AC power grid fault current waveform recording method

Abstract

The invention discloses an alternating current power grid fault current waveform recording method, and relates to the technical field of power grid fault detection, when power grid waveform distortion or a high-resistance fault symptom is monitored, denoising and wavelet enhancement are performed on multi-source wave recording and simulation data through a data acquisition and preprocessing module, and an initial data set with multi-label high-order features is generated; the multi-task learning module corrects data distribution deviation in combination with domain adaptation and adaptive weighted loss, integrates physical priori such as arc discharge, and recognizes composite faults such as harmonic distortion, short circuit and grounding; through cross validation and noise simulation evaluation, the model verifies robustness through high-order calculus measurement, characteristic attention distribution can be verified in a visual mode, incremental training is automatically triggered when real-time recording reaches a threshold value, a comparison memory mechanism is applied, new fault information is adaptively absorbed, and a multi-label diagnosis closed loop is formed. And the accuracy and continuous optimization capability of multi-scene fault identification of the smart power grid are effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power grid fault detection, and specifically to an AC power grid fault current waveform recording method. Background Technique

[0002] With the continuous evolution of the power system towards intelligence, distribution, and diversification, the access of large-scale renewable energy generation and power electronic terminals in the distribution link is increasing day by day, and the complexity and dynamics of power grid operation have increased significantly. Multiple harmonics, transient interference, and non-linear arc discharge characteristics are often superimposed in the voltage and current waveforms, and some lines will also show a time-varying state of repeated power flow inversion due to the reverse power flow of distributed power sources.

[0003] In this context, the fault forms are not only limited to the traditional single three-phase short circuit or single-phase grounding forms, but also multiple faults such as harmonic distortion, high-resistance grounding, and instantaneous overvoltage may occur concurrently during the same fault period, resulting in compound fault phenomena. Since these fault types are coupled or superimposed in the waveform domain, it is often impossible to accurately identify each sub-fault component based on a single feature or traditional simple criterion. On the other hand, the diversity of power grid scale and operating conditions also increases the difficulty of fault sample acquisition, data management, and real-time processing, making it an urgent issue to be broken through in the field of intelligent power grid fault diagnosis to deeply explore multi-source recording information and maintain the adaptive ability of the model to environmental changes.

[0004] In the Chinese invention patent with the authorization announcement number CN113625125B, a high-resistance grounding fault detection method, device, and equipment for a distribution network are disclosed. The method includes: periodically obtaining the magnetic field signal corresponding to the distribution network feeder; performing a first noise reduction operation on the magnetic field signal to generate multiple first-order basic components; performing a second noise reduction operation on each first-order basic component to generate multiple second-order basic components; constructing a target curve according to the numerically largest and second-largest second-order basic components selected from the multiple second-order basic components as the abscissa component and the ordinate component; and judging whether the distribution network feeder has a high-resistance grounding fault according to the comparison result between the target slope corresponding to the target curve and the preset slope coefficient, so as to quickly and accurately identify the high-resistance grounding fault of the distribution network feeder.

[0005] However, combined with the actual application scenario and the existing technology: In the current process of multi-label recognition and online optimization for complex faults, there is still a core problem: how to accurately identify rare faults (such as high-resistance faults) and concurrent fault types from the vast amount of collected fault waveform data, and achieve adaptive updates in a real-time environment. When the power grid's power flow direction, load characteristics, or system topology change, it is often difficult to track new fault evolution patterns in a timely manner using a static model for fault analysis. At the same time, the distribution deviation between simulation data and real recordings can lead to insufficient generalization ability of the model for on-site scenarios; even if the data scale is enhanced through physical simulation, it is difficult to avoid the differences between actual waveforms and simulations in terms of noise, edge sampling accuracy, and arc discharge non-linearity, resulting in fault discrimination deviations. In addition, multi-label fusion requires effective decomposition of different types of fault characteristics such as harmonic distortion, short circuit, and overvoltage within the same waveform. If the physical prior and online learning mechanism are not combined, it is extremely easy to cause a sharp drop in recognition results and missed detections when new fault patterns or fault distributions change.

[0006] To this end, the present invention provides a method for recording fault current waveforms of an AC power grid. Summary of the Invention

[0007] (I) Technical problems to be solved In view of the deficiencies of the prior art, the present invention provides a method for recording fault current waveforms of an AC power grid. The data acquisition and preprocessing module denoises and wavelet-enhances multi-source recordings and simulation data to generate an initial data set of multi-label high-order features. Subsequently, the multi-task learning module combines domain adaptation and adaptive weighted loss to correct data distribution deviations, and incorporates physical priors such as arc discharge to identify complex faults such as harmonic distortion, short circuit, and grounding. After cross-validation and noise simulation evaluation, the model verifies its robustness using high-order calculus metrics and can check the feature attention distribution through visualization. Finally, when the real-time recording reaches the threshold during online deployment, it automatically triggers incremental training and applies a contrast memory mechanism to adaptively absorb new fault information, forming a multi-label diagnosis closed-loop, effectively improving the accuracy and continuous optimization ability of multi-scenario fault recognition in smart grids, and solving the technical problems described in the background art.

[0008] (II) Technical solutions To achieve the above objectives, the present invention is realized through the following technical solutions: A method for recording the waveform of AC power grid fault current, including: when it is monitored that the waveform of the power grid is distorted or there are signs of high-resistance faults, the data acquisition and preprocessing module automatically extracts multi-source recording and simulation data when the monitoring result exceeds the pre-set fault threshold; by performing denoising and anomaly rejection based on physical laws, and then using the wavelet method to enhance and expand rare fault samples, and enriching sample identification through hierarchical indexing and label binding, finally generating an initial data set containing multi-label high-order features and unified formatted storage, which can be used for subsequent multi-label fault analysis; After the initial data set is formatted and enters the multi-label modeling stage, the multi-task learning module, after obtaining the high-order features and the input of the time-domain sampling branch, corrects the problem of uneven distribution through domain adaptation and adaptive weighted loss function, and incorporates physical priors such as arc discharge to limit the abnormal drift of network weights. At the same time, parallel convolution extraction is performed on the harmonic distortion features and the hidden vector activation value is output at the fusion layer, and finally a multi-label classification structure model for composite fault identification is generated; After the constructed multi-label model completes the preliminary training and enters the evaluation stage, the cross-validation and noise simulation module examines the stability of the model for various fault scenarios through the analysis of splitting and high-order calculus metric formulas, and checks the attention distribution of key features such as harmonic distortion and high-resistance grounding in the interpretability visualization. If there are no weights in the physical prior constraint layer that violate the laws of electricity, the intermediate model parameters with good robustness and the ability to analyze fault subdivision features are output; When the model completes the offline verification and needs to be put into the edge and cloud collaborative environment, the online deployment and continuous optimization module starts the fault identification inference after the real-time recording wave meets the trigger threshold, and monitors the identification quality of new or rare faults through the high-order deviation metric function. Once a significant drift is detected, it automatically uploads the error samples and labels to trigger incremental training, and at the same time retains historical knowledge through the comparison memory mechanism, and finally forms a long-term adaptive fault diagnosis closed loop.

