Power equipment fault prediction method based on support vector regression

Abstract

The invention discloses a power equipment fault prediction method based on support vector regression, and aims to improve the prediction accuracy. The method comprises the following steps: firstly, collecting multi-dimensional operation parameter data of power equipment, and constructing an original monitoring data set; a data set is segmented through a sliding time window, wavelet transform is used for noise reduction, and then a normalization method is applied to generate a standardized data set. Time domain and frequency domain features are further extracted, a feature vector space is constructed, an optimal feature subset is screened through a mutual information method, and a training data set is generated; performing preliminary training on a support vector regression model by using the training data set to generate candidate support vectors, and screening a support vector set with the maximum contribution through an optimization algorithm; a grid search method is adopted to adjust kernel function parameters and penalty factors, and a support vector set is further optimized. And evaluating the performance of the model through cross validation, applying the optimized support vector regression model to power equipment fault prediction, and generating a fault occurrence probability prediction value.

Description

Technical Field

[0001] The present invention relates to the technical field of fault prediction, and in particular to a method for predicting faults of electric power equipment based on support vector regression. Background Art

[0002] In the prior art, power equipment fault prediction is one of the important means to ensure the safe and reliable operation of the power system. Commonly used methods include statistical analysis based on historical data, rule-based reasoning based on expert knowledge, and machine learning algorithms that have developed rapidly in recent years. These methods collect and analyze equipment operation data to identify potential failure modes and risk signals, thereby realizing the prediction and early warning of equipment failures. In particular, machine learning models such as support vector machines (SVM) and support vector regression (SVR) have been gradually applied to power equipment fault prediction due to their ability to process high-dimensional data.

[0003] However, there are still some major problems in the existing technology. First, when faced with complex and multi-dimensional power equipment data, traditional fault prediction methods often find it difficult to extract effective feature information and are easily interfered by noise, thus affecting the accuracy of prediction. In addition, support vector machines and support vector regression models rely more on manual adjustments in parameter optimization and support vector screening, and lack a systematic optimization method, resulting in high model complexity and unstable prediction performance. Especially when dealing with different load conditions and environmental influences, the generalization ability and adaptability of the model are difficult to guarantee.

[0004] In view of the above problems, the present invention proposes a method for predicting power equipment faults based on support vector regression. Summary of the invention

[0005] The present application provides a method for predicting power equipment faults based on support vector regression to improve the accuracy of prediction.

[0006] The present application provides a method for predicting power equipment faults based on support vector regression, comprising:

[0007] Collect multi-dimensional operating parameter data of power equipment to construct an original monitoring data set, wherein the multi-dimensional operating parameters include equipment temperature, vibration frequency, load current, operating voltage and ambient humidity;

[0008] The original monitoring data set is segmented according to a preset sliding time window, and wavelet transform is applied to each segmented data for noise reduction to suppress interference signals; the noise-reduced data is standardized using a maximum-minimum normalization method to obtain a standardized data set;

[0009] Extracting temporal features and temporal features from the standardized data set to construct a comprehensive feature vector space; analyzing the correlation of each feature in the comprehensive feature vector space by mutual information method, screening features with rich information and low redundancy, and forming an optimal feature subset;

[0010] Constructing a training data set using the optimal feature subset; training a support vector regression model using the training data set to obtain a preliminarily trained support vector regression model; wherein the preliminarily trained support vector regression model includes a set of candidate support vectors for describing the regression boundary of the model;

[0011] Using an optimization algorithm to screen the candidate support vectors, select the support vectors that contribute most to the regression effect, and generate a preliminary optimal support vector set;

[0012] The kernel function parameters and penalty factors of the initially trained support vector regression model are adjusted by combining the grid search method to ensure that the model achieves the best balance between accuracy and complexity, and the initially selected support vector set is further searched and optimized to screen out the final support vector set that can best describe the data features, thus obtaining an adjusted and optimized support vector regression model.

[0013] The performance of the adjusted and optimized support vector regression model is evaluated by cross-validation method to verify the accuracy and stability of the model on the training data set and the test data set, so as to obtain the final optimized support vector regression model;

[0014] The real-time collected operating parameters of the electric power equipment are input into the finally optimized support vector regression model to generate a prediction value of the probability of failure.

[0015] Furthermore, the collecting of multi-dimensional operating parameter data of the power equipment and the construction of the original monitoring data set include:

[0016] Multi-dimensional operating parameter data of power equipment, including equipment temperature, vibration frequency, load current, operating voltage and ambient humidity, are synchronously collected through multiple independent sensors. The data collection time interval and sampling frequency of each sensor are dynamically adjusted according to the parameter characteristics to ensure synchronization and timing consistency of each parameter.

[0017] Furthermore, extracting temporal features and temporal features from the standardized data set to construct a comprehensive feature vector space includes:

[0018] Extracting time domain features from the standardized data set segment by segment to describe the basic statistical characteristics of the power equipment operating parameters, wherein the time domain features include mean, standard deviation, peak, skewness and kurtosis coefficient;

[0019] Performing frequency domain analysis on each segment of standardized data to extract frequency domain features to reveal the frequency components and energy distribution of the signal, wherein the frequency domain features include power spectrum density, frequency center, bandwidth, and frequency peak;

[0020] The extracted time domain features and frequency domain features are combined into a multi-dimensional feature vector to construct a comprehensive feature vector space.

[0021] Furthermore, the mutual information method is used to analyze the correlation of each feature in the comprehensive feature vector space, screen features with rich information and low redundancy, and form an optimal feature subset, including:

[0022] For each feature in the comprehensive feature vector space, the mutual information value is calculated with the target variable to quantify the degree of information sharing between each feature and the target variable; wherein the target variable includes the probability of failure and the health status of the equipment;

[0023] Based on the results of mutual information analysis, features that contribute more to the target variable information and have low redundancy are preferentially selected, and the optimal feature subset is gradually constructed.

[0024] Furthermore, constructing a training data set using the optimal feature subset includes:

[0025] Based on the formed optimal feature subset, add a time series label to each data point to ensure the consistency of each data point in the time dimension and avoid time series misalignment between features;

[0026] The optimal feature data after time alignment are combined into a training data set, where each sample data includes each feature data in the optimal feature subset and its corresponding target variable label.

[0027] Furthermore, the using the training data set to train the support vector regression model to obtain a preliminarily trained support vector regression model includes:

[0028] Input the constructed training data set into the support vector regression model, and perform preliminary training using the relationship between the features in the training data set and the target variable;

[0029] In the support vector regression model after preliminary training, a set of candidate support vectors is identified from the training data through the support vector selection mechanism within the model, wherein the set of candidate support vectors is used to define the boundary of the regression model and represent the key data points of the model in the feature space.

[0030] Furthermore, an optimization algorithm is used to screen the candidate support vectors, select the support vectors that contribute most to the regression effect, and generate a preliminary preferred support vector set, including:

[0031] For the candidate support vectors in the initially trained support vector regression model, the contribution of each candidate support vector to the model regression effect is calculated one by one, specifically including: removing each candidate support vector from the model one by one, and calculating the error increment of the model on the validation set before and after removing the support vector; wherein the error increment is used to evaluate the contribution of the support vector to the model regression effect, and the larger the error increment, the higher the contribution of the support vector to the model regression effect; sorting the candidate support vectors according to the size of the error increment, and selecting the support vector with the highest contribution as the object of further optimization;

[0032] For the initially screened high-contribution support vector set, a gradient descent-based optimization algorithm is applied for fine screening, including: in the initial state, all high-contribution support vectors are used as the current support vector set; in each iteration, the gradient of the current support vector set to the model prediction error is calculated, the weight of the support vector is gradually adjusted, and a penalty term is imposed on the support vector with lower contribution to further reduce its contribution to the model regression effect; when the weight of the specified support vector drops below the preset threshold, it is removed from the support vector set; and the iteration is continued until the change in the model error is lower than the preset convergence standard.

