Intelligent substation secondary circuit fault arc on-line monitoring method and device

Abstract

The application discloses a kind of intelligent substation secondary circuit fault arc online monitoring method and device, pulse optical signal is sent to the redundant optical fiber in the substation secondary circuit optical cable by optical time domain reflectometer, and Rayleigh backscattering signal is collected;Wavelet packet transform is carried out on scattering signal, the energy characteristics and entropy characteristics of each frequency band are calculated, and multidimensional feature vector is obtained;Non-dominated solution set is filtered by applying Pareto Front Optimization algorithm, and the Pareto optimal solution set of arc characteristics is obtained;The directed graph model of optical fiber sensing network is constructed, the arc candidate position is determined by applying directed minimum cut algorithm, and positioning is carried out by tree filling algorithm;Feature and position information are constructed into time sequence feature matrix, input into the CNN-LSTM hybrid network of Bayes online resource allocation optimization for classification identification;Locking decision strategy is generated based on fault diagnosis result, encapsulated as MMS message and sent to station control layer, and protection linkage is realized.

Description

Technical Field

[0001] This invention relates to the field of power system condition monitoring technology, specifically to the field of secondary circuit fault monitoring in intelligent substations, and particularly to a method and device for online monitoring of arc faults in secondary circuits of intelligent substations. Background Technology

[0002] Fault monitoring of secondary circuits in intelligent substations is a crucial aspect of ensuring the safe operation of power systems. As the nervous system of an intelligent substation, the secondary circuits bear the critical tasks of signal acquisition, transmission, and control execution. Their safe and stable operation directly impacts the reliability and safety of the entire substation.

[0003] Currently, fault monitoring in the secondary circuits of smart substations mainly employs two types of technologies: one is a monitoring method based on electrical parameters, which judges faults by measuring abnormalities in parameters such as voltage and current, but this type of method is not sensitive enough to small signal interference; the other is a temperature monitoring method, which monitors temperature abnormalities through technologies such as infrared thermography, but this type of method can only detect when the fault has developed to a severe stage.

[0004] The closest existing technology is the application of fiber optic sensing technology to power equipment condition monitoring. This technology uses specially designed optical fibers as distributed sensors to detect changes in physical quantities such as temperature and vibration around the power equipment. However, this technology requires the dedicated deployment of sensing optical fibers, increasing installation complexity; moreover, existing systems are mainly designed for primary equipment (such as transformers and GIS equipment), and the signal processing algorithms are not suitable for feature extraction of minute arc signals in secondary circuits, making it difficult to accurately identify weak arc faults in secondary circuits caused by insulation aging or poor contact.

[0005] The main problems with existing technologies include: first, they cannot utilize the existing optical cable resources of substations for monitoring, requiring the additional deployment of dedicated sensors, which results in high modification costs; second, the fault arc feature extraction and identification algorithms are not sensitive enough to weak arc signals, making it difficult to accurately distinguish fault arcs from normal vibration interference in complex environments, leading to a high false alarm rate. Summary of the Invention

[0006] The purpose of this invention is to solve the above-mentioned problems in the prior art and provide a method and device for online monitoring of secondary circuit fault arcs in intelligent substations. It utilizes the existing optical cable resources of the substation, eliminating the need for additional dedicated sensors, and achieves high-precision positioning and accurate identification of secondary circuit fault arcs through advanced signal processing and machine learning algorithms.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A method for online monitoring of arc faults in the secondary circuit of an intelligent substation involves transmitting pulsed optical signals to redundant optical fibers in the secondary circuit optical cable of the substation using an optical time-domain reflectometer (OTDR), collecting Rayleigh backscattered signals from the redundant fibers to obtain a scattered signal data stream; performing wavelet packet transform on the scattered signal data stream to decompose the signal into multiple frequency bands, calculating the energy and entropy characteristics of each frequency band to obtain a multidimensional feature vector; modeling feature selection as a multi-objective optimization problem based on the multidimensional feature vector, applying the Pareto front-end optimization algorithm to filter the non-dominated solution set, and obtaining the Pareto optimal solution set for the arc features; and constructing a directed graph of an optical fiber sensor network based on the Pareto optimal solution set. The model uses a directed minimum cut algorithm to determine candidate arc locations, and a tree-filling algorithm to locate them, obtaining arc location coordinates and location confidence. The Pareto optimal solution set, arc location coordinates, and location confidence are used to construct a time-series feature matrix, where the location confidence is used as a weighting coefficient for the location features. This matrix is ​​then weighted and input into a Bayesian online resource allocation optimization CNN-LSTM hybrid network for classification and recognition, yielding fault diagnosis results. Based on these results, the severity of the fault is assessed, and a blocking decision strategy is generated. This blocking decision strategy is encapsulated into an MMS message and sent to the station control layer to achieve protection linkage.

[0009] Optionally, the step of sending pulsed optical signals to redundant optical fibers in the secondary circuit optical cable of the substation via an optical time-domain reflectometer, acquiring Rayleigh backscattered signals in the redundant optical fibers, and obtaining a scattered signal data stream includes:

[0010] A narrow-linewidth laser is used to generate continuous light, which is then modulated into a pulsed light signal with a pulse width of 10-100ns by an acousto-optic modulator.

[0011] The pulsed optical signal is sent to the redundant optical fiber, and the Rayleigh backscattering signal in the redundant optical fiber is collected by a photodetector to obtain the original scattered signal.

[0012] The original scattered signal is digitally filtered and then averaged multiple times to obtain the scattered signal data stream.

[0013] Optionally, the step of performing wavelet packet transform on the scattered signal data stream to decompose the signal into multiple frequency bands, calculating the energy characteristics and entropy characteristics of each frequency band, and obtaining a multi-dimensional feature vector includes:

[0014] Applying db4 wavelet packet transform to the scattered signal data stream decomposes the signal into 5 layers, yielding wavelet packet coefficients in 32 frequency bands;

[0015] The energy value and Shannon entropy value are calculated for the wavelet packet coefficients of each frequency band to obtain the energy characteristics and entropy characteristics of each frequency band;

[0016] Calculate the rate of change of the energy feature and the entropy feature between adjacent time windows to obtain the rate of change of energy and the rate of change of entropy.

[0017] The energy feature, the entropy feature, the energy change rate, and the entropy change rate are combined to obtain the multidimensional feature vector.

[0018] Optionally, the step of modeling feature selection as a multi-objective optimization problem based on the multi-dimensional feature vector, and applying the Pareto front-end optimization algorithm to filter the non-dominated solution set to obtain the Pareto optimal solution set of the arc features includes:

[0019] The information gain and Fisher score are calculated for each feature in the multidimensional feature vector to obtain the feature importance scoring matrix;

[0020] Feature selection is modeled as a multi-objective optimization problem, with the optimization objectives defined as maximizing feature discriminative power and minimizing the number of features.

[0021] Based on the multidimensional feature vector, a candidate feature subset is generated using binary encoding, wherein each candidate feature subset is represented by a binary vector with a length equal to the dimension of the multidimensional feature vector;

[0022] Based on the feature importance scoring matrix, the Pareto front-end optimization algorithm is applied to traverse the candidate feature subsets and filter feature combinations that are not dominated by the optimization objective to obtain the Pareto optimal solution set.

[0023] Optionally, the step of constructing a directed graph model of the fiber optic sensor network based on the Pareto optimal solution set, applying the directed minimum cut algorithm to determine candidate arc positions, and locating the arc position using a tree-filling algorithm to obtain the arc position coordinates and location confidence includes:

[0024] The fiber optic sensor network is abstracted as a directed graph model, where nodes represent virtual sensing points on the fiber, edges represent the signal propagation relationship between adjacent sensing points, and edge weights are defined based on the correlation and intensity variation of the scattered signals.

[0025] In the directed graph model, a source point and a sink point are set, and the directed minimum cut algorithm is applied to calculate the minimum cut set from the source point to each candidate point. The point with the smallest cut capacity is determined as the candidate position of the electric arc.

[0026] A tree-like network structure is constructed based on the candidate locations of the electric arc. The signal strength residual of each node is calculated by applying the tree-filling algorithm, and the path with the minimum residual is determined to obtain the coordinates of the electric arc location and the location confidence.

[0027] Optionally, the step of constructing a directed graph model of the fiber optic sensor network based on the Pareto optimal solution set, applying the directed minimum cut algorithm to determine candidate arc positions, and locating the arc position using a tree-filling algorithm to obtain the arc position coordinates and location confidence includes:

[0028] Based on the Pareto optimal solution set, the anomaly score of each virtual sensing point on the optical fiber is calculated. The anomaly score is the weighted distance of each feature value from the mean of the normal state.

[0029] The virtual sensing points are filtered based on the anomaly score and adaptive threshold, and consecutive anomaly points are clustered to form candidate anomaly regions.

[0030] A local directed subgraph is constructed for the candidate anomaly region, and the directed minimum cut algorithm is applied to calculate the minimum cut set. The point with the smallest cut capacity is determined as the candidate arc location.

[0031] A tree-like network structure is constructed based on the candidate locations of the electric arc. A filling function is defined to represent the signal strength of each node. The filling function value of the internal node is calculated by weighted averaging of the child nodes.

[0032] Calculate the residual between the fill function value of each node and the measured signal, determine the path with the smallest residual, and perform interpolation positioning between adjacent sampling points to obtain the arc position coordinates and positioning confidence.

[0033] Optionally, the Pareto optimal solution set, the arc position coordinates, and the location confidence are constructed into a time-series feature matrix, wherein the location confidence is used as a weighting coefficient for the location features to weight the feature matrix, and the matrix is ​​then input into a Bayesian online resource allocation optimization CNN-LSTM hybrid network for classification and recognition to obtain fault diagnosis results, including:

[0034] Based on the features in the Pareto optimal solution set, the arc position coordinates, and the position confidence, a sliding window with a time window of 100ms is constructed to organize the continuous feature vectors into a time-series feature matrix.

[0035] The time series feature matrix is ​​standardized to obtain a standardized time series feature matrix;

[0036] The standardized temporal feature matrix is ​​input into a CNN network containing a multi-layer convolutional structure. Local feature patterns of the arc signal are extracted through convolution operations to obtain the CNN feature map.

[0037] The CNN feature map is input into a bidirectional LSTM network to capture the temporal dependencies of the arc signal and obtain a temporal feature representation.

[0038] The temporal feature representation is classified to obtain the fault diagnosis result, which includes arc intensity, location, and duration.

[0039] Optionally, the CNN-LSTM hybrid network is optimized through Bayesian online resource allocation, including:

[0040] Define a set of reusable resources in the CNN-LSTM hybrid network, which includes attention weights, inter-layer connection strengths, and feature channel activation strengths.

[0041] A value probability distribution model is established for each resource in the reusable resource set, and the parameters of the value probability distribution model include the prior mean and the prior variance.

[0042] In each identification task, based on the current input temporal feature matrix, the value estimate of each resource is sampled from the value probability distribution model based on Thompson sampling;

[0043] Based on the estimated value and resource usage cost, a resource allocation scheme is solved under budget constraints to determine the usage status of each resource. Based on the usage status of the resources, classification and identification are performed to obtain the current fault diagnosis result.

[0044] The loss function value is calculated based on the comparison between the current fault diagnosis result and the actual fault label. The loss function value is used as a feedback signal, and the parameters of the value probability distribution model are updated by applying the Bayesian update rule to realize the online optimization of the resource allocation strategy.

[0045] Optionally, the step of assessing the severity of the fault based on the fault diagnosis results and generating a blocking decision strategy, encapsulating the blocking decision strategy into an MMS message and sending it to the station control layer to achieve protection linkage, includes:

[0046] Based on the arc intensity, duration and location information in the fault diagnosis results, the potential risk value of the arc fault to the secondary system is calculated, and the fault severity assessment result is obtained.

[0047] Based on the severity assessment results of the fault, a multi-objective decision-making method is applied to find the balance between system security and power supply reliability, establish a blocking decision model that includes risk thresholds and decision rules, and generate a hierarchical blocking decision strategy.

[0048] The fault diagnosis results and the hierarchical interlocking decision strategy are encapsulated into MMS messages according to the IEC61850 standard, and the message priority and target address are configured.

[0049] The MMS message is sent to the protection device through the station control layer network. The protection device executes the corresponding protection action according to the hierarchical blocking decision strategy and returns the execution response result.

[0050] The effectiveness of the hierarchical locking decision strategy is evaluated based on the execution response results and subsequent monitoring data. The risk threshold and decision rules in the locking decision model are adjusted according to the evaluation results to achieve adaptive optimization of the locking decision.

[0051] Optionally, before transmitting the pulsed optical signal to the redundant optical fiber in the secondary circuit optical cable of the substation via the optical time domain reflectometer, the method further includes:

[0052] The redundant fiber cores in the secondary circuit optical cable of the substation are fused together with the optical time domain reflectometer through the optical cable splice box to form a sensing optical fiber network.

[0053] The pulse repetition frequency of the optical time domain reflectometer is configured to be 1-10kHz, and the sampling rate is above 100MHz.

[0054] This application provides an online monitoring device for arc faults in the secondary circuit of an intelligent substation, comprising:

[0055] The fiber optic scattering signal acquisition module is used to send pulsed optical signals to the redundant optical fibers in the secondary circuit optical cable of the substation via an optical time domain reflectometer, acquire the Rayleigh backscattering signals in the redundant optical fibers, and obtain the scattering signal data stream.

[0056] The feature extraction module is used to perform wavelet packet transform on the scattered signal data stream, decompose the signal into multiple frequency bands, calculate the energy features and entropy features of each frequency band, and obtain a multi-dimensional feature vector.

[0057] The Pareto optimization module is used to model feature selection as a multi-objective optimization problem based on the multi-dimensional feature vector, and apply the Pareto front-end optimization algorithm to filter the non-dominated solution set to obtain the Pareto optimal solution set of the arc feature.

[0058] The arc positioning module is used to construct a directed graph model of the fiber optic sensor network based on the Pareto optimal solution set, apply the directed minimum cut algorithm to determine the candidate position of the arc, and perform positioning through the tree filling algorithm to obtain the arc position coordinates and positioning confidence.

[0059] The fault identification module is used to construct a time-series feature matrix from the Pareto optimal solution set, the arc position coordinates, and the position confidence, wherein the position confidence is used as a weight coefficient of the position feature to weight the feature matrix, and input it into a CNN-LSTM hybrid network with Bayesian online resource allocation optimization for classification and identification, so as to obtain the fault diagnosis result.

[0060] The interlocking decision and communication module is used to assess the severity of the fault based on the fault diagnosis results and generate an interlocking decision strategy. The interlocking decision strategy is then encapsulated into an MMS message and sent to the station control layer to achieve protection linkage.

[0061] The beneficial effects of this invention include:

[0062] First, by utilizing the redundant fiber cores of existing optical cables in substations as distributed sensing media, there is no need to lay additional dedicated sensors, reducing system deployment costs and achieving a "plug-and-play" monitoring solution.

[0063] Second, Pareto front-end optimization theory is applied to arc feature extraction to construct a multi-objective optimization framework, seeking the optimal balance between feature discriminative power and computational complexity to improve feature selection efficiency.

[0064] Third, by applying a near-linear time directed minimum cut algorithm and a tree-fill approximation algorithm, the location of the electric arc is quickly and accurately located with a spatial resolution better than 0.5 meters.

[0065] Fourth, it innovatively combines Bayesian online reusable resource allocation theory with CNN-LSTM hybrid networks to optimize the allocation of computational resources for deep learning models and improve the accuracy and computational efficiency of arc feature recognition.

[0066] Fifth, an adaptive blocking decision mechanism based on a discard strategy is designed to seek the best balance between system security and power supply reliability, and to achieve precise protection against secondary circuit faults.

[0067] Sixth, the optical time domain reflectance technology is deeply integrated with the IEC61850 standard, and seamless connection with other intelligent devices in the substation is achieved through MMS messages, thus constructing a complete secondary circuit safety protection system. Attached Figure Description

[0068] Figure 1 This is a flowchart illustrating the overall technical solution of the present invention.

[0069] Figure 2 Flowchart for fiber optic scattering signal acquisition;

[0070] Figure 3 This is a schematic diagram of wavelet packet decomposition and multidimensional feature extraction.

