A method for diagnosing rolling bearing faults

Abstract

This invention discloses a method for diagnosing rolling bearing faults, including experimental setup, signal reconstruction, model extraction, fault diagnosis, and verification. The method involves building a rolling bearing fault simulation experimental setup and collecting the original vibration signals from the setup. A dual-tree complex wavelet packet transform is used to decompose the vibration signals into high-frequency wavelet coefficients. Threshold denoising is then applied to these high-frequency wavelet coefficients, and the denoised wavelet coefficients are reconstructed to obtain the denoised reconstructed signal. A cyclic spectrum coherence two-dimensional feature extraction model is established for the reconstructed signal. A multi-scale convolutional layer parallel structure is built using multi-scale convolutional kernels to extract input features at different scales. A multi-scale convolutional neural network is constructed for fault type diagnosis, and the diagnostic accuracy is determined. The diagnostic accuracy of the input and output shaft bearings under different operating conditions is compared to verify the stability and generalization ability of the method. This invention can improve the accuracy of rolling bearing fault diagnosis and has excellent generalization ability.

Description

Technical Field

[0001] This invention relates to the field of bearing fault diagnosis technology, and in particular to a method for diagnosing rolling bearing faults. Background Technology

[0002] The rolling bearing fault diagnosis method is based on dual-tree complex wavelet packet transform thresholding, cyclic spectral coherence, and multi-scale convolutional neural networks. Rolling bearings are widely used in mechanical equipment in key fields such as aerospace and industry. Due to their complex operating conditions and high workload, rolling bearings are prone to failure, and accurate fault diagnosis is crucial to ensuring the safe and efficient operation of equipment. The vibration signal of rolling bearings is affected by coupling interference from other components. Eliminating noise, obtaining appropriate features that are well-correlated with the bearing's health status, and making accurate predictions are current challenges. With the continuous development of technology, the manufacturing process requirements for rolling bearing fault diagnosis methods are also becoming increasingly stringent.

[0003] Existing methods for diagnosing rolling bearing faults have certain drawbacks. Current fault diagnosis methods often employ wavelet analysis to suppress noise when extracting signal features. However, this approach fails to guarantee effective noise reduction and strong feature separation and extraction capabilities when dealing with situations involving coordinated operation of components, environmental interference, and significant noise levels. When selecting time-frequency domain features from the denoised signal, a common method is to compare each group of signals and select based on feature distinctiveness. However, this approach requires re-experimenting to develop new features for different practical problems during preprocessing, resulting in poor generalization ability and consuming considerable time for manual selection, thus introducing subjectivity. Currently used two-dimensional feature representation methods include continuous wavelet transform, improved DS evidence fusion, grayscale images, and infrared thermal images. However, these methods are gradually deviating from vibration signal-based diagnostic methods, and surface images such as grayscale and infrared cannot monitor internal information. In state recognition, traditional theoretical methods such as ANN, SVM, k-nearest neighbor (kNN), and probabilistic graphical models (PGM) are limited by their shallow architectures. In increasingly complex application scenarios, they struggle to effectively learn distinguishing features from raw high-dimensional inputs, and their diagnostic performance depends on expert diagnostic experience and the selection of diagnostic features. In the process of classification and recognition using convolutional neural networks (CNNs), the uniqueness of the convolution kernel scale leads to the loss of spatiotemporal information when extracting features using convolutional layers, ignoring a large amount of valuable local information and causing a decline in model classification performance. This negatively impacts user experience. Therefore, we propose a rolling bearing fault diagnosis method. Summary of the Invention

[0004] Technical problem to be solved: In view of the shortcomings of the prior art, the present invention provides a rolling bearing fault diagnosis method, which can improve the accuracy of rolling bearing fault diagnosis, has good generalization ability, and can effectively solve the problems in the background art.

