Intelligent fault diagnosis method for spiral bevel gear under unbalanced data set

Abstract

The invention relates to a spiral bevel gear intelligent fault diagnosis method under an unbalanced data set, and the method comprises the following steps: collecting operation vibration signals of a spiral bevel gear in different health states, and constructing a data set; constructing an enhanced data set based on a data enhancement algorithm, fusing maximum pooling and average pooling operations to improve a deep auto-encoder, and completing characteristic pre-training of samples based on unsupervised learning; a pre-trained encoder module is adopted to process the unbalanced training data set, and data samples are mapped to a nonlinear feature space; constructing a classifier model based on a full connection layer to retrain the feature representation of the training data set, and constructing a loss function by adopting a segmented adjustable balance factor; and introducing an independent test data set on the basis of the verification set to carry out performance evaluation. The problem that the diagnosis performance is reduced due to the fact that an existing spiral bevel gear intelligent fault diagnosis method is difficult to guarantee extraction of effective fault information is solved, higher diagnosis precision is achieved, and good robustness and generalization are shown.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of cross-domain fault diagnosis, and particularly relates to a data-driven intelligent fault diagnosis method for spiral bevel gears under an unbalanced data set. BACKGROUND

[0002] In recent years, spiral bevel gears are widely used in fans, helicopters and aero-engines, and are developing towards large-scale, high-speed and precision. Long-term work in high temperature, high speed, heavy load and strong impact complex working conditions, spiral bevel gears will inevitably produce various fault modes. If the fault mode cannot be accurately detected and timely repaired, it may affect the overall performance and even cause major safety accidents. Compared with traditional spur gears and parallel gears, the vibration signal of spiral bevel gears in the meshing process has stronger non-stationarity, and the impact response caused by damage is more weak, which leads to that the fault feature extraction of spiral bevel gears is more dependent on advanced signal processing technology and researchers' diagnosis experience. In addition, the special contact mode and complex transmission path of spiral bevel gears make the dynamic modeling of various fault modes extremely challenging. Therefore, it is of great significance to study the end-to-end intelligent fault diagnosis method of spiral bevel gears.

[0003] However, spiral bevel gears are mostly in non-fault state, which leads to that the number of fault samples is much smaller than that of healthy samples. The unbalanced distribution of samples will lead to that the traditional deep learning end-to-end diagnosis model is difficult to learn strong robustness of generalization features from small data samples, and the diagnosis result is easy to be inclined to large sample healthy data, so it is necessary to pay attention to the intelligent fault diagnosis research based on deep learning under unbalanced data set. Some scholars respectively adopt data enhancement technology and fault sample loss weighting to solve the problem of insufficient spiral bevel gear fault samples, for example, adopting autoencoder, variational autoencoder, generating antithetical network to generate new fault samples, adopting Focal loss loss function to give greater punishment to fault sample error classification, etc., which have been widely applied to the fault diagnosis research of spiral bevel gears under unbalanced data, but there are still some problems to be improved: (1) generating fault samples by adopting methods such as training autoencoder and generating antithetical network will consume a lot of time cost, which does not meet the real-time demand of industrial monitoring. In addition, the diversity of generated samples depends on the number and quality of original samples, and it is still difficult to get high accuracy by relying on the traditional end-to-end training method to enhance the data set. (2) In actual problems, reducing the misclassification cost of small samples (referring to the loss or negative consequences brought by the classification model when the sample is incorrectly divided into a non-real category.) by using an improved loss function can make the training process continuous, but directly processing the original sample and trying to extract effective features still has challenges, especially when the designed classification model does not have strong robustness. Therefore, when the training data set is unbalanced, new technologies need to be introduced to accurately and efficiently diagnose various fault modes of spiral bevel gears. SUMMARY