[0071] Figure 4 Flowchart for optimizing feature selection for the Pareto front end;

[0072] Figure 5 A schematic diagram of the arc positioning process;

[0073] Figure 6 This is a schematic diagram of the fault diagnosis and identification process;

[0074] Figure 7 This is a structural diagram of an online monitoring device for arc faults in the secondary circuit of an intelligent substation. Detailed Implementation

[0075] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0076] Example 1

[0077] like Figure 1 As shown, this embodiment provides an online monitoring method for arc faults in the secondary circuit of an intelligent substation. A pulsed optical signal is sent to redundant optical fibers in the secondary circuit optical cable of the substation using an optical time-domain reflectometer (OTDR). Rayleigh backscattered signals from the redundant fibers are collected to obtain a scattered signal data stream. Wavelet packet transform is performed on the scattered signal data stream to decompose the signal into multiple frequency bands. The energy and entropy characteristics of each frequency band are calculated to obtain a multi-dimensional feature vector. Based on the multi-dimensional feature vector, feature selection is modeled as a multi-objective optimization problem. The Pareto front-end optimization algorithm is applied to filter the non-dominated solution set to obtain the Pareto optimal solution set for the arc features. An optical fiber sensor network is constructed based on the Pareto optimal solution set. A directed graph model is used, and the directed minimum cut algorithm is applied to determine the candidate positions of the electric arc. The position is then located using a tree-filling algorithm to obtain the arc position coordinates and position confidence. The Pareto optimal solution set, the arc position coordinates, and the position confidence are used to construct a time-series feature matrix. The position confidence is used as a weighting coefficient for the position features to weight the feature matrix. The matrix is ​​then input into a CNN-LSTM hybrid network with Bayesian online resource allocation optimization for classification and recognition to obtain fault diagnosis results. Based on the fault diagnosis results, the severity of the fault is assessed and a blocking decision strategy is generated. The blocking decision strategy is encapsulated into an MMS message and sent to the station control layer to achieve protection linkage.

[0078] S1: Acquisition of fiber optic scattering signals;

[0079] This step involves sending pulsed optical signals to the redundant optical fibers in the secondary circuit optical cable of the substation using an optical time domain reflectometer, collecting the Rayleigh backscattered signals in the redundant optical fibers, and obtaining the scattered signal data stream.

[0080] Specifically, redundant fiber cores in the substation's secondary circuit optical cable are first fused together with an optical time domain reflectometer (e.g., a phase-sensitive optical time domain reflectometer φ-OTDR) through an optical cable splice box, forming a ring or bus topology sensing fiber network as a continuous distributed sensing medium. This method of utilizing existing redundant fiber cores avoids the cost and construction difficulty of laying additional dedicated sensing fibers, enabling rapid deployment of the monitoring system.

[0081] Continuous light is generated using a narrow-linewidth laser with a linewidth of 0.1 nanometers or less. This continuous light is then modulated into a high-energy, narrow-pulse optical signal with a pulse width of ten to one hundred nanoseconds using an acousto-optic modulator. The pulse repetition frequency is set to 1-10 kilohertz to ensure sufficient spatial resolution and sensing distance. As the narrow-pulse optical signal propagates in the optical fiber, Rayleigh backscattering occurs at each location within the fiber. The intensity and phase of the scattered light carry information about the local state of the fiber.

[0082] A high-sensitivity photodetector and a high-speed analog-to-digital converter are used to collect Rayleigh backscattered light signals from optical fibers. The sampling rate is set to over 100 MHz to achieve high-speed acquisition of the scattered light signals. When an arc fault occurs in the secondary circuit, the high temperature generated by the arc causes a slight change in the refractive index of the nearby optical fiber, which in turn alters the characteristics of the Rayleigh scattering signal. This change can be captured by a high-precision photodetector system.

[0083] The acquired scattered signals undergo digital filtering and signal enhancement to eliminate environmental noise and system errors. Multiple accumulation and averaging operations are used to improve the signal-to-noise ratio, resulting in a high-quality scattered signal data stream. The digital filtering employs a bandpass filter to remove low-frequency drift and high-frequency random noise while preserving the characteristic frequency bands of the arc signal. Multiple accumulation and averaging operations are typically performed ten to fifty times to achieve a balance between improving the signal-to-noise ratio and maintaining real-time performance.

[0084] S2: Multidimensional feature vector extraction;

[0085] This step performs wavelet packet transform on the scattered signal data stream, decomposes the signal into multiple frequency bands, calculates the energy characteristics and entropy characteristics of each frequency band, and obtains a multidimensional feature vector.

[0086] The high-quality scattered signal data stream obtained in step S1 is subjected to db4 wavelet packet transform to decompose the signal into five layers, obtaining wavelet packet coefficients of 32 different frequency bands, forming the time-frequency representation of the signal. The db4 wavelet is chosen as the basis function because it has good time-frequency localization characteristics and can effectively capture transient changes and frequency features in the arc signal.

[0087] Wavelet packet transform is an extension of the traditional discrete wavelet transform. It not only decomposes the low-frequency part of the signal but also recursively decomposes the high-frequency part, thus providing a more refined frequency band division. When an electric arc occurs in an optical fiber, the high temperature generated by the arc causes a sudden change in the local refractive index of the fiber. This physical change leaves specific time-frequency characteristics in the scattered signal. Decomposing the signal to the fifth layer means that the original signal is divided into 32 frequency bands. This fine frequency band division can capture the characteristic frequencies of different types of electric arcs, such as micro-arcs, flashover arcs, and continuous arcs.

[0088] For each frequency band, the energy distribution and Shannon entropy are calculated using the wavelet packet coefficients, constructing a multidimensional feature vector containing energy features, entropy features, energy change rate, and entropy change rate. The energy feature represents the degree of energy concentration of the signal within that frequency band and is calculated by summing the squares of the wavelet packet coefficients. The Shannon entropy is used to quantify the complexity and uncertainty of the signal within each frequency band, reflecting the degree of disorder in the signal.

[0089] In addition to static features, energy change rate and entropy change rate are also calculated as dynamic features. These rates are obtained by dividing the difference between the energy and entropy features between adjacent time windows by the time interval, typically ten milliseconds. These dynamic features capture the temporal evolution of the arc signal, helping to distinguish the arc from other interference signals.

[0090] Finally, a vector containing 128 features is constructed, with 32 frequency bands and four features per band: energy, entropy, rate of energy change, and rate of entropy change, fully characterizing the energy and entropy characteristics of the arc signal. This feature combination can effectively distinguish the arc from other interference signals. In particular, the arc signal usually exhibits abrupt energy changes but relatively stable entropy, while mechanical vibration exhibits the characteristic of simultaneous changes in energy and entropy.

[0091] S3: Obtaining the Pareto optimal solution set;

[0092] Based on the aforementioned multidimensional feature vector, this step models feature selection as a multi-objective optimization problem, applies the Pareto front-end optimization algorithm to filter the non-dominated solution set, and obtains the Pareto optimal solution set for the arc feature.

[0093] First, the information gain and Fisher score are calculated for each feature in the multidimensional feature vector obtained in step S2 to obtain the feature importance scoring matrix. Information gain measures the contribution of a feature to the classification result and is obtained by calculating the difference between the entropy of the class variable and the conditional entropy of the class variable under known feature conditions. The Fisher score evaluates the ratio of inter-class separation to intra-class clustering of a feature and is obtained by calculating the ratio of the weighted sum of the differences in feature means between different classes to the weighted sum of the variances within each class.

[0094] By combining these two scoring methods, a two-dimensional score vector is constructed for each feature, creating a feature importance score matrix. This score matrix includes the performance of each feature across both evaluation dimensions, providing a quantitative basis for the subsequent Pareto optimization. Compared to traditional single-criteria feature evaluation methods, this multi-criteria evaluation can more comprehensively reflect the discriminative power of features.

[0095] Feature selection is modeled as a multi-objective optimization problem, defining maximizing feature discriminative power and minimizing the number of features as optimization objectives. The first objective is to maximize the discriminative power of the feature set, calculated by weighting the information gain and Fisher scores of the selected features. The second objective is to minimize the size of the feature set, i.e., the number of selected features. These two objectives are often conflicting; selecting more features can improve discriminative power but increases computational complexity.

[0096] Based on the multidimensional feature vector, candidate feature subsets are generated using binary encoding. Each candidate feature subset is represented by a binary vector with a length equal to the dimension of the multidimensional feature vector. Each element in the binary vector corresponds to a feature; a value of 1 indicates that the feature is selected, and a value of 0 indicates that it is not selected.

[0097] Based on the feature importance scoring matrix, the Pareto front-end optimization algorithm is applied to traverse the candidate feature subsets, filtering for feature combinations that are not dominated by any of the optimization objectives, thus obtaining the Pareto optimal solution set. In the Pareto front-end, no solution can be better than any other solution simultaneously on all objectives. The algorithm initializes an empty solution set, then traverses the candidate feature subset space. For each candidate solution, if no other solution is better than or equal to it on all objectives and strictly better than it on at least one objective, then the solution is added to the solution set, and all solutions dominated by it are removed.

[0098] To further optimize the solution set, the theory of tight universal boundaries of convex hulls is applied. This theory states that the Pareto front of a multi-objective optimization problem always lies on or inside the boundary of the convex hull in the objective space. By calculating the convex hull in the objective space, the Pareto front can be determined more efficiently.

[0099] Finally, an optimal solution balancing discriminative power and feature quantity is selected from the Pareto non-dominated solution set. The corresponding feature subset is the Pareto optimal solution set for the arc features. This feature subset typically contains twenty to thirty key features, maintaining high discriminative power and low computational complexity, providing high-quality feature input for subsequent arc localization and identification.

[0100] S4: Arc position positioning;

[0101] This step constructs a directed graph model of the fiber optic sensor network based on the Pareto optimal solution set, applies the directed minimum cut algorithm to determine the candidate positions of the electric arc, and uses the tree filling algorithm to locate the position, thereby obtaining the coordinates of the electric arc position and the location confidence.

[0102] The fiber optic sensor network is abstracted as a directed graph model, where nodes represent virtual sensing points on the fiber, edges represent the signal propagation relationships between adjacent sensing points, and edge weights are defined based on the correlation and intensity variations of the scattered signals. The vertex set represents several virtual sensing points distributed along the fiber, each point corresponding to a sampling location on the fiber. The physical distance between adjacent points is typically 0.5 meters, determined by the spatial resolution of the optical time-domain reflectometer system. The edge set represents the signal propagation relationships between the virtual sensing points.

[0103] For each edge, its weight is defined as a weighted combination of the similarity of the scattered signals and the physical distance. The similarity of the scattered signals is obtained by calculating the cosine similarity or Pearson correlation coefficient of the scattered signals at two sensing points, and the physical distance is the actual distance between the two sensing points. By adjusting the tradeoff factor, the influence of similarity and distance in the weight calculation can be balanced.

[0104] A source and sink node are introduced into the directed graph model. The source node is connected to all sensor points, and the sink node is also connected to all sensor points. This source-sink graph structure allows for the application of the minimum cut algorithm for arc localization. The edge weight from the source node to a sensor point is defined as the anomaly score for that point, calculated based on the arc features output in step S3. The anomaly score is obtained by calculating the weighted sum of the standardized distances of each feature value at the sensor point from the mean of the normal state. The feature weights are determined by Pareto optimization in step S3. The edge weight from the sensor point to the sink node is set to a fixed constant, representing the cost of identifying that point as a non-arc point.

[0105] A directed minimum cut algorithm is applied to calculate the minimum cut set from the source point to each candidate point, and the point with the smallest cut capacity is determined as the candidate arc location. An improved Karger algorithm is employed, which has a near-linear time complexity. For each candidate region, a local subgraph is constructed, each vertex is initialized into a separate group, and then edges are iteratively randomly selected and the groups containing the vertices are merged until only two groups remain, one containing the source point and the other containing the sink point. The cut capacity is calculated, and after repeating this process multiple times, the cut with the smallest capacity is selected as the result.

[0106] To improve the stability of the algorithm, a dynamic adjustment mechanism for edge weights is introduced. The edge weights in the graph are dynamically updated based on the temporal evolution pattern of the arc characteristics, ensuring that the edge weights reflect the correlation of feature changes. This dynamic adjustment can better capture the spatiotemporal evolution characteristics of the arc signal.

[0107] The minimum cut result divides the graph into two parts: one part contains the abnormal region corresponding to the source point, and the other part contains the normal region corresponding to the sink point. The vertices connected to the abnormal region by the cut edge set are the possible arc occurrence points, among which the point with the highest abnormality score is initially determined as the most likely location of the arc occurrence.

[0108] A tree-like network structure is constructed based on the candidate arc locations. A tree-filling algorithm is applied to calculate the signal strength residuals of each node, determining the path with the minimum residuals to obtain the arc location coordinates and location confidence. The network structure near the candidate arc point is simplified into a tree, where the root node corresponds to the candidate arc point and the leaf nodes correspond to surrounding sensor points. The tree construction is based on signal propagation paths and correlations. Starting from the candidate arc point, the main direction of signal propagation is determined based on signal correlation and time delay, and the tree structure is recursively constructed along this main direction.

[0109] A fill function is defined on the tree to represent the signal strength of each node. For leaf nodes, the fill function value is directly determined by the measured signal strength. For internal nodes, the fill function value is calculated by weighted averaging of its child nodes. The residual between the fill function value and the measured signal for each node is calculated, and the path with the smallest residual is found, as the nodes on this path are most likely to contain the arc occurrence point.

[0110] To further improve positioning accuracy, an interpolation algorithm is applied along the determined path to calculate the possible locations of sub-sampling points. Several virtual points are inserted between adjacent sampling points, and the signal strength at the interpolation points is estimated using linear interpolation. By analyzing the signal characteristics of the interpolation points, the positioning accuracy can be improved to five to ten centimeters.

[0111] Finally, the precise distance coordinates are combined with the wiring diagram of the substation's secondary circuit to convert them into actual physical location information, and the confidence level of the location is calculated. The confidence level is obtained by taking an exponential function over the minimum residual value; a higher confidence level indicates a more reliable location result. The output results include the fiber optic distance coordinates of the arc location, the corresponding physical location description, and the location confidence level.

[0112] S5: Fault diagnosis and identification;

[0113] This step constructs a time-series feature matrix from the Pareto optimal solution set, the arc position coordinates, and the position confidence. The position confidence is used as a weighting coefficient for the position features to weight the feature matrix. The matrix is ​​then input into a CNN-LSTM hybrid network with Bayesian online resource allocation optimization for classification and recognition, resulting in a fault diagnosis.

[0114] Based on the features in the Pareto optimal solution set, the arc position coordinates, and the position confidence, a sliding window with a time window of 100 milliseconds is constructed, and the continuous feature vectors are organized into a time-series feature matrix. For each time point, several sampling points within the first 100 milliseconds are extracted. For a sampling rate of 10 kHz, the number of sampling points is 1000, forming a feature time-series matrix.

[0115] To incorporate location information, the location features and confidence scores are expanded into vectors of the same length as the time series matrix and concatenated with the original feature matrix. This results in a time series feature matrix containing complete time series features and location information. The feature matrix is ​​then standardized to obtain a standardized time series feature matrix, ensuring that all feature values ​​fall within a similar numerical range. Standardization is achieved by subtracting the mean of each feature from the training set and dividing by the standard deviation.

[0116] To enhance the model's robustness to noise and variation, data augmentation strategies were implemented, including random shifts on the time axis, Gaussian noise addition, and random feature masking. Random shifts on the time axis were performed within a range of ±10 milliseconds, the signal-to-noise ratio of the Gaussian noise was set to 20 to 30 decibels, and random feature masking randomly set one to three feature values ​​to zero.

[0117] The standardized temporal feature matrix is ​​input into a CNN network containing multiple convolutional layers. Local feature patterns of the arc signal are extracted through convolutional operations to obtain the CNN feature map. The CNN feature extractor consists of three convolutional layers, each using an increasing number of convolutional kernels: 16, 32, and 64 kernels respectively. The kernel size is 3x3, and the stride is 1. After each convolutional layer, ReLU activation and batch normalization are applied, followed by max pooling with a 2x2 pooling window and a stride of 2.

[0118] To enhance the model's adaptability to different types of arc signals, a residual connection is introduced between the second and third layers. The output of the first layer is upsampled and then added to the output of the second layer. This residual structure helps the network retain low-level features from earlier layers, preventing the gradient vanishing problem.

[0119] To capture features at different time scales, a multi-scale convolution module was introduced. Different sized convolutional kernels, including 3x3, 5x5, and 7x7, were used in parallel in the third convolutional layer, and the results were stitched together after dimensionality reduction via 1x1 convolution. This design enables the network to capture both short-term and long-term arc features simultaneously.

[0120] Finally, global average pooling is applied to convert the final feature map into a fixed-dimensional feature vector, typically 64-dimensional, which represents the high-level feature abstraction extracted by the CNN from the temporal feature matrix.

[0121] The CNN feature maps are input into a bidirectional LSTM network to capture the temporal dependencies of the arc signal, resulting in a temporal feature representation. The bidirectional LSTM network contains 128 LSTM units; its bidirectional structure allows the network to consider both past and future information simultaneously. Each LSTM unit controls the information flow through forget gates, input gates, and output gates. The forget gate determines which information is discarded from the unit state, the input gate determines which new information is stored in the unit state, and the output gate determines which information is output.

[0122] In a bidirectional LSTM, the input sequence is processed twice, in both the forward and reverse directions, resulting in two sets of hidden states. The final hidden state is a connection between the two directions. To enhance the LSTM's ability to model the long-term dependence of electric arc signals, a residual LSTM structure is introduced. Residual connections are introduced into the LSTM layers to help the gradient propagate over long sequences.

