Fan fault diagnosis method under data imbalance condition
Through the combination of deep learning technology and improved SMOTE algorithm and CNN-GRU model, the problem of data imbalance in fan fault diagnosis is solved, and more efficient and accurate fault diagnosis is achieved.
Patent Information
- Application Number
- CN202510174814.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-18
- Publication Date
- 2025-05-30
AI Technical Summary
Fans often face data imbalance during operation, resulting in inefficient and low accuracy of traditional fault diagnosis methods.
Deep learning technology is adopted, combined with the improved SMOTE algorithm and CNN-GRU model, data preprocessing and feature extraction are carried out, the category distribution of the training set is balanced, and fault mode recognition is achieved through model training.
It significantly improves the accuracy and reliability of fan fault diagnosis, effectively solves the problem of data imbalance, and improves diagnosis efficiency.
Smart Images
Figure CN120067691A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of fan fault diagnosis, and more specifically, to a fan fault diagnosis method under data imbalance conditions. Background Art
[0002] As an indispensable key device in industrial production, the stable operation of a fan is directly related to production efficiency and safety. However, during long-term operation, due to factors introduced during the design, production, manufacturing, transportation processes, as well as internal and external factors during the service stage, various types of faults are inevitable for the fan. In particular, the fan often faces the problem of data imbalance during operation, that is, the amount of data in the normal operation state is much larger than that in the fault state, which poses a challenge to fault diagnosis. Traditional fan fault diagnosis methods mostly rely on time-domain and frequency-domain analysis of vibration signals, but there are problems of low efficiency and low accuracy when dealing with unbalanced data.
[0003] In order to improve the accuracy and efficiency of fan fault diagnosis, researchers have started to explore deep learning-based methods to solve the problem of inter-class imbalance. Although there are various methods, the field of fan fault diagnosis still faces some challenges and problems, and further research and exploration are needed, especially in-depth research and exploration in aspects such as data preprocessing, feature selection, and machine learning algorithm optimization, in order to achieve more accurate and efficient fault diagnosis and ensure the stable operation and production safety of fan equipment. Summary of the Invention
[0004] The present invention provides a fan fault diagnosis method under data imbalance conditions to solve the problem of sample imbalance encountered by industrial fans during operation.
[0005] By applying deep learning technology, this method can effectively learn from unbalanced data and identify potential fault patterns, significantly improving the accuracy and reliability of fan fault diagnosis.
[0006] According to one aspect of the present invention, there is provided a fan fault diagnosis method under data imbalance conditions, characterized by comprising the following steps:
[0007] Step 1: Install a vibration acceleration sensor on the fan bearing, and convert the collected vibration signal into an electrical signal through a professional data acquisition card;
[0008] Step 2: Perform data cleaning and fast Fourier transform on the collected data for subsequent analysis;
[0009] Step 3: Divide the data set into a majority-class sample set and a minority-class sample set according to the number of samples, where the minority-class samples are fault samples;
[0010] Step 4: Use the improved SMOTE algorithm to generate new samples in the minority class sample set of the training set X, and at the same time combine the undersampling method for the majority class to balance the training set, that is, obtain the new training set Xnew;
[0011] Step 5: Build a fault diagnosis model for fault diagnosis, and select the CNN-GRU combined model as the fault diagnosis model;
[0012] Step 6: Train the data generation model, and input the preprocessed and enhanced data in Step 3 into two models for training;
[0013] Step 7: Use the test set to verify the diagnostic effect of the model, and finally compare the fault diagnosis results before and after balancing the training set to verify the effectiveness of the improved SMOTE algorithm.
[0014] Preferably based on the above scheme, the vibration acceleration sensor in Step 1 is a piezoelectric acceleration sensor, which acquires vibration signals and stores data after filtering out noise and filtering and amplifying.