[0004] In view of the problem that the current intelligent fault diagnosis method of spiral bevel gears relies on the traditional end-to-end intelligent training framework, the loss function is modified to balance the classification cost and the data enhancement technology is difficult to ensure the extraction of effective fault information, thereby leading to the decline of the diagnosis performance in the spiral bevel gear fault diagnosis task, an intelligent fault diagnosis method of spiral bevel gears based on data driving under unbalanced data set is provided, which is used for accurately and efficiently identifying each fault mode of the spiral bevel gears when the training data set is unbalanced; a two-stage intelligent diagnosis framework is provided, the feature learning and mode classification processes are decoupled by using data enhancement and loss weighting, high-quality nonlinear feature representations of each fault mode are learned, the improved cross-entropy function is adjusted in sections to fine-tune the feature representations, the attention to the small fault samples is improved, the decision boundary is effectively smoothed, the robustness of the model is improved, higher diagnosis accuracy is realized in the unbalanced spiral bevel gear fault diagnosis and identification task, and good robust generalization is displayed, so that the technical problems involved in the background technology can be solved.

[0005] The technical scheme of the present application is as follows:

[0006] An intelligent fault diagnosis method of spiral bevel gears under unbalanced data set comprises the following steps:

[0007] Step S1, collect vibration signals of spiral bevel gears in different health states, and construct training data set, test data set and verification data set;

[0008] Step S2, construct an enhanced data set based on a data enhancement algorithm, fuse the maximum pooling and average pooling operations to design an improved deep autoencoder, and complete the characteristic pre-training of the sample based on unsupervised learning;

[0009] Step S3, process the unbalanced training data set by using the pre-trained encoder module, and map the data samples to a nonlinear feature space;

[0010] Step S4, construct a classifier model based on a full connection layer to retrain the feature representation of the training data set, and an improved cross-entropy loss function is constructed by using a segmented adjustable balance factor in the training process;

[0011] Step S5, introduce an independent test data set based on the verification set to evaluate the performance.

[0012] Further, step S2 specifically comprises the following steps:

[0013] Step S21, calculate the nearest k samples near the small fault sample based on the Euclidean distance of the n-dimensional feature space, generate new fault samples by using the synthetic minority oversampling technique, and construct an enhanced data set:

[0014]

[0015] wherein s represents the s-th feature dimension, respectively represent the value of the sample x i , x j in the s-th feature dimension, represent a new sample x i , x j generated by synthesis, i x j represents a fault data sample with a dimension of n, Dis(x i , x i ) represents the n-dimensional space Euclidean distance between two fault samples, argsort() represents sorting the Euclidean distance, and k nearest neighbors of x ξ is a random number between 0 and 1.

[0016] Step S22, based on the constructed enhanced data set, an improved deep autoencoder is designed by fusing the max-pooling and average-pooling operations, and an unsupervised learning algorithm is used to learn the nonlinear high-dimensional feature representation of the sample.

[0017]

[0018] wherein N represents the total number of samples, x i represents the i-th sample in the enhanced data set, represents the i-th reconstructed sample, is the mean square error loss of the enhanced data set.

[0019] Further, step S3 specifically includes the following steps:

[0020] Step S31, using a pre-trained encoding module in the improved deep convolutional autoencoder to process the unbalanced training data set.

[0021]

[0022] wherein pre_feature represents the feature map output by the convolution module, M represents the number of samples, F represents the number of convolution kernels, and Q represents the number of input feature channels, represents the feature map of the c-th channel of the i-th sample in the unbalanced training set, respectively represent the weight and bias matrix of the c-th channel of the f filters in the encoding module, and σ represents the ReLU activation function.

