A method for online real-time monitoring and early warning of hidden open circuits in relay protection secondary circuits

By combining physical simulation enhancement and adaptive feature selection with customized residual attention neural networks and fuzzy comprehensive evaluation method, the problems of low diagnostic accuracy and long-term accuracy decay of hidden open circuits in the secondary circuit of relay protection are solved, and real-time and efficient fault monitoring and early warning are realized.

CN123121406APending Publication Date: 2026-11-17INNER MONGOLIA HELIN POWER GENERATION CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611310996.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-27
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies suffer from high dimensionality and severe redundancy in the characteristic data of hidden open circuit faults in the secondary circuit of relay protection. Traditional methods lack effective screening and dimensionality reduction mechanisms, convolutional neural networks are not sensitive enough to weak features, and lack real-time situational awareness and adaptive update mechanisms for cascading faults, resulting in low diagnostic accuracy and long-term operational accuracy decay.

Method used

A data preprocessing method based on physical simulation enhancement and adaptive feature selection is adopted, combined with a customized residual attention neural network model and fuzzy comprehensive evaluation method for fault location and risk assessment, and the model is updated in a closed loop through cross-entropy confidence assessment and incremental training strategy.

Benefits of technology

It enables efficient screening of key fault features, enhances sensitivity to weak features, assesses the risk of cascading failures in real time, ensures diagnostic accuracy and early warning precision during long-term operation, and solves the problems of rigidity and accuracy decay of traditional methods.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123121406A_ABST
    Figure CN123121406A_ABST
Patent Text Reader

Abstract

This application relates to the field of power system relay protection technology and discloses a method for online real-time monitoring and early warning of hidden open circuits in relay protection secondary circuits, including a system initialization stage and an online operation stage. The system initialization stage includes the following steps: Step 1: Collect historical multi-dimensional state data and construct a one-dimensional initial feature vector; Step 2: Perform adaptive feature filtering on the enhanced dataset to obtain key fault feature sequences; Step 3: Input the three-dimensional time-series feature tensor into a customized residual attention neural network model for offline training. This method for online real-time monitoring and early warning of hidden open circuits in relay protection secondary circuits uses a dynamic low-variance filtering method to filter out low-variance features, and then uses an adaptive chi-square test model to remove features that are not significantly correlated with the fault category based on a dynamic threshold set according to the power grid operating topology. This dual adaptive filtering reduces the feature dimensionality and eliminates redundancy.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of power system relay protection technology, and in particular to a method for online real-time monitoring and early warning of hidden open circuits in the secondary circuit of relay protection. Background Technology

[0002] The secondary circuit of relay protection is a key system for ensuring the safe operation of the power grid. It realizes the acquisition, transmission and logical operation of electrical quantities through secondary equipment such as merging units, intelligent terminals and protection devices. During the operation of smart substations, the hidden open circuit faults in the secondary circuit are physically invisible and the message interaction is complex. Therefore, maintenance personnel need to manually investigate by monitoring the communication status and massive alarm information. Thus, an online real-time monitoring and early warning method is needed to improve the diagnostic efficiency of hidden open circuit faults in the secondary circuit.

[0003] Existing secondary circuit fault monitoring technologies mainly rely on network analyzer packet capture and expert rule bases of background monitoring systems. In terms of data processing, traditional convolutional neural network models are typically used to classify the collected fault features, directly inputting one-dimensional time-series data into the network. In terms of fault early warning, existing technologies use a single alarm judgment method based on fixed thresholds, and perform post-event analysis by matching historical fault case databases.

[0004] However, existing secondary circuit fault monitoring technologies still have the following problems that urgently need to be solved: I. The feature data of hidden line breakage faults are highly dimensional and redundant. Traditional methods lack effective screening and dimensionality reduction mechanisms for fault features, and the fixed threshold screening method cannot adapt to different power grid operating conditions, resulting in difficulty in model training and low diagnostic accuracy. Second, traditional convolutional neural networks lack multi-scale perception capabilities for power frequency characteristics when processing one-dimensional power time-series signals, and suffer from gradient vanishing and loss of key weak features when processing complex fault signals, and are not sensitive enough to the weak features of hidden line breaks. Third, the lack of a real-time situational awareness mechanism for cascading failures caused by secondary circuit disconnection, the fixed-weight risk calculation method cannot reflect the nonlinear coupling relationship between indicators, and the model cannot perform adaptive closed-loop updates based on the on-site troubleshooting results, resulting in the decline of the accuracy of the early warning system during long-term operation. Therefore, a method for online real-time monitoring and early warning of hidden open circuits in the secondary circuit of relay protection is proposed. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this application provides an online real-time monitoring and early warning method for hidden open circuits in relay protection secondary circuits. This method can solve the problems of low diagnostic accuracy due to rigid feature screening mechanisms in traditional methods, insufficient sensitivity of convolutional networks to weak features, and long-term operational accuracy degradation due to the lack of real-time evaluation and adaptive update mechanisms for cascading faults.

