Small current grounding system fault recording data generation method and system

Abstract

The invention discloses a small-current grounding system fault recording data generation method and system, and belongs to the technical field of power system fault analysis, and the method comprises the steps: obtaining simulation waveform data and on-site recording data, constructing a multi-dimensional condition vector input generative adversarial network, and employing a generator to carry out fault recording on the basis of maintaining the position of a fault point unchanged, outputting a residual waveform by taking a fault category and a fault point mask as conditions, and superposing the residual waveform with the simulation waveform to generate an enhanced waveform; a discriminator is utilized to execute waveform authenticity discrimination, fault position prediction and fault classification based on enhanced waveforms and on-site recording data, a bidirectional mapping generator constraint generation waveform of a simulation waveform domain and an on-site recording domain is constructed to keep fault semantic consistency, and the generator, the discriminator and the bidirectional mapping generator are jointly optimized to output fault recording data. The problem that the generalization ability of a downstream fault identification model is poor due to insufficient real field fault data is solved.

Description

Technical Field

[0001] This invention relates to a method and system for generating fault recording data in a low-current grounding system, belonging to the field of power system fault analysis technology. Background Technology

[0002] In power distribution network operation, accurate modeling and analysis of fault waveform data from low-current grounding systems are core technical supports for improving relay protection sensitivity and ensuring system safety and reliability. Currently, the acquisition of field waveform data faces a "double scarcity" dilemma: on the one hand, high-resistance grounding faults have a low probability of occurrence in actual power grids due to weak grounding currents and ambiguous characteristics, and the waveform recorder is limited by installation location, sampling accuracy, and trigger threshold, resulting in a severe shortage of usable real fault waveform data; on the other hand, although traditional simulation tools can generate grounding fault waveforms, they exhibit systematic deviations from real data in terms of temporal dynamic characteristics, spectral energy distribution, and interphase coupling relationships of three-phase signals. Directly using these for model training leads to a decline in the generalization ability of downstream fault identification algorithms. To alleviate the data shortage problem, the industry has attempted to introduce data augmentation techniques, such as noise injection, waveform translation, amplitude scaling, and other simple signal transformation methods, as well as the successful experience of generative adversarial networks in image / speech generation, to try to build fault waveform data generation models.

[0003] However, the scarcity of field waveform data and the offset distribution of simulation data make it difficult for traditional data augmentation methods to generate augmented data that simultaneously satisfies temporal consistency and spectral rationality. The generated waveforms are prone to "spectral distortion" or "destruction of physical coupling relationships." Generative adversarial networks (GANs) suffer from a lack of constraints in the application of power fault scenarios. Existing GAN models only focus on waveform morphology fitting and do not embed fault category labels, fault location masks, and three-phase physical coupling constraints. As a result, although the generated data looks realistic, the fault semantics and physical characteristics do not conform to engineering reality. In the multi-task learning framework, the discriminator needs to perform three tasks simultaneously: true / false discrimination, fault classification, and location prediction. However, the existing loss function design lacks a dynamic weight balancing mechanism, which can easily lead to task conflicts. Moreover, the cycle consistency loss only guarantees waveform appearance similarity and does not incorporate physical constraints such as Kirchhoff's laws, so it cannot ensure the dual consistency of the generated waveform in terms of fault semantics and physical characteristics. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for generating fault recording data in low-current grounding systems. By constructing a generative adversarial network with bidirectional mapping and multi-task constraints, the problem of poor generalization ability of downstream fault identification models caused by insufficient real-world fault data is solved.

[0005] To solve the above-mentioned technical problems, the present invention is implemented using the following technical solution.

[0006] In a first aspect, the present invention provides a method for generating fault recording data in a low-current grounding system, comprising:

[0007] Acquire simulated waveform data and on-site recorded waveform data;

[0008] A multidimensional conditional vector is constructed based on simulated waveform data and on-site recorded waveform data;

[0009] The multidimensional condition vector includes simulation waveforms, fault categories, fault point masks, and simulation parameters.

[0010] The multidimensional conditional vector is input into the trained generative adversarial network for processing, including:

[0011] Using a generator, while keeping the fault location unchanged, the residual waveform is output based on the fault category and fault point mask. The residual waveform is then superimposed with the simulation waveform to generate an enhanced waveform.

[0012] The discriminator uses the enhanced waveform and the recorded waveform data to simultaneously perform waveform authenticity determination, fault location prediction, and fault classification tasks.

[0013] A bidirectional mapping generator between the simulation waveform domain and the field recording waveform domain is constructed, and the generated waveform maintains fault semantic consistency during the bidirectional mapping process by utilizing the cyclic consistency loss constraint.

[0014] The generator, the discriminator, and the bidirectional mapping generator are jointly optimized, and the optimized generator outputs a generated waveform that conforms to the distribution of the field waveform data, which is used as fault waveform data.