[0005] Technical Solution: To achieve the above objectives, the technical solution adopted by this invention is: a method for diagnosing rolling bearing faults, comprising the following operational steps:

[0006] S1: Experimental bench setup: Set up a rolling bearing fault simulation experimental bench and collect the original vibration signals of the simulation experimental bench;

[0007] S2: Signal reconstruction: The vibration signal is decomposed using dual-tree complex wavelet packet transform to obtain high-frequency wavelet coefficients, then threshold denoising is performed on the high-frequency wavelet coefficients, and the denoised wavelet coefficients are reconstructed to obtain the denoised reconstructed signal.

[0008] S3: Model Extraction: Establish a two-dimensional feature extraction model for the reconstructed signal based on cyclic spectrum coherence;

[0009] S4: Fault Diagnosis: A multi-scale convolutional layer parallel structure is built using multi-scale convolutional kernels to extract input features at different scales, construct a multi-scale convolutional neural network for fault type diagnosis, and determine the diagnostic accuracy.

[0010] S5: Verification: Compare the diagnostic accuracy of the input and output shaft bearings under different operating conditions to verify the stability and generalization ability of this method.

[0011] As a preferred technical solution of this application, in step S1, the original vibration signal of the output shaft is measured by a vibration sensor. The original vibration signal includes five types: normal rolling bearing condition, outer ring crack, inner ring crack, roller crack, and mixed fault condition. The mixed fault condition includes at least two types of outer ring crack, inner ring crack, and roller crack.

[0012] As a preferred technical solution of this application, the dual-tree complex wavelet packet transform in step S2 is an improvement on the dual-tree complex wavelet transform, which realizes the decomposition and reconstruction of the signal through parallel transformation of the real and imaginary parts of the discrete wavelet.

[0013] As a preferred technical solution of this application, in step S2, the dual-tree complex wavelet transform decomposes the original vibration signal into 5 layers, uses the 'db14' wavelet to construct a first-level filter and a general filter, and selects a Q-shift filter with a length of 14 as the dual-tree filter.

[0014] As a preferred technical solution of this application, the S2 step uses a soft threshold denoising method to process the denoised signal. The threshold selection method is to use an unbiased likelihood estimation threshold estimator, which is an adaptive threshold selection based on the unbiased likelihood estimation principle.

[0015] As a preferred technical solution of this application, the specific steps for constructing a multi-scale convolutional neural network for fault identification in step S4 are as follows:

[0016] A1: Construct a multi-scale convolutional neural network;

[0017] A2: Training is performed using samples, wherein each fault signal has 280 sets of labeled samples, and the samples are divided into training set, validation set and test set in a ratio of 7:1:2;

[0018] A3: Visualize the t-SNE clustering distribution of activations in early max pooling layers, final convolutional layers, and final softmax layers of multi-scale convolutional neural networks;

[0019] A4: Use a confusion matrix to summarize the prediction results of the classification model.

[0020] As a preferred technical solution of this application, in step S4, a multi-scale convolutional neural network is constructed. The network structure has a total of 7 convolutional layers, followed by a pooling layer. The input two-dimensional features are transmitted to three convolutional layers with different kernel sizes in parallel to extract multi-scale features. The parallel structure includes three parallel branch structures, and each branch structure has different convolutional kernels with kernel sizes of 3 × 3, 5 × 5 and 7 × 7, respectively. Different kernel sizes extract features at different scales.

[0021] As a preferred technical solution of this application, the pooling operation in step S4 adopts a strategy combining max pooling and average pooling. Max pooling is used in the early network structure to preserve the significant features of the signal, while average pooling is added in the later network structure to consider all important feature elements of the signal.

[0022] As a preferred technical solution of this application, the specific steps of step S5 are as follows:

[0023] B1: Change the rolling bearing operating conditions and use the same method as above to classify and diagnose the five types of output shaft fault bearing data;

[0024] B2: To further verify the generalization ability of this method, the load and speed conditions were changed respectively, and the input shaft bearing was also included in the diagnostic conditions for verification. The same method was used to classify and diagnose the faulty bearing data under different conditions.

