A bearing fault diagnosis method and system based on deep learning

Abstract

The application discloses a bearing fault diagnosis method and system based on deep learning, and the method comprises the following steps: firstly, obtaining the original vibration signal of a rolling bearing, and based on the original vibration signal, processing the original vibration signal by adopting an AVMD algorithm and a TOTEO algorithm to obtain a data set of various fault types; dividing the data set into a training set and a test set; secondly, establishing a GRU-CNN-GAP network model, training the GRU-CNN-GAP network model by adopting the training set to obtain a bearing fault diagnosis model; and finally, testing the test set by utilizing the bearing fault diagnosis model to obtain a rolling bearing fault diagnosis classification result, and improving the fault diagnosis precision.

Description

Technical Field

[0001] This invention relates to the field of rolling bearing fault diagnosis technology, and specifically to a bearing fault diagnosis method and system based on deep learning. Background Technology

[0002] With the intelligent development of industrial equipment, rolling bearings, as an important component of mechanical equipment, are often affected by harsh environments such as dust, corrosion, and insufficient lubrication in highly integrated and complex operating environments, leading to equipment failure. In particular, if rolling bearing failures are not promptly detected and addressed, they can cause major safety accidents and significant economic losses to enterprises. Therefore, accurate and timely diagnosis of bearing failures is of paramount importance.

[0003] Currently, mainstream fault diagnosis methods can be divided into two main types. One type is fault diagnosis methods based on signal analysis, such as empirical mode decomposition, variational mode decomposition, and singular value decomposition. This method mainly relies on manual extraction of feature values ​​from fault signals to diagnose the fault. The other type is data-driven deep learning methods. This method does not require manual intervention, can automatically extract signal features, and has high accuracy, thus it is widely used in the field of fault diagnosis. Summary of the Invention

[0004] The purpose of this invention is to overcome the above-mentioned technical deficiencies and provide a testing method, device, equipment and medium for front-end and back-end input of gateway products, thereby solving the technical problem of low accuracy in bearing fault diagnosis in the prior art.

[0005] To achieve the above-mentioned technical objectives, the present invention adopts the following technical solution:

[0006] Firstly, the present invention provides a bearing fault diagnosis method based on deep learning, comprising the following steps:

[0007] The original vibration signal of the rolling bearing is obtained. Based on the original vibration signal, the AVMD algorithm and the TOTEO algorithm are used to process the original vibration signal in sequence to obtain a dataset of multiple fault types. The dataset is then divided into a training set and a test set.

[0008] A GRU-CNN-GAP network model is established, and the training set is used to train the GRU-CNN-GAP network model to obtain a bearing fault diagnosis model.

[0009] The bearing fault diagnosis model is used to test the test set to obtain rolling bearing fault diagnosis classification results.

[0010] In some embodiments, the original vibration signal includes at least one of the following: outer ring fault vibration signal, inner ring fault vibration signal, rolling element fault vibration signal, and normal state vibration signal.

[0011] In some embodiments, the original vibration signal is processed sequentially using the AVMD algorithm and the TOTEO algorithm to obtain a dataset of various fault types, including:

[0012] The original vibration signal is decomposed and reconstructed using the VMD algorithm to obtain multiple reconstructed signals. The energy ratios of each reconstructed signal are compared, and the reconstructed signal at which the energy ratio reaches its peak is taken as the fault signal. The VMD algorithm solves the optimal solution of the variational model through an iterative method to complete the adaptive decomposition of the original vibration signal.

[0013] The reconstructed fault signal is processed using the TOTEO algorithm to obtain multiple fault feature amplified signals. Based on the fault feature amplified signals, multiple fault vibration signal matrices are obtained. The fault vibration signal matrices are saved to a dataset to obtain a dataset with multiple fault types.

[0014] In some embodiments, the optimal solution of the VMD algorithm is calculated as follows:

[0015]

[0016]

[0017]

[0018] Among them, u k Let ω be the modal function. k Let λ be the center frequency of each mode, λ be the Lagrange multiplier, S be the signal to be analyzed, and ω be the frequency. In the frequency domain of the modal function, In the frequency domain, where the frequency is... For the frequency domain of Lagrange multipliers, In the frequency domain of the signal to be analyzed, i is the number of points, k is the number of decompositions, and α is the quadratic penalty factor.

[0019] In some embodiments, the energy ratio is calculated as follows:

[0020]

[0021]

[0022] Among them, E A This corresponds to the energy of the A-th component, where A is the number of decompositions, and E... Xδ is the original signal energy, δ is the energy ratio, E is the signal energy, F is the signal sequence, N is the sampling data length, and i is the number of points.

[0023] In some embodiments, the TOTEO algorithm is calculated as follows:

[0024] ψ[x(n)]=[x(n)] 3 -x(n-1) 2 x(n+1)+x(n+1) 2 x(n-1)-x(n)x(n-2)x(n+2),

[0025] Where ψ[x(n)] is the fault characteristic amplified signal, and x(n) is the discrete signal.

[0026] In some embodiments, establishing the GRU-CNN-GAP network model, and training the GRU-CNN-GAP network model using the training set to obtain a bearing fault diagnosis model, includes:

[0027] Step 1: Divide the training set into multiple small batches of training samples using the mini-batch training method;

[0028] Step 2: Input the mini-batch training samples into the GRU-CNN-GAP network model for forward propagation to obtain prediction results;

[0029] Step 3: Obtain the target label; based on the prediction result and the target label, determine the loss function.