[0123] A self-attention mechanism is introduced, enabling the network to dynamically focus on important parts of the sequence. Attention weights are obtained by calculating the energy score of the hidden state and normalizing it using the softmax function. The context vector is calculated by weighting the hidden state with attention weights, capturing the most relevant information in the sequence.

[0124] A hierarchical LSTM structure is adopted, stacking two LSTM layers. The first layer captures the basic temporal pattern, and the second layer models more advanced state transitions. Finally, the output of the LSTM network is the hidden state of the entire sequence. Each hidden state is a 256-dimensional vector containing connections between 128 forward LSTM units and 128 backward LSTM units.

[0125] The temporal feature representation is classified to obtain the fault diagnosis result containing arc intensity, location, and duration. The classification process is implemented through a fully connected layer and a softmax activation function, outputting the arc category probability distribution, including different types such as micro-arcs, flashover arcs, and continuous arcs, as well as information such as arc intensity level, precise location, and duration.

[0126] The CNN-LSTM hybrid network is optimized through Bayesian online resource allocation. A set of reusable resources is defined in the CNN-LSTM hybrid network, comprising attention weights, inter-layer connection strengths, and feature channel activation strengths. A value probability distribution model is established for each resource in the reusable resource set, with parameters including prior mean and prior variance.

[0127] In each identification task, based on the temporal feature matrix of the current input, the value estimate of each resource is sampled from the value probability distribution model using Thompson sampling. Thompson sampling is a Bayesian optimization method that balances exploration and utilization by sampling from the posterior distribution.

[0128] Based on the estimated value and resource usage costs, a resource allocation scheme is solved under budget constraints to determine the usage status of each resource. Classification and identification are then performed based on the resource usage status to obtain the current fault diagnosis result. The resource allocation problem is modeled as a knapsack problem, with the goal of maximizing the total value under budget constraints.

[0129] The loss function value is calculated based on the comparison between the current fault diagnosis result and the actual fault label. This loss function value is then used as a feedback signal to update the parameters of the value probability distribution model using a Bayesian update rule, thereby achieving online optimization of the resource allocation strategy. The Bayesian update rule updates the prior distribution based on the observed rewards, making the posterior distribution more accurately reflect the true value of the resources.

[0130] Through this Bayesian online resource allocation optimization, the CNN-LSTM hybrid network can dynamically adjust the use of computing resources, efficiently switch between different types of arc recognition tasks, and improve recognition accuracy and computational efficiency.

[0131] S6: Lockout decision-making and protection linkage;

[0132] This step assesses the severity of the fault based on the fault diagnosis results and generates a blocking decision strategy. The blocking decision strategy is then encapsulated into an MMS message and sent to the station control layer to achieve protection linkage.

[0133] Based on the arc intensity, duration, and location information from the fault diagnosis results, the potential risk value of the arc fault to the secondary system is calculated, yielding a fault severity assessment result. The potential risk value comprehensively considers factors such as the arc's energy, duration, and the importance of its location, and is calculated using weighted summation or fuzzy inference methods. Fault severity is categorized into four levels: minor, moderate, severe, and extremely severe.

[0134] Based on the severity assessment results of the faults, a multi-objective decision-making method is applied to find the balance point between system security and power supply reliability. A locking decision model incorporating risk thresholds and decision rules is established, generating a hierarchical locking decision strategy. The multi-objective decision-making method considers the two conflicting objectives of system security and power supply reliability, seeking the optimal balance point through Pareto optimization or weighted summation methods.

[0135] The interlocking decision model includes four levels of decision strategies: observation level, corresponding to minor faults, which only records information and does not take action; alarm level, corresponding to moderate faults, which issues alarm signals to remind maintenance personnel; mild interlocking level, corresponding to severe faults, which selectively interlocks non-critical protection functions; and severe interlocking level, corresponding to extremely severe faults, which interlocks all protection functions in the relevant area and activates the emergency plan.

[0136] The fault diagnosis results and the hierarchical interlocking decision strategy are encapsulated into MMS messages according to the IEC 61850 standard, with message priority and target address configured. MMS stands for Manufacturing Message Specification, a communication protocol defined in the IEC 61850 standard. Message encapsulation includes defining data objects, assigning attribute values, and selecting service types. Message priority is set according to the severity of the fault, with messages for extremely severe faults having the highest priority. The target address includes the network address of the station control layer monitoring system and related protection devices.

[0137] The MMS message is sent to the protection device via the station control layer network. The protection device executes the corresponding protection action according to the hierarchical interlocking decision strategy and returns the execution response result. The station control layer network typically uses Ethernet technology, supporting high-speed and reliable data transmission. After receiving the MMS message, the protection device parses the message content, executes the corresponding protection action according to the interlocking decision strategy, such as interlocking specific protection functions, disconnecting faulty circuits, and activating backup equipment, and encapsulates the execution result into an MMS message and returns it.

[0138] The effectiveness of the hierarchical interlocking decision-making strategy is evaluated based on the execution response results and subsequent monitoring data. The risk threshold and decision rules in the interlocking decision-making model are adjusted according to the evaluation results to achieve adaptive optimization of the interlocking decision. The effectiveness evaluation compares the actual fault development with the expected results to calculate the accuracy and timeliness of the decision. If the decision is found to be inaccurate or untimely, the parameters of the interlocking decision-making model are adjusted, such as lowering the risk threshold to improve sensitivity or modifying the decision rules to adapt to new fault modes.

[0139] Furthermore, the discard strategy refers to the strategy of selectively retaining or discarding certain protective functions during the interlocking decision-making process, based on the severity of the fault and the system operating status. Specifically, this includes:

[0140] First, discard non-critical protection functions: For minor faults, discard the blocking signals of non-critical protection functions, issue only an alarm, and maintain power supply continuity.

[0141] Second, selective discarding of redundant protection: When multiple protection redundancies exist, protection blocking signals far from the fault point are selectively discarded based on the fault location.

[0142] Third, timing discard strategy: For intermittent faults, a time window is set to discard repeated blocking signals within the window to avoid frequent operation of the protection device.

[0143] The goal of the drop-off strategy is to minimize unnecessary protection actions and improve power supply reliability while ensuring system safety.

[0144] Through this adaptive interlocking decision-making mechanism, the system can maintain power supply reliability to the maximum extent while ensuring safety, and achieve precise protection against secondary circuit faults.

[0145] This embodiment utilizes existing redundant fiber cores in the substation's optical cable as a distributed sensing medium, eliminating the need for additional dedicated sensors and reducing system deployment costs. By applying the Pareto front-end optimization algorithm for feature selection, an optimal balance is achieved between feature discriminative power and computational complexity. The application of the directed minimum cut algorithm and the tree-filling algorithm enables rapid and accurate arc location. Combining Bayesian online resource allocation theory with a CNN-LSTM hybrid network improves the accuracy and computational efficiency of arc feature recognition. An adaptive interlocking decision mechanism is designed to achieve precise protection against secondary circuit faults.

[0146] Example 2

[0147] This embodiment, based on Embodiment 1, further details the specific implementation method of fiber optic scattering signal acquisition.

[0148] like Figure 2 As shown, the step of sending pulsed optical signals to redundant optical fibers in the secondary circuit optical cable of the substation via an optical time-domain reflectometer, collecting Rayleigh backscattered signals in the redundant optical fibers, and obtaining a scattered signal data stream includes:

[0149] S2.1: Pulse light signal generation;

[0150] A narrow-linewidth laser is used to generate continuous light, which is then modulated into a pulsed light signal with a pulse width of 10 to 100 nanoseconds by an acousto-optic modulator.

[0151] Narrow-linewidth lasers employ distributed feedback laser diodes with a linewidth of less than 0.1 nanometers. The center wavelength is typically chosen to be 1550 nanometers, the standard wavelength for fiber optic communication, offering low loss and low dispersion. The continuous optical power output of the laser is typically 10 to 50 milliwatts, ensuring sufficient signal strength.

[0152] An acousto-optic modulator is a device that modulates light waves using the acousto-optic effect. It generates sound waves by applying a radio frequency signal to a crystal, which causes periodic changes in the crystal's refractive index, thus modulating the passing light waves. The frequency of the modulator's radio frequency drive signal is typically tens to hundreds of megahertz, and the modulation depth can reach over 90%.

[0153] By precisely controlling the drive signal of the acousto-optic modulator, continuous light can be modulated into narrow pulse optical signals with pulse widths ranging from 10 to 100 nanoseconds. The choice of pulse width requires a trade-off between spatial resolution and signal strength; narrower pulses offer higher spatial resolution but lower signal energy, while wider pulses have higher signal energy but lower spatial resolution. For substation secondary circuit monitoring, a pulse width of 50 nanoseconds is typically chosen, corresponding to a spatial resolution of approximately five meters.

[0154] S2.2: Scattering signal acquisition;

[0155] The pulsed optical signal is sent to the redundant optical fiber, and the Rayleigh backscattering signal in the redundant optical fiber is collected by a photodetector to obtain the original scattered signal.

[0156] Pulsed optical signals are injected into redundant optical fibers through an optical circulator or optical coupler. An optical circulator is a non-reciprocal optical device that allows optical signals to propagate in a specific direction, guiding the incident light into the fiber while simultaneously guiding the backscattered light in the fiber into the detector, thus achieving separation of transmission and reception.

[0157] When pulsed light propagates in an optical fiber, Rayleigh backscattering occurs at each location due to density fluctuations in the fiber material. The intensity of the scattered light is related to the local properties of the fiber; when the fiber is subjected to external disturbances such as temperature, strain, or vibration, the intensity and phase of the scattered light change. The high temperature generated by an electric arc can cause a local temperature increase in the fiber, changing its refractive index and thus altering the characteristics of the scattered light.

[0158] Backscattered light signals are acquired using a high-sensitivity photodetector. The photodetector employs an avalanche photodiode or photomultiplier tube, exhibiting high sensitivity and fast response characteristics. The detector's response time is typically less than one nanosecond, enabling it to capture rapidly changing scattered signals. The detector's output current signal is converted into a voltage signal by a transimpedance amplifier and then digitized by a high-speed analog-to-digital converter.

[0159] The sampling rate of the high-speed analog-to-digital converter (ADC) is set to 100 MHz or higher, typically 250 MHz or 500 MHz, to ensure accurate sampling of the scattered signal waveform. The ADC resolution is usually 12-bit or 14-bit, providing sufficient dynamic range. The acquired digital signal is the original scattered signal.

[0160] S2.3: Signal preprocessing;

[0161] The original scattered signal is digitally filtered and then averaged multiple times to obtain the scattered signal data stream.

[0162] Digital filtering employs finite impulse response (FIR) or infinite impulse response (IR) filters to remove low-frequency drift and high-frequency random noise. Low-frequency drift primarily originates from laser power fluctuations and ambient temperature variations, and is removed by a high-pass filter, with a cutoff frequency typically set to several kilohertz. High-frequency random noise mainly originates from detector thermal noise and amplifier noise, and is removed by a low-pass filter, with a cutoff frequency typically set to tens of megahertz.

[0163] Multiple accumulation and averaging is an effective method to improve the signal-to-noise ratio (SNR). For each spatial location, multiple scattered signals are acquired and averaged. Random noise cancels out each other, while the useful signal is accumulated and enhanced. The number of accumulations is typically ten to fifty, striking a balance between improving the SNR and maintaining real-time performance. The more accumulations, the more significant the improvement in SNR, but the lower the data update rate.

[0164] After digital filtering and multiple cumulative averaging processes, a high-quality scattered signal data stream is obtained, which serves as the input data for subsequent feature extraction and arc identification.

[0165] This embodiment provides a complete technical solution for the construction of an optical fiber scattering signal acquisition system by detailing the specific implementation methods of pulsed optical signal generation, scattering signal acquisition, and signal preprocessing.

[0166] Example 3

[0167] This embodiment, based on Embodiment 1, further details the specific implementation method of multidimensional feature vector extraction.

[0168] like Figure 3 As shown, the wavelet packet transform is performed on the scattered signal data stream to decompose the signal into multiple frequency bands. The energy characteristics and entropy characteristics of each frequency band are calculated to obtain a multi-dimensional feature vector, including:

[0169] S3.1: Wavelet packet decomposition;

[0170] The db4 wavelet packet transform was applied to the scattered signal data stream to decompose the signal into five layers, resulting in wavelet packet coefficients of 32 frequency bands.

[0171] The db4 wavelet is the fourth member of the Daubechies wavelet family. It has four vanishing moments and can effectively represent polynomial signals. The db4 wavelet has compact support properties and good localization performance in both the time and frequency domains, making it suitable for analyzing non-stationary signals.

[0172] Wavelet packet transform achieves multi-scale decomposition of a signal through a series of high-pass and low-pass filters. In the first decomposition level, the original signal is passed through a low-pass filter and a high-pass filter to obtain low-frequency and high-frequency components, respectively, and then each component is downsampled. In the second decomposition level, the low-frequency and high-frequency components obtained in the first level are again passed through low-pass and high-pass filters to obtain four frequency bands. This process is repeated recursively until the fifth level, resulting in 32 frequency bands.

[0173] Each frequency band corresponds to a specific frequency range, enabling the capture of the characteristics of electric arc signals in different frequency domains. The frequency characteristics of electric arc signals are closely related to the type of arc; micro-arcs typically range from 5 to 20 kHz, flashover arcs from 20 to 50 kHz, and continuous arcs from 50 to 200 kHz. By decomposing the signal into 32 frequency bands, the frequency characteristics of different types of electric arcs can be captured precisely.

[0174] S3.2: Calculation of energy and entropy characteristics;

[0175] The energy value and Shannon entropy value are calculated for the wavelet packet coefficients of each frequency band to obtain the energy characteristics and entropy characteristics of each frequency band.

[0176] Energy characteristics represent the degree of energy concentration of a signal within a specific frequency band, calculated by summing the squares of the wavelet packet coefficients. For the i-th frequency band, the energy characteristic is the sum of the squares of all wavelet packet coefficients in that band. Energy characteristics reflect the intensity of an electric arc signal within a specific frequency range; different types of electric arcs exhibit different energy distributions across different frequency bands.

[0177] Shannon entropy is used to quantify the complexity and uncertainty of a signal within each frequency band. The calculation of Shannon entropy first normalizes the squares of the wavelet packet coefficients to obtain a probability distribution, and then calculates the entropy value of the probability distribution. A higher entropy value indicates a more complex signal and greater uncertainty. Arc signals typically have high entropy values ​​because an electric arc is a complex nonlinear discharge phenomenon, while normal signals or simple mechanical vibrations have lower entropy values.

[0178] S3.3: Dynamic feature calculation;

[0179] Calculate the rate of change of the energy feature and the entropy feature between adjacent time windows to obtain the rate of change of energy and the rate of change of entropy.

[0180] The rate of energy change and the rate of entropy change are obtained by dividing the difference between the energy and entropy features between adjacent time windows by the time interval. The time window is typically ten milliseconds, with fifty percent overlap between adjacent windows, meaning the features are calculated every five milliseconds.

[0181] The rate of energy change reflects the temporal evolution of the arc signal energy. The energy rises rapidly at the start of the arc, remains relatively stable during steady combustion, and decreases rapidly upon extinguishing. By analyzing the rate of energy change, different stages of the arc can be identified.

[0182] The rate of entropy change reflects the temporal evolution of the complexity of an electric arc signal. The signal complexity changes rapidly during arc initiation and extinction, while it remains relatively stable during steady combustion. The rate of entropy change helps distinguish an electric arc from other interference signals; for example, mechanical vibrations typically exhibit rapid changes in both energy and entropy, while an electric arc shows rapid energy changes but relatively stable entropy.

[0183] S3.4: Feature vector combination;

[0184] The energy feature, the entropy feature, the energy change rate, and the entropy change rate are combined to obtain the multidimensional feature vector.

[0185] For the 32 frequency bands, each frequency band has four features: energy, entropy, rate of energy change, and rate of entropy change, for a total of 128 features. These features are combined into a 128-dimensional feature vector, which fully characterizes the energy entropy characteristics and time evolution characteristics of the arc signal.

[0186] The construction of the feature vector takes into account multiple dimensions of the arc signal, including frequency domain features, energy features, complexity features, and time evolution features, providing rich information for subsequent feature selection and arc identification.

[0187] This embodiment provides a complete technical solution for multidimensional feature vector extraction by detailing the specific implementation methods of wavelet packet decomposition, energy and entropy feature calculation, dynamic feature calculation, and feature vector combination.

[0188] Example 4

[0189] This embodiment, based on embodiment 1, further details the specific implementation of obtaining the Pareto optimal solution set.

[0190] like Figure 4 As shown, based on the multidimensional feature vector, feature selection is modeled as a multi-objective optimization problem. The Pareto front-end optimization algorithm is applied to filter the non-dominated solution set to obtain the Pareto optimal solution set for the arc features, including:

[0191] S4.1: Feature Importance Assessment;

[0192] The information gain and Fisher score are calculated for each feature in the multidimensional feature vector to obtain the feature importance scoring matrix.