[0025] Beneficial Effects: Compared with the prior art, the present invention provides a rolling bearing fault diagnosis method with the following beneficial effects: This rolling bearing fault diagnosis method can improve the accuracy of rolling bearing fault diagnosis and has good generalization ability. The method of the present invention is applicable to situations where components operate in coordination, the working environment is interfered with, and there is a lot of noise. It can effectively reduce noise in the signal and extract distinctive two-dimensional features. It uses a multi-scale convolutional neural network to achieve high classification performance, with a comprehensive diagnostic accuracy of over 95%. Moreover, the method of the present invention can achieve accurate classification under different working conditions and has good generalization ability. The overall rolling bearing fault diagnosis method is more effective than traditional methods. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the overall structure of a rolling bearing fault diagnosis method according to the present invention.

[0027] Figure 2 This is a diagram showing the overall structure of a multi-scale convolutional neural network in a rolling bearing fault diagnosis method of the present invention.

[0028] Figure 3 This is a diagram of the multi-scale convolutional neural network structure in a rolling bearing fault diagnosis method of the present invention. Implementation

[0029] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings and specific embodiments. However, those skilled in the art will understand that the embodiments described below are some embodiments of the present invention, but not all embodiments, and are only used to illustrate the present invention, and should not be regarded as limiting the scope of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Where specific conditions are not specified in the embodiments, conventional conditions or conditions recommended by the manufacturer shall be followed. Where the manufacturers of reagents or instruments are not specified, they are all conventional products that can be purchased commercially.

[0030] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0031] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0032] like Figure 1-3 As shown, a method for diagnosing rolling bearing faults includes the following steps:

[0033] S1: Experimental bench setup: Set up a rolling bearing fault simulation experimental bench and collect the original vibration signals of the simulation experimental bench;

[0034] S2: Signal reconstruction: The vibration signal is decomposed using dual-tree complex wavelet packet transform to obtain high-frequency wavelet coefficients. Then, threshold noise reduction is performed on the high-frequency wavelet coefficients, and the noise-reduced wavelet coefficients are reconstructed to obtain the noise-reduced reconstructed signal.

[0035] S3: Model Extraction: Establish a two-dimensional feature extraction model for cyclic spectrum coherence of the reconstructed signal;

[0036] S4: Fault Diagnosis: A multi-scale convolutional layer parallel structure is built using multi-scale convolutional kernels to extract input features at different scales, construct a multi-scale convolutional neural network for fault type diagnosis, and determine the diagnostic accuracy.

[0037] S5: Verification: Compare the diagnostic accuracy of input and output shaft bearings under different operating conditions to verify the stability and generalization ability of this method. The method of this invention is applicable to situations where components in the system operate in coordination, the working environment is disturbed, and there is a lot of noise. It can effectively reduce the noise of the signal and extract distinctive two-dimensional features. It uses a multi-scale convolutional neural network to achieve high classification performance, with a comprehensive diagnostic accuracy of over 95%. Moreover, the method of this invention can achieve accurate classification under different operating conditions and has a good generalization ability.

[0038] Furthermore, in step S1, the original vibration signal of the output shaft is measured by a vibration sensor. The original vibration signal includes five types: normal rolling bearing condition, outer ring crack, inner ring crack, roller crack, and mixed fault condition. The mixed fault condition includes at least two types of outer ring crack, inner ring crack, and roller crack.

[0039] Furthermore, in step S2, the dual-tree complex wavelet packet transform is an improvement on the dual-tree complex wavelet transform, which achieves signal decomposition and reconstruction through parallel transformation of the real and imaginary parts of the discrete wavelet.

[0040] Furthermore, in step S2, the dual-tree complex wavelet transform decomposes the original vibration signal into 5 layers. The 'db14' wavelet is used to construct a first-level filter and a general filter, and a Q-shift filter with a length of 14 is selected as the dual-tree filter.

[0041] Furthermore, in step S2, a soft thresholding method is used to process the denoised signal. The threshold selection method is to use an unbiased likelihood estimation threshold estimator, which is an adaptive threshold selection based on the principle of unbiased likelihood estimation.