[0015] Preferably based on the above scheme, Step 2 specifically includes the following steps:
[0016] Step 2.1: Data preprocessing. First, clean the data to remove noise samples;
[0017] Step 2.1.1: Use the method of linear interpolation to process missing values in the data. For sequence data, missing values can be filled with the weighted average of adjacent values. The specific formula is as follows:
[0018]
[0019] Among them, X missing is the missing value, t represents the current time point, and t - 1 and t + 1 represent the time points before and after the missing value;
[0020] Step 2.1.2: Use the method of Z-score outlier detection to detect and process outliers in the data. By quantifying the deviation degree of data points from the average value, identify those outliers that are significantly different from most of the data. The specific formula is as follows:
[0021]
[0022] Among them, X represents the current observed value, μ represents the mean of the feature column, σ represents the standard deviation of the feature column, which is used to measure the dispersion degree of the data, Z represents the Z-score value, which measures the deviation degree of the current observed value X from the mean μ. When |Z| > 3, this observed value is regarded as an outlier;
[0023] Step 2.1.2: Apply min-max normalization to normalize the data, scale the numerical range of all features to a specified interval (usually from 0 to 1) to eliminate the influence of the dimension between different features, making the model training more stable and efficient. The specific formula is as follows:
[0024]
[0025] Where X represents the original feature value, min(X) represents the minimum value of feature X column, max(X) represents the maximum value of feature X column, and X' represents the normalized feature value, with the value range of [0, 1];
[0026] Step 2.2: Perform a fast Fourier transform on the normalized vibration signal X' to obtain its frequency domain representation X freq [k], and the specific formula is:
[0027]
[0028] Where is the imaginary unit, is the index of the frequency domain, =0, 1, 2,..., N - 1.
[0029] Based on the above scheme, preferably, step 4 specifically includes:
[0030] Step 4.1, for the fault sample set, use the k-means++ algorithm for clustering, divide it into m clusters, select the center point of each clustering cluster as the oversampling point, and for each sample x i , use the Gaussian kernel function to calculate its local density ρ i :
[0031]
[0032] d ij represents the distance between the sample point x i and the sample point x j , d c represents the truncation distance;
[0033] Step 4.2, oversample the minority class samples according to the optimal sampling rate, and obtain the center point of each cluster as the most representative sample point and add it to the training set;
[0034] Step 4.3, use K-nearest neighbors to perform noise filtering on the majority class samples, and determine the k1 nearest neighbor samples of each majority class sample through the k-nearest neighbor algorithm; for each majority class sample, when all its k1 nearest neighbor samples belong to the minority class, then this majority class sample is judged as a noise sample and removed from the dataset. Among them, the K-nearest neighbor algorithm uses the commonly used distance metric method - Euclidean distance, and its specific formula is:
[0035]
[0036] In the Euclidean distance formula, x i and x j represent the coordinate vectors of two points in the feature space; specifically:
[0037] is a point in the feature space, where n is the dimension of the feature space, is the coordinate value of point x i on the first dimension;
[0038] is another point in the feature space, similar to x i and it also has n coordinate values, is the coordinate value of point x j on the first dimension;
[0039] Step 4.4, generate new synthetic samples between the minority class samples through the SMOTE algorithm, increase the number of minority class samples to be close to the number of majority class samples, perform undersampling on the majority class samples, randomly remove some samples from them to make the number of majority class samples close to the increased number of minority class samples, and integrate the minority class samples generated by SMOTE with the undersampled majority class samples to form a more balanced dataset.
