A gearbox fault diagnosis method based on domain adaptation and deep residual network

Abstract

The application discloses a gearbox fault diagnosis method based on field adaptation and a deep residual network, and comprises the following steps: collecting vibration signals of gears in different working conditions by using an acceleration sensor to establish a gear fault data set; converting one-dimensional vibration signals into two-dimensional images which are more easily extracted by a deep residual network by using a Grahm angle difference field, and dividing the two-dimensional images into a training set and a test set according to a certain proportion; establishing a fault diagnosis model, including a feature extractor, a label predictor and a field classifier, and aligning feature distribution differences between a source field and a target field by using a joint maximum average difference and the field classifier to enhance the data information integration capability of the two fields; inputting the training set with labels in the source field and the target field into the fault diagnosis model to train the fault diagnosis model, and obtaining a trained fault diagnosis model; inputting the test set without labels in the target field into the trained fault diagnosis model to verify the trained fault diagnosis model, so that accurate classification of different target field gear fault types is realized.

Description

Technical Field

[0001] This invention relates to the field of gear fault diagnosis technology, and specifically to a gearbox fault diagnosis method based on domain adaptation and deep residual networks. Background Technology

[0002] Gears, as core mechanical transmission components, are widely used in key industrial fields such as aerospace, wind power generation, shipbuilding, rail transportation, and mining machinery, bearing the heavy responsibility of transmitting motion and power. However, due to their long-term operation in harsh environments such as high temperature, high pressure, impact, and overload, they inevitably experience various failures such as fatigue, cracks, wear, and tooth breakage. These failures not only severely damage the performance of mechanical systems and significantly shorten equipment lifespan, but may also lead to serious production interruptions or safety accidents. Therefore, developing and deploying intelligent gear fault diagnosis models suitable for actual industrial needs to achieve early and accurate fault identification is of crucial strategic significance for reducing unplanned downtime, cutting maintenance costs, and improving system 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 the field of intelligent fault diagnosis; however, its widespread application is limited by its over-reliance on large amounts of high-quality labeled data. Furthermore, existing models typically can only handle data under a single operating condition. In complex and ever-changing real-world industrial environments, changes in equipment operating conditions (such as load and speed) directly lead to significant shifts in the feature distribution of the acquired signals, severely weakening the model's generalization ability. Faced with new operating conditions, traditional methods often require time-consuming model retraining and expensive, tedious manual labeling of new target domain data, undoubtedly resulting in a huge waste of time and economic resources. Therefore, there is an urgent need to develop an intelligent fault diagnosis model that can adaptively handle multiple complex operating conditions without requiring large amounts of labeled data, to significantly improve the practicality and deployment efficiency of diagnostic systems and meet the actual needs of dynamic industrial scenarios. Summary of the Invention

[0004] To address the issue of low fault diagnosis rate and efficiency of gears under complex operating conditions with limited labeled data, a gearbox fault diagnosis method based on domain adaptation and deep residual networks is proposed. This method extracts features from fault data using deep residual networks and attention mechanisms, and utilizes JMMD and a domain discriminator to reduce the feature distribution between the source and target domains. This reduces the dependence on labeled data in the target domain, improves diagnostic efficiency, and enables accurate classification of gear faults under different operating conditions.

[0005] This invention provides a gearbox fault diagnosis method based on domain adaptation and deep residual networks, comprising:

[0006] Step 1: Use an accelerometer to collect vibration signals of the gear under different working conditions and fault conditions to establish a gear fault dataset;

[0007] Step 2: The one-dimensional vibration signal is converted into a two-dimensional image that is easier for deep residual networks to extract features using the Gram angle difference field, and then divided into training set and test set according to a certain ratio;

[0008] Step 3: Establish a fault diagnosis model, including a feature extractor, a label predictor, and a domain classifier. Furthermore, enhance the integration capability of data information between the two domains by aligning the feature distribution differences between the source and target domains through the joint maximum average difference and domain classifier.

[0009] Step 4: Input the labeled training sets from the source and target domains into the fault diagnosis model for training to obtain the trained fault diagnosis model.

