A Slewing Bearing Fault Diagnosis Method Based on TimeGAN-SSGC-GAT

Abstract

The application provides a slewing bearing fault diagnosis method based on Time GAN-SSGC-GAT. The method performs data enhancement on original sample features through a Time GAN method, expands a training set, and solves the small sample problem of a training model. Through an SSGC method, a time sequence signal is converted into non-Euclidean structure diagram data. A GAT method with a multi-head attention mechanism is used to perform node classification on the graph structure data, different attention weights are allocated to different neighborhoods, the problem that a traditional deep learning method cannot fully utilize the spatial relationship between training samples is solved, and thus slewing bearing fault diagnosis under multiple working conditions and safety maintenance of a radar servo system are realized.

Description

Technical Field

[0001] This invention relates to a slewing bearing fault diagnosis method based on Time GAN-SSGC-GAT, belonging to the field of slewing bearing fault diagnosis. Background Technology

[0002] In major equipment fields such as lifting machinery, construction machinery, and military special vehicles, slewing bearings, as a new type of component, play a very important role in the mechanical field. Because slewing bearings operate at low speeds, have high load-bearing capacity, and undergo complex and variable working conditions, fault diagnosis research is essential to ensure the normal operation and smooth functioning of equipment, maintain the safety of machinery and personnel, and reduce unnecessary losses.

[0003] However, while existing fault diagnosis methods have shown excellent results in the field of slewing bearings, they have not addressed issues such as variable working conditions, small sample sizes, and multi-sensor fusion in experimental studies, nor have they explored the relationships and interdependencies between sample data features. Therefore, some studies have proposed displaying data in a new form of irregular graphs. Furthermore, these proposed methods differ from traditional methods, requiring the construction of complex graph data, which poses a significant challenge to existing machine learning algorithms. This also means that while important operations like convolution are widely used in Euclidean space, they are difficult to model graphical structural data in non-Euclidean spaces. Graph Neural Networks (GNNs) are a deep learning-based method for processing graph information, and they have attracted widespread attention due to their ability to analyze graphical structural data. Because GNNs can model the interdependencies between data and embed them into extracted features, this method has gradually become a research hotspot in the field of fault diagnosis. However, GNNs neglect the importance of input information and the interdependencies of data. The correlation of fault information varies across different neighborhoods. If the same weights are given, a certain amount of information will be lost, thus affecting the final fault diagnosis result. The GAT network overcomes the inherent shortcomings of GNNs by demonstrating its superiority through adaptive attention weights assigned to different connection nodes.

[0004] Furthermore, in the actual data signal acquisition process of slewing bearings, problems such as insufficient data samples and high acquisition costs are often encountered. Currently, using data-based methods to improve or augment data quality has become a common approach for processing small sample data. These methods include Variational Auto-Encoders (VAEs), Generative Adversarial Networks (GANs), transfer learning, Synthetic Minority Over Sampling Technique (SMOTE), and TrAdaboost models. However, the methods mentioned above do not fully consider the inherent correlation of time-series data and lack evaluation and improvement of data generation quality, resulting in low accuracy during fault data augmentation. Therefore, researching sample data augmentation methods in the context of small sample data, and fully considering the influence of time-series information on features during sample data augmentation, is of great significance for conducting slewing bearing fault diagnosis, prediction, and health management. Summary of the Invention

[0005] Objective: To address the shortcomings of existing technologies, this invention provides a slewing bearing fault diagnosis method based on Time GAN-SSGC-GAT. This method uses Time GAN to augment the features of the original samples, expanding the training set and solving the problem of small sample size in training models. The SSGC method converts time-series signals into non-Euclidean structural graph data. A GAT method with multi-head attention is employed to classify nodes in the graph structural data, assigning different attention weights to different neighborhoods. This solves the problem that traditional deep learning methods cannot fully utilize the spatial relationships between training samples, thereby enabling slewing bearing fault diagnosis under various operating conditions and ensuring the safe maintenance of radar servo systems.

[0006] Technical solution:

[0007] A slewing bearing fault diagnosis method based on Time GAN-SSGC-GAT specifically includes the following steps:

[0008] Step 1: Set up the experimental setup, which includes four accelerometers and one acoustic emission sensor. The four accelerometers are evenly arranged at 90° intervals along the inner ring of the slewing bearing, vertically above the surface. The acoustic emission sensor is located below the base of the fixed slewing bearing. The accelerometers include three unidirectional accelerometers and one tridirectional accelerometer.

[0009] Step 2: Collect condition monitoring data signals of slewing bearings under different states and working conditions, and divide the collected data signals into: healthy slewing bearing samples, slewing bearing samples with one rolling element failure, slewing bearing samples with inner ring failure, and slewing bearing samples with outer ring failure.

[0010] In step 2, the slewing bearing has four different states:

[0011] A healthy slewing bearing without damage, a slewing bearing with an inner ring failure with a cracked hole about 1 mm deep, a slewing bearing with an outer ring failure with an irregular scratch about 1 mm deep, and a slewing bearing with a rolling element failure due to roller wear.

[0012] The collected condition monitoring data signals of the slewing bearing under different states and working conditions include vibration signal Acc and acoustic emission signal AE;

[0013] Sample construction is performed based on the collected vibration signal Acc and acoustic emission signal AE. The sample construction process specifically includes the following steps:

[0014] Step 2.1: In the experiment, the sampling frequency was set to 50kHz, the sampling interval was 1 minute, and the duration of each sampling was 1 minute. Based on the determined sampling frequency, sampling time, and sampling interval, historical monitoring data of the slewing bearing under different operating conditions were obtained. The storage format of the collected key state data signal samples of the slewing bearing is shown below:

[0015] ;

[0016] In the formula, m represents the number of rows in each group of slewing bearing signal samples; each column represents the signal transmission channel of a sensor.

[0017] Step 2.2: Obtain all data sample signals collected under the above four states of slewing bearing, and divide the collected data signals into: healthy slewing bearing sample N, one rolling element faulty slewing bearing sample B1, faulty slewing bearing sample I, and outer ring faulty slewing bearing sample O.