[0040] Preferably based on the above scheme, step 5 specifically includes:
[0041] Step 5.1: Construct a convolutional neural network (CNN);
[0042] The convolutional neural network adopts the methods of local connection and weight sharing, and alternately uses convolutional layers and pooling layers to extract features from the vibration signal data of industrial equipment, and then learns local spatial patterns to obtain an effective low-dimensional feature representation;
[0043] Step 5.2: Construct a gated recurrent unit network (GRU);
[0044] The GRU network adds an update gate and a reset gate on the basis of the recurrent neural network, and controls the information state of the GRU network at each moment through these two gates;
[0045] The GRU network has three inputs at time t:
[0046] The input value x t of the network at the current moment, that is, the sequence data after feature extraction;
[0047] The output value h t-1 of the hidden layer of the GRU network at the previous moment;
[0048] The hidden state h at the previous moment t-1 ;
[0049] The GRU network has two outputs at time t: the output value h of the hidden layer at the current moment t and the updated hidden state h t ;
[0050] The update gate determines the amount of information of the hidden state h at the previous moment t-1 reserved to the current hidden state h t ;
[0051] The reset gate determines the influence degree of the input x of the network at the current moment t on the hidden state h at the previous moment t-1 ;
[0052] The output gate (usually merged with the update gate in GRU) determines the amount of information of the current hidden state h t as the output h of the hidden layer at the current moment t ;
[0053] Step 5.3: Combine the convolutional neural network (CNN) for feature extraction, and then use the gated recurrent unit (GRU) to process the time series information
[0054] Compared with the prior art, a fan fault diagnosis method under data imbalance of the present invention has the following advantages
[0055] (1) The concept of increasing local density is used to refine the distribution of fault samples, and oversampling is performed at samples with small local density, avoiding that the number of samples in the fault sample subset is too scarce or the distribution position deviates too far from the sample space and is treated as noise samples, resulting in within-class imbalance
[0056] (2) By oversampling the minority class samples at the optimal sampling rate, the center points of each cluster are obtained as the most representative sample points and added to the training set, thereby increasing the number of minority class samples and achieving the purpose of reducing the data imbalance rate. The problem that the SMOTE algorithm does not have a clear sampling rate is solved, and the algorithm has good efficiency
[0057] (3) Combining the undersampling method, the SMOTE algorithm generates new synthetic samples between the minority class samples, increasing the number of minority class samples to be close to the number of majority class samples. Then, undersampling is performed on the majority class samples, and some samples are randomly removed from them to make the number of majority class samples close to the number of increased minority class samples. Finally, the minority class samples generated by SMOTE are integrated with the undersampled majority class samples to form a more balanced data set. This method not only retains the characteristics of the minority class samples but also reduces data bias, helping to improve the classification performance of the model
[0058] (4) Select the CNN-GRU combined model, extract features by combining with CNN, and then use GRU to process time series information to further improve the ability to identify time series faults. Through this combination of multiple levels and multiple models, complex fan fault modes can be better processed. Brief Description of the Drawings
[0059] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for describing the embodiments will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings. In the drawings:
[0060] Figure 1 It is a flowchart of a fan fault diagnosis method under sample imbalance driven by deep learning of the present invention. Detailed Embodiments
[0061] The following will further describe in detail the specific embodiments of the present invention in conjunction with the accompanying drawings and embodiments. The following embodiments are used to illustrate the present invention, but are not used to limit the scope of the present invention.
[0062] It should be understood that when used in this specification and the appended claims, the term "comprising" indicates the presence of described features, wholes, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components, and / or groups.
[0063] To make the drawings concise, only the parts related to the present invention are schematically shown in each drawing, and they do not represent their actual structures as products. Additionally, to make the drawings concise and easy to understand, in some drawings, components with the same structure or function are only schematically shown for one of them, or only one of them is marked. In this article, "one" not only means "only this one", but also can mean "more than one" situation.
[0064] It should also be further understood that the term "and / or" used in the specification and appended claims of this application refers to any combination and all possible combinations of one or more of the related listed items, and includes these combinations.
[0065] In the embodiments shown in the drawings, the indication of directions (such as up, down, left, right, front, and back) is used to explain that the structures and movements of various components of the present invention are not absolute but relative. When these components are in the positions shown in the drawings, these explanations are appropriate. If the descriptions of the positions of these components change, the indication of these directions also changes accordingly.
[0066] In addition, in the description of this application, terms such as "first" and "second" are only used for differential description and cannot be construed as indicating or implying relative importance.
[0067] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the specific implementation manners of the present invention will be described below with reference to the accompanying drawings. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings and other implementation manners can be obtained.
[0068] Please refer to Figure 1 , a rolling fan diagnosis method under the condition of sample imbalance proposed by the present invention of the present invention includes the following steps:
[0069] Step 1: Data acquisition. Install a vibration sensor on the fan bearing, use a piezoelectric acceleration sensor to obtain the vibration signal of the fan bearing, and convert the collected vibration signal into an electrical signal through a professional data acquisition card. These electrical signals are processed by filtering out noise and filtering and amplifying to improve the signal quality, and then stored to provide a high-quality data basis for subsequent analysis.