[0006] According to the method for online real-time monitoring and early warning of hidden open circuits in the secondary circuit of relay protection provided in this application, the method includes: System initialization phase and online operation phase; The system initialization phase includes the following steps: Step 1: Collect historical multi-dimensional state data, construct a one-dimensional initial feature vector, establish the equivalent circuit equation of the secondary loop based on the implicit disconnection physical mechanism, and generate simulated fault waveform data by jointly sampling the disconnection time and transition resistance parameters. The simulated fault waveform data is then superimposed with the one-dimensional initial feature vector to complete data augmentation. Step 2: Adaptive feature selection is performed on the enhanced dataset to obtain the key fault feature sequence. The dataset is then resampled according to a fixed number of points within the power frequency cycle, and multi-domain features are extracted to construct a three-dimensional time-series feature tensor. Step 3: The customized residual attention neural network model constructed by inputting the three-dimensional temporal feature tensor is trained offline to generate a benchmark diagnostic model, a fault coding mapping dictionary, and feature mapping rules; The online operation phase includes: Step 4: Collect multi-dimensional status data during the operation of the secondary circuit of the relay protection in real time, and generate a real-time three-dimensional time-series feature tensor using the same normalization parameters and feature mapping rules as in the system initialization phase. Step 5: Input the real-time three-dimensional temporal feature tensor into the benchmark diagnostic model, and output the location result and fault code of the secondary circuit hidden open circuit fault; Step 6: Based on the fault location results, trigger the chain fault situation awareness module. Based on the pre-built chain fault event chain template, use the hierarchical string fuzzy matching algorithm to perform real-time risk assessment. Use the fuzzy comprehensive evaluation method with time-series penalty factor to calculate the comprehensive risk level. Introduce supplementary judgment criteria to determine the final risk level and output early warning information to the application platform layer. Step 7: Based on the cross-entropy confidence assessment mechanism, the consistency between the early warning information and the on-site defect elimination results is judged, high confidence samples are selected, the upper limit of the training set window is set, and the benchmark diagnostic model is updated in a closed loop by combining the incremental training strategy based on the performance monitoring of the validation set.

[0007] Preferably, the time of wire breakage in step one follows a uniform distribution within one power frequency cycle; the transition resistance parameter follows a log-normal distribution with a mean of 200. The standard deviation is 50. The value range is 10. Up to 1000 The joint sampling strategy accurately matches the poor contact physical transient characteristics of latent breaks, which distinguish them from metallic breaks; the specific formula for the data augmentation is: ; in For the enhanced feature vector, The original one-dimensional initial feature vector is the one-dimensional feature vector collected. For fault waveform characteristics generated based on circuit equation simulation, A random scaling factor ranging from 0.1 to 0.5; During the online operation phase, the simulation waveform feature dictionary generated during the system initialization phase is only used for table lookup during the online phase and does not perform electromagnetic transient simulation calculations.

[0008] Preferably, the adaptive feature selection in step two specifically involves: Step 1: Obtain the real-time operating topology status of the power grid. When the power grid is in grid-connected operation mode, set the first significance level dynamic threshold. When the power grid is in islanded operation mode, set the second significance level dynamic threshold. Step 2: Use dynamic low variance filtering to extract the variance distribution probability density function of each feature under historical normal operating conditions, take the lower limit of the 95% confidence interval as the dynamic variance threshold, and screen out feature quantities whose sample variance is less than the dynamic variance threshold. Step 3: Use an adaptive chi-square test model to calculate the chi-square statistic and corresponding P-value between each feature retained after low variance filtering and the fault category label, and filter out features with P-values ​​greater than the dynamic significance level threshold corresponding to the current running topology.

[0009] Preferably, the three-dimensional temporal feature tensor reconstruction method in step two is as follows: Step 1: Set the sampling rate of the original analog quantity to 4000Hz, and extract 80 time-series sampling points for the characteristic quantity in each channel within one power frequency cycle of 20ms; Step 2: Divide the key fault feature sequence into 12 channels: Channels 1 to 3 are the three-phase current and angle characteristics on the high-voltage side, Channels 4 to 6 are the three-phase current and angle characteristics on the low-voltage side, Channels 7 to 9 are the three-phase voltage and angle characteristics on the high-voltage side, and Channels 10 to 12 are the three-phase voltage and angle characteristics on the low-voltage side. Step 3: Extract three feature subspaces for the feature quantities in each channel: time domain features, frequency domain features, and wavelet packet energy features, and construct a three-dimensional time-series feature tensor with a size of 12×80×3.

[0010] Preferably, the customized residual attention neural network model described in step three includes an input layer, multiple customized attention residual blocks, an average pooling layer, and a fully connected layer; The customized attention residual block includes a parallel multi-scale convolution module, a first batch of normalized layers, a first activation function layer, a second batch of normalized layers, a convolution block attention module, and an identity mapping adder. The parallel multi-scale convolution module contains three parallel convolution branches, namely: A standard convolutional branch with a kernel size of 1×3 is used to extract the fundamental frequency features. A dilated convolutional branch with a kernel size of 1×5 and a dilation rate of 2 is used to extract low-frequency harmonic features; A deformable convolutional branch with a kernel size of 1×7 is used to adaptively adjust the sampling point weights to match the distortion start point features of the fault transient waveform; the outputs of the three branches are spliced ​​together in the channel dimension.

[0011] Preferably, both the offline training in step three and the closed-loop update in step seven use the FocalLoss loss function, with the following formula: ; in This represents the model's predicted probability for the true class. The focus parameter is set to 2. The category balance weight parameter is calculated using the following formula: ; in Let be the number of samples in class t, C be the total number of classes, and satisfy the normalization constraint. During the closed-loop update process, the global mean and variance statistical parameters of all batch standardized layers in the customized residual attention neural network model are frozen.

[0012] Preferably, the fault code mapping dictionary in step three uses a 16-bit binary encoding rule to generate hexadecimal fault codes: Bits D15 to D13 indicate the major fault category, bits D12 to D11 indicate the fault type of the sampling circuit, bits D10 to D8 indicate the type of associated device, bits D7 to D6 indicate the link type, bits D5 to D3 indicate the fault phase sequence, bits D2 to D1 indicate the high-voltage side or low-voltage side, and bit D0 indicates the sampling circuit link break status. The fault location result described in step five is used to parse the specific fault location from the fault code mapping dictionary.