[0018] The set of healthy slewing bearing samples is denoted as:

[0019] ;

[0020] In the formula, This represents the i-th group of healthy slewing bearing samples. This indicates the total number of healthy slewing bearing samples;

[0021] The set of slewing bearing samples with rolling element failures is denoted as:

[0022] ;

[0023] In the formula, This represents a slewing bearing sample with a rolling element failure in the i-th group. This represents the total number of slewing bearing samples with a rolling element failure.

[0024] The set of slewing bearing samples with inner ring failures is denoted as:

[0025] ;

[0026] In the formula, This represents the slewing bearing sample with inner ring failure in the i-th group. This indicates the total number of faulty slewing bearing samples;

[0027] The set of slewing bearing samples with outer ring failures is denoted as:

[0028] ;

[0029] In the formula, This represents the slewing bearing sample with outer race failure in the i-th group. This indicates the total number of slewing bearing samples with outer ring failures.

[0030] Step 3: Use the Person correlation coefficient to measure the degree of correlation between the data signals acquired by each sensor channel, and use this as the basis for sensor selection;

[0031] In step 3, the Person correlation coefficient is used to calculate the coefficient between each sensor channel to measure the degree of correlation between the data signals collected by each sensor channel. The sensor channel signals with a correlation coefficient of 0.7 or higher are used as the basis for selecting sensor acquisition signals.

[0032] Step 3.1: Calculate the Pearson correlation coefficient between each sensor channel based on the samples of each type of slewing bearing;

[0033] (1);

[0034] In the formula, X and Y are two random and different sensor signal channels; Let X be the covariance between channels X and Y. and Let X and Y be the standard deviations of their respective channels.

[0035] Step 3.2: Calculate the correlation coefficient between each sensor channel based on the signal characteristics collected by each sensor channel using the PEN correlation coefficient method, and draw its heat map.

[0036] Step 4: Extract the time domain, frequency domain, and time-frequency domain features of the dataset, label them accordingly, and establish the Real feature set of real sample data;

[0037] In step 4, the time-domain, frequency-domain, and time-frequency-domain features of the dataset are extracted, specifically including: absolute mean, peak value, root mean square value, root mean square amplitude, skewness value, kurtosis value, waveform factor, impulse factor, skewness factor, peak factor, margin factor, kurtosis factor, centroid frequency, mean square frequency, root mean square frequency, frequency variance, and eight wavelet packet feature components extracted after three-level wavelet packet decomposition, wherein the wavelet packet decomposition uses the db3 wavelet basis.

[0038] Step 4.1: Extract the slewing bearing signal features under different conditions and divide them into: healthy slewing bearing feature set, rolling element failure slewing bearing feature set, inner ring failure slewing bearing feature set, and outer ring failure slewing bearing feature set;

[0039] Step 4.2: Integrate the slewing bearing signal feature sets under four different conditions to form a real sample data feature set Real, and label the four features.

[0040] Step 5: Construct a Time GAN-based network to enhance and train the real sample data feature set and generate a virtual sample data feature set, Synthetic, which expands the sample features while retaining their unique temporal correlation.

[0041] In step 5, the constructed Time GAN model mainly consists of an autoencoder (AE) component and a generative adversarial network (GAN) component. The AE component includes an embedding network and a recovery network, and the GAN component includes a sequence generator and a sequence discriminator, wherein:

[0042] Step 5.1: Based on the data generation model principle of the Time GAN model, the real sample data feature set Real is divided into the static feature S and the time feature X of the slewing bearing;

[0043] Define vector for Static characteristics of a slewing bearing; vector for Time characteristics of a slewing bearing; and These are the vector spaces representing the static and temporal characteristics of the slewing bearing, respectively.

[0044] A set of randomly generated data within the time series length T Joint distribution with radar slewing bearing signal characteristics , Its time series length;

[0045] definition Let be the index of a single sample in the radar slewing bearing signal feature training data. Then the training data is represented as:

[0046] ;

[0047] The distribution of the characteristic time data of real radar slewing bearing signals is defined as follows:

[0048] ,

[0049] The distribution of the time data generated after model training is defined as follows:

[0050] ;

[0051] Step 5.2: In the AE component, the embedding network and the recovery network provide a mapping between the slewing bearing training features and the latent space, learn the latent temporal features of the training data through low-dimensional representation, and reduce the dimensionality of the slewing bearing data signal features;

[0052] definition and To the characteristic space of the slewing bearing signal and The corresponding latent feature vector space is then represented by the embedding function as:

[0053] ;

[0054] Transform its static and temporal features into latent spatial codes ,Right now ;

[0055] Conversely, the recovery function is expressed as:

[0056] ;

[0057] Reconstruct the static and time characteristics of the slewing bearing ,Right now ;

[0058] Therefore, the embedding network and recovery network functions can be defined as follows:

[0059] ;

[0060] In the formula, and These are embedded networks representing the static and temporal characteristics of the slewing bearing, respectively. For the static characteristics of slewing bearings and This represents the potential feature vector space for the static and temporal characteristics of a slewing bearing. and A recovery network for static and time-embedded slewing bearings; and It is the input data after returning static and temporary code to their characteristic representation;

[0061] In GAN components, the sequence generator network first outputs into the embedding space, defining... and Define a vector space with known distributions for the static and temporal characteristics of the slewing bearing, and plot random vectors from it as generators. and Given the input, the generator function is expressed as:

[0062] ;

[0063] The latent code is generated using static and time random vector sets of the slewing bearing. , is represented as:

[0064] ;

[0065] Sequence discriminator networks also operate from the embedding space, which in turn operates in the latent space, generating functions.