[0010] Step 5: Input the unlabeled test set of the target domain into the trained fault diagnosis model for verification, thereby achieving accurate classification of gear fault types in different target domains.

[0011] The gearbox fault diagnosis method based on domain adaptation and deep residual networks of the present invention has the following beneficial effects:

[0012] 1. By utilizing Gram angular difference field (GADF) to convert one-dimensional vibration signals into two-dimensional images, the powerful feature learning capabilities of two-dimensional convolutional neural networks can be fully utilized, effectively capturing and enhancing the temporal dependence and interaction within the signal, and enhancing the visualization and interpretability of features.

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

[0014] 3. By incorporating Joint Maximum Mean Difference (JMMD) and a domain discriminator into the model, the flexibility of adversarial learning and the explicitness of distance-based metrics can be synergistically utilized to overcome the limitations of single methods. This results in more stable, robust, and deeper cross-domain feature alignment, significantly improving the model's diagnostic performance and generalization ability across target domains under different operating conditions. Attached Figure Description

[0015] Figure 1 This is a flowchart of a gearbox fault diagnosis method based on domain adaptation and deep residual network according to the present invention;

[0016] Figure 2 This is a diagram illustrating the process of converting a one-dimensional signal into a two-dimensional image using GADF.

[0017] Figure 3 This is a schematic diagram of the fault diagnosis model of the present invention;

[0018] Figure 4 This is a schematic diagram of the multi-scale feature extraction module;

[0019] Figure 5 This is a schematic diagram of the attention residual module;

[0020] Figure 6 A schematic diagram of the improved attention feature fusion module;

[0021] Figure 7(a) shows the confusion matrix of DANN in the CB transfer task;

[0022] Figure 7(b) shows the confusion matrix of ResNet18 under the CB migration task;

[0023] Figure 7(c) shows the confusion matrix of MobileNetV3 under the CB transfer task;

[0024] Figure 7(d) is a confusion matrix diagram of the method of the present invention under the CB migration task. Detailed Implementation

[0025] like Figure 1 As shown, a gearbox fault diagnosis method based on domain adaptation and deep residual networks according to the present invention includes:

[0026] Step 1: Use an accelerometer to collect vibration signals of the gear under different working conditions and establish a gear fault dataset.

[0027] In practical implementation, a faulty gearbox experimental platform was used, consisting of a CL-100 gear testing machine, a dynamic signal testing system, a gearbox, and an acceleration sensor. This experimental platform can collect vibration signals of the gears under different operating conditions, thereby establishing a gear fault dataset. During specific operation, the motor speed was set to 600 r / min, 900 r / min, and 1200 r / min, and the sampling frequency was set to 25.6 Hz. To more comprehensively simulate gearbox faults in actual industrial production, healthy gears underwent preprocessing, resulting in the two most common fault types: tooth surface wear and tooth surface pitting.

[0028] Step 2: Use Gram Angular Difference Field (GADF) to convert the one-dimensional vibration signal into a two-dimensional image that is easier for deep residual networks to extract features from, and divide it into training and test sets in an 8:2 ratio. Figure 2 This demonstrates the process of converting a one-dimensional signal into a GADF. Specifically:

[0029] Step 2.1: The original one-dimensional vibration signal data X = (x) measured by the experimental platform iLet X be a time series with N samples (i = 1, 2, ..., N). Normalize the original one-dimensional vibration signal data X to scale the data range to the interval [-1, 1].

[0030]

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

[0032]

[0033] Step 2.3: Due to but The time series monotonically changes within the interval [0,π], thus making the encoding of the time series on polar coordinates a unique mapping. Under the encoding scheme in step 2.2, the direction of time evolution follows a path from the upper left to the lower right. The main diagonal elements contain the temporal information of the original gear vibration signal, while the off-diagonal regions reveal the correlation between signals at different times. This represents the angle value of the i-th sample. The two-dimensional image of the Gram angular difference field transformation is defined as follows:

[0034]

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

[0036] Table 1 shows the category labels and sample classifications of the experimental datasets under different working conditions.