[0033] Furthermore, the performance of the adjusted and optimized support vector regression model is evaluated by cross-validation method to verify the accuracy and stability of the model on the training data set and the test data set to obtain the final optimized support vector regression model, including:

[0034] The adjusted and optimized support vector regression model is applied to the cross-validation method, and the training data set is divided into several subsets to train and validate the model in a grouped manner;

[0035] Summarize the results of each round of cross-validation, calculate the prediction error of the model on each validation set, and take the average error on all validation sets as the overall performance indicator of the model;

[0036] If both the mean error and variance meet the preset accuracy and stability standards, the current support vector regression model is determined as the final optimization model; if not, return to adjust the model parameters or optimize the support vector set, and repeat the cross-validation steps until the final optimization model that meets the requirements is obtained.

[0037] The beneficial effects of the technical solution provided by this application include:

[0038] (1) By segmenting and denoising the multidimensional operating parameter data, combined with maximum-minimum normalization, the noise interference is significantly suppressed, thereby improving the accuracy of data processing and laying a high-quality data foundation for subsequent feature extraction and fault prediction. (2) The optimal feature subset with rich information and low redundancy is selected through the mutual information method, which effectively reduces the data dimension of model training and improves the training efficiency of the support vector regression model. At the same time, it ensures that the model can make full use of key features, thereby improving the accuracy of fault prediction. (3) The grid search method is used to optimize the kernel function parameters and penalty factors of the support vector regression model, so that the model achieves the best balance between accuracy and complexity. Further, the optimization algorithm is used to select the support vector set with the largest contribution, reduce the model complexity, and improve the generalization ability and prediction stability of the model. (4) The performance of the optimized support vector regression model is evaluated and verified by the cross-validation method to ensure the accuracy and stability of the model on different data sets. Finally, a fault prediction model with high generalization ability is obtained, which effectively improves the reliability of power equipment fault prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 Flowchart of a method for predicting power equipment faults based on support vector regression. DETAILED DESCRIPTION

[0040] Many specific details are described in the following description to facilitate a full understanding of the present application. However, the present application can be implemented in many other ways than those described herein, and those skilled in the art can make similar generalizations without violating the connotation of the present application, so the present application is not limited by the specific implementation disclosed below.

[0041] The first embodiment of the present application provides a method for predicting power equipment failure based on support vector regression. Figure 1 , which is a schematic diagram of the first embodiment of the present application. Figure 1 A first embodiment of the present application provides a method for predicting power equipment faults based on support vector regression and is described in detail.

[0042] Step S101: Collect multi-dimensional operating parameter data of power equipment to construct an original monitoring data set, wherein the multi-dimensional operating parameters include equipment temperature, vibration frequency, load current, operating voltage and ambient humidity.

[0043] In order to implement the operation of step S101, it is first necessary to establish a sensor monitoring system for real-time collection of various operating parameters of the power equipment. The system includes temperature sensors, vibration sensors, current sensors, voltage sensors, humidity sensors, etc. Each sensor must be installed in the key parts of the equipment so that reliable and accurate operating data can be obtained. The specific installation location can be optimized according to the equipment type and operating environment. For example, the temperature sensor can be installed in the area with higher heat generation, and the vibration sensor can be installed on the main load-bearing structure of the equipment.

[0044] During the data collection process, the temperature, vibration frequency, load current, operating voltage, and ambient humidity of the equipment should be collected according to the set sampling frequency and time interval to ensure that these data can reflect the actual situation of the equipment under different operating conditions. The sampling frequency can be set based on the operating speed and fault response time of the equipment. For example, for high-speed rotating equipment, it is recommended that the sampling frequency be no less than 1kHz to capture small vibration changes.

[0045] The data collected by each sensor is centrally transmitted to the central processing unit through the data acquisition module for preliminary data storage and organization. The data storage format can be selected in a common time series data format to facilitate subsequent data processing and analysis. In order to ensure the integrity and continuity of the data, a real-time data verification and redundancy mechanism can be set to ensure that no data loss or error occurs during the data transmission process.

[0046] When all sensor data collection is completed, the output data of different sensors are integrated to form an original monitoring data set containing multi-dimensional operating parameters. When constructing the original monitoring data set, the data of each sensor should be synchronized to ensure that all data has a unified timestamp so that parameters of different dimensions can reflect the equipment status at the same time point.

[0047] Furthermore, the collecting of multi-dimensional operating parameter data of the power equipment and the construction of the original monitoring data set include:

[0048] Multi-dimensional operating parameter data of power equipment, including equipment temperature, vibration frequency, load current, operating voltage and ambient humidity, are synchronously collected through multiple independent sensors. The data collection time interval and sampling frequency of each sensor are dynamically adjusted according to the parameter characteristics to ensure synchronization and timing consistency of each parameter.

[0049] First, a multi-sensor synchronous data acquisition system needs to be built to collect multi-dimensional parameters during the operation of power equipment. The system includes multiple independent sensors, each of which collects specific operating parameters such as temperature, vibration frequency, load current, operating voltage, and ambient humidity. These sensors need to be installed in key parts of the equipment to ensure that the collected data accurately reflects the actual operating status of the equipment. Each sensor should have a sampling frequency and collection interval suitable for its parameter characteristics. For example, the temperature sensor can be set to a relatively long collection interval, while the vibration sensor requires a higher sampling frequency to capture rapid vibration changes.

[0050] In actual implementation, data synchronization is particularly important. In order to achieve timing consistency between different sensors, the acquisition system needs to have a synchronization control module to ensure that all sensors collect data at the same time point. The synchronization control module can coordinate the sampling operations of each sensor through distributed clock synchronization or global clock management to ensure data consistency. In particular, the sampling frequency and acquisition interval of the sensor can be dynamically adjusted according to parameter changes. For example, when the equipment load changes greatly, the sampling frequency of the load current sensor can be automatically increased to capture more detailed data.

[0051] After the collection is completed, the data of each sensor will be centrally transmitted to the central processing unit, where preliminary data sorting and time stamp synchronization will be performed. The sorted data constitutes a complete original monitoring data set, including multi-dimensional equipment operation parameters. This data set can not only reflect the status of the equipment at each time point, but also ensure the time sequence consistency between the parameters, providing a reliable data basis for subsequent data processing and fault prediction.

[0052] Step S102: The original monitoring data set is segmented according to a preset sliding time window, and wavelet transform is applied to each segmented data for noise reduction to suppress interference signals; the noise-reduced data is standardized using a maximum-minimum normalization method to obtain a standardized data set.

[0053] In order to implement step S102, the original monitoring data set needs to be preprocessed to improve the data quality and eliminate interference. First, the length and step size of the sliding time window are set to divide the original monitoring data set into multiple overlapping or non-overlapping time periods. The length of the sliding time window should be set according to the operating characteristics of the equipment and the fault prediction requirements, such as the operating cycle of the equipment or the characteristics of the signal change. The choice of step size can also be adjusted according to the characteristics of the data, and should usually be less than the window length to ensure that the data has a certain continuity between different windows, so as to capture dynamic changes.

[0054] After the time window segmentation is completed, wavelet transform is applied to each segment of data for noise reduction. Wavelet transform is an effective time-frequency analysis method that can decompose the signal into different frequency components. First, select the appropriate wavelet basis function (such as Haar wavelet or Daubechies wavelet) and the number of decomposition layers. The choice of wavelet basis function can be determined according to the frequency characteristics of the signal and the noise reduction requirements, while the number of decomposition layers needs to balance the noise reduction effect and signal fidelity. In the noise reduction process of each segmented data, the wavelet transform decomposes the original signal into a low-frequency approximation part and a high-frequency detail part, where the high-frequency part usually contains interference signals. In the noise reduction process, threshold processing can be applied to the coefficients of the high-frequency part to eliminate the noise component and retain the main signal characteristics of the low-frequency part. Finally, the signal is restored by inverse wavelet transform to obtain the noise-reduced data.

