Rolling contact fatigue detection method based on hybrid neural network

Abstract

The present application relates to the technical field of rolling contact fatigue detection, and particularly relates to a rolling contact fatigue detection method based on a hybrid neural network, which comprises the following steps: constructing a corresponding hybrid neural network model as a fatigue prediction model based on a neural network and an SVM classifier; training the fatigue prediction model; inputting vibration data of a test sample to be detected into the trained fatigue prediction model, performing feature extraction and feature classification prediction by the neural network and the SVM classifier respectively, and then outputting corresponding fatigue prediction class probabilities; and taking the fatigue prediction class probabilities output by the fatigue prediction model as fatigue detection results of the corresponding test sample to be detected. In the present application, feature extraction can be accurately and effectively realized through the fatigue prediction model, and the generalization ability of the fatigue prediction model can be improved, so that the accuracy and effect of rolling contact fatigue detection are improved.

Description

Technical Field

[0001] This invention relates to the field of rolling contact fatigue detection technology, and more specifically to a rolling contact fatigue detection method based on a hybrid neural network. Background Technology

[0002] In mechanical engineering, some mechanical parts operate under rolling contact conditions, such as bearings, cams, and gears widely used in automobiles, ships, and trains. After a certain number of cyclic stresses, pits, shallow and deep spalling will occur on the contact surfaces of these parts, leading to a decrease in their durability. This phenomenon is called rolling contact fatigue, a typical phenomenon that significantly affects the mechanical properties of parts. Rolling contact fatigue in critical parts directly impacts machine operating safety. Therefore, rolling contact fatigue testing has emerged. Rolling contact fatigue testing is a test method that uses a cylindrical (or cylindrical convex) specimen and a co-existing specimen to assess the rolling contact fatigue performance of materials through rolling contact.

[0003] To address the issues of poor accuracy and low efficiency in existing rolling contact fatigue testing methods, Chinese Patent Publication No. CN111291466A discloses "An Analysis Method and System for Factors Affecting Bearing Fatigue Life," which includes: collecting bearing data over a period of time; importing the bearing data into a VAE network model to obtain bearing data for different major categories of influencing factors; importing the bearing data from each major category of influencing factors into an SVM classification model for multi-classification processing to obtain the degree of influence of different sub-categories of influencing factors under each major category on bearing fatigue life, so as to adjust the bearing usage plan based on the degree of influence.

[0004] The bearing fatigue life influencing factor analysis method in the aforementioned existing scheme is also a rolling contact fatigue detection method, which uses an SVM classification model to obtain the degree of influence of different factors on bearing fatigue life. This existing scheme is mainly used for detecting influencing factors of rolling contact fatigue life, not for detecting rolling contact fatigue itself. The applicant has found that the rolling contact fatigue state of a specimen generally includes three states: normal, running-in, and failure. Existing technologies commonly use neural network models to achieve fatigue detection of specimens. Because neural network models can extract features well, they can ensure the effectiveness of rolling contact fatigue detection. However, the Softmax classifier commonly used in neural network models requires a large amount of sample data during training to improve the accuracy of the recognition results. Excessive use of sample data can lead to the model performing well on the training set but poorly on other samples, i.e., it is prone to overfitting, resulting in low accuracy in rolling contact fatigue detection. Therefore, how to design a detection method that can improve the accuracy of rolling contact fatigue detection is an urgent technical problem to be solved. Summary of the Invention

[0005] In view of the shortcomings of the prior art, the technical problem to be solved by the present invention is: how to provide a rolling contact fatigue detection method based on a hybrid neural network, which can accurately and effectively extract features through a fatigue prediction model, and improve the generalization ability of the fatigue prediction model, thereby improving the accuracy and effect of rolling contact fatigue detection.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0007] A method for detecting rolling contact fatigue based on hybrid neural networks includes:

[0008] S1: Construct a hybrid neural network model based on neural networks and SVM classifiers as a fatigue prediction model;

[0009] S2: Train the fatigue prediction model;

[0010] S3: Input the vibration data of the sample to be tested into the trained fatigue prediction model, and perform feature extraction and feature classification prediction by the neural network and SVM classifier respectively, and then output the corresponding fatigue prediction class probability.

[0011] S4: Use the fatigue prediction class probability output by the fatigue prediction model as the fatigue test result of the corresponding test specimen.

