate system, where Encoding as the cosine of the angle Time node t i Encoding as the radius r i The conversion process is as follows:

[0032]

[0033] Step 2.3: Since Then Monotonically changes in the interval [0, π], and further makes the encoding of the time series in the polar coordinate a unique mapping relationship, under the encoding scheme of step 2.2, the time evolution direction follows the path from the upper left to the lower right, where the main diagonal elements contain the time sequence information of the original gear vibration signal, and the non-diagonal line area reveals the correlation between signals at different times, where represents the angle value of the i-th sample. The two-dimensional image defined by the Gram angle difference field conversion is as follows:

[0034]

[0035] Where, represents the angle value of the i-th sample.

[0036] The classification of category labels and samples of experimental data sets under different working conditions is shown in Table 1.

[0037] Table 1. Classification of category labels and samples of experimental data sets

[0038]

[0039] Step 3: Establish a fault diagnosis model, as shown in Figure 3 , including a feature extractor, a label predictor, and a domain classifier, and by jointly maximizing the average difference and aligning the feature distribution difference between the source domain and the target domain, the domain classifier enhances the integration ability of the data information of the two domains.

[0040] The feature extractor of the fault diagnosis model in step 3 is an improved deep residual network, the input is the generated two-dimensional image, including a multi-scale feature extraction module, multiple attention residual modules, and an attention feature fusion module. The specific structure of the multi-scale feature extraction module is as follows: Figure 4As shown, it is composed of three parallel convolutional layers with kernel sizes of 7x7, 9x9 and 11x11, which increases the receptive field of the network by using multi-scale large kernel convolution, captures more feature information of different scales and fuses them. The padding parameters of each convolution kernel are set to 3, 4 and 5, and a BachNorm layer is introduced after the convolutional layer to improve the generalization ability. For the input X GADF , it is first passed through the multi-scale feature extraction module, and the expression of the output of each branch is as follows:

[0041] G1=B(DwConv 7×7 (X GADF )) (4)

[0042] G2=B(DwConv 9×9 (X GADF )) (5)

[0043] G3=B(DwConv 11×11 (X GADF )) (6)

[0044] Where DwConv 7×7 , DwConv 9×9 and DwConv 11×11 represent large kernel depth separable convolution with kernel sizes of 7x7, 9x9 and 11x11 respectively, B represents the BN layer, G j (j=1,2,3) represents the output of the jth branch, and finally the outputs of each branch are fused to obtain the multi-scale feature as:

[0045] F=Concat[G1,G2,G3,dim=1] (7)

[0046] In the formula, F is the output of the multi-scale feature extraction module, and dim=1 represents connection in the channel dimension.

[0047] The fused feature information is input into four attention residual modules in series, and the specific structure is as shown in Figure 5 , which is composed of an improved residual module and an attention branch, which is used to better extract information obtained from various sensor signals and aggregate position information in each direction to improve the generalization performance of the model. A down-sampling layer composed of a BN layer and a convolutional layer with a step of 2 and a kernel of 2x2 is inserted between adjacent attention residual modules. For the multi-scale feature F, the feature extraction process of each improved residual module is as follows:

[0048] H1=δ r1 (f r1 (B(DwConv 9×9 (F)))) (8)

[0049] H2=δ r2 ((B(f r2 (H1)))) (9)H1 and H2 represent the outputs after dimensionality increase and dimensionality reduction, respectively, δ r1 δ r2 f represents the GELU activation function. r1 f r2 This represents a 1×1 convolutional layer, used to increase and decrease the number of channels, respectively.

[0050] In the attention branch of the attention residual module, two pooling kernels (H, 1) and (1, W) with different spatial ranges are used to aggregate the features of F along the horizontal and vertical coordinates, respectively.

[0051]

[0052] Among them, z w (w) and z h (h) The outputs of F at height h and width w are respectively, and their channel number is reduced to a lower dimension. Then, the normalized feature maps of the two channels are obtained by using depthwise separable convolution and the Sigmoid function.

[0053] g h =σ h1 (DWConv 3×3 (δ h (f h (z h (h))))) (12

[0054] g w =σ w1 (DWConv 3×3 (δ w (f w (z w (w))))) (13

[0055] σ w1 σ h1 δ represents the Sigmoid activation function. w δ h f represents the GELU activation function. w f h This represents two 1×1 convolutional layers, DWConv 3×3 It is a depthwise separable convolution with a kernel size of 3×3; normalized features are treated as feature-level weight maps to enhance feature representation; in addition, to preserve original information, a residual strategy is used to combine the enhanced features with the original features in two directions:

[0056]

[0057] Here represents Hadamard product, and then the fused information is extracted by using 3x3 depth separable convolution:

[0058]

[0059] where σ h2 , σ w2 represent Sigmoid activation function, and then the output v is obtained by re-weighting H2 using the output of attention branch:

[0060] v = H2 x s h x s w (18)

[0061] Finally, the output of each attention residual module is obtained by using 1x1 convolution layer:

[0062] O = B(f r3 (v)) + F (19)

[0063] where f r3 is 1x1 convolution, O e R C×H×W .