[0030] Step 4: Based on the loss function, use the backpropagation algorithm to determine the gradient of the GRU-CNN-GAP network model parameters;

[0031] Step 5: Based on the gradient of the parameters of the GRU-CNN-GAP network model, update the parameters of the GRU-CNN-GAP network model using the Adam optimizer;

[0032] Step 6: Repeat steps 2 to 5 until each batch of training samples completes the training of the GRU-CNN-GAP network model to obtain the bearing fault diagnosis model.

[0033] In some embodiments, the GRU-CNN-GAP network model includes at least an input layer, intermediate layers, and an output layer, wherein the input layer includes at least a gated recurrent unit, the intermediate layers include at least a convolutional neural network, and the output layer includes at least a global average pooling unit; a method for training the GRU-CNN-GAP network model using the mini-batch training samples includes:

[0034] Based on the mini-batch training samples, feature extraction is performed on the mini-batch training samples through the gated recurrent unit of the input layer to obtain sample features;

[0035] Based on the sample features, the sample features are deeply mined through the convolutional neural network of the intermediate layer to determine the differential data in the sample and extract the differential data.

[0036] The pooling kernel value is obtained by performing a pooling kernel operation on the sample data after the difference data is extracted through the global average pooling unit of the output layer.

[0037] Based on the pooling kernel values, the SoftMax classifier is used to classify the pooling kernel values ​​to complete the bearing fault classification.

[0038] In some embodiments, the calculation formula for the gated loop unit is:

[0039] z t =σ(W z ·[h t-1 x t ]),

[0040] r t =σ(W r ·[h t-1 x t ]),

[0041]

[0042]

[0043] Among them, Z t To update the gate, r t To reset the door, h t-1 x is the hidden unit value from the previous time step. t h represents the input data at the current moment. t The output data of the hidden layer. The candidate state is obtained by combining the previous hidden layer state and the current input, where σ is the sigmoid activation function and W is the current hidden layer state. z To update the gate neuron parameters, W r To reset the gate neuron parameters of the gate, W h These are the parameters of the gate neurons in the hidden layer.

[0044] Secondly, the present invention also provides a bearing fault diagnosis system based on deep learning, comprising:

[0045] The dataset acquisition module is used to acquire the original vibration signal of the rolling bearing. Based on the original vibration signal, the AVMD algorithm and the TOTEO algorithm are used to process the original vibration signal in sequence to obtain a dataset of multiple fault types, and the dataset is divided into a training set and a test set.

[0046] The network model training module is used to establish a GRU-CNN-GAP network model. The training set is used to train the GRU-CNN-GAP network model to obtain a bearing fault diagnosis model.

[0047] The bearing fault diagnosis result classification module is used to test the test set using the bearing fault diagnosis model to obtain rolling bearing fault diagnosis classification results.

[0048] Compared with existing technologies, the bearing fault diagnosis method and system based on deep learning provided by this invention uses the AVMD-TOTEO algorithm to process fault signals as input. Data fusion, data truncation, and normalization are performed to generate training samples readable by the GRU layer. The GRU layer extracts features from each training sample with temporal characteristics, capturing and memorizing relevant features. After feature extraction, the memorized information is passed down layer by layer to the next GRU network. A Dropout layer is added to each GRU layer to improve training speed while preventing overfitting. A CNN layer is used to deeply mine the output features of the GRU layer. By constructing multiple 1D-CNN layers, subtle differences between different faults can be extracted from the feature data, further improving the model's diagnostic accuracy. After the CNN layer, a 1D-GAP layer replaces the traditional fully connected layer, significantly reducing the number of model parameters while using global information, further improving the training and diagnosis speed. Finally, a SoftMax classifier is used to classify the output values ​​of the 1D-GAP layer and provide a diagnostic result, improving fault diagnosis accuracy. Attached Figure Description

[0049] Figure 1 This is a flowchart of an embodiment of the bearing fault diagnosis method based on deep learning provided by the present invention;

[0050] Figure 2 This is a flowchart of an embodiment of the bearing fault diagnosis system based on deep learning provided by the present invention;

[0051] Figure 3 This is a schematic diagram of the 1D-GAP pooling process in the bearing fault diagnosis method based on deep learning provided by the present invention.

[0052] Figure 4This is a schematic diagram showing the energy ratios of different reconstructed signals in the bearing fault diagnosis method based on deep learning provided by this invention.

[0053] Figure 5 This is a schematic diagram of the reconstructed signal after AVMD processing in the bearing fault diagnosis method based on deep learning provided by the present invention;

[0054] Figure 6 This is a schematic diagram of the TOTEO signal analysis results in the bearing fault diagnosis method based on deep learning provided by the present invention.

[0055] Figure 7 This is a schematic diagram comparing the kurtosis mean values ​​of different fault types in the bearing fault diagnosis method based on deep learning provided by this invention.

[0056] Figure 8 This is a schematic diagram of the error curve of model training in the bearing fault diagnosis method based on deep learning provided by the present invention.