[0193] Information gain measures the contribution of a feature to the classification result. For each feature, the entropy of the categorical variable is first calculated, which includes two classes: electric arc and non-electric arc. The entropy value represents the uncertainty of the class distribution; it is maximum when the number of samples in both classes is equal, and zero when there are only samples in one class. Then, the conditional entropy of the categorical variable is calculated given the feature value; this conditional entropy represents the uncertainty of the class distribution given the feature value. Information gain equals the categorical entropy minus the conditional entropy, indicating how much uncertainty the feature reduces. The larger the information gain, the greater the contribution of the feature to the classification.

[0194] The Fisher score assesses the ratio of between-class separation to within-class clustering for a feature. For each feature, the mean and variance of the arc class and the non-arc class are first calculated. Then, the between-class separation (the square of the difference between the two class means) and the within-class clustering (the weighted sum of the two class variances) are calculated. The Fisher score equals the between-class separation divided by the within-class clustering. A higher Fisher score indicates a stronger ability of the feature to distinguish between different classes.

[0195] By combining information gain and Fisher score, a two-dimensional score vector is constructed for each feature. For 128 features, a 128-row, two-column feature importance score matrix is ​​built, with each row corresponding to one feature. The first column is the information gain, and the second column is the Fisher score. This score matrix comprehensively reflects the performance of each feature across different evaluation dimensions.

[0196] S4.2: Multi-objective optimization modeling;

[0197] Feature selection is modeled as a multi-objective optimization problem, with the optimization objectives defined as maximizing feature discriminative power and minimizing the number of features.

[0198] The first optimization objective is to maximize the discriminative power of the feature set. The discriminative power of the feature set is calculated by a weighted sum of the information gain and Fisher scores of the selected features. For a feature subset, the sum of the information gains of all features in the subset is multiplied by a weight factor of 0.5, and then the sum of the Fisher scores of all features is multiplied by a weight factor of 0.5 to obtain the total discriminative power of that feature subset. The weight factors are chosen so that information gain and Fisher scores have equal importance in the total discriminative power.

[0199] The second optimization objective is to minimize the size of the feature set. The size of the feature set is the number of selected features. Selecting more features can improve discriminative power, but it increases computational complexity and the risk of overfitting. Selecting fewer features can reduce computational complexity, but it may sacrifice discriminative power. These two objectives are usually conflicting, requiring a multi-objective optimization method to find the optimal balance.

[0200] S4.3: Generation of candidate feature subsets;

[0201] Based on the multidimensional feature vector, a candidate feature subset is generated using binary encoding, wherein each candidate feature subset is represented by a binary vector with a length equal to the dimension of the multidimensional feature vector.

[0202] The binary encoding method uses a 128-bit binary vector to represent a feature selection scheme. Each element in the vector corresponds to a feature; a value of 1 indicates that the feature is selected, and a value of 0 indicates that the feature is not selected. For example, the vector 10010...0- indicates that the 1st, 4th, and 128th features are selected, and other features are not selected.

[0203] Several strategies can be employed to generate candidate feature subsets. One strategy is random generation, where each element is randomly set to either 0 or 1, generating a large number of candidate solutions. Another strategy is heuristic generation based on feature importance scores, prioritizing features with higher information gain and Fisher scores. Evolutionary algorithms can also be used, generating new candidate solutions from existing solutions through crossover and mutation operations.

[0204] To improve search efficiency, multiple strategies are typically combined to generate candidate feature subsets. These include randomly generated solutions to ensure search space coverage, heuristically generated solutions to improve search efficiency, and solutions generated by evolutionary algorithms to optimize the solution.

[0205] Due to the high dimensionality of the feature space (128 dimensions), exhaustive search is not feasible. The following hybrid strategy is used to generate candidate feature subsets:

[0206] First, heuristic initialization based on feature importance scores: sort by information gain and Fisher score, and select the top 50 features with the highest scores as the initial candidate feature pool.

[0207] Second, evolutionary algorithm search: NSGA-II (non-dominated sorting genetic algorithm II) is used for multi-objective optimization search, with the population size set to 100, crossover probability of 0.9, mutation probability of 0.1, and 100 generations of iteration.

[0208] Third, local search optimization: The Pareto front solution obtained by the evolutionary algorithm is further optimized by local search methods, including forward selection and backward elimination.

[0209] Using the above strategy, the algorithm can find a high-quality Pareto optimal solution set within an acceptable time.

[0210] S4.4: Pareto front-end optimization;

[0211] Based on the feature importance scoring matrix, the Pareto front-end optimization algorithm is applied to traverse the candidate feature subsets and filter feature combinations that are not dominated by the optimization objective to obtain the Pareto optimal solution set.

[0212] The core idea of ​​the Pareto front optimization algorithm is to find a set of non-dominated solutions. For two candidate solutions A and B, if A is not inferior to B on all optimization objectives and is strictly superior to B on at least one objective, then A is said to dominate B. Non-dominated solutions are those not dominated by any other solutions; these solutions constitute the Pareto front.

[0213] The algorithm first initializes an empty non-dominated solution set. Then, it iterates through all candidate feature subsets. For each candidate solution, it checks if there exists a solution in the non-dominated solution set that dominates the candidate solution. If such a solution exists, the candidate solution is dominated and not added to the non-dominated solution set. If not, the candidate solution is added to the non-dominated solution set, and it is checked whether the candidate solution dominates other solutions in the non-dominated solution set. If it does dominate, the dominated solution is removed from the non-dominated solution set.

[0214] To further optimize the solution set, the theory of tight universal boundaries of convex hulls is applied. This theory states that the Pareto front of a multi-objective optimization problem always lies on or inside the boundary of the convex hull in the objective space. By computing the convex hull in the objective space, the Pareto front can be determined more efficiently. Fast convex hull algorithms, such as the Graham scan algorithm or the Jarvis stepping algorithm, are used for convex hull computation.

[0215] After computing the convex hull, check if each solution in the non-dominated solution set lies on the convex hull boundary. Solutions on the convex hull boundary are true Pareto optimal solutions, while solutions inside the convex hull, although non-dominated, may not be globally optimal. This method can further refine the non-dominated solution set, yielding a higher-quality Pareto optimal solution set.

[0216] Finally, the optimal solution balancing discriminative power and feature quantity is selected from the Pareto optimal solution set. The selection method can employ the knee method, which finds the point with the largest curvature at the Pareto front; this point typically represents the optimal trade-off between the two objectives. Alternatively, weights can be set according to practical application requirements, and the optimal solution can be selected from the Pareto optimal solution set using a weighted summation method.

[0217] The subset of features corresponding to the selected best solution is the Pareto optimal solution set of the arc features. This feature subset typically contains twenty to thirty key features, which maintains high discriminative power while having low computational complexity, providing high-quality feature input for subsequent arc localization and identification.

[0218] This embodiment provides a complete technical solution for obtaining the Pareto optimal solution set by detailing the specific implementation methods of feature importance assessment, multi-objective optimization modeling, candidate feature subset generation, and Pareto front-end optimization.

[0219] Example 5

[0220] This embodiment, based on embodiment 1, further details the first specific implementation method of arc position positioning.

[0221] like Figure 5 As shown, the directed graph model of the fiber optic sensor network constructed based on the Pareto optimal solution set, the directed minimum cut algorithm is applied to determine the candidate positions of the electric arc, and the location is performed by the tree filling algorithm to obtain the coordinates of the electric arc position and the location confidence, including:

[0222] S5.1: Construction of directed graph models;

[0223] The fiber optic sensor network is abstracted as a directed graph model, where nodes represent virtual sensing points on the fiber, edges represent the signal propagation relationship between adjacent sensing points, and edge weights are defined based on the correlation and intensity variation of the scattered signals.

[0224] A directed graph model consists of a vertex set and an edge set. The vertex set represents a number of virtual sensing points distributed along the optical fiber, with each point corresponding to a sampling position on the fiber. The interval between sampling positions is determined by the spatial resolution of the optical time-domain reflectometer system, typically 0.5 meters. For an optical fiber one kilometer long, the number of virtual sensing points is approximately two thousand.

[0225] An edge set represents the signal propagation relationship between virtual sensing points. For two adjacent sensing points, if there is a signal propagation relationship between them, a directed edge is established between them. The direction of the edge indicates the direction of signal propagation, usually from the source end to the far end. For non-adjacent sensing points, if their scattered signals have a strong correlation, an edge can also be established; this edge is called a long-range edge and is used to capture the nonlocal correlation of the signal.

[0226] Edge weights are defined as a weighted combination of the similarity of the scattered signals and the physical distance. The similarity of the scattered signals is obtained by calculating the cosine similarity or Pearson correlation coefficient between the scattered signals at two sensing points. The cosine similarity measures the angle between two signal vectors, while the correlation coefficient measures the degree of linear correlation between the two signals. Higher similarity indicates that the signal characteristics of the two sensing points are more similar, and the edge weight is larger.

[0227] Physical distance refers to the actual distance between two sensing points. The greater the distance, the greater the signal attenuation and the smaller the edge weight. By adjusting the trade-off factor, the influence of similarity and distance in the weight calculation can be balanced. The choice of the trade-off factor needs to be optimized according to the actual application scenario; typically, the similarity trade-off factor is set to 0.7, and the distance trade-off factor is set to 0.3.

[0228] S5.2: Source and sink settings;

[0229] In the directed graph model, a source point and a sink point are set, and the directed minimum cut algorithm is applied to calculate the minimum cut set from the source point to each candidate point. The point with the smallest cut capacity is determined as the candidate position of the electric arc.

[0230] The source point is a virtual node connected to all sensing points. The edge weight from the source point to each sensing point is defined as the anomaly score for that point, calculated based on the arc features output in step S3. The anomaly score is obtained by calculating the weighted sum of the standardized distances of each feature value at the sensing point from the mean of the normal state. The standardized distance is the feature value minus the mean, divided by the standard deviation, representing the degree to which the feature value deviates from the normal state. The weights in the weighted sum are determined by Pareto optimization in step S3, with more important features receiving larger weights. A higher anomaly score indicates that the sensing point is more likely to be an arc occurrence point, and the edge weight from the source point to that point is larger.

[0231] The sink node is also a virtual node, to which all sensor points are connected. The edge weights from sensor points to the sink node are set to a fixed constant, representing the cost of identifying that point as a non-arc point. The choice of this constant affects the sensitivity of the algorithm; a larger constant makes the algorithm more likely to identify more points as arc points, resulting in higher sensitivity but also a higher false alarm rate. A smaller constant makes the algorithm more conservative, resulting in lower sensitivity but also a lower false alarm rate. Typically, this constant is set to the median of all outlier scores.

[0232] The goal of the directed minimum cut algorithm is to find a cut that divides a graph into two parts, one containing the source and the other containing the sink, such that the total weight of the cut edges is minimized. A cut edge is an edge that runs from the source to the sink. The minimum cut problem can be solved using the maximum flow algorithm; according to the maximum flow-minimum cut theorem, the minimum cut capacity of a graph equals its maximum flow.

[0233] An improved Karger algorithm is used to compute the minimum cut. The Karger algorithm is a randomized algorithm that gradually simplifies a graph by randomly merging edges until only two vertices remain; the edge between these two vertices is the cut edge. The algorithm's time complexity is close to linear, making it suitable for large-scale graphs. To improve the algorithm's accuracy, it is typically run multiple times, and the cut with the smallest capacity is selected as the result.

[0234] The minimum cut divides the graph into two parts: the part containing the source vertex corresponds to the anomalous region, and the part containing the sink vertex corresponds to the normal region. The vertices connected to the anomalous region by the set of cut edges are the possible arc occurrence points. From these candidate points, the point with the highest anomalous score is selected as the candidate arc location.

[0235] S5.3: Tree-based fill positioning;

[0236] A tree-like network structure is constructed based on the candidate locations of the electric arc. The signal strength residual of each node is calculated by applying the tree-filling algorithm, and the path with the minimum residual is determined to obtain the coordinates of the electric arc location and the location confidence.

[0237] The network structure near the candidate arc point is simplified into a tree. The root node of the tree corresponds to the candidate arc point, and the leaf nodes correspond to the surrounding sensor points. The tree is constructed based on the signal propagation path and correlation. Starting from the candidate arc point, the main direction of signal propagation is determined according to the signal correlation and time delay. The tree structure is recursively constructed along the main direction until all sensor points in the candidate region are covered.

[0238] A fill function is defined on the tree to represent the signal strength of each node. For leaf nodes, the fill function value is directly determined by the measured signal strength. For internal nodes, the fill function value is calculated as a weighted average of its child nodes. The weight of a child node is determined by the edge weight from the child node to the parent node; the larger the edge weight, the larger the weight of the child node. This bottom-up filling process ensures that the fill function value of an internal node incorporates the signal information of all its descendant nodes.

[0239] Calculate the residual between the fill function value and the measured signal for each node. The residual is the fill function value minus the absolute value of the measured signal strength, representing the fitting error of the fill function to the measured signal. The smaller the residual, the better the fill function fits the signal at that node.

[0240] Find the path with the minimum residual. Starting from the root node, select the child node with the minimum residual at each node, recursively moving downwards until the leaf node, forming a path from the root to the leaf. The nodes on this path are most likely to contain the arc occurrence point because the signal characteristics of these nodes best match the fill function, indicating that the signal propagation mode conforms to the theoretical model of arc occurrence.

[0241] To further improve positioning accuracy, an interpolation algorithm is applied along the determined path. Several virtual points are inserted between adjacent sampling points, typically nine virtual points, ensuring a distance of 0.05 meters between adjacent virtual points. The signal strength at the interpolation points is estimated using linear interpolation, i.e., a distance-weighted average based on the signal strengths of adjacent sampling points.

[0242] The residual between the fill function value of the interpolation point and the interpolated signal is calculated, and the interpolation point with the smallest residual is selected as the precise location of the electric arc. Using this interpolation method, the positioning accuracy can be improved from 0.5 meters to 0.05 meters, or five centimeters, which is far superior to the spatial resolution of the optical time domain reflectometer system itself.

[0243] Precise distance coordinates are combined with the substation's secondary circuit wiring diagram to convert them into actual physical location information. The wiring diagram contains the correspondence between fiber optic paths and secondary circuit equipment. By consulting the wiring diagram, it can be determined which equipment or terminal block the electric arc occurred near.

[0244] Calculate the confidence level of the location. The confidence level is obtained by taking an exponential function of the minimum residual value, specifically by multiplying the negative of the base of the natural logarithm by a scaling factor multiplied by the minimum residual value raised to the power of the scaling factor. The scaling factor is used to adjust the range of the confidence level, typically set to be between 0.5 and 1. The smaller the minimum residual value, the closer the confidence level is to 1, indicating a more reliable location result. The larger the minimum residual value, the closer the confidence level is to 0.5, indicating a lower reliability of the location result.

[0245] The output includes the fiber optic distance coordinates of the arc location, the corresponding physical location description, and the location confidence level. For example, the output might be a fiber optic distance of 325.35 meters, corresponding to the terminal block at the rear end of cabinet 2F of the 220 kV transformer protection panel, with a confidence level of 0.92.

[0246] This embodiment provides a complete technical solution for arc location positioning by detailing the specific implementation methods of directed graph model construction, source and sink point setting, and tree-filling positioning.

[0247] Example 6

[0248] This embodiment, based on embodiment 1, further details the second specific implementation method of arc position positioning.

[0249] The directed graph model of the fiber optic sensor network, constructed based on the Pareto optimal solution set, uses the directed minimum cut algorithm to determine candidate arc positions, and performs localization using the tree-filling algorithm to obtain the arc position coordinates and location confidence, including:

[0250] S6.1: Anomaly score calculation;

[0251] Based on the Pareto optimal solution set, an anomaly score is calculated for each virtual sensing point on the optical fiber. The anomaly score is the weighted distance of each feature value from the mean of the normal state.

[0252] For each virtual sensing point, the feature value of that point is extracted from the Pareto optimal solution set obtained in step S3. The Pareto optimal solution set contains twenty to thirty key features, each with a corresponding feature value.

[0253] Calculate the standardized distance of each feature value from the normal operating mean. The normal operating mean and standard deviation are obtained through statistical analysis of historical normal operating data. For each feature, a large number of feature values ​​during normal operation are collected, and their mean and standard deviation are calculated. The standardized distance, calculated by subtracting the mean from the current feature value and dividing by the standard deviation, represents the degree to which the current feature value deviates from the normal operating state, expressed in standard deviations.

[0254] The anomaly score for the sensor point is obtained by weighted summation of the standardized distances of all features. The weights are determined by Pareto optimization in step S3, with more important features receiving larger weights. The weighted summation ensures that the anomaly score comprehensively considers the deviation of all key features, thus fully reflecting the degree of anomaly at the sensor point.

[0255] A higher anomaly score indicates that the signal characteristics of the sensing point deviate further from the normal state, and it is more likely to be a point where an arc occurs. A negative anomaly score or one close to zero indicates that the point is in a normal state, while a positive and large anomaly score indicates that an anomaly exists at the point.

[0256] S6.2: Identification of candidate anomaly regions;

[0257] The virtual sensing points are filtered based on the anomaly score and adaptive threshold, and consecutive anomaly points are clustered to form candidate anomaly regions.