[0023] Further, step S4 specifically includes the following steps:

[0024] Step S41, a linear classifier is constructed to train the feature representation extracted by the encoder in the previous stage, and the parameter learning of the classifier is completed; the training in this stage takes the feature of the training set as the input, and takes the label as the supervision signal, and adopts the cross-entropy loss function to optimize the weight parameter; the output layer uses the Softmax activation function to convert the output vector into a class probability distribution, which is used for multi-class classification; the feature representation of the training data set is retrained;

[0025] Step S42, an improved cross-entropy loss function is constructed by using a segmented adjustable factor, and the feature representation is fine-tuned by using the linear classifier;

[0026]

[0027] f m =1 / (1+exp(-a m ))

[0028]

[0029] Wherein, class represents the total number of classes, a m represents the relative weight coefficient of the mth class sample, num represents the total number of all training samples, MCE represents the weighted cross-entropy loss, n m represents the number of samples of the mth health state, f m represents the weight penalty coefficient applied to the mth class sample, y tm and y pm respectively represent the real probability distribution and the predicted probability distribution of the mth class sample, represents the probability value of the predicted class, and γ is a segmented adjustable factor.

[0030] Further, step S5 is specifically: in order to verify the generalization performance of the intelligent fault recognition model trained under the condition of unbalanced data, an independent test data set is introduced on the basis of the verification set to evaluate the performance; the process includes standardization preprocessing of input samples, model weight loading, feature extraction based on deep network and fault class prediction.

[0031] Further, in step S1, the number of healthy samples in the training data set constructed is more than the number of fault samples.

[0032] Further, in step S5, the recognition effect of the model on different fault types is comprehensively evaluated through the accuracy, recall rate and F1-score performance indicators.

[0033] Further, the test data set is new data that does not participate in training and parameter tuning, and is used to measure the generalization ability and stability of the model in the actual application scene.

[0034] Preferably, in step S2, the improved deep convolutional autoencoder pre-training model is improved, and the hyperparameter settings are as follows: the batch size is 28, the initial learning rate is 0.0001, the maximum number of iterations epoch is set to 500 times, the mean square error loss is used as the cost loss function in the training process, and the Adam optimizer is used as the optimization algorithm.

[0035] Preferably, in step S4, the classifier model is constructed based on the full connection layer, and the hyperparameter settings are as follows: the batch size is 32, the initial learning rate is 0.001, the maximum number of iterations epoch is set to 100 times, the segmented adjustable improved cross-entropy loss is used as the cost loss function in the training process, and the Adam optimizer is used as the optimization algorithm.

[0036] The present application has the following advantages:

[0037] The high-quality distribution of the data set is efficiently learned, the attention to the small sample features of the fault is deepened, the decision boundary is effectively smoothed, and the recognition accuracy of the fault sample is improved. The method achieves higher diagnosis precision in the unbalanced spiral bevel gear fault diagnosis and recognition task, and has good robustness and generalization, which is superior to the prior art. Specifically, in step S2, the improved deep convolutional autoencoder is constructed based on global pooling and average pooling for information coding, which considers the mining of global information and local information, and effectively reflects the feature essence of the diagnosis signal. In addition, the mean square error loss is used for pre-training of the enhanced data set, which has the effect of pre-extracting effective signal features. In step S3, the pre-trained improved deep convolutional autoencoder coding module is used to map the unbalanced training set to generate high-dimensional features, which improves the nonlinear expression capability of the sample. In step S4, the segmented adjustable improved cross-entropy loss function is used to balance the classification loss cost of the fault sample features and the healthy sample features, effectively smooth the decision boundary, and improve the attention to the small sample of the fault. In addition, different loss functions are used in the feature pre-extraction and feature fine-tuning processes, which utilize the different properties of each other to improve the diagnosis performance of the neural network. The present application can realize high-precision classification of each fault mode of the spiral bevel gear under the condition of highly unbalanced training data set, and has more excellent diagnosis performance. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 The flowchart of the present application is a new framework for intelligent fault diagnosis of spiral bevel gears based on data-driven under unbalanced data set.

[0039] Figure 2 The experimental platform diagram of the present application is a spiral bevel gear box vibration signal collection experiment platform.

