A method and system for diagnosing a slight damage of a tapered roller bearing

Abstract

The application discloses a kind of weak damage diagnosis method and system of tapered roller bearing, belong to fault diagnosis technique and signal processing analysis technical field.The method of the present application proposes to obtain two-dimensional time-frequency image by introducing dynamic change window length short-time Fourier transform, this method has the advantages of weakening background interference, enhancing abnormal vibration signal characteristics, and easy to calculate, vibration signal characteristics are comprehensive, and the degree of self-adaptation is obvious;On this basis, the time-frequency image under different conditions is analyzed by using the generative adversarial network model, and the sample is expanded, which solves the problem that the fault characteristics of the weak damage vibration signal of the tapered roller bearing are difficult to extract;Further, in view of the strict requirements of convolutional neural network for training samples, while considering the subtle differences in sample size and the problem of fixed weight, the present application proposes a tapered roller bearing weak damage recognition model, which realizes intelligent detection of abnormal state of tapered roller bearing.

Description

Technical Field

[0001] This invention relates to a method and system for diagnosing minor damage in tapered roller bearings, belonging to the fields of fault diagnosis technology and signal processing and analysis technology. Background Technology

[0002] Tapered roller bearings, capable of withstanding both radial and axial loads, are widely used in the support systems of various rotating machinery. Failure in these bearings can severely impact the safe operation of the equipment. For raceway spalling failures in tapered roller bearings, classic methods such as envelope analysis can be used for effective diagnosis. However, in the early stages of raceway failure, the fault often manifests as non-through, localized spalling, with weak vibration signals, making direct diagnosis using classic methods difficult. Therefore, it is necessary to develop diagnostic methods for weak damage in tapered roller bearings.

[0003] Short-Time Fourier Transform (STFT) is an effective time-frequency analysis tool capable of clearly decomposing signals and extracting spectral features, thus intuitively revealing potential anomalous components. However, STFT faces the challenge of choosing the right window length when extracting weak anomalous features. Meanwhile, Convolutional Neural Networks (CNNs), as powerful deep learning models, are particularly well-suited for processing image data, automatically learning different types of features and improving detection accuracy and efficiency. However, CNNs also encounter the challenge of incomplete feature extraction in fault diagnosis. To address these issues, improving the feature representation of signals and the extraction of intrinsic features is crucial for accurately assessing the condition of tapered roller bearings. Summary of the Invention

[0004] This invention provides a method and system for diagnosing minor damage in tapered roller bearings, which solves the problem that the fault characteristics of the vibration signal of minor damage in tapered roller bearings are weak and difficult to extract, and realizes the diagnosis of minor damage in tapered roller bearings.

[0005] The technical solution of this invention is:

[0006] According to a first aspect of the present invention, a method for diagnosing minor damage in tapered roller bearings is provided, comprising the following steps:

[0007] Step 1: Collect acceleration vibration signals of tapered roller bearings and construct accelerator vibration signal datasets under different conditions;

[0008] Step 2: Extract time-frequency features from the vibration signal based on the short-time Fourier transform with a dynamically changing window length, and obtain two-dimensional time-frequency images of the acceleration vibration signal of the first tapered roller bearing under different states to obtain an acceleration vibration signal image sample set;

[0009] Step 3: Construct a generative adversarial network model. Train the model based on the acceleration vibration signal image sample set obtained in Step 2. Then use the trained model to generate a two-dimensional time-frequency image of the acceleration vibration signal of the second tapered roller bearing, which serves as an expanded image sample set of acceleration vibration signal.

[0010] Step 4: Merge the augmented image sample set of acceleration vibration signals into the image sample set of acceleration vibration signals to obtain a mixed sample set of acceleration vibration signals; preset the training, testing and validation ratios, and divide the mixed sample set of acceleration vibration signals into a training set, a test set and a validation set;

[0011] Step 5: Construct a weak damage identification model for tapered roller bearings for classification and identification;

[0012] Step 6: Input the training set obtained in Step 4 into the tapered roller bearing weak damage identification model constructed in Step 5 for training, and then perform hyperparameter adjustment based on the results of the validation set to obtain the trained tapered roller bearing weak damage identification model.

