A multi-stage overcurrent and ground fault protection coordination system

Abstract

A multi-level overcurrent and grounding fault protection collaborative system includes a sensing and sampling module, a signal conditioning and feature extraction module, a fault diagnosis and protection execution module, a data statistics and communication module, and a security and access control module. The sensing and sampling module is used to collect power system data; the signal conditioning and feature extraction module is used to preprocess the raw data and extract time-frequency domain features; the fault diagnosis and protection execution module is used to perform fault detection and multi-level collaborative decision execution; the data statistics and communication module is used to build a fault event database and communicate with adjacent protection devices in real time; and the security and access control module is used to provide user permission management and system interaction control functions. This invention proposes a fault signal feature extraction algorithm based on machine learning to extract fault signal features and a deep learning-based overcurrent and grounding fault detection algorithm to detect and identify faults.

Description

Technical Field

[0001] This invention relates to the fields of machine learning, deep learning, and fault detection technology, specifically to a multi-level overcurrent and grounding fault protection collaborative system. Background Technology

[0002] Machine learning technology is a method that uses pattern analysis and feature modeling on a large amount of historical operation and fault data. It aims to solve the problem that traditional power protection relies on manually setting thresholds and is difficult to fully extract the features of complex fault signals. In a multi-level overcurrent and ground fault protection collaborative system, machine learning technology can automatically extract features from overcurrent and ground fault signals, identify time-domain, frequency-domain and time-frequency-domain features from the original power waveform, and improve the robustness and distinguishability of fault signals in noisy environments.

[0003] Deep learning and fault detection technology is a method that uses multi-layer neural networks and intelligent classification models to fit and identify complex nonlinear relationships. It aims to solve the problem that traditional overcurrent and grounding protection relies on fixed logic criteria and is difficult to cope with diverse operating conditions and multiple types of faults. In a multi-level overcurrent and grounding fault protection collaborative system, deep learning and fault detection technology can intelligently detect and identify the power signals after feature extraction, realize the rapid and accurate judgment of different levels of overcurrent, phase-to-phase short circuit and single-phase grounding faults, and improve the sensitivity and selectivity of the system.

[0004] Existing multi-level overcurrent and ground fault protection collaborative systems lack intelligent data-driven capabilities and cannot flexibly adjust to dynamic changes under different operating conditions in the power system. This leads to insufficient detection accuracy and false positives / missed negatives in scenarios with high penetration of distributed power sources, nonlinear loads, and complex short circuits. Furthermore, in terms of protection action coordination, existing systems rely on fixed timing delay strategies and differential configurations, making it difficult to adaptively adjust to real-time changes in the power system topology and operating status. Consequently, there may be weak selectivity, slow action, and malfunctions among multi-level protection actions, affecting the stability and reliability of the power system. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-level overcurrent and ground fault protection collaborative system to address the problems of existing multi-level overcurrent and ground fault protection collaborative systems mentioned in the background art, which lack intelligent data-driven capabilities and cannot flexibly adjust according to dynamic changes under different operating conditions in the power system. This results in insufficient detection accuracy and false positives / false negatives in scenarios with high penetration of distributed power sources, nonlinear loads, and complex short circuits. Furthermore, in terms of protection action coordination, existing systems rely on fixed timing delay strategies and differential configurations, making it difficult to adaptively adjust according to real-time changes in the power system topology and operating status. Consequently, there may be weak selectivity, slow action, and malfunctions among multi-level protection actions, affecting the stability and reliability of the power system.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a multi-level overcurrent and grounding fault protection collaborative system, comprising a sensing and sampling module, a signal conditioning and feature extraction module, a fault diagnosis and protection execution module, a data statistics and communication module, and a security and access control module, characterized in that: the sensing and sampling module is used to collect current, voltage, zero-sequence quantity, and switch status data in the power system, and converts analog data into a high-precision digital sampling stream to ensure measurement accuracy and timing synchronization to support subsequent discrimination; the signal conditioning and feature extraction module proposes a fault signal feature extraction algorithm based on machine learning, used to perform anti-aliasing filtering, noise reduction, and transformer characteristic correction on the sampled signal, and extract time-domain and frequency-domain fault features to ensure transient information retention and feature discriminability and robustness; the fault diagnosis and protection execution module includes a fault detection and discrimination unit and a hierarchical coordination and decision-making unit, wherein the fault detection and discrimination unit proposes a deep learning-based overcurrent and grounding fault... The detection algorithm identifies and classifies overcurrent and grounding fault types and severity, ensuring high sensitivity and low false alarm rate in fault determination. The hierarchical coordination and decision-making unit is used to execute action priority determination and coordination logic based on protection level, timing curve, and adjacent end status information, ensuring selective cooperation and reliable yielding mechanism between local and remote protection. The data statistics and communication module includes a database construction unit and a communication and interconnection unit. The database construction unit is used to build a fault event database, structurally store event and operation statistics data, and provide retrieval and reporting interfaces to ensure data integrity and queryability to support fault analysis and operation and maintenance decisions. The communication and interconnection unit is used for real-time and non-real-time message exchange and manages link redundancy and time synchronization to ensure fast and reliable protection coordination and synchronization consistency. The security and access control module is used to manage user authentication, permission hierarchy, and system interaction control to ensure the integrity, traceability, and security of system configuration and operation.

[0007] Preferably, the sensing and sampling module is configured with current transformers, voltage transformers and zero-sequence transformers, an anti-aliasing analog front-end, a high-precision analog-to-digital converter and a time synchronization device to uniformly sample and time-label the current, voltage, zero-sequence quantity and switching status of the power system, so as to ensure that high-precision, low-latency and time-consistent digital raw data are obtained for subsequent processing and positioning analysis.

[0008] Preferably, the signal conditioning and feature extraction module proposes a fault signal feature extraction algorithm based on machine learning. By performing digital filtering, transient and harmonic separation and normalization on the original samples, and combining machine learning for time-frequency domain feature extraction, it ensures that feature vectors with discriminative power against overcurrent and grounding faults are extracted from noise and interference for use by subsequent modules.

