A Fault Diagnosis Method for Fan Bearings under Variable Working Conditions with Small Samples

Abstract

The present invention relates to the technical field of fan bearing fault diagnosis, and specifically discloses a small-sample variable-condition fan bearing fault diagnosis method, especially a method for extracting the fault characteristics of fan bearings. During the operation of the fan, rolling bearing faults are likely to occur due to external factors. However, due to the complexity of the fan bearing installation process, maintenance personnel usually replace the bearings in time before the bearing life expires or after a fault occurs to avoid downtime losses. A time series generative adversarial network is used to generate a part of virtual data that conforms to the distribution of the original fault data, enrich the fault characteristics of the bearing data set, and achieve the balance between fault samples. Aiming at the problem that the existing research ignores some time-domain characteristics of the fault data, a dual-channel fusion network model BiLSTM-RepVGG-ECA is constructed. This model strengthens the characteristics of the time-domain data through a bidirectional long short-term memory network and uses RepVGG-ECA to extract the spatial information of the time-frequency diagram after the constant Q transform.

Description

Technical Field

[0001] The present invention relates to the technical field of fan bearing fault diagnosis, and specifically to a small-sample variable-condition fan bearing fault diagnosis method. Background Art

[0002] A fan is a fluid machine widely used in industries, agriculture, construction, transportation and other fields. Its main function is to convert electrical energy or other forms of energy into the kinetic energy and pressure energy of gas flow, so as to achieve purposes such as gas transportation, pressurization, circulation, ventilation, etc. The fan bearing is an important part of the fan. It bears the weight and unbalanced force of the fan rotor and ensures the stable operation of the fan rotor. However, due to the influence of various factors during the operation of the fan bearing, such as load changes, poor lubrication, improper installation, material defects, environmental temperature and humidity changes, etc., various faults are likely to occur, such as wear, cracks, spalling, pitting, etc. These faults will lead to a decline in the performance of the fan bearing, and even cause the fan to shut down or an accident, bringing serious losses to production and safety. Therefore, timely and effectively detecting and diagnosing fan bearing faults is of great significance for improving the operation efficiency and reliability of the fan, extending the service life of the fan, reducing maintenance costs and shutdown losses, and ensuring production safety.

[0003] In recent years, with the development of technologies such as deep learning and sensor monitoring, the fault prediction and health management of core components of industrial equipment have become a popular research topic. The field of bearing fault diagnosis has thus developed rapidly, and research on data mining, machine learning, and deep learning has emerged in an endless stream, promoting industrial development. Therefore, fault diagnosis, monitoring and analyzing the bearing state of a fan bearing in a variable-condition environment can ensure the stability and reliability of the wind turbine generator set, and is of great significance for improving the efficiency of wind power generation. The classification model of deep learning focuses on extracting spatial information. Through the time-frequency diagram generated from the vibration signal, the spatial characteristics of the fault data can be extracted. However, the fan bearing fault data contains rich time-domain information, while only a part of the time-domain information is contained in the time-frequency diagram. In addition, the time-frequency analysis method breaks the continuity of the time-domain characteristics of the bearing fault data, and different parameter settings will result in different time-frequency characteristics. Most previous studies have ignored this problem and have not combined deep learning with time-domain characteristics well.

[0004] Due to the complexity of the fan bearing installation process, maintenance personnel usually replace the bearing in time before the bearing life expires or after a fault occurs to avoid shutdown losses, which makes it difficult to obtain fan bearing fault data. During the fan bearing fault diagnosis process, too few fault samples will lead to a weak generalization ability and low diagnosis accuracy of the trained deep learning model. How to effectively increase the number of fault samples and improve the accuracy of bearing fault classification is one of the main research contents of current fan bearing fault diagnosis. Summary of the Invention

[0005] The present invention provides a method for diagnosing faults of a small-sample variable-condition fan bearing, especially a method for extracting fault features of a fan bearing.