[0258] The adaptive threshold is dynamically determined based on the statistical characteristics of anomaly scores for all sensor points within the current time window. First, the mean and standard deviation of all anomaly scores are calculated. Then, the threshold is set to the mean plus a threshold coefficient multiplied by the standard deviation. The threshold coefficient is typically set between 2.5 and 3.5. A larger threshold coefficient makes the algorithm more conservative, identifying only points with very high anomalies, while a smaller threshold coefficient makes the algorithm more sensitive, identifying even minor anomalies.

[0259] This adaptive threshold effectively addresses variations in background noise levels. When ambient noise is high, the anomaly scores at all sensing points increase, and the adaptive threshold rises accordingly, preventing noise from being misidentified as an electric arc. When ambient noise is low, the adaptive threshold decreases, improving the detection sensitivity for weak electric arcs.

[0260] For each sensing point, if its anomaly score is greater than an adaptive threshold, it is marked as a candidate anomaly. Then, the candidate anomalies are clustered, grouping spatially contiguous anomalies into the same candidate anomaly region. Clustering uses a density-based method; if the distance between two anomalies is less than the cluster radius (typically set to five meters), they are grouped into the same region.

[0261] For each candidate anomalous region, its total anomalous score is calculated, which is the sum of the anomalous scores of all anomalous points within the region. Simultaneously, the temporal persistence and spatial continuity indices of the region are calculated. Temporal persistence is assessed by examining whether the anomalous region exists across multiple consecutive time windows; the longer the duration, the more likely it is to be a genuine arc rather than random noise. Spatial continuity is assessed by examining whether the distribution of anomalous points within the region is continuous; the more continuous the distribution, the more likely it is to be a genuine arc.

[0262] Candidate anomaly regions are sorted, and the top few regions with the highest total anomaly score and meeting the requirements for temporal persistence and spatial continuity are selected. Typically, the top three to five regions are chosen as the target areas for subsequent precise localization. This multi-candidate region strategy can handle complex situations where multiple electric arcs occur simultaneously.

[0263] S6.3: Directed minimum cut positioning;

[0264] A local directed subgraph is constructed for the candidate abnormal region, and the directed minimum cut algorithm is applied to calculate the minimum cut set. The point with the smallest cut capacity is determined as the candidate location of the electric arc.

[0265] For each candidate anomaly region, a local directed subgraph is constructed. The vertex set of the subgraph includes all sensor points within the region, as well as the source and sink points. The edge set of the subgraph includes edges between sensor points within the region, and edges from the source point to a sensor point and from a sensor point to the sink point. The edge weights are defined in the same way as in Example 5.

[0266] The directed minimum cut algorithm is applied to compute the minimum cut set. Since the size of the subgraph is much smaller than the complete fiber optic sensor network graph, the computational efficiency of the minimum cut algorithm is greatly improved. The Karger algorithm or other minimum cut algorithms are used to compute the minimum cut from the source to the sink.

[0267] The minimum cut divides the subgraph into two parts: the part containing the source vertex corresponds to the anomalous region, and the part containing the sink vertex corresponds to the normal region. The vertices connected to the anomalous region by the set of cut edges are the possible arc occurrence points. From these candidate points, the point with the highest anomalous score is selected as the candidate arc location for that anomalous region.

[0268] For multiple candidate anomaly regions, candidate arc locations are calculated separately. Then, based on a comprehensive evaluation of the anomaly score and confidence level, the final candidate arc locations are determined. If the anomaly scores of multiple regions are close, it may indicate the existence of multiple arc occurrence points, in which case multiple arc locations are output.

[0269] S6.4: Precise positioning for tree filling;

[0270] A tree-like network structure is constructed based on the candidate locations of the electric arc. A fill function is defined to represent the signal strength of each node, and the fill function value of the internal node is calculated by weighted averaging of the child nodes.

[0271] The method for constructing the tree network structure is the same as in Example 5. Starting from the candidate point of the electric arc as the root node, the tree structure is recursively constructed according to the signal propagation path and correlation until all sensing points in the candidate area are covered.

[0272] A fill function is defined to represent the signal strength of each node. The physical meaning of the fill function is a theoretical estimate of the node's signal strength based on the signal propagation model. For leaf nodes, the fill function value equals the measured signal strength because leaf nodes have no child nodes and cannot be derived from child nodes.

[0273] For internal nodes, the fill function value is calculated as a weighted average of its child nodes. Specifically, for each child node, its fill function value is multiplied by the edge weight from that child node to its parent node. Then, the weighted values ​​of all child nodes are summed, and finally, the result is divided by the sum of all edge weights to obtain the parent node's fill function value. This bottom-up filling process ensures that the fill function value of an internal node incorporates the signal information from all its descendant nodes.

[0274] The calculation of the filling function is actually a signal reconstruction process. It involves deriving the theoretical signals of the internal nodes layer by layer from the measured signals of the leaf nodes. If a node is the point where an arc occurs, its measured signal will differ significantly from the theoretical signal, resulting in a large residual.

[0275] S6.5: Residual Analysis and Interpolation Positioning;

[0276] Calculate the residual between the fill function value of each node and the measured signal, determine the path with the smallest residual, and perform interpolation positioning between adjacent sampling points to obtain the arc position coordinates and positioning confidence.

[0277] For each node in the tree, calculate the residual between its fill function value and the measured signal strength. The residual is the fill function value minus the absolute value of the measured signal strength, representing the difference between the theoretical estimate and the actual measurement. The smaller the residual, the more consistent the signal characteristics of the node are with the normal propagation model, and the less likely it is to be an arc point. The larger the residual, the more deviates the signal characteristics of the node from the normal propagation model, and the more likely it is to be an arc point.

[0278] Starting from the root node, select the child node with the smallest residual at each node, recursively moving downwards until a leaf node is reached, forming a path from the root to the leaf. This path is called the optimal path, and the nodes on the path represent the main direction of signal propagation. The arc point is usually not on the optimal path, but on a branch that deviates from the optimal path, because the arc changes the signal propagation characteristics.

[0279] Therefore, the node with the largest residual in the tree is most likely the point where the arc occurs. However, the node with the largest residual may be affected by noise. To improve the robustness of the localization, the top few nodes with the largest residuals, usually the top five nodes, are selected as candidate arc points.

[0280] For each candidate point, interpolation is performed between its adjacent sampling points to determine its location. The interpolation method is the same as in Example 5, where several virtual points are inserted between adjacent sampling points. The signal strength of the virtual points is estimated by linear interpolation, the residuals of the virtual points are calculated, and the virtual point with the largest residual is selected as the precise location of the arc.

[0281] By using interpolation methods, the positioning accuracy can be improved from 0.5 meters to 0.05 meters. Precise distance coordinates are combined with the wiring diagram of the substation's secondary circuits to convert them into actual physical location information.

[0282] Calculate the confidence score for the location. The confidence score considers multiple factors, including the maximum residual value, outlier scores for candidate points, temporal persistence, and spatial continuity. The confidence score is calculated using an exponential function, with parameters including a weighted combination of the maximum residual value, outlier scores, and persistence indicators. A higher confidence score indicates a more reliable location result.

[0283] The output includes the fiber optic distance coordinates of the arc location, the corresponding physical location description, and the location confidence level. If multiple candidate locations with high confidence exist, multiple arc locations are output, indicating that multiple arcs may exist.

[0284] This embodiment provides another complete technical solution for arc location by detailing the specific implementation methods of anomaly score calculation, candidate anomaly region determination, directed minimum cut positioning, tree-filled precise positioning, and residual analysis and interpolation positioning. Compared with Embodiment 5, this embodiment adds a preliminary screening step for candidate anomaly regions, improving positioning efficiency and accuracy.

[0285] Example 7

[0286] This embodiment, based on embodiment 1, further details the specific implementation method of fault diagnosis and identification.

[0287] like Figure 6 As shown, the Pareto optimal solution set, the arc position coordinates, and the positional confidence are constructed into a time-series feature matrix. The positional confidence is used as a weighting coefficient for the positional features, and the feature matrix is ​​weighted and input into a Bayesian online resource allocation optimization CNN-LSTM hybrid network for classification and recognition to obtain fault diagnosis results, including:

[0288] S7.1: Construction of the temporal feature matrix;

[0289] Based on the features in the Pareto optimal solution set, the arc position coordinates, and the position confidence, a sliding window with a time window of one hundred milliseconds is constructed, and the continuous feature vectors are organized into a time-series feature matrix.

[0290] Key features are extracted from the Pareto optimal solution set obtained in step S3. These features typically include twenty to thirty optimized frequency band energy entropy features. Each feature carries certain unique characteristics of the arc signal, such as the energy concentration in a specific frequency band, signal complexity, and time evolution characteristics.

[0291] Arc location information, including precise fiber distance coordinates and location confidence, is obtained from step S4 or S5. This location information is incorporated as contextual features into the feature matrix, enhancing the model's understanding of the spatial distribution characteristics of the arc. Arc location information is crucial for identifying arc types because arcs in different locations may exhibit different characteristics; for example, arcs in high-voltage areas are typically more intense, while arcs in low-voltage areas may be weaker.

[0292] A sliding window with a time window of 100 milliseconds is constructed. For each time point, continuous feature vectors within the previous 100 milliseconds are extracted. Assuming the temporal resolution of feature extraction is 0.1 milliseconds, that is, feature vectors are calculated once every 0.1 milliseconds, then the 100-millisecond time window contains 1000 feature vectors.

[0293] These 1000 feature vectors are arranged in chronological order to form a two-dimensional matrix. The matrix has 1000 rows, corresponding to 1000 time points, and the number of columns is the number of features, typically twenty to thirty. This two-dimensional matrix is ​​called the temporal feature matrix, which contains the complete temporal evolution information of the arc signal within a 100-millisecond time window.

[0294] To incorporate location information, the location coordinates and confidence scores are expanded into vectors of the same length as the time series matrix. Specifically, two vectors of length 1000 are created: the first vector contains all elements equal to the location coordinates, and the second vector contains all elements equal to the confidence scores. These two vectors are then added as new columns to the time series feature matrix, increasing the number of columns by two.

[0295] After expansion, the temporal feature matrix has a dimension of 1000 rows multiplied by the number of features plus two columns, containing complete temporal features and location information. This data organization method enables deep learning models to simultaneously learn the temporal patterns and spatial characteristics of electric arc signals.

[0296] S7.2: Characteristic matrix standardization;

[0297] The time series feature matrix is ​​standardized to obtain a standardized time series feature matrix.

[0298] Standardization is an important preprocessing step in deep learning model training. Its purpose is to adjust all feature values ​​to a similar range, preventing certain features from dominating model training due to excessively large value ranges, and ensuring that the model can learn all features fairly.

[0299] The standardization method used is Z-score standardization, also known as zero-mean standardization. For each column in the time-series feature matrix, i.e., each feature dimension, the mean and standard deviation of that column across all samples in the training set are first calculated. Then, for each element in that column, the mean is subtracted and the result is divided by the standard deviation to obtain the standardized value.

[0300] The standardized feature values ​​have a mean of zero and a standard deviation of one, with most values ​​ranging from -3 to +3. This distribution characteristic is beneficial for training neural networks because activation functions such as ReLU and Sigmoid have good gradient properties near zero.

[0301] It is important to note that the mean and standard deviation used for standardization must be calculated from the training set and then applied to the training, validation, and test sets. Calculating the mean and standard deviation separately for each dataset would lead to data leakage and affect the accuracy of model evaluation.

[0302] The same standardization process is applied to the two newly added features, location coordinates and confidence level. The mean and standard deviation of the location coordinates reflect the distribution characteristics of the arc location, while the mean and standard deviation of the confidence level reflect the reliability distribution of the positioning results.

[0303] After standardization, a standardized temporal feature matrix is ​​obtained, which serves as the input data for the CNN-LSTM hybrid network.

[0304] S7.3: CNN Feature Extraction;

[0305] The standardized temporal feature matrix is ​​input into a CNN network containing a multi-layer convolutional structure. Local feature patterns of the arc signal are extracted through convolution operations to obtain the CNN feature map.

[0306] A CNN network consists of three convolutional layers, each using an increasing number of kernels: 16, 32, and 64. The kernel size is 3x3, meaning three units are used in both the time and feature dimensions. The stride is 1, indicating that the kernel moves one unit at a time.

[0307] The first convolutional layer takes a standardized temporal feature matrix as input. The input dimension is 1000 times the number of features plus 2 times 1, where the last dimension 1 indicates a single channel. The first convolutional layer uses 16 3x3 convolutional kernels to perform convolution operations on the input, resulting in 16 feature maps. Each feature map captures a certain local pattern in the input data, such as a sudden increase in energy or a rapid change in entropy.

[0308] After the convolution operation, the ReLU activation function is applied. The ReLU function sets negative values ​​to zero and retains positive values, introducing non-linear characteristics that enable the network to learn complex non-linear relationships. Then, batch normalization is applied to standardize each feature map, making the mean of the feature map 0 and the variance 1, which accelerates training convergence and improves model stability.

[0309] After batch normalization, max pooling is applied with a 2x2 pooling window and a stride of 2. Max pooling selects the maximum value within each 2x2 region as the output, reducing the size of the feature map, decreasing computation, and preserving the most salient features. After the first convolution, activation, batch normalization, and pooling layer, the size of the feature map is halved.

[0310] The second convolutional layer receives the output of the first layer as input and uses 32 3x3 convolutional kernels to perform the same convolution, activation, batch normalization, and pooling operations. The second layer extracts higher-level features, such as combinations of multiple local patterns and trends in time series data.

[0311] The third convolutional layer uses 64 3x3 kernels to extract the highest-level abstract features. Instead of pooling after the third layer, global average pooling is applied. Global average pooling calculates the average of all elements in each feature map, compressing each feature map into a scalar, resulting in a 64-dimensional feature vector.

[0312] To enhance the model's adaptability to different types of arc signals, a residual connection is introduced between the second and third layers. The residual connection adds the output of the first layer, after upsampling, to the output of the second layer. Upsampling adjusts the size of the first layer's output to match that of the second layer's output through interpolation or deconvolution operations. The residual connection helps the network retain low-level features from earlier layers, preventing information loss and the vanishing gradient problem.

[0313] To capture features at different time scales, a multi-scale convolution module is introduced in the third convolutional layer. This module uses convolutional kernels of different sizes in parallel, including 3x3, 5x5, and 7x7. The 3x3 kernel captures short-term features, the 5x5 kernel captures mid-term features, and the 7x7 kernel captures long-term features. The outputs of the three kernels are then concatenated after dimensionality reduction via a one-to-one convolution to form a comprehensive multi-scale feature representation.

[0314] After processing by the CNN network, the CNN feature map is obtained, which is a 64-dimensional feature vector representing the high-level feature abstraction extracted by the CNN from the temporal feature matrix. These features capture key local patterns of the arc signal, providing refined input for subsequent LSTM time series analysis.

[0315] S7.4: LSTM timing modeling;

[0316] The CNN feature map is input into a bidirectional LSTM network to capture the temporal dependencies of the arc signal, thus obtaining a temporal feature representation.

[0317] A bidirectional LSTM network contains 128 LSTM units. LSTM is a special type of recurrent neural network that solves the gradient vanishing and gradient exploding problems of traditional recurrent neural networks through a gating mechanism, and can effectively capture long-term dependencies.

[0318] An LSTM cell contains three gates: a forget gate, an input gate, and an output gate. The forget gate determines which information is discarded from the cell state, the input gate determines which new information is stored in the cell state, and the output gate determines which information is output. The values ​​of these three gates are between zero and one, calculated using the sigmoid activation function.

[0319] The forget gate is calculated based on the current input and the hidden state from the previous time step, and then the sigmoid function is applied after a linear transformation of the weight matrix and bias vector. The output of the forget gate is multiplied by the cell state from the previous time step, determining how much historical information is retained.

[0320] The input gate is also computed based on the current input and the hidden state from the previous time step. Simultaneously, the candidate cell state is computed, and its value is constrained to between -1 and +1 using the tanh activation function. The output of the input gate is multiplied by the candidate cell state to determine how much new information is added.

[0321] The current cell state is equal to the product of the forget gate output and the cell state at the previous time step, plus the product of the input gate output and the candidate cell state. This additive operation allows gradients to propagate directly, avoiding the vanishing gradient problem.

[0322] The output gate is calculated based on the current input and the hidden state from the previous time step. The output of the gate is multiplied by the value of the current cell state after passing through the tanh function to obtain the hidden state at the current time step. The hidden state is then passed as the output of the LSTM to the next time step or the next layer.

[0323] A bidirectional LSTM consists of two LSTM layers: one processing the sequence in the forward direction and the other in the backward direction. The forward LSTM processes the sequence from beginning to end, capturing the dependencies between the past and the present. The backward LSTM processes the sequence from end to beginning, capturing the dependencies between the future and the present. The hidden states from both directions are concatenated to form the bidirectional hidden state, which contains complete information about the past and future.

