Capacitor fault recognition method based on parameter fusion

Abstract

The application provides a capacitor fault identification method based on parameter fusion. By collecting multi-dimensional characteristic quantity data such as capacitance value, equivalent series resistance value and dielectric loss tangent value, and inputting the characteristic fusion into a target network model, the accurate classification of the capacitor fault type is realized. Through the complementarity of multi-dimensional characteristic quantity and the extraction of time domain statistical characteristics combined with the feature fusion technology, the correlation between the characteristics is fully mined, and the representation ability of the model to the fault mode is enhanced. The target network model can specifically fuse a deep feedforward neural network, a convolutional neural network and a long short-term memory network to extract static characteristics, dynamic characteristics and time sequence dependency, respectively, and form a spatio-temporal joint feature expression. Therefore, the effectiveness, reliability and practicability of the capacitor fault diagnosis in engineering application are improved.

Description

Technical Field

[0001] This application relates to the field of electrical equipment testing, and in particular to a capacitor fault identification method based on parameter fusion. Background Technology

[0002] Metallized film capacitors are widely used in high-voltage direct current transmission, modular multilevel converters, and new energy power generation due to their excellent self-healing properties, high energy density, and good electrical characteristics. However, with long-term operation, capacitors gradually deteriorate in performance and may even fail due to dielectric aging, electrode degradation, the self-healing process, and environmental factors. Effective testing is necessary to determine the degree of degradation and fault condition of capacitors.

[0003] However, current capacitor fault identification methods typically rely on a single feature quantity for fault diagnosis. While this approach has some practicality in specific scenarios, its limitations are obvious. First, over-reliance on a single feature quantity makes it difficult to distinguish and diagnose multiple complex fault modes. Second, there may be correlations and complementarities between various features, and current methods lack in-depth exploration and utilization of these correlations, leading to insufficient diagnostic accuracy. Furthermore, there is a need for dynamic adaptability in the method; changes in the operating environment can significantly interfere with a single feature quantity, and current identification methods struggle to effectively eliminate the influence of environmental noise on the diagnostic results. Additionally, although some feasible implementations introduce simple trend analysis and threshold judgment, they are still insufficient in multi-feature quantity fusion and nonlinear problem handling.

[0004] In summary, the current capacitor fault identification process suffers from poor identification results due to limitations in feature data. Therefore, a capacitor fault identification method based on parameter fusion is needed to effectively utilize the static and dynamic characteristics of capacitors and improve the effectiveness of fault identification. Summary of the Invention

[0005] The purpose of this application is to at least solve one of the aforementioned technical defects, particularly the technical defect in the prior art where the ineffective use of feature data leads to poor capacitor fault identification.

[0006] In a first aspect, this application provides a capacitor fault identification method based on parameter fusion, the method comprising: Collect multi-dimensional characteristic data of the target capacitor; The target capacitor is used to indicate a capacitor in a preset fault state, and the multi-dimensional feature data includes capacitance value, equivalent series resistance value and dielectric loss tangent value. The multi-dimensional feature data is fused to generate target data; The multi-dimensional feature data and the target data are input into the target network model to obtain the fault classification result; The fault classification results include at least one of dielectric degradation faults, conductivity degradation faults, and insulation degradation faults.

[0007] As an optional implementation, before performing feature fusion on the multi-dimensional feature data to generate target data, the method further includes a data cleaning process, specifically including: Based on the isolated forest algorithm, outlier data points in the multi-dimensional feature data are removed. The missing data corresponding to the multi-dimensional feature data is completed by linear interpolation, and the completed multi-dimensional feature data is then normalized.

[0008] As an optional implementation, the step of removing outlier data points from the multi-dimensional feature data according to the Isolation Forest algorithm includes: Randomly select a subset of features from the multidimensional feature data to construct multiple random trees; Based on the minimum path length required for each of the multi-dimensional feature data to be isolated by random numbers, the isolation depth is calculated, and when the path length is less than the dynamic threshold, the corresponding multi-dimensional feature data is identified as an abnormal data point. The dynamic threshold is determined based on the isolated depth distribution of historical data using quantile or maximum entropy methods.

[0009] As an optional implementation, the step of performing feature fusion on the multi-dimensional feature data to generate target data includes: Extract the time-domain and / or frequency-domain features corresponding to the capacitance value, the equivalent series resistance value, and the dielectric loss tangent value; The time-domain features include maximum value, minimum value, average value, standard deviation, slope, kurtosis, and skewness. The time-domain features and / or frequency-domain features are normalized and weighted fusion processes to generate target data.

[0010] As an optional implementation, the target network model includes: A deep feedforward neural network is used to receive the multi-dimensional feature data and extract static features; A convolutional neural network is used to receive the target data and extract dynamic features; A long short-term memory network is used to fuse the outputs of the deep feedforward neural network and the convolutional neural network.

[0011] As an optional implementation, during the training process of the target network model, the acquisition of multi-dimensional feature data of the target capacitor includes: Identify a first number of long-term degraded fault capacitors and a second number of malfunctioning fault capacitors. Using a high-voltage bridge, multiple sets of capacitance values, equivalent series resistance values, and dielectric loss tangent values ​​are obtained for each of the long-term degradation fault capacitors and each of the abnormal operation fault capacitors under a preset time or preset capacity ratio.

[0012] As an optional implementation, the method further includes: Based on the fault classification results, the fault type and corresponding fault severity are determined; Based on the fault type and the fault severity, fault handling information is generated; The fault handling information includes at least one operation instruction among dielectric replacement, electrode repair, or insulation reinforcement.