[0006] A method for diagnosing faults of a small-sample variable-condition fan bearing according to the present invention adopts six stages, namely, expansion of fault data, optimization of variational mode decomposition parameters by a dung beetle optimization algorithm, variational mode decomposition, signal reconstruction, data set generation, and fault diagnosis. The dung beetle optimization algorithm is the dung beetle optimization algorithm, and its technological process is as follows:

[0007] Step 101: Start the fault diagnosis process of the variable-condition fan bearing;

[0008] Step 102: Read the bearing fault data of the Case Western Reserve data set, including inner race fault, outer race fault, and rolling element fault data;

[0009] Step 103: Initialize the parameters of the time series generative adversarial network;

[0010] Step 104: Obtain the training result of the time series generative adversarial network, that is, virtual data conforming to the distribution of the original fault data;

[0011] Step 105: Read the data set generated by the time series generative adversarial network, including 9 types of fault data and 1 type of healthy state data;

[0012] Step 106: Select the bearing vibration data of one state as the input data of the dung beetle optimization algorithm;

[0013] Step 107: Construct the fitness function of the dung beetle optimization algorithm;

[0014] Step 108: Initialize the parameters of the dung beetle optimization algorithm and start iterating the parameters of the dung beetle optimization algorithm;

[0015] Step 109: Obtain the operation result of the dung beetle optimization algorithm and substitute the result into the variational mode decomposition algorithm;

[0016] Step 110: Run the variational mode decomposition to obtain the corresponding modes and modal spectra;

[0017] Step 111: Extract fault features to obtain a reconstructed signal;

[0018] Step 112: Perform overlapping sampling and constant Q transform on the reconstructed signal to obtain a time-frequency diagram;

[0019] Step 113: Divide the data set;

[0020] Step 114: Use the divided data set to train the BiLSTM-RepVGG-ECA network to extract spatial information;

[0021] Step 115: Load the weights of the trained BiLSTM-RepVGG-ECA model and use the test set for prediction;

[0022] Step 116: Obtain the fault diagnosis result.

[0023] Preferably, the implementation process of the time series generative adversarial algorithm in steps 103 to 104 includes:

[0024] Step 201: Input the inner race fault, outer race fault, and rolling element fault data as random noise into the generator;

[0025] Step 202: Input the pseudo-data generated by the generator and the real data into the discriminator;

[0026] Step 203: The discriminator performs a binary classification task to determine whether the generated pseudo-data is true or false;

[0027] Step 204: The generator and the discriminator are alternately trained to continuously optimize their respective network parameters until the Nash equilibrium is finally reached;

[0028] Step 205: Obtain the training result of the time series generative adversarial network, that is, virtual data that conforms to the distribution of the original fault data;

[0029] Step 206: Obtain the visualization results of the principal component analysis and t-distributed stochastic neighbor embedding of the original data and the generated data, including the discriminant score and the prediction score;

[0030] Step 207: Verify the effect of the generated data using the discriminant score and the prediction score.

[0031] Preferably, the implementation process of the dung beetle optimization algorithm in steps 106 to 109 includes:

[0032] Step 301: Intercept the bearing fault data segment;

[0033] Step 302: Construct the average envelope entropy function of the data segment;

[0034] Step 303: Establish a dung beetle optimization algorithm model, initialize the parameters, and start M times of algorithm iterations;

[0035] Step 304: In one iteration process, start a loop with the loop number of the population size N and execute steps 305-308;

[0036] Step 305: Update the position information of the rolling ball dung beetle;

[0037] Step 306: Update the position information of the incubating ball;

[0038] Step 307: Update the position information of adult dung beetles;

[0039] Step 308: Update the position information of thief dung beetles;

[0040] Step 309: Increment the number of sub - loops by 1;

[0041] Step 310: After the current sub - loop ends, increment the number of main loops by 1;

[0042] Step 311: Obtain the global optimal position and the fitness function value.

[0043] Preferably, the implementation process of the variational mode decomposition and reconstruction in steps 110 to 111 includes:

[0044] Step 401: Use the result of the dung beetle optimization algorithm as the initialization parameter for variational mode decomposition;

[0045] Step 402: Run variational mode decomposition;

[0046] Step 403: Obtain the time - domain diagram and frequency - spectrum diagram of each mode;

[0047] Step 404: Calculate the kurtosis of each mode;

[0048] Step 405: Select the mode corresponding to the maximum kurtosis for envelope spectrum analysis;

[0049] Step 406: Select the mode corresponding to the maximum kurtosis as the reconstructed signal;

[0050] Step 407: End of fault feature extraction.

[0051] Preferably, the implementation process of the constant Q - transform and dataset production in steps 112 to 113 includes:

[0052] Step 501: Import the reconstructed signal data after fault feature extraction;

[0053] Step 502: Calculate the number of overlapping sampling samples;

[0054] Step 503: Perform a constant Q - transform on each sample;

[0055] Step 504: Produce a bearing dataset in 10 states;

[0056] Step 505: Divide the dataset according to 8:1:1.