[0013] Step 7: Input the time-frequency images of the test set / tapered roller bearing to be diagnosed obtained in Step 4 into the trained tapered roller bearing weak damage identification model obtained in Step 6 for testing / diagnosis.

[0014] Furthermore, the states in the accelerator vibration signal data set under different states are divided into five categories: single-point center damage, through-damage, circumferential edge damage, circumferential center damage, and normal.

[0015] Furthermore, step 2 specifically includes:

[0016] For the acceleration vibration signal samples in the accelerator vibration signal dataset under different states constructed in step 1, the data segments are divided according to the dynamically changing window length to obtain multiple short time frame signals.

[0017] Perform a Fast Fourier Transform on each short time frame signal to obtain the corresponding spectrum; combine the spectrum results of all frames to obtain a two-dimensional time-frequency image of the acceleration vibration signal of the first tapered roller bearing.

[0018] Furthermore, the initial window length is set to L1, and when n>1, the window length L is dynamically adjusted according to the energy distribution. n The rules are as follows:

[0019] If E(n-1) > E high Set window length L n =L min Where n>1;

[0020] If E(n-1) <Elow Set window length L n =L max Where n>1;

[0021] If E low ≤E(t)≤E high The window length is linearly interpolated based on the energy.

[0022]

[0023] In the formula, E hgih E low These represent the maximum and minimum preset energy threshold values, respectively; L max L min These represent the maximum and minimum values ​​of the preset sliding window length threshold, respectively; E(n-1) represents the energy of the (n-1)th short-time frame signal; L n The window length is the nth short time frame signal.

[0024] Furthermore, step 3 specifically includes:

[0025] Construct a generative adversarial network model, wherein the constructed generative adversarial network model includes a generator and a discriminator;

[0026] The parameters of the generative adversarial network model are adjusted according to the image resolution;

[0027] The established model is trained based on the acceleration vibration signal image sample set obtained in step 2, and then the trained model is used to generate a two-dimensional time-frequency image of the acceleration vibration signal of the second tapered roller bearing, which serves as an expanded image sample set of acceleration vibration signal.

[0028] Furthermore, the generator employs four convolutional layers with a training stride of 1 to 2. The resolution of the generated samples is equal to the image resolution in the acceleration vibration signal image sample set from step 2. The generator loss function Loss1 is calculated as follows:

[0029]

[0030] In the formula, N is the number of samples in each training iteration, and p j Let D represent the j-th original sample in each training iteration, and let D denote the discriminator. D(p) j ) represents the original sample p j The probability of it being true;

[0031] The discriminant loss function Loss2 is calculated using the following formula:

[0032]

[0033] In the formula, pj Let y represent the j-th original sample in each training iteration. j For the j-th generated sample generated by the generator each time it is trained, D(y) j ) represents the probability that the generated sample is fake, and λ represents the hyperparameter that controls the strength of the gradient penalty; This represents the value of the gradient penalty. p j and y j The uniform sampling is performed, and ε is obtained by random sampling from a uniform distribution. express L2 norm, For the discriminator pair The gradient of the output result.

[0034] Furthermore, the tapered roller bearing weak damage identification model includes multiple convolutional layers, multiple activation layers, multiple pooling layers, one dropout layer, multiple attention mechanism layers, and a fully connected layer. The time-frequency image is used as input, and after feature extraction by the first convolutional module, feature extraction is performed through three structurally identical branches. Each branch consists of a second convolutional module, a third convolutional module, and an attention mechanism layer connected sequentially. Multi-scale feature extraction is achieved through these three branches. The outputs of the three branches are then fused, passed through dropout layers and fully connected layers, and finally processed using the Softmax function to obtain the classification output. Each of the first, second, and third convolutional modules consists of one convolutional layer, one activation layer, and one pooling layer.

[0035] According to a second aspect of the present invention, a micro-damage diagnosis system for tapered roller bearings is provided, comprising a module of the micro-damage diagnosis method for tapered roller bearings described in any one of the above-described methods.

[0036] According to a third aspect of the present invention, a processor is provided for running a program, wherein the program, when running, performs the micro-damage diagnosis method for tapered roller bearings as described in any one of the preceding claims.