[0015] Furthermore, the generator is a one-dimensional convolutional neural network based on a temporal convolutional neural network and a residual structure, comprising:

[0016] The input layer is used to receive multidimensional condition vectors;

[0017] The convolutional coding layer is used to extract preliminary temporal features of the simulation waveform based on multi-dimensional conditional vectors through multiple layers of one-dimensional convolutional kernels;

[0018] The temporal convolutional neural network module includes multiple stacked causal dilated convolutional layers for receiving the initial temporal features, expanding the receptive field through progressively increasing dilation coefficients, capturing and modeling long-range dependencies in the waveform, and outputting long-range dependency features.

[0019] The residual block, connected after the temporal convolutional neural network module, is used to perform in-depth processing of the long-range dependent features and learn the local perturbation features in the waveform through identity mapping short-circuiting connections.

[0020] The conditional fusion module is used to embed simulation parameters and splice them with local perturbation features to obtain the conditionally fused features;

[0021] A gated layer based on fault location is used to receive and process the fault category and fault location mask. The gated mechanism controls the features after condition fusion to ensure that the fault location is not shifted during the generation process.

[0022] The decoding layer is used to perform deconvolution operations on the gated features and reconstruct them into residual waveforms of the same length as the simulated waveforms through upsampling operations.

[0023] An enhanced signal output layer is used to superimpose the residual waveform with the simulated waveform to generate an enhanced waveform.

[0024] Furthermore, the enhanced waveform is represented as follows:

[0025] ;

[0026] In the formula, Indicates the current time Enhanced waveform, Indicates the current time The simulated waveform, Indicates the current time The residual waveform.

[0027] Furthermore, the discriminator is a multi-task discriminator based on a one-dimensional convolutional neural network, comprising:

[0028] The shared feature extraction layer includes a cross-channel convolutional layer, a one-dimensional convolutional layer, a batch normalization layer, and a LeakyReLU function layer, which is used to extract the temporal features of the input waveform and the phase coupling features between the three phase signals using one-dimensional convolution. The input waveform includes the enhanced waveform and the field waveform recording data.

[0029] The waveform authenticity detection head includes an average pooling layer and a fully connected layer, which is used to perform waveform authenticity detection tasks and distinguish whether the input waveform comes from field waveform recording data or from the generator.

[0030] The fault location prediction head, including a deconvolution layer, a batch normalization layer, and a sigmoid function layer, is used to perform the fault location prediction task. The output is a point-by-point probability distribution consistent with the number of sampling points, predicting the fault location of the input waveform.

[0031] The fault category identification head, consisting of an average pooling layer and a fully connected layer, is used to perform fault classification tasks and verify the fault category of the input waveform.

[0032] Furthermore, the cross-channel convolutional layer employs cross-channel convolution operations to maintain the coherence and physical coupling of the enhanced waveform.

[0033] Furthermore, the total loss function is obtained by jointly optimizing the adversarial loss, cycle consistency loss, short-time Fourier transform loss, classification loss, and position loss, and the generative adversarial network is iteratively trained.

[0034] Furthermore, the total loss function is expressed as:

[0035] ;

[0036] In the formula, Represents the total loss function. Indicates resistance to loss The weighting coefficients, Indicates the loss of cycle consistency The weighting coefficients, Represents the short-time Fourier transform loss The weighting coefficients, Represents classification loss The weighting coefficients, Indicates position loss Weighting coefficients;

[0037] The adversarial loss is expressed as:

[0038] ;

[0039] In the formula, This represents the probability distribution of waveform data recorded on-site. Samples of on-site waveform data from the sampling site Take the expected value. Represents the logarithmic function. This indicates that the discriminator determines the on-site waveform data samples. It is the probability of the real sample. This indicates the data recorded from the fault waveform. probability distribution Sample fault recording data Take the expected value. This indicates that the discriminator determines the fault recording data sample. It is the probability of being a real sample;

[0040] The cycle consistency loss is expressed as:

[0041] ;

[0042] In the formula, This represents the loss of cycle consistency. [.] denotes the probability distribution from the simulated waveform data. Sampled simulated waveform data Take the expected value. Indicates from the real domain To the simulation domain generator, Indicates from the simulation domain To the real domain generator, Denotes the L1 norm;

[0043] The short-time Fourier transform loss is expressed as:

[0044] ;

[0045] In the formula, Indicates will A different The configured short-time Fourier transform loss is accumulated, where, express Configuration serial number, Indicates the first indivual The configured short-time Fourier transform loss;

[0046] The classification loss is expressed as:

[0047] ;

[0048] In the formula, Indicates the first The true category of the field waveform data samples obtained from the secondary classification process. Indicates the first The probability of the predicted category of the field waveform data sample obtained from the classification execution, where C represents the number of classification executions;

[0049] The position loss is expressed as:

[0050] ;

[0051] In the formula, Indicates the loss at the location of the fault. Indicates the number of experiments. Indicates the actual location of the fault. This indicates the predicted location of the fault.