[0057] Preferably, the implementation process of the BiLSTM - RepVGG - ECA algorithm in steps 114 to 116 includes:

[0058] Step 601: Add an ECA attention module after the fifth stage of RepVGG to form a lightweight RepVGG-ECA model;

[0059] Step 602: Connect the bidirectional long short-term memory network to the fully connected layer of RepVGG-ECA to form a two-channel fusion network model;

[0060] Step 603: Select the optimizers for the RepVGG-A0 network and the bidirectional long short-term memory network;

[0061] Step 604: Perform fault diagnosis on the fan bearing based on the BiLSTM-RepVGG-ECA model.

[0062] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0063] Taking the fault diagnosis of a fan bearing under variable working conditions as the background, based on the RepVGG network, an ECA attention module is added, and then the bidirectional long short-term memory network is connected to the fully connected layer of RepVGG-ECA to form a two-channel fusion network model. A method for fault diagnosis of a fan bearing based on a BiLSTM-RepVGG-ECA fusion model is proposed. Aiming at the problem that the existing research ignores some time-domain features of fault data, a two-channel fusion network model BiLSTM-RepVGG-ECA is constructed. This model strengthens the features of time-domain data through the bidirectional long short-term memory network and uses RepVGG-ECA to extract the spatial information of the time-frequency diagram after constant Q transform, which can effectively improve the results of bearing fault diagnosis under variable working conditions. During the training process of a deep learning model, if the samples are unbalanced, it often leads to problems such as low generalization ability and poor robustness of the model. To solve the problem of sample imbalance and deal with various faults that may occur in the actual operation of the bearing, a time series generative adversarial network is used to generate a part of virtual data that conforms to the distribution of the original fault data, enrich the fault features of the bearing data set, and achieve the balance between fault samples. At the same time, the original method of manually setting the variational mode decomposition parameters is abandoned, and the dung beetle optimization algorithm is used to accurately calculate the variational mode decomposition parameters of the fault data, effectively avoiding the mode mixing problem, realizing the intelligence and automation of fault feature extraction of the fan bearing under variable working conditions, improving the control level of fan bearing operation and maintenance, reducing the human resource cost, significantly improving the results of fan bearing fault diagnosis, and reducing the economic expenditure of fan management. If it is popularized and applied in China, it will generate great direct and indirect economic and social benefits. Description of the Drawings

[0064] The technical solutions of the present invention will be further described in detail below in conjunction with the drawings and embodiments.

[0065] Figure 1: Flow chart of variable working condition fan bearing fault diagnosis of the present invention;

[0066] Figure 2 : Flow chart of time series generative adversarial algorithm of the present invention;

[0067] Figure 3 : Flow chart of dung beetle optimization algorithm of the present invention;

[0068] Figure 4 : Flow chart of dung beetle optimization algorithm - variational mode decomposition fault feature extraction of the present invention;

[0069] Figure 5 : Flow chart of data set production of the present invention;

[0070] Figure 6 : Flow chart of fault diagnosis process of the present invention;

[0071] Figure 7 : Visualization results of principal component analysis and t-distributed stochastic neighbor embedding of fault data generated by the time series generative adversarial algorithm of the present invention;

[0072] Figure 8 : Loss curves of three models of bidirectional long short-term memory network, BiLSTM-RepVGG and BiLSTM-RepVGG-ECA of the present invention under the Case Western Reserve data set. Detailed implementation manners

[0073] The following will disclose multiple embodiments of the present invention in the form of diagrams. For the sake of clarity, many physical details will be described together in the following narrative. However, it should be understood that these physical details are not used to limit the present invention. That is to say, in some embodiments of the present invention, these physical details are not necessary. In addition, for the sake of simplifying the diagrams, some conventional structures and components will be shown in a simple schematic manner in the diagrams.

[0074] In addition, the technical solutions between various embodiments can be combined with each other, but it must be based on the fact that those of ordinary skill in the art can implement them. When the combination of technical solutions conflicts with each other or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the protection scope required by the present invention.