[0009] Preferably, the fault signal feature extraction algorithm based on machine learning is as follows: First, by implementing finite impulse response anti-aliasing filtering on the sampled discrete signal and performing precise timestamp alignment, it is ensured that there are no aliasing components in the frequency band during subsequent feature extraction and that the time base of each endpoint is consistent with the time synchronization requirements of the system. The specific formula is expressed as follows: ; in, Represented as a filtered discrete-time series, representing the first... The filtered output of each sampling point Represented as the first discrete sampled signal of the input original signal One data sample, The FIR filter is represented by the first... Order coefficient, This is expressed as the order of the FIR filter. Represented as the index of the number of discrete-time samples. This is represented by the FIR filter order index. By ensuring that the input waveforms received by each protection unit meet the requirements for phase and amplitude accuracy in both time synchronization and frequency band consistency for directional judgment and grading criteria, reliable input is provided for subsequent feature extraction based on machine learning. Secondly, by establishing a system response model for the secondary side of the transformer and performing regularized deconvolution processing in the frequency domain to correct the measurement response, the masking of fault features by the transformer's non-ideal frequency response and secondary saturation is eliminated. This ensures that the machine learning model reads the true primary signal features during the fault signal feature extraction stage. The specific formula is expressed as: ; in, Represented as the corrected frequency domain signal, Expressed as angular frequency, This is represented as the discrete-time Fourier transform of the filtered signal Y[p]. Expressed as the equivalent frequency response of the current transformer and sampling link, it represents the frequency domain response of the measurement channel to the actual primary quantity. Described as the complex conjugate of T(ω), Expressed as the square of the magnitude of T(ω), Represented as a regularization factor, precise transformer calibration can avoid directional misjudgments and stage conflicts caused by secondary saturation and uneven frequency response, thereby improving the algorithm's sensitivity and accuracy to the essential characteristics of overcurrent and grounding faults. This is represented as a time-domain corrected sample sequence. This is represented as an inverse Fourier transform, and then, through the corrected time-domain signal... A multi-level decomposition is performed using adjustable-factor wavelet transform to simultaneously preserve fault transient and harmonic components in the time and frequency domains. By calculating the kurtosis of each level and selecting several levels with the largest kurtosis as the most information-rich levels, the focus can be placed on sub-bands containing fault impact and arc features. This reduces redundancy and preserves fault discrimination information in machine learning feature extraction. The specific formula is as follows: ; in, Represented as the first The discrete coefficient sequence of wavelet subbands in the time domain represents the temporal energy distribution of that frequency band. Represented as the first The wavelet basis impulse response coefficients of the first layer represent the response coefficients used for the first layer. Layer-extracted filters, Represented as a decomposition layer index, Represented as the first The peak attitude quantity of the layer coefficient represents the first layer. The sharpness of the layer distribution, The larger the value, the more violent the transient. Represented as the first Layer coefficient sequence The number of samples, Represented as the first The sample mean of the layer coefficients, based on the selected layer. The output has targeted frequency band characteristics for use in downstream protection components and deep learning-based fault detectors, achieving synergy between time-frequency domain information and hierarchical protection logic. Secondly, by selecting wavelet layer coefficients... Calculate a set of time-domain and frequency-domain statistics to construct an initial set of feature vectors. The subband energy, normalized spectral entropy, and time-domain statistics are calculated to transform the original waveform into a concise and discriminative numerical representation that can be used by machine learning models. The specific formula is as follows: ; in, Represented as the first Total energy of the sub-bands Represented as the first Normalized energy percentage of the layer This represents the total number of layers used for feature calculation. Represented as the layer index used for feature calculation. Represented as the first Total energy of the sub-bands Expressed as spectral entropy, it reflects the uniformity of energy distribution between subbands. Represented as a logarithmic function, Expressed as the normalized spectral entropy, Represented as the root mean square value of the time-domain statistic. The kurtosis factor, expressed as a time-domain statistic, measures the ratio of the transient peak value to the root mean square. This represents the operation of taking the real part of a complex number. This is represented as taking the maximum value of the sequence, and then, by using the original statistical feature vector... The input is fed into a machine learning mapper, which generates a low-dimensional, highly discriminative final feature vector while preserving fault discrimination information. This vector can then be used as a protection criterion input in the system for combined use with traditional criteria by the deep learning fault detector. The specific formula is as follows: ; in, This is represented as the low-dimensional feature vector of the output. Represented as the ReLU nonlinear activation function, it transforms the result of a linear mapping into a nonlinear representation. Represented as a mapping matrix, Represented as a bias column vector, finally, by converting the final feature vector... This, along with a set of metadata including timestamp, measurement point ID, and preprocessing version number, is combined into a standardized feature package and sent to the fault diagnosis and protection execution module according to a fixed communication format. The specific formula is as follows: ; in, Represented as the sent feature packet vector, Represented as timestamp data, Represented as measurement point identification data, Represented as version number data, Represented as a vector transpose operation, the constructed feature package is sent to the fault diagnosis and protection execution module through a low-latency channel to ensure that the extracted fault features can be used by the system in real time to complete fault detection and protection decisions.

[0010] Preferably, the fault diagnosis and protection execution module includes a fault detection and discrimination unit. The fault detection and discrimination unit proposes an overcurrent and grounding fault detection algorithm based on deep learning. By performing real-time analysis of the time-frequency domain features provided by the signal conditioning and feature extraction module, a deep learning model is constructed within the system for reasoning and identification. This enables the identification of instantaneous faults and persistent faults, and the differentiation between phase-to-phase short-circuit faults and single-phase grounding faults. This ensures that high-sensitivity detection and reliable discrimination of multiple types of power faults can still be achieved under complex operating conditions and strong interference environments.