[0037] The beneficial effects of this invention are as follows: Addressing the problem that traditional methods rely on expert identification of abnormal features, resulting in diagnostic results influenced by prior knowledge, this invention proposes a short-time Fourier transform with dynamically varying window lengths to obtain two-dimensional time-frequency images. This method has the advantages of reducing background interference, enhancing abnormal vibration signal features, and is computationally convenient, providing comprehensive vibration signal features and exhibiting significant adaptability. Furthermore, a generative adversarial network model is used to perform feature analysis on time-frequency images under different states, expanding the sample and solving the problem of difficulty in extracting weak damage fault signals from tapered roller bearings. Moreover, considering the strict requirements of convolutional neural networks for training samples, as well as the subtle differences in sample size and fixed weights, this invention proposes a weak damage identification model for tapered roller bearings, achieving intelligent detection of abnormal states in tapered roller bearings. Attached Figure Description

[0038] Figure 1 This is a flowchart of the method steps of the present invention;

[0039] Figure 2 This is a schematic diagram of STFT image generation used in this invention;

[0040] Figure 3 This is a schematic diagram of GAN sample augmentation used in this invention;

[0041] Figure 4 This is a schematic diagram of the tapered roller bearing weak damage identification model of the present invention;

[0042] Figure 5 This is a schematic diagram of the channel-priority attention mechanism;

[0043] Figure 6 This is a display diagram of the STFT image used in this invention;

[0044] Figure 7 This is a sample image generated by the GAN used in this invention. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be arbitrarily combined with each other.

[0046] Example 1: As Figure 1-7 As shown, according to a first aspect of the present invention, a method for diagnosing minor damage to tapered roller bearings is provided, comprising the following steps:

[0047] Step 1: Collect acceleration vibration signals of tapered roller bearings and construct accelerator vibration signal datasets under different conditions;

[0048] Step 2: Extract time-frequency features from the vibration signal based on the short-time Fourier transform with a dynamically changing window length, and obtain two-dimensional time-frequency images of the acceleration vibration signal of the first tapered roller bearing under different states to obtain an acceleration vibration signal image sample set;

[0049] Step 3: Construct a generative adversarial network model. Train the model based on the acceleration vibration signal image sample set obtained in Step 2. Then use the trained model to generate a two-dimensional time-frequency image of the acceleration vibration signal of the second tapered roller bearing, which serves as an expanded image sample set of acceleration vibration signal.

[0050] Step 4: Merge the augmented image sample set of acceleration vibration signals into the image sample set of acceleration vibration signals to obtain a mixed sample set of acceleration vibration signals; preset the training, testing and validation ratios, and divide the mixed sample set of acceleration vibration signals into a training set, a testing set and a validation set;

[0051] Step 5: Construct a weak damage identification model for tapered roller bearings for classification and identification;

[0052] Step 6: Input the training set obtained in Step 4 into the tapered roller bearing weak damage identification model constructed in Step 5 for training, and then perform hyperparameter adjustment based on the results of the validation set to obtain the trained tapered roller bearing weak damage identification model.

[0053] Step 7: Input the test set / time-frequency image of the tapered roller bearing to be diagnosed obtained in Step 4 into the trained tapered roller bearing weak damage identification model obtained in Step 6 for testing / diagnosis. That is, input the test set obtained in Step 4 into the trained tapered roller bearing weak damage identification model obtained in Step 6 for testing, or input the time-frequency image of the tapered roller bearing to be diagnosed into the trained tapered roller bearing weak damage identification model obtained in Step 6 for diagnosis.

[0054] Furthermore, the states in the accelerator vibration signal data set under different states are divided into five categories: single-point center damage, through-damage, circumferential edge damage, circumferential center damage, and normal.

[0055] Furthermore, step 2 specifically includes:

[0056] For the acceleration vibration signal sample x in the accelerator vibration signal dataset under different states constructed in step 1, the data is divided into segments according to the dynamically changing window length to obtain multiple short-time frame signals. The nth short-time frame signal X nThe expression is: X n =x n ·w(t), where w(t) is the window function and n is the index of the time frame; x n This is the nth data segment;

[0057] Perform a Fast Fourier Transform on each short time frame signal to obtain the corresponding spectrum; combine the spectrum results of all frames to obtain a two-dimensional time-frequency image of the acceleration vibration signal of the first tapered roller bearing.