[0324] For arc recognition tasks, bidirectional LSTM can simultaneously consider the start, development, and end stages of an arc signal, comprehensively capturing the temporal evolution characteristics of the arc. For example, by observing the start stage of the arc signal, the type of arc can be determined; by observing the development stage, the intensity of the arc can be determined; and by observing the end stage, the duration of the arc can be determined.

[0325] To enhance the ability of LSTM to model the long-term dependence of electric arc signals, a residual LSTM structure is introduced. Residual LSTM adds residual connections to the standard LSTM, directly adding the current input to the hidden state. This structure helps gradients propagate over long sequences and prevents gradient vanishing.

[0326] A self-attention mechanism is introduced, enabling the network to dynamically focus on important parts of the sequence. The self-attention mechanism calculates the correlation between each time step and other time steps in the sequence, generating attention weights. Time steps with higher attention weights contribute more to the final output. For electric arc recognition, the self-attention mechanism can automatically focus on key stages of the arc, such as the start-up moment and the peak moment, improving recognition accuracy.

[0327] A hierarchical LSTM structure is employed, stacking two LSTM layers. The first LSTM layer receives CNN feature maps as input, capturing basic temporal patterns. The second LSTM layer receives the hidden states of the first layer as input, modeling more advanced state transitions. This hierarchical structure enables the network to learn multi-level temporal representations, from simple time patterns to complex state evolutions.

[0328] After processing by a bidirectional LSTM network, a temporal feature representation is obtained, which is a 256-dimensional vector containing the concatenated hidden states of 128 forward LSTM units and 128 backward LSTM units. This temporal feature representation captures the complete temporal dependencies of the arc signal, providing rich information for the final classification and recognition.

[0329] S7.5: Fault classification and identification;

[0330] The temporal feature representation is classified to obtain the fault diagnosis result, which includes arc intensity, location, and duration.

[0331] The training process of the CNN-LSTM hybrid network includes:

[0332] Training data preparation: The training data comes from arc signals collected at the substation site and arc signals simulated in the laboratory, including four types: micro-arcs, flashover arcs, continuous arcs, and intermittent arcs. Each type contains at least 500 samples, for a total of more than 2,000 labeled samples. The labeling information includes arc type, intensity level, location coordinates, and duration.

[0333] Model training configuration:

[0334] Loss function: Cross-entropy loss is used for classification tasks, and mean squared error loss is used for regression tasks. The total loss is the weighted sum of the two, with weights of 0.6 and 0.4, respectively.

[0335] Optimizer: Adam optimizer, with an initial learning rate set to 0.001 and employing a learning rate decay strategy;

[0336] Batch size: 32;

[0337] Training rounds: 100 rounds, using an early stopping strategy: training stops when the validation set loss does not decrease for 10 consecutive rounds.

[0338] Data partitioning: The ratio of training set, validation set, and test set is 7:1.5:1.5;

[0339] Model performance metrics: On the test set, the accuracy rate of arc type classification reaches over 92%, and the proportion of locations with a positioning error of less than 0.5 meters reaches over 85%.

[0340] Classification is achieved through fully connected layers and a softmax activation function. The fully connected layers map 256-dimensional temporal feature representations to a category space. For arc type classification, categories include no arc, micro-arc, flashover arc, and continuous arc, with a typical number of four to six categories. The weight matrix of the fully connected layer has a dimension of 256 multiplied by the number of categories, and the bias vector has a dimension equal to the number of categories.

[0341] The output of the fully connected layer is activated by the softmax function, which transforms the output value into a probability distribution. The softmax function normalizes the output value by taking the exponent, ensuring that the sum of the probabilities of all categories is one, with the probability of each category between zero and one. The category with the highest probability is the predicted arc type.

[0342] In addition to arc type classification, it is also necessary to predict continuous-value attributes such as arc intensity, location, and duration. These attributes are achieved through a regression task. An additional fully connected layer is added after the temporal feature representation to predict arc intensity, location deviation, and duration, respectively.

[0343] Arc intensity is predicted via regression, outputting a continuous value representing the arc's energy level. The intensity value is typically normalized to between zero and one, where zero indicates no arc and one indicates the strongest arc. Intensity prediction is based on the energy characteristics and temporal evolution of the arc signal; a stronger arc indicates more concentrated energy and a longer duration.

[0344] The position deviation is predicted via regression and outputs a continuous value representing the deviation between the actual arc position and the initial positioning position. Although step S4 or S5 has already provided the arc position, the deep learning model can further refine the position based on signal features, improving positioning accuracy. The position deviation is typically between -1 meter and +1 meter, indicating that the actual position is within one meter before or after the initial positioning position.

[0345] Duration is predicted via regression, outputting a continuous value representing the duration from arc initiation to extinction. Duration is a crucial indicator of arc severity; the longer the duration, the greater the damage to equipment. Duration prediction is based on the temporal evolution characteristics of the arc signal, determined by analyzing the signal's start and end points.

[0346] The classification and regression results are combined to obtain a complete fault diagnosis result. The fault diagnosis result includes information such as arc type, arc intensity, arc location, location reliability, and duration.

[0347] This multi-task learning approach enables the model to learn classification and regression tasks simultaneously, sharing underlying feature representations and improving the model's generalization ability and prediction accuracy.

[0348] This embodiment provides a complete technical solution for fault diagnosis and identification by detailing the specific implementation methods of temporal feature matrix construction, feature matrix standardization, CNN feature extraction, LSTM temporal modeling, and fault classification and identification.

[0349] Example 8

[0350] This embodiment, based on Embodiment 7, further details the Bayesian online resource allocation optimization method for CNN-LSTM hybrid networks.

[0351] The CNN-LSTM hybrid network is optimized through Bayesian online resource allocation, including:

[0352] S8.1: Reusable resource definition;

[0353] Define a set of reusable resources in the CNN-LSTM hybrid network, which includes attention weights, inter-layer connection strengths, and feature channel activation strengths.

[0354] In deep learning models, the allocation of computational resources has a significant impact on model performance. Traditional deep learning models use the same network structure and parameters for all inputs, making it impossible to dynamically adjust computational resources based on input characteristics. Bayesian online resource allocation methods treat certain components in the model as reusable resources and dynamically allocate these resources based on input characteristics, thereby improving the model's computational efficiency and prediction accuracy.

[0355] The reusable resource set includes three types of resources:

[0356] The first category is attention weights. In self-attention mechanisms, attention weights determine the contribution of each moment in the sequence to the final output. The critical moments may differ for different types of arcs. For example, for micro-arcs, the initial initiation moment is most important, while for continuous arcs, the stable combustion phase is most crucial. By dynamically adjusting the allocation of attention weights, the model can focus on the most relevant moments, improving recognition accuracy.

[0357] The second category is interlayer connectivity strength. In residual connections and skip connections, connectivity strength determines the fusion ratio of features at different levels. For some arc types, low-level local features are more important, while for other arc types, high-level abstract features are more important. By dynamically adjusting the interlayer connectivity strength, the model can select the optimal feature combination based on the input characteristics.

[0358] The third category is the activation intensity of feature channels. In a convolutional layer, each convolutional kernel corresponds to a feature channel, and different feature channels capture different patterns. For some types of electric arcs, certain feature channels are more relevant, while for other types, other feature channels are more relevant. By dynamically adjusting the activation intensity of feature channels, the model can focus on the most relevant features and suppress irrelevant features, thereby improving computational efficiency.

[0359] These three types of resources are organized into a resource set, with each resource having a unique identifier, as well as its corresponding use value and usage cost. The size of the resource set depends on the network structure and typically contains hundreds to thousands of resources.

[0360] S8.2: Establishment of value probability distribution;

[0361] A value probability distribution model is established for each resource in the reusable resource set, and the parameters of the value probability distribution model include the prior mean and the prior variance.

[0362] For each resource, its use value is unknown and may vary with the characteristics of the input data. A Bayesian approach is used to model resource value, treating value as a random variable and representing the uncertainty of value using a probability distribution.

[0363] The value probability distribution adopts a Gaussian distribution, determined by two parameters: mean and variance. The mean represents the expected estimate of the resource value, while the variance represents the uncertainty of the estimate. The larger the variance, the less we know about the resource value, and the higher the uncertainty.

[0364] Initially, the probability distribution of the value of all resources uses a prior distribution. The prior mean is set to zero, indicating that the value of the resources is unknown initially. The prior variance is set to a large value, such as one, indicating that the uncertainty about the value of the resources is high initially.

[0365] As the model is trained and used, the value probability distribution is continuously updated by observing the effects of resource usage, gradually bringing the mean closer to the true value of the resources, reducing the variance, and lowering the uncertainty. This online learning approach enables the model to adapt to different input data distributions and continuously optimize resource allocation strategies.

[0366] S8.3: Resource value sampling;

[0367] In each identification task, based on the temporal feature matrix of the current input, the value estimate of each resource is sampled from the value probability distribution model based on Thompson sampling.

[0368] Thompson sampling is a Bayesian optimization method that balances exploration and exploitation by sampling from a posterior distribution. Exploration refers to trying resources with high uncertainty to obtain more information, while exploitation refers to selecting high-value resources to achieve better performance. Thompson sampling naturally achieves this balance through random sampling.

[0369] For each resource, a value is randomly sampled from its value probability distribution as an estimate of the resource's value. Sampling uses a Gaussian random number generator to generate random numbers based on the current mean and variance. If the resource's variance is large, the sampled value will vary widely, making it more likely to sample a value significantly different from the mean, thus enabling exploration. If the resource's variance is small, the sampled value will be close to the mean, enabling utilization.

[0370] The sampling process is performed independently for each resource, yielding estimated values ​​for all resources. These estimates are then used for subsequent resource allocation decisions.

[0371] To account for the characteristics of the input data, the sampling distribution can be adjusted based on the temporal feature matrix of the current input. Specifically, a small neural network maps the temporal feature matrix to adjustment parameters of the resource value distribution, such as mean shift and variance scaling. This context-sensitive sampling allows resource allocation to adapt to different input characteristics.

[0372] S8.4: Solving resource allocation;

[0373] Based on the estimated value and resource usage cost, a resource allocation scheme is solved under budget constraints to determine the usage status of each resource. Based on the usage status of the resources, classification and identification are performed to obtain the current fault diagnosis result.

[0374] The resource allocation problem is modeled as a knapsack problem. The knapsack problem is a classic combinatorial optimization problem where the goal is to select items under capacity constraints to maximize total value. In the resource allocation problem, resources correspond to items, the budget corresponds to the knapsack capacity, the resource value corresponds to the item value, and the resource cost corresponds to the item weight.

[0375] Resource usage cost represents the computational cost required to use that resource. For attention weights, the cost is the computational cost of calculating the attention score. For inter-layer connections, the cost is the computational cost of the connection operation. For feature channels, the cost is the computational cost of the convolution operation. Cost is typically measured in floating-point operations or memory accesses.

[0376] Budget constraints represent the total available computing resources. The budget can be set according to the specific application scenario; a smaller budget is needed if real-time response is required, while a larger budget is needed if offline processing is possible. Budget constraints force the model to make optimal choices with limited resources, rather than using all resources.

[0377] The knapsack problem can be solved using dynamic programming or a greedy algorithm. Dynamic programming can find the optimal solution, but its computational complexity is high. The greedy algorithm sorts resources by their value-to-cost ratio, prioritizing those with higher value-to-cost ratios; it has low computational complexity, but may not be the optimal solution. In practical applications, an approximate solution from the greedy algorithm is usually used to strike a balance between computational efficiency and solution quality.

[0378] The solution yields a resource allocation scheme, determining whether each resource is used. Network components corresponding to used resources are activated, while those corresponding to unused resources are suppressed. For example, if a feature channel is not selected, the convolutional kernel for that channel does not participate in the computation, and the output is set to zero. If an inter-layer connection is not selected, the weight of that connection is set to zero.

[0379] Based on the resource usage status, forward propagation is performed to calculate the model's output and obtain the current fault diagnosis result. Since only a portion of the resources are used, the model's computational load is reduced, and the inference speed is improved. At the same time, by selecting the most relevant resources, the model's prediction accuracy may also be improved.

[0380] S8.5: Value distribution update;

[0381] The loss function value is calculated based on the comparison between the current fault diagnosis result and the actual fault label. The loss function value is used as a feedback signal, and the parameters of the value probability distribution model are updated by applying the Bayesian update rule to realize the online optimization of the resource allocation strategy.

[0382] During the training phase, each input sample has a corresponding true fault label. The model's prediction is compared with the true label, and the loss function value is calculated. Cross-entropy loss is used for classification tasks, and mean squared error loss is used for regression tasks. The smaller the loss function value, the more accurate the prediction.

[0383] The negative value of the loss function is used as the reward signal, representing the effectiveness of the current resource allocation scheme. Smaller losses correspond to larger rewards, and larger losses correspond to smaller rewards. The reward signal is used to update the probability distribution of resource value.

[0384] The Bayesian update rule updates the prior distribution based on the observed rewards, resulting in the posterior distribution. For a Gaussian distribution, the Bayesian update has an analytical solution. The update formula is as follows: the posterior mean equals the weighted average of the prior mean and the observed values, with the weights determined by the prior variance and the observation noise variance. The posterior variance equals the harmonic mean of the prior variance and the observation noise variance.

[0385] For a resource being used, its value distribution is updated based on the observed rewards. If the reward is high, the mean is adjusted upwards, indicating that the resource's value is validated. If the reward is low, the mean is adjusted downwards, indicating that the resource's value is questioned. Simultaneously, a decrease in variance indicates an increased understanding of the resource's value and a reduction in uncertainty.

[0386] For unused resources, their value distribution is not updated because no usage effect has been observed. This selective updating allows the model to focus on learning the value of used resources, improving learning efficiency.

[0387] Through continuous Bayesian updates, the probability distribution of resource value gradually converges to the true value. The model learns to select the optimal resource combination under different input conditions, achieving online optimization of resource allocation strategies.

[0388] During the inference phase, although there are no real labels, other feedback signals, such as user feedback and subsequent monitoring data, can be used to continue updating the value distribution and enable the model to learn and adapt continuously.

[0389] This embodiment provides a complete technical solution for Bayesian online resource allocation optimization of CNN-LSTM hybrid networks by detailing the specific implementation methods of reusable resource definition, value probability distribution establishment, resource value sampling, resource allocation solution, and value distribution update. This optimization method enables the model to dynamically adjust computational resources according to input characteristics, improving computational efficiency while maintaining high accuracy, and is particularly suitable for arc monitoring applications with high real-time requirements.

[0390] Example 9

[0391] This embodiment, based on embodiment 1, further details the specific implementation method of generating early warning signals based on fault diagnosis results.

[0392] The step of generating a warning signal based on the fault diagnosis result includes the arc type, severity level, recommended handling measures, and expected impact range, including:

[0393] S9.1: Arc type identification;

[0394] Based on the classification results in the fault diagnosis results, the arc type is determined, including micro-arc, flashover arc, continuous arc, and intermittent arc.

[0395] The classification result is extracted from the fault diagnosis result obtained in step S7. The classification result is a probability distribution representing the probability that the input signal belongs to each arc type. The category with the highest probability is selected as the final arc type.

[0396] Micro-arcs are electric arcs with low energy and short duration, usually caused by poor contact or minor insulation damage. Micro-arcs are characterized by energy concentrated in the high-frequency range, a duration of less than ten milliseconds, and a small signal amplitude. Although micro-arcs have low energy, frequent occurrences can lead to contact oxidation and aging of insulation materials, requiring timely intervention.

[0397] A flashover arc is a transient discharge caused by insulation breakdown. It has relatively high energy but a very short duration. Flashover arcs are characterized by energy concentrated in the mid-frequency range, a duration between ten and fifty milliseconds, a large signal amplitude, and a rapid rise time. Flashover arcs are usually caused by surface contamination of insulating materials, excessive humidity, or voltage spikes, and may damage the insulating materials, requiring immediate inspection.

[0398] A sustained electric arc is an electric arc that burns continuously, possessing high energy and a long duration. The characteristics of a sustained electric arc include energy distribution across a wide frequency band, a duration exceeding fifty milliseconds, and a large and relatively stable signal amplitude. Sustained electric arcs are typically caused by severe insulation breakdown or conductor fracture, and can result in serious equipment damage or even fires, requiring immediate attention.

[0399] Intermittent electric arcs are electric arcs that appear and disappear periodically, usually caused by poor contact or loose connections. Intermittent arcs are characterized by a periodic pulse pattern, with relatively regular pulse intervals; individual pulses resemble micro-arcs or flashover arcs. Intermittent arcs indicate a mechanical connection problem, requiring inspection of fasteners and contact points.

[0400] In addition to primary types, composite types can also be identified, such as the transformation of a micro-arc into a continuous arc, indicating an increase in arc intensity and a worsening situation. By analyzing changes in arc type over time, the development trend of the fault can be determined.

[0401] S9.2: Severity assessment;

[0402] Based on the arc intensity, duration, and location reliability in the fault diagnosis results, a severity score is calculated, and severity levels are classified according to the scoring threshold.

[0403] The severity score takes into account multiple factors, including arc intensity, duration, location reliability, and location importance.