[0055] After the noise reduction process is completed, in order to further standardize the scale of different segmented data, the maximum-minimum normalization method is used to standardize the noise-reduced data. The purpose of normalization is to map the numerical range of each segmented data to a fixed interval (such as 0 to 1), thereby reducing the impact of different feature value ranges on the model. In actual operation, the maximum and minimum values ​​in each segmented data are first obtained, and then each data point in the segment is scaled according to the formula so that all data points are within the set interval. The standardized data retains the signal characteristics after noise reduction and is convenient for subsequent feature extraction and model training.

[0056] Through the above steps, the standardized data set obtained is the basic data for subsequent analysis and model input, ensuring the quality and consistency of the data.

[0057] Furthermore, the original monitoring data set is segmented according to a preset sliding time window, and wavelet transform is applied to each segmented data for noise reduction to suppress interference signals, including:

[0058] Based on the time series of the original monitoring data set X = {x1, x2, ..., x n}, define the sliding time window T and step size s, and use the following formula (1) to determine the optimal segment length T * :

[0059]

[0060] Among them, N T is the number of data points in the sliding time window T; x i is the i-th data point in the time series X; x i+1 is the i+1th data point in the time series X; M(T) represents the mean of the data points in the sliding time window T; σ(T) represents the standard deviation of the data points in the sliding time window T;

[0061] For each optimal segment T * The data in the image are transformed by Haar wavelet transform to decompose each segmented signal into a low-frequency approximation part and a high-frequency detail part, and the segmented signal is obtained as shown in the following formula (2):

[0062]

[0063] Among them, J is the number of layers of wavelet decomposition, indicating the scale of wavelet decomposition; K j is the number of detail coefficients decomposed at the jth layer; d j,k is the j-th layer detail coefficient, representing the high-frequency signal part; a J is the approximate coefficient of the Jth layer, representing the low-frequency signal part; ψ j,k (t) is the wavelet basis function of the detail part of the jth layer, through ψ j,k (t) = 2 j / 2 ψ(2 j tk) generates; φ J (t) is the scaling function of the Jth layer, through φ J (t) = 2 J / 2 φ(2 J tk) is generated; where the function ψ(t) is the basis function of the Haar wavelet, defined according to the following formula (3):

[0064]

[0065] The function φ(t) is the scaling function of the Haar wavelet and is defined according to the following formula (4):

[0066]

[0067] According to the following formula (5), the adaptive interval value λ is calculated T* :

[0068]

[0069] Among them, α is the experience adjustment factor; N T* For window T * The number of data points in σ T* is the window T * The standard deviation of the data points within ;

[0070] According to the following formula (6), for the detail coefficient d j,k Perform threshold screening to generate a new detail coefficient d j ′ ,k ;:

[0071]

[0072] The new detail coefficient d j ′ ,k Substitute into formula (2) and replace the original detail coefficient d j,k , get the noise-reduced segmented signal And the noise-reduced segmented signal as the denoised data.

[0073] First, we need to analyze the time series X of the original monitoring data set = {x1, x2, …, x n} for processing. In this process, a sliding time window T and a step size s are set to segment the entire time series. The sliding time window is used to segment the time series into multiple continuous subsequences, so that noise reduction operations can be applied to each subsequence. * When , the formula (1) is used to calculate to find the segment that makes the signal feature clearest under a specific window.

[0074] Formula (1) is used to determine the optimal segment length T * :

[0075]

[0076] Among them, N T Represents the number of data points in the sliding time window T;

[0077] x i is the i-th data point in the time series X,

[0078] x i+1 is the next adjacent data point in the sequence.

[0079] The mean M(T) represents the average value of the data points in window T, which is used to measure the overall level of the data;

[0080] The standard deviation σ(T) represents the volatility of the data points within the window T and is used to evaluate the degree of dispersion of the data. The purpose of the formula is to obtain the optimal time window length T by maximizing the ratio of the signal change to its volatility. * .

[0081] After determining the optimal segment T * After that, Haar wavelet transform is applied to each segmented data to decompose the segmented signal into a low-frequency approximate part and a high-frequency detail part, which is expressed as the segmented signal in formula (2):

[0082]

[0083] In this formula, J represents the number of layers of wavelet decomposition, which defines the scale of the wavelet;

[0084] K j is the number of detail coefficients decomposed at the jth layer;

[0085] Detail coefficient d j,k Represents the high-frequency part of the signal at the jth layer, corresponding to the detailed information of the wavelet transform;

[0086] Approximation coefficient a J represents the low-frequency part, that is, the overall outline of the signal.

[0087] Wavelet basis function ψ j,k (t) is used to generate the high-frequency details of each layer, defined as ψ j,k (t) = 2 j / 2 ψ(2 j tk), and the scaling function φ J (t) is used for the low-frequency part and is defined as φ J (t) = 2 J / 2 φ(2 J tk).

[0088] For the wavelet basis function ψ(t) and the scale function φ(t), the Haar wavelet can be used to define its form, where:

[0089]

[0090] In order to remove high-frequency noise signals, the adaptive interval value is calculated according to formula (5)

[0091]

[0092] Among them, α is the empirical adjustment factor used to control the strictness of the interval value;

[0093] N (T*) For window T * The number of data points in ;

[0094] σ (T*) Represents the standard deviation of the data points within the window, which is used to measure the volatility of the data.

[0095] Finally, the detail coefficient d is calculated using formula (6) j,k Perform intermediate value screening to generate new detail coefficient d j ′ ,k , to complete the noise reduction process:

[0096]

[0097] The updated detail coefficient dj ′ ,k Substitute into formula (2) and recalculate the noise-reduced segmented signal Finally, the noise-reduced segmented signal It can be used as input for subsequent data processing to improve signal purity and prediction accuracy.

[0098] Step S103: extracting time domain features and temporal features from the standardized data set to construct a comprehensive feature vector space; analyzing the correlation of each feature in the comprehensive feature vector space by mutual information method, screening features with rich information and low redundancy, and forming an optimal feature subset.

[0099] In order to implement step S103, it is necessary to extract features from the standardized data set to construct a comprehensive feature vector space containing time domain and frequency domain information. First, the time domain features are extracted for each segmented data. The time domain feature extraction step includes calculating basic statistics such as the mean, standard deviation, kurtosis and skewness of the data. The mean can reflect the overall level of the segmented data, the standard deviation indicates the degree of fluctuation of the data, and the kurtosis and skewness can reflect the distribution form of the data, such as whether it presents a spike or deviates from symmetry. By extracting these time domain features, the operating status of the power equipment can be characterized from different angles.

[0100] After the time domain features are extracted, the frequency domain features of the data are then extracted. To do this, each data segment is first converted to the frequency domain, which can generally be achieved through a fast Fourier transform (FFT). This step decomposes the original time domain data into different frequency components to reveal the periodicity and frequency characteristics of the signal. In the frequency domain, the power spectral density and frequency center of each segment can be calculated. The power spectral density is used to measure the energy distribution of the data on each frequency component, while the frequency center represents the dominant frequency of the signal. These frequency domain features can provide the frequency characteristics of equipment vibration or current fluctuations, further revealing potential failure modes of the equipment.

[0101] After the time domain and frequency domain feature extraction is completed, these features are combined to form a comprehensive feature vector space. Each data point in this space corresponds to a segmented data feature, which contains multi-dimensional time domain and frequency domain information. At this time, the dimension of the feature vector space is high and may contain some redundant or low-information features, so further optimization is required. The mutual information method is used to analyze each feature in the feature vector space and calculate the correlation between each feature and the target variable. The mutual information method can identify features with a high degree of information sharing with the target variable, while eliminating redundant features that are highly correlated with other features, thereby forming an optimal feature subset with rich information and low redundancy.

[0102] Finally, through the above steps, an optimized feature vector space is obtained, which contains high-information and non-redundant features, and will provide high-quality input data for subsequent model training.

