Power transmission line fault identification method and device based on double enhancement and dynamic trajectory discrimination

CN122508382BActive Publication Date: 2026-09-18ZHEJIANG UNIV OF TECH +1
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202610995221.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-07-06
Publication Date
2026-09-18
Estimated Expiration
2046-07-06

AI Technical Summary

Technical Problem

[0005]为了解决现有输电线路智能故障辨识方法中故障样本稀缺与类别不均衡、多窗口特征工程体系不完善、难以有效保留并利用输电线路故障的时间相关性与动态演化信息、以及缺乏基于动态轨迹的级联判别机制等问题,本发明提出了一种基于双重增强与动态轨迹判别的输电线路故障辨识方法及装置

Benefits of technology

[0042] 1. This invention proposes a method, device, system, and storage medium for identifying transmission line faults based on dual enhancement and dynamic trajectory discrimination. At the data level, by constructing a dual enhancement mechanism that includes IMF-Mixing modal cross-recombination and intra-class perturbation/operating condition perturbation enhancement, data enhancement is achieved at both the original waveform level and the modal component level, effectively solving the problem of insufficient model generalization ability caused by the scarcity of transmission line fault samples and the extreme imbalance of categories.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122508382B_ABST
    Figure CN122508382B_ABST
Patent Text Reader

Abstract

The application discloses a power transmission line fault identification method and device based on double enhancement and dynamic trajectory discrimination. First, the original power transmission line fault recording data is collected and basic preprocessing is performed, and the standardized fault original sequence is output. Then, the double enhancement strategy is implemented on the standardized sequence, including in-class perturbation and working condition disturbance enhancement and mode cross-reorganization enhancement. The enhanced data set is subjected to multi-window slicing, and multi-dimensional features are extracted to form a feature vector set to retain the evolution information of the fault in different time windows. The multi-dimensional feature vector set is input into the CatBoost model for multi-class fault preliminary judgment, and the initial category prediction and confidence are output. The low confidence samples are subjected to dynamic trajectory discriminator cascade correction, and the fault cross-window evolution path is used to improve the identification ability of the easily confused faults. Finally, the final fault identification result after cascade discrimination is output. The method significantly improves the identification accuracy of the power transmission line fault.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of fault diagnosis technology for smart grids and power systems, and specifically relates to a method and device for identifying transmission line faults based on dual enhancement and dynamic trajectory discrimination. Background Technology

[0002] With the expansion of power grids and the rapid development of ultra-high voltage (UHV) transmission technology, the structural forms and operating conditions of transmission lines are becoming increasingly complex. State perception technologies based on fault recording devices and online monitoring systems have been widely applied. When a fault occurs, quickly and accurately distinguishing different fault types, such as external force damage, lightning strikes, and foreign objects, while ensuring safe and stable operation, is a crucial prerequisite for guiding fault location, maintenance strategy formulation, and operational mode adjustment. Traditional relay protection and fault recording analysis mainly rely on phasor measurements, setting values, and the experience of maintenance personnel. For complex scenarios with changing operating conditions and significant disturbances, there is often a risk of misjudgment and missed judgment, making it difficult to meet the needs of next-generation intelligent transmission lines for "refined fault identification." Furthermore, transmission line faults are not instantaneous phenomena at a single static moment, but rather processes with significant time correlation and dynamic evolution characteristics. From the transient abrupt changes before and after fault triggering, energy transfer, and subsequent attenuation and recovery, the voltage, current, zero-sequence quantities, and related auxiliary states of different fault types typically exhibit different change paths within a continuous time window.

[0003] In recent years, intelligent fault diagnosis technology has gradually emerged in the field of power transmission lines. It constructs multi-dimensional feature or spectral input models by analyzing recorded voltage and current waveforms to achieve automatic fault type identification. However, truly implementing these methods in real-world power grid applications still faces several challenges, mainly in the following aspects: a) Scarcity of fault samples and extremely uneven distribution of categories: In actual engineering, common fault samples such as grounding short circuits are relatively abundant, while scenarios such as external force damage, lightning strikes, and foreign object entrapment occur very infrequently and are affected by weather, channel environment, etc., resulting in inconsistent waveform quality. Existing methods often use simple oversampling or adding Gaussian noise to augment the data, but such enhancements have a weak correlation with electrical and physical mechanisms, easily introducing distorted waveforms, leading to insufficient generalization performance of the model and difficulty in supporting refined identification in engineering scenarios. b) Incomplete multi-window feature engineering system, making it difficult to characterize the fault evolution process: The discrimination information of power transmission line faults is not only reflected in the amplitude or spectral differences at a single moment, but also in the continuous change relationship between adjacent time windows, the transmission law of abrupt changes, and the sequence of transient responses. Existing research suffers from significant limitations in feature construction due to its reliance on a single dimension and single window: it either focuses on a few time-domain and frequency-domain statistical indicators within a single time window or calculates a small number of global features for the entire waveform. A systematic feature engineering framework oriented towards "multi-window, multi-dimensional, and multi-physical quantities" has not yet been formed, resulting in the compression of cross-window dynamic correlation information into discrete static representations. c) Lack of a cascaded discrimination mechanism based on dynamic trajectories: Most existing intelligent fault identification schemes adopt a one-step single-model decision-making mode, lacking refined utilization of the confidence level and uncertainty of the model output. More importantly, for easily confused fault types, relying solely on single-window features or static fusion probabilities often fails to reliably distinguish their evolutionary differences during triggering, development, and decay processes. Existing methods rarely perform refined fault type discrimination from the perspective of the cross-window evolution trajectory of the same fault event, resulting in a lack of reliable secondary discrimination basis for low-confidence samples. This makes it difficult for existing methods to further improve the identification capability and decision credibility of key confused pairs beyond overall accuracy, limiting their application value in actual power grid operation and maintenance decisions.

[0004] Therefore, an intelligent technical solution for transmission line fault identification is designed. This solution efficiently integrates multi-source information on fault evolution through a dual data augmentation strategy and multi-window, multi-dimensional feature engineering. It retains and utilizes continuous evolution information of the fault across multiple time windows at the event level, and further performs cascaded discrimination based on confidence level and dynamic trajectory, achieving high-precision identification of various fault types under complex operating conditions. While meeting engineering application requirements, this solution provides a reliable basis for power grid operation scheduling and maintenance decisions, and has significant application value and engineering significance for promoting the development of next-generation intelligent transmission line operation and maintenance technology. Summary of the Invention

[0005] To address the problems in existing intelligent fault identification methods for transmission lines, such as scarce fault samples and imbalanced categories, imperfect multi-window feature engineering system, difficulty in effectively preserving and utilizing the temporal correlation and dynamic evolution information of transmission line faults, and lack of a cascaded discrimination mechanism based on dynamic trajectory, this invention proposes a transmission line fault identification method and device based on dual enhancement and dynamic trajectory discrimination.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A method for identifying transmission line faults based on dual enhancement and dynamic trajectory discrimination includes the following steps:

[0008] S1: Collect raw transmission line fault recording data and perform basic preprocessing to output standardized raw fault sequences;

[0009] S2: Perform a dual enhancement strategy on the standardized original fault sequence to construct an enhanced multi-source heterogeneous waveform dataset (i.e., a high-fidelity multi-source heterogeneous waveform dataset).