[0066] ;

[0067] It receives static and time-encoded signals from the slewing bearing and outputs the classification results. , is represented as:

[0068] ;

[0069] The sequence generator and sequence discriminator functions can be defined as follows:

[0070] ;

[0071] In the formula, and Sequence generator networks for static and temporal characteristics of slewing bearings, respectively; and The data format for combining static and time random vectors of the slewing bearing into a latent code after passing through a generator network; and A sequence discriminator network for the static and temporal characteristics of slewing bearings; and The results are used to determine the static and time-dependent characteristics of the slewing bearing. and These are sequences of forward and backward hidden states, respectively; and For output layer functions;

[0072] Step 5.3: During the overall Time GAN model operation, firstly, the loss is reconstructed using AE. As a reversible mapping of slewing bearing signal characteristics and latent space, the embedded network and recovery network functions extract data from the original data. Learn latent representations from data to accurately reconstruct the data. ;

[0073] Then, unsupervised loss is used. Define a supervised loss between data based on real sequences and data from the sequence generator network;

[0074] Finally, through unsupervised GAN... It reflects the feedback achieved in the adversarial process as the data generated by the sequence generator continuously approximates the real sequence data;

[0075] The definitions of the three loss functions are as follows:

[0076]

[0077] in, It is a random static feature; These are the original static features; This represents the current random temporal feature sample; These are the original temporal feature samples; and These are the judgment results of the original static and time characteristic data of the slewing bearing obtained from the model, respectively. and The results are the static and time-based characteristic data of the slewing bearing, respectively. and These are the current random temporal characteristics and static characteristics of the slewing bearing, respectively. This represents the current random temporal characteristic vector of the slewing bearing.

[0078] Step 5.4: Obtain the Synthetic feature set of virtual sample data for slewing bearing generation.

[0079] Step 6: Construct graph structure data for the virtual sample data feature set using the SSGC graph construction method;

[0080] In step 6, the SSGC graph construction method is used to construct graph structure data for the feature set of virtual sample data; the collaborative similarity graph construction method integrates k-nearest neighbor, threshold-based cosine similarity, threshold-based Mahalanobis distance and minimum spanning tree to generate an adjacency matrix representing sample connections;

[0081] Step 6.1: Using k-nearest neighbors, threshold-based cosine similarity, threshold-based Mahalanobis distance, and MST techniques, respective adjacency matrices were constructed, as follows: , , , ;

[0082] Step 6.2: Calculate the similarity between adjacency matrices using the Frobenius norm. ;

[0083] Step 6.3: Calculate the weights between matrices based on their similarity. ;

[0084] Step 6.4: Connect the sample features to each other by calculating similarity in step 5.2 and assigning weights in step 5.3;

[0085] Use median thresholding Methods for merging adjacency matrices , , and The sample features are interconnected to form a graph, and a fusion adjacency matrix is ​​used. To represent the constructed graph, Defined as:

[0086] ;

[0087] Step 7: Divide the constructed graph structure data into training set and test set, and input the training set into the GAT fault diagnosis model constructed using multi-head attention mechanism for training;

[0088] In step 7, the ratio of the model training set to the test set is 8:2. The GAT fault diagnosis model mainly contains two GAT layers and two FC layers, and the model is updated through the ReLU activation function.

[0089] Step 7.1: Convert the input features into high-level features;

[0090] Step 7.2: Apply a shared linear transformation parameterized by the weight matrix W to all nodes;

[0091] Step 7.3: Execute the shared attention mechanism A on the node. After passing through the Leaky ReLU layer, obtain the attention score. ;

[0092]

[0093] In the formula: Indicates the connected operation; note the fraction. Represents a node For nodes The importance of the features; and Updated during iterations of the GAT layer. Activation functions for GAT and FC layers;

[0094] Step 7.4: All attention scores are normalized using the Softmax function. This makes it easy to compare between different nodes;

[0095] ;

[0096] In the formula, For nodes With nodes The normalized attention coefficients between them.

[0097] Step 8: Input the test set into the trained GAT model to achieve accurate classification of slewing bearings under multiple working conditions. Use T-SNE and confusion matrix methods to visualize the classification results of slewing bearings under multiple working conditions.

[0098] Beneficial effects:

[0099] (1) By using Time GAN to augment the features of real sample data, the training set was expanded, thus solving the problem of small sample size in training models;

[0100] (2) The time series signal is converted into non-Euclidean structure graph data by using the SSGC method; the GAT method is used to classify the nodes of the graph structure data and assign different attention weights to different neighborhoods. This solves the problem that traditional deep learning methods cannot make full use of the spatial relationship between training samples, thereby realizing the fault diagnosis of slewing bearings under multiple working conditions and improving its fault classification accuracy. Attached Figure Description

[0101] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0102] Figure 1 This is a flowchart illustrating the overall method of an embodiment of the present invention;

[0103] Figure 2 This is the original monitoring data of the slewing bearing under four different state types according to an embodiment of the present invention;

[0104] Figure 3 This is a heatmap of sensor correlation analysis used for the acquired signal data in this embodiment of the invention.

[0105] Figure 4 This is a comparison diagram of the feature sets of real and virtual sample data in an embodiment of the present invention;

[0106] Figure 5 This is a visualization of PCA and T-SNE for the feature sets of real and virtual sample data in embodiments of the present invention;

[0107] Figure 6 This is a visualization of the fault diagnosis model before and after training according to an embodiment of the present invention;

[0108] Figure 7 This is a confusion matrix diagram of the prediction results of the fault diagnosis model in an embodiment of the present invention;

[0109] Figure 8 This is a comparison chart of the fault diagnosis classification accuracy of the proposed method, CNN, LSTM, ANN, RNN, and SVM in this embodiment of the invention.

[0110] Figure 9 This is a comparison chart of the accuracy of slewing bearing fault diagnosis classification using six methods with rotational speed as the variable, according to an embodiment of the present invention.

[0111] Figure 10 This is a comparison chart of the accuracy of slewing bearing fault diagnosis classification using six methods with load as the variable, according to an embodiment of the present invention. Detailed Implementation

[0112] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0113] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and 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 this invention.