[0040] Figure 3 The present application is a multi-classification confusion matrix diagram of the fault diagnosis method for spiral bevel gear box fault diagnosis. DETAILED DESCRIPTION

[0041] The application will be described in detail below with reference to the drawings and specific embodiments. The embodiments are implemented on the premise of the technical scheme of the application, and detailed implementation modes and specific operation processes are given, but the protection scope of the application is not limited to the following embodiments.

[0042] A data-driven intelligent fault diagnosis method for spiral bevel gears under an unbalanced data set, as shown in the formula (1), comprises the following steps: Figure 1

[0043] Step S1, collect vibration signals of spiral bevel gears in different health states, construct training data set, test data set and validation data set, and the number of health samples in the training data set constructed based on this is much larger than the number of fault samples;

[0044] Step S2, based on the data enhancement algorithm, construct an enhanced data set, and construct an improved deep convolutional autoencoder based on unsupervised learning to complete the characteristic pre-training of the sample;

[0045] Step S2 specifically comprises the following steps:

[0046] Step S21, based on the Euclidean distance of the n-dimensional feature space, calculate the nearest k samples near the fault sample, generate a new fault sample by using the synthetic minority over-sampling technique, and construct an enhanced data set:

[0047]

[0048] wherein s represents the value of the s-th feature dimension, respectively represent the value of the s-th feature dimension of the sample x i , x j in the s-th feature dimension, x i , x j represents a fault data sample with a dimension of n, Dis(x i , x j ) represents the n-dimensional space Euclidean distance between two fault samples, argsort() represents sorting the Euclidean distance, and selecting k nearest neighbors of x i ξ is a random number between 0 and 1;

[0049] Step S22, based on the constructed enhanced data set, fuse the maximum pooling and average pooling operations to design an improved deep autoencoder, and use an unsupervised learning algorithm to learn the nonlinear high-dimensional feature representation of the sample;

[0050]

[0051] wherein N represents the total number of samples, x i ​​represents the i-th sample in the augmented dataset, represents the i-th reconstructed sample, is the mean square error loss of the augmented dataset;

[0052] Step S3, using the pre-trained encoder module to process the imbalanced training dataset, and mapping the data samples to a nonlinear feature space;

[0053] Step S3 specifically comprises the following steps:

[0054] Step S31, using the pre-trained improved deep convolutional autoencoder to process the imbalanced training dataset;

[0055]

[0056] wherein pre_feature represents the feature map output by the convolutional module, M represents the number of samples, F represents the number of convolutional kernels, and Q represents the number of input feature channels, represents the i-th sample in the imbalanced training set, and represents the feature map of the c-th channel of the i-th sample in the imbalanced training set, respectively represent the weight and bias matrix of the c-th channel of the f filters of the encoding module, and σ represents the ReLU activation function;

[0057] Step S4, constructing a classifier model to retrain the feature representation of the training dataset, and using a segmented adjustable balance factor to construct an improved cross-entropy loss function in the training process;

[0058] Step S4 specifically comprises the following steps:

[0059] Step S41, constructing a linear classifier to train the feature representation extracted by the previous stage encoder and complete the parameter learning of the classifier. The training in this stage takes the feature of the training set as input and the label as the supervision signal, and uses the cross-entropy loss function to optimize the weight parameters. The output layer uses the Softmax activation function to convert the output vector into a class probability distribution, which is used for multi-class classification; and the feature representation of the training dataset is retrained;

[0060] Step S42, using a segmented adjustable factor to construct an improved cross-entropy loss function, and using a linear classifier to fine-tune the feature representation;

[0061]

[0062] f m = 1 / (1+exp(-a m ))

[0063]

[0064] wherein class represents the total number of classes, am represents the relative weight coefficient of the mth class sample, num represents the total number of all training samples, MCE represents the weighted cross-entropy loss, n m represents the number of samples of the mth class health state, f m represents the weight penalty coefficient applied to the mth class sample, y tm and y pm respectively represent the real probability distribution and the predicted probability distribution of the mth class sample, represents the probability value of the predicted class, and γ is a piecewise adjustable factor.