[0010] S3: Perform multi-window slicing and multi-dimensional feature extraction on the enhanced multi-source heterogeneous waveform dataset to construct a multi-dimensional feature vector set;

[0011] S4: Input the multi-dimensional feature vector set into the classification model to perform preliminary judgment of multiple types of faults, and output the confidence score of the sample;

[0012] S5: Input samples with confidence levels below a preset threshold into the dynamic trajectory discriminator, perform discrimination correction, and obtain the transmission line fault identification results.

[0013] In this invention, firstly, raw transmission line fault recording data is acquired and basic preprocessing is performed to output a standardized original fault sequence. Then, a dual enhancement strategy is implemented on the standardized sequence, including intra-class perturbation and operating condition perturbation enhancement, and mode cross-recombination enhancement (IMF-Mixing). The enhanced dataset is then sliced ​​into multiple windows, and multi-dimensional features are extracted and concatenated to form a feature vector set to preserve the evolution information of the fault in different time windows. The multi-dimensional feature vector set is input into a CatBoost model for initial multi-class fault identification, outputting initial class predictions and confidence scores. Low-confidence samples are corrected using a cascaded dynamic trajectory discriminator, utilizing the fault's cross-window evolution path to improve the identification capability of easily confused faults. Finally, the final fault identification result after cascaded discrimination is output. This invention significantly improves the accuracy of transmission line fault identification, providing reliable technical support for power system fault diagnosis.

[0014] In step S1, the basic preprocessing includes data parsing, quality screening, fault event windowing, resampling, and DC detrending processing.

[0015] The fault recording devices deployed in the lines and substations automatically collect full-channel recording files, perform parsing, filtering, windowing and resampling processing on the raw recording data, and output standardized original fault sequences.

[0016] In step S2, a dual enhancement strategy is applied to the standardized original fault sequence to construct an enhanced multi-source heterogeneous waveform dataset. Specifically, this includes: enhancing the standardized original fault sequence with intra-class perturbations and operating condition disturbances, as well as modal cross-recombination enhancement (IMF-Mixing), to obtain the enhanced multi-source heterogeneous waveform dataset. The intra-class perturbation and operating condition disturbance enhancement module performs enhancement operator processing on the standardized original fault sequence. These enhancement operators include time shift, amplitude scaling, frequency drift, voltage sag / surge, current amplitude scaling, noise superposition, DC bias, and linear trend superposition, outputting a disturbance enhancement sequence that simulates the actual operating environment and measurement anomalies. More preferably, the intra-class perturbations include two or three of time shift, amplitude scaling, and noise superposition; the operating condition disturbances include two, three, or four of voltage sag / surge, current amplitude scaling, frequency drift, and DC bias.

[0017] In step S2, the modal cross-recombination enhancement includes: performing variational mode decomposition on the sequence to obtain intrinsic mode components (IMFs), and then cross-recombining the IMFs according to their frequency components from largest to smallest before reconstructing the sequence. More preferably, the modal cross-recombination enhancement module (IMF-Mixing) performs variational mode decomposition on the normalized fault original sequence to obtain multi-order IMFs, cross-recombines the low-frequency IMFs of the target sample with the high-frequency IMFs of similar samples, applies a softening gain coefficient to the recombined high-frequency components, and then reconstructs the sequence, outputting the reconstructed enhanced sequence.

[0018] In step S3, the multi-window includes 4 to 8 overlapping time windows regarding the fault development state; the calculation process for multi-window slicing of the high-fidelity multi-source heterogeneous waveform dataset is expressed by the following formula:

[0019]

[0020] in, This represents the input standardized fault waveform sequence; Number the time windows; Relative index within the window ( , (where the window length is the number of sampling points). Indicates from sequence According to the first The subsequence within the window obtained by capturing the nth window, whose nth... Each sampling point satisfies ; For the first A window in the sequence The starting sampling point index on, For the fault trigger point in the sequence The corresponding sampling point index, For the first Each window is relative to the fault trigger point. The initial offset (in units of sampling points), therefore ; For the first The end boundary index of each window, satisfying This allows for multi-window overlapping slicing based on the fault point, providing a unified windowed input for subsequent multi-dimensional feature extraction.

[0021] In step S3, the multi-dimensional features include time-domain features, frequency-domain features, time-frequency-domain features, statistical morphological features, and auxiliary metadata features. The multi-dimensional feature extraction module extracts time-domain, frequency-domain, time-frequency-domain, and statistical morphological features within each window. These multi-dimensional features include, but are not limited to: current impulse intensity, zero-sequence component level, energy and mutation rate, harmonic ratio, wavelet energy distribution, zero-sequence pulse characteristics, fault angle, correlation features, and auxiliary metadata features such as time context and switching status.

[0022] In step S3, the multi-source feature concatenation module concatenates window-level feature vectors extracted from different time window segments of the same event to form a multi-window, multi-dimensional feature vector set for the event. This vector set is not only used for subsequent CatBoost multi-class fault initial judgment, but also for constructing fault evolution trajectories in chronological order, providing window-level basic input for dynamic trajectory cascade discrimination.

[0023] In step S4, the classification model adopts the CatBoost classification model and uses a window weight weighted fusion strategy. The CatBoost multi-class fault initial judgment process is used to input a multi-window, multi-dimensional feature vector set into the CatBoost model and obtain the event-level prediction result through the window weight fusion strategy. Its calculation process is represented by the following formula:

[0024]

[0025] in This indicates that the CatBoost model is effective for events. The Window feature vectors Predicted as the first The probability of a type of failure; Indicates an event The Feature vectors of a window; An index representing a fault event; Indicates the window index. ; Indicates the fault category index, ; Indicates an event In the Event-level fusion probability over class-related faults; For the first The weight coefficients of each window satisfy the following conditions: + +...+ =1; Total number of windows; Total number of fault categories; For CatBoost model events The initial category prediction results; For the event The confidence level is the maximum value of the event-level fusion probability.

[0026] In step S4, the dynamic trajectory discriminator is triggered based on the confidence level of the initial CatBoost judgment result. The triggering condition is expressed by the following formula:

[0027]

[0028] in, For the event Confidence level; Index for fault events; The confidence threshold; For CatBoost model events The initial category prediction results; For the event The final category prediction result;

[0029] In step S5, the dynamic trajectory discriminator cascade process includes two stages: trajectory construction and standardization, and energy calculation and cascade discrimination. It is used to re-characterize and re-discriminate the evolution path of low-confidence samples across the window time dimension.

[0030] The first stage: trajectory construction and standardization, used to aggregate multi-window features by event, construct 4×F dynamic trajectories, and perform standardization. Its calculation process is represented by the following formula:

[0031]

[0032] in, This represents the trajectory discrimination feature set, which contains F trajectory discrimination feature extraction functions; Indicates the first A trajectory discrimination feature extraction function; ; Indicates the fault event index; Indicates window index ; Represents trajectory feature index ; Indicates an event In the The first window extracted Each trajectory discrimination feature value; for all and After arranging the eigenvalues, an event can be formed. of Dynamic trajectory matrix; Represents the standardized trajectory features; and They represent the first The mean and standard deviation of each trajectory discrimination feature on the training set.