[0103] Furthermore, extracting temporal features and temporal features from the standardized data set to construct a comprehensive feature vector space includes:

[0104] Extracting time domain features from the standardized data set segment by segment to describe the basic statistical characteristics of the power equipment operating parameters, wherein the time domain features include mean, standard deviation, peak, skewness and kurtosis coefficient;

[0105] Performing frequency domain analysis on each segment of standardized data to extract frequency domain features to reveal the frequency components and energy distribution of the signal, wherein the frequency domain features include power spectrum density, frequency center, bandwidth, and frequency peak;

[0106] The extracted time domain features and frequency domain features are combined into a multi-dimensional feature vector to construct a comprehensive feature vector space.

[0107] First, the standardized data is processed in segments, and each segment of data represents the operating status of the power equipment within a specific time window. For each segment of data, time domain features are extracted to obtain basic statistical information of the operating parameters. Time domain features include mean, standard deviation, peak, skewness and kurtosis coefficient, each of which describes different statistical characteristics of the data. The mean can reflect the overall level of the data in this segment, the standard deviation indicates the amplitude of data fluctuation, the peak indicates the maximum value of the data in this segment, and the skewness and kurtosis coefficients provide information on the distribution form of the data, describing the symmetry and sharpness of the data distribution respectively.

[0108] After the time domain features are extracted, each segment of data needs to be analyzed in the frequency domain to reveal the frequency characteristics of the signal. Frequency domain analysis usually uses methods such as fast Fourier transform (FFT) or wavelet transform to convert time series data into the frequency domain. For the extraction of frequency domain features, the main focus is on indicators such as power spectrum density, frequency center, bandwidth, and frequency peak. Power spectrum density is used to measure the energy distribution of different frequency components and reflects the strength of the signal at different frequencies. The frequency center indicates the location of the main frequency component of the signal, while the bandwidth describes the range of the signal frequency distribution, and the frequency peak is used to mark the frequency point with the highest energy. These frequency domain features can reveal the changes in periodic or specific frequency components in the operating state of power equipment, providing more in-depth signal characteristics for fault prediction.

[0109] Finally, all the extracted time domain and frequency domain features are combined to form a multidimensional feature vector, each of which represents the comprehensive status information of the equipment operation in a specific time period. All these feature vectors together constitute a comprehensive feature vector space, in which each data point contains the time domain and frequency domain information of the equipment in that time period. This comprehensive feature vector space provides high-quality input for the subsequent support vector regression model training, enabling the model to make full use of multidimensional feature data for fault prediction.

[0110] Furthermore, the mutual information method is used to analyze the correlation of each feature in the comprehensive feature vector space, screen features with rich information and low redundancy, and form an optimal feature subset, including:

[0111] For each feature in the comprehensive feature vector space, the mutual information value is calculated with the target variable to quantify the degree of information sharing between each feature and the target variable; wherein the target variable includes the probability of failure and the health status of the equipment;

[0112] Based on the results of mutual information analysis, features that contribute more to the target variable information and have low redundancy are preferentially selected, and the optimal feature subset is gradually constructed.

[0113] First, each feature in the comprehensive feature vector space is analyzed separately, and its mutual information value is calculated with the target variable. The mutual information value is a measure of the degree of information sharing between two variables, which can represent the correlation between the feature and the target variable. The target variable usually includes the probability of failure and the health status of the equipment. The former is used to estimate the possibility of equipment failure, and the latter is used to evaluate whether the equipment is operating normally.

[0114] When calculating the mutual information value between each feature and the target variable, the mutual information formula in information theory can be used to compare the distribution of the feature with the distribution of the target variable to quantify the correlation between them. The higher the mutual information value, the greater the degree of information sharing between the feature and the target variable, that is, the feature has a higher information contribution value to the target variable. In this way, through the mutual information value, it is possible to identify which features have greater information value for equipment fault prediction.

[0115] Next, based on the results of mutual information analysis, the features that contribute more information to the target variable are selected first. At the same time, considering the redundancy of the data, when constructing the optimal feature subset, it is also necessary to evaluate the correlation between each feature and other features to ensure that the selected features have low redundancy. By gradually screening and excluding features with high redundancy, an optimal feature subset with rich information and low redundancy is finally obtained.

[0116] This optimal feature subset can provide high-quality input data for the subsequent support vector regression model, enabling the model to focus on features directly related to fault prediction, effectively improving the accuracy and efficiency of prediction.

[0117] Step S104: construct a training data set using the optimal feature subset; train a support vector regression model using the training data set to obtain a preliminarily trained support vector regression model; wherein the preliminarily trained support vector regression model includes a set of candidate support vectors for describing the regression boundary of the model.

[0118] In order to implement step S104, it is first necessary to construct a training data set using the optimal feature subset. The process begins by classifying each record containing the optimal feature and ensuring that the time point corresponding to each record matches the target fault state information. These records are high-quality data after feature screening and can effectively represent the operating status of the power equipment.

[0119] Next, these records are arranged in chronological order to form time series training samples. When preparing the training data set, ensure that each sample includes feature values ​​extracted from the optimal feature subset. These samples will form a set of multidimensional data, where each dimension corresponds to a specific optimal feature, and the multidimensional feature value set of all samples is used as the input of the support vector regression model. To ensure the temporal dependence of the model on the samples, the creation of training samples needs to consider continuous data points within the time window, so as to maintain the dynamic change characteristics of the operating parameters of the power equipment during model training.

[0120] Then, the training data set is input into the support vector regression model to start preliminary training of the model. During the preliminary training process, the support vector regression model will gradually learn the characteristic patterns of the training samples and identify the relationship between the data features and the equipment failure status. During the training process, the model will automatically adjust the position of the support vector to form a preliminary regression boundary. In this process, the support vector regression model will identify a set of candidate support vectors, which are key data points that have a significant impact on the regression boundary of the model, and through them, the distribution of the model in the feature space can be described.

[0121] Finally, by performing this preliminary training on the model, a support vector regression model containing candidate support vectors is obtained. This model has basic fault prediction capabilities and provides a basis for subsequent optimization and screening.

[0122] Furthermore, constructing a training data set using the optimal feature subset includes:

[0123] Based on the formed optimal feature subset, add a time series label to each data point to ensure the consistency of each data point in the time dimension and avoid time series misalignment between features;

[0124] The optimal feature data after time alignment are combined into a training data set, where each sample data includes each feature data in the optimal feature subset and its corresponding target variable label.

[0125] First, each data point in the optimal feature subset needs to be processed to ensure its consistency in the time dimension. To this end, time series labels are first added to all data points in the optimal feature subset. These labels not only mark the time position of the data points, but also ensure that the time order of each feature is accurately synchronized. For example, during the data acquisition process, different operating parameters (such as temperature, vibration frequency, load current, etc.) may be collected at different frequencies or time intervals. At this time, time series labels can be used to align these data with different acquisition frequencies so that all features have logical consistency at the same time point. In this way, by adding time series labels to each data point, the problem of timing misalignment between features can be avoided, ensuring the continuity and consistency of the input data.

[0126] Next, for each time label, the optimal feature data at the same time point are combined to form a complete sample data. In this process, it is necessary to ensure that each sample data contains all the key features selected from the optimal feature subset, so as to provide sufficient information support for the model. In addition, it is necessary to add a target variable label to each sample data. The target variable usually represents the fault state or health state of the power equipment, such as "normal" or "fault". The introduction of this label provides a training target for the model, so that the model can learn the correlation between the optimal feature and the equipment status during training.

[0127] Finally, these time-aligned sample data containing target variable labels are combined into a training data set. During the construction process, a sliding window method can be used to generate samples to ensure that the training data has continuous time series information. For example, a certain window size and sliding step size are selected to gradually slide and generate samples on the time series, so that the dynamic changes in the operating status of the equipment can be captured. Through this sample generation method, the training data set not only retains the association between features and target variables, but also enhances the model's sensitivity to time series changes. After careful design and construction, this training data set can provide high-quality, time-series continuous and information-rich input data for the support vector regression model, providing a solid foundation for the accuracy of fault prediction.