[0012] Preferably, in step S1, the hybrid neural network model is constructed based on the VGG16 network model, wherein the fully connected layer and the Softmax layer of the VGG16 network model are removed, and the pooling layer of the last layer of the VGG16 network model is connected to the input layer of the SVM classifier to form a hybrid neural network model of VGG16-SVM.

[0013] Preferably, in the hybrid neural network model, the original ReLU activation function of the VGG16 network model is replaced with the Maxout activation function to form a hybrid neural network model of VGG16-Maxout-SVM.

[0014] Preferably, in step S2, the fatigue prediction model is trained through the following steps:

[0015] S201: Create training data for training the fatigue prediction model;

[0016] S202: Train a VGG16 network model based on training data to construct a hybrid neural network model;

[0017] S203: Fix the parameters of the convolutional and pooling layers of the VGG16 network model after training, then remove the fully connected and softmax layers of the VGG16 network model, and connect the pooling layer of the last layer to the input layer of the SVM classifier to form a hybrid neural network model.

[0018] S204: Train the parameters of the SVM classifier in the hybrid neural network model based on the training data until training is complete.

[0019] Preferably, in step S201, training data is generated through the following steps:

[0020] S2011: Perform rolling contact tests on training specimens;

[0021] S2012: Acquire image data and corresponding vibration data of the rolling contact surface of the training specimen;

[0022] S2013: Calculate the corresponding fatigue state based on the image data of the rolling contact surface of the training specimen;

[0023] S2014: The fatigue state and corresponding vibration data of the training specimens are used as a set of training data.

[0024] Preferably, in step S2013, the area of ​​the spalling point on the rolling contact surface of the training sample is calculated based on the image data, and then the training sample is classified into the corresponding fatigue state based on the area of ​​the spalling point.

[0025] Preferably, in step S2014, the vibration data corresponding to the fatigue state is converted into an SDP image with a size of 224×224 using the SDP algorithm, and then the SDP images of the fatigue state and the corresponding vibration data are used as a set of training data.

[0026] Preferably, in step S202, before training the VGG16 network model, the VGG16 network model is initialized using the Xavier initialization method.

[0027] The initialization of the VGG16 network model must satisfy the following:

[0028]

[0029] Where: n i Var(W) represents the number of neurons in the i-th layer of the neural network; i ) represents the weight variance of the i-th layer.

[0030] Preferably, in step S202, the VGG16 network model is trained using a training strategy based on L2 regularized momentum gradient descent. During the training process, only the state of the convolutional and pooling layers of the VGG16 network model is considered. The training of the VGG16 network model is completed when the cross-entropy loss approaches its minimum value.

[0031] Preferably, the training strategy based on L2 regularized momentum gradient descent updates the parameters using the following formula:

[0032] Gradient update: g′ t =g t +λθ t-1 ;

[0033] Weight update: θ t =θ t-1 -μg t -μλθ t-1 ;

[0034] Momentum update: d t =μd t-1 -αg′ t ;

[0035] In the above formula: θ represents the initial weight with regularization; d represents the initial momentum; t represents the gradient iteration number; g′ t θ represents the gradient after the t-th iteration update; t d represents the weight value after the t-th iteration update; t Let denot be the momentum after the t-th iteration update, α be the learning rate, μ be the momentum coefficient, and λ be the regularization coefficient.

[0036] Compared with the prior art, the rolling contact fatigue detection method of the present invention has the following advantages:

[0037] This invention constructs a hybrid neural network model based on neural networks and SVM classifiers as a fatigue prediction model. On one hand, the neural network in the fatigue prediction model can accurately and effectively extract features, thus ensuring the effectiveness of the model in rolling contact fatigue detection. On the other hand, the fatigue prediction model uses an SVM classifier combined with the features extracted by the neural network for classification and prediction. The SVM classifier, based on the principle of minimizing structural risk, minimizes the generalization error on unknown datasets through training on known datasets, effectively solving the problem of model overfitting and thus improving the model's generalization ability, thereby enhancing the accuracy of the model in rolling contact fatigue detection. Furthermore, this invention's method of implementing rolling contact fatigue detection based on vibration data ensures the effectiveness of rolling contact fatigue detection. Attached Figure Description

[0038] To make the objectives, technical solutions, and advantages of the invention clearer, the invention will now be described in further detail with reference to the accompanying drawings, wherein:

[0039] Figure 1 This is a logic block diagram of a rolling contact fatigue detection method.