[0011] Preferably, the deep learning-based overcurrent and ground fault detection algorithm is as follows: First, by constructing a normalized time series matrix from the multi-channel sub-band features and time-domain statistics from the signal conditioning and feature extraction modules in chronological order, the deep recurrent network can learn the temporal pattern of overcurrent and ground fault detection using sequence dependencies, ensuring that the network input simultaneously contains transient information and steady-state trends, thereby supporting the system's distinction between transient and persistent faults. The specific formula is expressed as follows: ; in, Represented as time Feature column vectors Represented as a time index, At any moment The Each feature data, Let the feature dimension be expressed as the feature dimension at each time step. Represented as time Independent data with different features Represented as a time series matrix input into a deep learning model. This is expressed as the length of the time window, used to capture short-time series information. Represented as time Independent column vectors of different features will be used in the system As a standard input format for deep learning models, hierarchical protection can obtain consistent time-series views both locally and remotely for collaborative judgment. Then, by processing the input matrix in an online runtime environment... The data is progressively fed into deep recurrent units to learn time-dependent features, and a gating mechanism is used to suppress the impact of long-term and short-term noise on the discrimination. This ensures that the temporal extraction capability of the deep learning model can identify transient changes and persistent overcurrents, thereby meeting the system's requirements for high sensitivity and low false alarm rate. The specific formula is as follows: ; in, This is represented as an update gate vector, used to control the proportion of new information written. This is represented as an element-wise Sigmoid activation function. This is represented as the input weight matrix of the update gate. This is represented as the hidden weight matrix of the update gate. Represented as The hidden state at any given moment This is represented as the update gate bias vector. This is represented as a reset gate vector, used to control the strength of the influence of historical information on candidate calculation. This is represented as element-wise nonlinear activation. This is represented as the input weight matrix of the reset gate. This is represented as the hidden weight matrix of the reset gate. This is represented as the reset gate bias vector. Represented as a candidate hidden state, This is represented as the candidate input weight matrix. Represented as the candidate hidden weight matrix, This is represented as a candidate bias vector. This is represented as element-wise multiplication. Represented as The hidden state update at each moment, by configuring this gating recursion as a low-latency edge instance, ensures that subsequent units can be updated based on this. The temporal evolution of the model determines the occurrence of faults and triggers hierarchical actions. Secondly, by retaining short-window energy and long-window trend quantities in parallel within the model and calculating the energy ratio, transient and persistent faults are distinguished. This is combined with deep hidden states. The magnitude of the sudden change ensures that while maintaining high sensitivity, false alarms caused by transient noise are suppressed, enabling the system to distinguish between transient faults and true persistent faults. This allows for the correct triggering of transient tripping and delayed backup actions in graded protection. The specific formula is as follows: ; in, Represented as short-window energy, This is expressed as the length of a long-term period. Represented as long window energy, Indicated as short-time duration, Represented as a time index, It can be represented as the square of the L2 norm of a vector. Represented as the instantaneous energy of all characteristics at that moment. Represented as a transient ratio indicator, when If the threshold is exceeded, it is judged as a momentary strong impact, which will trigger a momentary action candidate. If a high value is maintained for more than t+1000 durations, it is considered a persistent fault. Represented as regularization parameters, With depth hidden state The transient response is combined with subsequent decision-making units. In the event of a sudden, strong impact, instantaneous actions are prioritized. For persistent faults, timed and inverse-time backup actions are initiated to ensure selectivity and minimal sacrifice. Then, the energy feature sequence is input into a constructed deep convolutional recurrent hybrid network to enhance the detection accuracy of overcurrent and grounding faults under complex operating conditions. Energy mutation features are extracted within a short window and used as local input to the convolutional neural network to capture fast transient patterns. Simultaneously, the continuity of the time series is maintained within a long window, and a long short-term memory network is used to model the global evolution trend, ensuring that the algorithm can identify persistent overcurrent and slowly developing high-impedance grounding faults. The specific formula is expressed as follows: ; in, This is represented by the output vector of the deep learning model, indicating the probability prediction results of the deep convolutional recurrent hybrid network for transient and persistent faults. Represented as a convolutional recurrent hybrid network, This is used to determine the length of the data frame. This is represented as the final joint judgment indicator. This is represented as a joint weighting factor, used to balance the influence of physical indicators and deep learning output. Through this dual-path fusion of convolution and recursion, the constructed deep learning model can form a composite judgment mechanism that is sensitive to sudden changes and robust to slow trends. This enables the system to maintain stable fault identification capabilities even under extreme power disturbances. Finally, the output of the deep learning model is combined with time-frequency domain features and input into a multi-layer neural network classifier to achieve the final identification of fault types. The multi-layer neural network classifier contains multiple fully connected layers and an attention mechanism. The fully connected layers are responsible for extracting global statistical patterns, while the attention mechanism automatically focuses on key frequency bands and transient features, thereby improving the identification accuracy of complex fault modes in high-noise environments. The specific formula is expressed as follows: ; in, Represented as a concatenated input vector, this concatenates the deep learning model output with time-frequency features in a dimensional manner, serving as the raw input to the classifier. Represented as an attention weight vector, This can be represented as a function that exponentially normalizes the vector elements. Represented as an attention weight mapping matrix, This is represented as the attention layer bias vector. Represented as attention-weighted feature vectors, This is represented as the final classification probability vector, used to determine overcurrent faults, grounding faults, and mixed faults. Represented as a probability distribution transformation function, Represented as the output layer weight matrix, This is represented as the weight matrix of the first fully connected layer. This is represented as the first-level bias vector. Represented as an output layer bias vector, it is then weighted by attention and entered into two fully connected layers to activate the output structure. The system can map the probability results into explicit protection action priorities and strategies.

[0012] Preferably, the fault diagnosis and protection execution module includes a hierarchical coordination and decision-making unit. The hierarchical coordination and decision-making unit implements timer management, priority and backoff strategies, and interaction mechanisms with adjacent devices, and dynamically selects the execution scheme between instantaneous actions and inverse time-limited actions to ensure that the protection actions have hierarchical selectivity, backup reliability, and coordination consistent with the system operation manual.

[0013] Preferably, the data statistics and communication module includes a database construction unit. The database construction unit designs an architecture that combines a historical fault event database with real-time file storage, an event indexing mechanism, and supports efficient retrieval and export interfaces to ensure that fault records and operational statistics are available and complete in queries and long-term analysis.

[0014] Preferably, the data statistics and communication module includes a communication and interconnection unit. The communication and interconnection unit realizes low-latency, deterministic, and highly reliable real-time information interaction and control coordination with adjacent protection devices and multi-level overcurrent and ground fault protection collaborative systems by implementing Ethernet and fiber optic physical interfaces, link redundancy strategies, and combining precise time synchronization mechanisms.

[0015] Preferably, the security and access control module establishes hierarchical permission management, operation identity authentication, interaction process verification, and key operation error prevention mechanisms to control the entire process of user commands, system parameter adjustments, and protection action issuance, ensuring that system interaction is secure and controllable and avoiding protection failures and malfunctions caused by misoperation.

[0016] Compared with the prior art, the beneficial effects of the present invention are: 1. The signal conditioning and feature extraction module proposes a fault signal feature extraction algorithm based on machine learning. First, the algorithm implements anti-aliasing filtering on the sampled discrete signal and strictly aligns the timestamps to ensure that the waveforms input to each protection unit are highly consistent in frequency band and time reference. This provides a reliable phase and amplitude basis for directional judgment and grade difference coordination, reducing misjudgments and missed actions caused by sampling and synchronization errors. Second, by establishing a secondary side response model of the instrument transformer and implementing frequency domain regularized deconvolution correction, the algorithm can effectively counteract the masking of fault features by the non-ideal frequency response and secondary saturation of the instrument transformer. This allows the machine learning stage to read more accurate signal features closer to the primary quantity, thereby reducing directional errors and grade difference conflicts, and improving the sensitivity and robustness to grounding faults and nonlinear faults. Simultaneously, it employs adjustable factor wavelet multi-layer decomposition and selects the most information-rich decomposition layer based on kurtosis index, enabling the algorithm to simultaneously preserve fault transients, arcing, and higher harmonics in the time and frequency domains. The algorithm firstly identifies faults and then uses hierarchical selection to significantly reduce unnecessary computational overhead, meeting the requirements for real-time processing at both the station and device levels. Secondly, it constructs a compact and highly interpretable initial feature set by calculating multidimensional statistics including subband energy, normalized spectral entropy, root mean square, and peak factor. It then generates low-dimensional discriminative feature vectors using a machine learning mapper, preserving the interpretability required by traditional protection systems while enabling the system to adaptively identify complex fault modes. Furthermore, the algorithm packages the final features with timestamps, measurement point IDs, and version information and distributes them to the hierarchical coordination and decision-making unit via a low-latency channel. This supports collaborative decision-making between local criteria and deep learning discrimination by protection devices, resulting in faster action response, stronger selectivity, and lower false tripping rate. Overall, this algorithm improves the sensitivity and accuracy of fault detection, enhances coordination and traceability among multi-level protection systems, and lays a reliable foundation for data-driven fault location, condition monitoring, and predictive maintenance.