[0057] Figure 9 This is a comparison diagram of the bearing fault diagnosis method based on deep learning provided by this invention with other models. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0059] This invention provides a bearing fault diagnosis method based on deep learning, wherein AVMD refers to Adaptive Variational Mode Decomposition, TOTEO refers to the third-order Teager energy operator, GRU refers to a gated recurrent unit, CNN refers to a convolutional neural network, and GAP refers to a global mean pooling layer. Figure 1 This is a flowchart of the bearing fault diagnosis method based on deep learning provided in an embodiment of the present invention. Please refer to [link / reference]. Figure 1 The deep learning-based bearing fault diagnosis method includes the following steps:

[0060] S100. Obtain the original vibration signal of the rolling bearing. Based on the original vibration signal, process the original vibration signal sequentially using the AVMD algorithm and the TOTEO algorithm to obtain a dataset of multiple fault types. Divide the dataset into a training set and a test set.

[0061] S200. After establishing the GRU-CNN-GAP network model, the GRU-CNN-GAP network model is trained using the training set to obtain the bearing fault diagnosis model.

[0062] S300. Use the bearing fault diagnosis model to test the test set to obtain the rolling bearing fault diagnosis classification results.

[0063] In this embodiment, firstly, the original vibration signal of the rolling bearing is acquired. Based on the original vibration signal, the AVMD algorithm and the TOTEO algorithm are used sequentially to process the original vibration signal to obtain a dataset of various fault types. The dataset is then divided into a training set and a test set. Secondly, a GRU-CNN-GAP network model is established and trained using the training set to obtain a bearing fault diagnosis model. Finally, the bearing fault diagnosis model is used to test the test set to obtain the rolling bearing fault diagnosis classification results, thereby improving the fault diagnosis accuracy.

[0064] In some embodiments, in step S100, the original vibration signal of the rolling bearing is acquired. The original vibration signal includes at least one of the following: outer ring fault vibration signal, inner ring fault vibration signal, rolling element fault vibration signal, and normal state vibration signal. The original vibration signal contains nine fault types and one normal type. The nine fault types are rolling fault (B), inner ring fault (IR), and outer ring fault (OR). Each fault type is further divided into single-point pits of 0.18, 0.36, and 0.53 mm according to the fault diameter of the bearing. Therefore, there are ten fault labels in total. The original vibration signal is obtained by an accelerometer with a sampling frequency of 12 kHz.

[0065] In some embodiments, the original vibration signal is processed sequentially using AVMD and TOTEO algorithms to obtain a dataset of multiple fault types. First, the original vibration signal is decomposed and reconstructed using the VMD algorithm to obtain multiple reconstructed signals. The energy ratios of each reconstructed signal are compared to determine the reconstructed signal at which the energy ratio reaches its peak value as the fault signal. This process is called AVMD. Second, the reconstructed fault signal is processed using the TOTEO algorithm to obtain multiple fault feature amplification signals. Based on the fault feature amplification signals, multiple fault vibration signal matrices are obtained. Finally, the fault vibration signal matrices are saved to the dataset to obtain a dataset of multiple fault types.

[0066] VMD is an adaptive, fully non-recursive signal processing method that solves the optimal solution of a variational model through iterative methods, thereby achieving adaptive decomposition of the signal. To determine the K modal components with the frequency center and bandwidth, a quadratic penalty factor α and a Lagrange multiplier operator λ(t) are introduced to transform the problem into an unconstrained variational problem, and the alternating direction multiplier method is used to update u. k n+1 ω k n+1 and λ n+1Seeking the optimal solution, the final solution to the optimization problem is:

[0067]

[0068]

[0069]

[0070] Repeat the above steps until the iteration stopping condition is met. Finally, we can decompose the original input signal into K intrinsic mode functions using the VMD algorithm.

[0071] VMD decomposes signals by solving the optimal solution of the constraint variational equations, resulting in orthogonal components. Based on the principle of orthogonal signals, from an energy perspective, the sum of the energies of each intrinsic mode function component should equal the energy of the original signal. To address this principle, an adaptive method for determining the energy ratio of the decomposition iterations is used, known as Adaptive Variational Mode Decomposition (AVMD). The energy ratio is the ratio of the sum of the energies of all intrinsic mode functions obtained after AVMD decomposition to the sum of the energies of the original signal. The formula for calculating the energy ratio is:

[0072]

[0073] Among them, E A This corresponds to the energy of the A-th component, where A is the number of decompositions, and E... X δ is the original signal energy, and δ is the energy ratio. Since VMD is essentially Wiener filtering, the signal energy attenuation will occur in the frequency band far from the center frequency of the component. Therefore, the energy ratio δ is always less than 1.

[0074] The formula for calculating signal energy is:

[0075]

[0076] Where E is the signal energy, F is the signal sequence, N is the sampling data length, and i is the number of points.

[0077] When the K value is too small, the signal decomposition is incomplete, resulting in large energy loss and a small energy ratio. When the K value is too large, the noise will be decomposed into an effective signal. Since the energy of the noise signal is dispersed within the frequency band, the energy of the noise signal is relatively small. Therefore, when the K value is too large, the energy ratio increases slightly. When the energy ratio reaches its extreme value for the first time, K is the optimal decomposition layer. The advantage of the AVMD parameter selection method is that it can remove some interference from the fault signal while highlighting the impulse signal in the original signal.