[0013] Preferably, the hierarchical string fuzzy matching algorithm described in step six is ​​as follows: Step 1: Calculate the similarity between the real-time collected event strings and the template fields based on the edit distance algorithm; Step 2: Set a hierarchical matching similarity threshold. For critical events involving main protection actions and circuit breaker tripping, set a matching similarity threshold of 98%. For non-critical events involving abnormal alarms and backup protection actions, set a matching similarity threshold of 90%. Step 3: When the number of events matched according to the event sequence is equal to the total number of events in the event chain and each event meets the corresponding grade threshold, the match is considered successful.

[0014] Preferably, the specific process of calculating the comprehensive risk level using the fuzzy comprehensive evaluation method in step six is ​​as follows: Step 1: Establish the factor set U= , , ; in, For the probability of cascading failures, The percentage decrease in substation availability. Loss of load rate With the comment set V=1, 2, 3; Among them, Level 1 is critical, Level 2 is serious, and Level 3 is moderate; Step 2: Construct the membership degree of each index using the triangular membership function to obtain the fuzzy relation matrix R; Step 3: Construct the time-series penalty factor ,in The attenuation coefficient is... The absolute value of the deviation between the actual event occurrence time and the time set in the event chain template is used to determine the timing penalty factor. The weights are corrected as a prerequisite for the calculation of the dynamic weight vector A; Step 4: Through fuzzy transformation Calculate the comprehensive evaluation vector A; Step 5: When the difference between the largest and second largest membership degree is less than 0.1, the weighted average principle is used to calculate the comprehensive risk value to determine the risk level. When the difference is greater than or equal to 0.1, the largest membership degree principle is used to determine the risk level.

[0015] Preferably, the closed-loop update process in step seven is as follows: Step 1: Calculate the cross-entropy between the actual fault result label and the model output probability distribution. Set the cross-entropy confidence threshold to 0.2. When the cross-entropy is less than 0.2, it is judged as a high-confidence sample. Step 2: Introduce a sliding upper limit mechanism for the training set window, setting the maximum capacity of the training set to 10,000 samples. When a high-confidence sample is added to the training set, causing the total number to exceed 10,000, the oldest sample is eliminated according to the first-in-first-out principle. Step 3: After each incremental training, calculate the F1-score of the model using an independent validation set. If the F1-score of the updated model is lower than that before the update, trigger the model rollback mechanism to restore the parameters of the previous version of the model. Step 4: During incremental training, freeze the weight parameters of the parallel multi-scale convolutional modules in the customized residual attention neural network model, and only fine-tune and update the weight parameters of the convolutional block attention modules and fully connected layers.

[0016] Compared with the prior art, this application provides a method for online real-time monitoring and early warning of hidden open circuits in relay protection secondary circuits, which has the following beneficial effects: 1. The online real-time monitoring and early warning method for hidden open circuits in the secondary circuit of the relay protection uses a dynamic low-variance filtering method to filter out low-variance features, and then uses an adaptive chi-square test model to remove features that are not significantly correlated with the fault category based on the dynamic threshold set according to the power grid operating topology. The dual adaptive screening reduces the feature dimension and eliminates redundancy, which significantly reduces the difficulty of model training.

[0017] 2. The online real-time monitoring and early warning method for hidden open circuits in the secondary circuit of the relay protection uses three branches of the parallel multi-scale convolution module in the customized residual attention neural network model to extract the features of the power frequency fundamental wave, low-frequency harmonics, and transient distortion initiation point, respectively. After dual weighting by the channel and spatial attention of the convolution block attention module, the key fault features are highlighted. The input features and the processed features are added by the identity mapping adder and passed to the next layer, which effectively suppresses gradient vanishing and enhances the sensitivity to weak features.

[0018] 3. The online real-time monitoring and early warning method for hidden open circuits in the secondary circuit of relay protection uses a hierarchical string fuzzy matching algorithm to determine the similarity between real-time events and event chain templates, corrects the dynamic weight vector with a time-series penalty factor, outputs the risk level through fuzzy comprehensive evaluation, combines cross-entropy confidence assessment to screen high-confidence samples, and adopts an incremental training strategy with validation set performance monitoring to update the model in a closed loop, ensuring continuous optimization capability during long-term operation.

[0019] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this application, nor is it intended to limit the scope of this application. Other features of this application will become readily apparent from the following description. Attached Figure Description

[0020] To more clearly illustrate the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 A flowchart illustrating an online real-time monitoring and early warning method for hidden open circuits in the secondary circuit of a relay protection system, provided in an embodiment of this application. Figure 2 This is a schematic diagram of the structure of the customized residual attention neural network model provided in the embodiments of this application. Detailed Implementation

[0022] The following description, in conjunction with the accompanying drawings, illustrates exemplary embodiments of this application, including various details to aid understanding. These should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0023] The following describes, with reference to the accompanying drawings, an embodiment of the present application of a method for online real-time monitoring and early warning of hidden open circuits in the secondary circuit of relay protection. Figure 1 This is a flowchart illustrating a method for online real-time monitoring and early warning of hidden open circuits in the secondary circuit of a relay protection system, as provided in an embodiment of this application. Figure 1 As shown.

[0024] Example 1: To address the high dimensionality and redundancy issues of hidden disconnection characteristics in relay protection secondary circuits, as well as the requirement for physical mechanism matching, this embodiment provides a system initialization data preprocessing method based on physical simulation enhancement and adaptive feature selection.