[0013] Secondly, this application provides a capacitor fault identification device based on parameter fusion, the device comprising: The acquisition module is used to collect multi-dimensional feature data of the target capacitor; The target capacitor is used to indicate a capacitor in a preset fault state, and the multi-dimensional feature data includes capacitance value, equivalent series resistance value and dielectric loss tangent value. The processing module is used to perform feature fusion on the multi-dimensional feature data to generate target data; The processing module is also used to input the multi-dimensional feature data and the target data into the target network model to obtain the fault classification result; The fault classification results include at least one of dielectric degradation faults, conductivity degradation faults, and insulation degradation faults.

[0014] Thirdly, this application provides a computer device including one or more processors and a memory storing computer-readable instructions that, when executed by the one or more processors, perform the steps of the method described in the first aspect.

[0015] Fourthly, this application provides a storage medium storing computer-readable instructions that, when executed by one or more processors, cause the one or more processors to perform the steps of the method described in the first aspect.

[0016] As can be seen from the above technical solutions, the embodiments of this application have the following advantages: Based on any of the above embodiments, the method corresponding to this application systematically solves the shortcomings of traditional capacitor fault diagnosis methods in terms of accuracy, adaptability, and intelligence through multi-dimensional feature fusion, data preprocessing optimization, and hybrid neural network model design. First, it collects multi-dimensional features such as capacitance value, equivalent series resistance value, and dielectric loss tangent value. Combined with time-domain statistical feature extraction and weighted fusion, it overcomes the deficiency of single feature values ​​in representing complex fault modes, significantly enhancing the discriminative power of the feature space. Next, the data cleaning process uses the isolated forest algorithm to remove outliers and linear interpolation to complete missing data, solving the data distortion problem caused by environmental noise and acquisition errors. Improved input data quality significantly enhances the stability of model training. Through the target network model, i.e., the hybrid neural network model, it achieves collaborative analysis of static parameters, dynamic changes, and long-term trends. For example, the deep feedforward neural network can learn the nonlinear relationship between capacitance value and dielectric degradation, the convolutional neural network captures the local abrupt changes in the equivalent series resistance value, and the long short-term memory network models the long-term degradation trend of the dielectric loss tangent value. The integration of these three technologies enables the model to break through the limitations of traditional methods in classifying multiple fault types. Furthermore, since the training data covers the entire lifecycle of the capacitor, some implementation methods combine fault severity quantification with maintenance strategy matching, forming a closed-loop operation encompassing detection, diagnosis, and decision-making, significantly improving maintenance efficiency. In summary, this application enhances the effectiveness, reliability, and practicality of capacitor fault diagnosis in engineering applications. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating a capacitor fault identification method based on parameter fusion provided in one embodiment of this application; Figure 2 This is an internal structural diagram of a computer device provided in an embodiment of this application. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] Metallized film capacitors are widely used in high-voltage direct current transmission, modular multilevel converters, and new energy power generation due to their excellent self-healing properties, high energy density, and good electrical characteristics. However, with long-term operation, capacitors gradually deteriorate in performance and may even fail due to dielectric aging, electrode degradation, the self-healing process, and environmental factors. Effective testing is necessary to determine the degree of degradation and fault condition of capacitors.

[0021] For example, in the field of capacitor fault diagnosis, one feasible implementation includes a capacitor fault detection system based on capacitance value changes. This system determines whether a fault has occurred by monitoring the capacitor's capacitance value in real time and comparing it with the normal operating range. It primarily relies on the single characteristic of capacitance value, using a simple threshold for over-limit alarms, suitable for rapid detection of early degradation. Another feasible implementation diagnoses whether the capacitor has experienced dielectric aging or terminal corrosion by monitoring changes in the equivalent series resistance value. Utilizing the sensitivity of the equivalent series resistance value to the internal losses of the capacitor, it can accurately identify trends in dielectric degradation. However, this method has poor adaptability to dynamic operating environments; for example, in environments with large temperature fluctuations, the equivalent series resistance value is easily affected by noise, leading to a high false alarm rate. Furthermore, this method also cannot integrate other characteristic values, limiting its diagnostic range. Yet another feasible implementation proposes a capacitor fault detection method based on the measurement of the dielectric loss tangent. The dielectric loss tangent value reflects the loss status of the capacitor's insulating dielectric; through long-term monitoring and trend analysis of this value, signs of capacitor degradation can be detected earlier.

[0022] However, while fault diagnosis based solely on a single feature may have some practicality in specific scenarios, its limitations are obvious: over-reliance on a particular feature makes it difficult to distinguish and diagnose multiple complex fault modes; there may be correlations and complementarities between various features, and a single analysis lacks in-depth exploration and utilization of these correlations, resulting in insufficient diagnostic accuracy; poor dynamic adaptability, as changes in the operating environment (such as temperature and humidity) can significantly interfere with a single feature, making it difficult to effectively eliminate the influence of environmental noise on the diagnostic results; some implementation methods introduce simple trend analysis and threshold judgment, but they are still insufficient in multi-feature fusion and nonlinear problem handling, failing to fully utilize the advantages of artificial intelligence technology.

[0023] Based on the above analysis, although some feasible implementation methods provide various approaches to capacitor fault diagnosis, their diagnostic accuracy, adaptability, and intelligence level still have considerable room for improvement. To address these shortcomings, this application further proposes a capacitor fault intelligent identification method based on multi-feature fusion. This method can fully explore the complex relationships between feature quantities, achieve accurate identification of different capacitor fault types, and possesses good environmental adaptability and robustness, guiding equipment operation and maintenance.