[0052] Furthermore, the discriminator's total loss function is obtained by jointly optimizing the discriminator's adversarial loss, classification loss, position loss, and symmetric component constraint loss, and the discriminator is then iteratively trained.

[0053] Furthermore, the total loss function of the discriminator is expressed as:

[0054] ;

[0055] In the formula, This represents the total loss function of the discriminator. Indicates discriminator against loss The weighting coefficients, Indicates the classification loss of the discriminator. The weighting coefficients, Indicates discriminator position loss The weighting coefficients, Represents the symmetric component constraint loss of the discriminator. Weighting coefficients;

[0056] The discriminant adversarial loss is expressed as:

[0057] ;

[0058] In the formula, This indicates that the discriminator is susceptible to loss.

[0059] The classification loss of the discriminator is expressed as:

[0060] ;

[0061] In the formula, This represents the classification loss of the discriminator;

[0062] The discriminant position loss is expressed as:

[0063] ;

[0064] In the formula, Indicates the location loss at the fault point;

[0065] The discriminant's symmetric component constraint loss is expressed as:

[0066] ;

[0067] In the formula, This represents the symmetric component constraint loss of the discriminator. , , This represents the zero-sequence component, positive-sequence component, and negative-sequence component corresponding to the fault recording data. , , The table represents the zero-sequence component, positive-sequence component, and negative-sequence component corresponding to the field waveform data.

[0068] In a second aspect, the present invention provides a fault recording data generation system for a low-current grounding system, used to implement the fault recording data generation method for a low-current grounding system described in the first aspect.

[0069] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0070] 1. This invention constructs a multi-dimensional conditional vector containing simulated waveforms, fault categories, fault point masks, and simulation parameters. Based on the fixed constraints on fault point locations and residual waveform superposition mechanism of the generator in the generative adversarial network, as well as the multi-task collaboration of the discriminator and the cyclic consistency loss constraint of the bidirectional mapping generator, it achieves the generation of high-quality fault waveform data that conforms to the distribution of field waveform data while maintaining the semantic consistency of faults. This significantly improves the accuracy and data reliability of fault analysis in low-current grounding systems and solves the problem of poor generalization ability of downstream fault identification models due to insufficient real-world fault data.

[0071] 2. This invention utilizes a one-dimensional convolutional neural network architecture based on temporal convolutional neural networks and residual structures in the generator. By combining the progressively increasing expansion coefficients of multiple causal dilated convolutional layers, the receptive field is expanded, effectively capturing long-range dependencies of waveforms. In conjunction with a gated layer based on fault location, the control of fault categories and fault point masks on features is strictly constrained through the gated mechanism, ensuring that the fault location is not shifted during the generation and superposition of residual waveforms. This achieves high accuracy in fault location and significantly improves the accuracy of fault location analysis in low-current grounding systems.

[0072] 3. This invention utilizes a multi-task architecture based on a one-dimensional convolutional neural network in the discriminator. By extracting the temporal features and interphase coupling features of the input waveform through a shared feature extraction layer, it works in conjunction with a true / false discrimination head, a fault location prediction head, and a fault category identification head to perform waveform authenticity discrimination, fault location prediction, and fault classification tasks, respectively. Combined with a bidirectional mapping generator constrained by cyclic consistency loss, it maintains the consistency of fault semantics during the conversion between the simulated waveform domain and the field recorded waveform domain, achieving multi-task collaborative optimization and effectively improving the distribution consistency between the generated waveform and the field recorded waveform data, as well as the reliability of fault semantics.

[0073] 4. This invention iteratively trains the generative adversarial network by jointly optimizing the total loss function composed of adversarial loss, cycle consistency loss, short-time Fourier transform loss, classification loss, and location loss. Among them, the short-time Fourier transform loss ensures that the generated waveform matches the field waveform data in terms of frequency domain characteristics, while the classification loss and location loss enhance the accuracy of fault category identification and fault point location prediction, respectively. Finally, it outputs a high-quality generated waveform that conforms to the distribution of field waveform data, significantly enhancing the accuracy and practicality of fault waveform data generation for low-current grounding systems. Attached Figure Description

[0074] Figure 1 This is a flowchart illustrating a method for generating fault recording data in a low-current grounding system according to an embodiment of the present invention. Detailed Implementation

[0075] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0076] Example 1

[0077] like Figure 1 As shown in the figure, this embodiment introduces a method for generating fault recording data in a low-current grounding system, including:

[0078] Step 1: Obtain simulation waveform data and on-site recorded waveform data.

[0079] This invention provides a dual-source data foundation for subsequent data fusion and enhancement by acquiring simulated waveform data and on-site recorded waveform data, ensuring that the generated waveform has both the controllability of simulated data and the authenticity of on-site data.