[0058] Furthermore, the initial window length is set to L1, and when n>1, the window length L is dynamically adjusted according to the energy distribution. n The rules are as follows:

[0059] If E(n-1) > E high Set window length L n =L min Where n>1;

[0060] If E(n-1) <E low Set window length L n =L max Where n>1;

[0061] If E low ≤E(t)≤E high The window length is linearly interpolated based on the energy.

[0062]

[0063] In the formula, E hgih E low These represent the maximum and minimum preset energy threshold values, respectively; L max L min These represent the maximum and minimum values ​​of the preset sliding window length threshold, respectively;

[0064] Furthermore, the expression for the energy is as follows:

[0065]

[0066] Among them, L n Let L1 be the window length of the nth short-time frame signal, and L1 be the given initial window length; x n (i) represents the i-th data point in the n-th data segment; it should be noted that if the remaining data points are insufficient to divide a short time frame signal based on the window length, they are directly discarded.

[0067] Furthermore, step 3 specifically includes:

[0068] Construct a generative adversarial network model, wherein the constructed generative adversarial network model includes a generator and a discriminator;

[0069] The parameters of the generative adversarial network (GAN) model are adjusted according to the image resolution. When the resolution of the two-dimensional time-frequency image of the acceleration vibration signal of the first tapered roller bearing is greater than 512×512, the parameters of the generator and discriminator in the GAN model are increased based on their default values; otherwise, the parameters of the generator and discriminator in the GAN model are decreased based on their default values. In this embodiment of the invention, the resolution of the two-dimensional time-frequency image of the acceleration vibration signal of the first tapered roller bearing is 224×224, and the parameters of the first convolutional layer of the generator and discriminator are decreased, specifically: the convolutional kernel size is 7*7 (the default value is 9*9); for example, the increased value can be 13*13.

[0070] The established model is trained based on the acceleration vibration signal image sample set obtained in step 2, and then the trained model is used to generate a two-dimensional time-frequency image of the acceleration vibration signal of the second tapered roller bearing, which serves as an expanded image sample set of acceleration vibration signal.

[0071] Furthermore, the generator employs four convolutional layers with a training stride of 1 to 2. The resolution of the generated samples is equal to the image resolution in the acceleration vibration signal image sample set from step 2. The generator loss function is calculated using the following formula:

[0072]

[0073] In the formula, N is the number of samples in each training iteration (in this embodiment of the invention, the batch size is set to 64), p j Let D represent the j-th original sample in each training iteration, and let D denote the discriminator. D(p) j ) represents the original sample p j The closer the probability of it being true is to 1, the better.

[0074] The discriminator loss function is calculated using the following formula:

[0075]

[0076] In the formula, p j Let y represent the j-th original sample in each training iteration. j For the j-th generated sample generated by the generator each time it is trained, D(y) j ) represents the probability that the generated sample is fake, and λ represents the hyperparameter controlling the strength of the gradient penalty. In this invention, it is set to 10, and This represents the value of the gradient penalty. p j and yj The uniform sampling is calculated as follows: ε is obtained by random sampling from the uniform distribution Uniform(0,1). express L2 norm, For the discriminator pair The gradient of the output result.

[0077] Through simulation experiments, the similarity between the images in the augmented image sample set generated after training with a traditional adversarial generative network model and the original images was compared. This invention introduces the aforementioned loss function to train the generative adversarial network based on parameter adjustment, which can ensure that the images in the augmented image sample set generated by the model after image sample training have a higher similarity to the original images, thereby making the augmented samples more realistic.

[0078] For example, the preset training, testing, and verification ratio is 8:1:1.

[0079] Furthermore, the tapered roller bearing weak damage identification model includes multiple convolutional layers, multiple activation layers, multiple pooling layers, one dropout layer, multiple attention mechanism layers, and a fully connected layer (i.e., 7 convolutional layers, activation layers, pooling layers, and 3 attention mechanism layers). The time-frequency image is used as input, and after feature extraction by the first convolutional module, feature extraction is performed through three structurally identical branches. Each branch consists of a second convolutional module, a third convolutional module, and an attention mechanism layer connected sequentially. Multi-scale feature extraction is achieved through these three branches. The outputs of the three branches are fused, then passed through a dropout layer and a fully connected layer, and finally the softmax function is used to obtain the classification output. Each of the first, second, and third convolutional modules consists of one convolutional layer, one activation layer, and one pooling layer.