[0075] Please refer to Figure 1-8 , a small sample variable working condition fan bearing fault diagnosis method of the present invention, the flow of the small sample variable working condition fan bearing fault diagnosis method based on the TimeGAN-BiLSTM-RepVGG-ECA fusion model is as Figure 1It can be divided into six stages: the augmentation of fault data, the optimization of variational mode decomposition parameters by the dung beetle optimization algorithm, variational mode decomposition, signal reconstruction, dataset generation, and fault diagnosis. The technical process is to use a time series generative adversarial network to generate a part of virtual data that conforms to the distribution of the original fault data to enhance the fault samples; use the average envelope entropy function of the modes in the results of the variational mode decomposition algorithm as the fitness function of the dung beetle optimization algorithm, then set the parameters of the dung beetle optimization algorithm, run the dung beetle optimization algorithm to obtain the optimized variational mode decomposition parameters K and α, use the optimized parameters to run the variational mode decomposition algorithm, calculate the kurtosis value of each mode, perform envelope entropy analysis on the mode corresponding to the maximum kurtosis value, and use it as the reconstructed signal of the fault signal. Finally, use the methods of overlapping sampling and constant Q transform on the reconstructed signal to make a dataset for the fault diagnosis of the fan bearing. After dividing the dataset, train it through the BiLSTM-RepVGG-ECA model and conduct tests.

[0076] The entire process of small-sample variable-condition fan bearing fault diagnosis mainly consists of six stages: the augmentation of fault data, the optimization of variational mode decomposition parameters by the dung beetle optimization algorithm, variational mode decomposition, signal reconstruction, dataset generation, and fault diagnosis. The purpose of each stage is different, as follows:

[0077] (1) Stage of fault data augmentation: Use a time series generative adversarial network to achieve the augmentation of fault data.

[0078] The time series generative adversarial mainly consists of four parts: an embedding function, a recovery function, a sequence generator, and a sequence discriminator. The purpose of the embedding function and the recovery function is to form a mapping relationship between the feature space and the latent feature space, enabling the network to learn the data features in the latent feature space through low-dimensional data, thereby reducing the high-dimensional characteristics of the original generative adversarial network's learning space. The embedding function converts static features and temporal features into latent codes for learning data features in the shallow feature space, and the recovery function converts the latent codes into static features and temporal features to obtain the reconstructed data. The sequence generator generates a latent code composed of random vectors of static features and temporal features through a generation function, and the sequence discriminator is used to classify the latent code of the generator and return the probabilities of discriminating real data and generated data.

[0079] The time series generative adversarial network combines unlabeled and labeled learning and uses three loss functions to train the network, namely the reconstruction loss function Ι R 、the unlabeled loss function Ι U 、the labeled loss function Ι S 。

[0080] The training results of the time series generative adversarial network are as Figure 7As shown, the visualization results of principal component analysis and t-distributed stochastic neighbor embedding show that the distributions of the original data and the generated data are very similar, that is, the generated data can effectively simulate the distribution of the original data.

[0081] (2) The stage of optimizing the parameters of variational mode decomposition by the dung beetle optimization algorithm: It includes at least the modeling of the dung beetle optimization algorithm and the selection of the fitness function.

[0082] Model establishment of the dung beetle optimization algorithm: The dung beetle optimization algorithm mainly simulates the behaviors of dung beetles in nature, such as rolling balls, dancing, foraging, stealing, and reproducing.

[0083] (a) Rolling the ball

[0084] To simulate the behavior of rolling a ball, the dung beetle needs to move along a given direction in the entire search space. Using sunlight as the navigation for the dung beetle's movement. Assuming that the light source intensity also affects the dung beetle's path, during the rolling process, the calculation formula for updating the position of the dung beetle is:

[0085] x i (t + 1) = x i (t) + α × k × x i (t - 1) + b × |x i (t) - X w | (1)

[0086] In formula (1): t is the current iteration number; x i (t) is the position information of the dung beetle; k is a constant representing the deflection coefficient, and its value range is (0, 0.2]; b is in the value range of (0, 1); α is the natural coefficient; X w is the global worst position; |x i (t) - X w | is the change in light intensity;

[0087] In formula (1), the appropriate selection of parameters k and b is crucial. Usually, the value of k is 0.1 and the value of b is 0.3. α simulates the natural factor that causes the dung beetle to deviate from the original direction. When α is 1, it means the direction does not deviate. When α is -1, it means the direction deviates from the original direction. |x i (t) - X w | The larger the value, the weaker the light source. By controlling the size of the light source through X w , the search range can be expanded. The addition of this term enables the dung beetle optimization algorithm to search the entire problem space as much as possible during the optimization process while reducing the possibility of falling into local optima.

[0088] (b) Dancing

[0089] When the dung beetle encounters an obstacle and cannot move forward, it will reorient itself by dancing to obtain a new route. To simulate the dancing behavior, the tangent function is used to obtain a new rolling direction. After obtaining the new direction, the dung beetle will roll the ball in this direction, and the position update formula is shown in Equation (2).