[0037] Table 1. Category labels and sample classification of the experimental dataset

[0038]

[0039] Step 3: Establish a fault diagnosis model, such as Figure 3 As shown, it includes a feature extractor, a label predictor, and a domain classifier, and enhances the ability to integrate data information between the source and target domains by aligning the feature distribution differences between the source and target domains through the joint maximum average difference and the domain classifier.

[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, and it includes 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, the network consists of three parallel convolutional layers with kernel sizes of 7×7, 9×9, and 11×11. By employing multi-scale large kernel convolutions, the receptive field of the network is increased, capturing and fusing more feature information at different scales. The padding parameters for each convolutional kernel are set to 3, 4, and 5, respectively. A BachNorm layer is introduced after the convolutional layers to improve generalization ability. For input X... GADF First, the feature extraction module performs a multi-scale feature extraction process. The output of each branch is expressed 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] 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:

[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 indicates that a connection is made in the channel dimension.

[0047] The fused feature information is then input into four cascaded attention residual modules, the specific structure of which is as follows: Figure 5 As shown, it consists of improved residual modules and attention branches, used to better extract information from various sensor signals and aggregate positional information from various directions, improving the model's generalization performance. Downsampling layers consisting of BN layers and convolutional layers with a stride of 2 and a kernel size of 2×2 are inserted between adjacent attention residual modules. For multi-scale features 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 This represents the Hadamard product, and then a 3×3 depthwise separable convolution is used to extract features from the fused information:

[0058]

[0059] Where σ h2 σ w2 This represents the Sigmoid activation function, and then the output of the attention branch is used to reweight H2 to obtain the output v:

[0060] v = H2 × s h ×s w (18)

[0061] Finally, a 1×1 convolutional layer is used to obtain the output of each attention residual module:

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

[0063] Among them, f r3 For a 1×1 convolution, O∈R C×H×W .

[0064] Feature fusion is performed on the feature information extracted from each attention residual module, and its structure is as follows: Figure 6 As shown: For the feature O1 extracted by the current attention residual module and the feature O2 extracted by the previous attention residual module, the intermediate feature K is obtained after addition:

[0065] K = O1 + O2 (20)

[0066] Then, global feature extraction and local feature extraction are performed on K respectively. Global feature extraction requires global average pooling and pointwise dot product operation to reduce the number of channels to C / r, where r is the channel reduction rate. Local feature extraction does not require global average pooling.

[0067] K1=λ(B(PWConv1(GAP(K)))) (21)

[0068] K2=λ(B(PWConv2(K))) (22)

[0069] Here, λ is the ReLU activation function, GAP refers to the global average pooling layer; PWConv1 and PWConv2 both represent pointwise convolution; next, the Concat operation is used to connect K1 and K2 along the channel dimension, and pointwise dot products are used to extract fused features. Finally, the Sigmoid function is used to obtain the weight matrix.

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

[0071] Where σ represents the Sigmoid function, PWConv3 represents pointwise convolution, and dim represents concatenation along the channel dimension; the final result of the attention feature fusion module is obtained:

[0072]

[0073] The label predictor consists of fully connected layers. The source domain output is transformed using a Softmax function to conform to the probability distribution of different failure states of the class. The label classification loss is as follows:

[0074]

[0075] Among them, L c (θ f ,θ c G represents the label prediction loss. y Represents the label predictor. The i-th sample is obtained through the feature extractor parameters θ f The obtained feature representation, It is the true label index of the i-th sample.

[0076] The domain classifier improves domain adaptation by structuring the feature information of the two domains. It obfuscates fault feature information between the two domains, causing the data to compete against each other until the domain classifier can no longer distinguish the distribution between the two domains, thus completing the alignment training of the two domain data. The domain classification loss is as follows:

[0077]

[0078] 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.

[0079] This invention employs joint maximum mean difference for distribution adaptation, improving diagnostic performance in cross-domain scenarios. Specifically, a mapping function is introduced to embed high-level features from two domains into two points in the regenerating kernel Hilbert space. The distance between these two points represents the distribution distance between the features of the two domains. Matching of the data distributions of the two domains is achieved by minimizing this distance. The distribution loss of the feature representation between the two domains is defined as L.JMMD The calculation method is as follows:

[0080]

[0081] Where, n s n t These are the number of samples in the source domain and the target domain, respectively. i sl z j tl Let k represent the feature representations of the i-th and j-th samples in the source and target domains, respectively, in the l-th feature subspace. l It is the l-th kernel function.