[0078] In some embodiments, TOTEO is introduced based on the conventional TEO. The fault signal processed by AVMD is processed using the TOTEO algorithm to obtain the fault characteristic amplified signal. TEO, also known as a nonlinear signal analysis operator, can fully reflect the frequency, amplitude, and energy changes of the original signal. The operation of TEO can be divided into nonlinear signals and discrete signals according to the characteristics of the original signal. For a nonlinear signal x(t), TEO can be defined as:

[0079] ψ[x(t)]=[x′(t)] 2 -x(t)x″(t),

[0080] Where x′(t) is the first derivative of x(t), and x″(t) is the second derivative of x(t), for a discrete signal x(n), TEO can be defined as:

[0081] ψ[x(n)]=[x(n)] 2 -x(n-1)x(n+1),

[0082] As shown in the above formula, in the process of analyzing vibration signals, to calculate the Teager energy value of the discrete signal x(n) at time n, only three sample data points—the previous time, the current time, and the next time—are needed. However, since there is often significant noise in the bearing's working environment, in some cases, the noise severely overwhelms the fault impact signal. If TEO is used to analyze the fault signal, the algorithm's noise resistance will be weakened, and it will be unable to effectively highlight the impact component in the signal. Therefore, TOTEO is introduced, and its TOTEO algorithm calculation formula is as follows:

[0083] ψ[x(n)]=[x(n)] 3 -x(n-1) 2 x(n+1)+x(n+1) 2 x(n-1)-x(n)x(n-2)x(n+2),

[0084] Wherein, ψ[x(n)] is the fault characteristic amplified signal, and x(n) is the discrete signal.

[0085] TOTEO is derived from the formula of second-order TEO while ensuring mathematical symmetry. Compared with conventional second-order TEO, TOTEO uses 5 symmetrical sample data in terms of calculation. Therefore, TOTEO can better distinguish the instantaneous changes of the signal in terms of time resolution, detect more instantaneous impulse components in the signal, and amplify the impulse part in the fault signal, thereby further expanding the differences between different faults.

[0086] In some embodiments, based on the outer ring fault vibration signal of the original vibration signal, each outer ring fault vibration signal is decomposed and reconstructed sequentially by the VMD algorithm with the number of decompositions K ranging from 2 to 9. The energy ratios of each reconstructed signal are compared, and the reconstructed signal with the energy ratio reaching the peak value is selected as the fault signal. After the fault is processed by the TOTEO algorithm, the fault feature amplified signal corresponding to the outer ring fault signal is obtained. The outer ring fault vibration signal matrix is ​​obtained based on multiple fault feature amplified signals.

[0087] Based on the original vibration signal, the inner ring fault vibration signal is processed by AVMD and TOTEO algorithm according to the outer ring fault vibration signal processing method. Then, the fault feature amplification signal corresponding to the inner ring fault vibration signal is obtained. The inner ring fault vibration signal matrix is ​​obtained based on multiple fault feature amplification signals.

[0088] Based on the original vibration signal, the rolling element fault vibration signal is processed by AVMD and TOTEO algorithm according to the outer ring fault vibration signal processing method described above. The resulting fault feature amplification signal is then obtained, and the rolling element fault vibration signal matrix is ​​obtained based on the multiple fault feature amplification signals.

[0089] Based on the original vibration signal, the normal vibration signal and the outer ring fault vibration signal processing method described above are used to process each normal state vibration signal sequentially using AVMD and TOTEO algorithms to obtain the normal vibration signal matrix.

[0090] Save the normal matrix, outer ring fault matrix, inner ring fault matrix, and rolling element fault matrix to the dataset.

[0091] In some embodiments, the Min-Max function is used to normalize the data to speed up the gradient descent process to find the optimal solution. The dataset is divided into a training set and a test set according to the aforementioned partitioning method. From each fault set, 70% is randomly selected as the training set and 30% is selected as the test set.

[0092] In some embodiments, in step S200, a GRU-CNN-GAP network model is established. The GRU-CNN-GAP network model includes at least an input layer, an intermediate layer, and an output layer. The input layer includes at least a gated recurrent unit (GRU), the intermediate layer includes at least a convolutional neural network (CNN), and the output layer includes at least a global average pooling unit (GAP). In the GRU-CNN-GAP network model, the GRU layer in the input layer is used to extract features from training samples with time series data. The CNN layer in the intermediate layer performs deep feature mining. The GAP layer in the output layer prevents overfitting during model training and significantly reduces the number of model parameters. The specific implementation of the GRU-CNN-GAP network model is as follows: a network model containing two GRU layers, three 1D-CNN layers, and one 1D-GAP layer is established. The number of neurons in the GRU layer is set to 128 and 64, and the output dimension of the 1D-CNN layer is set to 64, 32, and 10.