[0009] Preferably, fault waveform data is obtained from multiple sources, and after the obtained raw data is preliminarily cleaned, it is named initial cleaned data, where: the rejection operation is performed on the waveform segments with obvious distortion; the samples with abnormal timestamps or seriously inconsistent with the known working condition information are marked and temporarily stored in the abnormal library; the common-sense error samples are filtered according to the prior physical constraints; Preferably, based on the rare fault types or composite fault waveforms in the initial cleaned data, targeted data enhancement and high-order feature extraction are performed, and an enhanced feature data set is generated. For the fault types with insufficient sample size, synthetic fault data with real waveform statistical characteristics is constructed, and part of the enhanced data obtained is integrated with the raw data and incorporated into the enhanced feature data set; Introduce high-order waveform energy measure to identify the key time-frequency characteristics of composite faults, and define the high-order waveform energy function , and its form can be based on the integral accumulation of high-order derivatives in the wavelet domain:

[0010] In the formula: represents the discrete wavelet transform coefficients of the fault waveform , is the layer scale, is the translation variable; is the order of the high-order derivative ; is the upper limit of the wavelet translation range; Preferably, keywords such as waveform sampling points, labels, GAN enhancement identifiers, and high-order energy metrics are stored in a hierarchical table; the enhanced feature data is randomly sampled according to the required ratio to obtain three subsets: training, validation, and testing; the consistency of the diversified samples is checked to make the high-order waveform energy correspond one-to-one with the original waveform, generating preprocessed output data; Preferably, multiple parallel input branches are defined to process time-domain sampling data and high-order eigenvalues respectively; After completing the parallel input, a basic convolutional layer is introduced to extract local patterns, and a feature fusion layer is set up to map the features extracted from each branch to the same hidden vector space to form a fused hidden vector; according to the set of target fault labels in the power grid, after retaining the number of channels matching the number of labels, a preliminary classification activation output is generated; Preferably, an adaptively updatable weight coefficient is incorporated into the loss function, and the following multi-label weighted loss function is defined :

[0011] Where: represents the classification activation output; is the corresponding true label; is the weighted coefficient of the fault category, which can be dynamically updated with iteration; is a single-label sub-loss function for measuring the difference between the activation value and the true label; A cross-domain mapping function is introduced to align the fused hidden vectors, and the cross-domain difference metric is defined as follows:

[0012] In the formula: and represent the input features in the simulation domain and the real domain respectively; is a learnable cross-domain mapping function, represents an output value or component at a certain dimension after mapping the feature vector h to the latent space and is the domain of definition of the latent space index; is the fractional-order (or higher-order) derivative of the variable z, where is the fractional-order; when it degenerates into the conventional first-order derivative; when it is the fractional-order derivative; is the balance coefficient, represents the vector norm or absolute value operation; Preferably, monotonicity or non-linear transition constraints are added to the activation function of the convolutional channel, which is achieved by imposing an additional regular constraint function on the weights of each convolutional kernel: In the formula: W is the set of weights of all convolutional kernels in this sub-channel;

[0013] represents the positive activation weight determined to correspond to the high-resistance fault mechanism in the m-th convolutional kernel; is the predefined constraint coefficient, is the overall weight of the physical prior regular term; Under the physical prior constraint, the domain adaptation module and the multi-label adaptive weighted loss function are retrained or fine-tuned, and the initial multi-label model, including the network structure and initial weights, is output after several rounds of iteration; Preferably, a subset with the widest coverage of fault label categories and diverse time-frequency features is screened from the training dataset and loaded to match the initial multi-label model; for categories with scarce sample sizes such as high-resistance faults, oversampling or increasing the training frequency is performed on the preprocessed output data of the original dataset, and combined with the adaptive weighted loss function, the attention to rare categories is improved; The prepared training set and the initial multi-label model are fed into the deep learning framework, and several rounds of iteration are performed according to the comprehensive loss function to obtain the preliminarily optimized network weights and generate the first training intermediate model; Preferably, the training set is divided into subsets, and one of them is used as the validation set one by one, and the remaining subsets are used as the training set, and a training-validation loop is performed on the first training intermediate model, and the multi-label classification results are recorded after each training; During the training or validation phase, additional noise simulation is introduced, amplitude perturbation or random high-frequency interference injection is performed on the waveform, and the following robustness metric function is introduced: :

[0014] In the formula: represents the response value shown by the model for a certain key performance indicator when the noise amplitude is under the -fold cross-validation; Different correspond to different intensities or types of noise injection, is the value range of the noise amplitude or interference intensity; represents the th-order derivative of with respect to the noise amplitude ; ; represents the value of the continuous wavelet transform coefficient of on the noise dimension at one or several scales ; , is the balance coefficient; combining the cross-validation results and the robustness metric function , further fine-tuning and iterating the intermediate training model one until the cross-validation metrics for each fold converge basically, and finally forming a more reliable model weight and structure configuration, and outputting the intermediate training model two; Preferably, by drawing an attention distribution map, showing the high-order response correlation of the key intervals concerned by the model at the waveform time domain or frequency domain level: checking the change slope or change curve of the model output label when the input fault waveform is fine-tuned on the key features, to judge whether the model strictly conducts identification around the physical prior. After interpretability verification and physical completeness verification, an interpretable training model is obtained; Preferably, install a fault identification and inference module on the edge device of the nearest substation to implement low-latency fault detection, and retain the global data aggregation and analysis function on the cloud platform for storing historical fault records and subsequent evaluation; Import the network structure and weights of the interpretable training model into the edge inference engine, and set the real-time sampling frequency, batch size, and number of core threads according to the grid operation parameters; continuously monitor the fault identification latency, accuracy, and resource occupancy within the set time period, collect the multi-label output results of online identification, and compare them with the known fault events, and record the real-time error list; Preferably, compare the real-time identification results reported by the edge side with the fault event records, count the accuracy and missed detection rate for each sub-fault type and centrally save them in the incrementally updated model; Define the high-order deviation metric function , and the formula is as follows:

[0015] In the formula: Characterize the multi-label fault identification deviation at time ; Denote the -th order derivative operation with respect to time is the fractal dimension metric function, is the trade-off coefficient, is the time window size; When the high-order deviation metric function exceeds the deviation threshold, an update is triggered; if the high-order deviation metric function exceeds the deviation threshold or a new fault category is detected, the corresponding data and labels are packaged and sent back to the cloud; For retraining or incremental training on new data, a new incremental update model is generated; after training is completed, the incremental update model is sent back to the edge side and automatically replaces the currently used fault identification model to form a deployed update model; Preferably, during training, key data and its feature distribution from the previous stage are appropriately retained, or regularization constraints are used to suppress large offsets of network weights, and a contrastive memory constraint function is introduced, as follows:

[0016] where: is the weight set of the previous stable model version, and W is the weight after the current retraining; is the -th order derivative to measure the distribution pattern of the difference between the weights of the new and old models in the weight dimension When it is confirmed that the model identification result is incorrect or a misjudgment occurs, this data is archived and manually labeled to form operation and maintenance labeled data, and the operation and maintenance labeled data is periodically incorporated into the retraining set; after multiple rounds of updates, memory merging, and operation and maintenance feedback, a deployed update model is integrated and obtained; (III) Beneficial effects The present invention provides a method for recording the fault current waveform of an AC power grid, having the following beneficial effects: This solution provides an overall solution idea with high precision, high robustness, and sustainable evolution for composite fault identification from multi-source data acquisition and preprocessing to online deployment and continuous optimization, including: By first obtaining initial cleaned data and enhanced feature data and outputting preprocessed output data, it can effectively balance the sample distributions of common faults and rare faults (such as high-resistance grounding), and with the assistance of wavelet transform or GAN data augmentation, it strengthens the extraction of complex waveform features such as transient spikes and harmonic distortions, not only retaining the diversity of real fault scenarios, but also highlighting key fault signals with high-order energy metrics or fractal features, so that the model training stage is no longer limited to a single feature dimension; The initial multi-label model realizes the parallel discrimination of multiple fault types such as harmonic distortion, short circuit, and grounding at the levels of fusing hidden vectors and classification activation outputs through parallel feature embedding and domain adaptation modules. An adaptive weighting factor and physical prior regularization are introduced into the loss function to enhance the model's adaptability to high-resistance faults and the real-simulation domain differences, significantly reducing the misjudgment probability of the model in the face of multi-label compound fault scenarios and making the network structure more interpretable; The training of intermediate model one and intermediate model two are sequentially tested through cross-validation and noise simulation. The robustness is measured by high-order calculus metrics, and the consistency between the power mechanism and network weights is verified under the visualization mechanism. The finally output interpretable training model not only avoids overfitting relying on a single scenario but also ensures stable recognition performance under diverse working conditions; The constructed real-time detection and online update process enables the real-time error list to monitor the fault recognition deviation at any time. Once a significant deviation is found, it automatically triggers cloud incremental training, generates an incremental update model, and quickly deploys it as the final deployed model. Combined with the operation and maintenance annotation and contrast memory mechanism, the model can continuously adapt to new fault samples and environmental changes while avoiding forgetting the learned knowledge. Overall; This solution effectively solves the problems of data imbalance, simulation-real domain differences, model interpretability, and online update in multi-label compound fault recognition, and can be widely applied to the detection and diagnosis of multi-scenario faults in smart grids. Brief Description of the Drawings

[0017] Figure 1 It is a schematic flow chart of the method for recording the fault current waveform of the AC power grid of the present invention. Detailed Embodiments

[0018] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0019] Please refer to Figure 1 , the present invention provides a method for recording the fault current waveform of an AC power grid, including, In a smart grid, the fault types are becoming increasingly diverse and prone to overlap with each other, thus forming compound faults and leading to complex waveform features. In order to achieve more accurate multi-label fault type recognition in subsequent steps, it is necessary to perform sufficient and detailed preprocessing and preliminary analysis on multi-source fault waveform data in the early stage; among them: Step 1. When it is detected that waveform distortion occurs in the power grid or signs of high-resistance faults appear, after the detection threshold of the data acquisition and preprocessing module reaches the set threshold, multi-source recorded wave and simulation data are automatically extracted; By performing denoising and anomaly rejection based on physical laws, then using the wavelet method to enhance and expand rare fault samples, and enriching sample identification through hierarchical indexing and label binding, an initial data set containing multi-label high-order features and unified formatted storage, which can be used for subsequent multi-label fault analysis, is finally generated; The above Step 1 includes the following contents: Step 101. Acquisition and preliminary cleaning of multi-source waveform data Fault waveform data are obtained from multiple sources, including collecting fault waveforms triggered under various operating conditions through a real power grid fault recorder (or online monitoring system), and generating synthetic fault data using a high-fidelity simulation platform under different distribution network parameters, fault point positions, and interference conditions; After preliminary cleaning of the obtained original data, it is uniformly named initial cleaned data, where: For waveform segments with obvious distortion (such as a constant curve caused by a sensor failure), rejection operations are performed; samples with abnormal timestamps or seriously inconsistent with known operating condition information are marked and temporarily stored in the anomaly library; common-sense error samples are filtered according to prior physical constraints (such as intervals where current and voltage cannot be negative at the same time); The data quality can be initially ensured, and invalid information caused by sensor defects or communication failures can be excluded; at the same time, real and simulation waveforms are uniformly incorporated into the initial cleaned data; by fusing real power grid recorded waves and simulation data, the diversity of fault samples is improved, and obvious abnormal or invalid waveforms can be effectively removed through preliminary cleaning and physical rule filtering, reducing interference in subsequent analysis, and laying a high-quality and wide-coverage data foundation for subsequent data enhancement and feature extraction; Step 102. Data enhancement and high-order feature extraction Based on rare fault types (such as high-resistance faults) or composite fault waveforms in the initial cleaned data, targeted data enhancement and high-order feature extraction are performed, and an enhanced feature data set is generated, focusing on the expansion of rare samples and the mining of key features. The core idea is to achieve a deeper characterization of composite fault signals through high-order transformation measures to support subsequent multi-label recognition models, where: For fault types with insufficient sample sizes such as high-resistance faults, the generative adversarial network (GAN) or wavelet perturbation method is used to construct synthetic fault data with real waveform statistical characteristics, balance the distribution of different fault categories in the initial cleaned data, and integrate some of the enhanced data obtained with the original data into the enhanced feature data set; In the generated waveform samples, a high-order waveform energy measure is introduced to identify the key time-frequency characteristics of composite faults. To depict more sensitive transient details, a high-order waveform energy function is defined , and its form can be based on the integral accumulation of the high-order derivative in the wavelet domain:

[0020] In the formula: represents the discrete wavelet transform coefficients of the fault waveform , is the scale of the th layer, is the translation variable; is the order of the high-order derivative , which is used to highlight the drastic changes in the waveform in the high-frequency or spike region; is the upper limit of the wavelet translation range, which is determined according to the waveform sampling duration and resolution; This high-order waveform energy is used to quantify the transient feature intensity of composite faults at a specific wavelet scale. When embedding it as a key feature into the subsequent fault identification model, it can enhance the sensitivity to waveform variations caused by the superposition of multiple types of faults such as harmonic distortion and short circuits; when used, it can significantly increase the diversity and quantity of rare fault waveforms, so that the subsequent model training can avoid being overly biased towards common fault types. Through high-order waveform energy and other advanced waveform measures, it can more effectively distinguish the transient characteristics of various faults and avoid subtle differences that cannot be captured by simple time-domain or frequency-domain methods. Using wavelet transform or GAN to perform data augmentation on rare samples such as high-resistance faults, balancing the fault category distribution, and extracting key features such as the superposition of transients and harmonics by defining high-order energy measures, making the composite fault features more prominent in the high-frequency or spike region.