[0025] The specific implementation steps of this embodiment are as follows: Step S1: Collect historical multi-dimensional status data through the process layer equipment of the intelligent substation; analog measurement data includes three-phase current and angle on the high and low voltage sides, positive and negative zero-sequence current and angle, three-phase voltage and angle on the high and low voltage sides, and positive and negative zero-sequence voltage and angle; digital alarm information includes GOOSE link break alarm, SV sampling link interruption alarm, and device lockout alarm; concatenate the above 119-dimensional data into a one-dimensional row vector according to the time sequence to construct a one-dimensional initial feature vector; Step S2: Based on the physical mechanism of implicit wire breakage, namely the decrease in effective current value, phase shift, and high-frequency transient components caused by poor contact or increased transition resistance, establish the equivalent circuit equation for the secondary loop; in the simulation model, the wire breakage time is randomly selected according to a uniform distribution within one power frequency cycle, and the transition resistance at the wire breakage point follows a mean of [value missing]. The standard deviation is 50. The log-normal distribution has a range of values ​​of 10. Up to 1000 This joint sampling strategy accurately matches the physical transient characteristics of poor contact that distinguish latent open circuits from metallic open circuits; by changing the above parameters, electromagnetic transient simulations are performed, generating a large amount of simulated fault waveform data covering different operating conditions and forming a waveform feature dictionary; the formula is used. The simulated fault waveform features are superimposed with the original one-dimensional initial feature vector. Data augmentation is performed using a random scaling factor ranging from 0.1 to 0.5.

[0026] Step S3: Perform adaptive feature filtering on the enhanced dataset; obtain the real-time operating topology of the power grid; when the power grid is in grid-connected operation mode, set a dynamic threshold for the first significance level; when the power grid is in islanded operation mode, set a dynamic threshold for the second significance level; use dynamic low variance filtering to normalize the maximum and minimum values ​​of each feature in the dataset and calculate the sample variance of each feature; extract the variance distribution probability density function of each feature under historical normal operating conditions, and take the lower limit of the 95% confidence interval as the dynamic variance threshold; filter out features with sample variance less than the dynamic variance threshold, reducing the 119-dimensional features to 79-dimensional features.

[0027] Step S4: Use an adaptive chi-square test model to perform secondary screening on the 79-dimensional features; calculate the chi-square statistic and corresponding P-value between each feature quantity and the fault category label; filter out features whose P-value is greater than the dynamic threshold of the significance level corresponding to the current running topology, and finally obtain the key fault feature sequence composed of 69 key fault features, and record the above normalization parameters and feature mapping rules for use in the online stage.

[0028] Example 2: To address the shortcomings of traditional neural networks in terms of sensitivity to subtle features of hidden line breaks and dimensionality adaptation, this embodiment provides a method for constructing a fault location model and conducting online diagnosis based on a customized residual attention neural network.

[0029] This embodiment is implemented based on the key fault feature sequence obtained in Embodiment 1, and the specific steps are as follows: Step S5: Set the sampling rate of the original analog quantity to 4000Hz, and extract 80 time-series sampling points for the characteristic quantity in each channel within one power frequency cycle of 20ms; considering that the sampling rate of 4000Hz and the 80 time-series sampling points correspond to the matching relationship between the Nyquist sampling theorem and the highest frequency of the 50th harmonic analysis within the 20ms power frequency cycle, this parameter selection satisfies the fidelity requirements of the high-frequency transient components and avoids the calculation delay caused by data redundancy; The key fault characteristic sequence is divided into 12 channels according to the category of physical quantities: channels 1 to 3 are the three-phase current and angle characteristics of the high-voltage side, channels 4 to 6 are the three-phase current and angle characteristics of the low-voltage side, channels 7 to 9 are the three-phase voltage and angle characteristics of the high-voltage side, and channels 10 to 12 are the three-phase voltage and angle characteristics of the low-voltage side. For each channel, three feature subspaces are extracted: time-domain features, frequency-domain features, and wavelet packet energy features. A subspace of size [size missing] is constructed. The three-dimensional temporal feature tensor.

[0030] Step S6: Construct a customized residual attention neural network model; Figure 2 This is a schematic diagram of the structure of the customized residual attention neural network model provided in the embodiments of this application, as shown below. Figure 2 As shown: The model consists of an input layer, multiple customized attention residual blocks, an average pooling layer, and a fully connected layer. Each customized attention residual block contains a parallel multi-scale convolution module, a first batch of normalized layers, a first activation function layer, a second batch of normalized layers, a convolution block attention module, and an identity mapping adder. The parallel multi-scale convolution module contains three parallel convolution branches: a standard convolution branch with a kernel size of 1×3 for extracting power frequency fundamental wave features, a dilated convolution branch with a kernel size of 1×5 and a dilation rate of 2 for expanding the receptive field to capture low-frequency harmonic features, and a deformable convolution branch with a kernel size of 1×7 for adaptively capturing transient irregular disturbance features of electric arcs. The outputs of the three branches are concatenated in the channel dimension. For deformable convolutional branches, by introducing an offset learning mechanism, it is used in this scenario to adaptively adjust the sampling point weights to match the distortion start point features of the fault transient waveform. Its offset corresponds in a physical sense to the non-rigid deformation of the current phase shift caused by the fluctuation of the hidden open circuit contact resistance and the high-frequency distortion position, thereby transforming it from a general image processing module into a dedicated feature extraction module designed for the distortion features of transient waveforms of power time series signals. The identity mapping adder adds the input features directly to the features processed by the convolutional block attention module.

