A fault auxiliary analysis method and system based on distributed sampling

Abstract

A fault auxiliary analysis method and system based on distributed sampling. The operation data and waveform signal data of the power grid are collected in real time at different locations of the power grid; characteristic values ​​that may represent faults are extracted from the pre-processed collected data; a variety of machine learning basic models are used to perform preliminary fault identification on the characteristic values; and a comprehensive judgment is made on the fault type based on the results of the integrated learning to ensure the synergistic effect of multiple algorithms. The present invention combines the integrated learning method of multiple machine learning algorithms, which can combine the advantages of multiple algorithms and enhance the robustness of fault identification. Through weighted voting or stacking methods, the system can make more accurate comprehensive judgments based on the judgment results of each algorithm, reducing the risk of judgment errors of a single algorithm.

Description

Technical Field

[0001] The present invention belongs to the technical field of power system monitoring, and in particular relates to a fault auxiliary analysis method and system based on distributed sampling. Background Art

[0002] With the continuous development of power systems, grids are becoming larger and larger, and the types of equipment are becoming increasingly diverse. Failures in equipment such as transformers, switches, and conductors can disrupt system operations, causing significant economic losses and safety hazards. Traditional fault detection methods, particularly in complex power systems, rely primarily on manual inspections or single monitoring methods. These methods are prone to blind spots, making it difficult to detect potential faults in a timely manner and preventing rapid system recovery.

[0003] In recent years, with the rapid development of big data, machine learning, and artificial intelligence technologies, data-driven fault diagnosis methods have gradually become a research hotspot. These methods collect real-time equipment operating data and use machine learning models to assess equipment status, enabling more accurate predictions of equipment operating status and fault trends. However, existing technologies still face the following challenges in fault diagnosis: various machine learning algorithms have limited applicability in power equipment fault diagnosis and are unable to cope with the diverse fault types in complex environments. Furthermore, they lack intelligent and dynamic adjustment mechanisms, preventing them from timely adjusting fault diagnosis criteria based on actual conditions, resulting in inaccurate fault diagnosis.

[0004] For example, the patent application with publication number CN118607390A discloses a single centralized data processing architecture for intelligent diagnosis of distribution network line faults, which makes it difficult to accurately judge various fault types in complex environments. Summary of the Invention

[0005] The purpose of this section is to summarize some aspects of the embodiments of the present invention and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section and the abstract and title of this application to avoid obscuring the purpose of this section, the abstract and the title of the invention, and such simplifications or omissions should not be used to limit the scope of the present invention.

[0006] In view of the problems existing in the above-mentioned prior art, the present invention is proposed. In a first aspect, the present invention discloses a fault auxiliary analysis method based on distributed sampling, comprising:

[0007] Real-time distributed sampling of grid data at multiple locations on the grid;

[0008] Preprocess the sampled data and extract the feature values ​​of the set dimension from the preprocessed data;

[0009] The extracted feature values ​​are input into different machine learning basic models, and each machine learning basic model makes independent predictions;

[0010] An integrated learning algorithm is used to fuse the prediction results of all machine learning basic models to obtain a comprehensive fault identification result;

[0011] The evaluation value of the comprehensive fault identification result is calculated, and the ensemble learning algorithm is optimized based on the reward and penalty mechanism to obtain the optimal comprehensive fault identification result.

[0012] Preferably, the eigenvalue of the set dimension is extracted from the preprocessed data, and the extraction formula is:

[0013] ;

[0014] in, is the characteristic function, indicating the The eigenvalues ​​of the set dimension, Collection time point The time interval is used for feature normalization processing; for The preprocessed data at time for In scale The wavelet coefficients on For the The scale parameter of the setting dimension; K is the number of setting dimensions.

[0015] Preferably, the extracted feature values Input into different machine learning basic models to obtain the confidence level of each machine learning basic model for the fault type, which is expressed as the following formula:

[0016] ;

[0017] in, For the Machine learning based models for fault types The confidence level of , the value range is [0,1]; To adjust the nonlinear scaling parameters of the machine learning basic model output; Indicates the The basic machine learning model The characteristic value of the set dimension The predicted output after calculation is, is the number of basic machine learning models.

[0018] Preferably, all machine learning basic models are integrated through the ensemble learning algorithm to Confidence level of each fault type , and obtain the comprehensive fault identification result, which is expressed by the following formula:

[0019] ;

[0020] in, is the comprehensive fault identification result, Indicates the The weights of the machine learning base model, For the The weight adjustment parameters of the basic machine learning model.