[0033] The second stage: energy calculation and cascade discrimination, used to calculate trajectory energy and energy difference, and to make discrimination corrections based on thresholds and direction rules. The calculation process is represented by the following formula:

[0034]

[0035] when and season Otherwise, let .

[0036] in, Indicates the fault category index; Indicates an event Relative to category The final trajectory energy; , and Representing events respectively Relative to category Geometric distance, morphological similarity distance, and physical constraint distance; , and These represent the weight coefficients of the geometric distance term, the morphological similarity distance term, and the physical constraint distance term, respectively. Indicates the category determined by CatBoost. In addition, the alternative categories that minimize the final trajectory energy; This represents the final trajectory energy corresponding to the initial classification. This represents the final trajectory energy corresponding to the candidate category; This represents the energy difference between the initial classification and the candidate classification. Indicates the energy difference threshold; Indicates an event The result of the correction direction rule is set to 1 if the correction direction rule is satisfied, and 0 otherwise. This indicates that CatBoost handles events. The initial category prediction results; Indicates an event The final category prediction results.

[0037] In step S5, samples with confidence levels below a preset threshold are input into the dynamic trajectory discriminator for discrimination correction to obtain the transmission line fault identification result. Specifically, this includes: constructing a dynamic trajectory matrix based on multi-window features of events for samples with confidence levels below the preset threshold and standardizing it; calculating the energy value and energy difference of the trajectory relative to each category of pattern prototypes; and performing discrimination correction according to threshold and direction rules to obtain the transmission line fault identification result. The calculation of the energy value and energy difference of the trajectory relative to each category of pattern prototypes uses energy calculation, which includes a weighted combination of geometric distance, morphological similarity, and physical constraint distance.

[0038] The fault identification result of the transmission line is the fault identification output after unifying the CatBoost initial judgment result and the dynamic trajectory discrimination correction result. The dynamic trajectory cascade discrimination focuses on using the time correlation and evolution characteristics of fault events across windows to correct low confidence samples.

[0039] A transmission line fault identification device includes a processor and a memory. The memory stores a computer program. When the processor executes the computer program, it implements the transmission line fault identification method based on dual enhancement and dynamic trajectory discrimination.

[0040] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned transmission line fault identification method based on dual enhancement and dynamic trajectory discrimination.

[0041] Compared with existing methods, the beneficial effects of the present invention are mainly reflected in:

[0042] 1. This invention proposes a method, device, system, and storage medium for identifying transmission line faults based on dual enhancement and dynamic trajectory discrimination. At the data level, by constructing a dual enhancement mechanism that includes IMF-Mixing modal cross-recombination and intra-class perturbation / operating condition perturbation enhancement, data enhancement is achieved at both the original waveform level and the modal component level, effectively solving the problem of insufficient model generalization ability caused by the scarcity of transmission line fault samples and the extreme imbalance of categories.

[0043] 2. In terms of feature extraction design, a multi-window slicing and multi-dimensional feature extraction mechanism was introduced. This design segments the recorded waveform sequence around the fault trigger point, allowing the transient response, energy distribution, statistical morphology, and auxiliary operational information such as time context and switching status of the same fault event in different time windows to be preserved synchronously. Compared with single-window or whole-segment static representation, this mechanism can more completely describe the temporal correlation and dynamic evolution characteristics of transmission line faults, providing a reliable foundation for subsequent construction of fault dynamic trajectories.

[0044] 3. In the classification and correction phase, a CatBoost initial judgment and dynamic trajectory cascade correction mechanism was designed. The CatBoost multi-class classifier first performs event-level fusion of multi-window, multi-dimensional features, and the dynamic trajectory discriminator then performs secondary discrimination for low-confidence samples by utilizing the difference between their cross-window feature change paths and the class prototype trajectories. This mechanism is not a simple superposition of existing classification results, but rather aims to compensate for the insufficient utilization of continuous fault evolution information by static fusion results, thereby improving the robustness of discrimination and the reliability of decision-making for easily confused fault pairs.

[0045] 4. In model optimization and decision-making applications, a confidence level splitting and cascade correction strategy is adopted to guide model optimization. By directly outputting high-confidence samples and correcting the trajectory of low-confidence samples, the overall identification accuracy is significantly improved. The final fault identification results can accurately characterize the current fault type and status, providing a reliable decision-making reference for intelligent fault diagnosis and emergency handling in transmission line operation and maintenance systems.

[0046] This invention achieves a superior balance between model accuracy and robustness. While ensuring identification accuracy, it effectively improves the model's ability to distinguish easily confused faults, making it well-suited for practical application scenarios where transmission line fault samples are scarce and the categories are unbalanced. It has strong practical application value. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of the overall process in an embodiment of the present invention.

[0048] Figure 2 This is a waveform comparison example of an external force-induced damage fault sample after being enhanced by intraclass perturbation in a 3I0 (3 times zero-sequence current) channel.

[0049] Figure 3 This is a waveform comparison example of an external force-induced damage fault sample after being enhanced by operating condition disturbance under the 3I0 (3 times zero-sequence current) channel.

[0050] Figure 4 This is a waveform comparison example of a foreign object fault sample after being enhanced by intraclass perturbation in the 3I0 (3 times zero-sequence current) channel.

[0051] Figure 5 This is a waveform comparison example of a foreign object fault sample after being enhanced by operating condition disturbance under the 3I0 (3 times zero sequence current) channel.

[0052] Figure 6 This is a waveform comparison example of a lightning strike fault sample after being enhanced by intra-class perturbation in the 3I0 (3 times zero-sequence current) channel.

[0053] Figure 7 This is a waveform comparison example of a lightning strike fault sample after being enhanced by operating condition disturbance under the 3I0 (3 times zero-sequence current) channel.

[0054] Figure 8 This is an example of a background event waveform (providing a low-frequency power frequency shape) obtained by decomposing a foreign object fault sample A using IMF-Mixing.

[0055] Figure 9 This is an example of a low-frequency background event time-frequency diagram derived from foreign object fault sample A after IMF-Mixing.

[0056] Figure 10 Example of a characteristic event waveform diagram (providing high-frequency fault shape) obtained by IMF-Mixing decomposing foreign object fault sample B.

[0057] Figure 11 This is an example of a time-frequency diagram of high-frequency characteristic events obtained by decomposing foreign object fault sample B using IMF-Mixing.

[0058] Figure 12 A visualization example of the t-SNE distribution of the original sample and the IMF-Mixing enhanced sample of the foreign object class.

[0059] Figure 13 Example of background event waveform (providing low-frequency power frequency shape) obtained by IMF-Mixing decomposition of lightning strike fault sample A.

[0060] Figure 14 This is an example of a low-frequency background event time-frequency diagram derived from lightning strike fault sample A after IMF-Mixing.

[0061] Figure 15 Example of characteristic event waveforms (providing high-frequency fault shapes) derived from lightning strike fault sample B after IMF-Mixing.

[0062] Figure 16 This is an example of a time-frequency diagram of high-frequency characteristic events extracted from lightning strike fault sample B after IMF-Mixing.

[0063] Figure 17A visualization example of the t-SNE distribution for original lightning strike samples and IMF-Mixing-enhanced samples.

[0064] Figure 18 This is an example of a background event waveform (providing a low-frequency power frequency shape) obtained by IMF-Mixing decomposing a fault sample A caused by external force damage.