[0017] 2. The fault detection and discrimination unit proposes a deep learning-based overcurrent and grounding fault detection algorithm. First, the algorithm constructs a multi-dimensional input matrix with a clear temporal structure by standardizing and organizing multi-channel sub-band features and time-domain statistics. This allows the deep recurrent network to fully capture the dynamic patterns and contextual dependencies in current and voltage signals. This temporal modeling mechanism not only preserves transient change information but also integrates steady-state trend changes, thus providing a reliable data foundation for distinguishing between transient and persistent faults. This enables the system to perform collaborative analysis and judgment based on a consistent, high-quality temporal view both locally and remotely. Second, the algorithm processes the input sequence frame by frame through a gated recurrent mechanism, dynamically adjusting the information flow using update and reset gates to effectively suppress the influence of noise interference and non-fault transient fluctuations on model judgment. This mechanism ensures the model's high sensitivity to real fault signals while reducing the false alarm rate. Internally, the algorithm uses parallel energy calculations for short and long time windows and introduces an energy ratio index to assist in distinguishing between transient impacts and persistent overcurrents, as well as changes in physical characteristics and neural network hidden states. The algorithm deeply integrates trends to form a composite criterion, ensuring that the system can accurately distinguish between transient disturbances and real faults, thereby triggering corresponding instantaneous actions and delayed backup protection, effectively balancing protection speed and selectivity. Furthermore, the algorithm combines the advantages of convolutional structures and recurrent networks to construct a hybrid architecture that can capture both local rapid transient patterns and model global evolution trends. This design enables the algorithm to have good detection capabilities for slowly developing faults in high-impedance grounding, expanding the system's applicable operating conditions. The model output is also concatenated with frequency domain features and used for decision-making via a multi-layer classifier with an attention mechanism. The attention mechanism can adaptively focus on the most discriminative feature regions, improving the identification accuracy of complex and mixed faults, and exhibiting excellent robustness under different noise environments. In summary, the deep learning-based overcurrent and grounding fault detection algorithm not only achieves high-precision and high-reliability detection of overcurrent and grounding faults, but also supports effective collaboration among multi-level protection systems, improving the rationality of fault response and the overall protection level of the multi-level overcurrent and grounding fault protection collaborative system. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the structure of the present invention. Detailed Implementation

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

[0020] Please see Figure 1 This invention provides a multi-level overcurrent and grounding fault protection collaborative system, including a sensing and sampling module, a signal conditioning and feature extraction module, a fault diagnosis and protection execution module, a data statistics and communication module, and a security and access control module. The sensing and sampling module collects current, voltage, zero-sequence quantity, and switch status data from the power system and converts analog data into a high-precision digital sampling stream to ensure measurement accuracy and timing synchronization for subsequent discrimination. The signal conditioning and feature extraction module proposes a machine learning-based fault signal feature extraction algorithm to perform anti-aliasing filtering, noise reduction, and transformer characteristic correction on the sampled signals, and extract time-domain and frequency-domain fault features, ensuring transient information retention and feature discriminability and robustness. The fault diagnosis and protection execution module includes a fault detection and discrimination unit and a hierarchical coordination and decision-making unit. The fault detection and discrimination unit proposes a deep learning-based overcurrent and grounding fault detection algorithm. The system categorizes and classifies overcurrent and grounding faults by type and severity, ensuring high sensitivity and low false alarm rate in fault diagnosis. A hierarchical coordination and decision-making unit executes action priority determination and coordination logic based on protection level, timing curves, and adjacent end status information, ensuring selective cooperation and reliable failover mechanisms between local and remote protection systems. The data statistics and communication module includes a database construction unit and a communication and interconnection unit. The database construction unit builds a fault event database, structurally storing event and operational statistics and providing retrieval and reporting interfaces, ensuring data integrity and queryability to support fault analysis and maintenance decisions. The communication and interconnection unit handles real-time and non-real-time message exchange and manages link redundancy and time synchronization, ensuring fast and reliable protection coordination and synchronization consistency. The security and access control module manages user authentication, permission levels, and system interaction control, ensuring the integrity, traceability, and security of system configuration and operation.

[0021] See Figure 1 Furthermore, the sensing and sampling module, by configuring current transformers, voltage transformers, zero-sequence transformers, anti-aliasing analog front-ends, high-precision analog-to-digital converters, and time synchronization devices, performs unified sampling and time labeling of the current, voltage, zero-sequence quantities, and switching states of the power system, ensuring that high-precision, low-latency, and time-consistent digital raw data is obtained for subsequent processing and positioning analysis.

[0022] See Figure 1 Furthermore, the signal conditioning and feature extraction module proposes a fault signal feature extraction algorithm based on machine learning. By performing digital filtering, transient and harmonic separation and normalization on the original samples, and combining machine learning for time-frequency domain feature extraction, it ensures that feature vectors with discriminative power against overcurrent and grounding faults are extracted from noise and interference for use by subsequent modules.