[0031] Step S7: Perform channel attention and spatial attention calculations in the attention residual block; The channel attention module performs global average pooling and global max pooling operations on the input feature map, inputs the two resulting feature maps into a shared multilayer perceptron, adds the two feature maps output by the multilayer perceptron element by element, and generates a channel attention feature map by passing it through the sigmoid activation function. The spatial attention module performs average pooling and max pooling along the channel dimension on the result of multiplying the channel attention feature map with the input feature map. The two feature maps are then concatenated along the channel dimension and passed through a 7×7 convolutional layer and a Sigmoid activation function to generate a spatial attention feature map.

[0032] Step S8: Input the reconstructed 3D temporal feature tensor into the customized residual attention neural network model constructed above for offline training; The model training process uses the FocalLoss loss function to address the sample imbalance problem. The formula for the FocalLoss loss function is: ; in This represents the model's predicted probability for the true class. The focus parameter is set to 2. The category balance weight parameter is calculated using the following formula:

[0033] in Let be the number of samples in class t, C be the total number of classes, and satisfy the normalization constraint. The closed-loop update process freezes the global mean and variance statistics of all batch standardized layers in the customized residual attention neural network model. An Adam optimizer with a learning rate cosine annealing strategy is used for parameter updates. After the model is trained, a baseline diagnostic model is generated, and a fault code mapping dictionary is generated according to the sixteen-bit binary encoding rule: Bits D15 to D13 indicate the major fault category, bits D12 to D11 indicate the fault type of the sampling circuit, bits D10 to D8 indicate the type of associated device, bits D7 to D6 indicate the link type, bits D5 to D3 indicate the fault phase sequence, bits D2 to D1 indicate the high-voltage side or low-voltage side, and bit D0 indicates the sampling circuit link break status. During the online operation phase, real-time data directly uses the same normalization parameters and feature mapping rules as in the initialization phase to generate real-time three-dimensional time-series feature tensors. The simulation waveform feature dictionary generated in the system initialization phase is only used for table lookup in the online phase and does not perform electromagnetic transient simulation calculations, thus ensuring the 20ms warning time limit.

[0034] Example 3: To address the lack of early warning for cascading faults caused by secondary circuit disconnection and the problem of nonlinear risk assessment, this embodiment provides a situational awareness method based on hierarchical fuzzy matching and fuzzy comprehensive evaluation.

[0035] This embodiment is triggered during the online operation phase after the benchmark diagnostic model outputs a fault code. The specific implementation method is as follows: Step S9: Parse the specific fault location based on the fault code mapping dictionary and construct a cascading fault event chain template; Based on protection principles and operating conditions, power grid topology, substation main wiring and protection configuration, a primary equipment adjacency matrix is ​​defined. Analyze the set of causes for protection malfunctions and construct a protection failure fault tree. Based on the primary equipment adjacency matrix, find the set of adjacent primary equipment and adjacent protection equipment of the faulty equipment, and generate a chain fault event template containing event name, triggering condition, action behavior and event description.

[0036] Step S10: Conduct a real-time risk assessment; Real-time acquisition of protection action signals, circuit breaker opening signals, alarm signals, protection settings and hard and soft pressure plate status of the secondary circuit of relay protection; and calculation of string similarity between real-time events and template events using a hierarchical string fuzzy matching algorithm based on edit distance. Set tiered matching similarity thresholds: For critical events involving main protection actions and circuit breaker tripping, the matching similarity threshold is set to 98%; for non-critical events involving abnormal alarms and backup protection actions, the matching similarity threshold is set to 90%. A match is considered successful when the number of events matched according to the event sequence is equal to the total number of events in the event chain and each event meets the corresponding grade threshold. Step S11: Calculate the overall risk level of the successfully matched chain of cascading failure events; Calculate the probability of cascading failures: ; Calculate the substation availability rate decline ratio Calculate the load shedding rate ; Establish a factor set: U= , , ; Comments set: V=1, 2, 3; Level 1 is critical, Level 2 is serious, and Level 3 is normal.

[0037] The fuzzy relation matrix is ​​obtained by constructing the membership degrees of each index using the triangular membership function. .

[0038] Step S12, construct the time-series penalty factor ,in The attenuation coefficient is... The absolute value of the deviation between the actual event occurrence time and the time set in the event chain template is used to determine the timing penalty factor. The weights are corrected as a prerequisite for the calculation of the dynamic weight vector A; The dynamic weight vector During the system initialization phase, the analytic hierarchy process (AHP) is pre-calculated and stored based on typical operating modes. When the power grid topology undergoes significant changes, the operation and maintenance personnel trigger a recalculation. During the online evaluation phase, the corresponding weight vector is only called by looking up the table based on the current operating mode, without performing AHP eigenvalue decomposition calculation in real time, thus achieving a balance between adaptability and real-time performance. Through fuzzy transformation Calculate the comprehensive evaluation vector B; considering that in multidimensional risk assessment, when the difference between the largest and second largest membership degree is less than 0.1, the membership degree of each risk level tends to be uniform, and the maximum membership degree principle is prone to boundary oscillation distortion, the weighted average principle is introduced to perform dimensionality reduction analysis to improve robustness. Therefore, when the difference is less than 0.1, the weighted average principle is used to determine the risk level, and when it is greater than or equal to 0.1, the maximum membership degree principle is used. Finally, the early warning information is output to the application platform layer.

[0039] Example 4: To improve the system's adaptability and prevent noisy data from interfering with the model's accuracy, this embodiment proposes a closed-loop feedback update method based on confidence assessment and performance monitoring.

[0040] This embodiment implements the alert by outputting warning information at the application platform layer and having it executed by operations and maintenance personnel. The specific steps are as follows: Step S13: Based on the early warning information output by the application platform layer, the maintenance personnel carry out on-site inspection and troubleshooting of the physical equipment of the relay protection secondary circuit of the intelligent station, and obtain the actual fault result label.