[0404] Arc intensity reflects the energy of the arc; the greater the intensity, the more severe the damage to the equipment. The intensity score is calculated by multiplying the arc intensity value by a weighting coefficient, which is set to 0.4, indicating that intensity accounts for 40% of the severity assessment.

[0405] Duration reflects the duration of the electric arc; the longer the duration, the greater the cumulative damage. The duration score is calculated by multiplying the normalized duration value by a weighting factor set to 0.3. The normalization method involves dividing the duration by a reference duration set to 100 milliseconds; durations exceeding 100 milliseconds have a normalized value greater than 1.

[0406] Location reliability reflects the reliability of the location results. The higher the confidence level, the clearer the location of the arc and the more targeted the treatment measures. The confidence score is the location reliability value multiplied by a weighting coefficient, which is set to 0.2.

[0407] Location importance reflects the criticality of the arcing location; arcs at certain locations, such as those in core equipment like main transformer protection and busbar protection, are more dangerous. Location importance is determined by querying an equipment importance database, which assigns an importance level to each piece of equipment or area, ranging from one to five, with five being the most important. The location importance score is calculated by dividing the importance level by five and then multiplying by a weighting coefficient set to 0.1.

[0408] The severity score is the sum of the intensity score, duration score, confidence score, and location importance score. The score ranges from zero to two, where zero indicates no danger and two indicates extreme danger.

[0409] The severity is graded based on a scoring threshold. There are four levels:

[0410] Level 1 (Low Risk): Score less than 0.5, corresponding to a micro-arc with a short duration and unimportant location. Routine inspections are recommended; record the event but immediate action is not necessary.

[0411] Level 2 (Medium Risk): Score between 0.5 and 1.0, corresponding to flashover arcs or micro-arcs of relatively long duration, or micro-arcs in important locations. It is recommended to schedule an inspection within 24 hours to identify the cause of the fault and develop a remedial plan.

[0412] Level 3 (High Risk): Score between 1.0 and 1.5, corresponding to a continuous arc or a flashover arc in a significant location. An on-site inspection is recommended within four hours, and power outage procedures should be arranged if necessary.

[0413] Level 4 (Emergency Risk): Score greater than 1.5, corresponding to high-intensity continuous electric arc or electric arc in core equipment. It is recommended to immediately dispatch an emergency response team, activate the emergency plan, and, if necessary, immediately cut off the power supply to prevent the accident from escalating.

[0414] The severity level, as the core information of the warning signal, determines the subsequent response speed and handling measures.

[0415] S9.3: Generation of treatment measures;

[0416] Based on the arc type and the severity level, matching treatment measures are retrieved from a predefined treatment measure knowledge base to generate suggested treatment measures.

[0417] The handling measures knowledge base is a rule base containing standard handling procedures corresponding to different arc types and severity levels. The knowledge base is built based on power system operation and maintenance experience and industry standards, and has undergone expert review and practical verification.

[0418] The knowledge base is organized as a two-dimensional table, with rows corresponding to arc types and columns corresponding to severity levels. Each cell contains the corresponding handling measures for that combination. The handling measures include detailed information such as inspection items, handling steps, required tools, and safety precautions.

[0419] For example, in the case of a micro-arc and a level 1 risk, the following measures should be taken: 1. Record the time and location of the event; 2. Inspect the terminals at that location during the next planned power outage, clean the oxide layer, and tighten them again; 3. Check the insulation material for minor damage and replace it if necessary; 4. Measure the contact resistance to ensure it meets the standards.

[0420] For situations involving continuous electric arcs and a level four risk, the following measures should be taken: 1. Immediately notify on-duty personnel and the emergency response team;

[0421] 2. Assess whether it is necessary to immediately cut off the power. If the arc is located in core equipment and shows a tendency to expand, cut off the power immediately.

[0422] 3. Dispatch an emergency response team carrying insulated tools, fire extinguishers, etc., to the scene;

[0423] IV. Isolate the faulty area and prevent personnel from approaching it;

[0424] 5. Conduct a detailed inspection using a thermal imager and an insulation resistance tester;

[0425] 6. After determining the cause of the malfunction, formulate a repair plan and replace the damaged insulation material or conductor;

[0426] 7. After the repair is completed, conduct insulation tests and withstand voltage tests. Power supply will be restored after the tests are confirmed to be satisfactory.

[0427] 8. Compile fault reports, analyze the causes of faults, and formulate preventive measures.

[0428] The knowledge base search performs exact matching based on arc type and severity level. If an exact match is found, the solution is returned directly. If no exact match is found, similarity matching is used to select the closest solution, and a message indicating that it is an approximate match is added.

[0429] The generated recommended actions are output in a structured format, including fields such as action number, action description, priority, estimated time, and required resources. Actions are sorted by priority, with higher priority actions listed first to ensure critical steps are executed first.

[0430] The handling measures also include a dynamic adjustment mechanism. If the actual situation deviates from expectations during implementation, such as increased arc intensity or additional damage, the system will reassess the severity based on the new information and update the handling measures. This dynamic adjustment allows the handling process to adapt to complex and changing field conditions.

[0431] S9.4: Prediction of the scope of impact;

[0432] Based on the arc location coordinates and the substation secondary circuit topology, the equipment and circuits that the arc may affect are analyzed, and the expected impact range is determined.

[0433] The substation secondary circuit topology includes the connection relationships of all secondary equipment, including protection devices, monitoring and control devices, communication equipment, and power supply systems. The topology is represented in the form of a graph, where nodes represent equipment and edges represent connections.

[0434] Based on the coordinates of the electric arc, locate the node or edge where the arc occurred in the topology graph. If the arc occurred inside a device, locate the device node. If the arc occurred on a connection line, locate the corresponding edge.

[0435] Starting from the node at the arc's location, a graph traversal algorithm, such as breadth-first search or depth-first search, is applied to traverse all devices connected to that node. The traversal depth is determined based on the arc type and intensity. For micro-arcs, the traversal depth is one, considering only directly connected devices. For continuous arcs, the traversal depth is two to three, considering indirectly connected devices.

[0436] During the process, the impact is assessed based on device type and connection method. Directly connected devices are most affected and may malfunction or be damaged due to electromagnetic interference from the arc, voltage fluctuations, or insulation failure. Indirectly connected devices are less affected, but may still be impacted if the arc causes a power outage or communication interruption.

[0437] For protection devices, assess whether an electric arc could cause malfunction or failure to operate. If the arc occurs in the current or voltage circuit of the protection device, it may cause the device to receive incorrect signals, triggering a malfunction. If the arc occurs in the power supply circuit of the protection device, it may cause the device to lose power and malfunction.

[0438] For measurement and control devices, assess whether an electric arc could cause erroneous measurement data or malfunctioning control commands. If the arc occurs in the signal circuit of the measurement and control device, it may cause abnormal measurement values, affecting the judgment of operators. If the arc occurs in the control circuit, it may cause control commands to fail to execute, affecting equipment operation.

[0439] For communication equipment, assess whether the electric arc could cause communication interruption or data loss. Electromagnetic interference generated by the arc can affect the transmission of communication signals, leading to data errors or communication link disruptions.

[0440] An impact scope report is generated by synthesizing information from all affected equipment. The report includes a list of affected equipment, an impact assessment, and a possible consequence analysis. The impact level is divided into three levels:

[0441] Direct impact: The equipment may be damaged or malfunction, requiring immediate inspection and handling.

[0442] Indirect impact: The equipment may be subject to interference, or its functions may malfunction, requiring monitoring and verification.

[0443] Potential impact: The equipment may be affected in the future, such as accelerated insulation aging, which needs to be included in subsequent inspection plans.

[0444] The impact report also includes a critical business impact analysis. If the affected equipment includes main protection or critical monitoring and control devices, it may lead to missing protection functions or monitoring blind spots, affecting the safe and stable operation of the power grid. The report will clearly identify these critical impacts and remind maintenance personnel to take additional safety measures, such as deploying backup protection and increasing manual inspections.

[0445] S9.5: Warning signal output;

[0446] The arc type, severity level, recommended handling measures, and expected impact range are integrated into an early warning signal, which is then sent to maintenance personnel and the monitoring center through multiple communication channels.

[0447] The warning signal uses a standardized data format, such as JSON or XML, and includes the following fields:

[0448] Event Number: A unique identifier for this warning event, formatted as date plus serial number, such as 202401-5000.

[0449] Time of occurrence: The precise time at which the electric arc is detected, accurate to the millisecond.

[0450] Arc types: micro-arc, flashover arc, continuous arc, or intermittent arc.

[0451] Severity levels: Level 1 to Level 4, and corresponding risk descriptions.

[0452] Arc location: Fiber optic distance coordinates and corresponding physical location description, such as the terminal block at the rear end of cabinet 2F of the 220 kV No. 2 transformer protection panel.

[0453] Location reliability: A value between 0 and 1, indicating the reliability of the location result.

[0454] Arc intensity: A value between 0 and 1, representing the energy level of the electric arc.

[0455] Duration: The duration in milliseconds.

[0456] Recommended handling measures: A structured list of handling steps, including a description of the measures, priority, and estimated time.

[0457] Expected impact: List of affected equipment, including equipment name, degree of impact, and possible consequences.

[0458] Additional information: Relevant waveform data, spectrograms, historical event records, etc., for detailed analysis.

[0459] Early warning signals are sent through multiple communication channels to ensure that maintenance personnel can receive them in a timely manner.

[0460] The primary channel is the graphical interface of the monitoring center. An alert window pops up on the main interface of the substation monitoring system, displaying the location and severity of the electric arc, along with recommended measures. The window uses color coding: blue for level one risk, yellow for level two, orange for level three, and red for level four, clearly indicating the urgency level. The window also includes a confirmation button; after maintenance personnel click to confirm, the system records the confirmation time and the person who confirmed.

[0461] The second channel is SMS notification. For level 3 and 4 risks, the system automatically sends SMS messages to relevant maintenance personnel. The SMS message includes the event number, arc type, location, and severity, reminding personnel to take immediate action. The list of SMS recipients is dynamically updated according to the duty schedule to ensure that information is delivered to the currently on-duty personnel.

[0462] The third channel is voice alarm. A voice alarm is played at the monitoring center, stating, "Level X arc detected, location X, please handle immediately." The volume and repetition frequency of the voice alarm are adjusted according to the severity; for level four risks, the maximum volume is used and the alarm is played continuously until confirmed.

[0463] The fourth channel is mobile application push notifications. Maintenance personnel have a dedicated application installed on their mobile devices. This application receives alert push notifications, displays detailed alert information and corresponding handling measures. The mobile application also supports on-site personnel uploading inspection photos and filling out handling records, achieving closed-loop management.

[0464] The fifth channel is email. The system sends detailed warning emails to relevant managers, including complete warning information, waveform graphs, spectrum graphs, and other attachments for post-event analysis and archiving.

[0465] All warning signals are recorded in the event database, including complete information such as sending time, receiving personnel, confirmation time, processing procedure, and processing result. The event database supports querying and statistical analysis, and can generate statistical reports on arc events, analyze the patterns and trends of arc occurrence, and provide data support for equipment maintenance and improvement.

[0466] This embodiment provides a complete technical solution for generating early warning signals based on fault diagnosis results by detailing the specific implementation methods of arc type identification, severity assessment, treatment measure generation, impact range prediction, and early warning signal output.

[0467] Example 10

[0468] This embodiment, based on Embodiment 1, further details the system's adaptive optimization and continuous learning mechanisms.

[0469] The substation secondary circuit arc fault diagnosis method based on distributed optical fiber sensing also includes adaptive optimization and continuous learning steps:

[0470] S10.1: Verification of diagnostic results;

[0471] Collect feedback information from maintenance personnel regarding the fault diagnosis results. The feedback information includes an evaluation of the accuracy of the diagnosis, the actual fault type, and the handling result.

[0472] After maintenance personnel complete on-site inspections and troubleshooting, they submit feedback information via a mobile application or monitoring system interface. The feedback information uses a structured form and includes the following fields:

[0473] Event Number: Associated with the corresponding warning event.

[0474] Diagnostic Accuracy Evaluation: Maintenance personnel evaluate the accuracy of the system's diagnostic results. Options include completely accurate, mostly accurate, partially accurate, and inaccurate. Completely accurate indicates that the arc type, location, and severity are consistent with the actual situation. Mostly accurate indicates that the main information is correct, but some details are inaccurate. Partially accurate indicates that some information is correct and some is incorrect. Inaccurate indicates that the diagnostic results do not match the actual situation.

[0475] Actual Fault Type: The actual fault type confirmed by maintenance personnel on-site may differ from the arc type diagnosed by the system. For example, the system may diagnose a continuous arc, but the actual fault may be an intermittent arc caused by poor contact.

[0476] Actual fault location: The actual fault location confirmed on-site by maintenance personnel, accurate to the specific equipment or terminal block.

[0477] Cause of failure: The root cause of the failure analyzed by maintenance personnel, such as insulation aging, loose wiring, contamination, mechanical damage, etc.

[0478] Handling measures: The actual handling measures taken by maintenance personnel, such as replacing insulation materials, re-tightening wiring, and cleaning contaminants.

[0479] Processing result: The effect after processing, with options including fault eliminated, fault mitigated, and fault not eliminated.

[0480] Additional notes: Other observations and suggestions from maintenance personnel, such as on-site environmental conditions and equipment status.

[0481] Feedback can also include on-site photos and videos to visually demonstrate the fault symptoms and handling process. Photos and videos are uploaded via a mobile application and automatically linked to the corresponding event logs.

[0482] The collected feedback information is stored in the feedback database and linked with the original warning information and diagnostic results to form a complete closed-loop record of events. The feedback database supports multi-dimensional queries and statistics, and can analyze the system's performance indicators such as diagnostic accuracy, false alarm rate, and false negative rate.

[0483] S10.2: Model performance evaluation;

[0484] Based on the feedback information, performance indicators such as diagnostic accuracy, location error, and severity assessment bias are calculated to evaluate the current performance of the model.

[0485] Diagnostic accuracy is the most important performance indicator, defined as the number of events whose diagnostic accuracy is completely or substantially accurate divided by the total number of events. Diagnostic accuracy reflects the overall reliability of the system; the higher the accuracy, the more reliable the system.

[0486] Diagnostic accuracy can be further subdivided into accuracy rates for different arc types. For each arc type, the diagnostic accuracy for that type of event is calculated, identifying which types the system performs well on and which types require improvement. For example, the system may have a high accuracy rate for identifying continuous arcs, exceeding 95%, but a low accuracy rate for identifying intermittent arcs, only 70%, indicating a need to enhance feature extraction and modeling for intermittent arcs.

[0487] Positioning error is defined as the absolute value of the difference between the system's diagnostic position coordinates and the actual fault position coordinates. The positioning error is expressed in meters and reflects the accuracy of the positioning algorithm. The average and maximum positioning errors for all events are calculated to evaluate the stability of the positioning performance.

[0488] The distribution characteristics of positioning errors can also be analyzed. If the positioning error follows a normal distribution, it indicates that the positioning algorithm is unbiased, and the error mainly comes from random noise. If the positioning error exhibits a systematic shift, such as always being too large or too small, it indicates that the positioning algorithm has systematic errors and needs calibration.

[0489] Severity assessment bias is defined as the difference between the severity level assessed by the system and the actual severity level judged by the operations and maintenance personnel. A bias of zero indicates an accurate assessment, a positive bias indicates an overestimation of severity, and a negative bias indicates an underestimation of severity.

[0490] Overestimating the severity of an issue can lead to unnecessary emergency responses and increased operational costs, but it does not pose a security risk. Underestimating the severity of an issue may result in delayed processing and increased security risks. Therefore, system design tends to slightly overestimate the severity of an issue, employing a conservative strategy to ensure security.

[0491] A confusion matrix for severity assessment is calculated. Rows in the matrix represent the system's assessed severity level, columns represent the actual severity level, and each cell represents the number of events in the corresponding combination. By analyzing the confusion matrix, we can identify which severity levels the system is prone to confusion with, and then improve the assessment algorithm accordingly.

[0492] In addition to accuracy, positioning error, and severity bias, other performance metrics are also calculated:

[0493] False alarm rate: The proportion of events that trigger a system alarm but are actually not faulty. A high false alarm rate can lead to alarm fatigue among maintenance personnel and reduce their willingness to respond.

[0494] False alarm rate: The proportion of events that actually cause a fault but for which the system does not trigger an alarm. The false alarm rate directly affects security and must be strictly controlled.

[0495] Response time: The time interval from the occurrence of an electric arc to the system issuing a warning. The shorter the response time, the more time is available for maintenance personnel to handle the situation.

[0496] Computational resource consumption: This refers to the computational resources consumed by the system, including CPU utilization, memory usage, and storage space. Reducing resource consumption while maintaining performance can improve the system's scalability and cost-effectiveness.

[0497] All performance metrics are recorded in time series, and performance trend charts are plotted to monitor changes in system performance. If a performance metric shows a downward trend, the cause is analyzed promptly and improvement measures are taken.