[0021] Step 103, Standardize data format and diversity After completing data augmentation and high-order feature extraction, it is necessary to perform unified formatting and diversity operations on the enhanced feature data to form the final preprocessed output data, which includes high-order feature values such as high-order waveform energy , label information (the fault type or combination corresponding to each waveform), and denoised time-frequency domain sampling data. The specific execution process is as follows: Store keyword fields such as waveform sampling points, labels, GAN enhancement flags, and high-order energy metrics as a consistent data structure, such as a hierarchical table or binary file format, to ensure that multi-label and time-frequency features can be successfully indexed in the same structure; randomly sample the enhanced feature data according to the required ratio (such as 6:2:2), and ensure that the distribution of each fault category is relatively balanced to obtain three subsets for training, validation, and testing; perform consistency checks on the diversified samples to ensure that each sample has complete waveform data and corresponding labels, and ensure that the high-order waveform energy corresponds one-to-one with the original waveform; Through format unification and diversification, it provides preprocessed output data for subsequent multi-label model training, and realizes the functional division of data in the form of training, validation, and test sets. Synchronously verify the high-order energy metric and the original waveform, reduce confusion or incorrect calls when using data, improve the reliability of model training and evaluation, and write the GAN enhancement flag into the metadata, which helps to perform differential evaluation in the subsequent model training and validation stages and observe the impact of real waveforms and synthetic waveforms on the recognition performance.

[0022] Step 101 focuses on preliminary cleaning and multi-source data fusion, providing a good foundation for data enhancement and feature extraction in Step 102: Step 102 performs targeted expansion and high-order measure extraction on the basis of the cleaned data, and outputs a richer and more discriminative waveform set. Step 103 then normalizes and diversifies its results into preprocessed output data.

[0023] Step 2: After the initial dataset is formatted and enters the multi-label modeling stage, the multi-task learning module, after obtaining high-order features and the input of the time-domain sampling branch, corrects the problem of unbalanced distribution through an adaptive weighted loss function, incorporates physical priors such as arc discharge to limit abnormal drift of network weights, and simultaneously performs parallel convolution extraction on harmonic distortion features and outputs the activation value of the hidden vector at the fusion layer, and finally generates a multi-label classification structure model for composite fault recognition; The said Step 2 includes the following contents: Step 201, Multi-label Feature Embedding and Primary Network Construction Based on the multi-source waveform data and high-order features provided by the preprocessed output data, construct the initial input layer and feature embedding layer of the network, and uniformly map the high-order waveform energy metric and the original waveform time-domain sampling points from the preprocessed output data to the input vector of the neural network to ensure that the model pays attention to both transient features and global features at the same time; Input layer configuration: Define multiple parallel input branches to process time-domain sampling data and high-order feature values respectively. For example, convolutional kernels more sensitive to high-frequency transients can act on the high-order feature branch, while convolutional kernels more sensitive to the overall waveform trend can act on the original waveform branch; Basic Convolution and Fusion Layer: After parallel input is completed, a basic convolutional layer (which can be a one-dimensional convolution Conv1D or a two-dimensional convolution Conv2D, depending on the waveform data arrangement) is introduced to extract local patterns. Subsequently, a feature fusion layer is established to map the features extracted from each branch to the same hidden vector space, forming a fused hidden vector H201. Initial Multi-Label Mapping: Based on the set of target fault labels in the power grid (such as harmonics, short circuits, high-resistance grounding, etc.), after retaining the number of channels matching the number of labels, an initial classification activation output is generated. Fusing the original waveform and high-order features at an early stage of the network makes the model more sensitive to weak or rapidly changing signals generated by complex faults. Through parallel input and feature fusion, the model's ability to capture different types of fault features is effectively improved. With the help of a multi-channel parallel structure and high-order feature fusion, the limitation that traditional single-channel convolution may dilute complex fault features is overcome.

[0024] Step 202, Multi-Label Weighted Loss Function and Domain Adaptation Module On top of the classification activation output, considering the characteristics of multi-label scenarios and the uneven data distribution of different fault categories, an adaptive weighted loss function is defined. To reduce the possible distribution differences between real waveforms and simulated waveforms, a domain adaptation module is introduced to make the model more suitable for the actual power system operation environment. The execution process is as follows: In multi-label recognition problems, each sample often contains multiple target fault labels simultaneously. To account for the error impact of rare fault categories (such as high-resistance faults), a weight coefficient that can be adaptively updated needs to be incorporated into the loss function. , the following multi-label weighted loss function is defined :

[0025] Where: represents the classification activation output; is the corresponding true label (1 if this fault type occurs, otherwise 0); is the weighted coefficient of the fault category, which can be dynamically updated with iterations; is a single-label sub-loss function that measures the difference between the activation value and the true label (such as based on logistic regression or cosine similarity, etc., which more specifically measures the discrimination of different sub-fault types); This design can adaptively increase the weight of rare categories during training and reduce the risk of common categories dominating the loss; Regarding the problem of distribution differences between simulated data and real data, a cross-domain mapping function is introduced to align the fused hidden vector H201, making the model have a more consistent feature representation when processing simulated data or real data, and defining the cross-domain difference metric As follows:

[0026] In the formula: and respectively represent the input features in the simulation domain and the real domain; is a learnable cross - domain mapping function, represents that after mapping the feature vector h to the latent space, at dimension a certain output value or component; is the domain of definition of the latent space index; is the fractional - order (or higher - order) derivative of the variable z, where is the fractional - order; when it degenerates into the conventional first - order derivative; when it is the fractional - order derivative; When , is the balance coefficient, used to weigh the importance of the numerical difference of the mapping output itself (the first term) and the change difference of the output in the latent space (the second term); if is closer to 1, more attention is paid to the numerical gap; if k is closer to 0, more attention is paid to aligning the deformation process of the mapping function; represents the vector norm or absolute - value operation (the norm, norm, etc. can be used according to specific requirements) to ensure the non - negativity of the difference description; Cross - domain difference metric measures the distance between the features of the two domains in the mapped space. The sampling range Z depends on the dimension of the latent vector. Finally, is incorporated into the overall loss, aiming to minimize this value, thereby reducing the difference between the simulation and the real data distribution and improving the performance of the model in the actual scenario. The combined loss obtained by combining multi - label weighting and domain adaptation is:

[0027] where is the balance coefficient of the domain - adaptation term, used to adjust the relative importance of multi - label classification and cross - domain alignment.