[0080] In the above, the activation layer uses the ReLU activation function; the training stride of the convolutional and pooling layers is set to 2 to improve the data dimensionality reduction speed and reduce the model training time. Furthermore, the attention mechanism layer adopts a channel-first attention mechanism. Compared to the channel-first attention mechanism, the channel-first attention mechanism can support the dynamic distribution of attention weights in both channel and spatial dimensions, thereby improving the model's recognition efficiency.

[0081] Based on the results of each training and testing iteration, the parameters of the tapered roller bearing weak damage identification model were optimized. A decaying learning rate was set to adaptively adjust the step size, thereby optimizing the training results. The momentum optimizer Adam was used to accelerate the convergence process of the convolutional neural network model and reduce oscillations. L2 regularization was added to the loss function. Where loss data The value of the cross-entropy loss function. The regularization coefficient is 10 in this invention. -5, ∑ w w 2 This represents the sum of squares of all weights, which helps to improve the overfitting problem that occurs during model training and obtain the optimal model parameters.

[0082] Furthermore, in step 7, the trained tapered roller bearing weak damage identification model obtained in step 6 is evaluated using accuracy, precision, recall, and F1-score as evaluation indicators.

[0083] According to a second aspect of the present invention, a weak damage diagnosis system for tapered roller bearings is provided, comprising modules of any one of the weak damage diagnosis methods for tapered roller bearings described above. Specifically, it comprises a first module for performing step 1: acquiring acceleration vibration signals of the tapered roller bearing and constructing an accelerator vibration signal dataset under different states; a second module for performing step 2: extracting time-frequency features from the vibration signals based on a short-time Fourier transform with a dynamically changing window length, obtaining two-dimensional time-frequency images of the acceleration vibration signals of the first tapered roller bearing under different states, and obtaining an acceleration vibration signal image sample set; a third module for performing step 3: constructing a generative adversarial network model, training the constructed model based on the acceleration vibration signal image sample set obtained in step 2, and then using the trained model to generate a two-dimensional time-frequency image of the acceleration vibration signals of the second tapered roller bearing as an augmented acceleration vibration signal image sample set; and a fourth module for performing step 4: expanding the acceleration vibration signal image sample set. The system incorporates an acceleration vibration signal image sample set into a mixed acceleration vibration signal image sample set. A preset training, testing, and validation ratio is used to divide the mixed acceleration vibration signal image sample set into a training set, a testing set, and a validation set. The fifth module executes step 5: constructing a tapered roller bearing weak damage identification model for classification and recognition. The sixth module executes step 6: inputting the training set obtained in step 4 into the tapered roller bearing weak damage identification model constructed in step 5 for training, and then performing hyperparameter adjustments based on the validation set results to obtain a trained tapered roller bearing weak damage identification model. The seventh module executes step 7: inputting the test set / time-frequency image of the tapered roller bearing to be diagnosed obtained in step 4 into the trained tapered roller bearing weak damage identification model obtained in step 6 for testing / diagnosis. For details not described in the above modules, please refer to the relevant descriptions in this embodiment.

[0084] According to a third aspect of the present invention, a processor is provided for running a program, wherein the program executes the micro-damage diagnosis method for tapered roller bearings as described in any one of the preceding embodiments.

[0085] Example 2: The following simulation further illustrates an optional embodiment of the present invention:

[0086] 1) Collect acceleration vibration signal data of laboratory tapered roller bearings under different conditions; the conditions are divided into five categories: single-point center damage, through-damage, full-circumference edge damage, full-circumference center damage and normal; in this embodiment of the invention, each condition has 10 samples, for a total of 50 samples;

[0087] 2) Based on the short-time Fourier transform with a dynamically changing window length, time-frequency features of the vibration signal are extracted to obtain two-dimensional time-frequency images of the acceleration vibration signal of the first tapered roller bearing under different states, thus obtaining an acceleration vibration signal image sample set; the short-time Fourier transform (STFT) conversion method is as follows: Figure 2 ( Figure 2 The bottom shows four random short-time Fourier transforms; this step yields 50 two-dimensional time-frequency images as a sample set of acceleration vibration signal images; such as Figure 6 The two-dimensional time-frequency RGB images under different states are further shown after short-time Fourier transform with dynamically varying window lengths.