[0065] Figure 19 This is an example of a low-frequency background event time-frequency diagram obtained by IMF-Mixing decomposing a fault sample A caused by external force damage.

[0066] Figure 20 This is an example of a characteristic event waveform diagram (providing high-frequency fault shapes) obtained by IMF-Mixing decomposing a fault sample B of the external force damage type.

[0067] Figure 21 This is an example of a time-frequency diagram of high-frequency characteristic events obtained by IMF-Mixing decomposing a fault sample B of the external force damage type.

[0068] Figure 22 This provides a visualization example of the t-SNE distribution for original samples and IMF-Mixing-enhanced samples from the external force damage category.

[0069] Figure 23 This is a flowchart of the preprocessing and dual enhancement strategy in an embodiment of the present invention.

[0070] Figure 24 This is a detailed flowchart of the implementation of Modal Cross-Recombination Enhancement (IMF-Mixing) in an embodiment of the present invention.

[0071] Figure 25 This is a detailed flowchart illustrating the implementation of intra-class perturbations and operational condition perturbations in an embodiment of the present invention.

[0072] Figure 26 This is a flowchart of the multi-window slicing and multi-dimensional feature extraction process in an embodiment of the present invention.

[0073] Figure 27 This is a flowchart of CatBoost's initial fault assessment and confidence level routing in an embodiment of the present invention.

[0074] Figure 28 This is a flowchart of the cascaded correction process for the dynamic trajectory discriminator in an embodiment of the present invention.

[0075] Figure 29 This is a comparison diagram of the confusion matrix between CatBoost initial judgment and dynamic trajectory concatenation judgment in an embodiment of the present invention.

[0076] Figure 30 This is a radar chart showing the accuracy improvement effect of each fault category in the embodiments of the present invention. Detailed Implementation

[0077] To make the objectives, design concepts, and technical solutions of the embodiments of the present invention clearer, the present invention will be further described below with reference to the accompanying drawings.

[0078] Reference Figure 1 This invention proposes a method and apparatus for identifying transmission line faults based on dual enhancement and dynamic trajectory discrimination. The method includes the following steps:

[0079] Step S1: Collect raw transmission line fault recording data and perform basic preprocessing.

[0080] Reference Figure 23 In a preferred embodiment of the present invention, step S1 is used to acquire raw waveform data for transmission line fault monitoring, and to normalize it into a standardized raw fault sequence through basic preprocessing, providing a unified and reliable input for subsequent dual enhancement and multi-window feature engineering. This step specifically includes the following sub-steps:

[0081] Step S101: Acquisition and Analysis of Raw Fault Recording Data

[0082] To construct a heterogeneous fault data input for transmission line fault identification, the original transmission line fault recording data is first automatically collected by a fault recording device in the actual power grid, and the recording files are then parsed, specifically including:

[0083] Data acquisition: Fault recording devices are installed in the transmission lines and their substations. When the line experiences abnormal operating conditions such as short circuit, grounding, or open circuit and triggers the protection or monitoring threshold, the recording device automatically starts and records a full-channel recording file containing multi-channel measurement information such as current and voltage.

[0084] Data analysis: The acquired waveform files are parsed to read and unify the following key information: sampling rate, number of sampling points, recording duration; channel name, phase and physical quantity type; fault trigger time marker, protection action flag and other auxiliary information.

[0085] Step S102: Data quality screening and outlier removal

[0086] Reference Figure 23 To ensure the reliability of the modeling data, a quality screening process is performed on the analyzed waveform data, removing samples that do not meet the quality requirements. This step sequentially checks the waveform files for analytical integrity, channel completeness, and coverage. This integrated check ensures that all samples meet the requirements of readable analysis, complete data, complete channels, and sufficient coverage.

[0087] Step S103: Windowing and Time Alignment of Fault Events

[0088] Reference Figure 23 To standardize the representation of fault samples across time scales, event-level windowing and time-axis alignment are performed on valid waveform recordings. This step centers on the fault trigger point, extracts an event-level time window covering the entire process, and ensures the comparability of different fault samples across time scales by using a uniform time window length, thus providing a foundation for subsequent multi-window slicing and dynamic trajectory construction.

[0089] Step S104: Resampling and Amplitude Normalization

[0090] Reference Figure 23 This step maps sequences from different sampling rates to a standard sampling rate and performs unit unification and DC detrending processing on the measurements of each channel. Through this processing, a standardized fault sequence with uniformity across the time axis, sampling rate, channel set, and amplitude scale is finally obtained, providing a stable and reliable input for the subsequent dual enhancement strategy.

[0091] Step S2: Perform a dual enhancement strategy on the standardized original fault sequences to construct a high-fidelity multi-source heterogeneous waveform dataset.

[0092] Based on the standardized original fault sequence obtained in step S1, referring to Figure 23 The process involves performing Modal Cross-Recombination Enhancement (IMF-Mixing) and Intra-class Perturbation / Operating Condition Perturbation Enhancement, respectively, to generate complementary enhanced sequences. These enhanced sequences are then aggregated with the original sequences to construct a high-fidelity multi-source heterogeneous waveform dataset. This step alleviates the problems of scarce fault samples and class imbalance, providing a more representative data foundation for subsequent multi-window slicing, multi-dimensional feature extraction, and dynamic trajectory cascade discrimination. Specifically, it includes the following sub-steps:

[0093] Step S201: Modal cross-recombination enhancement (IMF-Mixing)

[0094] In step S201, refer to Figure 24 The normalized fault sequence is enhanced by mode crossover and recombination based on variational mode decomposition (VMD). For any normalized fault sequence... First, it is decomposed into several intrinsic mode components arranged in descending order of frequency using VMD. , ( =0 represents the highest frequency. = -1 is the lowest frequency), and its reconstruction form can be expressed as:

[0095]

[0096] in, This represents the original sequence of standardized faults at time [time]. The possible values ​​of ; The first mode obtained by variational mode decomposition is... Each intrinsic mode component at time... The possible values ​​of ; Indicates modal index, ; This indicates the number of modes obtained from VMD decomposition; This represents the time index. Preferably, the lowest frequency mode mainly carries the fault background and slowly varying power frequency information, while the high / medium frequency modes mainly carry the fault texture and transient details.

[0097] Select target samples from the same fault category sample set. and similar reference samples VMD decomposition was performed to obtain and The lowest frequency mode of A is used as the background component, and the remaining high / mid frequency modes of B are used as detail components, with a softening gain coefficient introduced. Enhanced sequences are constructed through modal crossover and recombination. The calculation process can be expressed as follows:

[0098]

[0099] in, This indicates that the enhanced sequence obtained through modal crossover and recombination is at time [time value missing]. The possible values ​​of ; Indicates target sample The Each modal component at time... The possible values ​​of ; Indicates reference sample The Each modal component at time... The possible values ​​of ; Represents the low-frequency modal index set; Represents the set of high-frequency and mid-frequency mode indices; Indicates modal index; Represents the softening gain coefficient, satisfying It is used to control the intensity of the introduced high-frequency components and avoid excessive amplification of noise or abnormal spikes; This represents the time index. Preferably, IMF-Mixing is applied to channels U0 and 3I0, while the target sample structure remains unchanged in other channels. By preserving the low-frequency background of the target sample and introducing high-frequency details from similar samples, a reconstructed enhancement sequence with rich detail variations can be generated without altering the physical nature of the fault. All sequences generated by IMF-Mixing together constitute the reconstructed enhancement sequence set.