[0042] Furthermore, the specific steps for constructing a multi-scale convolutional neural network for fault identification in step S4 are as follows:

[0043] A1: Construct a multi-scale convolutional neural network;

[0044] A2: Training is performed using samples, where each fault signal has 280 sets of labeled samples, and the samples are divided into training set, validation set and test set in a ratio of 7:1:2;

[0045] A3: Visualize the t-SNE clustering distribution of activations in early max pooling layers, final convolutional layers, and final softmax layers of multi-scale convolutional neural networks;

[0046] A4: Use a confusion matrix to summarize the prediction results of the classification model.

[0047] Furthermore, in step S4, a multi-scale convolutional neural network is constructed. The network structure has a total of 7 convolutional layers, followed by a pooling layer. The input two-dimensional features are transmitted to three convolutional layers with different kernel sizes in parallel to extract multi-scale features. The parallel structure contains three parallel branch structures, and each branch structure has a different convolutional kernel with kernel sizes of 3 × 3, 5 × 5 and 7 × 7, respectively. Different kernel sizes extract features at different scales.

[0048] Furthermore, in step S4, the pooling operation adopts a strategy that combines max pooling and average pooling. Max pooling is used in the early network structure to preserve the salient features of the signal, while average pooling is added in the later network structure to consider all important feature elements of the signal.

[0049] Furthermore, the specific steps of step S5 are as follows:

[0050] B1: Change the rolling bearing operating conditions and use the same method as above to classify and diagnose the five types of output shaft fault bearing data;

[0051] B2: To further verify the generalization ability of this method, the load and speed conditions were changed respectively, and the input shaft bearing was also included in the diagnostic conditions for verification. The same method was used to classify and diagnose the faulty bearing data under different conditions. Example

[0052] like Figure 1 As shown, the specific steps of the rolling bearing fault diagnosis method of the present invention are as follows:

[0053] Step 1: Build a rolling bearing fault simulation test bench, collect vibration data from the simulation test bench, and obtain the original vibration signals of the output shaft in five types, including the normal state of the rolling bearing, outer ring crack, inner ring crack, roller crack, and mixed fault state.

[0054] Step 2: The vibration signal is decomposed using Dual-Tree Complex Wavelet Packet Transform (DTCWPT) to obtain high-frequency wavelet coefficients. Then, threshold denoising is applied to the high-frequency wavelet coefficients, and the denoised wavelet coefficients are reconstructed to obtain the denoised reconstructed signal. The steps are as follows:

[0055] (1) Based on the dual-tree complex wavelet transform, the decomposition and reconstruction of the signal are achieved by parallel transformation of the real and imaginary parts of the discrete wavelet. The mathematical model is as follows:

[0056] make and Let represent the wavelet function and represent the corresponding scaling function, respectively. The complex wavelet transform can be expressed as the following wavelet function and scaling function:

[0057]

[0058] in It is a real even number. They are imaginary numbers and odd numbers;

[0059] The choice of the two real and imaginary parts makes them form a Hilbert transform pair:

[0060]

[0061] (2) The general process of dual-tree complex wavelet packet transform (DTCWPT) is based on a filter bank, which consists of three types of filters:

[0062] First-stage filter bank and Dual-tree filter and Universal Filter ;

[0063] To improve the efficiency of complementary branch coefficients and enhance the reliability of the transform, the low-pass filters of the two branches must be delayed by approximately half a sample, as shown in the following equation:

[0064]

[0065] Among them, the filter It is a fractional delay filter that can obtain (or approximately obtain) half of the sampling delay, where L is... The latency, The filter to be determined is shown by the asterisk (*), which indicates convolution.

[0066] The filter block structure and decomposition and reconstruction process of Dual-Tree Complex Wavelet Packet Transform (DTCWPT) are as follows: Figure 2 As shown.

[0067] (3) An unbiased likelihood estimation (SURE) threshold estimator is adopted, and an adaptive threshold selection based on the principle of unbiased likelihood estimation (quadratic equation) is performed. Its mathematical expression is as follows:

[0068]

[0069] Command signal It is a discrete time series. In the formula, is An ascending sequence.