[0021] Preferably, according to the comprehensive fault identification results , judged by the set decision threshold The corresponding fault is expressed by the following formula:

[0022] ;

[0023] in, represents the combination of decision thresholds set for the fault type, and ; is the number of decision thresholds to set.

[0024] Preferably, the decision threshold is set by using a supervised learning-based training model to train based on historical fault sample data to obtain a threshold combination. .

[0025] Preferably, by the formula Obtain an evaluation value of the comprehensive fault identification result; wherein, Comprehensive fault identification results The assessed value of For real results.

[0026] Preferably, optimizing the ensemble learning algorithm using a reward and penalty mechanism includes:

[0027] If the evaluation value of the comprehensive fault identification result Greater than the set reference threshold , then rewards are given, no weight adjustment is made, and the optimization ends;

[0028] If the evaluation value of the comprehensive fault identification result Less than the set reference threshold , then the penalty is triggered and the weight of each machine learning basic model in the integrated learning algorithm is adjusted.

[0029] Preferably, the weights of each machine learning basic model are readjusted according to the penalty rule, and the weight adjustment formula is:

[0030] ;

[0031] in, and Respectively The basic machine learning model Round and The weight of the round; is the learning rate, which is used to control the magnitude of weight adjustment; is the set reference threshold.

[0032] A second aspect of the present invention discloses a fault assisted analysis system based on distributed sampling, which runs the above-mentioned fault assisted analysis method based on distributed sampling, including:

[0033] Distributed data acquisition module, used for real-time distributed collection of power grid data;

[0034] Data preprocessing module, used to preprocess the collected data;

[0035] A fault identification module is used to identify faults based on preprocessed data;

[0036] The optimization module is used to optimize the fault identification process to obtain the best fault identification result.

[0037] Beneficial effects of the present invention:

[0038] 1. This invention acquires real-time data from diverse environments and operating conditions through distributed data sampling at multiple locations in the power grid. Combining independent predictions from multiple machine learning models, this method employs an ensemble learning approach, leveraging the strengths of multiple algorithms to enhance the robustness of fault identification. This invention can handle the complexity and diversity of power grid environments, reducing the risk of misjudgment by a single algorithm. By integrating multiple models, the system can accurately identify different types of faults, thereby improving the accuracy and adaptability of fault diagnosis.

[0039] 2. This invention optimizes the integrated algorithm by employing a reward and penalty mechanism. Based on the comprehensive fault identification results after each fault diagnosis, the system adjusts the weights of the machine learning model to ensure that the system can flexibly adapt to different fault scenarios. This dynamic adjustment mechanism eliminates the reliance on fixed rules for fault diagnosis, allowing for timely optimization based on actual diagnostic feedback. The fault diagnosis criteria are continuously adjusted as the grid's operating status changes, improving the real-time and accuracy of diagnosis and avoiding the limitations of manually set rules that struggle to cope with complex changes.

[0040] 3. The present invention compares each fault identification result with a set reference threshold to assess the accuracy of fault diagnosis and initiate an optimization process when necessary. This self-learning capability based on real-time data feedback enables the system to continuously improve during actual operation, self-optimizing without human intervention. Over long-term operation, the system can continuously accumulate and feedback learning to optimize the model's weights and judgment criteria, thereby continuously reducing the probability of misjudgments and missed judgments, gradually improving the equipment's fault diagnosis capabilities and enhancing the reliability and safety of power equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 The figure is a schematic diagram of the overall process of the fault-assisted analysis method based on distributed sampling. DETAILED DESCRIPTION

[0042] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. The described embodiments are only part of the embodiments of the present invention, not all of them. Based on the spirit of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of the present invention.

[0043] Example 1

[0044] This embodiment describes a specific implementation of a fault auxiliary analysis method based on distributed sampling proposed by the present invention.

[0045] A fault-assisted diagnosis and analysis method based on distributed sampling includes the following steps:

[0046] Step 1: Multiple distributed sampling devices are installed at various locations on the power grid to synchronously collect real-time grid operational data and waveform signal data, including time-series waveform data such as three-phase voltage signals and three current signals. Data collected by all sampling devices is transmitted to a central processing system, providing raw information for subsequent analysis.

[0047] Step 2: Preprocess the collected data, including denoising and filtering, to ensure data quality. Analyze the preprocessed data in the time and frequency domains to obtain the eigenvalues ​​under the specified dimensions, which serve as the input for the basic machine learning model.