[0023] See Figure 1 Furthermore, the fault signal feature extraction algorithm based on machine learning is as follows: First, by implementing finite impulse response anti-aliasing filtering on the sampled discrete signal and performing precise timestamp alignment, it is ensured that there are no aliasing components in the frequency band during subsequent feature extraction and that the time base of each endpoint is consistent with the time synchronization requirements of the system. The specific formula is expressed as follows: ; in, Represented as a filtered discrete-time series, representing the first... The filtered output of each sampling point Represented as the first discrete sampled signal of the input original signal One data sample, The FIR filter is represented by the first... Order coefficient, This is expressed as the order of the FIR filter. Represented as the index of the number of discrete-time samples. This is represented by the FIR filter order index. By ensuring that the input waveforms received by each protection unit meet the requirements for phase and amplitude accuracy in both time synchronization and frequency band consistency for directional judgment and grading criteria, reliable input is provided for subsequent feature extraction based on machine learning. Secondly, by establishing a system response model for the secondary side of the transformer and performing regularized deconvolution processing in the frequency domain to correct the measurement response, the masking of fault features by the transformer's non-ideal frequency response and secondary saturation is eliminated. This ensures that the machine learning model reads the true primary signal features during the fault signal feature extraction stage. The specific formula is expressed as: ; in, Represented as the corrected frequency domain signal, Expressed as angular frequency, This is represented as the discrete-time Fourier transform of the filtered signal Y[p]. Expressed as the equivalent frequency response of the current transformer and sampling link, it represents the frequency domain response of the measurement channel to the actual primary quantity. Described as the complex conjugate of T(ω), Expressed as the square of the magnitude of T(ω), Represented as a regularization factor, precise transformer calibration can avoid directional misjudgments and stage conflicts caused by secondary saturation and uneven frequency response, thereby improving the algorithm's sensitivity and accuracy to the essential characteristics of overcurrent and grounding faults. This is represented as a time-domain corrected sample sequence. This is represented as an inverse Fourier transform, and then, through the corrected time-domain signal... A multi-level decomposition is performed using adjustable-factor wavelet transform to simultaneously preserve fault transient and harmonic components in the time and frequency domains. By calculating the kurtosis of each level and selecting several levels with the largest kurtosis as the most information-rich levels, the focus can be placed on sub-bands containing fault impact and arc features. This reduces redundancy and preserves fault discrimination information in machine learning feature extraction. The specific formula is as follows: ; in, Represented as the first The discrete coefficient sequence of wavelet subbands in the time domain represents the temporal energy distribution of that frequency band. Represented as the first The wavelet basis impulse response coefficients of the first layer represent the response coefficients used for the first layer. Layer-extracted filters, Represented as a decomposition layer index, Represented as the first The peak attitude quantity of the layer coefficient represents the first layer. The sharpness of the layer distribution, The larger the value, the more violent the transient. Represented as the first Layer coefficient sequence The number of samples, Represented as the first The sample mean of the layer coefficients, based on the selected layer. The output has targeted frequency band characteristics for use in downstream protection components and deep learning-based fault detectors, achieving synergy between time-frequency domain information and hierarchical protection logic. Secondly, by selecting wavelet layer coefficients... Calculate a set of time-domain and frequency-domain statistics to construct an initial set of feature vectors. The subband energy, normalized spectral entropy, and time-domain statistics are calculated to transform the original waveform into a concise and discriminative numerical representation that can be used by machine learning models. The specific formula is as follows: ; in, Represented as the first Total energy of the sub-bands Represented as the first Normalized energy percentage of the layer This represents the total number of layers used for feature calculation. Represented as the layer index used for feature calculation. Represented as the first Total energy of the sub-bands Expressed as spectral entropy, it reflects the uniformity of energy distribution between subbands. Represented as a logarithmic function, Expressed as the normalized spectral entropy, Represented as the root mean square value of the time-domain statistic. The kurtosis factor, expressed as a time-domain statistic, measures the ratio of the transient peak value to the root mean square. This represents the operation of taking the real part of a complex number. This is represented as taking the maximum value of the sequence, and then, by using the original statistical feature vector... The input is fed into a machine learning mapper, which generates a low-dimensional, highly discriminative final feature vector while preserving fault discrimination information. This vector can then be used as a protection criterion input in the system for combined use with traditional criteria by the deep learning fault detector. The specific formula is as follows: ; in, This is represented as the low-dimensional feature vector of the output. Represented as the ReLU nonlinear activation function, it transforms the result of a linear mapping into a nonlinear representation. Represented as a mapping matrix, Represented as a bias column vector, finally, by converting the final feature vector... This, along with a set of metadata including timestamp, measurement point ID, and preprocessing version number, is combined into a standardized feature package and sent to the fault diagnosis and protection execution module according to a fixed communication format. The specific formula is as follows: ; in, Represented as the sent feature packet vector, Represented as timestamp data, Represented as measurement point identification data, Represented as version number data, Represented as a vector transpose operation, the constructed feature package is sent to the fault diagnosis and protection execution module through a low-latency channel to ensure that the extracted fault features can be used by the system in real time to complete fault detection and protection decisions.

[0024] See Figure 1 Furthermore, the fault diagnosis and protection execution module includes a fault detection and discrimination unit. This unit proposes an overcurrent and grounding fault detection algorithm based on deep learning. By performing real-time analysis of the time-frequency domain features provided by the signal conditioning and feature extraction module, a deep learning model is constructed within the system for reasoning and identification. This allows the system to identify instantaneous and persistent faults, and distinguish between phase-to-phase short-circuit faults and single-phase grounding faults. This ensures high-sensitivity detection and reliable discrimination of various types of power faults even under complex operating conditions and strong interference environments.

[0025] See Figure 1Furthermore, the deep learning-based overcurrent and ground fault detection algorithm is as follows: First, by constructing a normalized time series matrix from the multi-channel sub-band features and time-domain statistics from the signal conditioning and feature extraction modules in chronological order, the deep recurrent network can learn the temporal pattern of overcurrent and ground fault detection using sequence dependencies, ensuring that the network input simultaneously contains transient information and steady-state trends, thereby supporting the system's distinction between transient and persistent faults. The specific formula is expressed as follows: ; in, Represented as time Feature column vectors Represented as a time index, At any moment The Each feature data, Let the feature dimension be expressed as the feature dimension at each time step. Represented as time Independent data with different features Represented as a time series matrix input into a deep learning model. This is expressed as the length of the time window, used to capture short-time series information. Represented as time Independent column vectors of different features will be used in the system As a standard input format for deep learning models, hierarchical protection can obtain consistent time-series views both locally and remotely for collaborative judgment. Then, by processing the input matrix in an online runtime environment... The data is progressively fed into deep recurrent units to learn time-dependent features, and a gating mechanism is used to suppress the impact of long-term and short-term noise on the discrimination. This ensures that the temporal extraction capability of the deep learning model can identify transient changes and persistent overcurrents, thereby meeting the system's requirements for high sensitivity and low false alarm rate. The specific formula is as follows: ; in, This is represented as an update gate vector, used to control the proportion of new information written. This is represented as an element-wise Sigmoid activation function. This is represented as the input weight matrix of the update gate. This is represented as the hidden weight matrix of the update gate. Represented as The hidden state at any given moment This is represented as the update gate bias vector. This is represented as a reset gate vector, used to control the strength of the influence of historical information on candidate calculation. This is represented as element-wise nonlinear activation. This is represented as the input weight matrix of the reset gate. This is represented as the hidden weight matrix of the reset gate. This is represented as the reset gate bias vector. Represented as a candidate hidden state, This is represented as the candidate input weight matrix. Represented as the candidate hidden weight matrix, This is represented as a candidate bias vector. This is represented as element-wise multiplication. Represented as The hidden state update at each moment, by configuring this gating recursion as a low-latency edge instance, ensures that subsequent units can be updated based on this. The temporal evolution of the model determines the occurrence of faults and triggers hierarchical actions. Secondly, by retaining short-window energy and long-window trend quantities in parallel within the model and calculating the energy ratio, transient and persistent faults are distinguished. This is combined with deep hidden states. The magnitude of the sudden change ensures that while maintaining high sensitivity, false alarms caused by transient noise are suppressed, enabling the system to distinguish between transient faults and true persistent faults. This allows for the correct triggering of transient tripping and delayed backup actions in graded protection. The specific formula is as follows: ; in, Represented as short-window energy, This is expressed as the length of a long-term period. Represented as long window energy, Indicated as short-time duration, Represented as a time index, It can be represented as the square of the L2 norm of a vector. Represented as the instantaneous energy of all characteristics at that moment. Represented as a transient ratio indicator, when If the threshold is exceeded, it is judged as a momentary strong impact, which will trigger a momentary action candidate. If a high value is maintained for more than t+1000 durations, it is considered a persistent fault. Represented as regularization parameters, With depth hidden state The transient response is combined with subsequent decision-making units. In the event of a sudden, strong impact, instantaneous actions are prioritized. For persistent faults, timed and inverse-time backup actions are initiated to ensure selectivity and minimal sacrifice. Then, the energy feature sequence is input into a constructed deep convolutional recurrent hybrid network to enhance the detection accuracy of overcurrent and grounding faults under complex operating conditions. Energy mutation features are extracted within a short window and used as local input to the convolutional neural network to capture fast transient patterns. Simultaneously, the continuity of the time series is maintained within a long window, and a long short-term memory network is used to model the global evolution trend, ensuring that the algorithm can identify persistent overcurrent and slowly developing high-impedance grounding faults. The specific formula is expressed as follows: ; in, This is represented by the output vector of the deep learning model, indicating the probability prediction results of the deep convolutional recurrent hybrid network for transient and persistent faults. Represented as a convolutional recurrent hybrid network, This is used to determine the length of the data frame. This is represented as the final joint judgment indicator. This is represented as a joint weighting factor, used to balance the influence of physical indicators and deep learning output. Through this dual-path fusion of convolution and recursion, the constructed deep learning model can form a composite judgment mechanism that is sensitive to sudden changes and robust to slow trends. This enables the system to maintain stable fault identification capabilities even under extreme power disturbances. Finally, the output of the deep learning model is combined with time-frequency domain features and input into a multi-layer neural network classifier to achieve the final identification of fault types. The multi-layer neural network classifier contains multiple fully connected layers and an attention mechanism. The fully connected layers are responsible for extracting global statistical patterns, while the attention mechanism automatically focuses on key frequency bands and transient features, thereby improving the identification accuracy of complex fault modes in high-noise environments. The specific formula is expressed as follows: ; in, Represented as a concatenated input vector, this concatenates the deep learning model output with time-frequency features in a dimensional manner, serving as the raw input to the classifier. Represented as an attention weight vector, This can be represented as a function that exponentially normalizes the vector elements. Represented as an attention weight mapping matrix, This is represented as the attention layer bias vector. Represented as attention-weighted feature vectors, This is represented as the final classification probability vector, used to determine overcurrent faults, grounding faults, and mixed faults. Represented as a probability distribution transformation function, Represented as the output layer weight matrix, This is represented as the weight matrix of the first fully connected layer. This is represented as the first-level bias vector. Represented as an output layer bias vector, it is then weighted by attention and entered into two fully connected layers to activate the output structure. The system can map the probability results into explicit protection action priorities and strategies.