[0070] (4) Using the soft thresholding method makes the denoised signal smoother and has better continuity. The expression for soft thresholding denoising is:

[0071]

[0072] In the formula These are threshold wavelet coefficients; These are the wavelet coefficients of the decomposition. It is a threshold;

[0073] After threshold denoising of the high-frequency wavelet coefficients after dual-tree complex wavelet packet decomposition, the denoised wavelet coefficients are reconstructed to obtain the denoised reconstructed signal.

[0074] Step 3: Apply cyclic spectral coherence analysis to the reconstructed signal to separate interference, detect and identify hidden periodic behavior in the signal, and finally extract two-dimensional fault features of the signal. The steps are as follows:

[0075] (1) Define a cyclic stationary signal The second moment is the instantaneous autocorrelation function (ACF) with a cyclic period T, and its formula is:

[0076]

[0077] Where * denotes complex conjugation, T The time lag is E, and the statistical average is E.

[0078] (2) Obtain the Fourier coefficients of the instantaneous autocorrelation function (ACF) as follows:

[0079]

[0080] in Let be the cycle frequency. From equation (7), it can be observed that the cyclic ACF represents... Relative to time-delay signals Fourier coefficients.

[0081] (3) The second-order statistical descriptor for cyclic stationarity is called the cyclic spectral correlation (CSC), which is calculated by performing a Fourier transform on the cyclic ACF, as shown in equation (8):

[0082]

[0083] CSC is also defined as the double Fourier transform of a signal, which is the frequency spectrum. and A function of the cycle frequency. Utilizing an additional frequency dimension, and The carrier wave and its modulation were revealed respectively. When When it equals 0, it corresponds to the classical power spectrum. On the other hand, when... When it is not equal to 0, it represents the power spectrum of that particular cyclic component.

[0084] (4) Use CSCoh to measure the correlation between two spectral components, and estimate it using equation (9):

[0085]

[0086] CSCoh can be interpreted as CSC for whitening signals. It tends to equalize regions with very different energy levels and amplify weak cyclic stationary signals.

[0087] Step 4: Construct a multi-convolutional layer parallel structure using multi-scale convolutional kernels, extract input features at different scales, build a multi-scale convolutional neural network (MSCNN) for fault type diagnosis, and determine the diagnostic accuracy. The steps are as follows:

[0088] (1) Construct a multi-scale convolutional neural network (MSCNN), such as Figure 2 As shown;

[0089] (2) Training is performed using samples, where each fault signal has 280 sets of labeled samples, and the samples are divided into training set, validation set and test set in a ratio of 7:1:2.

[0090] (3) Visualize the t-SNE clustering distribution of the activations of the early max pooling layer, the final convolutional layer, and the final softmax layer in the multi-scale convolutional neural network (MSCNN);

[0091] (4) Use confusion matrix to summarize the prediction results of classification model.

[0092] Convolutional Neural Networks (CNNs) have three main characteristics: local receptive fields, shared weights, and pooling. A CNN typically consists of multiple alternating convolutional and pooling layers, ultimately connected to a fully connected layer. The mathematical expression for the convolution process is as follows:

[0093]

[0094] in This represents the element in the i-th row and j-th column of the k-th output feature map of the l-th layer. and It is the element at the corresponding position of the k-th input feature map and the convolution kernel. It is each Size, The weights are represented by '*', and '*' represents the convolution operation. It is an activation function.

[0095] Next, downsampling is typically performed using max pooling or average pooling to preserve salient features, reduce feature dimensionality, and increase the receptive field. This invention employs a combined max pooling and average pooling strategy. Max pooling is used in the early stages of the network structure to retain salient features, while average pooling is added in the later stages to consider all important feature elements. This combined application of the two pooling methods accelerates CNN computation while preserving more discriminative features. The general expression for the pooling method is:

[0096]

[0097] in It is max pooling. It is average pooling. It is an element of the k-th output feature map after pooling. Each is of size The pooling region;

[0098] After transformations by convolutional and pooling layers, the deep features are finally fed into fully connected (FC) layers. To obtain a sufficient number of non-linear features, the FC layers process the feature maps, concentrating a large number of parameters in the neural network, which can easily lead to overfitting. To avoid overfitting, a Dropout layer is placed at the beginning of the network, randomly activating network units according to selected probabilities, weakening the dependencies between neurons and increasing the network's generalization ability.