[0128] Furthermore, the using the training data set to train the support vector regression model to obtain a preliminarily trained support vector regression model includes:

[0129] Input the constructed training data set into the support vector regression model, and perform preliminary training using the relationship between the features in the training data set and the target variable;

[0130] In the support vector regression model after preliminary training, a set of candidate support vectors is identified from the training data through the support vector selection mechanism within the model, wherein the set of candidate support vectors is used to define the boundary of the regression model and represent the key data points of the model in the feature space.

[0131] First, the constructed training data set needs to be input into the support vector regression model. The training data set contains multiple samples, each of which consists of a set of key features and target variables. These features are derived from the multidimensional data analysis of the operating status of power equipment. By inputting the training data set into the model, the support vector regression model will gradually learn the mapping relationship between each feature and the target variable. The core goal of the initial training is to enable the model to identify the relationship between the feature and the fault state and form a basic regression boundary, so as to have the initial ability to predict power equipment failures.

[0132] During the model training process, the support vector regression model will use the internal optimization algorithm to gradually adjust its parameters so that it can better fit the relationship between the features and target values ​​in the training data. In the support vector regression model, the structure of the model depends on a set of key support vectors. These support vectors are a small number of data points in the training data, which play a decisive role in the regression boundary of the model. Specifically, the model will identify these data points that play a key role in the formation of the boundary during the training process and mark them as candidate support vectors. This set of candidate support vectors not only represents the core distribution of the model in the feature space, but also determines the location and shape of the regression boundary, thereby providing a high-precision boundary in future fault prediction.

[0133] After the initial training, the set of candidate support vectors in the model has been preliminarily determined. These candidate support vectors are actually the "key anchor points" of the model in the feature space, and they can effectively represent the model's ability to identify equipment failures. This step ensures that the model not only has predictive capabilities, but also maintains a high computational efficiency due to the relatively small number of candidate support vectors. This provides a basis for subsequent further optimization and candidate support vector screening, achieving a preliminary balance between the accuracy and complexity of the model.

[0134] Step S105: using an optimization algorithm to screen the candidate support vectors, selecting the support vectors that contribute most to the regression effect, and generating a preliminary preferred support vector set.

[0135] Step S105 needs to use an optimization algorithm to screen the candidate support vectors to select the support vectors that contribute the most to the model regression effect, and then generate a preferred support vector set. First, in the support vector regression model that has been preliminarily trained, the candidate support vector set contains a large number of data points, which have different degrees of influence on the regression boundary of the model. Therefore, each candidate support vector needs to be analyzed to determine its contribution to the model prediction accuracy.

[0136] During the screening process, an iterative optimization algorithm, such as stepwise elimination or stepwise addition, can be used to evaluate the contribution of each candidate support vector. Specifically, for each candidate support vector, after temporarily removing the support vector, the change in the prediction error of the model on the validation data set is calculated. If the prediction error of the model increases significantly after removing the support vector, it means that the support vector contributes more to the regression effect and should be retained in the preliminary preferred support vector set. On the contrary, if removing the support vector has little effect on the prediction error, it can be considered to be removed to simplify the model.

[0137] In order to ensure the accuracy of the screening process, the contribution of each candidate support vector can be evaluated multiple times, and its performance under different data distributions can be recorded to ensure that the retained support vectors have a high prediction contribution under various conditions. This screening method gradually eliminates redundant support vectors while gradually optimizing the structure of the model, so that the final set of preliminary preferred support vectors can achieve ideal regression effects with a smaller number of support vectors.

[0138] Through the above steps, the preliminary optimal support vector set formed not only effectively reduces the complexity of the model, but also improves the prediction accuracy and stability of the model, laying a solid foundation for subsequent parameter optimization and final model construction.

[0139] Furthermore, the optimization algorithm is used to screen the candidate support vectors, select the support vectors that contribute most to the regression effect, and generate a preliminary preferred support vector set, including:

[0140] For the candidate support vectors in the initially trained support vector regression model, the contribution of each candidate support vector to the model regression effect is calculated one by one, specifically including: removing each candidate support vector from the model one by one, and calculating the error increment of the model on the validation set before and after removing the support vector; wherein the error increment is used to evaluate the contribution of the support vector to the model regression effect, and the larger the error increment, the higher the contribution of the support vector to the model regression effect; sorting the candidate support vectors according to the size of the error increment, and selecting the support vector with the highest contribution as the object of further optimization;

[0141] For the initially screened high-contribution support vector set, a gradient descent-based optimization algorithm is applied for fine screening, including: in the initial state, all high-contribution support vectors are used as the current support vector set; in each iteration, the gradient of the current support vector set to the model prediction error is calculated, the weight of the support vector is gradually adjusted, and a penalty term is imposed on the support vector with lower contribution to further reduce its contribution to the model regression effect; when the weight of the specified support vector drops below the preset threshold, it is removed from the support vector set; and the iteration is continued until the change in the model error is lower than the preset convergence standard.

[0142] First, for each candidate support vector, calculate its specific contribution to the model regression effect one by one. The specific calculation method is to remove each candidate support vector separately and observe the change in the prediction error of the model on the validation set before and after removal. By calculating the error increment, the contribution of each support vector to the overall regression effect of the model can be measured. The larger the error increment, the greater the impact of the support vector on the prediction accuracy of the model, and therefore the higher its contribution.

[0143] Next, all candidate support vectors are sorted according to the size of the error increment, and the support vectors with higher contribution are retained as the initial preferred support vector set for further optimization in the next step. In this process, irrelevant or low-contribution support vectors can be effectively reduced, thereby simplifying the model structure, while ensuring that the retained support vectors can effectively improve the regression accuracy.

[0144] Then, for the initially screened high-contribution support vector set, a gradient descent-based optimization algorithm is used for more refined screening. This step aims to further adjust the weights of these support vectors and optimize the model through an iterative process. The specific implementation method is to use all high-contribution support vectors as the initial state of the current support vector set. Then, in each iteration, the weights of the current support vector set are adjusted by calculating the gradient of the model prediction error, focusing on those support vectors that contribute less to the model, and applying penalty terms to these support vectors to reduce their weights, so that their contribution to the model gradually decreases.

[0145] When the weight of a support vector falls below the preset threshold, it is removed from the support vector set. This process is repeated until the error change of the model is lower than the preset convergence standard, indicating that the model has reached a stable state. The final support vector set is not only simplified in terms of the structure of the model, but also can achieve the optimal balance of regression effect while ensuring high prediction accuracy.

[0146] Step S106: The kernel function parameters and penalty factors of the support vector regression model that has been preliminarily trained are adjusted in combination with the grid search method to ensure that the model achieves the best balance between accuracy and complexity, and the preliminarily selected support vector set is further searched and optimized to screen out the final support vector set that can best describe the data features, thereby obtaining an adjusted and optimized support vector regression model.

[0147] In order to implement step S106, the support vector regression model that has been preliminarily trained and supported vectors screened needs to be further optimized to ensure that it achieves the best balance between accuracy and complexity. First, the grid search method is used to adjust the key parameters of the model, including kernel function parameters and penalty factors. The grid search method is a systematic parameter optimization method that first defines a combination of multiple candidate values ​​within a preset parameter range. The kernel function parameters control the mapping method of data points in the support vector regression model in the feature space, while the penalty factor determines the tolerance of the model to errors. By traversing these candidate parameter combinations, the effect of each combination is tested on the validation data set in turn, and its prediction accuracy and complexity performance are recorded.

[0148] After testing all parameter combinations, select the parameter combination that minimizes the prediction error, use it as the optimal kernel function parameter and penalty factor, and apply this parameter combination to the support vector regression model. This step ensures that the model maintains low complexity while ensuring accuracy, thereby improving generalization ability and avoiding overfitting.