[0024] In summary, the technical concept of this application lies in systematically solving the shortcomings of traditional capacitor fault diagnosis methods in terms of accuracy, adaptability, and intelligence through multi-dimensional feature fusion, data preprocessing optimization, and hybrid neural network model design. First, multi-dimensional features such as capacitance, equivalent series resistance, and dielectric loss tangent are collected. Combined with time-domain statistical feature extraction and weighted fusion, this overcomes the deficiency of single features in representing complex fault modes, significantly enhancing the discriminative power of the feature space. Next, the data cleaning process employs the isolated forest algorithm to remove outliers and linear interpolation to complete missing data, solving the data distortion problem caused by environmental noise and acquisition errors. Improved input data quality significantly enhances the stability of model training. Through the target network model, i.e., the hybrid neural network model, collaborative analysis of static parameters, dynamic changes, and long-term trends is achieved. For example, the deep feedforward neural network can learn the nonlinear relationship between capacitance and dielectric degradation, the convolutional neural network captures the local abrupt changes in the equivalent series resistance, and the long short-term memory network models the long-term degradation trend of the dielectric loss tangent. The integration of these three elements enables the model to surpass the limitations of traditional methods in classifying multiple fault types. Furthermore, since the training data covers the entire lifecycle of the capacitor, some implementation methods combine fault severity quantification with maintenance strategy matching, forming a closed-loop operation encompassing detection, diagnosis, and decision-making, significantly improving maintenance efficiency. In summary, this application enhances the effectiveness, reliability, and practicality of capacitor fault diagnosis in engineering applications.

[0025] The methods provided in this application will be described in detail below based on the corresponding implementation methods in some practical application scenarios.

[0026] Please see Figure 1 , Figure 1 This is a flowchart illustrating a capacitor fault identification method based on parameter fusion provided in one embodiment of this application, as shown below. Figure 1 As shown, the method includes: S101. Collect multi-dimensional characteristic data of the target capacitor; The target capacitor is used to indicate a capacitor in a preset fault state, and the multi-dimensional feature data includes capacitance value, equivalent series resistance value and dielectric loss tangent value. Specifically, the target capacitors refer to those capacitors that have been determined to be likely to be in a preset fault state, or capacitors that have a significant impact on the electrical system during operation. The multi-dimensional characteristic data covers multiple key parameters related to the fault state, such as capacitance value, equivalent series resistance value, and dielectric loss tangent. In application, the multi-dimensional characteristic data of the capacitor to be evaluated can be obtained through different offline or online methods. For example, a high-voltage bridge can be used for offline measurement. This method can accurately obtain the capacitance value and loss tangent parameters by applying a standard voltage signal and detecting the response current. Alternatively, online calculation can be performed by real-time monitoring of voltage and current values. Specifically, synchronous sampling technology can be used to achieve high-precision acquisition of voltage and current waveforms, and combined with digital signal processing algorithms (such as Fast Fourier Transform) to calculate the equivalent series resistance value, thereby obtaining the required data.

[0027] S102. Perform feature fusion on the multi-dimensional feature data to generate target data; In this application, the specific operations of feature fusion include, but are not limited to, multiple sub-processes such as feature expansion, feature data cleaning, and feature data processing. For example, feature expansion can achieve dimensionality improvement through time series expansion (such as sliding window sampling) and spatial feature construction (such as parameter ratio calculation); feature data cleaning requires outlier detection (using the 3σ criterion or the isolated forest algorithm) and missing value imputation (based on K-nearest neighbor interpolation or LSTM prediction models); feature data processing includes methods such as standardization (Z-score transformation) and dimensionality reduction (principal component analysis or t-SNE algorithm). Through the comprehensive application of these processes, the effectiveness and accuracy of feature data can be effectively improved, thereby generating more reliable target data, which facilitates efficient processing of the target data by the subsequent target network model.

[0028] S103. Input the multi-dimensional feature data and the target data into the target network model to obtain the fault classification result; Specifically, the fault classification results may include at least one type among dielectric degradation faults, conductivity degradation faults, and insulation degradation faults. In this application, a multi-form basic model architecture is specifically adopted for the design of the target network model. A deep feedforward neural network is responsible for extracting static features (such as absolute parameter values), a convolutional neural network is used to capture dynamic features (such as parameter change rate features), and a long short-term memory network focuses on mining temporal dependencies (such as parameter drift trends). These three networks are weighted and fused through an attention mechanism to form a spatiotemporally joint feature representation. This architecture aims to achieve deep fusion of multiple features, further enhancing the application value of feature data and the overall performance of the model. Specific implementation methods and details will be described in detail in subsequent sections.

[0029] As an optional implementation, the data collection methods during the training of the target network model include: Identify a first number of long-term degraded fault capacitors and a second number of malfunctioning fault capacitors. Using a high-voltage bridge, multiple sets of capacitance values, equivalent series resistance values, and dielectric loss tangent values ​​are obtained for each of the long-term degradation fault capacitors and each of the abnormal operation fault capacitors under a preset time or preset capacity ratio.

[0030] Multiple sets of data for capacitors with long-term degradation and abnormal operation were incorporated into the model training, covering the entire lifecycle of capacitors from early deterioration to complete failure. Traditional training data often focuses on single fault types or short-term data, resulting in insufficient model generalization ability. By collecting multiple sets of data through a high-voltage bridge at preset time or capacity ratios, the gradual and abrupt failure scenarios of capacitors in actual operation were simulated. This data acquisition strategy can improve the model's accuracy in detecting early degradation and reduce the misclassification rate of unknown fault types.

[0031] Specifically, in practical application scenarios, these can include: Collect long-term degraded capacitor samples. Select at least 20 sets of metallized film capacitor equipment that have been in service for a long time as test samples. If this is difficult to achieve, a long-term withstand voltage test can be carried out at one DC voltage and 0.5 AC voltage of the rated capacitor. Place the capacitor in an oven with a constant temperature of 25°C and humidity of 50% for a long time until the capacitance drops to 95%.