[0040] Figure 2 This is a network structure diagram of a hybrid neural network model; in the diagram, C represents a convolutional layer and P represents a pooling layer.

[0041] Figure 3 The network structure diagram for the Maxout activation function;

[0042] Figure 4 This is a schematic diagram of the main and auxiliary test system; in the diagram, 1 represents the main sample and 2 represents the auxiliary sample.

[0043] Figure 5 SDP images for three fatigue states; in the figure, Figure 5 (a) indicates the normal state. Figure 5 (b) indicates the running-in state. Figure 5 (c) indicates a failure state;

[0044] Figure 6 The figures show the epochs and accuracy curves for the four models. Detailed Implementation

[0045] The following detailed explanation illustrates the specific implementation methods:

[0046] Example:

[0047] This embodiment discloses a rolling contact fatigue detection method based on a hybrid neural network.

[0048] like Figure 1 As shown, the rolling contact fatigue detection method based on hybrid neural networks includes the following steps:

[0049] S1: Construct a hybrid neural network model based on neural networks and SVM classifiers as a fatigue prediction model;

[0050] S2: Train the fatigue prediction model;

[0051] S3: Input the vibration data of the sample to be tested into the trained fatigue prediction model, and perform feature extraction and feature classification prediction by the neural network and SVM classifier respectively, and then output the corresponding fatigue prediction class probability.

[0052] In this embodiment, both feature extraction and feature classification prediction using the neural network and SVM classifier are implemented using existing mature methods, and this invention does not make any improvements to them. Specifically, the fatigue prediction model extracts corresponding feature vector values ​​based on vibration data through the neural network, and then inputs the feature vector values ​​into the SVM classifier. The tensor in the last layer of the SVM classifier undergoes linear transformation, biasing, and activation to obtain the corresponding fatigue prediction class probability.

[0053] S4: Use the fatigue prediction class probability output by the fatigue prediction model as the fatigue test result of the corresponding test specimen.

[0054] This invention constructs a hybrid neural network model based on neural networks and SVM classifiers as a fatigue prediction model. On one hand, the neural network in the fatigue prediction model can accurately and effectively extract features, thus ensuring the effectiveness of the model in rolling contact fatigue detection. On the other hand, the fatigue prediction model uses an SVM classifier combined with the features extracted by the neural network for classification and prediction. The SVM classifier, based on the principle of minimizing structural risk, minimizes the generalization error on unknown datasets through training on known datasets, effectively solving the problem of model overfitting and thus improving the model's generalization ability, thereby enhancing the accuracy of the model in rolling contact fatigue detection. Furthermore, this invention's method of implementing rolling contact fatigue detection based on vibration data ensures the effectiveness of rolling contact fatigue detection.

[0055] In the specific implementation process, combined with Figure 2 As shown, the hybrid neural network model is built based on the VGG16 network model. The fully connected layer and softmax layer of the VGG16 network model are removed, and the pooling layer of the last layer of the VGG16 network model is connected to the input layer of the SVM classifier to form a hybrid neural network model of VGG16-SVM.

[0056] Specifically, in the hybrid neural network model, the original ReLU activation function of the VGG16 network model is replaced with the Maxout activation function to form a hybrid neural network model of VGG16-Maxout-SVM.

[0057] The Maxout function combines multiple activation functions and continuously updates the activation parameters simultaneously, ensuring iterative gradient training. Figure 3 As shown, each Maxout unit of the Maxout function consists of several activation functions, all of which are trainable activation units. Its output is the maximum value after activation, so it can fit any convex function, including activation functions such as ReLU.

[0058] For a given input x (input vector or hidden layer state vector), the following operations are performed: activation and maximum selection: z ij=x T W ij +b ij ; In the formula: W ij ∈R d×m×k b ij ∈R m×k And both represent trainable variables, representing the number of hidden units, h i Indicates the selection of activation function z ij Maximum value operation.

[0059] This invention replaces the original fully connected layers and Softmax layers of the VGG16 network model with an SVM classifier. This approach, while ensuring the accuracy and effectiveness of feature extraction, further leverages the characteristics of the SVM classifier to address the problem of model overfitting, thereby improving the model's generalization ability.