[0099] Step 5: Compare the diagnostic accuracy of the input and output shaft bearings under different operating conditions to verify the stability and generalization ability of this method. The steps are as follows:

[0100] (1) Change the rolling bearing operating conditions and use the same method as above to classify and diagnose the five types of output shaft fault bearing data;

[0101] (2) To further verify the generalization ability of this method, the load and speed conditions were changed respectively, and the input shaft bearing was also added to the diagnostic conditions for verification. The same method was used to classify and diagnose the faulty bearing data under different conditions.

[0102] The embodiment combines experiments to comprehensively apply the above methods for rolling bearing fault diagnosis. Vibration data is collected by vibration sensors mounted on the housing, with the sensor for collecting output shaft vibration signals mounted on the upper side of the output shaft, and a sampling frequency of 8kHz. The experiment collects data from an output shaft bearing fault simulation kit, with the motor speed set to 1350 r / min and the load torque to 3 N / m.

[0103] The original waveform of the signal acquired by the vibration sensor is used. DTCWPT is used to extract frequency-band features from the original time-domain signal, with a decomposition level of 5. SURE threshold estimation is used to perform soft thresholding noise reduction on the high-frequency wavelet coefficients, and the reconstructed signal is obtained after inverse transformation.

[0104] After signal denoising and reconstruction, CSCoh was used for two-dimensional feature extraction. In the CSCoh results, NC exhibits significant axial frequencies and harmonics in the 1000-2000 Hz spectral range; OR and IR both show different defect frequencies and harmonics, with better characterization of axial frequencies and their harmonics; RC and HF, while reflecting axial frequencies and harmonics, show differentiated fault frequencies and harmonics in the lower spectral frequency range. By comparing the CSCoh processing results of the original and reconstructed signals, it can be concluded that the two-dimensional characterization of the reconstructed signal significantly improves the defect frequency characterization compared to the direct processing of the original signal, and the distinguishability of various fault features is greater.

[0105] A multi-scale convolutional neural network (MSCNN) is constructed for fault identification. The structure of the constructed MSCNN is as follows: Figure 3As shown, the network structure contains 7 convolutional layers, each followed by a pooling layer. The input 2D features are fed into three convolutional layers with different kernel sizes in parallel to extract multi-scale features. First, max pooling is performed on convolutional layers with kernel sizes of 5×5 and 3×3, followed by fusion. The other branch of the parallel structure uses the same structure and fusion method. After the first stage of fusion, a pair of parallel fusion layers are obtained, followed by identical convolutional and average pooling layers to accelerate the process. After pooling, the resulting pair of feature values ​​are fused a second time. Finally, the data fused with multi-scale features, after convolution and pooling, enters the Dropout layer before classification, with a dropout rate set to 0.4. After constructing the Multi-Scale Convolutional Neural Network (MSCNN), it is trained using samples. Each signal class has 280 labeled samples, which are divided into training, validation, and test sets in a 7:1:2 ratio. Finally, the confusion matrix was used to summarize the prediction results of the classification model, and the overall accuracy reached 95.7%, proving that the method can accurately classify five types of bearing faults.