[0065] Step S5, in order to verify the generalization performance of the intelligent fault recognition model trained under the unbalanced data condition, the present application introduces an independent test data set for performance evaluation on the basis of the verification set. The process includes standardization preprocessing of input samples, model weight loading, feature extraction based on a deep network, and fault class prediction. Finally, the recognition effect of the model on different fault types is comprehensively evaluated through performance indicators such as accuracy (Accuracy), recall (Recall), and F1-score. The test data set, which is new data not involved in training and parameter tuning, is mainly used to measure the generalization ability and stability of the model in actual application scenarios.

[0066] Embodiment

[0067] Referring to Figure 1 , the content of the present application can be mainly divided into four parts. The first part is to collect vibration acceleration signals of spiral bevel gearboxes in different health states, divide them into training data sets, test data sets and verification data sets after mean removal and normalization, and the number of health samples in the training data set constructed based on this is much more than the number of fault samples; the second part is to design an improved deep convolutional autoencoder by fusing average pooling and maximum pooling, to preprocess the enhanced data set, and then to map the training data samples to feature representations using the encoding module; the third part is to design a piecewise adjustable improved cross-entropy loss function, to construct a simple classifier based on a fully connected neural layer for fine-tuning of sample features, and the fourth part is to use the verification data set to test the diagnosis performance of the trained spiral bevel gear intelligent fault diagnosis new framework.

[0068] In the embodiment, the original vibration signals are collected using a spiral bevel gear fault simulation test bench, and are divided into training samples, test samples and verification samples. In order to simulate industrial noise interference, -5dB Gaussian white noise is added to the collected original vibration signals. The test bench is composed of a variable speed drive motor, a coupler, a bevel gear box, a brake, etc. (see Figure 2The motor speed is constant at 600 rpm, the sampling frequency is 10240 Hz, and the three-axis vibration acceleration sensor installed on the box is used to measure the vibration acceleration in x, y and z directions after the motor speed is stable. The spiral bevel gear box fault simulation test bench simulates 8 groups of different health states of working signals, including health state, and 20% crack of bevel gear, 30% crack of bevel gear, 40% crack of bevel gear, pitting of bevel gear, missing tooth of bevel gear, missing block of bevel gear and wear of bevel gear.

[0069] An improved deep convolutional autoencoder pre-training model based on average pooling and maximum pooling is constructed, and the hyperparameters are set as follows: the batch size is 28, the initial learning rate η is 0.0001, the maximum number of iterations epoch is set to 500 times, the mean square error loss is used as the cost loss function in the training process, and the Adam optimizer is used as the optimization algorithm. Table 1 shows the network structure and parameter setting of the improved deep convolutional autoencoder pre-training model used in the application (the convolutional layer step is 2 by default).

[0070] Table 1 Improved deep convolutional autoencoder network structure parameters

[0071]

[0072] The pre-trained improved deep autoencoder encoding module is used to process the unbalanced training data set to obtain the nonlinear feature representation of the training sample; a classifier model is constructed based on the full connection layer, and the hyperparameters are set as follows: the batch size is 32, the initial learning rate η is 0.001, the maximum number of iterations epoch is set to 100 times, the segmented adjustable improved cross-entropy loss is used as the cost loss function in the training process, and the Adam optimizer is used as the optimization algorithm; Table 2 shows the network result parameters of the classifier model used in the application; Table 3 shows the comparison results of the intelligent identification methods of the application method and the data enhancement method, the loss weighting method, the data enhancement method (1) using the nearest neighbor algorithm to expand the sample data, (2) using the K-means class and the synthetic minority over-sampling technique to expand the sample data, and the loss weighting method (1) using the improved Focal loss loss function, (2) using the cost-sensitive loss function.