[0088] 3) The acceleration vibration signal image sample set obtained in step 1) is input into the GAN generative adversarial network model to generate a large number of similar samples to solve the problem of small sample size. The resolution of the two-dimensional time-frequency image of the acceleration vibration signal of the first tapered roller bearing is 224×224. The specific parameters of the first convolutional layer of the generator and discriminator are: the convolution kernel size is 7*7. Through this step, 50 samples are generated for each state, that is, a total of 250 two-dimensional time-frequency images, which serve as the expanded image sample set of the acceleration vibration signal; for example... Figure 7 The two-dimensional time-frequency RGB images of the acceleration vibration signal of the second tapered roller bearing generated by the GAN generative adversarial network model under different states are further presented.

[0089] 4) The augmented image sample set of the acceleration vibration signal is merged into the image sample set of the acceleration vibration signal to obtain a mixed sample set of acceleration vibration signal images, totaling 300 two-dimensional time-frequency images; these are then used to create training, testing, and validation sets in an 8:1:1 ratio. The sample creation process is as follows: Figure 3 .

[0090] As can be seen from the above technical solution, the present invention combines STFT with a generative adversarial network model, which enables image samples to have advantages such as diverse features, obvious differences in sample space, and direct observation.

[0091] 5) Constructing a weak damage identification model for tapered roller bearings for classification and recognition; the weak damage identification model for tapered roller bearings constructed in this invention uses a multi-scale convolutional network model as its basic structure, as shown in the figure below. Figure 4As shown in Table 1, the relevant parameters are as follows: the entire network has few layers, requires few training parameters, has a fast training speed, and achieves good classification results. Then, a channel-first attention mechanism is added to the network, and its network structure is as follows: Figure 5 The addition of this attention mechanism can dynamically adjust the attention weights in the channel and spatial dimensions, thereby improving the model's recognition efficiency; the output of the fully connected layer uses the Softmax function.

[0092] Table 1 Initial parameters of the model

[0093]

[0094] 6) Input the training set obtained in step 4) into the tapered roller bearing weak damage identification model constructed in step 5) for training, and then perform hyperparameter adjustment based on the results of the validation set to obtain the trained tapered roller bearing weak damage identification model.

[0095] 7) Input the test set obtained in step 4) into the trained tapered roller bearing weak damage identification model obtained in step 6) for testing. Use accuracy, precision, recall, and F1-score as evaluation indicators to evaluate the trained tapered roller bearing weak damage identification model obtained in step 6). Specific evaluation indicators are shown in Table 2.

[0096] Table 2. Test results of acceleration vibration signals of laboratory tapered roller bearings.

[0097] Accuracy / % Accuracy / % Recall rate / % F1-score / % 98 97 97 97

[0098] As shown in Table 2, the present invention achieves intelligent detection of abnormal conditions in tapered roller bearings with high accuracy.