[0028] The multi - label adaptive weighted loss function can effectively improve the model's attention to rare fault types, avoid over - fitting to common faults caused by unbalanced sample distribution, align the simulation and real waveforms at the latent vector layer, and can alleviate the recognition bias caused by distribution differences during the actual deployment of the model. The organic combination of multi - label weighting and domain adaptation improves the recognition accuracy and generalization ability of the model in complex power scenarios. The adaptively updated weight The mechanism can dynamically adjust the attention to each fault category during the training process, enabling the model to spontaneously adapt to different changes in fault distributions.

[0029] Step 203: Deep mapping incorporating power physics priors Incorporate the physical priors of specific fault mechanisms in the power system into the network structure or activation function, so that the model does not violate the basic power system mechanisms when identifying complex faults. Driven by the obtained comprehensive loss, further perform structural constraints and parameter corrections on the network, and finally output the model structure and preliminary training weights; According to the characteristics of power mechanisms such as arc discharge and line coupling, add physical prior constraints to the sub-channels specifically set for high-resistance faults or harmonic distortion faults in the network. Among them, add a certain monotonicity or non-linear transition limit to the activation function of the convolutional channel to avoid the network learning mappings that conflict with the arc discharge law. Such priors can be achieved by imposing an additional regularization constraint function on the weights of each convolutional kernel:

[0030] In the formula: \(W\) is the set of weights of all convolutional kernels in this sub-channel; represents the positive activation weight in the \(m\)-th convolutional kernel determined to correspond to the high-resistance fault mechanism; \(\lambda\) is a predefined constraint coefficient used to limit the sign flip or abnormal sudden increase of the weights in the convolutional kernel that violate the physical characteristics of arc discharge; \(\alpha\) is the overall weight of the physical prior regularization term to prevent excessive interference with the normal fitting ability of the network; Under the physical prior constraints, retrain or fine-tune the domain adaptation module and the multi-label adaptive weighted loss function, so that the network not only takes into account rare faults and cross-domain alignment, but also strictly adheres to the basic fault laws of the power system; finally, after several rounds of iteration, output the initial multi-label model as the model result of this step, including the network structure (such as the size of each layer's convolutional kernel and the number of feature channels) and preliminary weights; Through physical prior constraints, significantly reduce the mappings learned by the model that conflict with the actual operation laws of the power grid, enhance the interpretability and robustness of complex faults, and form a complement to the domain adaptation mechanism to ensure that the model can still maintain recognition consistency in a wider range of power operating conditions; Sub-step 201 completes multi-source feature fusion in the early stage of the network, providing the basic activation values for Sub-step 202 for multi-label weighting and domain adaptation; Sub-step 202 uses the adaptive weighted loss function and the cross-domain alignment mechanism to effectively handle rare faults and the simulation-real distribution deviation. Sub-step 203 further incorporates power physics priors on this basis, enabling the model to have a higher degree of fit to the real fault mechanism, and at the same time outputting preliminary training weights, laying a solid foundation for subsequent training and evaluation.

[0031] For example, the high-resistance fault current shows pulsed or fuse-like non-linear fluctuations in the low-value range. It can be declared that the fluctuation range is approximately 0 - 75A, and the activation function is clipped in combination with a specific threshold; For harmonic faults, a ratio limit of the harmonic component to the fundamental wave amplitude can be introduced in the network (such as not exceeding a certain safety factor), so that the features learned by the convolution kernel or the mapping layer do not violate the basic distribution law of harmonics in the power grid; Taking arc discharge as an example, the typical arc current frequency band observed in simulation and actual tests can be disclosed (such as inter-harmonic resonance peaks, etc.), and it is explained how to impose restrictions on the network weights corresponding to this frequency band in the regularization term.

[0032] Step 3: After the constructed multi-label model completes the preliminary training and enters the evaluation stage, the cross-validation and noise simulation module examines the stability of the model for various fault scenarios through the analysis of the splitting and high-order calculus metric formula, and verifies the attention distribution of key features such as harmonic distortion and high-resistance grounding in the interpretability visualization; If there are no weights in the physical prior constraint layer that violate the laws of electricity, the intermediate model parameters with good robustness and the ability to analyze fault subdivision features are output; The said Step 3 includes the following contents: Step 301: Multi-source training data integration and distribution balance strategy Based on the preprocessed output data and the initial multi-label model, first perform the distribution balance processing on the training set, and then start the initial training iteration. Specifically, it includes: According to the training data set (including real and simulated fault waveforms, multi-label annotations, and high-order features), select the subset with the widest coverage of fault label categories and diverse time-frequency features from the training data set and match it with the initial multi-label model for loading; For categories with scarce sample sizes such as high-resistance faults, by performing oversampling or increasing the training frequency on the preprocessed output data of the original data set, the proportion of each fault label in the training batch is made more balanced, so as to cooperate with the adaptive weighted loss function and increase the attention to rare categories; Send the prepared training set and the initial multi-label model into the deep learning framework together, and perform several rounds of iteration according to the comprehensive loss function (including multi-label weighting and domain adaptation) to obtain the preliminarily optimized network weights and generate the first training intermediate model; When in use, the combination of oversampling and adaptive weighting focuses on solving the problem of insufficient training volume for rare category faults, effectively uses the preprocessed output data and the initial multi-label model for the initial training iteration to obtain the first training intermediate model. The fusion of multi-source data ensures that the model covers more working conditions and initially reduces the risk of overfitting; Step 302: Cross-validation and noise reliability assessment Based on the training of the intermediate model one, the generalization ability and robustness of the model are tested through cross-validation and noise simulation, and the model parameters are gradually iterated during the training process to generate a more reliable intermediate result for training the intermediate model two. To achieve a higher-level metric, a high-order calculus-based evaluation formula is adopted to measure the consistency and effectiveness of the model's performance in different data analyses; where: The training set is divided into folds (folds), where (a common choice is 5 or 10). One fold is taken as the validation set one by one, and the remaining folds are used as the training set to perform a training-validation loop on the intermediate model one. After each training, the multi-label classification results (such as accuracy, precision, etc.) are recorded and uniformly summarized into the metric items; During the training or validation phase, additional noise simulation is introduced (corresponding to the noise removal in step one), amplitude perturbation or random high-frequency interference injection is performed on the waveform, and the degree of decrease in the model recognition rate is observed. To quantitatively represent the reliability performance, the following robustness metric function is introduced :