[0032] Collect abnormal faulty capacitor samples, and select at least 20 sets of retired capacitor equipment that have abnormal damage but have not burned out as test samples.

[0033] Data on capacitance, equivalent series resistance, and dielectric loss tangent were obtained by testing 40 groups of test samples using a high-voltage bridge at 100h, 200h, 300h, ..., 2000h, or selected up to the service time when the capacitance drops to 95%.

[0034] Capacitance value: Changes in this value can reflect whether the capacitor has experienced dielectric aging, short circuit, or open circuit faults.

[0035] Equivalent series resistance value: An abnormally high value usually indicates dielectric aging or corrosion of internal connection terminals.

[0036] Dielectric loss tangent: reflects the loss characteristics of the capacitor's insulating dielectric. An increase in tangent may indicate problems such as dielectric deterioration or moisture absorption.

[0037] In the data acquisition process, the effectiveness of basic data can also be improved through voting or weighting mechanisms using multi-sensor arrays.

[0038] In this embodiment, by collecting multi-dimensional feature data such as capacitance, equivalent series resistance, and dielectric loss tangent, and then fusing these features before inputting them into the target network model, accurate classification of capacitor fault types is achieved. Through the complementarity of multi-dimensional features and the extraction of temporal statistical features using feature fusion technology, the correlations between features are fully explored, enhancing the model's ability to represent fault modes. Specifically, the target network model can integrate deep feedforward neural networks, convolutional neural networks, and long short-term memory networks to extract static features, dynamic features, and temporal dependencies, forming a spatiotemporally joint feature representation. This improves the effectiveness, reliability, and practicality of capacitor fault diagnosis in engineering applications.

[0039] As an optional implementation, before performing feature fusion on the multi-dimensional feature data to generate target data, the method further includes a data cleaning process, specifically including: Based on the isolated forest algorithm, outlier data points in the multi-dimensional feature data are removed. The missing data corresponding to the multi-dimensional feature data is completed by linear interpolation, and the completed multi-dimensional feature data is then normalized.

[0040] In this implementation, the Isolation Forest algorithm is introduced to remove outlier data points, and linear interpolation is used to fill in missing data. Combined with normalization, this effectively improves the quality and consistency of the input data. Traditional methods often introduce interference into the model due to environmental noise or acquisition errors causing data anomalies or missing values, reducing diagnostic reliability. The Isolation Forest algorithm, based on unsupervised learning, quickly identifies outliers in high-dimensional data through random tree construction and path length calculation. Dynamic threshold settings further enhance the adaptability of anomaly detection. Linear interpolation fills in missing values ​​using the continuity of adjacent data points, avoiding model training bias caused by data interruptions. Normalization eliminates dimensional differences, ensuring the balance of multiple features during fusion and model input. This improves model training efficiency and diagnostic reliability.

[0041] As an optional implementation, the step of removing outlier data points from the multi-dimensional feature data according to the Isolation Forest algorithm includes: Randomly select a subset of features from the multidimensional feature data to construct multiple random trees; Based on the minimum path length required for each of the multi-dimensional feature data to be isolated by random numbers, the isolation depth is calculated, and when the path length is less than the dynamic threshold, the corresponding multi-dimensional feature data is identified as an abnormal data point. The dynamic threshold is determined based on the isolated depth distribution of historical data using quantile or maximum entropy methods.

[0042] Specifically, in practical application scenarios, this implementation method may include: Outlier removal: An anomaly detection method based on the isolated forest algorithm removes outliers during the data collection process.

[0043] Isolation Forest is an unsupervised anomaly detection algorithm based on a tree structure. This method can automatically detect and remove outliers in data without supervision, such as outliers caused by equipment failure or sampling errors. It is particularly suitable for outlier removal in high-dimensional data.

[0044] The specific steps are as follows: Modeling: Randomly select several feature subsets to construct multiple random trees. Each tree forms different partitioning paths by randomly dividing the data space.

[0045] Anomaly scoring: Calculate the "isolation depth" for each data point, which is the path length required for that point to be isolated by a random tree. Outliers typically have shorter isolation depths.

[0046] Removal: Set a threshold for abnormal scores (such as based on quantiles or maximum entropy methods), mark points with scores exceeding the threshold as abnormal points and remove them.

[0047] Furthermore, the aforementioned data completion process may specifically include: For missing data caused by equipment failure or acquisition error, a linear interpolation algorithm is used to fill in the missing data.

[0048] Linear interpolation is a simple and effective data completion method. It can effectively fill in data gaps within a short timeframe, ensuring the continuity of subsequent analysis. It is suitable for situations with high sampling frequencies and strong data continuity. The specific steps are as follows: Identify missing points: Mark the locations of missing values ​​in the dataset due to equipment problems or acquisition errors.

[0049] Linear interpolation: For each missing point, the interpolation is calculated using the values ​​of the two valid data points before and after it, according to the following formula: x_missing = x_before + (x_after - x_before) × (t_missing - t_before) / (t_before - t_before).

[0050] Where xbefore and xafter are the values ​​of the adjacent points before and after the missing point, respectively, and tmissing is the timestamp of the missing point.

[0051] This implementation details the steps of the Isolation Forest algorithm, including random feature subset selection, construction of multiple random trees, and dynamic threshold determination. Isolation Forest utilizes the ease with which outliers are isolated by randomly partitioning the data space, quantifying isolation depth by path length, and optimizing the judgment criteria through dynamic thresholding combined with historical data distribution to avoid over-removal or under-removal issues caused by fixed thresholds. In scenarios with multi-dimensional features in capacitors, the efficiency and effectiveness of anomaly detection are improved, while ensuring real-time processing of high-dimensional data.