[0064] The feature information extracted by each attention residual module is fused, and the structure is shown in Figure 6 : for the feature O1 extracted by the current attention residual module and the feature O2 extracted by the last attention residual module, the intermediate feature K is obtained after the addition operation:

[0065] K = O1 + O2 (20)

[0066] Then K is respectively subjected to global feature extraction and local feature extraction, wherein the global feature extraction needs to be subjected to global average pooling operation first, and the channel number is reduced to C / r by using point-wise dot product operation, r is the channel reduction rate; and the local feature extraction does not need to be subjected to global average pooling;

[0067] K1 = λ(B(PWConv1(GAP(K))) (21)

[0068] K2 = λ(B(PWConv2(K))) (22)

[0069] where λ is ReLU activation function, GAP represents global average pooling layer; PWConv1 and PWConv2 represent point-wise convolution; then K1 and K2 are connected in channel dimension by using Concat operation, and the fused feature is extracted by using point-wise dot product, and then the weight matrix is obtained by using Sigmoid function:

[0070] K' = sigma (PWConv3 (Concat (K1, K2, dim = 1))) (23)

[0071] Wherein, sigma indicates Sigmoid function, PWConv3 indicates pointwise convolution, and dim indicates splicing in channel dimension;Finally, the fusion result of the attention feature fusion module is obtained:

[0072]

[0073] The label predictor is composed of a full connection layer, and the output result of the source domain is transformed by a Softmax function to make it conform to the probability distribution of different fault conditions of the class, and the label classification loss is as follows:

[0074]

[0075] Wherein, L c (θ f ,θ c ) is the label prediction loss, G y Indicates the label predictor, is the feature representation of the i-th sample obtained by the feature extractor parameter θ f , is the real label index of the i-th sample.

[0076] The domain classifier improves the effect of domain adaptation by structuring the feature information of two domains;It promotes the mutual confrontation of two domain data by confusing the fault feature information between two domains, until the domain classifier cannot distinguish the distribution of two domains, and then completes the alignment training of two domain data, and the domain classification loss is as follows:

[0077]

[0078] Wherein, L d (θ f ,θ d ) is the domain classification loss, G d Indicates the domain classifier, and G f Is a feature extractor, And Respectively indicate the sample features from different source domains and target domains, d i Is the domain label.

[0079] The present application adopts joint maximum average difference to carry out distribution adaptation, and improves the diagnostic performance under the cross-domain scene. Specifically, the mapping function is introduced to embed the high-level features of two domains into two points in the reproducing kernel Hilbert space, and the distance between the two points is used to represent the distribution distance between the two domain features, and the matching of the distribution of the data of two domains is realized by minimizing the distance, and the distribution loss of the feature representation between two domains is defined as LJMMD , the calculation method is as follows:

[0080]

[0081] Wherein, n s , n t respectively the number of samples of the source domain and the target domain, z i sl , z j tl respectively represent the feature representation of the i, j sample of the source domain and the target domain in the lth feature subspace, k l is the lth kernel function.

[0082] Therefore, the total classification loss is:

[0083] L MDARDAJM = L c + λ 1 L JMMD - λ 2 L d (28)

[0084] Wherein, λ 1 is the weighting parameter of the joint maximum average difference loss function, λ 2 is the adaptive weighting parameter of the domain discriminator loss function.

[0085] Step 4: input the labeled training set in the source domain and the target domain into the fault diagnosis model for training, and obtain the trained fault diagnosis model.

[0086] Step 5: input the unlabeled test set of the target domain into the trained fault diagnosis model for verification, so as to realize accurate classification of different target domain gear fault types.

[0087] The accuracy Accuray is used as the model performance evaluation index, and the specific calculation formula is as follows:

[0088]

[0089] Wherein, TP represents true positive; FP represents false positive; TN represents true negative; FN represents false negative.

[0090] In specific implementation, the motor speed is set as condition A at 600r / min, condition B at 900r / min, and condition C at 1200r / min during the experiment. In order to better reflect the superiority of the performance of the method, three popular models DANN, Resnet18 and MobileNetV3 are selected for comparison experiment, and table 2 shows the performance of the migration task of different models.