[0033] In the formula: represents the response value shown by the model for a certain key performance indicator (such as multi-label accuracy) when the noise amplitude is during the fold validation (or the th group of data); different correspond to different intensities or types of noise injection, is the value range of the noise amplitude or interference intensity; represents the th order (or fractional order) derivative of with respect to the noise amplitude , ; If takes a non-integer value, the fractional order derivative defined by Riemann-Liouville or Caputo can be used to capture more complex progressive characteristics in high-resistance fault or arc discharge scenarios; represents the value of the continuous wavelet transform (CWT) coefficients of in the noise dimension at one or several scales , which is used to measure the local oscillation degree of the response with respect to the change in noise; is the wavelet scale or frequency band, which is commonly used to identify violent fluctuations of different magnitudes. When is small, it focuses on depicting rapid detailed fluctuations, When it is relatively large, focus on the overall change trend; , is the balance coefficient, which is used to make a trade-off between the fractional derivative metric term and the wavelet metric term; when is close to 1, the evaluation of the model robustness depends more on the measurement of the slope or step change of in the noise dimension; when α is close to 0, more emphasis is placed on capturing mutations or detail jitters in the wavelet domain; Combining the cross-validation results and the robustness metric function , further fine-tuning and iterating the intermediate training model one until the validation metrics for each fold basically converge, finally forming a more reliable model weight and structure configuration, and the output is the intermediate training model two; During use, cross-validation fully examines the performance of the model under different fault scenarios and different data splitting methods on multi-fold data, which helps to discover potential overfitting problems or biases towards certain types of faults; through the robustness metric function measuring the change curve of accuracy or other metrics when injecting noise, the tolerance of the model to interference can be evaluated more precisely. Introducing fractional derivatives into the evaluation of the performance metrics with respect to noise variation breaks through the limitations of traditional accuracy difference or error variance measurement, making the reliability test more in-depth and sensitive.

[0034] Step 303, Interpretability analysis and model completeness verification Using the robust network weights and structure of the intermediate training model one, perform interpretability analysis on the model; at the same time, combined with the mechanism constraints of the power system, verify the completeness of the model to ensure that it will not produce abnormal decisions that significantly violate physical laws in real application scenarios. Finally, output the interpretable training model as the final result of step three; among them: By drawing the attention distribution map, show the high-order response correlation of the key intervals concerned by the model at the waveform time domain or frequency domain level: check the change slope or change curve of the model output label when the input fault waveform is slightly adjusted at the key features (such as arc discharge features) to determine whether the model strictly identifies around the physical prior; Combined with power physics prior constraints (such as the non-linear characteristics of arc discharge), sample and detect the weight distribution of the internal convolution kernels or mapping functions of the model to check whether there are extreme weights that violate the prior; for fault types with high mutual correlation (such as harmonic distortion and overvoltage), check the prediction coherence of the model when both types of faults occur simultaneously to prevent the model from only identifying one while ignoring the actual occurrence of the other fault; for example: if convolution kernels or feature channels are used to detect arc features, The range is defined as the high-frequency region of interest for the arc. The amplitude of the convolution kernel weights in this frequency band is regarded as the key monitoring object. For example, when most of the activated channels fall in the lower frequency band but there are some high-amplitude high-frequency convolution kernel weights, and their distribution cannot be explained by conventional faults (such as three-phase short circuit, low-frequency harmonics), it is determined that the prior of the arc nonlinear characteristics may be violated; when weights with a strong corresponding relationship with this harmonic band are found in the convolution kernel or feature map layer but do not match the annotation, it can be determined that there is an illegal distribution. For example, if the expected harmonics are strong around the 5th order, but the network unreasonably amplifies at the 10th order and above, it may indicate that the model confuses other fault characteristics or there are weight deformities.

[0035] In the inference stage, multi-label results are output for each sample, such as Fault Type 1: Harmonic distortion; Fault Type 2: Overvoltage; Fault Type 3: High-resistance fault, etc. If some labels have strong correlations (such as harmonic distortion and overvoltage) that usually exist simultaneously in the true data annotation, but only a single label is given in this inference, there may be a missed fault. Therefore, a layer of coherence check can be added after the inference. Define the rule: When the correlation between label A and B exceeds R% (obtained based on historical statistics or electrical physics mechanisms), if A is detected but the confidence of B is extremely low, then recalculate or refer to the associated fault probability to improve the recognition accuracy for the scenario where both exist simultaneously.

[0036] After the interpretability verification and physical completeness check, the final form and parameters of the obtained model are named the interpretable training model. The interpretable training model has both the ability to recognize complex fault scenarios and a certain degree of interpretability; When in use, through interpretability analysis and physical constraint check, when the model faces complex or new power faults, it can provide a reasonable explanation for the decision-making process, reduce the uncertainty in actual applications, ensure the logical association and consistency between different faults in multi-label fault recognition, and avoid the phenomenon of one-sided recognition or ignoring associated faults. Combining high-order response correlation analysis with electrical physics prior provides a more systematic means for the interpretability evaluation of the model under complex waveform characteristics.

[0037] For the distribution balance, noise reliability, and interpretability during the training process, the actual applicability and understandability of the model are optimized in a hierarchical and progressive manner. The output interpretable training model not only ensures the high-precision recognition ability of multi-label composite faults but also takes into account the robustness and interpretability consistent with the electrical system mechanism. Combining the attention mechanism or Grad-CAM, visualize the key points of attention of the model in the waveform time domain or feature domain, and detect the distribution of network parameters based on electrical physics prior to prevent abnormal decisions that seriously violate the mechanism.

[0038] Step 4: When the model completes offline verification and needs to be deployed in an edge-cloud collaborative environment, the online deployment and continuous optimization module starts fault identification inference after the real-time waveform recording meets the trigger threshold, and monitors the identification quality of new or rare faults through a high-order deviation metric function. Once a significant drift is detected, it automatically uploads the error samples and labels to trigger incremental training, and retains historical knowledge through a comparison memory mechanism, ultimately forming a long-term adaptive fault diagnosis closed-loop; The content of Step 4 is as follows: Step 401: Online deployment of the model and initial operation monitoring Install a fault identification inference module on the edge device of the nearby substation to achieve low-latency fault detection. Retain the global data aggregation and analysis function on the cloud platform for storing historical fault records and subsequent evaluation; the fault identification inference module can be used to make a preliminary determination of waveform distortion, abnormal current or harmonic content, decide when to call the inference process, and at what level of detection threshold to perform emergency inference or regular inference. For example, a high-resistance fault current threshold of 10A - 30A (combined with factors such as the system voltage level) is configured in the substation automation system. When the current in the data stream is continuously detected to exceed this threshold and shows an atypical waveform distortion, the inference module is triggered to start fault analysis.