[0070] Step 2: Data preprocessing. Perform data cleaning and fast Fourier transform (FFT) on the collected data to facilitate subsequent analysis.
[0071] Step 2.1: In data preprocessing, first perform cleaning on the data to remove noise samples to avoid interpolating based on noise points and thus forming more noise points. It mainly includes several steps such as missing value processing, outlier detection, and normalization processing.
[0072] Step 2.1.1, use the method of linear interpolation to process the missing values of the data to ensure the integrity and consistency of the data set, thereby improving the accuracy and reliability of data analysis. The specific formula is as follows:
[0073]
[0074] For sequence data, the missing values can be filled with the weighted average of adjacent values. Among them, X missing is the missing value, t represents the current time point, and t - 1 and t + 1 represent the time points before and after the missing value;
[0075] Step 2.1.2, use the method of Z-score outlier detection to detect and process outliers in the data, and identify those outliers that are significantly different from most of the data by quantifying the deviation degree of the data points from the average value. The specific formula is as follows:
[0076]
[0077] Where X represents the current observed value, μ represents the mean of the feature column, σ represents the standard deviation of the feature column, which is used to measure the degree of data dispersion, Z represents the Z-score value, which measures the deviation degree of the current observed value X from the mean μ. When |Z| > 3, this observed value is usually regarded as an outlier;
[0078] Step 2.1.3: Apply min-max normalization to normalize the data, and scale the numerical range of all features to a specified interval (usually from 0 to 1) to eliminate the influence of the dimension between different features, making the model training more stable and efficient. The specific formula is as follows:
[0079]
[0080] Where X represents the original feature value, min(X) represents the minimum value of the feature X column, max(X) represents the maximum value of the feature X column, and X' represents the normalized feature value, and its value range is [0, 1];
[0081] Step 2.2: Perform a fast Fourier transform on the normalized vibration signal X' to convert the time-domain signal to the frequency domain to analyze its frequency components and obtain its frequency-domain representation X freq [k]. The specific formula is:
[0082]
[0083] Where j is the imaginary unit, k is the index of the frequency domain, k = 0, 1, 2,..., N - 1.
[0084] Step 3: Dataset classification. After normalization and FFT processing, the dataset is divided into a majority-class sample set and a minority-class sample set according to the number of samples. The majority-class sample set usually represents the normal state, while the minority-class sample set represents the fault state. This classification simulates the problem of unbalanced vibration data of the fan. Through this division, we can better simulate and process the unbalanced dataset encountered in practical applications, providing an accurate data basis for subsequent data augmentation, model training, and fault diagnosis;
[0085] Step 4: Data augmentation. Use the improved SMOTE algorithm to generate new samples in the minority class sample set of the training set X. These new samples not only expand the originally insufficient minority class, but also increase the diversity of samples by interpolating in the feature space, avoiding the overfitting problem caused by simple replication. This method effectively creates new and representative data points without changing the original data distribution, thus balancing the class distribution of the training set, that is, obtaining the new training set Xnew. The specific improved SMOTE algorithm is as follows;
[0086] Step 4.1. For the fault sample set, use the k-means++ algorithm for clustering, divide it into m clusters, and select the center point of each clustering cluster as the oversampling point, which solves the problem of insufficient importance of samples. At the same time, the concept of local density is added to refine the distribution of fault samples. For each sample x i , use the Gaussian kernel function to calculate its local density ρ i :
[0087]
[0088] d ij is the distance between the sample point x i and the sample point x j , and d c is the truncation distance.