[0052] As an optional implementation, the step of performing feature fusion on the multi-dimensional feature data to generate target data includes: Extract the time-domain and / or frequency-domain features corresponding to the capacitance value, the equivalent series resistance value, and the dielectric loss tangent value; The time-domain features include maximum value, minimum value, average value, standard deviation, slope, kurtosis, and skewness. The time-domain features and / or frequency-domain features are normalized and weighted fusion processes to generate target data.

[0053] Furthermore, in some feasible implementations, frequency domain features can be incorporated to jointly perform data fusion. For example, the aforementioned time-domain features can be converted into frequency-domain features using a Fast Fourier Transform (FFT), extracting the dominant frequency amplitude and harmonic component ratio. A gated recurrent unit can then dynamically assign weights to the time-domain and frequency-domain features to generate a fused feature vector. The dynamic weight assignment method can include: calculating the correlation between the time-domain and frequency-domain features using a time attention mechanism, assigning weight coefficients to each feature based on the correlation, and sparsifying low-weight features.

[0054] In specific implementations, the feature extraction process and the physical meaning of each data point may include: Five time-domain features were extracted from the capacitance value C(t), the equivalent series resistance ESR(t), and the dielectric loss tangent tan δ(t).

[0055] The maximum and minimum values ​​are: the maximum value of the equivalent series resistance may represent the limit of current loss, and the minimum value of the capacitance reflects the lowest performance state of the capacitor, which usually occurs when the capacitor is severely degraded.

[0056] The average value, the average value of each curve, reflects the long-term performance of the capacitor. If the capacitor deteriorates, the average capacitance will typically show a downward trend, while the equivalent series resistance will increase.

[0057] C(t): Average value C avg =1 / n∑C(t i ), ESR(t): Average ESRavg =1 / n∑ESR(t i ), tanδ(t): Average value tanδ avg =1 / n∑tanδ(t i ).

[0058] n is the number of data points.

[0059] Standard deviation measures the volatility of each characteristic value. Capacitor losses and uneven mass can lead to significant volatility.

[0060] C(t): , ESR(t): , tan δ(t): .

[0061] The slope, calculated through linear regression analysis, represents the trend of change for each curve. The slope reflects the rate of change in capacitor performance; a positive slope indicates performance improvement, while a negative slope indicates performance degradation.

[0062] The time series curve of the capacitance was fitted using linear regression, C(t): , where ac is the slope, and the equivalent series resistance is similar to the tangent of the dielectric loss angle.

[0063] Kurtosis and skewness: Kurtosis: Measures the degree to which data is peaked. High kurtosis indicates that the data distribution is relatively concentrated, while low kurtosis indicates that the data distribution is relatively flat.

[0064] Skewness: Measures the degree of skewness in the distribution of data. Positive skewness indicates that the data is concentrated on the left, while negative skewness indicates that the data is concentrated on the right.

[0065] Feature fusion: Normalization, to ensure that features have the same dimension, maps all feature values ​​to the range [0,1], using the Min-Max normalization method to eliminate dimensional differences between features.

[0066] The normalization formula is: x norm =(xx min ) / (x max -x min ).

[0067] Weighted fusion involves merging each feature of each curve using a weighted average. The weighting coefficients can be determined based on experience or through data-driven methods (such as feature importance).

[0068] Weighted average: Assuming 5 features are extracted for each parameter, totaling 15 features (5 features for C, ESR, and tanδ), weights can be assigned to each feature, and a weighted sum can be calculated: F fusion = w1 F1+w2 F2+ +w15 F15.

[0069] Where w1, w2, ..., w15 are the corresponding weights.

[0070] This implementation addresses the problem of insufficient representational power of single features by extracting time-domain features and performing normalized weighted fusion. Traditional methods directly use the original features, neglecting the indicative role of time-domain statistical characteristics in fault modes. Specifically, the standard deviation of capacitance values ​​reflects capacitor performance fluctuations, the slope quantifies the rate of degradation, and kurtosis and skewness reveal the correlation between data distribution patterns and faults. Weighted fusion optimizes the expressive power of feature combinations by assigning differentiated weights to different features. The fused target data can improve the model's sensitivity in identifying various capacitor faults and its resistance to noise interference.

[0071] As an optional implementation, the target network model includes: A deep feedforward neural network is used to receive the multi-dimensional feature data and extract static features; A convolutional neural network is used to receive the target data and extract dynamic features; A long short-term memory network is used to fuse the outputs of the deep feedforward neural network and the convolutional neural network.

[0072] During neural network training, the cross-entropy loss function can be used to measure the error, the Adam optimizer can be used to dynamically adjust the learning rate, and the robustness of the model can be enhanced by adversarial example generation strategies, such as injecting Gaussian noise or obtaining adversarial examples based on generative adversarial networks. In addition, early stopping can be used to terminate training as needed.

[0073] Specifically, this application designs a multi-stage neural network architecture that combines deep feedforward neural networks (DNN), convolutional neural networks (CNN), and long short-term memory networks (LSTM) to gradually extract the static and dynamic relationships between features and ultimately complete the fault classification task.

[0074] Deep feedforward neural networks (DNNs) serve as the first-stage network, receiving preprocessed static features. Each fusion parameter has a clear feature label, corresponding to dielectric degradation faults, conductive degradation faults, and insulation degradation faults, and learns the direct relationship between them and fault types.