[0041] Step S14: Calculate the cross-entropy between the actual fault result label and the output probability distribution of the customized residual attention neural network model. Set the cross-entropy confidence threshold to 0.2. When the calculated cross-entropy is less than 0.2, it is judged as a high-confidence sample; when the calculated cross-entropy is greater than or equal to 0.2, it is judged as a low-confidence sample. For low-confidence samples, trigger the manual review process to remove noise data caused by abnormal sensor sampling.

[0042] Step S15: Add high-confidence samples to the training set to trigger model closed-loop update; A sliding upper limit mechanism for the training set window is introduced, setting the maximum capacity of the training set to 10,000 samples. This value corresponds to the balance between the annual fault frequency of a typical substation and the minimum number of samples required for incremental model convergence. This ensures that the training set covers enough new fault distribution features while preventing the degradation of inference speed caused by the infinite accumulation of samples. When the total number of samples exceeds 10,000 due to the addition of high-confidence samples to the training set, the oldest samples are eliminated according to the first-in-first-out principle. After each incremental training, the F1-score of the model is calculated using an independent validation set. If the F1-score of the updated model is lower than that before the update, the model rollback mechanism is triggered to restore the parameters of the previous version of the model. If the F1-score is equal to or higher than that before the update, the updated parameters are retained.

[0043] Step S16: The closed-loop update process adopts an incremental training strategy, freezing the weight parameters of the parallel multi-scale convolutional modules in the customized residual attention neural network model, and freezing the global mean and variance statistics parameters of all batch normalized layers. Only the weight parameters of the convolutional block attention modules and fully connected layers are fine-tuned and updated. The fine-tuning learning rate is set to one-tenth of the initial learning rate, i.e., 0.0001. This mechanism ensures that the model has the ability to learn continuously when faced with new fault features, while effectively preventing catastrophic forgetting and deterioration of reasoning speed.

[0044] Example 5: To verify the objective effect of the above technical solution, this embodiment designed a control experiment and analyzed the principle.

[0045] I. Experimental Design and Scheme: This experiment constructed a hybrid test set containing real historical operating data and electromagnetic transient simulation data, totaling 5000 samples; to verify the synergistic effect of each core technical feature, three control schemes were set up: Comparison Scheme A: The traditional Gaussian white noise superposition scheme without the introduction of physical simulation enhancement is adopted, while retaining the other features of this application; Comparison Scheme B: Using the standard ResNet18 network without introducing parallel multi-scale convolution modules, while retaining the remaining features of this application; Comparative Option C: A risk assessment method with a fixed threshold is adopted, but the fuzzy comprehensive evaluation method with the introduction of a time-series penalty factor is not used, while retaining the other features of this application.

[0046] The complete solution in this application was tested in the same hardware and software environment as the three sets of comparative solutions mentioned above.

[0047] II. Experimental Data: The test results are shown in the table below:

[0048] The experimental data shows that the complete solution proposed in this application achieves optimal results in terms of diagnostic accuracy, false alarm rate control, and early warning timeliness. By analyzing the underlying logic using data, the complete solution achieves an accuracy rate of 99.9%, while the solution using a single replacement module only achieves 96%~97%. This is attributed to the serial synergistic effect of various technical features. First, the dual dynamic feature screening mechanism, through dynamic low-variance filtering and adaptive chi-square test, eliminates interference and redundant terms unrelated to hidden disconnections before the data is input into the model. This significantly improves the signal-to-noise ratio of the multi-scale convolution module during training and inference, avoiding the erroneous induction of attention weights by noisy data. This is the fundamental premise for subsequent high-precision diagnosis. Second, the multi-scale convolution uses 1×3 standard convolution to extract the fundamental frequency wave, 1×5 dilated convolution to extract low-frequency harmonics, and 1×7 deformable convolution to adaptively match the distortion start point features of the fault transient waveform, forming a rigorous feature mapping relationship. This further refines weak fault features within the denoised feature space.

[0049] Compared to Scheme C, which has a risk warning lead time of only 5ms, this application achieves 20ms. This is because the nonlinear fuzzy evaluation method that introduces a timing penalty factor avoids false alarms and false alarms in complex working conditions by penalizing events that deviate from the timing logic. As a result, a reliable warning can be given in the early stage of fault evolution.

[0050] The above data and principle analysis fully demonstrate that the combination of the various technical features of this application produces an unexpected technical effect where "the whole is greater than the sum of its parts".

[0051] In summary, this method for online real-time monitoring and early warning of hidden open circuits in relay protection secondary circuits uses a dynamic low-variance filtering method to filter out low-variance features, and then uses an adaptive chi-square test model to eliminate features that are not significantly correlated with the fault category based on a dynamic threshold set according to the power grid operating topology. This dual adaptive screening reduces the feature dimensionality and eliminates redundancy, significantly reducing the difficulty of model training.

[0052] Furthermore, this method for online real-time monitoring and early warning of hidden open circuits in the secondary circuit of relay protection extracts the features of the fundamental frequency, low-frequency harmonics, and transient distortion starting point through three branches of the parallel multi-scale convolution module in the customized residual attention neural network model. The features are then weighted by the channel and spatial attention of the convolution block attention module to highlight key fault features. The input features and the processed features are added by the identity mapping adder and passed to the next layer, effectively suppressing gradient vanishing and enhancing the sensitivity to weak features.