[0106] It should be noted that, in this document, relational terms such as first and second (number one, number two), etc., are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0107] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A method for diagnosing rolling bearing faults, characterized in that: The following steps are included: S1: Experimental bench setup: Set up a rolling bearing fault simulation experimental bench and collect the original vibration signals of the simulation experimental bench; S2: Signal reconstruction: The vibration signal is decomposed using dual-tree complex wavelet packet transform to obtain high-frequency wavelet coefficients, then threshold denoising is performed on the high-frequency wavelet coefficients, and the denoised wavelet coefficients are reconstructed to obtain the denoised reconstructed signal. S3: Model Extraction: Establish a two-dimensional feature extraction model for the reconstructed signal based on cyclic spectrum coherence; S4: Fault Diagnosis: A multi-scale convolutional layer parallel structure is constructed using multi-scale convolutional kernels to build a multi-scale convolutional neural network. The network structure has a total of 7 convolutional layers, each followed by a pooling layer. The input two-dimensional features are transmitted in parallel to three convolutional layers with different kernel sizes to extract multi-scale features. This parallel structure includes three parallel branches, each with different convolutional kernels of sizes 3 × 3, 5 × 5, and 7 × 7. Different kernel sizes extract features at different scales. The 5 × 5 and 3 × 7 kernels are used first to extract the input features. After the convolutional layers complete max pooling, they are fused. The other branch of the parallel structure also adopts the same structure and fusion method. After the first stage of fusion, a pair of parallel fusion layers are obtained. The same convolutional layers and average pooling layers are added after them to speed up the operation. After pooling, the resulting pair of feature values ​​are fused for the second time. Finally, the data with multi-scale features is fused and then entered into the Dropout layer before classification. The pooling operation adopts a strategy that combines max pooling and average pooling. In the early stage of the network structure, max pooling is used to preserve the significant features of the signal. In the later stage of the network structure, average pooling is added to consider all important feature elements of the signal. A multi-scale convolutional neural network is constructed to diagnose fault types and determine the diagnostic accuracy. S5: Verification: Compare the diagnostic accuracy of the input and output shaft bearings under different operating conditions to verify the stability and generalization ability of this method.

2. The rolling bearing fault diagnosis method according to claim 1, characterized in that: In step S1, the original vibration signal is measured by a vibration sensor. The original vibration signal includes five types: normal rolling bearing condition, outer ring crack, inner ring crack, roller crack, and mixed fault condition. The mixed fault condition includes at least two types of outer ring crack, inner ring crack, and roller crack.

3. The rolling bearing fault diagnosis method according to claim 1, characterized in that: In step S2, the dual-tree complex wavelet packet transform is an improvement on the dual-tree complex wavelet packet transform. It achieves signal decomposition and reconstruction through parallel transformation of the real and imaginary parts of discrete wavelets.

4. The rolling bearing fault diagnosis method according to claim 1, characterized in that: In step S2, the dual-tree complex wavelet packet transform decomposes the original vibration signal into 5 layers. The 'db14' wavelet is used to construct a first-level filter and a general filter. A Q-shift filter with a length of 14 is selected as the dual-tree filter.

5. The rolling bearing fault diagnosis method according to claim 1, characterized in that: In step S2, a soft threshold denoising method is used to process the denoised signal. The threshold selection method is to use an unbiased likelihood estimation threshold estimator, which is an adaptive threshold selection based on the principle of unbiased likelihood estimation.

6. The rolling bearing fault diagnosis method according to claim 1, characterized in that: The specific steps for constructing a multi-scale convolutional neural network for fault identification in step S4 are as follows: A1: Construct a multi-scale convolutional neural network; A2: Training is performed using samples, wherein each fault signal has 280 sets of labeled samples, and the samples are divided into training set, validation set and test set in a ratio of 7:1:2; A3: Visualize the t-SNE clustering distribution of activations in early max pooling layers, final convolutional layers, and final softmax layers of multi-scale convolutional neural networks; A4: Use a confusion matrix to summarize the prediction results of the classification model.

7. The rolling bearing fault diagnosis method according to claim 1, characterized in that: The specific steps of step S5 are as follows: B1: Change the rolling bearing operating conditions and use the same method to classify and diagnose the five types of output shaft fault bearing data; B2: To further verify the generalization ability of this method, the load and speed conditions were changed respectively, and the input shaft bearing was also added to the diagnostic conditions for verification. The same method was used to classify and diagnose the faulty bearing data under different conditions.

Images