[0090] x i (t + 1) = x i (t) + tan(θ)|x i (t) - x i (t - 1)| (2)

[0091] In Equation (2): θ is the deflection angle, and its value range is [0, π]; x i (t) is the position information of the dung beetle;

[0092] In the above formula, when θ takes the values of 0, π / 2, and π, it is considered that the position has not been updated. When taking other values, it is considered that the position of the dung beetle has been updated.

[0093] (c) Reproduction

[0094] To provide a safe environment for the offspring, it is crucial for the dung beetle to choose a suitable oviposition site. To simulate the oviposition area of the female dung beetle, a boundary selection strategy is proposed, and its conceptual definition is shown in Equation (3).

[0095]

[0096] In Equation (3): X * is the current local optimal position; Lb * , Ub * are the lower and upper bounds of the oviposition area; Lb, Ub are the lower and upper bounds of the optimization problem; T max represents the maximum number of iterations;

[0097] Once the oviposition area is determined, the female dung beetle will choose to lay eggs in the egg ball in this area. In the dung beetle optimization algorithm, it is assumed that each female dung beetle will lay only one egg in each iteration. According to Equation (3), it can be seen that the boundary range of the oviposition area changes dynamically with the value. Therefore, the position of the egg ball also changes dynamically during the iteration process, and its update formula is as follows:

[0098] B i (t + 1) = X * + b1×(B i (t) - Lb * ) + b2×(B i(t) - Ub * ) (4)

[0099] In Equation (4): B i(t) is the position information at the t-th iteration; b1 and b2 are two independent random vectors; X * is the current local optimal position; Lb * , Ub * are the lower and upper bounds of the spawning area;

[0100] (d) Foraging

[0101] Based on the foraging process of adult dung beetles, an optimal foraging area model is established, as shown in Equation (6):

[0102]

[0103] In Equation (5): X b is the global optimal position; Lb b , Ub b are the lower and upper bounds of the best foraging area; Lb and Ub are the lower and upper bounds of the optimization problem;

[0104] According to the principle of updating the breeding position, the dung beetle foraging position update formula can be obtained as follows:

[0105] x i (t + 1) = x i (t) + C1 × (x i (t) - Lb b ) + C2 × (x i (t) - Ub b ) (6)

[0106] In Equation (6): C1 and C2 are random numbers following a normal distribution; Lb b , Ub b are the lower and upper bounds of the best foraging area; x i (t) is the position information of the dung beetle;

[0107] (e) Stealing

[0108] In the dung beetle population, a part of the dung beetles will steal the dung balls of other dung beetles, which are called thief dung beetles. From Equation (5), it can be seen that X b is the optimal food source. Therefore, it is assumed that the surrounding of X b is the best position to compete for food. The position information of the thief dung beetles is updated according to the following formula:

[0109] x i (t + 1) = X b + S × g × (|x i (t) - X * | + |x i (t) - X b |) (7)

[0110] In Equation (7): S is a constant; g is a random vector subject to a normal distribution; x i (t) is the position information of the dung beetle; X b is the global optimal position; X * is the current local optimal position;

[0111] Selection of the fitness function: The fitness function selected in the present invention is the envelope entropy of the mode after variational mode decomposition as the fitness function for dung beetle optimization, and its calculation formula is as follows:

[0112]

[0113] In Equation (8): a(j) is the envelope signal after Hilbert transform demodulation; p(j) is the normalized representation of a(j); E p is the envelope entropy; j is 1, 2, 3, 4.....N, and N is a natural number;

[0114] (3) Variational mode decomposition stage: This stage includes performing variational mode decomposition on the input data using the result calculated by the dung beetle optimization algorithm. The principle of variational mode decomposition is: First is the establishment process of the variational model.

[0115] Decompose the signal f(t) into several modal component signals u k (t), as shown in Equation (9). Demodulate each mode to obtain the single-sided spectrum

[0116]

[0117] In Equation (9): σ(t) is the unit impulse function; u k (t) is the modal component signal;

[0118] By introducing the center frequency w k , modulate the spectrum of each demodulated signal to the corresponding baseband, as shown in Equation (10)

[0119]

[0120] Calculate the square norm of the gradient of Equation (10) and estimate the bandwidth corresponding to each modal component to obtain the following constrained variational problem model

[0121]

[0122] In Equation (11): u k are the k modal components after decomposition; w k are the center frequencies of the k modal components; j is the imaginary unit; e is a constant; f(t) is the input signal;

[0123] Then there is the problem of solving the variational model. To solve the constraint problem established in Equation (12), a quadratic penalty factor α and a Lagrange multiplier operator λ are introduced, transforming this problem into an unconstrained variational problem as follows.