[0075] DNN is the most basic deep learning model, composed of multiple layers of fully connected neurons. Its core idea is to extract high-level features from the data layer by layer by performing non-linear transformations on the input data through weights and activation functions.

[0076] Mathematical representation: h (l) =f(W (l) h (l-1) +b (l) ).

[0077] Among them, h (l) For the output of layer l, W (l) and b (l) Here, denoted as weights and biases, respectively, and f is the ReLU activation function.

[0078] Convolutional Neural Networks (CNNs) are second-stage networks that input time-series features (such as the rate of change of tanδ) into convolutional layers to extract local dynamic patterns (such as rapidly deteriorating time segments). CNNs receive static features extracted by DNNs as prior information and combine static features with time-series features through feature concatenation, thereby enhancing the richness of feature representation.

[0079] CNNs extract local patterns and spatial features from data through local receptive fields and weight sharing mechanisms, and are often used to process two-dimensional images or one-dimensional signal sequences.

[0080] Long Short-Term Memory (LSTM) network, as the third-stage network, further processes time series features, capturing the long-term trends and dynamic changes of capacitor features. The input of LSTM includes local features extracted by CNN and global static features extracted by DNN, forming a spatiotemporal joint feature representation.

[0081] LSTM is a special type of recurrent neural network (RNN) that solves the gradient vanishing and exploding problems of traditional RNNs when dealing with long-term dependencies by introducing gating mechanisms (input gate, forget gate, output gate).

[0082] Integration and Training: In the final stage, the outputs of DNN, CNN and LSTM are fused through ensemble learning to output the fault classification results.

[0083] Supervised learning was employed, and the model training was completed through the following steps: 1. The classification error was measured using the cross-entropy loss function; 2. Algorithm optimization: The Adam optimizer was used to dynamically adjust the learning rate to accelerate convergence; 3. Training and validation: The data was divided into a training set (70%), a validation set (20%), and a test set (10%) to ensure the stability and generalization ability of the model performance, and three fault type labels were output respectively: dielectric degradation fault, conductivity degradation fault, and insulation degradation fault.

[0084] This implementation refines the specific architecture of the target network model, employing a hybrid model architecture of deep feedforward neural networks, convolutional neural networks, and long short-term memory networks to extract static features, dynamic features, and temporal dependencies, respectively. Traditional single-network models struggle to simultaneously capture both the static parameters and dynamic trends of capacitors. The deep feedforward neural network learns the nonlinear mapping between features and fault types through fully connected layers, the convolutional neural network extracts local dynamic patterns using convolutional kernels, and the long short-term memory network models long-term degradation trends through a gating mechanism. The integration of these three technologies achieves multi-level feature fusion, improving the model's classification accuracy for complex fault modes.

[0085] As an optional implementation, the method further includes: Based on the fault classification results, the fault type and corresponding fault severity are determined; Based on the fault type and the fault severity, fault handling information is generated; The fault handling information includes at least one operation instruction among dielectric replacement, electrode repair, or insulation reinforcement.

[0086] In some implementations, fault handling information can also be generated in the form of a knowledge graph.

[0087] For example, the construction of a fault handling knowledge graph may include: extracting fault-cause-measure triples from maintenance manuals, adding weight labels to the triples based on maintenance records, and embedding expert experience rules through knowledge distillation techniques.

[0088] Furthermore, the logic for generating fault handling information may include: associating fault types with preset maintenance knowledge graphs, dynamically adjusting maintenance priorities based on equipment operating status and spare parts inventory, and updating the knowledge graph through graph neural networks.

[0089] This implementation generates specific operation instructions based on the fault classification results, achieving closed-loop management from diagnosis to maintenance. Some implementations only output the fault type, lacking direct support for operation and maintenance decisions. This implementation can quantify the fault severity and match maintenance strategies with an expert knowledge base, improving operation and maintenance efficiency. For example, when the insulation degradation fault severity is moderate, an insulation strengthening instruction is automatically triggered; if it is severe, capacitor replacement is recommended. In practical applications, this process significantly reduces fault handling response time and maintenance costs, improving the effectiveness and practicality of fault identification.

[0090] This application embodiment also provides a flexible DC fault active clearing device, the device comprising: The acquisition module is used to collect multi-dimensional feature data of the target capacitor; The target capacitor is used to indicate a capacitor in a preset fault state, and the multi-dimensional feature data includes capacitance value, equivalent series resistance value and dielectric loss tangent value. The processing module is used to perform feature fusion on the multi-dimensional feature data to generate target data; The processing module is also used to input the multi-dimensional feature data and the target data into the target network model to obtain the fault classification result; The fault classification results include at least one of dielectric degradation faults, conductivity degradation faults, and insulation degradation faults.

[0091] In this embodiment, by collecting multi-dimensional feature data such as capacitance, equivalent series resistance, and dielectric loss tangent, and then fusing these features before inputting them into the target network model, accurate classification of capacitor fault types is achieved. Through the complementarity of multi-dimensional features and the extraction of temporal statistical features using feature fusion technology, the correlations between features are fully explored, enhancing the model's ability to represent fault modes. Specifically, the target network model can integrate deep feedforward neural networks, convolutional neural networks, and long short-term memory networks to extract static features, dynamic features, and temporal dependencies, forming a spatiotemporally joint feature representation. This improves the effectiveness, reliability, and practicality of capacitor fault diagnosis in engineering applications.

[0092] As an optional implementation, the processing module is further configured to perform a data cleaning process after the acquisition module collects the multi-dimensional feature data of the target capacitor, specifically including: Based on the isolated forest algorithm, outlier data points in the multi-dimensional feature data are removed. The missing data corresponding to the multi-dimensional feature data is completed by linear interpolation, and the completed multi-dimensional feature data is then normalized.