[0082] Therefore, the total classification loss is:

[0083] L MDARDAJM =L c +λ1L JMMD -λ2L d (28)

[0084] Where λ1 is the trade-off parameter of the joint maximum average difference loss function, and λ2 is the adaptive trade-off parameter of the domain discriminator loss function.

[0085] Step 4: Input the labeled training sets from the source and target domains into the fault diagnosis model for training to 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, thereby achieving accurate classification of gear fault types in different target domains.

[0087] Accuracy is used as the performance evaluation metric for the model, and the specific calculation formula is as follows:

[0088]

[0089] Where TP represents a true positive; FP represents a false positive; TN represents a true negative; and FN represents a false negative.

[0090] In specific implementation, the motor speed was set to 600 r / min as condition A, 900 r / min as condition B, and 1200 r / min as condition C during the experiment. To better demonstrate the superior performance of the proposed method, three popular models—DANN, ResNet18, and MobileNetV3—were selected for comparative experiments. Table 2 shows the transfer task performance of different models.

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

[0092]

[0093] As can be seen, the method proposed in this invention outperforms other models in accuracy under all transfer scenarios, achieving 100% accuracy in AB, BA, and CB tasks, and accurately identifying various faults. To more clearly visualize the proportion of different fault type samples that are misclassified, Figures 7(a), (b), (c), and (d) show the confusion matrix diagrams of the four methods under the CB transfer task. It can be clearly seen that DANN misclassifies most fault types with a label of 0.

[0094] ResNet18 misclassifies label 2 to some extent; MobileNetV3 classifies the three fault types relatively accurately, but also exhibits varying degrees of misclassification; while the method proposed in this paper accurately classifies the three fault types, demonstrating its superior performance.

[0095] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A gearbox fault diagnosis method based on domain adaptation and deep residual networks, characterized in that, include: Step 1: Use an accelerometer to collect vibration signals of the gear under different working conditions and fault conditions to establish a gear fault dataset; Step 2: The one-dimensional vibration signal is converted into a two-dimensional image that is easier for deep residual networks to extract features using the Gram angle difference field, and then divided into training set and test set according to a certain ratio; Step 3: Establish a fault diagnosis model, including a feature extractor, a label predictor, and a domain classifier. Furthermore, enhance the integration capability of data information between the two domains by aligning the feature distribution differences between the source and target domains through the joint maximum average difference and domain classifier. The feature extractor of the fault diagnosis model is an improved deep residual network. The input is a generated two-dimensional image, which includes a multi-scale feature extraction module, multiple attention residual modules and attention feature fusion module. For input First, it goes through a multi-scale feature extraction module, the expression of which is as follows: (4) (5) (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: (7) In the formula, F is the output of the multi-scale feature extraction module, and dim=1 indicates that a connection is made in the channel dimension; The fused multi-scale features are input into multiple attention residual modules. For the multi-scale feature F, the feature extraction process of each improved residual module is as follows: (8) (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 These represent 1×1 convolutional layers, used to increase and decrease the number of channels, respectively. 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. (10) (11) in, and The outputs of F at height h and width w are respectively. Then, their channel number is reduced to a lower dimension, and the normalized feature maps of the two channels are obtained by using depthwise separable convolution and the Sigmoid function. (12) (13) 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 regarded as feature-level weight maps to enhance feature representation, and a residual strategy is adopted to combine the enhanced features with the original features in two directions; (14) (15) here This represents the Hadamard product, and then a 3×3 depthwise separable convolution is used to extract features from the fused information: (16) (17) Where σ h2 σ w2 This represents the Sigmoid activation function, and then the output of the attention branch is used to reweight H2 to obtain the output v: (18) Finally, a 1×1 convolutional layer is used to obtain the output of each attention residual module: (19) Among them, f r3 For a 1×1 convolution, O ; For the feature information O extracted from each attention residual module, feature fusion is performed: For the features to be fused O1 and O2, the intermediate feature K is obtained: (20) Then, global feature extraction and local feature extraction are performed on K respectively; (21) (22) here The ReLU activation function is used, and GAP refers to the global average pooling layer. PWConv1 and PWConv2 both represent pointwise convolutions. The Concat operation is then used to connect K1 and K2 along the channel dimension, and pointwise dot products are used to extract fused features. Finally, the Sigmoid function is used to obtain the weight matrix. (23) Where σ represents the Sigmoid function, PWConv3 represents pointwise convolution, and dim represents concatenation along the channel dimension; the final result of the attention feature fusion module is obtained: (24) Step 4: Input the labeled training sets from the source and target domains into the fault diagnosis model for training to obtain the trained fault diagnosis model. Step 5: Input the unlabeled test set of the target domain into the trained fault diagnosis model for verification, thereby achieving accurate classification of gear fault types in different target domains.