[0114] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can include direct contact between the first and second features, or contact between the first and second features through another feature between them. Furthermore, "above," "over," and "on top" of the second feature includes the first feature directly above or diagonally above the second feature, or simply indicates that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature includes the first feature directly below or diagonally below the second feature, or simply indicates that the first feature is at a lower horizontal level than the second feature.

[0115] like Figure 1 As shown, step 1 involves assembling the experimental setup, which includes four accelerometers and one acoustic emission sensor. The four accelerometers are evenly arranged at 90° intervals along the inner ring of the slewing bearing, vertically above the surface. The acoustic emission sensor is located below the base of the fixed slewing bearing. The accelerometers include three unidirectional accelerometers and one tridirectional accelerometer.

[0116] Step 2: First, using rotational speed and overturning force as variables, set different working conditions and collect vibration and acoustic emission signals of the slewing bearing under different states and working conditions. The collected data signals are divided into: healthy slewing bearing samples, slewing bearing samples with one rolling element failure, slewing bearing samples with inner ring failure, and slewing bearing samples with outer ring failure.

[0117] The slewing bearing has four different states:

[0118] A healthy slewing bearing without damage, a slewing bearing with an inner ring failure with a cracked hole about 1 mm deep, a slewing bearing with an outer ring failure with an irregular scratch about 1 mm deep, and a slewing bearing with a rolling element failure due to roller wear.

[0119] The acquired data signals include vibration signals (ACC) and acoustic emission signals (AE). Samples are constructed based on the acquired vibration signals (ACC) and acoustic emission signals (AE). The sample construction process includes the following steps:

[0120] Step 2.1: In the experiment, the sampling frequency was set to 50kHz, the sampling interval was 1 minute, and the duration of each sampling was 1 minute. Based on the determined sampling frequency, sampling time, and sampling interval, historical monitoring data of the slewing bearing under different operating conditions were obtained. The storage format of each set of key state data signal samples of the slewing bearing is shown below:

[0121] (1)

[0122] In the formula, m represents the number of rows in each group of slewing bearing signal samples; each column represents the signal transmission channel of a sensor.

[0123] Step 2.2: Obtain all data sample signals collected under each state of the slewing bearing, such as... Figure 2 As shown, the collected data signals are divided into: healthy slewing bearing sample (N), one rolling element faulty slewing bearing sample (B1), faulty slewing bearing sample (I), and outer ring faulty slewing bearing sample (O).

[0124] The set of healthy slewing bearing samples is denoted as...

[0125] , This represents the i-th group of healthy slewing bearing samples. This indicates the total number of healthy slewing bearing samples;

[0126] The set of rolling element fault slewing bearing samples is denoted as

[0127] , This represents a slewing bearing sample with a rolling element failure in the i-th group. This represents the total number of slewing bearing samples with a rolling element failure.

[0128] The set of slewing bearing samples with inner ring failures is denoted as...

[0129] , This represents the slewing bearing sample with inner ring failure in the i-th group. This indicates the total number of faulty slewing bearing samples;

[0130] The set of slewing bearing samples with outer ring failures is denoted as...

[0131] , This represents the slewing bearing sample with outer race failure in the i-th group. This indicates the total number of slewing bearing samples with outer ring failures.

[0132] Step 3: Use the Person correlation coefficient to calculate the coefficient between each sensor channel to measure the correlation between the data signals acquired by each sensor channel. Sensor channel signals with a correlation coefficient of 0.7 or higher are used as the basis for selecting sensor acquisition signals.

[0133] Step 3.1: Calculate the Pearson correlation coefficient between each sensor channel based on the samples of each type of slewing bearing;

[0134] (2);

[0135] In the formula, X and Y are two random and different sensor signal channels; Let X be the covariance between channels X and Y. and Let X and Y be the standard deviations of their respective channels.

[0136] Step 3.2: Based on the signal characteristics acquired by each sensor channel, calculate the correlation coefficient between each sensor channel using the PEN correlation coefficient method, and draw its heatmap for easy observation. Specifically, as shown below... Figure 3 As shown in the figure, the horizontal and vertical axes represent the signal characteristic variables of each sensor channel, and the color intensity of the heat map on the right represents the correlation level of the features. Darker colors represent close to 1 (positive correlation), and lighter colors represent close to -1 (negative correlation); the darker the color, the stronger the correlation. The figure shows that the correlation level of the data signal characteristics between each sensor is very high, indicating a strong correlation between the attached sensors. Furthermore, the diagonal elements are always equal to 1 because the sensors are perfectly correlated with themselves.

[0137] Step 4: Extract the time-domain, frequency-domain, and time-frequency-domain features of the dataset, label them accordingly, and form a feature set of real sample data. The extracted features specifically include: absolute mean, peak value, root mean square value, root mean square amplitude, skewness value, kurtosis value, waveform factor, impulse factor, skewness factor, peak factor, margin factor, kurtosis factor, centroid frequency, mean square frequency, root mean square frequency, frequency variance, and eight wavelet feature components extracted after three-level wavelet decomposition. The specific feature extraction formulas are shown in Table 1.

[0138] Table 1: Characteristic Formulas

[0139]

[0140] Step 4.1: Based on the above feature formula, extract the slewing bearing signal features under different conditions and divide them into: healthy slewing bearing feature set, rolling element failure slewing bearing feature set, inner ring failure slewing bearing feature set, and outer ring failure slewing bearing feature set.

[0141] Step 4.2: Integrate the slewing bearing signal feature sets under four different conditions to form a real sample data feature set Real, and number the labels of the four features.

[0142] Step 5: Construct a Time GAN-based network to enhance and train the real sample data feature set and generate a virtual sample data feature set. This expands the sample features while preserving their unique temporal correlation. The Time GAN model parameter settings are shown in Table 2.

[0143] Table 2: Time GAN Model Parameter Table

[0144]

[0145] The Time GAN model mainly consists of an autoencoder (AE) component and a generative adversarial network (GAN) component. The former includes an embedding network and a recovery network, while the latter includes a sequence generator and a sequence discriminator.

[0146] Step 5.1: Based on the data generation model principle of the Time GAN model, the real sample data feature set Real is divided into the static feature S and the time feature X of the slewing bearing.