[0149] Next, to further optimize the initially selected support vector set, a search algorithm based on local features can be used to ensure that the final support vector set can more accurately describe the characteristics of the data. Specifically, by analyzing the contribution of each support vector in the initially selected support vector set to the prediction results, redundant or less contributing support vectors can be identified and eliminated, thereby retaining the support vectors that are most helpful to the model effect. This process not only reduces the number of support vectors, but also further optimizes the computational efficiency of the model.

[0150] Finally, through the above parameter adjustment and further optimization of the support vector set, an adjusted and optimized support vector regression model is obtained. This model can achieve the best balance between accuracy and complexity, providing a reliable basis for subsequent fault prediction.

[0151] Furthermore, the grid search method is combined to adjust the kernel function parameters and penalty factors of the initially trained support vector regression model to ensure that the model achieves the best balance between accuracy and complexity, and the initially selected support vector set is further searched and optimized to screen out the final support vector set that can best describe the data characteristics, and obtain the adjusted and optimized support vector regression model, including:

[0152] Step P101: for the initially trained support vector regression model, define the search range of the kernel function parameter γ and the penalty factor C, and optimize them by an adaptive grid search method;

[0153] Step P102: In the optimized support vector regression model, further screening support vectors by local density distribution of data to obtain a more representative set of support vectors;

[0154] Wherein, the step P101 specifically includes the following sub-steps:

[0155] Sub-step P101-1: Set the initial search range of kernel function parameter γ and penalty factor C, which are γ={γ1,γ2,…,γ T1} and C={C1,C2,…,C T2}, where γ controls the width of the kernel function in the support vector regression model; C controls the penalty intensity of the model for errors; T1 and T2 are the sizes of the initial search ranges of the kernel function parameter γ and the penalty factor C, respectively;

[0156] Sub-step P101-2: Update the search range by using the following formulas (8) and (9) to gradually narrow it to a range with a smaller model error:

[0157]

[0158] Among them, γ t+1 and C t+1 Respectively represent the kernel function parameters and penalty factors in the t+1th iteration; γ t and C t denote the kernel function parameters and penalty factors in the tth iteration respectively; δ is the adjustment factor, which controls the speed of shrinking the search range; N is the number of samples in the validation set; y i is the true value of the i-th sample; f(x i ; γ t ,C t ) indicates that the parameter γ t and C t The following model takes input x i The predicted value of ; σ(y) is the standard deviation of the true value y in the validation set;

[0159] Sub-step P101-3: Select the parameter combination (γ * ,C * ) as the final optimized kernel function parameters and penalty factors, and apply them to the support vector regression model;

[0160] Wherein, the step P102 specifically includes the following sub-steps:

[0161] Sub-step P102-1: Calculate the local density of each support vector to identify the high-density area of ​​the data. The support vector m j,k The local density ρ(m j,k ) is defined as follows:

[0162]

[0163] in, represents the support vector m j,k Neighborhood set of ; ∥∥m j,k -m j,l ∥∥ represents the support vector m j,k and m j,l The Euclidean distance between them; ε is a smoothing parameter used to control density attenuation;

[0164] Sub-step P102-2: retain the support vectors with higher local density to form the final support vector set to ensure the generalization ability and feature representativeness of the model.

[0165] First, the kernel function parameter γ and penalty factor C of the support vector regression model are optimized by grid search. This optimization process aims to balance the accuracy and complexity of the model, thereby improving its prediction performance. Specifically, the grid search method disperses the range of kernel function parameters and penalty factors into multiple candidate value sets, selects the optimal combination from these sets, and gradually adjusts the prediction error of the model.

[0166] In the first step (step P101), the search range of the kernel function parameter γ and the penalty factor C is defined for the support vector regression model. Specifically, the initial search range is set to γ ​​= {γ1,γ2,…,γ T1} and C={C1,C2,…,C T2}, where γ controls the width of the kernel function, that is, how the similarity of the data decays as the distance increases; C controls the penalty for errors, and a larger C value indicates a lower tolerance for errors. Parameters T1 and T2 are the initial search ranges of the kernel function parameters and the penalty factor, respectively.

[0167] In the grid search method, each combination of the initial search range is evaluated, and in each iteration, the parameter range is gradually narrowed by an adaptive formula. The shrinkage formulas are:

[0168]

[0169] In the above formula, γ t+1 and C t+1 are the kernel function parameters and penalty factors for the t+1th iteration, γ t and C tare the kernel function parameters and penalty factors for the tth iteration respectively. The adjustment factor δ is used to control the shrinking speed; N is the number of samples in the validation set; y i is the true value of the i-th sample, f(x i ; γ t ,C t ) means that the input x is i The predicted value of ; σ(y) is the standard deviation of the validation set. The above formula improves the adaptability of the model and reduces the error by gradually reducing the parameter range.

[0170] Next, we select the combination with the smallest average error on the validation set (γ * ,C * ) is selected, which finally optimizes the grid search process. The selected optimal parameters will be applied to the support vector regression model.

[0171] In the second step (step P102), the optimized support vector set is further screened to ensure the generalization ability and representativeness of the model. The high-density regions in the data are identified by calculating the local density of each support vector. The local density ρ(m (j,k) ) is calculated as follows:

[0172]

[0173] In this formula, represents the support vector m j,k The neighborhood set of contains other support vectors that are close to the support vector; ∥∥m j,k -m j,l ∥∥ represents the support vector m j,k and m j,l The Euclidean distance between them describes the spatial distance between them. The parameter ∈ is a smoothing parameter used to control the rate of density decay and ensure that the density calculation can accurately reflect the local distribution characteristics of the support vector.

[0174] In m j,k and m j,l In , the subscripts j, k, and l represent specific identifiers related to the organization and selection of support vectors. The following is an explanation of their meanings:

[0175] j: This subscript usually represents a level or a layer in a hierarchical structure, such as the multiscale decomposition used in wavelet analysis. In the context of support vector selection, j may represent different levels of detail, or different stages in the optimization process, such as selecting high-contribution support vectors in the early stages and further refining them in the later stages.

[0176] k: This subscript represents the unique identifier of the support vector in a specific level j, indicating the specific location or instance of a support vector in this level. In the set of support vectors in this level, k is used to distinguish each support vector.

[0177] l: This subscript is used to denote another support vector in the same level or neighborhood set for pairwise comparisons, such as computing the Euclidean distance. l enables comparisons between support vectors to assess their neighborhood density or redundancy.

[0178] Therefore, m j,k represents the support vector at level j and position k, and m j,l Then represents another support vector at the same level but at position l. This representation is often used for distance calculation between pairs of support vectors to evaluate which support vectors are closer in the feature space.

[0179] Finally, a set of support vectors with higher local density is retained to build a more representative model, and support vectors with lower density are removed to further improve the generalization performance of the model.

[0180] Step S107: The performance of the adjusted and optimized support vector regression model is evaluated by cross-validation method to verify the accuracy and stability of the model on the training data set and the test data set, so as to obtain the final optimized support vector regression model.

[0181] In order to implement step S107, it is necessary to conduct a comprehensive performance evaluation of the adjusted and optimized support vector regression model through cross-validation to ensure the prediction accuracy and stability of the model on different data sets. First, the training data set is divided into several subsets of equal size, usually five or ten subsets. Next, one of the subsets is selected as the validation set, and the remaining subsets are used as training sets to train the support vector regression model. After the training is completed, the model is used to predict the validation set and the prediction error is recorded.

[0182] This process needs to be repeated several times to ensure that each subset can be used as a validation set in turn for a comprehensive evaluation. Specifically, the combination of the training set and the validation set is updated in each cross-validation to ensure that the model is tested under different data distributions. After each round of validation, a set of prediction error values ​​will be obtained. These error values ​​are averaged to obtain the overall prediction error of the model. The smaller the average error, the higher the prediction performance of the model. This average error can be used to measure the overall accuracy of the model.

[0183] In addition to the average error, the error variance of each validation can also be calculated to evaluate the stability of the model. The error variance reflects the consistency of the model's performance on different data sets. The smaller the variance, the more consistent the model's performance on each data subset and the higher its stability. This process ensures that the model has a high generalization ability on both the training and test data sets.