[0093] Among them, the GRU network, as a variant of the Long Short-Term Memory network, can effectively solve the gradient explosion and gradient vanishing problems in recurrent neural networks. Compared with the Long Short-Term Memory network, the GRU network has a simpler structure. By merging the forget gate and the input gate into the update gate, the number of model parameters can be reduced, thereby improving its computational speed. When faced with a large dataset, the GRU network can save a lot of training and testing time. The calculation formula of its gated recurrent unit is:

[0094] z t =σ(W z ·[h t-1 x t ]),

[0095] r t =σ(W r ·[h t-1 x t ]),

[0096]

[0097]

[0098] Among them, Z t To update the gate, r t To reset the door, h t-1 x is the hidden unit value from the previous time step. t h represents the input data at the current moment. t The output data of the hidden layer. The candidate state is obtained by combining the previous hidden layer state and the current input, where σ is the sigmoid activation function and W is the current hidden layer state. z To update the gate neuron parameters, Wr To reset the gate neuron parameters of the gate, W h These are the parameters of the gate neurons in the hidden layer.

[0099] As a typical feedforward neural network, the CNN model contains fewer parameters than a regular fully connected neural network, which is beneficial for model training and optimization. The typical structure of a CNN usually consists of an input layer, convolutional layers, pooling layers, fully connected layers, and an output layer. The input layer can be used to store the array transformed from the input data. The role of the convolutional layer is to extract features from the data in the input layer and obtain effective information. The pooling layer is responsible for further selecting effective information from the obtained features and reducing the dimensionality of the feature information to reduce the computation time. The fully connected layer is responsible for connecting all neurons in the feature map and finally outputting the result through the hidden layer.

[0100] A novel pooling structure called one-dimensional global mean pooling (1D-GAP) is used on top of the CNN model. (See [link to relevant documentation]). Figure 3 It is mainly used to reduce the number of model parameters and improve the running speed of the algorithm. It can also adaptively match the dimensionality transformation between CNN layers and SoftMax layers. Assuming that the output feature map size of the last group of 1D convolutional layers is [h×n], where h represents the feature map height and n represents the width, after n [h×n] 1D-GAP pooling kernel operations, each pooling kernel can obtain a value [1×n]. This value is input into the SoftMax classifier to complete the target classification.

[0101] In some embodiments, the training set is input into the established GRU-CNN-GAP network model for training to obtain a bearing fault diagnosis model. This model employs a mini-batch training method and the Adam optimizer to train the GRU-CNN-GAP network model. The training set data is input into the GRU-CNN-GAP network model for training and learning the model parameters. Parameters are adjusted through iterative calculations of forward and backward propagation. The specific steps include:

[0102] Step 1: Divide the training set into multiple small batches of training samples using the mini-batch training method;

[0103] Step 2: Input a small batch of training samples into the GRU-CNN-GAP network model for forward propagation to obtain prediction results;

[0104] Step 3: Obtain the target label and determine the loss function based on the prediction results and the target label;

[0105] Step 4: Based on the loss function, use the backpropagation algorithm to determine the gradient of the GRU-CNN-GAP network model parameters;

[0106] Step 5: Based on the gradient of the GRU-CNN-GAP network model parameters, update the GRU-CNN-GAP network model parameters using the Adam optimizer;

[0107] Step 6: Repeat steps 2 to 5 until each small batch of training samples completes the training of the GRU-CNN-GAP network model to obtain the bearing fault diagnosis model.

[0108] In some embodiments, a method for training the GRU-CNN-GAP network model using mini-batch training samples includes:

[0109] Based on mini-batch training samples, feature extraction is performed on the mini-batch training samples through the gated recurrent unit of the input layer to obtain the relevant features of the samples;

[0110] Based on the features of the samples, the convolutional neural network in the intermediate layer is used to perform in-depth mining of the features of the samples, identify the differential data in the samples, and extract the differential data.

[0111] The pooling kernel value is obtained by performing a pooling kernel operation on the sample data after extracting the differential data through the global average pooling unit of the output layer.

[0112] Based on the pooling kernel values, the SoftMax classifier is used to classify the pooling kernel values ​​to complete the bearing fault classification.

[0113] In some embodiments, in step S300, the test set is input into the bearing fault diagnosis model to perform fault diagnosis and obtain fault assessment and diagnosis results.

[0114] In some embodiments, the provided rolling bearing fault diagnosis model was experimentally verified and its performance analyzed. To verify the effectiveness of the method of the present invention, an experimental analysis was conducted using a dataset of rolling bearing faults at the fan end. Unidirectional acceleration sensors were installed at the drive end (DE), fan end (FE), and base (BA) of the motor to detect the vibration signals of the motor under different faults. Four load conditions were also included: load 1 (1796 r / min), load 2 (1772 r / min), load 3 (1748 r / min), and load 4 (1722 r / min). The fault sampling frequency was 12 kHz, and the sampling time was 10 seconds. Each fault type yielded a tensor matrix of [3, 120000], where 3 represents the vibration signals at three different locations, and 120000 represents the total length of the vibration signals. The rolling bearing fault dataset is shown in Table 1. Due to the large size of the bearing fault dataset, it was impossible to analyze all fault signals. Therefore, only the first segment of the drive end fault signal with fault label 0 in load 1 was used as an example for data processing and analysis.