[0100] Reference Figures 8 to 11 , Figures 13 to 16 as well as Figures 18 to 21 It is evident that, in the three types of faults—foreign objects, lightning strikes, and external force damage—IMF-Mixing enhancement is based on modal cross-recombination of two similar fault samples. One similar sample provides low-frequency background components and main energy distribution contours, while the other similar sample provides high-frequency fault feature components and time-frequency texture details. This allows the enhanced sample to obtain richer transient detail information while maintaining the original fault's main structural features.

[0101] Reference Figure 12 , Figure 17 and Figure 22 As can be seen, after IMF-Mixing enhancement, all types of enhanced samples form a reasonable extended region around the original sample in the t-SNE feature space, exhibiting a distribution characteristic of coexistence of intra-class clustering and moderate diffusion. This indicates that the enhancement strategy improves sample diversity and model generalization ability while maintaining the semantic consistency of categories.

[0102] Step S202: Enhancement of intra-class perturbations and operating condition perturbations

[0103] In step S202, refer to Figure 25 Preset enhancement operators are applied to the standardized fault original sequence to generate intra-class perturbation enhancement sequences and operating condition perturbation enhancement sequences, so as to improve the model's generalization ability to measurement errors and operating condition fluctuations.

[0104] In a preferred embodiment, for any standardized original fault sequence Randomly select enhancement operators and their parameter vectors Constructing Enhancement Sequences Its form can be abstractly represented as:

[0105]

[0106] in, This represents the original sequence of standardized faults at time [time]. The possible values ​​of ; This indicates the enhanced sequence obtained after the enhancement operator is applied at time [time value missing]. The possible values ​​of ; This indicates the specific combination of operators in the enhancement operator; This represents the set of parameters for the enhancement operator, such as time shift, amplitude scaling factor, sag, frequency offset, DC bias, and linear trend slope. Indicates a time index.

[0107] Intraclass perturbation enhancement operators specifically include:

[0108] ① Time shift operator: A small time shift is applied by cyclic shifting to simulate subtle deviations in sampling time;

[0109] ② Uniform amplitude scaling operator: Apply uniform small amplitude scaling to all channels to simulate PT / CT ratio drift;

[0110] ③ High signal-to-noise ratio noise superposition operator: In White noise is superimposed within the range to simulate a background noise environment.

[0111] The operating condition disturbance enhancement operator specifically includes:

[0112] ① Frequency drift operator: In The system frequency drift is introduced within a certain range and is achieved by linear stretching / compression of the time axis to simulate the effect of frequency fluctuations on the waveform phase.

[0113] ② Trend Bias Operator: Overlaid on all channels DC bias and The linear trend simulates sensor temperature drift or long-term integration error.

[0114] ③ Current scaling operator: applies a scaling operator to the current path. The large value scaling within the range simulates current amplitude distortion caused by abnormal CT ratio or measurement system failure.

[0115] ④ Voltage sag / surge operator: The voltage channel has a finite time window amplitude change, and the sag range is... The temporary upgrade range is Simulates voltage transient processes.

[0116] By randomly applying the above perturbations to samples of the same type, enhanced samples are obtained that maintain the fault category but differ in time alignment, amplitude level, noise texture, and operating condition background, thus achieving intra-class distribution expansion and cross-operating condition simulation. All samples generated by the above operators together constitute the intra-class perturbation / operating condition perturbation enhancement sequence set in this invention.

[0117] Reference Figures 2 to 7 This demonstrates a waveform comparison under two enhancement methods: intra-class perturbation enhancement and operating condition perturbation enhancement, for typical faults such as external force damage, foreign objects, and lightning strikes. Specifically, Figure 2 , Figure 4 , Figure 6 The waveform changes of fault samples involving external force damage, foreign objects, and lightning strikes after intra-class perturbation enhancement are shown respectively. Figure 3 , Figure 5 , Figure 7 The waveform changes of the three types of fault samples after being enhanced by operating condition disturbances are shown respectively. Taking external force failure as an example, Figure 2The enhanced samples show slight changes in local details such as phase and amplitude compared to the original samples, but the overall fault waveform profile remains consistent, indicating that intra-class perturbation enhancement can expand the intra-class sample distribution without changing the original fault semantics. Figure 3 The enhanced sample, while maintaining the basic fault pattern, exhibits more significant changes in amplitude magnitude and local fluctuation characteristics, thus simulating response differences under complex operating conditions and measurement disturbances. Similarly, such as Figures 4 to 7 As shown, by implementing the above two types of enhancements on foreign object and lightning strike fault samples, a sub-sample set that combines the stability of the main structure with the diversity of local details can be generated. This effectively enriches the working condition background of the high-fidelity multi-source heterogeneous waveform dataset and provides data support for the subsequent CatBoost model to learn robust fault identification rules.

[0118] Through steps S201 and S202, the reconstructed enhanced sequences and the intra-class perturbation / operating condition perturbation enhanced sequences obtained respectively exhibit complementary characteristics in terms of temporal structure, spectral details, and operating condition background. These two types of enhanced sequences, together with the standardized original fault sequences, constitute the high-fidelity multi-source heterogeneous waveform dataset described in this invention. This dataset provides a diverse and physically interpretable training and recognition sample foundation for subsequent multi-window slicing, multi-dimensional feature extraction, and CatBoost multi-class fault initial judgment and dynamic trajectory cascade discrimination.

[0119] Step S3: Multi-window slicing and multi-dimensional feature extraction to construct a multi-dimensional feature vector set;

[0120] Reference Figure 26 The high-fidelity multi-source heterogeneous waveform dataset constructed in step S2 is then subjected to multi-window slicing and multi-dimensional feature extraction to form a multi-dimensional feature vector set for subsequent classification modeling. This process includes a multi-window slicing module, a multi-dimensional feature extraction module, and a multi-source feature concatenation module.

[0121] Step S301: Multi-window slicing module

[0122] The multi-window slicing module is used to extract multiple overlapping time window segments covering the entire fault evolution process around the fault trigger point. The multiple time windows are configured according to preset start and end relative positions so that they respectively cover the pre-fault, fault transient, and fault post-fault stages; at the same time, the overlapping of windows ensures that key fault details are not segmented and omitted, thereby enhancing the stability and separability of the features.

[0123] In a preferred embodiment, the multi-window slicing module further divides each standardized waveform segment (preferably from 40ms before the fault to 100ms after the fault, resampled to 5000Hz, resulting in a sequence of 700 points) centered on the fault trigger point. Preferably, the sequence is divided into four overlapping time windows to cover the leading, middle, and trailing stages of the fault evolution.

[0124] Preferably, all four windows are 350 points in length and are divided according to fixed start and end points as follows: window W1: [0, 350), window W2: [100, 450), window W3: [200, 550), and window W4: [350, 700]. The fault trigger point is preferably located near point 200 in the 700-point sequence. Therefore, the windows correspond to observation perspectives at different stages before and after the fault. By overlapping the windows, key transient features of the fault can be avoided from falling at the window boundaries, thus preventing information loss and improving the robustness of subsequent feature extraction and classification.