[0039] Import the network structure and weights of the interpretable training model into the edge inference engine, set the real-time sampling frequency, batch size, and number of core threads according to the grid operation parameters to ensure fast response; continuously monitor the fault identification latency, accuracy, and resource occupancy within a set time period, collect the multi-label output results of online identification, and compare them with known fault events (if any) for verification, and record the real-time error list; Through the collaborative deployment of the edge and the cloud, it can balance real-time performance and global data management, verify the inference performance of the model in the initial stage in a real environment, lay a foundation for subsequent continuous optimization, and practically verify the multi-label feature extraction and power physical constraints in the interpretable training model. The output error set, the real-time error list, is used for model deviation detection and update triggering; For online inference, the priority is usually higher than batch retraining; the computing tasks can be queued and scheduled on the cloud or the edge. Set the inference task to a high priority to ensure real-time performance. When a new fault mode appears, only key data segments need to be uploaded for small-batch supplementary training, and there is no need to frequently transmit large amounts of full-scale fault data. For example, how much computing power core should be reserved for online inference in a power substation or an edge server: the training task should run in the background and automatically cede resources to the inference thread.

[0040] Step 402: Intelligent deviation detection and dynamic retraining trigger Monitor the possible accuracy degradation or fault mode drift of the model by using the collected online recognition results and error information. When significant deviations or new fault modes are found, automatically trigger the retraining or calibration process to maintain the model performance. The specific steps are as follows: Compare the real-time recognition results reported by the edge side with the fault event records, count the accuracy and missed detection rate for each sub-fault type, and centrally save them in the incrementally updated model; define a high-order deviation metric function , to capture the differences and trends between the model and the actual faults within different time windows. The example formula is as follows:

[0041] In the formula: Characterizes the multi-label fault recognition deviation at time (for example, a mapping value of a certain high-level error metric or multi-label error set), which can be defined as the difference degree between the model output and the true label according to the actual scenario; Represents the -th order derivative operation with respect to time , can be an integer order or a fractional order (such as Caputo, Riemann-Liouville, etc.), used to characterize the sudden change, slow change or memory characteristics of over time. When , it can more sensitively capture the metastable or sudden change region of the deviation curve. If , it is more sensitive to the severe jitter of the second order and above; is a fractal dimension metric function, used to evaluate the complexity or irregularity degree of the deviation curve at a very small scale. Common definition methods include box dimension-based, fractal box counting-based, or wavelet decomposition-based local fractal dimension; For example, the wavelet domain fractal metric can be adopted:

[0042] Among them, is the continuous wavelet coefficient of , is the wavelet scale, and are adjustable coefficients. This integral reflects the fractal structure of the deviation curve at different scales; is a trade-off coefficient, which is distributed between the derivative term and the fractal dimension metric term. When is close to 1, more attention is paid to the calculus properties of the deviation over time (such as the severity of multiple order derivatives). When is close to 0, more attention is paid to the roughness or detail complexity of the deviation itself in the fractal sense; is the time window size, which is used to accumulate the overall trend of deviations during the period from to ; in the power system, it can be set according to the average duration of fault events, sampling period, etc. When the high-order deviation metric function exceeds the deviation threshold, it indicates that there is a large or sudden recognition error in the model recently, triggering an update. If the high-order deviation metric function exceeds the deviation threshold or a new fault category is detected (such as a special harmonic overvoltage mode that has not appeared before), the corresponding data and labels will be packed and sent back to the cloud. Load the network structure and weights of the interpretable training model, start retraining or incremental training for new data, and generate a new incremental update model; after training is completed, the incremental update model will be sent back to the edge side and automatically replace the currently used fault recognition model to form a deployed update model. Capture and quantify the accuracy change of the model in a dynamic environment in real time, avoid automatically triggering retraining when being out of calibration for a long time, introduce new fault mode data in a timely manner, so that the model has the ability of self-iteration and improvement. The high-order deviation metric function adopts fractional derivative integration, which can more sensitively capture the prediction error fluctuations caused by the drift of fault modes or the surge of noise interference, avoid being out of calibration for a long time, and enable the model to quickly adapt to the dynamic changes of the power grid.

[0043] Step 403, Knowledge Consolidation and Collaborative Update To prevent the model from forgetting old fault modes or high-resistance fault characteristics during incremental training, key data and their feature distributions in the early stage can be appropriately retained during training, or regularization constraints can be used to suppress large offsets of network weights, and a contrastive memory constraint function is introduced, as follows:

[0044] In the formula: is the weight set of the previous stable model version, and W is the weight after the current retraining; is the -order derivative (optional fractional order) to measure the distribution form of the difference between the weights of the new and old models in the weight dimension ; When the operation and maintenance engineer or the fault investigation record confirms that the model recognition result is incorrect or a misjudgment occurs, this data is archived and manually labeled to form operation and maintenance labeled data. The operation and maintenance labeled data is periodically incorporated into the retraining set to strengthen the learning depth for error-prone scenarios or difficult faults, and improve the overall accuracy of the model. After multiple rounds of updates, merged memories, and operation and maintenance feedback, the cloud integrates the merged model version to obtain the deployed updated model, which is deployed at the edge and enters normal operation. The deployed updated model retains the ability to identify past fault modes and has adaptability to new fault modes, becoming the model finally put into use in the fourth step. By comparing the memory constraint function , the problem of historical knowledge forgetting caused by frequent updates is effectively alleviated, the recognition accuracy of known fault modes is maintained, and the data can be gradually purified and the recognition performance for rare and complex scenarios can be strengthened by using the secondary labeling of operation and maintenance engineers. The false alarms of error-prone faults are significantly reduced, and the deployed updated model output integrates new and old knowledge in continuous iteration, supporting the recognition of multiple fault scenarios in the smart grid in the long term.

[0045] Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are executed in hardware or software depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of this application.

[0046] Those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working processes of the systems, devices, and units described above can refer to the corresponding processes in the foregoing method embodiments and will not be repeated here.

[0047] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. For example, the device embodiments described above are only illustrative. For example, the division of the units is only for some logical function divisions, and there can be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed couplings, direct couplings, or communication connections to each other can be through some interfaces. The indirect couplings or communication connections of the devices or units can be in electrical, mechanical, or other forms.