[0073] Table 2 Classifier model structure parameters

[0074] Functional layer Parameter setting Activation function Flattening layer Input size: 20x64, output size: 1280x1 / First fully connected layer Input size: 1280x1, output size: 128x1 Swish Second fully connected layer Input size: 128x1, output size: 32x1 Swish Classification layer Input size: 32x1, output size: 8x1 Softmax

[0075] Table 3 Comparison of prediction results

[0076] Fault identification method Input form Ten-time average recognition accuracy The method of the present invention Normalized original time-domain vibration signal 98.37％ Data enhancement method (1) Normalized original time-domain vibration signal 86.17％ Data enhancement method (2) Normalized original time-domain vibration signal 88.46％ Loss weighting method (1) Normalized original time-domain vibration signal 89.26％ Loss weighting method (2) Normalized original time-domain vibration signal 85.25％

[0077] The deep learning intelligent diagnostic framework used in the data augmentation and loss weighting methods has the same number of convolutional layers and network structure parameters as the framework of this invention. The input is the same as the method of this invention: the original time-domain vibration signal with added Gaussian white noise. Specifically, analyzing the identification results of one of the 10 runs of the method of this invention, the overall classification accuracy of intelligent fault identification is 98.37%, and the multi-class confusion matrix is ​​as follows: Figure 3 As shown in Table 2 and Figure 3 As shown, the method proposed in this invention has the highest average recognition accuracy and can accurately predict the fault states of spiral bevel gears based on noisy original vibration time-domain signals when the training set is unbalanced.

[0078] Reference Figure 1 As shown, the spiral bevel gearbox fault simulation test bench has two rolling bearing supports at both ends of the bevel gear. In this coupled model, the motion of each component of the spiral bevel gearbox system interacts and couples with each other.

[0079] Reference Figure 3 As shown, the multi-class confusion matrix diagram for testing and identification using the method of this invention is presented. The horizontal axis represents the predicted state label, and the vertical axis represents the true state label. The numbers on the main diagonal represent the recognition accuracy for each category. Labels 0-7 represent bevel gear samples with 20% cracking, 30% cracking, 40% cracking, pitting, missing teeth, missing blocks, wear, and healthy condition, respectively.

[0080] The above-described embodiments are merely one implementation of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention should be determined by the appended claims.

Claims

1. A spiral bevel gear intelligent fault diagnosis method under an unbalanced data set, characterized by, The method comprises the following steps: Step S1, collecting spiral bevel gear vibration signals under different health conditions to construct a training data set, a test data set and a verification data set; Step S2, constructing an enhanced data set based on a data enhancement algorithm, fusing a maximum pooling and an average pooling operation to design an improved deep autoencoder, and completing sample characteristic pre-training based on unsupervised learning; Step S3, processing the unbalanced training data set by using the pre-trained encoder module to map the data samples to a nonlinear feature space; Step S4, constructing a classifier model based on a full connection layer to retrain the feature representation of the training data set, and constructing an improved cross-entropy loss function by using a segmented adjustable balance factor in the training process; Step S5, introducing an independent test data set based on the verification set to evaluate the performance.

2. The intelligent fault diagnosis method for spiral bevel gears under imbalanced data set according to claim 1, characterized in that, Step S2 specifically comprises the following steps: Step S21, calculating the nearest k samples near the small fault samples based on the Euclidean distance of the n-dimensional feature space, generating new fault samples by using a synthetic minority oversampling technique, and constructing an enhanced data set: where s represents the s-th feature dimension, respectively represent the sample x i , x j the value on the s-th feature dimension, represent a new sample x i , x j represent a fault data sample with a dimension of n, Dis(x i , x j ) represents an n-dimensional space Euclidean distance between two fault samples, argsort() represents sorting the Euclidean distance, and k nearest neighbors of x i ξ is a random number between 0 and 1;​ Step S22, based on the constructed enhanced data set, fusing a maximum pooling and an average pooling operation to design an improved deep autoencoder, and learning the nonlinear high-dimensional feature representation of the samples by using an unsupervised learning algorithm; where N denotes the total number of samples, x i represents the i-th sample in the enhanced dataset, represents the i-th reconstructed sample, is the mean squared error loss of the enhanced dataset.