[0125] Step S302: Multi-dimensional feature extraction module

[0126] like Figure 26 As shown, in step S302, the multi-dimensional feature extraction module constructs a multi-dimensional feature engineering system within each time window segment obtained in step S301 and outputs a feature subset for subsequent CatBoost modeling. In a preferred embodiment, event-level features are shared and aligned to the corresponding window records across windows of the same event. Preferably, the following waveform features (consistent with the feature columns actually used in the training phase) are extracted within each window, including but not limited to:

[0127] ① Time / Environment Features: Extract time features such as the month in which the event occurs to characterize seasonal operational differences.

[0128] ② Event-level switch quantity characteristics: Introduce switch quantity statistical characteristics related to fault events as multi-source supplementary information, including reclosing success flag, multiple trip flag and power outage ratio index, to characterize fault handling behavior and severity.

[0129] ③ Time-domain characteristics: Extract features such as current impulse intensity, zero-sequence component level, energy and mutation rate to characterize fault transient intensity and evolution speed.

[0130] ④ Frequency domain features: Extract feature harmonic ratios and DC bias features to characterize nonlinear textures and waveform bias.

[0131] ⑤ Time-frequency domain features: Frequency band energy distribution and energy entropy features are extracted through wavelet decomposition to characterize the differences in the aggregation of transient energy in different frequency bands.

[0132] ⑥ Statistical and morphological characteristics: Extract features such as zero-sequence pulse width and zero-sequence voltage-current coupling relationship to capture the differences in intermittent impact and coupling morphology, and supplement the characterization of waveform distribution and change trend.

[0133] By extracting features from multiple dimensions such as time domain, frequency domain, time-frequency domain, statistical morphological features, and supplementary information, it is possible to comprehensively characterize the differences in fault waveform morphology, spectral structure, energy distribution, physical mechanism, and operational handling, effectively enhancing inter-class separability and intra-class robustness, and providing rich and interpretable discrimination basis for subsequent CatBoost multi-type fault initial judgment.

[0134] Step S303: Multi-source feature concatenation module, constructing a multi-dimensional feature vector set.

[0135] Reference Figure 26 In step S303, the multi-source feature stitching module stitches the multi-dimensional features extracted from each window in step S302 in a preset order, and further stitches together the features from multiple windows of the same fault event to form the final multi-dimensional feature vector. Its form can be expressed by the following formula:

[0136]

[0137] in, Indicates the fault event index; Indicates the number of time windows extracted around the fault trigger point; , , , and These represent fault events. In the Extracted time-domain, frequency-domain, time-frequency-domain, statistical morphology, and supplementary feature vectors within each window. Indicates a fault event In the The comprehensive feature vector on each window; Indicates a fault event The final event-level multi-dimensional feature vector; This indicates a vector concatenation operation.

[0138] Preferably, the multi-source feature stitching module further includes feature standardization processing, which performs a standardization transformation with a mean of 0 and a variance of 1 on each feature dimension to eliminate the adverse effects of differences in feature dimensions on subsequent classification modeling.

[0139] Step S4: CatBoost Initial Fault Assessment and Confidence-Based Traffic Splitting

[0140] like Figure 27As shown, the CatBoost model described in step S4 is a highly efficient classification algorithm based on gradient boosting decision trees, possessing advantages such as native support for categorical features, strong resistance to overfitting, and fast training speed. In this invention, the model receives a multi-dimensional feature vector set from step S3, first independently classifying and predicting the feature vectors for each time window, and then aggregating the window-level prediction results into event-level discrimination results through a specific window weight weighted fusion strategy. Step S4 realizes the complete classification process from a multi-dimensional feature vector set to fault category prediction, including three sub-steps: CatBoost multi-class fault initial judgment, window weight weighted fusion strategy, and confidence-based triage decision. The mathematical description of the entire process can be expressed by the following formula:

[0141]

[0142] in This indicates that the CatBoost model is effective for events. The Window feature vectors Predicted as the first The probability of a type of failure; Indicates an event The Feature vectors of a window; Indicates the fault event index; Indicates the window index. ; Indicates the fault category index, ; Indicates an event In the Event-level fusion probability over class-related faults; For the first The weight coefficients of each window satisfy the following conditions: ; Total number of windows; Total number of fault categories; For CatBoost model events The initial category prediction results; For the event The confidence level is the maximum value of the event-level fusion probability.

[0143] Step S401: Implementation of Initial Fault Judgment for Multiple Types in CatBoost

[0144] In step S401, the CatBoost model receives a multi-dimensional feature vector set from step S3, independently classifies and predicts the feature vectors for each time window, and outputs a window-level prediction probability distribution. In terms of model structure design, the tree depth is set to 6 layers, the learning rate to 0.03, and the number of iterations to 800. These parameters have been experimentally verified to achieve an optimal balance between model expressive power and overfitting risk, while ensuring sufficient model convergence.

[0145] Step S402: Weighted fusion strategy based on window weight

[0146] In step S402, the system weights and fuses the window-level predicted probabilities according to preset window weights, aggregating them into an event-level probability distribution. By assigning different weight coefficients to different time windows, key information in the fault evolution process can be better captured. Preferably, based on the temporal evolution characteristics of the fault waveform, a symmetrical window weight distribution is adopted: windows W2 and W3 (indices 1 and 2) cover the core time period of fault occurrence, containing the richest fault feature information, with a weight set to 0.35; windows W1 and W4 (indices 0 and 3) capture the fault precursors and subsequent evolution respectively, with a weight set to 0.15, preserving complete sequence information.

[0147] This design follows temporal bias, ensuring the model has a balanced perception of the fault evolution process, while guaranteeing that the total weights are strictly equal to 1, facilitating subsequent confidence calculations and threshold comparisons. Furthermore, this weighting scheme considers the balance between computational complexity and performance, achieving high-quality information fusion through simple weighted averaging, avoiding the overfitting risks that complex nonlinear fusion algorithms may bring.

[0148] Step S403: Confidence Calculation and Triage Decision

[0149] Step S403 calculates the confidence level based on the event-level probability distribution from step S402 and performs intelligent triage decision-making to determine whether subsequent dynamic trajectory cascade discrimination needs to be triggered. First, it determines whether the dynamic trajectory discriminator needs to be triggered based on the confidence level of the initial CatBoost judgment result. The triggering condition is expressed by the following formula:

[0150]

[0151] in, For the event Confidence level; The confidence threshold; For CatBoost model events The initial category prediction results; For the event The final category prediction result; This serves as an index for fault events; when the confidence level is higher than a threshold, the initial CatBoost judgment result is directly output as the final predicted category; when the confidence level is lower than a threshold, the sample is marked as a low-confidence sample and enters the dynamic trajectory discriminator cascade process. In one embodiment, the confidence threshold... The threshold can be preset based on the event-level confidence distribution on the training or validation set, the false positive rate of low-confidence samples, and the trigger ratio of the dynamic trajectory discriminator; alternatively, a validation set grid search can be performed within the candidate threshold set to select a threshold that allows the validation set classification accuracy, recall, or overall evaluation metric to meet preset requirements. Preferably, It can be 0.7, but is not limited to that value.

[0152] This binary triage mechanism has clear decision boundaries, avoiding the complexity and error accumulation that can result from multi-level judgments. At the same time, this mechanism provides a clear processing path for each fault event, facilitating fault analysts to track and verify the diagnostic process.