[0060] Meanwhile, this invention replaces the original ReLU activation function of the VGG16 network model with the Maxout activation function. The Maxout activation function can select the maximum value at the same position of multiple feature maps as the final output, which reduces the number of feature maps. This solves the problem that the gradient of ReLU functions is prone to vanishing during tensor propagation, thus increasing the model training cost. This further improves the efficiency of rolling contact fatigue detection.

[0061] In the specific implementation process, the fatigue prediction model is trained through the following steps:

[0062] S201: Create training data for training the fatigue prediction model;

[0063] S202: Train a VGG16 network model based on training data to construct a hybrid neural network model;

[0064] During training, a 224×224 image is input into the VGG16 model, and only the states of the convolutional and pooling layers are monitored. When the corresponding LOSS curve becomes flat and the fluctuations are small, it indicates that the LOSS is close to its minimum, and the VGG16 network model training is complete. Specifically, the VGG16 network model is trained using a momentum gradient descent method based on L2 regularization. During training, only the states of the convolutional and pooling layers of the VGG16 network model are monitored, and the training of the VGG16 network model is complete when the cross-entropy loss approaches its minimum.

[0065] In this embodiment, the VGG16 network model is trained using the deep learning backpropagation algorithm. Deep learning backpropagation seeks the minimum loss value, L2 plasmonic prevents overfitting, and momentum gradient descent accelerates convergence. Backpropagation is needed to update parameters because the convolutional layer parameters of VGG need to be adjusted.

[0066] S203: Fix the parameters of the convolutional and pooling layers of the VGG16 network model after training, then remove the fully connected and softmax layers of the VGG16 network model, and connect the pooling layer of the last layer to the input layer of the SVM classifier to form a hybrid neural network model.

[0067] S204: Train the parameters of the SVM classifier in the hybrid neural network model based on the training data until training is complete.

[0068] In this embodiment, the training of the SVM classifier is achieved using mature training methods in the existing technology.

[0069] This invention trains the parameters of the convolutional and pooling layers of the VGG16 network model and the parameters of the SVM classifier separately through step-by-step training. By fixing the parameters of the convolutional and pooling layers of the VGG16 network model after training, the feature extraction effect of the VGG16 network model is guaranteed. Then, combined with the training of the SVM classifier, the accuracy and effectiveness of feature extraction can be guaranteed while taking into account the classification and prediction effects, thereby further improving the training effect of the hybrid neural network model.

[0070] Specifically, training data is created through the following steps:

[0071] S2011: Perform rolling contact tests on training specimens;

[0072] In this embodiment, rolling contact fatigue testing is performed using a rolling contact fatigue testing machine, which mainly includes a master testing system and a data acquisition system. The master testing system is used to perform rolling contact testing on the specimen, and the specimen is rolled by two servo motors. Figure 4As shown, the main and auxiliary samples are rotated by rotating shafts driven by servo motors to simulate a rolling contact environment under loaded load conditions. The main sample is a roller made of 40Cr, 20CrNiMo, or 18CrNiMo7-6 materials with a hardness of 58+1.5HRC and a surface roughness of 0.8Ra. The roller contact surface widths are 3mm, 5mm, 8mm, and 10mm. When the contact surface width is 5mm, the main and auxiliary roller samples are in line contact; when the contact surface width is 10mm, the main and auxiliary roller samples are in point contact. During the experiment, the main and auxiliary roller samples are in contact with a metered amount of oil-lubricated medium, and their contact surfaces are in a near-elastohydrodynamic lubrication state. The rotation speed is set to 1200r / min-1500r / min, and a load of 1350N is applied.

[0073] S2012: Acquire image data and corresponding vibration data of the rolling contact surface of the training specimen;

[0074] In this embodiment, image data and vibration data of the roller (training sample) are acquired through a data acquisition system.

[0075] The image acquisition section primarily acquires images of the roller sample contact surface. The hardware used includes a LINEA series linear scan camera from DALSA (Canada) and an MLM-3XMP lens from COMPUTAR (Japan). After image acquisition, the image detection algorithm can identify the peeling points on the roller contact surface and calculate the peeling area, achieving a detection accuracy of approximately 95%. The vibration data acquisition section primarily acquires the vibration data of the roller during the test. The hardware used is a TRZD-VA-59 piezoelectric vibration sensor with a sampling frequency of 10kHz.