[0124]

[0125] In Equation (12): λ(t) is the Lagrange multiplier operator; < > represents the inner product operation; α is the penalty parameter;

[0126] In Equation (12), by continuously updating λ(t) n+1 the numerical values of the three terms, the optimal solution of the variational problem can be obtained. λ(t) n+1 The update formulas are respectively Equations (13) to (15).

[0127]

[0128]

[0129]

[0130] In Equations (13) to (15): f(ω) is the Fourier transform of f(t); u i (ω) is the Fourier transform of u i (t); λ(ω) is the Fourier transform of λ(t); u k is the k decomposed modal components; ω k is the center frequency of the k modal components; 1 is the noise tolerance; α is the penalty parameter;

[0131] According to the above steps, the mode, spectrum, and center frequency of the signal can be solved.

[0132] (4) Signal reconstruction stage: Calculate the kurtosis of the modes of the dung beetle optimization algorithm - variational mode decomposition. The calculation formula (16) is as follows:

[0133]

[0134] x i is the vibration amplitude corresponding to the discrete sequence points of the time-domain waveform; is the average amplitude of the discrete sequence; N is the number of discrete sequence points; K is the kurtosis calculated from the decomposed modes;

[0135] Select the mode with the maximum kurtosis as the extracted bearing fault feature.

[0136] (5) Dataset generation stage: Use the overlapping sampling method for the extracted fault features to generate a constant Q transform time-frequency diagram. The calculation formula (17) is as follows:

[0137]

[0138] In Equation (18): L is the data length of overlapping sampling; T is the length of the constant Q transform points; S is the moving step; N is the generated data set;

[0139] Divide the generated data set into a training set, a validation set, and a test set according to the ratio of 8:1:1.

[0140] (6) Fault diagnosis stage: Input the training set and the validation set into the model. After 60 times of training, save the weights. During testing, load the weights to perform bearing fault diagnosis.

[0141] The BiLSTM-RepVGG-ECA model adds an ECA attention module after the fifth stage of RepVGG to form a RepVGG-ECA lightweight model, which strengthens the channel feature extraction ability while reducing the number of model parameters. Then, connect the bidirectional long short-term memory network and the fully connected layer of RepVGG-ECA to form a two-channel fusion network model: BiLSTM-RepVGG-ECA. This model uses the bidirectional long short-term memory network as the feature extraction network for bearing fault time-domain data, and RepVGG-ECA as the feature extraction network for time-frequency space information.

[0142] The training curves of the three models, namely the bidirectional long short-term memory network, BiLSTM-RepVGG, and BiLSTM-RepVGG-ECA, under the Case Western Reserve dataset are as shown. From the Figure 8 Figure 8 loss curve shown, it can be seen that the training effect of the original bidirectional long short-term memory network is poor. After fusing with the RepVGG network, the performance has been significantly improved. After adding the ECA attention, the convergence speed of the model becomes very fast, and the accuracy has also been improved to a certain extent, indicating that the improvements made by BiLSTM-RepVGG-ECA are all effective.

[0143] The working process and steps of the present invention are as follows:

[0144] Step 101: Start the variable-condition fan bearing fault diagnosis process

[0145] Step 102: Read the bearing fault data of the Case Western Reserve dataset, including inner race fault, outer race fault, and rolling element fault data;

[0146] Step 103: Initialize the parameters of the time series generative adversarial network;

[0147] Step 104: Obtain the training result of the time series generative adversarial network, that is, virtual data conforming to the distribution of the original fault data;

[0148] Step 105: Read the dataset generated by the time series generative adversarial network, including 9 types of fault data and 1 type of healthy state data;

[0149] Step 106: Select the bearing vibration data of one state as the input data for the dung beetle optimization algorithm;

[0150] Step 107: Construct the fitness function of the dung beetle optimization algorithm;

[0151] Step 108: Initialize the parameters of the dung beetle optimization algorithm and start iterating the parameters of the dung beetle optimization algorithm;

[0152] Step 109: Obtain the running result of the dung beetle optimization algorithm and substitute the result into the variational mode decomposition algorithm;

[0153] Step 110: Run the variational mode decomposition to obtain the corresponding modes and modal spectra;

[0154] Step 111: Extract fault features to obtain the reconstructed signal;

[0155] Step 112: Perform overlapping sampling and constant Q transform on the reconstructed signal to obtain the time-frequency diagram;

[0156] Step 113: Divide the dataset;

[0157] Step 114: Use the divided dataset to train the BiLSTM-RepVGG-ECA network to extract spatial information;

[0158] Step 115: Load the weights of the trained BiLSTM-RepVGG-ECA model and use the test set for prediction;

[0159] Step 116: Obtain the fault diagnosis result.