[0147] Define vector for Static characteristics of a slewing bearing; vector for The time characteristics of a slewing bearing. and These are vector spaces representing the static and temporal characteristics of the slewing bearing, respectively. A set of... Joint distribution with certain radar slewing bearing signal characteristics , Its time series length. Definition Let be the index of a single sample in the radar slewing bearing signal feature training data, then the training data is represented as: The distribution of the characteristic time data of real radar slewing bearing signals is defined as follows: The distribution of the time data generated after model training is defined as follows: .

[0148] The learning objective of the Time GAN model is to use training data Make Closer .

[0149] Step 5.2: In the AE component, the embedding and recovery network provides a mapping between the slewing bearing training features and the latent space, learns the latent temporal features of the training data through low-dimensional representation, and reduces the dimensionality of the slewing bearing data signal features.

[0150] definition and To the characteristic space of the slewing bearing signal and The corresponding latent feature vector space. Then the embedding function... Transform its static and temporal features into latent spatial codes ,Right now Conversely, the recovery function Reconstruct the static and time characteristics of the slewing bearing ,Right now Therefore, the embedding network and recovery network functions can be defined as follows:

[0151] (3)

[0152] In the formula, and These are embedded networks representing the static and temporal characteristics of the slewing bearing, respectively. and This represents the potential feature vector space for the static and temporal characteristics of a slewing bearing. and A recovery network for static and time-embedded slewing bearings; and It is the input data after returning static and temporary code to their characteristic representation.

[0153] In GAN components, the sequence generator network does not directly generate the synthesized output in the slewing bearing signal feature space, but instead first outputs it into the embedding space. (Definition) and Define a vector space with known distributions for the static and temporal characteristics of the slewing bearing, and plot random vectors from it as generators. and The input. Then the generator function. The latent code is generated using static and time random vector sets of the slewing bearing. ,Right now Sequence discriminator networks also operate from the embedding space, but they operate in the latent space, generating functions. It receives static and time-encoded signals from the slewing bearing and outputs the classification results. ,Right now Therefore, the sequence generator and sequence discriminator functions can be defined as follows:

[0154] (4)

[0155] In the formula, and Sequence generator networks for static and temporal characteristics of slewing bearings, respectively; and The data format for combining static and time random vectors of the slewing bearing into a latent code after passing through a generator network; and A sequence discriminator network for the static and temporal characteristics of slewing bearings; and The results are used to determine the static and time-dependent characteristics of the slewing bearing. and These are sequences of forward and backward hidden states, respectively; and This is the output layer function.

[0156] Step 5.3: During the overall Time GAN model operation, firstly, use AE to reconstruct the loss. As a reversible mapping of slewing bearing signal characteristics and latent space, the embedded network and recovery network functions extract data from the original data. Learn latent representations from data to accurately reconstruct the data. Then, unsupervised loss is used. We define a supervised loss between data based on real sequences and data from the sequence generator network. Finally, we apply an unsupervised loss method using GANs. This reflects the feedback achieved in the adversarial process as the data generated by the sequence generator continuously approximates the real sequence data. The definitions of the above three loss functions are as follows:

[0157] (5)

[0158] in, It is a random static feature; These are the original static features; This represents the current random temporal feature sample; These are the original temporal feature samples; and These are the judgment results of the original static and time characteristic data of the slewing bearing obtained from the model, respectively. and The results are the static and time-based characteristic data of the slewing bearing, respectively. and These are the current random temporal characteristics and static characteristics of the slewing bearing, respectively. This represents the current random temporal characteristic vector of the slewing bearing.

[0159] Step 5.4: Obtain the virtual sample data feature set Synthetic for the slewing bearing and compare it with the real sample data feature set Real. See details below. Figure 4 As shown; and the generated data is evaluated visually, such as... Figure 5 As shown;

[0160] Figure 4In this study, after 50,000 iterations of the training model, the original signal feature data of the slewing bearing were visualized over time steps. In each graph, the horizontal axis represents the specific time step, and the vertical axis represents the corresponding normalization value. The blue solid line represents the real sample features, and the orange dashed line represents the newly generated virtual sample features. The first six rows show the trends of 24 features, and the last row shows the trends of four labels. From a short-term perspective, while the newly generated sample features do not strictly match the real sample features, they share certain trends, preserving the temporal correlation and transient distribution characteristics of the real sample features. Numerically speaking, based on the initial normalization, the difference between the generated data and the real data effectively expands the training dataset, enriching the data and enhancing its diversity.

[0161] Considering that information with short-term or instantaneous trend changes is difficult to analyze the effectiveness of the generated time-series sample data features from a multidimensional or high-dimensional overall distribution perspective, PCA and T-SNE are used to perform dimensionality reduction analysis on the data. The visualization results are as follows: Figure 5 As shown in the figure, black represents the original data, and red represents the synthesized data. It can be observed that the red samples (generated time-series features) and the black samples (original time-series features) are almost completely synchronized. Overall, the feature sequences generated by Time GAN conform to the trends of the real sample feature sequences, and the distribution of the time-series feature sequences is largely the same as that of the real sample feature sequences. Therefore, it can be used to train models when real time-series data is lacking.

[0162] Step 6: Construct graph structure data for the virtual sample data feature set using the SSGC graph construction method; the collaborative similarity graph construction method utilizes the advantages of k-nearest neighbor (k-NN), threshold-based cosine similarity (cos), threshold-based Mahalanobis distance and minimum spanning tree (MST) techniques, integrates these adjacency matrices, and generates an adjacency matrix representing sample connections;

[0163] Step 6.1: Using k-nearest neighbors, threshold-based cosine similarity, threshold-based Mahalanobis distance, and MST techniques, respective adjacency matrices were constructed, as follows: , , , .

[0164] Step 6.2: Calculate the similarity between adjacency matrices using the Frobenius norm. .

[0165] Step 6.3: Calculate the weights between matrices based on their similarity. .