[0093] In this implementation, the Isolation Forest algorithm is introduced to remove outlier data points, and linear interpolation is used to fill in missing data. Combined with normalization, this effectively improves the quality and consistency of the input data. Traditional methods often introduce interference into the model due to environmental noise or acquisition errors causing data anomalies or missing values, reducing diagnostic reliability. The Isolation Forest algorithm, based on unsupervised learning, quickly identifies outliers in high-dimensional data through random tree construction and path length calculation. Dynamic threshold settings further enhance the adaptability of anomaly detection. Linear interpolation fills in missing values ​​using the continuity of adjacent data points, avoiding model training bias caused by data interruptions. Normalization eliminates dimensional differences, ensuring the balance of multiple features during fusion and model input. This improves model training efficiency and diagnostic reliability.

[0094] As an optional implementation, the specific method by which the processing module removes outlier data points from the multi-dimensional feature data according to the isolated forest algorithm includes: Randomly select a subset of features from the multidimensional feature data to construct multiple random trees; Based on the minimum path length required for each of the multi-dimensional feature data to be isolated by random numbers, the isolation depth is calculated, and when the path length is less than the dynamic threshold, the corresponding multi-dimensional feature data is identified as an abnormal data point. The dynamic threshold is determined based on the isolated depth distribution of historical data using quantile or maximum entropy methods.

[0095] This implementation details the steps of the Isolation Forest algorithm, including random feature subset selection, construction of multiple random trees, and dynamic threshold determination. Isolation Forest utilizes the ease with which outliers are isolated by randomly partitioning the data space, quantifying isolation depth by path length, and optimizing the judgment criteria through dynamic thresholding combined with historical data distribution to avoid over-removal or under-removal issues caused by fixed thresholds. In scenarios with multi-dimensional features in capacitors, the efficiency and effectiveness of anomaly detection are improved, while ensuring real-time processing of high-dimensional data.

[0096] As an optional implementation, the specific method by which the processing module performs feature fusion on the multi-dimensional feature data to generate target data includes: Extract the time-domain and / or frequency-domain features corresponding to the capacitance value, the equivalent series resistance value, and the dielectric loss tangent value; The time-domain features include maximum value, minimum value, average value, standard deviation, slope, kurtosis, and skewness. The time-domain features and / or frequency-domain features are normalized and weighted fusion processes to generate target data.

[0097] This implementation addresses the problem of insufficient representational power of single features by extracting time-domain features and performing normalized weighted fusion. Traditional methods directly use the original features, neglecting the indicative role of time-domain statistical characteristics in fault modes. Specifically, the standard deviation of capacitance values ​​reflects capacitor performance fluctuations, the slope quantifies the rate of degradation, and kurtosis and skewness reveal the correlation between data distribution patterns and faults. Weighted fusion optimizes the expressive power of feature combinations by assigning differentiated weights to different features. The fused target data can improve the model's sensitivity in identifying various capacitor faults and its resistance to noise interference.

[0098] As an optional implementation, the target network model includes: A deep feedforward neural network is used to receive the multi-dimensional feature data and extract static features; A convolutional neural network is used to receive the target data and extract dynamic features; A long short-term memory network is used to fuse the outputs of the deep feedforward neural network and the convolutional neural network.

[0099] This implementation refines the specific architecture of the target network model, employing a hybrid model architecture of deep feedforward neural networks, convolutional neural networks, and long short-term memory networks to extract static features, dynamic features, and temporal dependencies, respectively. Traditional single-network models struggle to simultaneously capture both the static parameters and dynamic trends of capacitors. The deep feedforward neural network learns the nonlinear mapping between features and fault types through fully connected layers, the convolutional neural network extracts local dynamic patterns using convolutional kernels, and the long short-term memory network models long-term degradation trends through a gating mechanism. The integration of these three technologies achieves multi-level feature fusion, improving the model's classification accuracy for complex fault modes.

[0100] As an optional implementation, during the training process of the target network model, the acquisition module is further configured to: Identify a first number of long-term degraded fault capacitors and a second number of malfunctioning fault capacitors. Using a high-voltage bridge, multiple sets of capacitance values, equivalent series resistance values, and dielectric loss tangent values ​​are obtained for each of the long-term degradation fault capacitors and each of the abnormal operation fault capacitors under a preset time or preset capacity ratio.

[0101] This implementation incorporates multiple sets of data from long-term deterioration and malfunctioning capacitors during model training, covering the entire lifecycle of capacitors from early degradation to complete failure. Traditional training data often focuses on single fault types or short-term data, leading to insufficient model generalization ability. By collecting multiple sets of data using a high-voltage bridge at preset times or capacity ratios, the gradual and abrupt failure scenarios of capacitors in actual operation are simulated. This data acquisition strategy improves the model's accuracy in detecting early degradation and reduces the misclassification rate for unknown fault types.

[0102] As an optional implementation, the processing module is further configured to: Based on the fault classification results, the fault type and corresponding fault severity are determined; Based on the fault type and the fault severity, fault handling information is generated; The fault handling information includes at least one operation instruction among dielectric replacement, electrode repair, or insulation reinforcement.

[0103] This implementation generates specific operation instructions based on the fault classification results, achieving closed-loop management from diagnosis to maintenance. Some implementations only output the fault type, lacking direct support for operation and maintenance decisions. This implementation can quantify the fault severity and match maintenance strategies with an expert knowledge base, improving operation and maintenance efficiency. For example, when the insulation degradation fault severity is moderate, an insulation strengthening instruction is automatically triggered; if it is severe, capacitor replacement is recommended. In practical applications, this process significantly reduces fault handling response time and maintenance costs, improving the effectiveness and practicality of fault identification.