[0048] Specifically, in step 2, multidimensional features are extracted from the processed data. The calculation formula for feature extraction is as follows:

[0049] ;

[0050] in, is the characteristic function, indicating the The eigenvalues ​​of the set dimension, Collection time point The time interval is used for feature normalization processing; for The preprocessed data at time for In scale The wavelet coefficients on For the The scale parameter of the setting dimension; K is the number of setting dimensions.

[0051] Step 3: Use ensemble learning methods to perform preliminary fault identification on the processed data. The operation process is as follows:

[0052] First, the feature values ​​extracted in step 2 are input into multiple machine learning basic models, and then an ensemble learning strategy is used to perform preliminary fault identification.

[0053] Each algorithm makes its own judgment independently, while ensemble learning fuses the results of multiple machine learning models to obtain a preliminary fault identification result.

[0054] Finally, for each fault type, ensemble learning will perform weighted voting or make decisions based on the prediction results of each algorithm to comprehensively determine the type of fault to ensure the synergistic effect of multiple algorithms.

[0055] Specifically, the basic machine learning models that can be input include but are not limited to: decision tree model, SVM model, KNN model; each machine learning model will generate confidence levels for several fault types.

[0056] The extracted feature values Input into multiple machine learning basic models to obtain the The basic machine learning model The confidence level of each fault type is calculated as follows:

[0057] ;

[0058] in, For the The basic machine learning model The confidence level of each fault type, the range is [0,1]; To adjust the nonlinear scaling parameters of the machine learning basic model output; Indicates the The basic machine learning model The characteristic value of the set dimension The predicted output after calculation is, is the number of basic machine learning models.

[0059] For the decision tree based model, The output is based on the characteristics A path through the tree to output a prediction (possibly some type of failure).

[0060] For the SVM basic model, The output of is classified by the support vector boundary and a category prediction is output.

[0061] For the KNN basic model, The output is based on the characteristics The distance between the training data and the neighboring samples is used to make classification predictions.

[0062] Fusion of multiple basic models for fault types Confidence , the comprehensive fault identification result is obtained through the integrated learning strategy, and the calculation formula is as follows:

[0063] ;

[0064] in, is the comprehensive fault identification result, Indicates the The weight of the machine learning basic model, based on the model Historical performance distribution; For the The weight adjustment parameters of the basic machine learning model are used to amplify the contribution of high-performance models. .

[0065] According to the comprehensive fault identification results , judge whether it is a certain type of fault by setting the decision threshold:

[0066] ;

[0067] in, represents the combination of decision thresholds set for the fault type, and ; is the number of decision thresholds to set.

[0068] Each decision threshold defines an evaluation value interval. If the comprehensive identification evaluation value Located in decision threshold Hedi decision threshold The interval between the two corresponds to the fault .

[0069] The decision threshold can be set based on the accuracy of historical classification. Based on the large amount of historical fault sample data, a supervised learning-based training model (such as logistic regression, neural network) is used for training, and the optimal threshold combination is learned based on the impact of different thresholds on the classification results. Among them, historical fault data is obtained through power grid operation logs and manual annotation.

[0070] As a prominent substantive feature of the present invention and one of the significant advances it brings to the prior art, the present invention combines the advantages of multiple algorithms and enhances the robustness of fault identification through an integrated learning method that combines multi-feature fusion analysis with multiple machine learning algorithms.

[0071] Step 4: Calculate an evaluation value for the comprehensive fault identification results, reward correct answers and penalize incorrect answers. Through this feedback mechanism, the judgment logic of integrated learning is continuously optimized, further improving the accuracy of fault identification. At the same time, the algorithm is more flexible and can quickly adapt to different power grid environments.

[0072] Evaluation value of comprehensive fault identification results The calculation formula is as follows:

[0073] ;

[0074] in, is the comprehensive fault identification result; The real result is determined by continuity mapping based on the physical characteristics of historical fault data or manually annotated fault data.