[0048] The unit described as a separation component may or may not be physically separated. The component shown as a unit may or may not be a physical unit, that is, it may be located in one place or distributed across multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0049] As described above, the foregoing is only a specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present application can easily think of changes or substitutions, which should all be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for recording fault current waveform of an AC power grid, characterized in that: include, When waveform distortion or high-resistance fault signs are detected in the power grid, and the monitoring results exceed the preset fault threshold, multi-source recording and simulation data are extracted, and denoising and anomaly elimination based on physical laws are performed to obtain rare fault samples, which are then enhanced and expanded to generate initial data; After the initial data is formatted, it enters the multi-label modeling stage, corrects the uneven distribution through domain adaptation and adaptive weighted loss, and extracts key features in parallel in the convolutional layer to generate a multi-label classification structure model; After completing the initial training and entering the evaluation phase, the model robustness is tested for multi-fold subsets and specific interference injections, and the attention distribution of harmonic distortion and high-resistance grounding characteristics is verified by visualization. If there are no weights that conflict with physical priors, the intermediate model parameters are output; Fault identification reasoning is performed after the real-time recorded data meets the trigger threshold, and a high-order deviation metric function is used to monitor new or rare faults. If a significant drift is detected, the misjudged samples and labels are automatically uploaded for incremental training, forming an adaptive diagnostic closed loop under the comparative memory mechanism.

2. The method for recording an AC power grid fault current waveform according to claim 1, characterized in that: Acquire fault waveform data from multiple sources, perform preliminary cleaning on the acquired raw data and obtain initial cleaned data; Based on the rare fault types or complex fault waveforms in the initial cleaned data, targeted data enhancement and high-order feature extraction are performed to generate an enhanced feature data set; For fault types with insufficient sample size, synthetic fault data with real waveform statistical characteristics are constructed, and the acquired partial enhanced data are integrated with the original data and included in the enhanced feature data set.

3. The method for recording an AC power grid fault current waveform according to claim 2, characterized in that: After introducing high-order waveform energy metrics to identify the key time-frequency characteristics of compound faults, waveform sampling points, labels, GAN enhancement identifiers, and high-order energy metrics are stored as a hierarchical table; The enhanced feature data is randomly sampled according to the required proportion to obtain three subsets: training, verification and testing; the consistency of the divided samples is checked to make the high-order waveform energy measurement correspond to the original waveform one-to-one, and the pre-processed output data is generated.

4. The method for recording an AC power grid fault current waveform according to claim 3, characterized in that: Define multiple parallel input branches to process time domain sampling data and high-order eigenvalues ​​respectively; After completing the parallel input, the basic convolution layer is introduced to extract local patterns, and a feature fusion layer is set up to map the features extracted from each branch to the same latent vector space to form a fused latent vector. According to the target fault label set in the power grid, the preliminary classification activation output is generated after retaining the number of channels that matches the number of labels.

5. The method for recording fault current waveform of an AC power grid according to claim 4, characterized in that: Incorporate adaptively updated weight coefficients into the loss function to define a multi-label weighted loss function; A cross-domain mapping function is introduced to align the fused hidden vectors, and a cross-domain difference metric is defined. The cross-domain difference metric is incorporated into the overall loss. After minimizing the value, the difference between the simulation and real data distribution is narrowed.

6. The method for recording fault current waveform of an AC power grid according to claim 5, characterized in that: Add monotonicity or nonlinear transition restrictions to the activation function of the convolution channel, which is achieved by applying additional regularization constraints to the weights of each convolution kernel; Under physical prior constraints, the domain adaptation module and the multi-label adaptive weighted loss function are retrained or fine-tuned, and the initial multi-label model, including the network structure and preliminary weights, is output after several rounds of iterations.

7. The method for recording fault current waveform of an AC power grid according to claim 6, characterized in that: Filter subsets with wide coverage of fault label categories and diverse time-frequency characteristics from the training data set and match them with the initial multi-label model for loading; for categories with scarce sample quantities such as high-resistance faults, perform oversampling or strengthen the training frequency in the output data of the original data set preprocessing, and use an adaptive weighted loss function to increase the focus on scarce categories; The prepared training set and the initial multi-label model are sent into the deep learning framework, and several rounds of iterations are performed according to the comprehensive loss function to obtain the preliminary optimized network weights and generate the training intermediate model 1.

8. The method for recording fault current waveform of an AC power grid according to claim 7, characterized in that: Divide the training set into folds, one of which is used as the validation set, and the rest The fold is used as the training set, and the training-validation cycle is performed on the training intermediate model. The multi-label classification results are recorded after each training. In the training or verification stage, additional noise simulation is introduced to perform amplitude perturbation or random high-frequency interference injection on the waveform, and a robustness measurement function is introduced: the cross-validation results and the robustness measurement function are combined to further fine-tune and iterate the training intermediate model 1 until the verification indicators of each fold basically converge, and finally a more reliable model weight and structural configuration are formed, which are output as the training intermediate model 2.

9. The method for recording fault current waveform of an AC power grid according to claim 8, characterized in that: By drawing an attention distribution diagram, the correlation of high-order responses of the key intervals that the model focuses on in the waveform time domain or frequency domain is displayed: when the input fault waveform is fine-tuned on the key features, the changing slope or changing curve of the output label is checked to determine whether the changing slope or changing curve of the output label is strictly identified around the physical prior. After interpretability verification and physical completeness verification, an interpretable training model is obtained.

10. The method for recording fault current waveform of an AC power grid according to claim 8, characterized in that: Install a fault identification and reasoning module on the nearest substation edge device to store historical fault records and conduct subsequent evaluations; import the network structure and weights of the interpretable training model into the edge inference engine, and set the real-time sampling frequency, batch size, and number of core threads according to the grid operation parameters; Continuously monitor fault identification delay, accuracy and resource usage within a set period of time, collect multi-tag output results of online identification, check them with known fault events, and record a real-time error list.

11. The method for recording fault current waveform of an AC power grid according to claim 10, characterized in that: Compare the real-time recognition results reported by the edge side with the fault event records, and calculate the accuracy and missed detection rate of each sub-fault type and save them in the incremental update model; if the defined high-order deviation metric function exceeds the deviation threshold or a new fault category is detected, the corresponding data and labels are packaged and sent back to the cloud; Retrain or incrementally train on new data to generate a new incremental update model; After the training is completed, the incremental update model is transmitted back to the edge side and automatically replaces the currently used fault identification model to form a deployed update model.

12. The method for recording fault current waveform of an AC power grid according to claim 11, characterized in that: During training, appropriate retention of early key data and their characteristic distribution, or use of regular constraints to suppress large deviations in network weights, and introduction of contrast memory constraint functions to mitigate catastrophic forgetting and retain important fault patterns learned in the power system; When it is confirmed that the model recognition result is wrong or missed, the data is archived and manually labeled to form operation and maintenance annotation data, and the operation and maintenance annotation data is periodically incorporated into the retraining set; after multiple rounds of updates, merged memory and operation and maintenance feedback, the deployed updated model is integrated to obtain.

Images