[0076] S2013: Calculate the corresponding fatigue state based on the image data of the rolling contact surface of the training specimen;

[0077] In this embodiment, the area of ​​the spalling point on the rolling contact surface of the training specimen is calculated based on the image data of the rolling contact surface, and then the training specimen is classified into the corresponding fatigue state based on the area of ​​the spalling point. Specifically, the fatigue state is classified according to "YB / T5345-2014 Test Method for Rolling Contact Fatigue of Metallic Materials": Normal state: no spalling point, that is, the area of ​​the spalling point is 0 mm. 2 Running-in stage: Peeling occurs, with peeling area ranging from 0 to 3 mm. 2 Failure condition: The area of ​​the spalling point is greater than or equal to 3mm. 2 .

[0078] S2014: The fatigue state and corresponding vibration data of the training specimens are used as a set of training data.

[0079] In this embodiment, the area of ​​the spalling point on the rolling contact surface of the training sample is calculated based on the image data of the rolling contact surface, and then the training sample is divided into the corresponding fatigue state based on the area of ​​the spalling point.

[0080] First, the vibration data is preprocessed and converted to a uniform range of 0-255. Second, the preprocessed vibration data is converted into an SDP image of size 56×56 using the SDP algorithm. Finally, the 56×56 SDP image is retopologically converted into the minimum input image size of 224×224 for the VGG16 model through two image topology transformations.

[0081] In this embodiment, for three different rolling contact fatigue states, each state contains 600 sets of sample data, with 1024 vibration data points per set. The training and test sets are divided at a 5:1 ratio, meaning each fatigue state contains 500 training samples and 100 test samples. Furthermore, one set of sample data corresponds to one SDP image, as shown in Table 1 and... Figure 5 As shown.

[0082] Table 1 SDP Image Dataset

[0083]

[0084] This invention can effectively acquire training data containing the correspondence between fatigue states and vibration data, thereby helping to ensure the training effect of the hybrid neural network model. Furthermore, the method of classifying fatigue states in this invention complies with relevant regulations and annotations, thus making it more suitable for practical applications.

[0085] In the specific implementation process, before training the VGG16 network model, it is first initialized using the Xavier initialization method. The Xavier initialization method is a mature existing technique, and its basic idea is to maintain the consistency of the variance of the gradient and the activation value.

[0086] Suppose that the linear unit f is differentiable at 0 and its derivative is 1, then the output of the i-th convolutional layer can be expressed as:

[0087] s i =z i W i +b i and z i+1 =f(s) l );

[0088] The output s of the convolutional layer is evaluated using the loss function Loss. i and w i Differentiate:

[0089]

[0090]

[0091] Since the outputs of the convolutional layers are all 0 during model initialization, most of the outputs at this point lie within the linear variation region of f. Assuming the weights of the convolutional layers are initialized independently and identically distributed, we can obtain:

[0092]

[0093]

[0094] In the formula, n i z represents the number of neurons in the i-th layer of the neural network; x represents the input image of the model. i The variance can be derived from the variance of x and n. i Var(W i The product of these products is obtained by multiplying them together. Based on this recurrence relation, for any d-layer network, we can obtain:

[0095]

[0096]

[0097] During the forward propagation phase of the model, excessively large output variance will cause gradient explosion, while excessively small variance will lead to gradient vanishing. Therefore, to ensure that the gradient is not too large or too small during model initialization, the variance between the output and input must be kept constant, i.e.: Var(z) i )≡Var(z i′ );

[0098] Similarly, during the backpropagation phase:

[0099] Therefore, we can conclude that:

[0100] Therefore, the initialization of the VGG16 network model must satisfy the following conditions.

[0101] In the above formula: n i Var(W) represents the number of neurons in the i-th layer of the neural network; i ) represents the weight variance of the i-th layer.

[0102] The VGG16 network model is initialized using the following formula;

[0103]

[0104] in, For probability theory to satisfy Var(W) i The uniform distribution of variance is expressed as follows: the weights w follow an interval... A uniform distribution on the surface.

[0105] This invention utilizes the Xavier initialization method to effectively initialize hybrid neural network models, thereby avoiding the problem of models failing to converge due to getting stuck in local optima during training, and thus improving the training effect.