[0075] Specifically, the feedback mechanism includes: designing reward and penalty rules for each recognition result. The reward and penalty are based on the evaluation value. The value size and its relative to the set reference threshold Deviation of reference threshold The setting is based on historical classification accuracy. The following is the specific plan:

[0076] Evaluation value The calculation formula represents the accuracy, i.e. and The smaller the deviation between, the higher the accuracy; the larger the deviation, the closer the accuracy is to 0; if Infinitely close to , then the deviation is almost 0, and the accuracy is the highest, indicating that the prediction is correct; if And if the deviation is too large, the accuracy is reduced, indicating that the prediction is wrong;

[0077] If S> , rewards are given without weight adjustment, and the feedback optimization ends; if S< , triggering penalties and adjusting weights.

[0078] According to the penalty rules, the weights of each machine learning basic model are readjusted. The weight adjustment formula is:

[0079] ;

[0080] in, and Respectively The basic machine learning model Round and The weight of the round; is the learning rate, which is used to control the magnitude of weight adjustment; is the set reference threshold.

[0081] When the predicted evaluation value of the equipment When the set decision threshold is reached, the system will trigger an early warning. The specific data recorded includes but is not limited to: device status, predicted fault type, specific time when the early warning is triggered, other device parameters and environmental information.

[0082] When the maintenance personnel arrive at the equipment site, they will input the preliminary inspection results of the equipment through the mobile terminal or other input devices; at this time, the maintenance personnel's judgment results will be recorded as the manually entered inspection evaluation value and the predicted evaluation value of the system Make a comparison;

[0083] If the inspection evaluation value is consistent with or highly matches the system prediction evaluation value, the warning is considered accurate, the fault identification result is directly adopted, and corresponding emergency or treatment measures are taken; if there is a large difference, it indicates a false alarm, and the entire process needs to be traced back, and further manual inspection is carried out to confirm the actual status of the equipment and adjust the data collection.

[0084] Example 2

[0085] This embodiment provides a fault auxiliary analysis system based on distributed sampling, which runs the above-mentioned fault auxiliary analysis method based on distributed sampling. Figure 1 , Figure 1 This is a schematic diagram of the overall process of the fault auxiliary analysis method based on distributed sampling proposed by the present invention. This system includes:

[0086] Distributed data acquisition module, used for real-time distributed collection of power grid data;

[0087] Data preprocessing module, used to preprocess the collected data, including denoising and filtering;

[0088] Fault feature extraction module, extracts feature values ​​that may represent faults from processed data;

[0089] The preliminary fault identification module uses multiple different machine learning models to perform preliminary fault identification based on feature values ​​and analyze potential fault modes;

[0090] The comprehensive judgment module integrates the preliminary fault identification results output by all machine learning basic models to obtain comprehensive fault identification results;

[0091] The fault scoring and feedback module evaluates the comprehensive fault identification results and optimizes them based on the reward and penalty mechanism to obtain the best fault identification results.

[0092] The monitoring and management module is used to monitor and manage the entire system in real time to ensure the normal operation of the entire process of data collection, analysis, and identification.

[0093] To validate the present invention, an ensemble learning method was used to diagnose transformer faults in a power grid. The following is a comparison of the accuracy, false alarm rate, and other indicators of the ensemble learning method and the single model method based on experimental data.

[0094] Experimental dataset:

[0095] Training set: 1000 data points; fault data: 200 (including transformer short circuit, overload, ground fault, etc.); normal data: 800. Basic learning algorithm models: Decision Tree (CART), Support Vector Machine (SVM), K-Nearest Neighbor (KNN).

[0096] Table 1 Comparison of ensemble learning and single model accuracy and other performance indicators

[0097]

[0098] As can be seen from the table above, ensemble learning achieved an accuracy of 92%, significantly higher than single models (e.g., 90% for decision trees, 88% for support vector machines, and 85% for k-NN models). By combining the prediction results of multiple models, ensemble learning effectively reduces both missed and false positives.

[0099] False Alarm Rate: The ensemble learning model achieved a false alarm rate of 2.5%, significantly lower than that of a single model (decision tree: 3.75%, SVM: 5%, KNN: 6.25%). This indicates that ensemble learning can better distinguish between normal and faulty data, reducing false alarms.

[0100] False negative rate: The false negative rate of ensemble learning is 5%, which is a significant improvement compared to other models (decision tree 10%, SVM 12%, KNN 12.5%). This reduction in false negative rate means that ensemble learning can more accurately identify fault data.

[0101] F1 value: The F1 value is the harmonic mean of precision and recall. The F1 value of ensemble learning is 0.90, higher than the 0.88 (decision tree), 0.84 (SVM), and 0.82 (KNN) values ​​of other single models. This shows that ensemble learning performs better in balancing precision and recall.