[0089] Oversampling is also required at samples with relatively small local density to avoid the situation where the number of samples in the fault sample subset is too small or the distribution position is too far from the sample space and is treated as a noise sample, resulting in intra-class imbalance;
[0090] Step 4.2. Oversample the minority class samples at the optimal sampling rate to obtain the center point of each cluster as the most representative sample point and add it to the training set, thereby increasing the number of minority class samples and achieving the purpose of reducing the data imbalance rate; solve the problem that the SMOTE algorithm does not have a clear sampling rate, and the algorithm has good efficiency;
[0091] The SMOTE algorithm only operates on minority class samples. Here, use K-nearest neighbors to filter noise from majority class samples. Specifically, in the dataset, determine the k1 nearest neighbor samples of each majority class sample through the k-nearest neighbor algorithm. For each majority class sample, if all its k1 nearest neighbor samples belong to the minority class, then this majority class sample is judged as a noise sample and removed from the dataset. Here, the K-nearest neighbor algorithm uses the commonly used distance metric method - Euclidean distance, and its specific formula is:
[0092]
[0093] In the Euclidean distance formula, x i and x j represent the coordinate vectors of two points in the feature space.
[0094] Specifically:
[0095] is a point in the feature space, where n is the dimension of the feature space, is the coordinate value of point x i on the first dimension.
[0096] is another point in the feature space, similar to x i and it also has n coordinate values, is the coordinate value of point x j on the first dimension;
[0097] Step 4.4: Generate new synthetic samples between the minority class samples through the SMOTE algorithm to increase the number of minority class samples to be close to the number of majority class samples. Then, perform undersampling on the majority class samples and randomly remove some samples from them to make the number of majority class samples close to the increased number of minority class samples. Finally, integrate the minority class samples generated by SMOTE with the undersampled majority class samples to form a more balanced dataset. This method not only retains the characteristics of the minority class samples but also reduces data bias, helping to improve the classification performance of the model.
[0098] Step 5 specifically includes the following content:
[0099] Step 5.1: Construct a Convolutional Neural Network (CNN);
[0100] The Convolutional Neural Network adopts the methods of local connection and weight sharing. By alternately using convolutional layers and pooling layers, it extracts features from the vibration signal data of industrial equipment, then learns local spatial patterns, obtains an effective low-dimensional feature representation, and provides an input for the Gated Recurrent Unit (GRU).
[0101] Step 5.2: Construct a Gated Recurrent Unit Network (GRU);
[0102] The GRU network adds an update gate and a reset gate on the basis of the Recurrent Neural Network, and controls the information state of the GRU network at each moment through these two gates;
[0103] The GRU network has three inputs at time t:
[0104] The input value x t of the network at the current moment, that is, the sequence data after feature extraction;
[0105] The output value h of the hidden layer of the GRU network at the previous momentt-1 ;
[0106] The hidden state h at the previous moment t-1 ;
[0107] The GRU network has two outputs at time t: the output value h of the hidden layer at the current moment t and the updated hidden state h t ;
[0108] The update gate determines the amount of information of the hidden state h at the previous moment t-1 retained in the current hidden state h t ;
[0109] The reset gate determines the degree of influence of the input x of the network at the current moment t on the hidden state h at the previous moment t-1 ;
[0110] The output gate (usually merged with the update gate in GRU) determines the amount of information of the current hidden state h t as the output h of the hidden layer at the current moment t ;
[0111] Further illustrate that the update gate, reset gate, candidate hidden state, and final hidden state are gated through the following formulas:
[0112] z t = σ(W z [h t-1 , x t + b z );
[0113] r t = σ(W r [h t-1 , x t + b r );
[0114]
[0115] z t , r t : are the update gate and the reset gate respectively;
[0116] b z , b r , b h : are the bias terms of the corresponding units respectively;
[0117] W z , W r , W h : are the weight matrices of the corresponding units respectively;
[0118] ⊙: represents element-wise multiplication;
[0119] σ represents the sigmoid activation function;
[0120] tanh represents the hyperbolic tangent activation function;
[0121] Step 5.3: Combining with a Convolutional Neural Network (CNN) for feature extraction and then using a Gated Recurrent Unit (GRU) to process time series information can significantly improve the recognition accuracy of time series faults. Through this multi-level and multi-model fusion strategy, complex wind turbine fault features can be effectively analyzed, thus diagnosing and predicting fault patterns more accurately.