[0106] In practice, the VGG16 network model is trained using a training strategy based on L2 regularized momentum gradient descent. The parameters are updated using the following formula:

[0107] Gradient update: g′ t =g t +λθ t-1 ;

[0108] Weight update: θ t =θ t-1 -μg t -μλθ t-1 ;

[0109] Momentum update: d t =μd t-1 -αg′ t ;

[0110] In the above formula: θ represents the initial weight with regularization; d represents the initial momentum; t represents the gradient iteration number; g′ t θ represents the gradient after the t-th iteration update; t d represents the weight value after the t-th iteration update; t Let denot be the momentum after the t-th iteration update, α be the learning rate, μ be the momentum coefficient, and λ be the regularization coefficient.

[0111] During model training, insufficient training data or excessive model parameters can easily lead to overfitting, resulting in weak model generalization ability. To prevent overfitting, regularization strategies are typically used to fine-tune the model parameters and improve generalization.

[0112] L2 regularization adds a regularization term Ω(θ) = 1 / 2 × ||θ to the objective function J. 2 ||, as shown in the following formula:

[0113] In the formula, λ∈[0,+∞) is the regularization parameter, which controls the influence of the regularization term on the objective function J.

[0114] The addition of a regularization term causes the model to favor smaller weights during training, resulting in a "weight decay" effect. Therefore, a larger λ indicates greater weight decay. After adding the regularization term, the weight update becomes:

[0115] After sorting:

[0116] Before each gradient update after weight decay, the weights are multiplied by a constant factor (1-λα), which reduces the bias of the objective function in directions where gradient changes are not significant, thereby accelerating the search for the optimal solution.

[0117] The purpose of momentum gradient descent is to accelerate convergence by adjusting the direction of gradient updates using momentum.

[0118] Let μ∈[0,1] be the momentum coefficient, then at the t-th iteration: m t =μm t-1 +g t In the formula, m t-1 To update the gradient direction, m t This represents the updated gradient direction.

[0119] At this point, the update formula for momentum gradient descent is: θ t =θ t-1 -ηm t ;

[0120] Momentum adjusts the direction of gradient updates, preventing gradients from approaching zero and parameters from getting trapped in local optima, thus improving oscillations. After introducing L2 regularization into momentum gradient descent, its loss function is:

[0121] Take the partial derivatives of the loss function with respect to the weights θ and the bias b:

[0122] The partial derivatives of the weights change after L2 regularization, but not the biases. This indicates that this method only affects the training of the model's weights θ, but not the biases b. Since model training primarily focuses on training the weights θ to achieve the fit, only the weights θ are derived:

[0123]

[0124]

[0125] The gradient update at this point is: g′ t =g t +λθ t-1 ;

[0126] The weight update for momentum gradient descent with L2 regularization is then: θ t =θ t-1 -ηg′ t =θ t-1 -ηg t -ηλθ t-1 .

[0127] The training of the VGG16 network model is generally completed using the layer-by-layer backpropagation algorithm (a specific application of gradient descent).

[0128] Assume the dimension of the CNN input layer is m, and the input vector is x = (x1, x2, ..., xn). m ) T Each convolutional layer and fully connected layer has n k There are n neurons (k = 1, 2, ..., R), and their corresponding vector representations are as follows: The output layer has c neurons, and the output vector is o = (o1, o2, ..., o3). c ) T Let the weight coefficient matrices of the (k-1)th layer and the kth layer be set as follows: The activation output of each neuron in each layer is then represented as:

[0129]

[0130] In the formula, b 1 b K b R+1 These are the bias values ​​for each layer; σ o These are the activation functions for each layer.

[0131] Let there be L training samples (x) l y l (1≤l≤L), the input is Expected output is The actual output is The cross-entropy function is chosen as the loss function. As the error increases, the gradient also increases, leading to faster weight updates (w) and accelerated training. This avoids the reduction in learning rate associated with the mean squared error loss function. The error is expressed as:

[0132] make The backpropagation error of the l-th sample in each layer of the network can be obtained as follows:

[0133] The recursive form is expressed as:

[0134] In the formula, This represents the multiplication of each element in the matrix.

[0135] Ultimately, through Update weight W k and bias value b k ;

[0136] in,

[0137]

[0138]

[0139]

[0140] This invention trains a hybrid neural network model using a momentum gradient descent training strategy based on L2 regularization. On the one hand, it can optimize model parameters using L2 regularization, thereby improving the model's generalization ability; on the other hand, it can accelerate convergence by adjusting the direction of gradient updates using momentum, thereby improving the model's training efficiency and ensuring the model's training effect.