[0102] Table 2 Comparison of accuracy data after optimization

[0103]

[0104] After 20 rounds of self-learning, the accuracy of the ensemble learning method increased by 7%, from the initial 85% to 92%; the false alarm rate and missed alarm rate were also significantly improved, with the false alarm rate reduced from 3.75% to 2.5%, and the missed alarm rate reduced from 10% to 5%.

[0105] The present disclosure may be a system, method and / or computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for causing a processor to implement various aspects of the present disclosure.

[0106] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the scope of protection of the claims of the present invention.

Claims

1. A fault auxiliary analysis method based on distributed sampling, characterized in that: The following steps are involved: Real-time distributed sampling of grid data at multiple locations on the grid; Preprocess the sampled data and extract the feature values ​​of the set dimension from the preprocessed data; The extracted feature values ​​are input into different machine learning basic models. Each machine learning basic model makes independent predictions to obtain the confidence level of each machine learning basic model for the fault type, which is expressed as the following formula: ； in, For the The basic machine learning model The confidence level of each fault type, the range is [0,1]; To adjust the nonlinear scaling parameters of the machine learning basic model output; Indicates the The basic machine learning model The characteristic value of the set dimension The predicted output after calculation is, is the number of basic machine learning models; Use ensemble learning algorithms to fuse the prediction results of all machine learning basic models to obtain a comprehensive fault identification result ; ； Indicates the The weights of the machine learning base model, For the The weight adjustment parameters of the basic machine learning model; According to the comprehensive fault identification results , judged by the set decision threshold The corresponding fault; the decision threshold is set by using a supervised learning-based training model to train based on historical fault sample data to obtain a threshold combination; if Located in decision threshold Hedi decision threshold The interval between the two corresponds to the fault ; Design a reward and penalty mechanism for each recognition result. The reward and penalty are determined based on the evaluation value. and its relative to the set reference threshold deviations; Calculate the evaluation value of the comprehensive fault identification results and optimize the ensemble learning algorithm based on the reward and penalty mechanism; if the evaluation value S> , rewards are given without weight adjustment, and the feedback optimization ends; if S< , triggering the penalty, and readjusting the weights of each machine learning basic model according to the penalty rules to obtain the optimal comprehensive fault identification result.

2. The fault-assisted analysis method based on distributed sampling according to claim 1, characterized in that: Extract the eigenvalues ​​of the set dimension from the preprocessed data. The extraction formula is: ； in, is the characteristic function, indicating the The eigenvalues ​​of the set dimension, Collection time point The time interval is used for feature normalization processing; for The preprocessed data at time for In scale The wavelet coefficients on For the The scale parameter of the setting dimension; K is the number of setting dimensions.

3. The fault-assisted analysis method based on distributed sampling according to claim 1, characterized in that: According to the comprehensive fault identification results , judged by the set decision threshold The corresponding fault is expressed by the following formula: ； in, represents the combination of decision thresholds set for the fault type, and ; is the number of decision thresholds to set.

4. The fault-assisted analysis method based on distributed sampling according to claim 1, characterized in that: By formula Obtaining an evaluation value of the comprehensive fault identification result; wherein, The comprehensive fault identification result The assessed value of For real results.

5. The fault-assisted analysis method based on distributed sampling according to claim 1, characterized in that: Optimizing the ensemble learning algorithm using a reward and penalty mechanism includes: If the evaluation value of the comprehensive fault identification result Greater than the set reference threshold , then rewards are given, no weight adjustment is made, and the optimization ends; If the evaluation value of the comprehensive fault identification result Less than the set reference threshold , then the penalty is triggered and the weight of each machine learning basic model in the integrated learning algorithm is adjusted.

6. The fault-assisted analysis method based on distributed sampling according to claim 5, characterized in that: According to the penalty rules, the weights of each machine learning basic model are readjusted. The weight adjustment formula is: ； in, and Respectively The basic machine learning model Round and The weight of the round; is the learning rate, which is used to control the magnitude of weight adjustment; is the set reference threshold.

7. A fault assisted analysis system based on distributed sampling, running a fault assisted analysis method based on distributed sampling according to any one of claims 1 to 6, characterized in that: include: Distributed data acquisition module, used for real-time distributed collection of power grid data; Data preprocessing module, used to preprocess the collected data; A fault identification module is used to identify faults based on preprocessed data; The optimization module is used to optimize the fault identification process to obtain the best fault identification result.

Images