[0026] See Figure 1 Furthermore, the fault diagnosis and protection execution module includes a hierarchical coordination and decision-making unit. This unit implements timer management, priority and backoff strategies, and an interaction mechanism with adjacent devices. It dynamically selects the execution scheme between instantaneous actions and inverse time-limited actions to ensure that the protection actions have hierarchical selectivity, backup reliability, and coordination consistent with the system operation manual.

[0027] See Figure 1 Furthermore, the data statistics and communication module includes a database construction unit. This database construction unit designs an architecture that combines a historical fault event database with real-time file storage, an event indexing mechanism, and supports efficient retrieval and export interfaces, ensuring that fault records and operational statistics are available and complete in queries and long-term analysis.

[0028] See Figure 1 Furthermore, the data statistics and communication module includes a communication and interconnection unit. The communication and interconnection unit realizes low-latency, deterministic, and highly reliable real-time information interaction and control coordination with adjacent protection devices and multi-level overcurrent and ground fault protection collaborative systems by implementing Ethernet and fiber optic physical interfaces, link redundancy strategies, and combining precise time synchronization mechanisms.

[0029] See Figure 1 Furthermore, the security and access control module establishes hierarchical permission management, operation identity authentication, interaction process verification, and key operation error prevention mechanisms to control the entire process of user commands, system parameter adjustments, and protection action issuance, ensuring that system interaction is secure and controllable and avoiding protection failures and malfunctions caused by misoperation.

[0030] In practical application, firstly, the sensing and sampling module collects current, voltage, zero-sequence quantities, and switch status data from the power system, converting analog data into a high-precision digital sampling stream to ensure measurement accuracy and timing synchronization for subsequent discrimination. Secondly, the signal conditioning and feature extraction module proposes a machine learning-based fault signal feature extraction algorithm to perform anti-aliasing filtering, noise reduction, and transformer characteristic correction on the sampled signals, extracting time-domain and frequency-domain fault features to ensure transient information retention and feature discriminability and robustness. Then, the fault diagnosis and protection execution module includes a fault detection and discrimination unit and a hierarchical coordination and decision-making unit. The fault detection and discrimination unit proposes a deep learning-based overcurrent and grounding fault detection algorithm to identify and classify overcurrent and grounding fault types and severity, ensuring high sensitivity and low false alarm rate fault judgment, and hierarchical coordination. The coordination and decision-making unit is used to execute action priority determination and coordination logic based on protection level, timing curve, and adjacent end status information, ensuring selective cooperation and reliable yielding mechanism between local and remote protection. Secondly, the data statistics and communication module includes a database construction unit and a communication and interconnection unit. The database construction unit is used to build a fault event database, structurally store event and operational statistics data, and provide retrieval and reporting interfaces, ensuring data integrity and queryability to support fault analysis and operation and maintenance decisions. The communication and interconnection unit is used for real-time and non-real-time message exchange and manages link redundancy and time synchronization, ensuring fast and reliable protection coordination and synchronization consistency. Finally, the security and access control module is used to manage user authentication, permission levels, and system interaction control, ensuring the integrity, traceability, and security of system configuration and operation.

[0031] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-level overcurrent and ground fault protection collaborative system, comprising a sensing and sampling module, a signal conditioning and feature extraction module, a fault diagnosis and protection execution module, a data statistics and communication module, and a security and access control module, characterized in that: The sensing and sampling module is used to collect current, voltage, zero-sequence quantity, and switch status data in the power system, and converts analog data into a high-precision digital sampling stream to achieve measurement accuracy and timing synchronization to support subsequent discrimination. The signal conditioning and feature extraction module proposes a fault signal feature extraction algorithm based on machine learning, used to perform anti-aliasing filtering, noise reduction, and transformer characteristic correction on the sampled signal, and extract time-domain and frequency-domain fault features, achieving transient information preservation and feature discriminability and robustness. The fault diagnosis and protection execution module includes a fault detection and discrimination unit and a hierarchical coordination and decision-making unit. The fault detection and discrimination unit proposes a deep learning-based overcurrent and grounding fault detection algorithm to identify and classify overcurrent and grounding fault types and severity, achieving high sensitivity and low false alarm rate fault judgment. The hierarchical coordination and decision-making unit... The policy unit is used to execute action priority determination and coordination logic based on protection level, timing curve, and adjacent end status information, realizing selective cooperation and reliable yielding mechanism between local and remote protection; the data statistics and communication module includes a database construction unit and a communication and interconnection unit. The database construction unit is used to build a fault event database, structurally store event and operation statistics data, and provide retrieval and reporting interfaces to achieve data integrity and queryability to support fault analysis and operation and maintenance decisions. The communication and interconnection unit is used for real-time and non-real-time message exchange and manage link redundancy and time synchronization to achieve fast and reliable protection coordination and synchronization consistency; the security and access control module is used to manage user identity authentication, permission classification, and system interaction control to achieve the integrity, traceability, and security of system configuration and operation.