[0099] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A method for diagnosing minor damage in tapered roller bearings, characterized in that, Includes the following steps: Step 1: Collect acceleration vibration signals of tapered roller bearings and construct acceleration vibration signal datasets under different conditions; Step 2: Extract time-frequency features from the vibration signal based on the short-time Fourier transform with a dynamically changing window length, and obtain two-dimensional time-frequency images of the acceleration vibration signal of the first tapered roller bearing under different states to obtain an acceleration vibration signal image sample set; Step 3: Construct a generative adversarial network model. Train the model based on the acceleration vibration signal image sample set obtained in Step 2. Then use the trained model to generate a two-dimensional time-frequency image of the acceleration vibration signal of the second tapered roller bearing, which serves as an expanded image sample set of acceleration vibration signal. Step 4: Merge the augmented image sample set of acceleration vibration signals into the image sample set of acceleration vibration signals to obtain a mixed sample set of acceleration vibration signals; preset the training, testing and validation ratios, and divide the mixed sample set of acceleration vibration signals into a training set, a test set and a validation set; Step 5: Construct a weak damage identification model for tapered roller bearings for classification and identification; Step 6: Input the training set obtained in Step 4 into the tapered roller bearing weak damage identification model constructed in Step 5 for training, and then perform hyperparameter adjustment based on the results of the validation set to obtain the trained tapered roller bearing weak damage identification model. Step 7: Input the time-frequency images of the test set / tapered roller bearing to be diagnosed obtained in Step 4 into the trained tapered roller bearing weak damage identification model obtained in Step 6 for testing / diagnosis. The states in the acceleration vibration signal dataset under different states are divided into five categories: single-point center damage, through-damage, circumferential edge damage, circumferential center damage, and normal. Step 2 specifically involves: For the acceleration vibration signal samples in the acceleration vibration signal dataset under different states constructed in step 1, the data segments are divided according to the dynamically changing window length to obtain multiple short time frame signals. Perform a fast Fourier transform on each short time frame signal to obtain the corresponding spectrum; combine the spectrum results of all frames to obtain a two-dimensional time-frequency image of the acceleration vibration signal of the first tapered roller bearing; The initial window length is set to L1. When n>1, the window length L is dynamically adjusted according to the energy distribution. n The rules are as follows: If E(n-1) > E high Set window length L n =L min Where n>1; If E(n-1) <E low Set window length L n =L max Where n>1; If E low ≤E(n-1)≤E high The window length is linearly interpolated based on the energy. In the formula, E high E low These represent the maximum and minimum preset energy threshold values, respectively; L max L min These represent the maximum and minimum values ​​of the preset sliding window length threshold, respectively; E(n-1) represents the energy of the (n-1)th short-time frame signal; L n The window length is the nth short time frame signal.

2. The method for diagnosing minor damage to tapered roller bearings according to claim 1, characterized in that, Step 3 specifically involves: Construct a generative adversarial network model, wherein the constructed generative adversarial network model includes a generator and a discriminator; The parameters of the generative adversarial network model are adjusted according to the image resolution; The established model is trained based on the acceleration vibration signal image sample set obtained in step 2, and then the trained model is used to generate a two-dimensional time-frequency image of the acceleration vibration signal of the second tapered roller bearing, which serves as an expanded image sample set of acceleration vibration signal.

3. The method for diagnosing minor damage to tapered roller bearings according to claim 2, characterized in that, The generator employs four convolutional layers with a training stride of 1 to 2. The resolution of the generated samples is equal to the resolution of the acceleration vibration signal image sample set from step 2. The generator loss function Loss1 is calculated as follows: In the formula, N is the number of samples in each training iteration, and p j Let D represent the j-th original sample in each training iteration, and let D denote the discriminator. D(p) j ) represents the original sample p j The probability of it being true; The discriminator loss function Loss2 is calculated as follows: In the formula, p j Let y represent the j-th original sample in each training iteration. j For the j-th generated sample generated by the generator each time it is trained, D(y) j ) represents the probability that the generated sample is fake, and λ represents the hyperparameter that controls the strength of the gradient penalty; This represents the value of the gradient penalty. p j and y j The uniform sampling is performed, and ε is obtained by random sampling from a uniform distribution. express L2 norm, For the discriminator pair The gradient of the output result.

4. The method for diagnosing minor damage to tapered roller bearings according to claim 1, characterized in that, The tapered roller bearing weak damage identification model includes multiple convolutional layers, multiple activation layers, multiple pooling layers, one dropout layer, multiple attention mechanism layers, and a fully connected layer. The time-frequency image is used as input, and after feature extraction by the first convolutional module, it is further processed through three structurally identical branches. Each branch consists of a second convolutional module, a third convolutional module, and an attention mechanism layer connected sequentially. Multi-scale feature extraction is achieved through these three branches. The outputs of the three branches are then fused, passed through dropout layers and fully connected layers, and finally processed using the Softmax function to obtain the classification output. Each of the first, second, and third convolutional modules consists of one convolutional layer, one activation layer, and one pooling layer.

5. A micro-damage diagnosis system for tapered roller bearings, characterized in that, The module includes the method for diagnosing minor damage to tapered roller bearings as described in any one of claims 1-4.

6. A processor, characterized in that, The processor is used to run a program, wherein the program executes the method for diagnosing minor damage to tapered roller bearings according to any one of claims 1-4.

Images