[0153] Step S5: Input the low-confidence samples into the dynamic trajectory discriminator and perform cascaded correction.

[0154] like Figure 28 As shown, step S5 follows the triage decision in step S4, specifically handling complex and confusing fault samples with confidence levels below a threshold. This step achieves precise discrimination of boundary samples by constructing dynamic trajectory features and calculating energy distance. The dynamic trajectory cascading process comprises two core stages: trajectory construction and standardization, and energy calculation and cascading discrimination. The entire process is based on rigorous mathematical theory, ensuring reliable diagnosis in complex fault scenarios.

[0155] Step S501: Trajectory Construction and Standardization

[0156] In step S501, multi-window features are aggregated according to fault events to construct a 4×F dynamic trajectory and perform standardization processing. The calculation process is expressed by the following formula:

[0157]

[0158] in, For the event In the The first window extracted Individual trajectory discrimination feature values; For the event In the Input feature vectors on each window; For the first Trajectory feature extraction function; A set of trajectory feature extraction functions; Index for windows ; For trajectory feature index; F is the total number of trajectory discriminative features; for all and corresponding Arranged in order, they constitute events. of Dynamic trajectory matrix. The standardized trajectory features; and The first The mean and standard deviation of each trajectory feature on the training set.

[0159] Trajectory Feature Extraction Function Set It is the core component of the dynamic trajectory discriminator. Each trajectory feature extraction function Window-level input feature vector As input, output the corresponding trajectory discrimination features. This type of trajectory discrimination feature differs from the general feature extraction in step S3, focusing more on characterizing the dynamic change pattern of the fault during the cross-window evolution process.

[0160] In one implementation, the trajectory discrimination feature set The mean value can be pre-selected during the training phase or determined based on the feature importance, class discrimination, and physical interpretability on the training set. For the same fault event, the system constructs a fixed-dimensional 4×F dynamic trajectory matrix according to the window order; and standard deviation It is obtained from statistics of the training set and remains fixed during the inference phase.

[0161] Step S502: Energy calculation and cascaded discriminant decision

[0162] like Figure 28 As shown, step S502 calculates the events based on standardized trajectory features. Relative to category The geometric distance, morphological similarity distance, and physical constraint distance are calculated and weighted to obtain the final trajectory energy. Based on this, cascaded discrimination correction is completed according to the energy difference and correction direction rules.

[0163] Geometric distance Used for quantifying events Standardized trajectory features and categories The formula for calculating the geometric deviation between trajectory pattern prototypes is as follows:

[0164]

[0165] in, For the event With category The geometric distance between them; For the first The first window, the Joint weight coefficients corresponding to each trajectory feature; Standardized trajectory features; For category In the The first window Pattern prototype values ​​of trajectory features; For window indexing; For trajectory feature indexing; This represents the total number of trajectory discrimination features.

[0166] In one implementation, the category Trajectory pattern prototype Can be classified by category in the training set The patterns are obtained through standardized dynamic trajectory statistics or clustering, for example, by using the mean trajectory of the same type of trajectory, cluster center, or mean trajectory within a cluster as the pattern prototype. Each category can correspond to one or more pattern prototypes to characterize the typical evolution path under different working conditions.

[0167] This distance ensures an accurate measurement of trajectory features in multidimensional space, through weighting coefficients. This approach achieves importance allocation across different time windows and feature dimensions. The weighting coefficients can be pre-set based on statistical analysis or validation results during the training phase to reflect the differences in contribution of each window and feature to fault evolution discrimination.

[0168] Morphological similarity distance The correlation coefficient is used to quantify the consistency between the trajectory temporal shape and the pattern prototype, which is used to characterize the degree of morphological matching of the fault evolution trajectory in the time dimension.

[0169]

[0170] in, For the event With category The morphological similarity distance between them; The total number of trajectory discrimination features; This represents the Pearson correlation coefficient; Indicates the event The Trajectory features in Standardized trajectory values ​​on each window , , and The constructed time-series trajectory; Indicates by category The Trajectory feature pattern prototype value , , and The pattern prototype trajectory is formed. Among them, Similar to the aforementioned model prototype Corresponding to, as category In the A cross-window prototype sequence on a trajectory feature. The smaller the distance, the higher the consistency of the trajectory morphology.

[0171] Physical constraint distance The penalty term is calculated based on the envelope boundary constraints, comprising two parts: magnitude constraint penalty and inter-window difference constraint penalty. The calculation process can be expressed as follows:

[0172]

[0173] in, For the event In category The amplitude constraint penalty energy; For the event In category The energy penalty for the difference constraint; For the event Relative to category The total physical constraint distance is the sum of the magnitude penalty and the differential penalty. For the event In the The first window Standardized trajectory values ​​for each trajectory feature; The standardized trajectory difference value of this feature between adjacent windows can be expressed as: ; and Categories In the The first window The lower and upper bounds of the amplitude of each trajectory feature; This represents the corresponding amplitude standard deviation; and Categories In the The first window The lower and upper bounds of the trajectory feature difference values; This represents the corresponding standard deviation of the difference; For window indexing; For trajectory feature indexing; This represents the total number of trajectory discrimination features.

[0174] In one implementation, the amplitude boundary and the difference boundary can be determined by categories in the training set. The standardized dynamic trajectory is obtained through statistical analysis, for example, by statistical analysis of the mean, standard deviation and adjacent window difference of similar trajectories, and is kept fixed during the inference stage to constrain whether the event to be identified is within the typical physical change range of that category.

[0175] The physical constraint distance, by setting reasonable threshold boundaries, restricts trajectory features to the physically feasible domain of each category. The portion exceeding the boundaries is penalized and accumulated after normalization according to the standard deviation, ensuring the physical reasonableness of the classification results. This constraint mechanism is based on statistical analysis of a large amount of historical fault data, extracting typical variation ranges of each category across different feature dimensions to form reliable physical prior knowledge.

[0176] The final trajectory energy is obtained by weighted fusion of the three distance components:

[0177]

[0178] in, For the event Relative to category The final trajectory energy; These are the weighting coefficients for the geometric distance term; For the event With category The geometric distance between them; This represents the weighting coefficient for the morphological similarity distance term; For the event With category The morphological similarity distance between them; These are the weighting coefficients for the physical constraint distance term; For the event With category The physical constraint distance between them.

[0179] After obtaining the final trajectory energy corresponding to each category, candidate categories are further determined and the energy difference is calculated. The final cascaded discrimination is then completed based on the threshold and correction direction rules. The calculation process can be expressed as follows:

[0180]

[0181] when and season Otherwise, let .

[0182] in, Indicates the category determined by CatBoost. In addition, the alternative categories that minimize the final trajectory energy; This represents the final trajectory energy corresponding to the initial classification. This represents the final trajectory energy corresponding to the candidate category; This represents the energy difference between the initial classification and the candidate classification. Indicates the energy difference threshold; Indicates an event The result of the correction direction rule is set to 1 if the correction direction rule is satisfied, and 0 otherwise. This indicates that CatBoost handles events. The initial category prediction results; Indicates an event The final category prediction results.

[0183] In one implementation, the weighting coefficients of geometric distance, morphological similarity distance, and physical constraint distance can be preset based on statistical analysis or verification results during the training phase, in order to balance the contributions of geometric distance, morphological similarity distance, and physical constraint distance to the final trajectory energy.