2. The multi-stage overcurrent and ground fault protection coordinated system according to claim 1, characterized in that: The sensing and sampling module, by configuring current transformers, voltage transformers, zero-sequence transformers, anti-aliasing analog front-ends, high-precision analog-to-digital converters, and time synchronization devices, performs unified sampling and time labeling of the current, voltage, zero-sequence quantities, and switch states of the power system, thereby obtaining high-precision, low-latency, and time-consistent digital raw data for subsequent processing and location analysis.

3. The multi-stage overcurrent and ground fault protection coordinated system according to claim 1, characterized in that: The signal conditioning and feature extraction module proposes a fault signal feature extraction algorithm based on machine learning. By performing digital filtering, transient and harmonic separation and normalization on the original samples, and combining machine learning for time-frequency domain feature extraction, it can extract feature vectors that can distinguish overcurrent and grounding faults from noise and interference for use by subsequent modules.

4. The multi-stage overcurrent and ground fault protection coordinated system according to claim 3, characterized in that, First, by implementing finite impulse response anti-aliasing filtering on the sampled discrete signal and performing precise timestamp alignment, it is ensured that subsequent feature extraction has no aliasing components within the frequency band and that the time base of each endpoint is consistent with the time synchronization requirements of the system. The specific formula is expressed as follows: ； in, Represented as a filtered discrete-time series, representing the first... The filtered output of each sampling point Represented as the first discrete sampled signal of the input original signal One data sample, The FIR filter is represented by the first... Order coefficient, This is expressed as the order of the FIR filter. Represented as the index of the number of discrete-time samples. This is represented by the FIR filter order index. By ensuring that the input waveforms received by each protection unit meet the requirements for phase and amplitude accuracy in both time synchronization and frequency band consistency for directional judgment and grading criteria, reliable input is provided for subsequent feature extraction based on machine learning. Secondly, by establishing a system response model for the secondary side of the transformer and performing regularized deconvolution processing in the frequency domain to correct the measurement response, the masking of fault features by the transformer's non-ideal frequency response and secondary saturation is eliminated. This ensures that the machine learning model reads the true primary signal features during the fault signal feature extraction stage. The specific formula is expressed as: ； in, Represented as the corrected frequency domain signal, Expressed as angular frequency, This is represented as the discrete-time Fourier transform of the filtered signal Y[p]. Expressed as the equivalent frequency response of the current transformer and sampling link, it represents the frequency domain response of the measurement channel to the actual primary quantity. Described as the complex conjugate of T(ω), Expressed as the square of the magnitude of T(ω), Represented as a regularization factor, precise transformer calibration can avoid directional misjudgments and stage conflicts caused by secondary saturation and uneven frequency response, thereby improving the algorithm's sensitivity and accuracy to the essential characteristics of overcurrent and grounding faults. This is represented as a time-domain corrected sample sequence. This is represented as an inverse Fourier transform, and then, through the corrected time-domain signal... A multi-level decomposition is performed using adjustable-factor wavelet transform to simultaneously preserve fault transient and harmonic components in the time and frequency domains. By calculating the kurtosis of each level and selecting several levels with the largest kurtosis as the most information-rich levels, the focus can be placed on sub-bands containing fault impact and arc features. This reduces redundancy and preserves fault discrimination information in machine learning feature extraction. The specific formula is as follows: ； in, Represented as the first The discrete coefficient sequence of wavelet subbands in the time domain represents the temporal energy distribution of that frequency band. Represented as the first The wavelet basis impulse response coefficients of the first layer represent the response coefficients used for the first layer. Layer-extracted filters, Represented as a decomposition layer index, Represented as the first The peak attitude quantity of the layer coefficient represents the first layer. The sharpness of the layer distribution, The larger the value, the more violent the transient. Represented as the first Layer coefficient sequence The number of samples, Represented as the first The sample mean of the layer coefficients, based on the selected layer. The output has targeted frequency band characteristics for use in downstream protection components and deep learning-based fault detectors, achieving synergy between time-frequency domain information and hierarchical protection logic. Secondly, by selecting wavelet layer coefficients... Calculate a set of time-domain and frequency-domain statistics to construct an initial set of feature vectors. The subband energy, normalized spectral entropy, and time-domain statistics are calculated to transform the original waveform into a concise and discriminative numerical representation that can be used by machine learning models. The specific formula is as follows: ； in, Represented as the first Total energy of the sub-bands Represented as the first Normalized energy percentage of the layer This represents the total number of layers used for feature calculation. Represented as the layer index used for feature calculation. Represented as the first Total energy of the sub-bands Expressed as spectral entropy, it reflects the uniformity of energy distribution between subbands. Represented as a logarithmic function, Expressed as the normalized spectral entropy, Represented as the root mean square value of the time-domain statistic. The kurtosis factor, expressed as a time-domain statistic, measures the ratio of the transient peak value to the root mean square. This represents the operation of taking the real part of a complex number. This is represented as taking the maximum value of the sequence, and then, by using the original statistical feature vector... The input is fed into a machine learning mapper, which generates a low-dimensional, highly discriminative final feature vector while preserving fault discrimination information. This vector can then be used as a protection criterion input in the system for combined use with traditional criteria by the deep learning fault detector. The specific formula is as follows: ； in, This is represented as the low-dimensional feature vector of the output. Represented as the ReLU nonlinear activation function, it transforms the result of a linear mapping into a nonlinear representation. Represented as a mapping matrix, Represented as a bias column vector, finally, by converting the final feature vector... This, along with a set of metadata including timestamp, measurement point ID, and preprocessing version number, is combined into a standardized feature package and sent to the fault diagnosis and protection execution module according to a fixed communication format. The specific formula is as follows: ； in, Represented as the sent feature packet vector, Represented as timestamp data, Represented as measurement point identification data, Represented as version number data, Represented as a vector transpose operation, the constructed feature package is sent to the fault diagnosis and protection execution module through a low-latency channel to ensure that the extracted fault features can be used by the system in real time to complete fault detection and protection decisions.

5. A multi-stage overcurrent and ground fault protection coordinated system according to claim 1, characterized in that: The fault diagnosis and protection execution module includes a fault detection and discrimination unit and a hierarchical coordination and decision-making unit. The fault detection and discrimination unit proposes an overcurrent and ground fault detection algorithm based on deep learning. By performing real-time analysis of the time-frequency domain features provided by the signal conditioning and feature extraction module, a deep learning model is built within the system for reasoning and identification. This algorithm can identify instantaneous and persistent faults, and distinguish between phase-to-phase short-circuit faults and single-phase ground faults, thereby ensuring high-sensitivity detection and reliable discrimination of multiple types of power faults even under complex operating conditions and strong interference environments. The hierarchical coordination and decision-making unit implements timer management, priority and backoff strategies, and interaction mechanisms with adjacent devices. It dynamically selects the execution scheme between instantaneous actions and inverse-time actions, ensuring that the protection actions have hierarchical selectivity, backup reliability, and coordination consistent with the system operation manual.