3. The intelligent fault diagnosis method for spiral bevel gears under imbalanced data set according to claim 1, characterized in that, Step S3 specifically comprises the following steps: Step S31, processing the unbalanced training data set by using the pre-trained encoder module in the improved deep convolutional autoencoder; wherein pre_feature represents the feature map output by the convolution module, M represents the number of samples, F represents the number of convolution kernels, and Q represents the number of channels of the input feature, represents the feature map of the cth channel of the ith sample of the unbalanced training set, respectively represent the weight and bias matrices on the cth channel of the f filters of the encoding module, and σ represents the ReLU activation function.

4. The intelligent fault diagnosis method for spiral bevel gears under imbalanced data set according to claim 1, characterized in that, Step S4 specifically comprises the following steps: Step S41, constructing a linear classifier to train the feature representation extracted by the encoder in the previous stage and complete the parameter learning of the classifier; the training in this stage takes the feature of the training set as the input and the label as the supervision signal, and uses a cross-entropy loss function to optimize the weight parameters; the output layer uses a Softmax activation function to convert the output vector into a class probability distribution for multi-class classification; the feature representation of the training data set is retrained; Step S42, using a segmented adjustable factor to construct an improved cross-entropy loss function, and fine-tuning the feature representation by using a linear classifier; wherein class represents the total number of classes, a m represents the relative weight coefficient of the mth class sample, num represents the total number of all training samples, MCE represents the weighted cross-entropy loss, n m represents the number of samples of the mth class health status, f m represents the weight penalty coefficient applied to the mth class sample, y tm and y pm respectively represent the real probability distribution and the predicted probability distribution of the mth class sample, represents the probability value of the predicted class, and γ is a piecewise adjustable factor.

5. The intelligent fault diagnosis method for spiral bevel gears under imbalanced data set according to claim 1, characterized in that, Step S5 specifically is: to evaluate the generalization performance of the intelligent fault recognition model trained under the unbalanced data condition, an independent test data set is introduced based on the verification set to evaluate the performance; this process includes standardization preprocessing of input samples, model weight loading, feature extraction based on a deep network and fault class prediction.

6. The intelligent fault diagnosis method for spiral bevel gears under imbalanced data set according to claim 1, characterized in that, In step S1, the number of healthy samples in the constructed training data set is more than the number of fault samples.

7. The intelligent fault diagnosis method for spiral bevel gears under imbalanced data set according to claim 1, characterized in that, In step S5, the recognition effect of the model on different fault types is comprehensively evaluated through accuracy, recall rate and F1-score performance indicators.

8. The intelligent fault diagnosis method of spiral bevel gears under imbalanced dataset according to claim 1, characterized in that, The test data set is new data that does not participate in training and parameter tuning, and is used to measure the generalization ability and stability of the model in actual application scenarios.

9. The intelligent fault diagnosis method for spiral bevel gears under imbalanced data set according to claim 1, characterized in that, In step S2, the improved deep convolutional autoencoder pre-training model is as follows: the batch size is 28, the initial learning rate η is 0.0001, the maximum number of iterations epoch is set to 500 times, the mean square error loss is used as the cost loss function in the training process, and the Adam optimizer is used as the optimization algorithm.

10. The intelligent fault diagnosis method of spiral bevel gears under imbalanced dataset according to claim 1, characterized in that, In step S4, the classifier model is constructed based on the full connection layer, and the hyperparameters are set as follows: the batch size is 32, the initial learning rate η is 0.001, the maximum number of iterations epoch is set to 100 times, the segmented adjustable improved cross-entropy loss is used as the cost loss function in the training process, and the Adam optimizer is used as the optimization algorithm.

Images