[0080] Step 2: Construct a multidimensional conditional vector based on simulated waveform data and on-site recorded waveform data.

[0081] In this invention, the multidimensional condition vector includes simulation waveforms, fault categories, fault point masks, and simulation parameters.

[0082] This invention constructs a multi-dimensional condition vector based on simulation waveforms, fault categories, fault point masks, and simulation parameters, realizing the structured integration of multi-dimensional information and providing richer condition constraints and feature guidance for generative adversarial networks.

[0083] Step 3: Input the multidimensional conditional vector into the trained generative adversarial network for processing, including:

[0084] Using a generator, while keeping the fault location unchanged, a residual waveform is output based on the fault category and fault location mask. The residual waveform is then superimposed with the simulation waveform to generate an enhanced waveform. A gating mechanism is used to ensure that the fault location remains unchanged, thereby achieving precise control of fault location.

[0085] By using a discriminator based on the enhanced waveform and the recorded waveform data, the waveform authenticity determination task, fault location prediction task, and fault classification task are performed simultaneously, thereby improving the globality and task coordination of waveform discrimination.

[0086] A bidirectional mapping generator between the simulation waveform domain and the field recording waveform domain is constructed. By using the cyclic consistency loss constraint, the generated waveform maintains the semantic consistency of the fault during the bidirectional mapping process, ensuring that the generated waveform retains the core fault characteristics after cross-domain conversion.

[0087] By jointly optimizing the generator, the discriminator, and the bidirectional mapping generator, the optimized generator outputs a generated waveform that conforms to the distribution of the field waveform data, which is used as fault waveform data, thereby improving the practicality and reliability of the fault waveform data.

[0088] This invention significantly increases the sample size of faults in low-current grounding systems while improving the temporal consistency, spectral rationality, and multi-channel physical coupling of the enhanced data. This significantly improves the accuracy and robustness of downstream fault identification and location models, making it suitable for scenarios such as distribution network grounding fault analysis, relay protection testing, and smart grid operation and maintenance. It provides key technical support for intelligent fault diagnosis in power systems.

[0089] Example 2

[0090] Based on the same inventive concept as Embodiment 1, this embodiment introduces the implementation steps of a method for generating fault recording data in a low-current grounding system, including:

[0091] Step 1: Obtain simulation waveform data and on-site recorded waveform data.

[0092] This embodiment uses power system simulation software, such as PSCAD, to generate simulated waveform data with a sampling frequency of 12000Hz. The data includes three-phase waveform information of voltage and current, covering various typical operating conditions and fault types, such as single-phase ground faults, two-phase short-circuit faults, and three-phase short-circuit faults. Simultaneously, on-site intelligent waveform recording devices are used to collect on-site waveform data, also with a sampling frequency of 12000Hz. The collected voltage and current three-phase waveform information is consistent with the simulated data.

[0093] After acquiring the field waveform data, preprocessing is performed, including normalization, window segmentation, and fault point marking. This includes: using maximum absolute value normalization to map all voltage and current waveforms in the field waveform data to the [-1, 1] interval for normalization; windowing the normalized field waveform data, with a window length of N=480 points, to ensure complete coverage of the fault transient process; and marking the fault initiation point based on the segmented field waveform data using manual or automatic detection algorithms, generating corresponding label vectors. Data segments containing faults are labeled as 1; normal data is labeled as 0. Considering potential marking errors, a three-segment marking method is adopted: 10 data points before and after the marked point are considered the fault occurrence segment and labeled as 1; data before the fault is labeled as 0; and data after the fault is labeled as 2. This effectively reduces interference from marking errors.

[0094] Simulated waveform data mainly consists of fundamental components, while real waveforms typically contain significant higher harmonics and random noise. Furthermore, in the simulation, the three-phase voltage amplitudes are strictly equal and the phase difference is maintained at 120°, whereas voltage imbalance exists in real power grids, leading to an increase in sequence components. To bridge the distribution difference between simulated and real data, this embodiment designs a data distribution mapping method that maps simulated waveform data to the real data distribution through nonlinear transformation.

[0095] Step 2: Construct a multidimensional conditional vector based on the simulated waveform data and the field recorded waveform data.

[0096] In this embodiment, after preprocessing the simulated waveform data and the recorded waveform data on site, the three-phase voltage, waveform current, and waveform are directly used as the basis for constructing a multi-dimensional condition vector.

[0097] In this embodiment, the multidimensional condition vector includes simulation waveforms, fault categories, fault point masks, and simulation parameters. The simulation parameters include simulation system parameters, line parameters, and fault parameters to ensure the consistency between fault waveform data and fault waveform data in real operating scenarios.

[0098] In this embodiment, the system parameters include grounding method, harmonic content, and unbalanced voltage; the line parameters include line length, cable ratio, overhead line length, and cable length; and the fault parameters include transition resistance, fault location, fault phase, fault type, and initial phase angle of the fault.