[0115] Table 1

[0116]

[0117] Using the number of points per revolution of the bearing as a reference, an appropriate sample length was selected. Based on the sampling frequency and the corresponding rotational speeds for the four loads, the range of points per revolution was calculated to be 400.9–418.1. To obtain sufficient data margin, the length of each training sample was ultimately set to 500 data points. After setting the training samples, the number of training samples needed to be increased to improve the model's generalization ability. A sliding window overlapping sampling method was used to generate the training samples, and the sample generation formula is as follows:

[0118]

[0119] Where L is the number of newly generated samples, I is the number of data points in the original fault signal, W is the length of the new training samples, B is the step size of the sliding window, and λ is the sampling overlap rate. Setting the step size of the sliding window to 100 (i.e., the sampling overlap rate is 80%), the number of newly generated samples is calculated to be 1195. To facilitate statistical analysis of the results, the first 1000 samples are taken as training samples. Therefore, the size of each fault set generated after the data truncation operation is [1000, 2500]. Since the normal signal only has sampling data from the drive end and fan end, the data from the base end in the fault signal is discarded to ensure data structure consistency.

[0120] In some embodiments, after preprocessing the raw data, the proposed AVMD method is first used to process the fault signal of the first segment of the drive end in fault label 0. For the energy ratios corresponding to different reconstructed signals, please refer to [link to relevant documentation]. Figure 4 .

[0121] from Figure 4 It can be seen that when the number of decompositions K is set to 7, the energy ratio of the reconstructed signal reaches its peak, and it can retain the impulse signal of the fault signal to the maximum extent while eliminating some interference. For the reconstructed signal after AVMD processing, please refer to [link to AVMD documentation]. Figure 5 .

[0122] from Figure 5 As can be seen, the reconstructed signal after AVMD processing removes some interference and retains the main impulse signals, thereby enhancing the fault characteristics of the original signal. Then, the improved TOTEO algorithm is used to perform signal analysis on the reconstructed fault signal. For the analyzed signal, please refer to [link to analysis]. Figure 6 .

[0123] from Figure 6 It can be seen that the TOTEO algorithm further amplifies the stronger part of the impact signal and conversely, further reduces the weaker part of the impact signal. This can amplify the characteristics of the fault signal and thus increase the differences between various faults. After processing the fault signal of the first segment of the drive end in fault label 0, the same method is used to process other fault signals.

[0124] Kurtosis, as a dimensionless parameter, is highly sensitive to impact signals from rotating machinery. When rotating machinery malfunctions, the impact component increases, leading to a greater kurtosis value in the signal. Furthermore, the larger the impact amplitude, the greater the kurtosis value. The kurtosis values ​​of the signal during processing are recorded and compared; the comparison results can be found in [reference needed]. Figure 7 .

[0125] from Figure 7 As shown in the figure, the kurtosis mean values ​​at AVMD, AVMD-TEO, and AVMD-TOTEO times in fault signal processing were compared. It can be seen from the figure that the kurtosis mean value of the fault signal after AVMD processing changes little, which makes the signal characteristics between different faults not obvious. However, the kurtosis mean values ​​of the fault signals after AVMD-TEO processing are enhanced to varying degrees. This is because the TEO algorithm can fully reflect the changes in frequency, amplitude, and energy of the original signal. On this basis, the proposed AVMD-TOTEO method further enhances the kurtosis mean value of the fault signal, further expanding the differences between faults, thereby improving the detection accuracy of the fault diagnosis model.

[0126] Finally, after fault signal processing, the size of each new fault set is [1000, 2500]. From each fault set, 70% is randomly selected as the training set and 30% as the test set. Simultaneously, the Min-Max function is used to normalize the data to accelerate the gradient descent process and find the optimal solution. The newly generated bearing fault dataset is shown in Table 2.

[0127] Table 2

[0128]

[0129]

[0130] In the design of the GRU-CNN-GAP fault diagnosis model, setting different hyperparameters can affect the diagnostic results. Therefore, repeated parameter tuning experiments were conducted on hyperparameters such as the number of network layers, the number of neurons, and the number of training iterations. Finally, a fault diagnosis model containing two GRU layers and three 1D-CNN layers was established. The number of neurons in the GRU layers was set to 128 and 64, and the output dimensions of the 1D-CNN layers were set to 64, 32, and 10. The fault diagnosis model first uses two GRU layers to extract features from training samples with time series data, capturing and memorizing relevant features. Then, the extracted and memorized features are input into three 1D-CNN layers for deep feature mining, extracting subtle differences between different faults from the data, thereby further improving the model's diagnostic accuracy. After the CNN layers, a 1D-GAP layer replaces the traditional fully connected layer, which can prevent overfitting during model training and significantly reduce the number of model parameters, further improving the training and diagnosis speed. Finally, a SoftMax layer is used to complete bearing fault classification. The hyperparameters of the fault diagnosis model are shown in Table 3.

[0131] Table 3

[0132]

[0133]

[0134] To improve the model's training performance, a mini-batch training method and the Adam optimizer were used in the experiment. The batch size was set to 4 samples per iteration, and the number of iterations was set to 200. For the model training error curve, please refer to [link / reference needed]. Figure 8 As can be seen, the model basically converged and stabilized after 200 training iterations, and no overfitting problem occurred.

[0135] After training the fault diagnosis model through experimental results and comparative analysis, the test set is fed into the designed model for fault diagnosis. In order to verify the feasibility and superiority of the proposed method, the present invention compares the proposed method with the fault diagnosis model before the improvement and the latest fault diagnosis model.