[0160] The implementation process of the time series generative adversarial algorithm in steps 103 to 104 includes:

[0161] Step 201: Input the inner race fault, outer race fault, and rolling element fault data into the random noise generator;

[0162] Step 202: Input the pseudo data generated by the generator and the real data into the discriminator;

[0163] Step 203: The discriminator performs a binary classification task to judge whether the generated pseudo data is true or false;

[0164] Step 204: Alternately train the generator and the discriminator, continuously optimize their respective network parameters, and finally reach the Nash equilibrium;

[0165] Step 205: Obtain the training result of the time series generative adversarial network, that is, virtual data conforming to the distribution of the original fault data;

[0166] Step 206: Obtain the visualization results of the principal component analysis and t-distributed stochastic neighbor embedding of the original data and the generated data respectively;

[0167] Step 207: Verify the effect of the generated data using discriminant scores (D-Scores) and prediction scores (P-Scores).

[0168] The implementation process of the dung beetle optimization algorithm in steps 106 to 109 includes:

[0169] Step 301: Intercept the bearing fault data segment;

[0170] Step 302: Construct the average envelope entropy function of the data segment;

[0171] Step 303: Establish a dung beetle optimization algorithm model, initialize the parameters, and start the algorithm iteration for M times;

[0172] Step 304: In one iteration process, start a loop with the loop number being the population size N, and execute steps 305 - 308;

[0173] Step 305: Update the position information of the rolling ball dung beetle;

[0174] Step 306: Update the position information of the incubating ball;

[0175] Step 307: Update the position information of the adult dung beetle;

[0176] Step 308: Update the position information of the thief dung beetle;

[0177] Step 309: Increment the number of sub-loops by 1;

[0178] Step 310: After the current loop ends, increment the number of main loops by 1;

[0179] Step 311: Obtain the global optimal position and the fitness function value.

[0180] The implementation process of variational mode decomposition and reconstruction in steps 110 to 111 includes:

[0181] Step 401: Use the result of the dung beetle optimization algorithm as the initialization parameter of the variational mode decomposition;

[0182] Step 402: Run the variational mode decomposition;

[0183] Step 403: Obtain the time domain diagram and frequency spectrum diagram of each mode;

[0184] Step 404: Calculate the kurtosis of each mode;

[0185] Step 405: Select the mode corresponding to the maximum kurtosis for envelope spectrum analysis;

[0186] Step 406: Select the mode corresponding to the maximum kurtosis as the reconstructed signal;

[0187] Step 407: End of fault feature extraction.

[0188] The implementation process of the above-mentioned steps 112 to 113 for constant Q transform and dataset production includes:

[0189] Step 501: Import the reconstructed signal data after fault feature extraction;

[0190] Step 502: Calculate the number of overlapping sampling samples;

[0191] Step 503: Perform constant Q transform on each sample;

[0192] Step 504: Produce a bearing dataset with 10 states;

[0193] Step 505: Divide the dataset according to 8:1:1.

[0194] The implementation process of the above-mentioned steps 114 to 116 for the BiLSTM-RepVGG-ECA algorithm includes:

[0195] Step 601: Add an ECA attention module after the fifth stage of RepVGG to form a lightweight RepVGG-ECA model;

[0196] Step 602: Connect the bidirectional long short-term memory network and the fully connected layer of RepVGG-ECA to form a two-channel fusion network model;

[0197] Step 603: Select the optimizers for the RepVGG-A0 network and the bidirectional long short-term memory network;

[0198] Step 604: Perform fault diagnosis on the fan bearing based on the BiLSTM-RepVGG-ECA model.

[0199] The above are only the embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the scope of the claims of the present invention.