[0166] Step 6.4: By calculating similarity in Step 6.2 and assigning weights in Step 6.3, the sample features can be concatenated together. Then, median thresholding is used. The method involves fusing these adjacency matrices, connecting sample features to form a graph, and then using the fused adjacency matrix. To represent the constructed graph. Then... It can be defined as:

[0167] (6)

[0168] Step 7: Divide the graph structure data constructed in Step 5 into a training set and a test set, and input the training set into the GAT fault diagnosis model constructed using the multi-head attention mechanism for training, so as to solve the problem that traditional deep learning methods cannot make full use of the spatial relationship between training samples.

[0169] During the training process of the fault diagnosis model, 80% of the total samples are randomly selected as the training set and 20% as the test set.

[0170] Step 7.1: Convert the input features into higher-level features that have better expressive power;

[0171] Step 7.2: Apply a shared linear transformation parameterized by the weight matrix W to all nodes;

[0172] Step 7.3: Execute shared attention mechanism A on the node, after... After the layer, attention scores are obtained. ;

[0173] (7)

[0174] In the formula: Indicates the connected operation; note the fraction. Represents a node For nodes The importance of the features; and Updated during iterations of the GAT layer;

[0175] Step 7.4: All attention scores are normalized using the Softmax function. This makes it easy to compare between different nodes;

[0176] (8)

[0177] In the formula, For nodes With nodes The normalized attention coefficients between them.

[0178] The GAT fault diagnosis model mainly consists of two GAT layers and two FC layers, and the model is updated using the ReLU activation function. The specific GAT model initialization parameter settings are shown in Table 3.

[0179] Table 3: GAT Model Parameters

[0180]

[0181] Step 8: Input the test set into the trained GAT model to achieve accurate classification of slewing bearings under multiple working conditions. Then, use T-SNE and confusion matrix methods to visualize the classification results of slewing bearings under multiple working conditions. See [link to T-SNE visualization results]. Figure 6 As shown, the visualization results of the confusion matrix are available in [link to visualization]. Figure 7 As shown;

[0182] In the T-SNE diagram, different colors represent different slewing bearing modes, and their corresponding labels are uniquely encoded. Specifically, "N" represents a healthy slewing bearing, "B1" represents a rolling element failure slewing bearing, "I" represents an inner ring failure slewing bearing, and "O" represents an outer ring failure slewing bearing. The T-SNE visualization shows that the features after training have significantly improved discriminative power compared to before training, with only a few scattered feature points overlapping.

[0183] To visually evaluate the performance of the proposed method, fault diagnosis results under various operating conditions of slewing bearings were studied. In the confusion matrix, the vertical axis and the horizontal axis represent the actual health condition and predicted health condition of the test samples, respectively. The figure shows that the proposed method exhibits excellent fault diagnosis performance for slewing bearings, indicating a low risk of false alarms. Furthermore, the proposed method maintains high fault diagnosis accuracy under multiple operating conditions, which is sufficient to meet the practical needs of the industry.

[0184] To verify the effectiveness of the proposed method, five different models were compared: CNN (Convolutional Neural Network), LSTM (Long Short-Term Memory Network), ANN (Artificial Neural Network), RNN (Recurrent Neural Network), and SVM (Support Vector Machine). The specific parameters of the comparison model method were as follows: (1) CNN has 3 convolutional layers, 3 pooling layers, 1 exit layer, and 1 output layer, of which the first two pooling layers are max pooling layers and the third is a global average pooling layer; (2) LSTM has 2 long short-term memory network layers, 2 exit layers, and 1 output layer; (3) ANN has 1 input layer, 3 hidden layers, 1 exit layer, and 1 output layer; (4) RNN has 2 recurrent network layers and 1 output layer; (5) The kernel function selected in SVM is RBF (Radial Basis Function of Gaussian), the penalty factor is 16, the coefficient of the kernel function is set to the default, and multiple classifiers need to be built to meet the requirements.

[0185] The final fault diagnosis classification results obtained from different models are shown in Table 4, and the classification results are illustrated in the bar chart below. Figure 8 As shown in the figure. Experimental results show that the proposed method outperforms the other four model methods in predicting slewing bearing failure modes, and the classification accuracy of the task prediction reaches over 90%, demonstrating good generalization performance.

[0186] Table 4: Fault diagnosis experimental results based on slewing bearing dataset under multiple working conditions

[0187]

[0188] Furthermore, to fully verify the effectiveness of different fault diagnosis methods, we considered different rotational speeds as variables and divided all the set operating conditions into three groups. Each of these three groups had the same rotational speed but different overturning force variables. To simplify these three groups of symbols, we abbreviated them as 2 rpm, 6 rpm, and 12 rpm. The fault diagnosis results of different classification methods are shown in Table 5. Figure 9 As shown.

[0189] Table 5: Experimental Results of Fault Diagnosis Based on Slewing Bearing Dataset under Rotational Speed ​​Variation

[0190]

[0191] like Figure 9As shown in the figure, the proposed method achieves excellent fault diagnosis performance on all datasets. The figure reveals that, compared to the fault diagnosis results at 2 rpm, the fault diagnosis accuracy of all methods except the proposed method is significantly improved at 6 rpm and 12 rpm. Further analysis shows that the traditional Support Vector Machine (SVM) fault classification method has the worst fault diagnosis performance for radar slewing bearings under all operating conditions. This is because SVM was initially used to solve binary classification problems; multi-class problems limit its fault diagnosis performance during the testing phase, leading to poor classification results. In contrast, learning methods based on CNN (Convolutional Neural Network), LSTM (Long Short-Term Memory Network), ANN (Artificial Neural Network), and RNN (Recurrent Neural Network) achieve relatively high fault diagnosis accuracy compared to the SVM method. However, due to the inherently weak feature extraction capability inherited from local convolution operations, their fault diagnosis performance is also not ideal.