[0053] Furthermore, this method for online real-time monitoring and early warning of hidden open circuits in relay protection secondary circuits uses a hierarchical string fuzzy matching algorithm to determine the similarity between real-time events and event chain templates, corrects the dynamic weight vector with a time-series penalty factor, outputs the risk level through fuzzy comprehensive evaluation, combines cross-entropy confidence assessment to screen high-confidence samples, and employs an incremental training strategy with validation set performance monitoring to update the model in a closed loop, ensuring continuous optimization capabilities during long-term operation. This solves the problems of low diagnostic accuracy due to rigid feature selection mechanisms in traditional methods, insufficient sensitivity of convolutional networks to weak features, and long-term accuracy decay due to the lack of real-time evaluation and adaptive update mechanisms for cascading faults.

[0054] The relevant modules involved in this system are all hardware system modules or functional modules that combine computer software programs or protocols with hardware in the prior art. The computer software programs or protocols involved in these functional modules are technologies known to those skilled in the art and are not improvements to this system. The improvement of this system lies in the interaction or connection between the modules, that is, in improving the overall structure of the system to solve the corresponding technical problems that this system aims to address.

[0055] Although embodiments of this application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of this application, the scope of which is defined by the appended claims and their equivalents.

[0056] The various numerical designations such as "first," "second," etc., used in this application are merely for ease of description and are not intended to limit the scope of the embodiments of this application, nor do they indicate the order of events.

[0057] At least one in this application can also be described as one or more, and multiple can be two, three, four or more, and this application does not impose any limitation. In the embodiments of this application, for a technical feature, the technical features in that technical feature are distinguished by "first", "second", "third", "A", "B", "C" and "D", and there is no order or size among the technical features described by "first", "second", "third", "A", "B", "C" and "D".

[0058] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this application can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this application can be achieved, and this is not limited herein.

[0059] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for online real-time monitoring and early warning of a relay protection secondary circuit hidden disconnection, characterized in that, This includes the system initialization phase and the online operation phase; The system initialization phase includes the following steps: Step 1: Collect historical multi-dimensional state data, construct a one-dimensional initial feature vector, establish the equivalent circuit equation of the secondary loop based on the implicit disconnection physical mechanism, and generate simulated fault waveform data by jointly sampling the disconnection time and transition resistance parameters. The simulated fault waveform data is then superimposed with the one-dimensional initial feature vector to complete data augmentation. Step 2: Adaptive feature selection is performed on the enhanced dataset to obtain the key fault feature sequence. The dataset is then resampled according to a fixed number of points within the power frequency cycle, and multi-domain features are extracted to construct a three-dimensional time-series feature tensor. Step 3: The customized residual attention neural network model constructed by inputting the three-dimensional temporal feature tensor is trained offline to generate a benchmark diagnostic model, a fault coding mapping dictionary, and feature mapping rules; The online operation phase includes: Step 4: Collect multi-dimensional status data during the operation of the secondary circuit of the relay protection in real time, and generate a real-time three-dimensional time-series feature tensor using the same normalization parameters and feature mapping rules as in the system initialization phase. Step 5: Input the real-time three-dimensional temporal feature tensor into the benchmark diagnostic model, and output the location result and fault code of the secondary circuit hidden open circuit fault; Step 6: Based on the positioning results, trigger the chain failure situation awareness module, use the hierarchical string fuzzy matching algorithm to perform real-time risk assessment based on the pre-built chain failure event chain template, use the fuzzy comprehensive evaluation method with time-series penalty factor to calculate the comprehensive risk level, introduce supplementary judgment criteria to determine the final risk level, and output early warning information to the application platform layer. Step 7: Based on the cross-entropy confidence assessment mechanism, the consistency between the early warning information and the on-site defect elimination results is judged, high confidence samples are selected, the upper limit of the training set window is set, and the benchmark diagnostic model is updated in a closed loop by combining the incremental training strategy based on the performance monitoring of the validation set.

2. The method for online real-time monitoring and early warning of hidden open circuits in relay protection secondary circuits according to claim 1, characterized in that, The disconnection time in step one obeys uniform distribution in a power frequency cycle; the transition resistance parameter obeys lognormal distribution, the mean value of the lognormal distribution is 200 , the standard deviation is 50 , the value range is 10 -1000 ; the joint sampling strategy accurately matches the contact failure physical transient characteristics of the implicit disconnection distinguished from the metallic disconnection; the specific formula of the data enhancement is: ; wherein is the enhanced feature vector, is the one-dimensional initial feature vector collected originally, is the fault waveform feature generated based on circuit equation simulation, is a random scaling factor with a value range of 0.1 to 0.

5. During the online operation phase, the simulation waveform feature dictionary generated during the system initialization phase is only used for table lookup during the online phase and does not perform electromagnetic transient simulation calculations.

3. The method according to claim 1, characterized in that, The adaptive feature selection described in step two specifically involves: Step 1: Obtain the real-time operating topology status of the power grid. When the power grid is in grid-connected operation mode, set the first significance level dynamic threshold. When the power grid is in islanded operation mode, set the second significance level dynamic threshold. Step 2: Use dynamic low variance filtering to extract the variance distribution probability density function of each feature under historical normal operating conditions, take the lower limit of the 95% confidence interval as the dynamic variance threshold, and screen out feature quantities whose sample variance is less than the dynamic variance threshold. Step 3: Use an adaptive chi-square test model to calculate the chi-square statistic and corresponding P-value between each feature retained after low variance filtering and the fault category label, and filter out features with P-values ​​greater than the dynamic significance level threshold corresponding to the current running topology.