[0091] Table 2. Comparison of migration task accuracy of different models

[0092]

[0093] It can be seen that the accuracy of the method proposed in the present application is better than that of the remaining models in any migration condition, and the accuracy reaches 100% in A-B, B-A, C-B tasks, which can accurately identify various faults. In order to more clearly visualize the proportion of different fault type samples being misclassified, FIGS. 7(a), (b), (c), (d) show the confusion matrix diagrams of the four methods in the C-B migration task, and it can be clearly seen that DANN misclassifies most of the fault types with label 0;

[0094] Resnet18 has certain misclassification for label 2; MobileNetV3 classifies the three fault types relatively accurately, but there are different degrees of misclassification; and the method proposed in the present application classifies the three fault types accurately, which reflects its superior performance.

[0095] The above only describes the preferred embodiments of the present application and is not used to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A gearbox fault diagnosis method based on domain adaptation and deep residual network, characterized in that, The application relates to a method for realizing fault type classification of gears in different target domains. Step 1: collecting vibration signals of gears in different working conditions by using an acceleration sensor to establish a gear fault data set; Step 2: converting one-dimensional vibration signals into two-dimensional images which are easier to be extracted by a deep residual network by using a Gram angle difference field, and dividing the two-dimensional images into a training set and a test set according to a certain proportion; Step 3: establishing a fault diagnosis model, including a feature extractor, a label predictor and a domain classifier, and aligning feature distribution differences between a source domain and a target domain by using a joint maximum average difference and the domain classifier to enhance the data information integration capability of the two domains; Step 4: inputting the labeled training set in the source domain and the target domain into the fault diagnosis model for training to obtain a trained fault diagnosis model; Step 5: inputting the unlabeled test set in the target domain into the trained fault diagnosis model for verification, so that the fault type of gears in different target domains can be accurately classified.

2. The gearbox fault diagnosis method based on domain adaptation and deep residual network according to claim 1, characterized in that, The step 2 is specifically as follows: Step 2.1: Assume that the original one-dimensional vibration signal data is X = (x i , i = 1, 2,..., N) is a time series with N samples, normalize the original one-dimensional vibration signal data X, and scale the data range to the [-1, 1] interval: Step 2.2: Transform the scaled time series data from Cartesian coordinate system to polar coordinate system, where Encode as angle cosine Time node t i Encode as radius r i The transformation process is as follows: Step 2.3: Since then monotonically in the interval [0, π], and thus the encoding of time series in polar coordinates is a unique mapping relationship. Under the encoding scheme of Step 2.2, the time evolution direction follows the path from the upper left to the lower right, where the main diagonal elements contain the time sequence information of the original gear vibration signal, and the off-diagonal area reveals the correlation between signals at different times. The two-dimensional image defined by the Gram angle difference field conversion is as follows: wherein, represents the angle value of the i-th sample. 3.The gearbox fault diagnosis method based on domain adaptation and deep residual network according to claim 1, characterized in that, The feature extractor of the fault diagnosis model in the step 3 is an improved deep residual network, the input is the generated two-dimensional image, and the improved deep residual network comprises a multi-scale feature extraction module, a plurality of attention residual modules and an attention feature fusion module; For input X GADF , first pass through the multi-scale feature extraction module, whose expression is as follows: G1 = B(DwConv 7×7 (X GADF )) (4) G2 = B(DwConv 9×9 (X GADF )) (5) G3 = B(DwConv 11×11 (X GADF )) (6) Among them, DwConv 7×7 ,DwConv 9×9 and DwConv 11×11 These represent deep separable convolutions with kernel sizes of 7×7, 9×9, and 11×11, respectively. B represents a BN layer, and G... j (j = 1, 2, 3) represents the output of the j-th branch. Finally, the outputs of each branch are fused to obtain the multi-scale features: F = Concat [G1, G2, G3, dim = 1] (7) In the formula, F is the output of the multi-scale feature extraction module, and dim = 1 represents that connection is performed in the channel dimension; The fused multi-scale features are input into the plurality of attention residual modules, and for the multi-scale features F, the feature extraction process of each improved residual module is as follows: H1 = δ r1 (f r1 (B(DwConv 9×9 (F)))) (8) H2= δ r2 ((B(f r2 (H1)))) (9) H1, H2 represent the output after dimensionality increase and dimensionality decrease respectively, δ r1 , δ r2 represents the GELU activation function, f r1 , f r2 represents a 1x1 convolution layer, respectively used to increase and decrease the number of channels; In the attention branch of the attention residual module, two different spatial range pooling kernels (H, 1) and (1, W) are used to aggregate the features of F along the horizontal coordinate and the vertical coordinate respectively; where z w (w) and z h (h) respectively output at height h and width w, then reduce their channel number to lower dimension, and obtain two channel normalized feature maps using depthwise separable convolution and Sigmoid function; g h = σ h1 (DWConv 3×3 (δ h (f h (z h (h))))) (12) g w = σ w1 (DWConv 3×3 (δ w (f w (z w (w))))) (13) σ w1 , σ h1 denotes the Sigmoid activation function, δ w , δ h denotes the GELU activation function, f w , f h denotes two 1x1 convolutional layers, DWConv 3×3 is a depth separable convolution with a kernel size of 3x3; the normalized features are regarded as feature-level weight maps to enhance the feature representation, and a residual strategy is adopted to combine the enhanced features with the original features in two directions; Here denotes the Hadamard product, followed by feature extraction using a 3x3 depthwise separable convolution on the fused information: where σ h2 , σ w2 denotes a Sigmoid activation function, and the output v is obtained by re-weighting H2 using the output of the attention branch. v = H2x s h x s w (18) Finally, a 1*1 convolution layer is used to obtain the output of each attention residual module: O = B(f r3 (v))+F (19) where f r3 is a 1 x 1 convolution, O e R C×H×W ; The feature information O extracted by each attention residual module is fused: for the to-be-fused feature O1 and O2, intermediate features K are obtained: K = O1 + O2 (20) Then, global feature extraction and local feature extraction are performed on K respectively; K1 = lambda (B (PWConv1 (GAP (K))) (21) K2 = lambda (B (PWConv2 (K))) (22) Here, lambda is a ReLU activation function, GAP represents a global average pooling layer, and PWConv1 and PWConv2 represent point-wise convolution; then, K1 and K2 are connected in the channel dimension by using a Concat operation, and point-wise dot product is used to extract and fuse the features, and then a Sigmoid function is used to obtain a weight matrix: K' = sigma (PWConv3 (Concat (K1, K2, dim = 1))) (23) Wherein, sigma represents a Sigmoid function, PWConv3 represents point-wise convolution, and dim represents that the connection is performed in the channel dimension; finally, the fusion result of the attention feature fusion module is obtained:

4. The gearbox fault diagnosis method based on domain adaptation and deep residual network according to claim 1, characterized in that, The label predictor is composed of a full connection layer, the output result of the source domain is transformed by using a Softmax function, so that the probability distribution of different fault conditions of the class is met, and the label classification loss is as follows: where L c (θ f ,θ c ) is the label prediction loss, G y denotes the label predictor, is the feature representation of the i-th sample obtained by the feature extractor parameter θ f , and is the true label index of the i-th sample.

5. The gearbox fault diagnosis method based on domain adaptation and deep residual network according to claim 4, characterized in that, The field classifier improves the effect of field self-adaption by structurally processing the two-field feature information; and it promotes the mutual confrontation of the two-field data by confusing the fault feature information between the two fields, until the field classifier cannot distinguish the distribution of the two fields, and then completes the alignment training of the two-field data, and the field classification loss is as follows: Among them, L d (θ f ,θ d G represents the domain classification loss. d G represents the domain classifier. f It is a feature extractor. and Let d represent the sample features from different source and target domains, respectively. i It is a domain tag.

6. The gearbox fault diagnosis method based on domain adaptation and deep residual network according to claim 5, characterized in that, Adopting joint maximum mean discrepancy for distribution adaptation, a mapping function is introduced to embed the high-level features of two domains into two points in the reproducing kernel Hilbert space, and the distance between the two points is used to represent the distribution distance between the features of two domains. The matching of the distribution of data in two domains is realized by minimizing the distance. The distribution loss of the features between two domains is defined as L JMMD , and the calculation is as follows: where n s , n t are the number of samples in the source domain and the target domain, respectively, z i sl , z j tl denote the feature representation of the i-th, j-th sample in the source domain and the target domain in the l-th feature subspace, respectively, k l is the l-th kernel function; The total classification loss is obtained as follows: L MDARDAJM = L c + λ1L JMMD - λ2L d (28) Wherein, λ1 is the weighting parameter of the joint maximum average difference loss function, and λ2 is the adaptive weighting parameter of the domain discriminator loss function.

7. The gearbox fault diagnosis method based on domain adaptation and deep residual network according to claim 1, characterized in that, The accuracy Accuray is used as the model performance evaluation index in step 5, and the specific calculation formula is as follows: Wherein, TP represents the true positive example; FP represents the false positive example; TN represents the true negative example; and FN represents the false negative example.

Images