[0184] Finally, through the above cross-validation process, it is confirmed whether the support vector regression model has reached the expected standards in terms of accuracy and stability, thereby obtaining a verified final optimization model. This model can not only accurately predict the operating status of power equipment, but also has robustness under different environments and data distributions, and is suitable for actual fault prediction applications.

[0185] Furthermore, the performance of the adjusted and optimized support vector regression model is evaluated by cross-validation method to verify the accuracy and stability of the model on the training data set and the test data set to obtain the final optimized support vector regression model, including:

[0186] The adjusted and optimized support vector regression model is applied to the cross-validation method, and the training data set is divided into several subsets to train and validate the model in a grouped manner;

[0187] Summarize the results of each round of cross-validation, calculate the prediction error of the model on each validation set, and take the average error on all validation sets as the overall performance indicator of the model;

[0188] If both the mean error and variance meet the preset accuracy and stability standards, the current support vector regression model is determined as the final optimization model; if not, return to adjust the model parameters or optimize the support vector set, and repeat the cross-validation steps until the final optimization model that meets the requirements is obtained.

[0189] In this embodiment, the performance of the adjusted and optimized support vector regression model needs to be evaluated to verify its accuracy and stability on different data sets. To this end, a cross-validation method is used to comprehensively evaluate the performance of the model. Specifically, the training data set is divided into several subsets of equal size, and one of the subsets is selected as the validation set in each round of cross-validation, and the remaining subsets are used as training sets. In this way, the model can be trained and verified round by round, so that each subset has the opportunity to be used as a validation set, ensuring that the performance of the model is not affected by specific data segmentation.

[0190] In each round of cross-validation, the prediction error of the model on the validation set is calculated. This error value can reflect the accuracy of the model on a specific data set. After all cross-validation rounds are completed, the validation error values ​​of each round are summarized and the average error of the model on all validation sets is calculated. This average error, as an indicator of the overall performance of the model, can reflect the overall accuracy of the model. In addition, in order to evaluate the consistency of the model on different data sets, the variance of the validation errors of each round needs to be calculated. A lower variance indicates that the model performs more consistently on each subset, that is, it has higher stability.

[0191] After the cross-validation is completed, if the average error and error variance of the model reach the preset accuracy and stability standards, it can be considered that the support vector regression model has been optimized to the best state and can be determined as the final optimized model. If the average error or variance does not meet the requirements, it is necessary to further adjust the model parameters or optimize the support vector set, and then repeat the cross-validation steps until the performance indicators of the model meet the expected requirements.

[0192] Step S108: inputting the real-time collected operating parameters of the power equipment into the finally optimized support vector regression model to generate a predicted value of the fault occurrence probability.

[0193] In order to implement step S108, it is necessary to first collect operating parameter data from the power equipment in real time and input these data into the final optimized support vector regression model to generate a probability prediction value of the fault. The operating parameter data collected in real time should include key indicators such as the temperature, vibration frequency, load current, operating voltage, and ambient humidity of the equipment. The collected data will be continuously updated over time to ensure that it reflects the current operating status of the equipment.

[0194] Before inputting real-time data into the model, ensure that the data format is consistent with the data format used for model training. In particular, in the preprocessing stage, apply the same normalization method as the training data (such as maximum-minimum normalization) to process the real-time data. This step ensures that the range and scale of the input data are consistent with those used during model training, thereby avoiding prediction bias caused by inconsistent data.

[0195] The standardized real-time data will be input into the support vector regression model. Based on the input multi-dimensional parameters, the model uses its internal trained regression boundaries and support vectors to calculate the probability prediction value of the equipment failing in the current state. This probability value reflects the possibility of failure of the power equipment under given real-time operating parameters and is a health status assessment based on current data.

[0196] Once the predicted value of the probability of failure is generated, it can be compared with the preset threshold to determine the health status of the equipment. When the predicted value is higher than the set threshold, the system can trigger a fault warning to remind maintenance personnel to pay attention to abnormal conditions of the equipment, and even further predict the possible time of failure, so as to take preventive maintenance measures in advance, thereby reducing the risk of equipment damage and downtime.

[0197] A second embodiment of the present application provides an electronic device, the electronic device comprising:

[0198] processor;

[0199] The memory is used to store a program, and when the program is read and executed by the processor, it executes a method for predicting power equipment faults based on support vector regression provided in the first embodiment of the present application.

[0200] The third embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon. When the program is executed by a processor, a method for predicting power equipment faults based on support vector regression provided in the first embodiment of the present application is executed.

[0201] Although the present application is disclosed as above in the form of a preferred embodiment, it is not intended to limit the present application. Any technical personnel in this field may make possible changes and modifications without departing from the spirit and scope of the present application. Therefore, the scope of protection of the present application shall be based on the scope defined by the claims of the present application.

Claims

1. A method for predicting power equipment faults based on support vector regression, characterized in that: include: Collect multi-dimensional operating parameter data of power equipment to construct an original monitoring data set, wherein the multi-dimensional operating parameters include equipment temperature, vibration frequency, load current, operating voltage and ambient humidity; The original monitoring data set is segmented according to a preset sliding time window, and wavelet transform is applied to each segmented data for noise reduction to suppress interference signals; the noise-reduced data is standardized using a maximum-minimum normalization method to obtain a standardized data set; Extracting temporal features and temporal features from the standardized data set to construct a comprehensive feature vector space; analyzing the correlation of each feature in the comprehensive feature vector space by mutual information method, screening features with rich information and low redundancy, and forming an optimal feature subset; Constructing a training data set using the optimal feature subset; training a support vector regression model using the training data set to obtain a preliminarily trained support vector regression model; wherein the preliminarily trained support vector regression model includes a set of candidate support vectors for describing the regression boundary of the model; Using an optimization algorithm to screen the candidate support vectors, select the support vectors that contribute most to the regression effect, and generate a preliminary optimal support vector set; The kernel function parameters and penalty factors of the initially trained support vector regression model are adjusted by combining the grid search method to ensure that the model achieves the best balance between accuracy and complexity, and the initially selected support vector set is further searched and optimized to screen out the final support vector set that can best describe the data features, thus obtaining an adjusted and optimized support vector regression model. The performance of the adjusted and optimized support vector regression model is evaluated by cross-validation method to verify the accuracy and stability of the model on the training data set and the test data set, so as to obtain the final optimized support vector regression model; The real-time collected operating parameters of the electric power equipment are input into the finally optimized support vector regression model to generate a prediction value of the probability of failure.

2. The method for predicting power equipment faults based on support vector regression according to claim 1, characterized in that: The collecting of multi-dimensional operating parameter data of the power equipment and constructing the original monitoring data set includes: Multi-dimensional operating parameter data of power equipment, including equipment temperature, vibration frequency, load current, operating voltage and ambient humidity, are synchronously collected through multiple independent sensors. The data collection time interval and sampling frequency of each sensor are dynamically adjusted according to the parameter characteristics to ensure synchronization and timing consistency of each parameter.

3. The method for predicting power equipment faults based on support vector regression according to claim 1, characterized in that: The step of extracting temporal features and temporal features from the standardized data set and constructing a comprehensive feature vector space comprises: Extracting time domain features from the standardized data set segment by segment to describe the basic statistical characteristics of the power equipment operating parameters, wherein the time domain features include mean, standard deviation, peak, skewness and kurtosis coefficient; Performing frequency domain analysis on each segment of standardized data to extract frequency domain features to reveal the frequency components and energy distribution of the signal, wherein the frequency domain features include power spectrum density, frequency center, bandwidth, and frequency peak; The extracted time domain features and frequency domain features are combined into a multi-dimensional feature vector to construct a comprehensive feature vector space.