6. A multi-stage overcurrent and ground fault protection coordinated system according to claim 5, characterized in that, First, by constructing a normalized time-series matrix from the multi-channel sub-band features and time-domain statistics from the signal conditioning and feature extraction modules in chronological order, the deep recurrent network can learn the temporal patterns of overcurrent and ground fault detection using sequence dependencies. This allows the network input to simultaneously contain transient information and steady-state trends, thereby supporting the system's ability to distinguish between transient and persistent faults. The specific formula is as follows: ； in, Represented as time Feature column vectors Represented as a time index, At any moment The Each feature data, Let the feature dimension be expressed as the feature dimension at each time step. Represented as time Independent data with different features Represented as a time series matrix input into a deep learning model. This is expressed as the length of the time window, used to capture short-time series information. Represented as time Independent column vectors of different features will be used in the system As a standard input format for deep learning models, hierarchical protection can obtain consistent time-series views both locally and remotely for collaborative judgment. Then, by processing the input matrix in an online runtime environment... The data is progressively fed into deep recurrent units to learn time-dependent features, and a gating mechanism is used to suppress the impact of long-term and short-term noise on the discrimination. This ensures that the temporal extraction capability of the deep learning model can identify transient changes and persistent overcurrents, thereby meeting the system's requirements for high sensitivity and low false alarm rate. The specific formula is as follows: ； in, This is represented as an update gate vector, used to control the proportion of new information written. This is represented as an element-wise Sigmoid activation function. This is represented as the input weight matrix of the update gate. This is represented as the hidden weight matrix of the update gate. Represented as The hidden state at any given moment This is represented as the update gate bias vector. This is represented as a reset gate vector, used to control the strength of the influence of historical information on candidate calculation. This is represented as element-wise nonlinear activation. This is represented as the input weight matrix of the reset gate. This is represented as the hidden weight matrix of the reset gate. This is represented as the reset gate bias vector. Represented as a candidate hidden state, This is represented as the candidate input weight matrix. Represented as the candidate hidden weight matrix, This is represented as a candidate bias vector. This is represented as element-wise multiplication. Represented as The hidden state update at each moment, by configuring this gating recursion as a low-latency edge instance, ensures that subsequent units can be updated based on this. The temporal evolution of the model determines the occurrence of faults and triggers hierarchical actions. Secondly, by retaining short-window energy and long-window trend quantities in parallel within the model and calculating the energy ratio, transient and persistent faults are distinguished. This is combined with deep hidden states. The magnitude of the sudden change ensures that while maintaining high sensitivity, false alarms caused by transient noise are suppressed, enabling the system to distinguish between transient faults and true persistent faults. This allows for the correct triggering of transient tripping and delayed backup actions in graded protection. The specific formula is as follows: ； in, Represented as short-window energy, This is expressed as the length of a long-term period. Represented as long window energy, Indicated as short-time duration, Represented as a time index, It can be represented as the square of the L2 norm of a vector. Represented as the instantaneous energy of all characteristics at that moment. Represented as a transient ratio indicator, when If the threshold is exceeded, it is judged as a momentary strong impact, which will trigger a momentary action candidate. If a high value is maintained for more than t+1000 durations, it is considered a persistent fault. Represented as regularization parameters, With depth hidden state The transient response is combined with subsequent decision-making units. In the event of a sudden, strong impact, instantaneous actions are prioritized. For persistent faults, timed and inverse-time backup actions are initiated to ensure selectivity and minimal sacrifice. Then, the energy feature sequence is input into a constructed deep convolutional recurrent hybrid network to enhance the detection accuracy of overcurrent and grounding faults under complex operating conditions. Energy mutation features are extracted within a short window and used as local input to the convolutional neural network to capture fast transient patterns. Simultaneously, the continuity of the time series is maintained within a long window, and a long short-term memory network is used to model the global evolution trend, ensuring that the algorithm can identify persistent overcurrent and slowly developing high-impedance grounding faults. The specific formula is expressed as follows: ； in, This is represented by the output vector of the deep learning model, indicating the probability prediction results of the deep convolutional recurrent hybrid network for transient and persistent faults. Represented as a convolutional recurrent hybrid network, This is used to determine the length of the data frame. This is represented as the final joint judgment indicator. This is represented as a joint weighting factor, used to balance the influence of physical indicators and deep learning output. Through this dual-path fusion of convolution and recursion, the constructed deep learning model can form a composite judgment mechanism that is sensitive to sudden changes and robust to slow trends. This enables the system to maintain stable fault identification capabilities even under extreme power disturbances. Finally, the output of the deep learning model is combined with time-frequency domain features and input into a multi-layer neural network classifier to achieve the final identification of fault types. The multi-layer neural network classifier contains multiple fully connected layers and an attention mechanism. The fully connected layers are responsible for extracting global statistical patterns, while the attention mechanism automatically focuses on key frequency bands and transient features, thereby improving the identification accuracy of complex fault modes in high-noise environments. The specific formula is expressed as follows: ； in, Represented as a concatenated input vector, this concatenates the deep learning model output with time-frequency features in a dimensional manner, serving as the raw input to the classifier. Represented as an attention weight vector, This can be represented as a function that exponentially normalizes the vector elements. Represented as an attention weight mapping matrix, This is represented as the attention layer bias vector. Represented as attention-weighted feature vectors, This is represented as the final classification probability vector, used to determine overcurrent faults, grounding faults, and mixed faults. Represented as a probability distribution transformation function, Represented as the output layer weight matrix, This is represented as the weight matrix of the first fully connected layer. This is represented as the first-level bias vector. Represented as an output layer bias vector, it is then weighted by attention and entered into two fully connected layers to activate the output structure. The system can map the probability results into explicit protection action priorities and strategies.

7. A multi-stage overcurrent and ground fault protection coordinated system according to claim 1, characterized in that, The data statistics and communication module includes a database construction unit and a communication and interconnection unit. The database construction unit designs an architecture that combines a historical fault event database with real-time file storage and an event indexing mechanism, and supports efficient retrieval and export interfaces, so that fault records and operational statistics are available and complete in querying and long-term analysis. The communication and interconnection unit implements Ethernet and fiber optic physical interfaces, link redundancy strategies, and a precise time synchronization mechanism to achieve low-latency, deterministic, and highly reliable real-time information interaction and control coordination with adjacent protection devices and multi-level overcurrent and ground fault protection collaborative systems.

8. A multi-stage overcurrent and ground fault protection coordinated system according to claim 1, characterized in that: The security and access control module establishes hierarchical permission management, operation identity authentication, interaction process verification, and key operation error prevention mechanisms to control the entire process of user commands, system parameter adjustments, and protection action issuance, thereby achieving secure and controllable system interaction and avoiding protection failures and erroneous actions caused by misoperation.

Images