[0099] Step 3: Input the multidimensional conditional vector into the trained generative adversarial network for processing, including:

[0100] Step 3.1: Using the generator, while keeping the fault location unchanged, output the residual waveform based on the fault category and fault point mask, and superimpose the residual waveform with the simulation waveform to generate an enhanced waveform.

[0101] In this embodiment, the residual waveform is predicted based on the simulated waveform and superimposed on the residual waveform, thereby improving the waveform realism and phase consistency while keeping the fault point location unchanged.

[0102] In this embodiment, the generator is a one-dimensional convolutional neural network based on a temporal convolutional neural network and a residual structure, including an input, a convolutional coding layer, a temporal convolutional neural network module, a residual block, a conditional fusion module, a gate layer based on the fault location, a decoding layer, and an enhanced signal output layer.

[0103] In this embodiment, the input layer is used to receive a multi-dimensional conditional vector; the convolutional coding layer is used to extract preliminary temporal features of the simulation waveform based on the multi-dimensional conditional vector through multiple one-dimensional convolutional kernels; the temporal convolutional neural network module includes multiple stacked causal dilated convolutional layers, used to receive the preliminary temporal features, expand the receptive field through progressively increasing dilation coefficients, capture and model long-range dependencies in the waveform, and output long-range dependency features; the residual block, connected after the temporal convolutional neural network module, is used to perform in-depth processing of the long-range dependency features, and learn the waveform through identity mapping short-circuiting connections. The simulation parameters are used to obtain the local perturbation features. The conditional fusion module is used to embed simulation parameters and concatenate them with the local perturbation features to obtain the conditionally fused features. The fault point location-based gating layer is used to receive and process the fault category and fault point mask, and control the conditionally fused features through a gating mechanism to ensure that the fault point location is not shifted during the generation process. The decoding layer is used to perform deconvolution operation on the gated features and reconstruct them into a residual waveform of the same length as the simulation waveform through an upsampling operation. The enhanced signal output layer is used to superimpose the residual waveform and the simulation waveform to generate an enhanced waveform.

[0104] In this embodiment, the enhanced waveform is represented as:

[0105] ;

[0106] In the formula, Indicates the current time Enhanced waveform, Indicates the current time The simulated waveform, Indicates the current time The residual waveform.

[0107] Step 3.2: Using the discriminator, based on the enhanced waveform and the on-site recorded waveform data, simultaneously perform waveform authenticity discrimination, fault location prediction, and fault classification tasks.

[0108] In this embodiment, the discriminator is a multi-task discriminator based on a one-dimensional convolutional neural network, including a shared feature extraction layer, a true / false discrimination head, a fault location prediction head, and a fault category identification head.

[0109] In this embodiment, the shared feature extraction layer includes a cross-channel convolutional layer, a one-dimensional convolutional layer, a batch normalization layer, and a LeakyReLU function layer. It is used to extract the temporal features of the input waveform and the interphase coupling features between the three-phase signals using one-dimensional convolution. The input waveform includes the enhanced waveform and the field-recorded waveform data. The authenticity discrimination head includes an average pooling layer and a fully connected layer, used to perform waveform authenticity discrimination tasks, distinguishing whether the input waveform comes from the field-recorded waveform data or from the generator. The fault location prediction head includes a deconvolutional layer, a batch normalization layer, and a Sigmoid function layer, used to perform fault location prediction tasks, outputting a point-by-point probability distribution consistent with the number of sampling points, and predicting the fault location of the input waveform. The fault category identification head includes an average pooling layer and a fully connected layer, used to perform fault classification tasks, verifying the fault category of the input waveform.

[0110] In this embodiment, the cross-channel convolutional layer uses cross-channel convolution operations to maintain the coherence and physical coupling of the enhanced waveform.

[0111] In this embodiment, the discriminator's total loss function is obtained by jointly optimizing the discriminator's adversarial loss, discriminator classification loss, discriminator position loss, and discriminator symmetric component constraint loss, and the discriminator is then iteratively trained.

[0112] In this embodiment, the total loss function of the discriminator is expressed as:

[0113] ;

[0114] In the formula, This represents the total loss function of the discriminator. Indicates discriminator against loss The weighting coefficients, Indicates the classification loss of the discriminator. The weighting coefficients, Indicates discriminator position loss The weighting coefficients, Represents the symmetric component constraint loss of the discriminator. Weighting coefficients;

[0115] In this embodiment, the discriminator adversarial loss is expressed as:

[0116] ;

[0117] In the formula, This indicates that the discriminator is susceptible to loss.