[0192] Simultaneously, considering different rotational speeds as variables, all set operating conditions were divided into three groups. Each of these three groups had the same load but different rotational speeds. To simplify these three groups of symbols, we abbreviated them as overturning force 0%, overturning force 50%, and overturning force 100%. The fault diagnosis results obtained from these different classification methods are shown in Table 6. Figure 10 As shown.

[0193] Table 6: Fault diagnosis experimental results based on slewing bearing dataset under load variation

[0194]

[0195] like Figure 10 As shown, the proposed method achieves excellent fault diagnosis performance on all datasets. Overall, the proposed method achieves high and stable fault classification accuracy in rolling bearing fault diagnosis, demonstrating its superiority.

[0196] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0197] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A slewing bearing fault diagnosis method based on Time GAN-SSGC-GAT, characterized in that: Specifically, the steps include the following: Step 1: Set up the experimental equipment, which includes 4 accelerometers and 1 acoustic emission sensor. The 4 accelerometers are evenly arranged at 90° intervals along the inner ring of the slewing bearing in the vertical direction above the surface. The acoustic emission sensor is located below the base of the fixed slewing bearing. The accelerometers include three unidirectional accelerometers and one tridirectional accelerometer. Step 2: Collect condition monitoring data signals of slewing bearings under different states and working conditions, and divide the collected data signals into: healthy slewing bearing samples, slewing bearing samples with one rolling element failure, slewing bearing samples with inner ring failure, and slewing bearing samples with outer ring failure. Step 3: Use the Person correlation coefficient to measure the degree of correlation between the data signals acquired by each sensor channel, and use this as the basis for sensor selection; Step 4: Extract the time domain, frequency domain, and time-frequency domain features of the dataset, label them accordingly, and establish the Real feature set of real sample data; Step 5: Construct a Time GAN-based network to enhance and train the real sample data feature set and generate a virtual sample data feature set, Synthetic, which expands the sample features while retaining their unique temporal correlation. Step 6: Construct graph structure data for the virtual sample data feature set using the SSGC graph construction method; In step 6, the SSGC graph construction method is used to construct graph structure data for the feature set of virtual sample data; the collaborative similarity graph construction method integrates k-nearest neighbor, threshold-based cosine similarity, threshold-based Mahalanobis distance and minimum spanning tree to generate an adjacency matrix representing sample connections; Step 6.1: Using k-nearest neighbors, threshold-based cosine similarity, threshold-based Mahalanobis distance, and MST techniques, respective adjacency matrices were constructed, as follows: , , , ; Step 6.2: Calculate the similarity between adjacency matrices using the Frobenius norm. ; Step 6.3: Calculate the weights between matrices based on their similarity. ; Step 6.4: Connect the sample features to each other by calculating similarity in step 5.2 and assigning weights in step 5.3; Use median thresholding Methods for merging adjacency matrices , , and The sample features are interconnected to form a graph, and a fusion adjacency matrix is ​​used. To represent the constructed graph, Defined as: ； Step 7: Divide the constructed graph structure data into training set and test set, and input the training set into the GAT fault diagnosis model constructed using multi-head attention mechanism for training; In step 7, the ratio of the model training set to the test set is 8:

2. The GAT fault diagnosis model contains two GAT layers and two FC layers, and the model is updated through the ReLU activation function. Step 7.1: Convert the input features into high-level features; Step 7.2: Apply a shared linear transformation parameterized by the weight matrix W to all nodes; Step 7.3: Execute the shared attention mechanism A on the node. After passing through the Leaky ReLU layer, obtain the attention score. ; ； In the formula: Indicates the connected operations; note the fractions. Represents a node For nodes The importance of the features; and Updated during iterations of the GAT layer. Activation functions for GAT and FC layers; Step 7.4: All attention scores are normalized using the Softmax function. This makes it easy to compare between different nodes; ； In the formula, For nodes With nodes Normalized attention coefficients between them; Step 8: Input the test set into the trained GAT model to achieve accurate classification of slewing bearings under multiple working conditions.

2. The slewing bearing fault diagnosis method based on Time GAN-SSGC-GAT as described in claim 1, characterized in that: In step 2, the slewing bearing has four different states: A healthy slewing bearing without damage, a slewing bearing with an inner ring failure with a cracked hole about 1 mm deep, a slewing bearing with an outer ring failure with an irregular scratch about 1 mm deep, and a slewing bearing with a rolling element failure due to roller wear. The collected condition monitoring data signals of the slewing bearing under different states and working conditions include vibration signal Acc and acoustic emission signal AE; Sample construction is performed based on the collected vibration signal Acc and acoustic emission signal AE. The sample construction process specifically includes the following steps: Step 2.1: In the experiment, the sampling frequency was set to 50kHz, the sampling interval was 1 minute, and the duration of each sampling was 1 minute. Based on the determined sampling frequency, sampling time, and sampling interval, historical monitoring data of the slewing bearing under different operating conditions were obtained. The storage format of the collected key state data signal samples of the slewing bearing is shown below: ； In the formula, m represents the number of rows in each group of slewing bearing signal samples; each column represents the signal transmission channel of a sensor. Step 2.2: Obtain all data sample signals collected under the above four states of slewing bearing, and divide the collected data signals into: healthy slewing bearing sample N, one rolling element faulty slewing bearing sample B1, faulty slewing bearing sample I, and outer ring faulty slewing bearing sample O; The set of healthy slewing bearing samples is denoted as: ； In the formula, This represents the i-th group of healthy slewing bearing samples. This indicates the total number of healthy slewing bearing samples; The set of slewing bearing samples with rolling element failures is denoted as: ； In the formula, This represents a slewing bearing sample with a rolling element failure in the i-th group. This represents the total number of slewing bearing samples with a rolling element failure. The set of slewing bearing samples with inner ring failures is denoted as: ； In the formula, This represents the slewing bearing sample with inner ring failure in the i-th group. This indicates the total number of faulty slewing bearing samples; The set of slewing bearing samples with outer ring failures is denoted as: ； In the formula, This represents the slewing bearing sample with outer race failure in the i-th group. This indicates the total number of slewing bearing samples with outer ring failures.