4. The method for predicting power equipment faults based on support vector regression according to claim 1, characterized in that: The mutual information method is used to analyze the correlation of each feature in the comprehensive feature vector space, select features with rich information and low redundancy, and form an optimal feature subset, including: For each feature in the comprehensive feature vector space, the mutual information value is calculated with the target variable to quantify the degree of information sharing between each feature and the target variable; wherein the target variable includes the probability of failure and the health status of the equipment; Based on the results of mutual information analysis, features that contribute more to the target variable information and have low redundancy are preferentially selected, and the optimal feature subset is gradually constructed.

5. The method for predicting power equipment faults based on support vector regression according to claim 1, characterized in that: The step of constructing a training data set using the optimal feature subset includes: Based on the formed optimal feature subset, add a time series label to each data point to ensure the consistency of each data point in the time dimension and avoid time series misalignment between features; The optimal feature data after time alignment are combined into a training data set, where each sample data includes each feature data in the optimal feature subset and its corresponding target variable label.

6. The method for predicting power equipment faults based on support vector regression according to claim 1, characterized in that: The step of training the support vector regression model using the training data set to obtain a preliminarily trained support vector regression model includes: Input the constructed training data set into the support vector regression model, and perform preliminary training using the relationship between the features in the training data set and the target variable; In the support vector regression model after preliminary training, a set of candidate support vectors is identified from the training data through the support vector selection mechanism within the model, wherein the set of candidate support vectors is used to define the boundary of the regression model and represent the key data points of the model in the feature space.

7. The method for predicting power equipment faults based on support vector regression according to claim 1, characterized in that: The candidate support vectors are screened using an optimization algorithm to select the support vectors that contribute most to the regression effect, and a preliminary set of preferred support vectors is generated, including: For the candidate support vectors in the initially trained support vector regression model, the contribution of each candidate support vector to the model regression effect is calculated one by one, specifically including: removing each candidate support vector from the model one by one, and calculating the error increment of the model on the validation set before and after removing the support vector; wherein the error increment is used to evaluate the contribution of the support vector to the model regression effect, and the larger the error increment, the higher the contribution of the support vector to the model regression effect; sorting the candidate support vectors according to the size of the error increment, and selecting the support vector with the highest contribution as the object of further optimization; For the initially screened high-contribution support vector set, a gradient descent-based optimization algorithm is applied for fine screening, including: in the initial state, all high-contribution support vectors are used as the current support vector set; in each iteration, the gradient of the current support vector set to the model prediction error is calculated, the weight of the support vector is gradually adjusted, and a penalty term is imposed on the support vector with lower contribution to further reduce its contribution to the model regression effect; when the weight of the specified support vector drops below the preset threshold, it is removed from the support vector set; and the iteration is continued until the change in the model error is lower than the preset convergence standard.

8. The method for predicting power equipment faults based on support vector regression according to claim 1, characterized in that: The cross-validation method is used to evaluate the performance of the adjusted and optimized support vector regression model, verify the accuracy and stability of the model on the training data set and the test data set, and obtain the final optimized support vector regression model, including: The adjusted and optimized support vector regression model is applied to the cross-validation method, and the training data set is divided into several subsets to train and validate the model in a grouped manner; Summarize the results of each round of cross-validation, calculate the prediction error of the model on each validation set, and take the average error on all validation sets as the overall performance indicator of the model; If both the mean error and variance meet the preset accuracy and stability standards, the current support vector regression model is determined as the final optimization model; if not, return to adjust the model parameters or optimize the support vector set, and repeat the cross-validation steps until the final optimization model that meets the requirements is obtained.

9. The method for predicting power equipment faults based on support vector regression according to claim 1, characterized in that: The original monitoring data set is segmented according to a preset sliding time window, and wavelet transform is applied to each segmented data to perform noise reduction to suppress interference signals, including: Based on the time series of the original monitoring data set X = {x1, x2, ..., x n }, define the sliding time window T and step size s, and use the following formula (1) to determine the optimal segment length T * : Among them, N T is the number of data points in the sliding time window T; x i is the i-th data point in the time series X; x i+1 is the i+1th data point in the time series X; M(T) represents the mean of the data points in the sliding time window T; σ(T) represents the standard deviation of the data points in the sliding time window T; For each optimal segment T * The data in the image are transformed by Haar wavelet transform to decompose each segmented signal into a low-frequency approximation part and a high-frequency detail part, and the segmented signal is obtained as shown in the following formula (2): Among them, J is the number of layers of wavelet decomposition, indicating the scale of wavelet decomposition; K j is the number of detail coefficients decomposed at the jth layer; d j,k is the j-th layer detail coefficient, representing the high-frequency signal part; a J is the approximate coefficient of the Jth layer, representing the low-frequency signal part; ψ j,k (t) is the wavelet basis function of the detail part of the jth layer, through ψ j,k (t) = 2 j / 2 ψ(2 j tk) generates; φ J (t) is the scaling function of the Jth layer, through v J (t) = 2 J / 2 φ(2 J tk) is generated; where the function ψ(t) is the basis function of the Haar wavelet, defined according to the following formula (3): The function φ(t) is the scaling function of the Haar wavelet and is defined according to the following formula (4): According to the following formula (5), the adaptive interval value is calculated Among them, α is the experience adjustment factor; For window T * The number of data points in ; is the window T * The standard deviation of the data points within ; According to the following formula (6), for the detail coefficient d j,k Perform threshold screening to generate a new detail coefficient d j ′ ,k ;: The new detail coefficient d j ′ ,k Substitute into formula (2) and replace the original detail coefficient d j,k , get the noise-reduced segmented signal And the noise-reduced segmented signal as the denoised data.

10. The method for predicting power equipment faults based on support vector regression according to claim 1, characterized in that: The grid search method is combined to adjust the kernel function parameters and penalty factors of the initially trained support vector regression model to ensure that the model achieves the best balance between accuracy and complexity, and further search and optimize the initially selected support vector set to screen out the final support vector set that can best describe the data characteristics, and obtain the adjusted and optimized support vector regression model, including: Step P101: for the initially trained support vector regression model, define the search range of the kernel function parameter γ and the penalty factor C, and optimize them by an adaptive grid search method; Step P102: In the optimized support vector regression model, further screening support vectors by local density distribution of data to obtain a more representative set of support vectors; Wherein, the step P101 specifically includes the following sub-steps: Sub-step P101-1: Set the initial search range of kernel function parameter γ and penalty factor C, which are γ={γ1,γ2,…,γ T1 } and C={C1,C2,…,C T2 }, where γ controls the width of the kernel function in the support vector regression model; C controls the penalty intensity of the model for errors; T1 and T2 are the sizes of the initial search ranges of the kernel function parameter γ and the penalty factor C, respectively; Sub-step P101-2: Update the search range by using the following formulas (8) and (9) to gradually narrow it to a range with a smaller model error: Among them, γ t+1 and C t+1 Respectively represent the kernel function parameters and penalty factors in the t+1th iteration; γ t and C t denote the kernel function parameters and penalty factors in the tth iteration respectively; δ is the adjustment factor, which controls the speed of shrinking the search range; N is the number of samples in the validation set; y i is the true value of the i-th sample; f(x i ; γ t ,C t ) indicates that the parameter γ t and C t The following model takes input x i The predicted value of ; σ(y) is the standard deviation of the true value y in the validation set; Sub-step P101-3: Select the parameter combination (γ * ,C * ) as the final optimized kernel function parameters and penalty factors, and apply them to the support vector regression model; Wherein, the step P102 specifically includes the following sub-steps: Sub-step P102-1: Calculate the local density of each support vector to identify the high-density area of ​​the data. The support vector m j,k The local density ρ(m j,k ) is defined as follows: in, represents the support vector m j,k Neighborhood set of ; ∥∥m j,k -m j,l ∥∥ represents the support vector m j,k and m j,l The Euclidean distance between them; ε is a smoothing parameter used to control density attenuation; Sub-step P102-2: retain the support vectors with higher local density to form the final support vector set to ensure the generalization ability and feature representativeness of the model.

Images