[0136] First, to verify the feasibility of the proposed method, it was compared and analyzed with the GRU-CNN-FC and GRU-CNN-GAP models. GRU-CNN-FC removed the 1D-GAP layer from the original model and replaced it with one Flatten layer and three fully connected layers, setting the number of neurons in the fully connected layers to 128, 64, and 10, respectively. The GRU-CNN-GAP model, on the other hand, was trained and tested without changing the original model. Both models used the original Load 1 bearing fault dataset for their training and testing sets. The diagnostic results of each model are shown in Table 4.

[0137] Table 4

[0138]

[0139] Table 4 uses time and accuracy metrics to evaluate the results, where P represents precision, R represents sensitivity, and F1 represents the harmonic mean. It can be seen that using the GAP layer reduces training and testing time; in terms of diagnostic accuracy, the GRU-CNN-GAP model achieves higher feature extraction capability and fault identification rate; and after incorporating the AVMD-TOTEO signal processing method, the F1 mean is further improved to 100% based on the AVMD-TOTEO and GRU-CNN-GAP methods.

[0140] To further verify the superiority of the proposed method, it was compared and analyzed with the ACD-K-Means diagnostic model proposed by Lanjun Wan, the SN-CNN diagnostic model proposed by Dhiraj Neupane, and the WGWOA-VMD-SVM diagnostic model proposed by Junbo Zhou. A confusion matrix was used to quantify the diagnostic results. The diagnostic results can be found in [link to relevant documentation]. Figure 9 .

[0141] The diagnostic results show that using the latest ACD-K-Means model, SN-CNN model, and WGWOA-VMD-SVM model can improve the fault diagnosis accuracy to 97.8%, 99.4%, and 99.66%, respectively. The proposed bearing fault diagnosis method achieves 100% accuracy compared to the other three models. Experimental results demonstrate that the fault diagnosis method based on AVMD-TOTEO and GRU-CNN-GAP has certain advantages over other existing models.

[0142] Based on the above-described deep learning-based bearing fault diagnosis method, this invention also provides a deep learning-based bearing fault diagnosis system. Please refer to [link to relevant documentation]. Figure 2 The deep learning-based bearing fault diagnosis system 500 includes a dataset acquisition module 510, a network model training module 520, and a bearing fault diagnosis result classification module 530.

[0143] The dataset acquisition module 510 is used to acquire the original vibration signal of the rolling bearing. Based on the original vibration signal, the AVMD algorithm and TOTEO algorithm are used to process the original vibration signal in sequence to obtain datasets of various fault types, and the dataset is divided into training set and test set.

[0144] The network model training module 520 is used to build a GRU-CNN-GAP network model. The training set is used to train the GRU-CNN-GAP network model to obtain a bearing fault diagnosis model.

[0145] The bearing fault diagnosis result classification module 530 is used to test the test set using the bearing fault diagnosis model in order to obtain the rolling bearing fault diagnosis classification result.

[0146] In summary, the bearing fault diagnosis method and system based on deep learning provided by this invention first acquires the original vibration signal of the rolling bearing. Based on the original vibration signal, the AVMD algorithm and the TOTEO algorithm are used sequentially to process the original vibration signal to obtain a dataset of various fault types, and the dataset is divided into a training set and a test set. Second, a GRU-CNN-GAP network model is established and trained using the training set to obtain a bearing fault diagnosis model. Finally, the bearing fault diagnosis model is used to test the test set to obtain the rolling bearing fault diagnosis classification results, thereby improving the fault diagnosis accuracy.

[0147] The specific embodiments of the present invention described above do not constitute a limitation on the scope of protection of the present invention. Any other corresponding changes and modifications made in accordance with the technical concept of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A bearing fault diagnosis method based on deep learning, characterized in that, Includes the following steps: The original vibration signal of the rolling bearing is obtained. Based on the original vibration signal, the AVMD algorithm and the TOTEO algorithm are used to process the original vibration signal in sequence to obtain a dataset of multiple fault types. The dataset is divided into a training set and a test set. The original vibration signal includes at least one of the following: outer ring fault vibration signal, inner ring fault vibration signal, rolling element fault vibration signal, and normal state vibration signal. The fault types include outer ring fault, inner ring fault, and rolling fault. The original vibration signal is processed sequentially using the AVMD algorithm and the TOTEO algorithm to obtain datasets of various fault types, including: The original vibration signal is decomposed and reconstructed using the VMD algorithm to obtain multiple reconstructed signals. The energy ratios of each reconstructed signal are compared, and the reconstructed signal at which the energy ratio reaches its peak is taken as the fault signal. The VMD algorithm solves the optimal solution of the variational model through an iterative method to complete the adaptive decomposition of the original vibration signal. The reconstructed fault signal is processed using the TOTEO algorithm to obtain multiple fault feature amplified signals. Based on the fault feature amplified signals, multiple fault vibration signal matrices are obtained. The fault vibration signal matrices are saved to a dataset to obtain a dataset with multiple fault types. A GRU-CNN-GAP network model is established, and the training set is used to train the GRU-CNN-GAP network model to obtain a bearing fault diagnosis model. The GRU-CNN-GAP network model includes at least an input layer, an intermediate layer, and an output layer. The input layer includes at least a gated recurrent unit, the intermediate layer includes at least a convolutional neural network, and the output layer includes at least a global average pooling unit. The bearing fault diagnosis model is used to test the test set to obtain rolling bearing fault diagnosis classification results.