3. The slewing bearing fault diagnosis method based on Time GAN-SSGC-GAT as described in claim 1, characterized in that: In step 3, the Person correlation coefficient is used to calculate the coefficient between each sensor channel to measure the degree of correlation between the data signals collected by each sensor channel. The sensor channel signals with a correlation coefficient of 0.7 or higher are used as the basis for selecting sensor acquisition signals. Step 3.1: Calculate the Pearson correlation coefficient between each sensor channel based on the samples of each type of slewing bearing; （1）； In the formula, X and Y are two random and different sensor signal channels; Let X be the covariance between channels X and Y. and Let X and Y be the standard deviations of their respective channels. Step 3.2: Calculate the correlation coefficient between each sensor channel based on the signal characteristics collected by each sensor channel using the PEN correlation coefficient method, and draw its heat map.

4. The slewing bearing fault diagnosis method based on Time GAN-SSGC-GAT as described in claim 1, characterized in that: In step 4, the time domain, frequency domain, and time-frequency domain features of the dataset are extracted, specifically including: absolute mean, peak value, root mean square value, root mean square amplitude, skewness value, kurtosis value, waveform factor, impulse factor, skewness factor, peak factor, margin factor, kurtosis factor, centroid frequency, mean square frequency, root mean square frequency, frequency variance, and 8 wavelet packet feature components extracted after 3-level wavelet packet decomposition, wherein the wavelet packet decomposition uses the db3 wavelet basis; Step 4.1: Extract the slewing bearing signal features under different conditions and divide them into: healthy slewing bearing feature set, rolling element failure slewing bearing feature set, inner ring failure slewing bearing feature set, and outer ring failure slewing bearing feature set; Step 4.2: Integrate the slewing bearing signal feature sets under four different conditions to form a real sample data feature set Real, and label the four features.

5. The slewing bearing fault diagnosis method based on Time GAN-SSGC-GAT as described in claim 1, characterized in that: In step 5, the constructed Time GAN model consists of an autoencoder (AE) component and a generative adversarial network (GAN) component. The AE component includes an embedding network and a recovery network, and the GAN component includes a sequence generator and a sequence discriminator, wherein: Step 5.1: Based on the data generation model principle of the Time GAN model, the real sample data feature set Real is divided into the static feature S and the time feature X of the slewing bearing; Define vector for Static characteristics of a slewing bearing; vector for Time characteristics of a slewing bearing; and These are the vector spaces representing the static and temporal characteristics of the slewing bearing, respectively. A set of randomly generated data within the time series length T Joint distribution with radar slewing bearing signal characteristics , Its time series length; definition Let be the index of a single sample in the radar slewing bearing signal feature training data. Then the training data is represented as: ； The distribution of the characteristic time data of real radar slewing bearing signals is defined as follows: ； The distribution of the time data generated after model training is defined as follows: ； Step 5.2: In the AE component, the embedding network and the recovery network provide a mapping between the slewing bearing training features and the latent space, learn the latent temporal features of the training data through low-dimensional representation, and reduce the dimensionality of the slewing bearing data signal features; definition and To the characteristic space of the slewing bearing signal and The corresponding latent feature vector space is then represented by the embedding function as: ； Transform its static and temporal features into latent spatial codes ,Right now ; Conversely, the recovery function is expressed as: ； Reconstruct the static and time characteristics of the slewing bearing ,Right now ; Therefore, the embedding network and recovery network functions can be defined as follows: ； In the formula, and These are embedded networks representing the static and temporal characteristics of the slewing bearing, respectively. For the static characteristics of slewing bearings and This represents the potential feature vector space for the static and temporal characteristics of a slewing bearing. and A recovery network for static and time-embedded slewing bearings; and It is the input data after returning static and temporary code to their characteristic representation; In GAN components, the sequence generator network first outputs into the embedding space, defining... and Define a vector space with known distributions for the static and temporal characteristics of the slewing bearing, and plot random vectors from it as generators. and Given the input, the generator function is expressed as: ； The latent code is generated using static and time random vector sets of the slewing bearing. , is represented as: ； Sequence discriminator networks also operate from the embedding space, which in turn operates in the latent space, generating functions. ； It receives static and time-encoded signals from the slewing bearing and outputs the classification results. , is represented as: ； The sequence generator and sequence discriminator functions can be defined as follows: ； In the formula, and Sequence generator networks for static and temporal characteristics of slewing bearings, respectively; and The data format for combining static and time random vectors of the slewing bearing into a latent code after passing through a generator network; and A sequence discriminator network for the static and temporal characteristics of slewing bearings; and The results of the discrimination of static and time-based generation characteristics of slewing bearings; and These are sequences of forward and backward hidden states, respectively; and For output layer functions; Step 5.3: During the overall Time GAN model operation, firstly, the loss is reconstructed using AE. As a reversible mapping of slewing bearing signal characteristics and latent space, the embedded network and recovery network functions extract data from the original data. Learn latent representations from data to accurately reconstruct the data. ; Then, unsupervised loss is used. Define a supervised loss between data based on real sequences and data from the sequence generator network; Finally, through unsupervised GAN... It reflects the feedback achieved in the adversarial process as the data generated by the sequence generator continuously approximates the real sequence data; The definitions of the three loss functions are as follows: ； in, It is a random static feature; These are the original static features; This represents the current random temporal feature sample; These are the original temporal feature samples; and These are the judgment results of the original static and time characteristic data of the slewing bearing obtained from the model, respectively. and The results are the static and time-based characteristic data of the slewing bearing, respectively. and These are the current random temporal characteristics and static characteristics of the slewing bearing, respectively. This represents the current random temporal characteristic vector of the slewing bearing; Step 5.4: Obtain the Synthetic feature set of virtual sample data for slewing bearing generation.

6. The slewing bearing fault diagnosis method based on Time GAN-SSGC-GAT as described in claim 1, characterized in that: In step 8, the T-SNE and confusion matrix methods are used to visualize the classification results of slewing bearings under multiple working conditions.

Images