[0104] It should be noted that the division of the various modules in the above device is merely a logical functional division. In actual implementation, they can be fully or partially integrated into a single physical entity, or they can be physically separated. Furthermore, these modules can be implemented entirely in software via processing element calls; they can be fully implemented in hardware; or some modules can be implemented by processing element calls to software, while others are implemented in hardware. For example, a processing module can be a separate processing element, or it can be integrated into a chip within the device. Alternatively, it can be stored as program code in the device's memory, and its functions can be called and executed by a processing element. The implementation of other modules is similar. Moreover, these modules can be fully or partially integrated together, or they can be implemented independently. The processing element here can be an integrated circuit with signal processing capabilities. During implementation, each step of the above method or each of the above modules can be completed through integrated logic circuits in the hardware of the processor element or through software instructions.

[0105] Indicatively, such as Figure 2 As shown, Figure 2 This is a schematic diagram of the internal structure of a computer device 300 provided in an embodiment of this application. The computer device 300 can be provided as a server. (Refer to...) Figure 2 The computer device 300 includes a processing component 302, which further includes one or more processors, and memory resources represented by memory 301 for storing instructions, such as application programs, that can be executed by the processing component 302. The application programs stored in memory 301 may include one or more modules, each corresponding to a set of instructions. Furthermore, the processing component 302 is configured to execute instructions to perform the methods of any of the embodiments described above.

[0106] The computer device 300 may also include a power supply component 303 configured to perform power management of the computer device 300, a wired or wireless network interface 304 configured to connect the computer device 300 to a network, and an input / output (I / O) interface 305. The computer device 300 may operate on an operating system stored in memory 301, such as Windows Server™, Mac OS X™, Unix™, Linux™, Free BSD™, or similar.

[0107] Those skilled in the art will understand that Figure 2 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0108] This application provides a storage medium storing computer-readable instructions that, when executed by one or more processors, cause the one or more processors to perform the method provided in any embodiment.

[0109] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0110] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can be referred to each other.

[0111] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A capacitor fault identification method based on parameter fusion, characterized in that, The method includes: Collect multi-dimensional characteristic data of the target capacitor; The target capacitor is used to indicate a capacitor in a preset fault state, and the multi-dimensional feature data includes capacitance value, equivalent series resistance value and dielectric loss tangent value. The multi-dimensional feature data is fused to generate target data; The multi-dimensional feature data and the target data are input into the target network model to obtain the fault classification results.

2. The method according to claim 1, characterized in that, Before performing feature fusion on the multi-dimensional feature data to generate the target data, the method further includes: Based on the isolated forest algorithm, outlier data points in the multi-dimensional feature data are removed. The missing data corresponding to the multi-dimensional feature data is completed by linear interpolation, and the completed multi-dimensional feature data is then normalized.

3. The method according to claim 2, characterized in that, The step of removing outlier data points from the multi-dimensional feature data according to the isolated forest algorithm includes: Randomly select a subset of features from the multidimensional feature data to construct multiple random trees; Based on the minimum path length required for each of the multi-dimensional feature data to be isolated by random numbers, the isolation depth is calculated, and when the path length is less than the dynamic threshold, the corresponding multi-dimensional feature data is identified as an abnormal data point. The dynamic threshold is determined based on the isolated depth distribution of historical data using quantile or maximum entropy methods.

4. The method according to claim 1, characterized in that, The step of fusing the multi-dimensional feature data to generate target data includes: Extract the time-domain and / or frequency-domain features corresponding to the capacitance value, the equivalent series resistance value, and the dielectric loss tangent value; The time-domain features and / or frequency-domain features are normalized and weighted fusion processes to generate target data.

5. The method according to claim 1, characterized in that, The target network model includes: A deep feedforward neural network is used to receive the multi-dimensional feature data and extract static features; A convolutional neural network is used to receive the target data and extract dynamic features; A long short-term memory network is used to fuse the outputs of the deep feedforward neural network and the convolutional neural network.

6. The method according to claim 5, characterized in that, During the training process of the target network model, the method further includes: Identify a first number of long-term degraded fault capacitors and a second number of malfunctioning fault capacitors. Using a high-voltage bridge, multiple sets of capacitance values, equivalent series resistance values, and dielectric loss tangent values ​​are obtained for each of the long-term degradation fault capacitors and each of the abnormal operation fault capacitors under a preset time or preset capacity ratio.

7. The method according to any one of claims 1-6, characterized in that, The method further includes: Based on the fault classification results, the fault type and corresponding fault severity are determined; Based on the fault type and the fault severity, fault handling information is generated; The fault handling information includes at least one operation instruction among dielectric replacement, electrode repair, or insulation reinforcement.

8. A capacitor fault identification device based on parameter fusion, characterized in that, The device includes: The acquisition module is used to collect multi-dimensional feature data of the target capacitor; The target capacitor is used to indicate a capacitor in a preset fault state, and the multi-dimensional feature data includes capacitance value, equivalent series resistance value and dielectric loss tangent value. The processing module is used to perform feature fusion on the multi-dimensional feature data to generate target data; The processing module is also used to input the multi-dimensional feature data and the target data into the target network model to obtain the fault classification result.

9. A computer device, characterized in that, The method includes one or more processors and a memory storing computer-readable instructions that, when executed by the one or more processors, perform the steps of the method as described in any one of claims 1-7.

10. A storage medium, characterized in that, The storage medium stores computer-readable instructions that, when executed by one or more processors, cause the one or more processors to perform the steps of the method as described in any one of claims 1-7.

Images