[0122] Step 6: In the data generation model training stage, we first reasonably divide the vibration data preprocessed and enhanced in Step 3 into a training set and a test set. The training set is used for model learning and parameter adjustment, while the test set is used to evaluate the generalization ability of the model. Then, we input the training set data into the CNN-GRU model and carefully adjust the model's parameters during training, including the learning rate, batch size, number of iterations, regularization coefficient, etc., to optimize the model performance and prevent overfitting. By adjusting these parameters, the CNN-GRU model can more effectively extract the temporal and spatial features in the data, thereby improving the prediction accuracy and the generalization ability of the model. By monitoring the loss function and evaluation metrics during training, such as accuracy and recall rate, we continuously iterate the model training until a satisfactory diagnostic effect is achieved. In addition, we may also apply early stopping and hyperparameter optimization techniques, such as grid search or random search, to find the best model configuration. Finally, we save the trained model and prepare for deployment to perform efficient wind turbine fault diagnosis in practical applications;
[0123] Step 7: After the model training is completed, we use the test set to evaluate the model and focus on the model's performance in practical applications. The evaluation metrics may include accuracy, recall rate, F1 score, etc., which can comprehensively reflect the diagnostic effect of the model. To verify the effectiveness of the improved SMOTE algorithm, we need to compare the fault diagnosis results before and after balancing the training set. Through this comprehensive evaluation and comparative analysis, we can verify whether the improved SMOTE algorithm effectively improves the accuracy and reliability of wind turbine fault diagnosis.
[0124] The present invention proposes a method for diagnosing fan faults in the case of data imbalance. This method first collects the vibration signals of the fan bearing through a vibration acceleration sensor and converts them into electrical signals, cleans and performs fast Fourier transform on the data to optimize the data quality, uses an improved SMOTE algorithm to oversample the minority fault samples to balance the class distribution of the dataset, and uses a CNN-GRU model for fault diagnosis, verifying the effectiveness of the improved SMOTE algorithm. This method significantly improves the accuracy and efficiency of fan fault diagnosis and effectively solves the problem of sample imbalance encountered by industrial fans during operation.
[0125] Finally, the method of the present application is only a preferred implementation and is not used to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.
Claims
1. A method for diagnosing fan faults under data imbalance, characterized in that: The following steps are involved: Step 1: Install the vibration acceleration sensor on the fan bearing and convert the collected vibration signal into an electrical signal through a professional data acquisition card; Step 2: Perform data cleaning and fast Fourier transform on the collected data to facilitate subsequent analysis; Step 3: Divide the data set into majority class sample sets and minority class sample sets according to the number of samples, where the minority class samples are fault samples; Step 4: Use the improved SMOTE algorithm to generate new samples in the minority class sample set of the training set X, and combine the undersampling method of the majority class to balance the training set, that is, to obtain a new training set Xnew; Step 5: Build a fault diagnosis model for fault diagnosis, and select the CNN-GRU combined model as the fault diagnosis model; Step 6: Data generation model training, input the preprocessed and enhanced data in step 3 into the two models for training; Step 7: Use the test set to verify the model diagnosis effect, and finally compare the fault diagnosis results before and after the balanced training set to verify the effectiveness of the improved SMOTE algorithm.
2. A method for diagnosing fan faults in the case of data imbalance as claimed in claim 1, characterized in that: The vibration acceleration sensor in step 1 is a piezoelectric acceleration sensor, which obtains vibration signals and stores data after noise filtering, filtering and amplification.