[0118] In this embodiment, the discriminator classification loss is expressed as:

[0119] ;

[0120] In the formula, This represents the classification loss of the discriminator;

[0121] In this embodiment, the discriminator position loss is expressed as:

[0122] ;

[0123] In the formula, Indicates the location loss at the fault point;

[0124] In this embodiment, the discriminator symmetric component constraint loss is expressed as:

[0125] ;

[0126] In the formula, This represents the symmetric component constraint loss of the discriminator. , , This represents the zero-sequence component, positive-sequence component, and negative-sequence component corresponding to the fault recording data. , , The table represents the zero-sequence component, positive-sequence component, and negative-sequence component corresponding to the field waveform data.

[0127] Step 3.3: Construct a bidirectional mapping generator between the simulation waveform domain and the field recording waveform domain, and use the cyclic consistency loss constraint to generate waveforms that maintain fault semantic consistency during the bidirectional mapping process.

[0128] This embodiment constructs a bidirectional mapping generator between the simulated waveform domain and the field recorded waveform domain. By utilizing the cyclic consistency loss constraint, the generated waveform maintains fault semantic consistency during the bidirectional mapping process, realizing the mapping from simulated data to real data, and from real data to simulated data. At the same time, the cyclic consistency loss constraint ensures that the generated waveform can recover the fault semantics after bidirectional mapping, effectively improving the ability of the generated data to fit the characteristics of the real recorded waveform and alleviating the problem of the difference between simulation and real distribution.

[0129] Step 3.4: Jointly optimize the generator, the discriminator and the bidirectional mapping generator, and use the optimized generator to output a generated waveform that conforms to the distribution of the field waveform data as fault waveform data.

[0130] In this embodiment, the total loss function is obtained by jointly optimizing the adversarial loss, cycle consistency loss, short-time Fourier transform loss, classification loss, and position loss, and the generative adversarial network is iteratively trained.

[0131] In this embodiment, the total loss function is expressed as:

[0132] ;

[0133] In the formula, Represents the total loss function. Indicates resistance to loss The weighting coefficients, Indicates the loss of cycle consistency The weighting coefficients, Represents the short-time Fourier transform loss The weighting coefficients, Represents classification loss The weighting coefficients, Indicates position loss Weighting coefficients;

[0134] In this embodiment, the adversarial loss is expressed as:

[0135] ;

[0136] In the formula, This represents the probability distribution of waveform data recorded on-site. Samples of on-site waveform data from the sampling site Take the expected value. Represents the logarithmic function. This indicates that the discriminator determines the on-site waveform data samples. It is the probability of the real sample. This indicates the data recorded from the fault waveform. probability distribution Sample fault recording data Take the expected value. This indicates that the discriminator determines the fault recording data sample. It is the probability of being a real sample;

[0137] In this embodiment, the cycle consistency loss is expressed as:

[0138] ;

[0139] In the formula, This represents the loss of cycle consistency. [.] denotes the probability distribution from the simulated waveform data. Sampled simulated waveform data Take the expected value. Indicates from the real domain To the simulation domain generator, Indicates from the simulation domain To the real domain generator, Denotes the L1 norm;

[0140] In this embodiment, the short-time Fourier transform loss is expressed as:

[0141] ;

[0142] In the formula, Indicates will A different The configured short-time Fourier transform loss is accumulated, where, express Configuration serial number, Indicates the first indivual The configured short-time Fourier transform loss;

[0143] In this embodiment, the classification loss is expressed as:

[0144] ;

[0145] In the formula, Indicates the first The true category of the field waveform data samples obtained from the secondary classification process. Indicates the first The probability of the predicted category of the field waveform data sample obtained from the classification execution, where C represents the number of classification executions;

[0146] In this embodiment, the position loss is expressed as:

[0147] ;

[0148] In the formula, Indicates the loss at the location of the fault. Indicates the number of experiments. Indicates the actual location of the fault. This indicates the predicted location of the fault.

[0149] Example 3

[0150] Based on the same inventive concept as other embodiments, this embodiment provides a fault recording data generation system for low-current grounding systems, used for the fault recording data generation method for low-current grounding systems as described in Embodiment 1 or 2.

[0151] Example 4

[0152] Based on the same inventive concept as other embodiments, this embodiment describes a computer-readable storage medium having computer instructions stored thereon, which, when executed by a processor, implement the steps of the methods of Embodiment 1 or 2 described above.

[0153] Example 5

[0154] Based on the same inventive concept as other embodiments, this embodiment introduces a computer program product, including computer instructions that, when executed by a processor, implement the steps of the methods described in Embodiment 1 or 2 above.

[0155] In summary, this invention constructs a multi-dimensional conditional vector containing simulated waveforms, fault categories, fault point masks, and simulation parameters. Based on the fixed constraints on fault point locations and residual waveform superposition mechanism of the generator in the generative adversarial network, as well as the multi-task collaboration of the discriminator and the cyclic consistency loss constraint of the bidirectional mapping generator, it achieves the generation of high-quality fault waveform data that conforms to the distribution of field waveform data while maintaining the semantic consistency of faults. This significantly improves the accuracy and data reliability of fault analysis in low-current grounding systems and solves the problem of poor generalization ability of downstream fault identification models due to insufficient real-world fault data.