Claims

1. A fault diagnosis method for a small-sample variable-condition fan bearing, which adopts six stages: expansion of fault data, optimization of variational mode decomposition parameters by the dung beetle optimization algorithm, variational mode decomposition, signal reconstruction, dataset generation, and fault diagnosis. The technological process of the dung beetle optimization algorithm is as follows: Step 101: Start the variable-condition fan bearing fault diagnosis process; Step 102: Read the bearing fault data of the Case Western Reserve dataset, including inner race fault, outer race fault, and rolling element fault data; Step 103: Initialize the parameters of the time series generative adversarial network; Step 104: Obtain the training result of the time series generative adversarial network, that is, virtual data conforming to the distribution of the original fault data; Step 105: Read the dataset generated by the time series generative adversarial network, including 9 types of fault data and 1 type of healthy state data; Step 106: Select the bearing vibration data of one state as the input data of the dung beetle optimization algorithm; Step 107: Construct the fitness function of the dung beetle optimization algorithm; Step 108: Initialize the parameters of the dung beetle optimization algorithm and start iterating the parameters of the dung beetle optimization algorithm; Step 109: Obtain the operation result of the dung beetle optimization algorithm and substitute the result into the variational mode decomposition algorithm; Step 110: Run the variational mode decomposition to obtain the corresponding modes and modal spectra; Step 111: Perform fault feature extraction to obtain the reconstructed signal; Step 112: Perform overlapping sampling and constant Q transform on the reconstructed signal to obtain the time-frequency diagram; Step 113: Divide the dataset; Step 114: Use the divided dataset to train the BiLSTM-RepVGG-ECA network to extract spatial information; Step 115: Load the weights of the trained BiLSTM-RepVGG-ECA model and use the test set for prediction; Step 116: Obtain the fault diagnosis result; The implementation process of the variational mode decomposition and reconstruction from Step 110 to Step 111 includes: Step 401: Use the result of the dung beetle optimization algorithm as the initialization parameter of the variational mode decomposition; Step 402: Run the variational mode decomposition; Step 403: Obtain the time-domain diagram and frequency-spectrum diagram of each mode; Step 404: Calculate the kurtosis of each mode; Step 405: Select the mode corresponding to the maximum kurtosis for envelope spectrum analysis; Step 406: Select the mode corresponding to the maximum kurtosis as the reconstructed signal; Step 407: End the fault feature extraction; The implementation process of the BiLSTM-RepVGG-ECA algorithm from Step 114 to Step 116 includes: Step 601: Add an ECA attention module after the fifth stage of RepVGG to form a lightweight RepVGG-ECA model; Step 602: Connect the bidirectional long short-term memory network and the fully connected layer of RepVGG-ECA to form a two-channel fusion network model; Step 603: Select the optimizers of the RepVGG-A0 network and the bidirectional long short-term memory network; Step 604: Perform fault diagnosis on the fan bearing based on the BiLSTM-RepVGG-ECA model.

2. The small-sample variable-condition fault diagnosis method for a fan bearing according to claim 1, wherein: The implementation process of the time series generative adversarial algorithm from step 103 to step 104 includes: Step 201: Input the inner race fault, outer race fault, and rolling element fault data into the generator as random noise. Step 202: Input the pseudo-data generated by the generator and the real data into the discriminator. Step 203: The discriminator performs a binary classification task to determine whether the generated pseudo-data is true or false. Step 204: The generator and the discriminator are alternately trained to continuously optimize their respective network parameters until the Nash equilibrium is finally reached. Step 205: Obtain the training result of the time series generative adversarial network, that is, the virtual data that conforms to the distribution of the original fault data. Step 206: Obtain the visualization results of the principal component analysis and t-distributed stochastic neighbor embedding of the original data and the generated data, including the discriminant score and the prediction score. Step 207: Verify the effect of the generated data using the discriminant score and the prediction score.

3. A small-sample variable-condition fan bearing fault diagnosis method according to claim 1, characterized in that: The implementation process of the dung beetle optimization algorithm from step 106 to step 109 includes: Step 301: Intercept the bearing fault data segment. Step 302: Construct the average envelope entropy function of the data segment. Step 303: Establish a dung beetle optimization algorithm model, initialize the parameters, and start the algorithm iteration for M times. Step 304: In one iteration process, start a loop with the loop number equal to the population size N, and execute steps 30, 5-308. Step 305: Update the position information of the rolling ball dung beetle. Step 306: Update the position information of the incubating ball. Step 307: Update the position information of the adult dung beetle. Step 308: Update the position information of the thief dung beetle. Step 309: Increment the number of sub-loops by 1. Step 310: After the current loop ends, increment the number of main loops by 1. Step 311: Obtain the global optimal position and the fitness function value.

4. A small-sample variable-condition fan bearing fault diagnosis method according to claim 1, characterized in that: The implementation process of the constant Q transform and dataset production from step 112 to step 113 includes: Step 501: Import the reconstructed signal data after fault feature extraction. Step 502: Calculate the number of overlapping sampling samples. Step 503: Perform a constant Q transform on each sample. Step 504: Produce a bearing dataset with ten states. Step 505: Divide the dataset according to 8:1:1.

Images