3. A method for diagnosing fan faults in the case of data imbalance as claimed in claim 1, characterized in that: The step 2 specifically includes the following steps: Step 2.1: Data preprocessing First, clean the data and remove noise samples; Step 2.1.1: Use linear interpolation to process missing values. For sequence data, the weighted average of adjacent values can be used to fill missing values. The specific formula is as follows: Among them, X missing is a missing value, t represents the current time point, and t-1 and t+1 represent the time points before and after the missing value; Step 2.1.2: Use the Z-score outlier detection method to detect and process outliers in the data. By quantifying the degree of deviation of the data point from the mean, we can identify outliers that are significantly different from the majority of the data. The specific formula is as follows: Where X: represents the current observation value, μ represents the mean of the feature column, σ represents the standard deviation of the feature column, which is used to measure the degree of dispersion of the data, and Z represents the Z-score value, which measures the degree of deviation of the current observation value X from the mean μ. When |Z|>3, the observation value is regarded as an outlier. Step 2.1.3: Apply min-max standardization to standardize the data and scale the numerical range of all features to a specified interval. The specific formula is as follows: Where X represents the original eigenvalue, min(X) represents the minimum value of the feature column X, max(X) represents the maximum value of the feature column X, and X' represents the standardized eigenvalue, which ranges from [0,1]. Step 2.2: Perform fast Fourier transform on the standardized vibration signal X' to obtain its frequency domain representation X freq [k], the specific formula is: Wherein, j is the imaginary unit, k is the index of the frequency domain, k = 0, 1, 2, ..., N-1.
4. A method for diagnosing fan faults in the case of data imbalance as claimed in claim 1, characterized in that: Step 4 specifically includes: Step 4.1: For the fault sample set, use the k-means++ algorithm to cluster it into m clusters, select the center point of each cluster as the oversampling point, and for each sample x i , use the Gaussian kernel function to calculate its local density ρ i : d ij Represents the sample point x i With sample point x j The distance between c represents the cutoff distance; Step 4.2: Oversample the minority class samples at the optimal sampling rate to obtain the center point of each cluster and add it to the training set as the most representative sample point; Step 4.3, use K nearest neighbor to filter out the noise of majority class samples, and determine the k1 nearest neighbor samples of each majority class sample through the k nearest neighbor algorithm; for each majority class sample, when all its k1 nearest neighbor samples belong to the minority class, the majority class sample is judged as a noise sample and is removed from the data set. The K nearest neighbor algorithm uses the commonly used distance measurement method - Euclidean distance, and its specific formula is: In the Euclidean distance formula, x i and x j Represents the coordinate vector of two points in the feature space; specifically: is a point in the feature space, where n is the dimension of the feature space, It is point x i Coordinate value in the lth dimension; is another point in the feature space, which is similar to x i Similarly, it also has n coordinate values, It is point x j Coordinate value in the lth dimension; In step 4.4, new synthetic samples are generated between minority class samples through the SMOTE algorithm, the number of minority class samples is increased to be close to the number of majority class samples, the majority class samples are undersampled, and some samples are randomly removed from them to make the number of majority class samples close to the number of minority class samples after the increase, and the minority class samples generated by SMOTE are integrated with the undersampled majority class samples to form a more balanced data set.
5. A method for diagnosing fan faults in the case of data imbalance as claimed in claim 4, characterized in that: Step 5 specifically includes: Step 5.1: Build a convolutional neural network (CNN); The convolutional neural network uses local connections and weight sharing, and alternately uses convolutional layers and pooling layers to extract features from the vibration signal data of industrial equipment, and then learns local spatial patterns to obtain effective low-dimensional feature representations. Step 5.2: Construct a gated recurrent unit network (GRU); The GRU network adds an update gate and a reset gate on the basis of the recurrent neural network, and controls the information state of the GRU network at each moment through these two gates; The GRU network has 3 inputs at time t: The input value x of the network at the current moment t , that is, the sequence data after feature extraction; The output value h of the hidden layer of the GRU network at the previous moment t-1 ; The implicit state h at the previous moment t-1 ; The GRU network has two outputs at time t: the output value h of the hidden layer at the current time t and the updated hidden state h t ; The update gate determines the implicit state h of the previous moment t-1 Keep to the current hidden state h t The amount of information; The reset gate determines the input x of the network at the current moment t For the previous implicit state h t-1 the extent of the impact; The output gate (usually combined with the update gate in GRU) determines the current hidden state h t As the hidden layer output h at the current moment t The amount of information. Step 5.3: Combine convolutional neural network (CNN) for feature extraction and then use gated recurrent unit (GRU) to process time series information.