2. The gearbox fault diagnosis method based on domain adaptation and deep residual networks according to claim 1, characterized in that, Step 2 specifically involves: Step 2.1: Assume the original one-dimensional vibration signal data is as follows: For 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]. (1) Step 2.2: Convert the scaled time series data from Cartesian coordinates to polar coordinates, where... Encoded as angle cosine φ i Time node t i Encoded as radius r i The conversion process is as follows: (2) Step 2.3: Due to , then φ i The time series monotonically changes within the interval [0, π], thus making the encoding of the time series on polar coordinates a unique mapping relationship. Under the encoding scheme in step 2.2, the time evolution direction follows a path from the upper left to the lower right. The main diagonal elements contain the temporal information of the original gear vibration signal, while the off-diagonal regions reveal the correlation between signals at different times. The two-dimensional image of the Gram angle difference field conversion is defined as follows: (3) Where, φ i This represents the angle value of the i-th sample.

3. The gearbox fault diagnosis method based on domain adaptation and deep residual networks according to claim 1, characterized in that, The label predictor consists of fully connected layers. The source domain output is transformed using a Softmax function to conform to the probability distribution of different failure states of the class. The label classification loss is as follows: (25) in, Predict loss for labels, Represents the label predictor. The i-th sample is obtained through the feature extractor parameters θ f The obtained feature representation, It is the true label index of the i-th sample.

4. The gearbox fault diagnosis method based on domain adaptation and deep residual networks according to claim 3, characterized in that, The domain classifier improves domain adaptation by structuring the feature information of the two domains. It obfuscates fault feature information between the two domains, causing the data to compete against each other until the domain classifier can no longer distinguish the distribution between the two domains, thus completing the alignment training of the two domain data. The domain classification loss is as follows: (26) in, For domain classification loss, Represents a domain classifier. It is a feature extractor. and These represent sample features from different source and target domains, respectively. It is a domain tag.

5. The gearbox fault diagnosis method based on domain adaptation and deep residual networks according to claim 4, characterized in that, Distribution adaptation is achieved using joint maximum mean difference. A mapping function is introduced to embed high-level features from two domains into two points in the regenerating kernel Hilbert space. The distance between these two points represents the distribution distance between the features of the two domains. Matching of the data distributions of the two domains is achieved by minimizing this distance. The distribution loss of the feature representation between the two domains is defined as... The calculation method is as follows: (27) Where, n s n t These are the number of samples in the source domain and the target domain, respectively. i sl z j tl Let k represent the feature representations of the i-th and j-th samples in the source and target domains, respectively, in the l-th feature subspace. l It is the l-th kernel function; Therefore, the total classification loss is: (28) Where λ1 is the trade-off parameter of the joint maximum average difference loss function, and λ2 is the adaptive trade-off parameter of the domain discriminator loss function.

6. The gearbox fault diagnosis method based on domain adaptation and deep residual networks according to claim 1, characterized in that, In step 5, accuracy is used as the model performance evaluation metric, and the specific calculation formula is as follows: (29) Where TP represents a true positive; FP represents a false positive; TN represents a true negative; and FN represents a false negative.

Images