[0141] To better illustrate the advantages of the hybrid neural network model in this invention, the following experiments are disclosed in this embodiment.

[0142] This experiment selected four models: VGG16-Maxout-SVM (the hybrid neural network model in this invention), CNN (1D-CNN), VGG16-LeakyReLU-SVM, and VGG16-SELU-SVM. Among them, 1D-CNN is a LeNet-5 structure, with each layer consisting of multiple one-dimensional vectors, and each vector consisting of multiple neurons. The vibration signal is directly input into the network as a one-dimensional vector.

[0143] VGG16-LeakyReLU-SVM, VGG16-SELU-SVM, and VGG16-Maxout-SVM were trained using the training dataset created in this invention. Since 1D-CNN directly inputs one-dimensional vibration data, to ensure the consistency of the training data, each state was set to contain 600 sets of sample data (1024 vibration data points per set), and the training and test sets were divided in a 5:1 ratio, so that the vibration data in each set is identical to a single SDP image.

[0144] Before training, the 1D-CNN was initialized with random parameters, while other models were initialized using the Xavier initialization method proposed in this invention. During training, all models adopted a momentum gradient descent training strategy based on L2 regularization, with an initial learning rate set to 0.001. To avoid overfitting, early stopping was employed in this experiment.

[0145] During model training, monitor the changes in epochs and accuracy curves, such as... Figure 6As shown in the figure, the accuracy curves reveal that the accuracy of 1D-CNN, VGG16-LeakyReLU-SVM, VGG16-SELU-SVM, and VGG16-Maxout-SVM tends to level off at 130 epochs, 175 epochs, 100 epochs, and 210 epochs, respectively. These epochs are set as the iteration number for the corresponding models, and the training parameters and weights are saved every 25 epochs. The model with the highest accuracy is then selected.

[0146] Table 2 lists the classification performance of each model for the three rolling contact fatigue states after 10-fold cross-validation. Five performance evaluation metrics were used: accuracy, precision, recall, F1-score, and training time.

[0147] Table 2 Classification performance of the four models

[0148]

[0149] As can be seen from Table 2, the VGG16-Maxout-SVM model used in this invention performs better, with its accuracy, precision, recall, and F1-Score being the best overall.

[0150] In terms of accuracy, VGG16-Maxout-SVM shows performance improvements of 9.9%, 3.8%, and 1% compared to 1D-CNN, VGG16-LeakyReLU-SVM, and VGG16-SELU-SVM, respectively. In terms of precision, VGG16-Maxout-SVM also shows improvements of 8.8%, 4.5%, and 0.6% compared to 1D-CNN, VGG16-LeakyReLU-SVM, and VGG16-SELU-SVM, respectively. In terms of recall, VGG16-Maxout-SVM shows performance improvements of 12%, 9.2%, and 2.8% compared to 1D-CNN, VGG16-LeakyReLU-SVM, and VGG16-SELU-SVM, respectively. In terms of F1 score, VGG16-Maxout-SVM shows performance improvements of 13%, 4.5%, and 1.5% compared to 1D-CNN, VGG16-LeakyReLU-SVM, and VGG16-SELU-SVM, respectively. However, VGG16-Maxout-SVM takes the longest to train, at 7.5 hours, which is 2.5 hours, 2 hours, and 2 hours longer than 1D-CNN, VGG16-LeakyReLU-SVM, and VGG16-SELU-SVM, respectively.

[0151] Therefore, the VGG16-Maxout-SVM used in this invention can more accurately classify different rolling contact fatigue states.

[0152] Since accuracy, precision, recall, and F1-score do not follow a normal distribution, they cannot be verified using a t-test. Therefore, to assess whether there are significant differences between different models, this experiment used the Wilcoxon signed-rank test for pairwise comparisons, and the comparison results are shown in Table 3.

[0153] Table 3 Statistical differences among the four models

[0154]

[0155] Note: VS refers to VGG16-SVM; ++ indicates p < 0.01; + indicates p < 0.05; * indicates p > 0.05

[0156] Table 3 shows that VGG16-Maxout-SVM differs significantly from 1D-CNN (p < 0.01), VGG16-LeakyReLU-SVM (p < 0.01), and VGG16-SELU-SVM (p < 0.01) in the classification of rolling contact fatigue states on the experimental dataset. However, due to the mathematical similarity between the LeakyReLU and SELU activation functions, their differences are not as large as those between the LeakyReLU and Maxout activation functions, although the SELU activation function exhibits better performance.