4. The method according to claim 1, characterized in that, The three-dimensional temporal feature tensor reconstruction method described in step two is as follows: Step 1: Set the sampling rate of the original analog quantity to 4000Hz, and extract 80 time-series sampling points for the characteristic quantity in each channel within one power frequency cycle of 20ms; Step 2: Divide the key fault feature sequence into 12 channels: Channels 1 to 3 are the three-phase current and angle characteristics on the high-voltage side, Channels 4 to 6 are the three-phase current and angle characteristics on the low-voltage side, Channels 7 to 9 are the three-phase voltage and angle characteristics on the high-voltage side, and Channels 10 to 12 are the three-phase voltage and angle characteristics on the low-voltage side. Step 3: Extract three feature subspaces for the feature quantities in each channel: time domain features, frequency domain features, and wavelet packet energy features, and construct a three-dimensional time-series feature tensor with a size of 12×80×3.

5. The method according to claim 1, characterized in that, The customized residual attention neural network model described in step three includes an input layer, multiple customized attention residual blocks, an average pooling layer, and a fully connected layer; The customized attention residual block includes a parallel multi-scale convolution module, a first batch of normalized layers, a first activation function layer, a second batch of normalized layers, a convolution block attention module, and an identity mapping adder. The parallel multi-scale convolution module contains three parallel convolution branches, namely: A standard convolutional branch with a kernel size of 1×3 is used to extract the fundamental frequency features. A dilated convolutional branch with a kernel size of 1×5 and a dilation rate of 2 is used to extract low-frequency harmonic features; A deformable convolutional branch with a kernel size of 1×7 is used to adaptively adjust the sampling point weights to match the distortion start point features of the fault transient waveform; the outputs of the three branches are spliced ​​together in the channel dimension.

6. The method for online real-time monitoring and early warning of hidden open circuits in relay protection secondary circuits according to claim 1, characterized in that, Both the offline training described in step three and the closed-loop update described in step seven use the FocalLoss loss function, with the following formula: ; in This represents the model's predicted probability for the true class. The focus parameter is set to 2. The category balance weight parameter is calculated using the following formula: ; in Let be the number of samples in class t, C be the total number of classes, and satisfy the normalization constraint. During the closed-loop update process, the global mean and variance statistical parameters of all batch standardized layers in the customized residual attention neural network model are frozen.

7. The method for online real-time monitoring and early warning of hidden open circuits in relay protection secondary circuits according to claim 1, characterized in that, Step 3 describes the use of a 16-bit binary encoding rule to generate hexadecimal fault codes in the fault code mapping dictionary. Bits D15 to D13 indicate the major fault category, bits D12 to D11 indicate the fault type of the sampling circuit, bits D10 to D8 indicate the type of associated device, bits D7 to D6 indicate the link type, bits D5 to D3 indicate the fault phase sequence, bits D2 to D1 indicate the high-voltage side or low-voltage side, and bit D0 indicates the sampling circuit link break status. The location of the hidden open circuit fault in the secondary circuit described in step five is determined by parsing the fault code mapping dictionary to obtain the specific fault location.

8. The method for online real-time monitoring and early warning of hidden open circuits in relay protection secondary circuits according to claim 1, characterized in that, The hierarchical string fuzzy matching algorithm described in step six is ​​as follows: Step 1: Calculate the similarity between the real-time collected event strings and the template fields based on the edit distance algorithm; Step 2: Set a hierarchical matching similarity threshold. For critical events involving main protection actions and circuit breaker tripping, set a matching similarity threshold of 98%. For non-critical events involving abnormal alarms and backup protection actions, set a matching similarity threshold of 90%. Step 3: When the number of events matched according to the event sequence is equal to the total number of events in the event chain and each event meets the corresponding grade threshold, the match is considered successful.

9. The method for online real-time monitoring and early warning of hidden open circuits in relay protection secondary circuits according to claim 1, characterized in that, The specific process of calculating the comprehensive risk level using the fuzzy comprehensive evaluation method described in step six is ​​as follows: Step 1: Establish the factor set U= , , ; in, For the probability of cascading failures, The percentage decrease in substation availability. Loss of load rate With the comment set V=1, 2, 3; Among them, Level 1 is critical, Level 2 is serious, and Level 3 is moderate; Step 2: Construct the membership degree of each index using the triangular membership function to obtain the fuzzy relation matrix R; Step 3: Construct the time-series penalty factor ,in The attenuation coefficient is... The absolute value of the deviation between the actual event occurrence time and the time set in the event chain template is used to determine the timing penalty factor. The weights are corrected as a prerequisite for the calculation of the dynamic weight vector A; Step 4: Through fuzzy transformation Calculate the comprehensive evaluation vector A; Step 5: When the difference between the largest and second largest membership degree is less than 0.1, the weighted average principle is used to calculate the comprehensive risk value to determine the risk level. When the difference is greater than or equal to 0.1, the largest membership degree principle is used to determine the risk level.

10. The method for online real-time monitoring and early warning of hidden open circuits in relay protection secondary circuits according to claim 1, characterized in that, The closed-loop update process in step seven is as follows: Step 1: Calculate the cross-entropy between the actual fault result label and the model output probability distribution. Set the cross-entropy confidence threshold to 0.

2. When the cross-entropy is less than 0.2, it is judged as a high-confidence sample. Step 2: Introduce a sliding upper limit mechanism for the training set window, setting the maximum capacity of the training set to 10,000 samples. When a high-confidence sample is added to the training set, causing the total number to exceed 10,000, the oldest sample is eliminated according to the first-in-first-out principle. Step 3: After each incremental training, calculate the F1-score of the model using an independent validation set. If the F1-score of the updated model is lower than that before the update, trigger the model rollback mechanism to restore the parameters of the previous version of the model. Step 4: During incremental training, freeze the weight parameters of the parallel multi-scale convolutional modules in the customized residual attention neural network model, and only fine-tune and update the weight parameters of the convolutional block attention modules and fully connected layers.