2. The bearing fault diagnosis method based on deep learning according to claim 1, characterized in that, The optimal solution of the VMD algorithm is calculated as follows: ， ， ， in, For modal functions, The center frequency of each mode, For Lagrange multipliers, For the signal to be analyzed, For frequency, In the frequency domain of the modal function, In the frequency domain, where the frequency is... For the frequency domain of Lagrange multipliers, For the frequency domain of the signal to be analyzed, It's points. For the number of decompositions, It is a secondary penalty factor.

3. The bearing fault diagnosis method based on deep learning according to claim 1, characterized in that, The energy ratio calculation formula is: ， ， in, It corresponds to the first One unit of energy It is the number of decompositions. It is the original signal energy, at the same time It's the energy ratio. It is signal energy. It is a signal sequence. It is the length of the sampled data. It's points.

4. The bearing fault diagnosis method based on deep learning according to claim 1, characterized in that, The TOTEO algorithm calculation formula is as follows: ， in, This is an amplified signal representing the fault characteristics. It is a discrete signal.

5. The bearing fault diagnosis method based on deep learning according to claim 1, characterized in that, The establishment of the GRU-CNN-GAP network model, and the training of the GRU-CNN-GAP network model using the training set to obtain the bearing fault diagnosis model, includes: Step 1: Divide the training set into multiple small batches of training samples using the mini-batch training method; Step 2: Input the mini-batch training samples into the GRU-CNN-GAP network model for forward propagation to obtain prediction results; Step 3: Obtain the target label, and determine the loss function based on the prediction result and the target label; Step 4: Based on the loss function, use the backpropagation algorithm to determine the gradient of the GRU-CNN-GAP network model parameters; Step 5: Based on the gradient of the parameters of the GRU-CNN-GAP network model, update the parameters of the GRU-CNN-GAP network model using the Adam optimizer; Step 6: Repeat steps 2 to 5 until each batch of training samples completes the training of the GRU-CNN-GAP network model to obtain the bearing fault diagnosis model.

6. The bearing fault diagnosis method based on deep learning according to claim 5, characterized in that, The method for training the GRU-CNN-GAP network model using the mini-batch training samples includes: Based on the mini-batch training samples, feature extraction is performed on the mini-batch training samples through the gated recurrent unit of the input layer to obtain sample features; Based on the sample features, the sample features are deeply mined through the convolutional neural network of the intermediate layer to determine the differential data in the sample and extract the differential data. The pooling kernel value is obtained by performing a pooling kernel operation on the sample data after extracting the differential data through the global average pooling unit of the output layer. Based on the pooling kernel values, the SoftMax classifier is used to classify the pooling kernel values ​​to complete the bearing fault classification.

7. The bearing fault diagnosis method based on deep learning according to claim 6, characterized in that, The calculation formula for the gated loop unit is: ， ， ， ， in, To update the door, To reset the door, The value of the hidden unit in the previous time step. The input data at the current moment, The output data of the hidden layer. The candidate state is obtained by performing a composite operation on the previous hidden layer state and the current input. It is the sigmoid activation function. To update the gate neuron parameters of the gate, To reset the gate neuron parameters, These are the parameters of the gate neurons in the hidden layer.

8. A bearing fault diagnosis system based on deep learning, characterized in that, include: The dataset acquisition module is used to acquire the original vibration signal of the rolling bearing. Based on the original vibration signal, the AVMD algorithm and the TOTEO algorithm are used to process the original vibration signal in sequence to obtain a dataset of multiple fault types. The dataset is divided into a training set and a test set. The original vibration signal includes at least one of the following: outer ring fault vibration signal, inner ring fault vibration signal, rolling element fault vibration signal, and normal state vibration signal. The fault types include outer ring fault, inner ring fault, and rolling fault. The original vibration signal is processed sequentially using the AVMD algorithm and the TOTEO algorithm to obtain datasets of various fault types, including: The original vibration signal is decomposed and reconstructed using the VMD algorithm to obtain multiple reconstructed signals. The energy ratios of each reconstructed signal are compared, and the reconstructed signal at which the energy ratio reaches its peak is taken as the fault signal. The VMD algorithm solves the optimal solution of the variational model through an iterative method to complete the adaptive decomposition of the original vibration signal. The reconstructed fault signal is processed using the TOTEO algorithm to obtain multiple fault feature amplified signals. Based on the fault feature amplified signals, multiple fault vibration signal matrices are obtained. The fault vibration signal matrices are saved to a dataset to obtain a dataset with multiple fault types. The network model training module is used to establish a GRU-CNN-GAP network model. The training set is used to train the GRU-CNN-GAP network model to obtain a bearing fault diagnosis model. The bearing fault diagnosis result classification module is used to test the test set using the bearing fault diagnosis model to obtain rolling bearing fault diagnosis classification results. The GRU-CNN-GAP network model includes at least an input layer, an intermediate layer, and an output layer. The input layer includes at least a gated recurrent unit, the intermediate layer includes at least a convolutional neural network, and the output layer includes at least a global average pooling unit.

Images