[0157] In summary, the VGG16-Maxout-SVM model used in this invention has the strongest classification performance in rolling contact fatigue state. At the same time, the VGG16-Maxout-SVM model also takes the most time. This is because the neuron-like structure of the Maxout activation function leads to increased model training time.

[0158] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described with reference to preferred embodiments, those skilled in the art should understand that various changes in form and detail can be made without departing from the spirit and scope of the invention as defined in the appended claims. Furthermore, common knowledge such as specific structures and characteristics known in the embodiments is not described in detail here. Finally, the scope of protection claimed by the present invention should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A rolling contact fatigue detection method based on a hybrid neural network, characterized by, The method comprises the following steps: S1: a corresponding hybrid neural network model is constructed based on a neural network and an SVM classifier as a fatigue prediction model; In step S1, the hybrid neural network model is constructed based on a VGG16 network model, wherein the full connection layer and the Softmax layer of the VGG16 network model are removed, and the pool layer of the last layer of the VGG16 network model is connected with the input layer of the SVM classifier to constitute a VGG16-SVM hybrid neural network model; in the hybrid neural network model, the original ReLU activation function of the VGG16 network model is replaced by a Maxout activation function to constitute a VGG16-Maxout-SVM hybrid neural network model; S2: model training is performed on the fatigue prediction model; In step S2, the fatigue prediction model is trained by the following steps: S201: training data for training the fatigue prediction model is prepared; S202: the VGG16 network model used to constitute the hybrid neural network model is trained based on the training data; In step S202, the VGG16 network model is initialized by the Xavier initialization method before being trained; Wherein, the VGG16 network model initialization needs to meet: where: n i represents the number of neurons in the i-th layer of the neural network; Var(W i ) denotes the weight variance of the i-th layer; In step S202, the VGG16 network model is trained by a training strategy based on the momentum gradient descent method of L2 regularization; only the states of the convolution layer and the pool layer of the VGG16 network model are concerned during the training process; The training strategy based on the momentum gradient descent method of L2 regularization updates the parameters by the following formula: Gradient update: g' t = g t + l t-1 ; Weight update: θ t = θ t-1 - μg t - μλθ t-1 ; Momentum update: d t = μd t-1 - αg′ t ; In the above formula: θ represents the initial weight with a regularization term; d represents the initial momentum; t represents the gradient iteration number; g t represents the updated gradient after the tth iteration; θ t represents the updated weight value after the tth iteration; d t represents the updated momentum after the tth iteration, α represents the learning rate; μ represents the momentum coefficient; λ represents the regularization coefficient; S203: the parameters of the trained convolution layer and pool layer of the VGG16 network model are fixed, then the full connection layer and the Softmax layer of the VGG16 network model are removed, and the pool layer of the last layer is connected with the input layer of the SVM classifier to constitute the hybrid neural network model; S204: the parameters of the SVM classifier in the hybrid neural network model are trained based on the training data until the training is completed; S3: the vibration data of the test sample to be detected is input into the trained fatigue prediction model, and feature extraction and feature classification prediction are performed by the neural network and the SVM classifier respectively, and then the corresponding fatigue prediction class probability is output; S4: the fatigue prediction class probability output by the fatigue prediction model is taken as the fatigue detection result of the corresponding test sample to be detected.

2. The hybrid neural network-based rolling contact fatigue detection method of claim 1, wherein, In step S201, the training data is prepared by the following steps: S2011: a rolling contact test is performed on the training sample; S2012: image data and corresponding vibration data of the rolling contact surface of the training sample are obtained; S2013: the corresponding fatigue state of the training sample is calculated based on the image data of the rolling contact surface of the training sample; S2014: the fatigue state of the training sample and the corresponding vibration data are taken as a group of training data.

3. The hybrid neural network-based rolling contact fatigue detection method of claim 2, wherein: In step S2013, the spalling point area of the rolling contact surface of the training sample is calculated based on the image data, and then the training sample is divided into the corresponding fatigue state based on the spalling point area.

4. The hybrid neural network-based rolling contact fatigue detection method of claim 2, wherein: In step S2014, the vibration data corresponding to the fatigue state is converted into an SDP image with a size of 224x224 by the SDP algorithm, and then the fatigue state and the SDP image of the corresponding vibration data are taken as a set of training data.

Images