[0184] Furthermore, the energy difference threshold The correction direction rule can be preset based on the validation set classification results, energy difference distribution, or grid search results; The initial judgment is modified to the candidate category only when the energy difference exceeds the energy difference threshold and the correction direction rule is met; otherwise, the initial judgment remains unchanged.

[0185] By synergistically fusing geometric distance, morphological similarity distance, and physical constraint distance, and using intelligent threshold discrimination based on energy differences, step S502 successfully achieved accurate classification correction for complex samples with low confidence. This process, using the cross-time evolution trajectory composed of multi-window features as the core of discrimination, can more fully utilize the temporal correlation and dynamic change patterns of transmission line faults, providing key technical support for the high-precision operation of the entire fault diagnosis system.

[0186] Performance evaluation and results analysis

[0187] Based on the event-level discrimination results obtained in steps S4 and S5, the fault identification performance is statistically evaluated, and the final classification results are compiled and output. This includes the following aspects:

[0188] First, the system reads the generated window-level feature table, the CatBoost event-level prediction results obtained in step S4, and the final prediction category after dynamic trajectory cascade discrimination in step S5, and constructs an event-level result set containing fields such as true label, initial judgment result, cascade result, energy difference, and confidence level.

[0189] Secondly, based on the above event-level results, the overall recognition performance is quantitatively evaluated. The overall accuracy of the CatBoost initial judgment result and the cascaded final result are calculated on all five types of fault samples. The performance changes of the system before and after cascading on all samples are compared. At the same time, detailed statistical results such as confusion matrix and classification report are generated on the test set.

[0190] Finally, the final category label obtained from the cascaded discrimination is used as the output of the fault identification method of this invention. For each fault event, the system not only provides the final fault category, but also retains auxiliary information such as the CatBoost initial judgment category, confidence level, and whether dynamic trajectory cascaded discrimination was triggered, providing an interpretable basis for subsequent manual review and misjudgment analysis.

[0191] The transmission line fault identification method based on dual enhancement and dynamic trajectory discrimination adopted in this invention has been fully verified through a series of experiments. Figure 29 and Figure 30 The comprehensive performance evaluation results of the method proposed in this invention are presented.

[0192] Figure 29 The confusion matrices of the CatBoost and cascaded models on the test set are compared. The confusion matrices show that the cascaded model significantly improves upon the CatBoost model, with the most significant improvement in the "bird damage" category, where accuracy jumped from 57.1% to 85.7%. The CatBoost model misclassified 28.6% of "bird damage" samples as "foreign objects" and 22.2% of "external force damage" samples as "foreign objects." The cascaded model reduced the misclassification rate for "external force damage" from 22.2% to 11.1% and completely eliminated the misclassification of "bird damage" as "foreign objects." This demonstrates that by introducing cascaded discrimination based on cross-window dynamic trajectories, the system can more fully utilize the evolutionary differences of fault events over time, thereby effectively improving the ability to distinguish easily confused categories.

[0193] Figure 30 The radar chart shows the accuracy comparison of the two models across various categories. The cascaded model (green line) outperforms or is no less than the CatBoost model (blue line) in almost all categories, with the most significant improvements in the "External Force Damage" and "Bird Damage" categories: the accuracy for "External Force Damage" increased from 0.667 to 0.889, and the accuracy for "Bird Damage" increased from 0.571 to 0.857. The accuracy for the remaining categories remained largely unchanged. The overall accuracy improved from 83.64% to 90.91%, meeting the stringent accuracy requirements for power system fault diagnosis.

[0194] Detailed analysis of the confusion matrix revealed that the main confusion in the cascaded model occurred when the "external force damage" category was misclassified as the "foreign object" category (11.1%). Despite a small number of misclassifications, the cascaded discriminator improved the overall accuracy of the model from 83.64% to 90.91% overall.

[0195] In summary, a comprehensive analysis of the confusion matrix and radar chart demonstrates that the cascaded discriminator method proposed in this invention exhibits superior performance in fault diagnosis tasks. The model demonstrates high accuracy and powerful feature extraction and discrimination capabilities, particularly excelling in handling easily confused fault categories. This cascaded architecture, based on the collaboration of CatBoost and a dynamic trajectory discriminator, provides a solid algorithmic foundation for power system fault diagnosis, effectively improving the accuracy and reliability of fault diagnosis.

Claims

1. A method for power transmission line fault identification based on dual enhancement and dynamic trajectory discrimination, characterized in that, Includes the following steps: S1: Collect raw transmission line fault recording data and perform basic preprocessing to output standardized raw fault sequences; S2: Perform a dual enhancement strategy on the standardized fault original sequence to construct an enhanced multi-source heterogeneous waveform dataset. Specifically, this includes enhancing the standardized fault original sequence with intra-class perturbation and operating condition perturbation, as well as enhancing it with modal cross-recombination, to obtain the enhanced multi-source heterogeneous waveform dataset. The intraclass perturbation includes two or three of the following: time shift, amplitude scaling, and noise superposition; the operating condition perturbation includes two, three, or four of the following: voltage surge / dip, current amplitude scaling, frequency drift, and DC bias. The modal crossover and recombination enhancement includes: performing variational mode decomposition on the sequence to obtain intrinsic mode components, and then reconstructing the intrinsic mode components by crossover and recombination according to the frequency components from largest to smallest; S3: Perform multi-window slicing and multi-dimensional feature extraction on the enhanced multi-source heterogeneous waveform dataset to construct a multi-dimensional feature vector set; S4: Input the multi-dimensional feature vector set into the classification model to perform preliminary judgment of multiple types of faults, and output the confidence score of the sample; The classification model used is the CatBoost classification model, which employs a window weight weighted fusion strategy. S5: Input samples with confidence levels below a preset threshold into the dynamic trajectory discriminator, perform discrimination correction, and obtain the transmission line fault identification results, specifically including: Samples with confidence levels below a preset threshold are aggregated into a dynamic trajectory matrix based on event-based multi-window features and then standardized. The energy value and energy difference of the trajectory relative to the prototype of each type of pattern are calculated, and the judgment and correction are performed according to the threshold and direction rules to obtain the transmission line fault identification results.

2. The method according to claim 1, characterized in that, In step S1, the basic preprocessing includes data parsing, quality screening, fault event windowing, resampling, and DC detrending processing.

3. The method according to claim 1, characterized in that, In step S3, the multi-window includes 4 to 8 overlapping time windows regarding the fault development state; The multi-dimensional features include time-domain features, frequency-domain features, time-frequency-domain features, statistical morphological features, and auxiliary metadata features.

4. The method according to claim 1, characterized in that, In step S5, the energy value and energy difference of the trajectory relative to the prototype of each type of pattern are calculated using energy calculation, which includes a weighted combination of geometric distance, morphological similarity and physical constraint distance.

5. A transmission line fault identification device, comprising a processor and a memory, characterized in that, The memory stores a computer program, and when the processor executes the computer program, it implements the method as described in any one of claims 1 to 4.

Citation Information

Patent Citations

  • Lithium battery fault diagnosis method and device based on CatBoost and electronic equipment

    CN114386537A

  • Method and process of fault detection in power systems using machine learning feature extraction

    US20260118399A1