[0156] This invention utilizes a one-dimensional convolutional neural network architecture based on temporal convolutional neural networks and residual structures in the generator. By combining multi-layer causal dilated convolutional layers with progressively increasing expansion coefficients, the receptive field is expanded, effectively capturing long-range dependencies of waveforms. In conjunction with a gated layer based on fault location, the fault category and fault point mask are strictly constrained to control the features through a gating mechanism, ensuring that the fault location is not shifted during the generation and superposition of residual waveforms. This achieves high accuracy in fault location and significantly improves the accuracy of fault location analysis in low-current grounding systems.

[0157] This invention utilizes a multi-task architecture based on a one-dimensional convolutional neural network in the discriminator. By extracting the temporal features and interphase coupling features of the input waveform through a shared feature extraction layer, it works in conjunction with a true / false discrimination head, a fault location prediction head, and a fault category identification head to perform waveform authenticity discrimination, fault location prediction, and fault classification tasks, respectively. Combined with a bidirectional mapping generator constrained by cyclic consistency loss, it maintains fault semantic consistency during the conversion between the simulated waveform domain and the field recorded waveform domain, achieving multi-task collaborative optimization and effectively improving the distribution consistency between the generated waveform and the field recorded waveform data, as well as the reliability of the fault semantics.

[0158] This invention iteratively trains the generative adversarial network by jointly optimizing the total loss function, which consists of adversarial loss, cycle consistency loss, short-time Fourier transform loss, classification loss, and location loss. Among them, the short-time Fourier transform loss ensures that the generated waveform matches the field waveform data in terms of frequency domain characteristics, while the classification loss and location loss enhance the accuracy of fault category identification and fault point location prediction, respectively. Finally, it outputs a high-quality generated waveform that conforms to the distribution of field waveform data, significantly improving the accuracy and practicality of fault waveform data generation for low-current grounding systems.

[0159] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0160] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0161] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0162] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0163] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for generating fault recording data in a low-current grounding system, characterized in that, include: Acquire simulated waveform data and on-site recorded waveform data; A multidimensional conditional vector is constructed based on simulated waveform data and on-site recorded waveform data; The multidimensional condition vector includes simulation waveforms, fault categories, fault point masks, and simulation parameters. The multidimensional conditional vector is input into the trained generative adversarial network for processing, including: Using a generator, while keeping the fault location unchanged, the residual waveform is output based on the fault category and fault point mask. The residual waveform is then superimposed with the simulation waveform to generate an enhanced waveform. The discriminator uses the enhanced waveform and the recorded waveform data to simultaneously perform waveform authenticity determination, fault location prediction, and fault classification tasks. A bidirectional mapping generator between the simulation waveform domain and the field recording waveform domain is constructed, and the generated waveform maintains fault semantic consistency during the bidirectional mapping process by utilizing the cyclic consistency loss constraint. The generator, the discriminator, and the bidirectional mapping generator are jointly optimized, and the optimized generator outputs a generated waveform that conforms to the distribution of the field waveform data, which is used as fault waveform data.

2. The method for generating fault recording data in a low-current grounding system according to claim 1, characterized in that, The generator is a one-dimensional convolutional neural network based on a temporal convolutional neural network and a residual structure, including: The input layer is used to receive multidimensional condition vectors. The convolutional coding layer is used to extract preliminary temporal features of the simulation waveform based on multi-dimensional conditional vectors through multiple layers of one-dimensional convolutional kernels; The temporal convolutional neural network module includes multiple stacked causal dilated convolutional layers for receiving the initial temporal features, expanding the receptive field through progressively increasing dilation coefficients, capturing and modeling long-range dependencies in the waveform, and outputting long-range dependency features. The residual block, connected after the temporal convolutional neural network module, is used to perform in-depth processing of the long-range dependent features and learn the local perturbation features in the waveform through identity mapping short-circuiting connections. The conditional fusion module is used to embed simulation parameters and splice them with local perturbation features to obtain the conditionally fused features; A gated layer based on fault location is used to receive and process the fault category and fault location mask. The gated mechanism controls the features after condition fusion to ensure that the fault location is not shifted during the generation process. The decoding layer is used to perform deconvolution operations on the gated features and reconstruct them into residual waveforms of the same length as the simulated waveforms through upsampling operations. An enhanced signal output layer is used to superimpose the residual waveform with the simulated waveform to generate an enhanced waveform.

3. The method for generating fault recording data in a low-current grounding system according to claim 2, characterized in that, The enhanced waveform is represented as follows: ； In the formula, Indicates the current time Enhanced waveform, Indicates the current time The simulated waveform, Indicates the current time The residual waveform.

4. The method for generating fault recording data in a low-current grounding system according to claim 1, characterized in that, The discriminator is a multi-task discriminator based on a one-dimensional convolutional neural network, including: The shared feature extraction layer includes a cross-channel convolutional layer, a one-dimensional convolutional layer, a batch normalization layer, and a LeakyReLU function layer, which is used to extract the temporal features of the input waveform and the phase coupling features between the three phase signals using one-dimensional convolution. The input waveform includes the enhanced waveform and the field waveform recording data. The waveform authenticity detection head includes an average pooling layer and a fully connected layer, which is used to perform waveform authenticity detection tasks and distinguish whether the input waveform comes from field waveform recording data or from the generator. The fault location prediction head, including a deconvolution layer, a batch normalization layer, and a sigmoid function layer, is used to perform the fault location prediction task. The output is a point-by-point probability distribution consistent with the number of sampling points, predicting the fault location of the input waveform. The fault category identification head, consisting of an average pooling layer and a fully connected layer, is used to perform fault classification tasks and verify the fault category of the input waveform.

5. The method for generating fault recording data in a low-current grounding system according to claim 4, characterized in that, The cross-channel convolutional layer uses cross-channel convolution operations to maintain the coherence and physical coupling of the enhanced waveform.

6. The method for generating fault recording data in a low-current grounding system according to claim 1, characterized in that, The total loss function is obtained by jointly optimizing the adversarial loss, cycle consistency loss, short-time Fourier transform loss, classification loss, and location loss, and the generative adversarial network is iteratively trained.

7. The method for generating fault recording data in a low-current grounding system according to claim 6, characterized in that, The total loss function is expressed as: ； In the formula, Represents the total loss function. Indicates resistance to loss The weighting coefficients, Indicates the loss of cycle consistency The weighting coefficients, Represents the short-time Fourier transform loss The weighting coefficients, Represents classification loss The weighting coefficients, Indicates position loss Weighting coefficients; The adversarial loss is expressed as: ； In the formula, This represents the probability distribution of waveform data from the field. Samples of field waveform data from the sampling site Take the expected value. Represents the logarithmic function. This indicates that the discriminator determines the on-site waveform data samples. It is the probability of the real sample. This indicates the data recorded from the fault waveform. probability distribution Sample fault recording data Take the expected value. This indicates that the discriminator determines the fault recording data sample. It is the probability of being a real sample; The cycle consistency loss is expressed as: ； In the formula, This represents the loss of cycle consistency. [.] denotes the probability distribution from the simulated waveform data. Sampled simulated waveform data Take the expected value. Indicates from the real domain To the simulation domain generator, Indicates from the simulation domain To the real domain generator, Denotes the L1 norm; The short-time Fourier transform loss is expressed as: ； In the formula, Indicates will A different The configured short-time Fourier transform loss is accumulated, where, express Configuration serial number, Indicates the first indivual The configured short-time Fourier transform loss; The classification loss is expressed as: ； In the formula, Indicates the first The true category of the field waveform data samples obtained from the secondary classification process. Indicates the first The probability of the predicted category of the field waveform data sample obtained from the classification execution, where C represents the number of classification executions; The position loss is expressed as: ； In the formula, Represents the total number of moments. Indicates the current time The actual location of the fault point Indicates the current time The predicted location of the fault point.

8. The method for generating fault recording data in a low-current grounding system according to claim 7, characterized in that, The discriminator's total loss function is obtained by jointly optimizing the discriminator's adversarial loss, classification loss, position loss, and symmetric component constraint loss, and the discriminator is then iteratively trained.

9. The method for generating fault recording data in a low-current grounding system according to claim 8, characterized in that, The total loss function of the discriminator is expressed as: ； In the formula, This represents the total loss function of the discriminator. Indicates discriminator against loss The weighting coefficients, Indicates the classification loss of the discriminator. The weighting coefficients, Indicates discriminator position loss The weighting coefficients, Represents the symmetric component constraint loss of the discriminator. Weighting coefficients; The discriminant adversarial loss is expressed as: ； In the formula, This indicates that the discriminator is susceptible to loss. The classification loss of the discriminator is expressed as: ； In the formula, This represents the classification loss of the discriminator; The discriminant position loss is expressed as: ； In the formula, Indicates the location loss at the fault point; The discriminant's symmetric component constraint loss is expressed as: ； In the formula, This represents the symmetric component constraint loss of the discriminator. , , This represents the zero-sequence component, positive-sequence component, and negative-sequence component corresponding to the fault recording data. , , The table represents the zero-sequence component, positive-sequence component, and negative-sequence component corresponding to the field waveform data.

10. A fault recording data generation system for a low-current grounding system, characterized in that, This method is used to implement the fault recording data generation method for low-current grounding systems according to any one of claims 1-9.

Images