[0498] S10.3: Sample Importance Assessment;

[0499] Analyze the samples with diagnostic errors in the feedback data, assess the importance of the samples based on the error type, frequency and impact, and identify the difficult samples that need to be focused on for learning.

[0500] Error samples are an important resource for model improvement. By analyzing error samples, we can identify the weaknesses in the model and make targeted improvements.

[0501] Error samples are categorized by error type:

[0502] Type error: The arc type is incorrectly identified, such as misidentifying a continuous arc as a flashover arc. Type errors may lead to inappropriate handling measures and affect the effectiveness of fault handling.

[0503] Location error: The arc location is incorrect, and the positioning error exceeds the acceptable range. This location error may cause maintenance personnel to inspect the fault at the wrong location, delaying troubleshooting.

[0504] Severity error: Errors in severity assessment, particularly underestimating severity. Underestimating errors can lead to delays in processing and increase security risks.

[0505] False alarm: The system alarms but there is no actual fault. False alarms may be caused by noise interference, environmental changes, or model overfitting.

[0506] Missed detection: A fault actually exists but is not detected by the system. Missed detection is the most serious error and must be analyzed and improved.

[0507] For each error type, its frequency of occurrence is statistically analyzed to identify high-frequency errors. High-frequency errors indicate that the model systematically fails in certain scenarios and requires priority improvement.

[0508] Assess the impact of the error. The impact is determined based on the severity of the consequences of the error. For example, an error that underestimates a level 4 risk as a level 1 risk has the highest impact because it could lead to a serious safety incident. An error that falsely reports a level 1 risk has a lower impact, only increasing the workload of operations and maintenance.

[0509] The importance score of a sample is calculated by considering the error type, frequency, and impact. Samples with high importance scores are considered difficult samples and require focused study. The formula for calculating the importance score is:

[0510] Importance score = Error frequency × Impact coefficient × Novelty coefficient;

[0511] Error frequency is the percentage of errors of that type among all errors. The impact coefficient is determined based on the error type, with the highest coefficient (3) for missed detections and underestimation of severity, 2 for type errors, and 1 for location errors and false alarms.

[0512] The novelty coefficient represents the degree of difference between a sample's features and those of existing training samples. A high novelty coefficient indicates that a sample can provide new knowledge to the model if its features are rarely seen in the training set. The novelty coefficient is determined by calculating the distance between a sample and its nearest neighbors in the training set; the greater the distance, the higher the novelty.

[0513] Error samples are ranked according to importance scores, and the top 20% with the highest scores are selected as the hard sample set. The hard sample set is used for incremental training of the model, focusing on improving the model's performance in challenging scenarios.

[0514] S10.4: Incremental learning training;

[0515] The difficult samples are added to the training dataset, and the CNN-LSTM hybrid network is incrementally trained to update the network parameters and improve the ability to identify difficult samples.

[0516] Incremental learning is an online learning method that updates model parameters using new samples without retraining the entire model, enabling the model to adapt to new data distributions.

[0517] Incremental learning training employs the mini-batch gradient descent method. The difficult sample set is divided into several mini-batches, each containing 16 to 32 samples. For each batch, forward propagation is performed to calculate the model output and the loss function, followed by backpropagation to calculate the gradient and update the network parameters.

[0518] To prevent catastrophic forgetting—that is, the model forgetting old knowledge when learning new samples—a resilient weight consolidation method is employed. Resilient weight consolidation adds a regularization term to the loss function, penalizing large changes in important parameters. Important parameters are those that significantly impact performance on older tasks, and are determined by calculating the second derivative of the parameter with respect to the loss on older tasks.

[0519] Through this regularization, the changes in important parameters are constrained when the model learns new samples, preserving its performance on old tasks, while unimportant parameters can be freely adjusted to adapt to new tasks.

[0520] The learning rate for incremental learning is set to one-tenth of the initial learning rate to avoid excessive parameter updates that could lead to model instability. The number of training epochs is determined based on the size of the hard sample set, typically five to ten epochs, until the model's loss on the hard samples converges.

[0521] Monitor the model's performance on the validation set during training to prevent overfitting. Stop training early if the validation set performance starts to decline.

[0522] After incremental learning training is complete, evaluate the performance of the updated model on the test set, compare the performance metrics before and after the update to verify the effectiveness of incremental learning. If the performance improvement is significant, deploy the updated model to the production system. If the performance improvement is not significant or even declines, analyze the reasons and adjust the training strategy or sample selection method.

[0523] S10.5: Parameter adaptive adjustment;

[0524] Based on the model performance evaluation results, the feature extraction parameters, localization algorithm parameters, and classification threshold are adaptively adjusted to optimize the overall system performance.

[0525] Feature extraction parameters include the number of wavelet decomposition layers, the number and width of frequency bands, and the window length for entropy calculation. These parameters affect the discriminative power and computational complexity of the features.

[0526] Based on performance evaluation results, if the recognition accuracy for a certain type of electric arc is low, the spectral characteristics of that type of arc are analyzed, and the frequency band allocation is adjusted to better cover the characteristic frequency range of that type of arc. For example, if the recognition accuracy for intermittent arcs is low, it may be because the energy of intermittent arcs is concentrated in a specific low-frequency band, and the current frequency band allocation does not have sufficient resolution in the low-frequency band. By increasing the number of frequency bands in the low-frequency band, the ability to distinguish intermittent arcs can be improved.

[0527] The number of wavelet decomposition layers affects the balance between time and frequency resolution. More layers result in higher frequency resolution but lower time resolution. The number of layers should be adaptively adjusted based on the time-frequency characteristics of the arc signal. For arcs with long durations, more layers can be used to achieve higher frequency resolution. For arcs with short durations, fewer layers should be used to maintain higher time resolution.

[0528] The parameters of the localization algorithm include the edge weight calculation formula for the directed graph model, the parameters of the minimum cut algorithm, and the weight coefficients of the tree filling algorithm.

[0529] Based on statistical analysis of location positioning errors, if the positioning results are found to be systematically biased towards a certain direction, the trade-off factors in the side weight calculation formula can be adjusted to correct the systematic error. For example, if the positioning results are always biased towards the light source, it may be because the weight of signal similarity is too large, causing the algorithm to tend to select locations with strong signals. This bias can be corrected by reducing the similarity weight and increasing the distance weight.

[0530] The weighting coefficients in the tree-filling algorithm affect the contribution of child nodes to the filling value of their parent nodes. The weighting coefficients are adaptively adjusted based on the signal propagation characteristics of different arc types. For arcs with concentrated energy, the weights of nearby terminal nodes are increased, while the weights of distant terminal nodes are decreased. For arcs with dispersed energy, a more balanced weight distribution is used.

[0531] Classification thresholds include scoring thresholds for severity levels and budget constraints for Bayesian resource allocation.

[0532] Based on statistical analysis of severity assessment bias, adjust the scoring thresholds. If the system is found to frequently underestimate severity, lower the thresholds for each level to make the assessment more conservative. If the system is found to frequently overestimate severity, raise the thresholds to reduce unnecessary emergency responses.

[0533] Threshold adjustment employs a dynamic optimization method. The optimization objective is defined as minimizing the number of errors that underestimate severity, while controlling the number of errors that overestimate severity within an acceptable range. Grid search or Bayesian optimization methods are used to search for the optimal combination of thresholds in the parameter space.

[0534] Budget constraints in Bayesian resource allocation affect the computational efficiency and accuracy of the model. Adjusting the budget constraint based on the response time requirements and computational resource limitations of the actual application is crucial. If response time requirements are stringent, reducing the budget constraint allows the model to use fewer resources and speeds up inference. Conversely, if accuracy requirements are stringent, increasing the budget constraint allows the model to use more resources and improves prediction quality.

[0535] All parameter adjustments are recorded in the parameter history database, including the adjustment time, reason for adjustment, parameter values ​​before and after adjustment, and performance indicators. By analyzing the parameter history, patterns in parameter adjustments can be summarized, a strategy library for adaptive parameter adjustment can be established, and more intelligent automatic adjustment can be achieved.

[0536] This embodiment provides a complete technical solution for the system's adaptive optimization and continuous learning by detailing the specific implementation methods of diagnostic result verification, model performance evaluation, sample importance assessment, incremental learning training, and adaptive parameter adjustment. This continuous learning mechanism enables the system to constantly adapt to new failure modes and operating environments, maintaining long-term high performance.

[0537] Example 11

[0538] This embodiment provides a substation secondary circuit arc fault diagnosis device based on distributed optical fiber sensing, used to implement the methods described in embodiments 1 to 10 above, such as... Figure 7 As shown, the device includes:

[0539] The data acquisition module 101 is used to acquire Rayleigh scattering signals through a distributed optical fiber sensing system deployed along the secondary circuit of the substation. The distributed optical fiber sensing system includes an optical time domain reflectometer and a single-mode optical fiber. The single-mode optical fiber is laid along the cable trays, protection cabinets and terminal blocks of the secondary circuit.

[0540] The data acquisition module 101 includes an optical time domain reflectometer hardware interface, a data buffer unit, and a data preprocessing unit.

[0541] The hardware interface of the optical time domain reflectometer (OTDR) is responsible for communicating with the OTD device and receiving raw Rayleigh scattering signal data. The interface supports multiple communication protocols, such as Ethernet, serial port, and Fibre Channel, to adapt to OTD devices from different manufacturers. The interface is also responsible for configuring device parameters, such as sampling rate, pulse width, and gain, to optimize device performance according to application requirements.

[0542] The data buffer unit provides high-speed data caching to temporarily store the acquired raw signal data. Because the optical time-domain reflectometer (OTDR) generates hundreds of megabytes of data per second, a large-capacity buffer is required to prevent data loss. The buffer unit employs a circular buffer structure, where new data overwrites the oldest data, maintaining a fixed buffer capacity.

[0543] The data preprocessing unit performs preliminary processing on the raw signals, including data format conversion, timestamp addition, and data integrity verification. Format conversion transforms device-specific data formats into standard formats for easier subsequent processing. Timestamp addition adds a precise timestamp to each data point for time-series analysis. Integrity verification detects data transmission errors, such as packet loss and out-of-order delivery, ensuring data quality.

[0544] The data acquisition module 101 also includes a synchronization control unit for the synchronous acquisition of multi-channel data. If the substation has multiple fiber optic loops and needs to acquire data from multiple channels simultaneously, the synchronization control unit ensures time alignment of the data from each channel, facilitating multi-source data fusion and analysis.

[0545] The signal processing module 102 is used to filter, denoise, and normalize the Rayleigh scattering signal, and to extract multi-scale time-frequency features by applying wavelet packet decomposition to obtain a multi-dimensional feature vector.

[0546] The signal processing module 102 includes a filtering and noise reduction unit, a normalization processing unit, a wavelet packet decomposition unit, and a feature extraction unit.

[0547] The filtering and noise reduction unit implements the filtering and noise reduction method described in Example 2, including a high-pass filter, an adaptive filter, and a wavelet threshold denoiser. The filter parameters are automatically adjusted according to the signal characteristics to adapt to different noise environments.

[0548] The normalization processing unit implements the normalization method described in Example 2, including amplitude normalization and baseline correction. The normalization parameters are statistically obtained from historical data and are updated periodically to adapt to equipment aging and environmental changes.

[0549] The wavelet packet decomposition unit implements the wavelet packet decomposition method described in Example 3, supporting various wavelet basis functions, such as Daubechies wavelet, Symlet wavelet, and Coiflet wavelet. The wavelet basis functions and the number of decomposition levels are configurable, and the optimal parameters can be selected according to application requirements.

[0550] The feature extraction unit implements the feature extraction method described in Example 3, calculating the energy entropy, permutation entropy, and sample entropy for each frequency band. Feature calculation is performed in parallel, utilizing multi-core CPUs or GPUs for acceleration, thus improving processing speed.

[0551] The signal processing module 102 also includes a feature caching unit, which caches the extracted feature vectors for constructing a temporal feature matrix. The feature caching uses a sliding window mechanism to maintain feature data from the most recent period, supporting real-time analysis.

[0552] The feature optimization module 103 is used to model feature selection as a multi-objective optimization problem based on the multi-dimensional feature vector, and apply the Pareto front-end optimization algorithm to filter the non-dominated solution set to obtain the Pareto optimal solution set of the arc feature.

[0553] The feature optimization module 103 includes a feature evaluation unit, a multi-objective modeling unit, a Pareto optimization unit, and a feature selection unit.

[0554] The feature evaluation unit implements the feature importance evaluation method described in Example 4, calculates information gain and Fisher score, and generates a feature importance scoring matrix. The evaluation unit supports incremental updates, quickly updating the scoring matrix when new training data becomes available.

[0555] The multi-objective modeling unit implements the multi-objective optimization modeling method described in Example 4, defining two optimization objectives: feature discriminative power and feature quantity, and constructing an optimization problem. The modeling unit supports custom objective functions, and other optimization objectives, such as computational complexity and robustness, can be added according to application requirements.

[0556] The Pareto optimization unit implements the Pareto front-end optimization algorithm described in Example 4, including steps such as candidate solution generation, non-dominated sorting, and convex hull calculation. The optimization unit employs efficient algorithms, such as NSGA-II and SPEA2 evolutionary algorithms, supporting optimization of large-scale feature sets.

[0557] The feature selection unit selects the optimal subset of features from the Pareto optimal solution set using either the knee method or a weighted summation method. The selection unit provides a visual interface displaying the Pareto front-end graph and supports interactive user selection of the optimal solution.

[0558] The feature optimization module 103 also includes an optimization result storage unit, which stores historical optimization results and supports result comparison and trend analysis. The storage unit adopts version management, recording the parameters and results of each optimization, which facilitates backtracking and auditing.

[0559] The location module 104 is used to construct a directed graph model of the optical fiber sensor network based on the Pareto optimal solution set, apply the directed minimum cut algorithm to determine the candidate position of the electric arc, and perform positioning by the tree filling algorithm to obtain the electric arc position coordinates and positioning confidence.

[0560] The location module 104 includes a graph model building unit, a minimum cutting unit, a tree building unit, a fill calculation unit, and a location output unit.

[0561] The graph model building unit implements the directed graph model building methods described in Examples 5 and 6, including node definition, edge definition, and weight calculation. The building unit supports dynamic graph updates, automatically updating the graph model when the fiber optic network topology changes.

[0562] The minimum cutting unit implements the directed minimum cutting algorithm described in Examples 5 and 6, including source and sink node settings, flow network construction, and minimum cutting solution. The cutting unit employs efficient maximum flow algorithms, such as the Dinic algorithm and the Push-Relabel algorithm, supporting fast solutions for large-scale graphs.

[0563] The tree construction unit implements the tree network structure construction method described in Examples 5 and 6, recursively constructing trees based on signal propagation paths and correlations. The construction unit supports multiple tree construction strategies, such as minimum spanning tree and shortest path tree, and selects the optimal strategy according to application requirements.

[0564] The filling calculation unit implements the tree filling algorithm described in Examples 5 and 6, calculating the filling function value and residual of each node. The calculation unit uses a bottom-up dynamic programming method to efficiently calculate the filling value.

[0565] The positioning output unit implements the interpolation positioning and confidence calculation methods described in Embodiments 5 and 6, outputting accurate location coordinates and confidence scores. The output unit is also responsible for mapping location coordinates to physical locations, querying the wiring diagram database, and generating readable location descriptions.

[0566] The location positioning module 104 also includes a positioning result verification unit, which performs a reasonableness check on the positioning result, such as checking whether the location is within the fiber optic coverage area, whether the confidence level is within a reasonable range, and filtering out abnormal results.

[0567] The fault diagnosis module 105 is used to construct a time-series feature matrix from the Pareto optimal solution set, the arc position coordinates and the position confidence, wherein the position confidence is used as a weight coefficient of the position feature to weight the feature matrix, and input it into a CNN-LSTM hybrid network with Bayesian online resource allocation optimization for classification and recognition to obtain the fault diagnosis result.

[0568] The fault diagnosis module 105 includes a matrix construction unit, a normalization unit, a CNN feature extraction unit, an LSTM time series modeling unit, a classification and recognition unit, and a resource optimization unit.

[0569] The matrix construction unit implements the time-series feature matrix construction method described in Example 7, including sliding window extraction, positional feature expansion, and matrix organization. The construction unit supports multiple window strategies, such as fixed window and adaptive window, and selects the optimal window based on signal characteristics.

[0570] The standardization unit implements the feature matrix standardization method described in Example 7, using Z-score standardization. The standardization unit maintains the statistical parameters of the training set to ensure that the test data uses the same standardization parameters.

[0571] The CNN feature extraction unit implements the CNN network structure described in Example 7, including multi-layer convolution, activation, batch normalization, pooling, and residual connections. The extraction unit is implemented using deep learning frameworks such as TensorFlow and PyTorch, and supports GPU acceleration.

[0572] The LSTM timing modeling unit implements the bidirectional LSTM network structure described in Example 7, including LSTM layers, residual connections, and a self-attention mechanism. The modeling unit supports various LSTM variants, such as GRU and Peephole LSTM, and the optimal structure can be selected according to application requirements.

[0573] The classification and recognition unit implements the classification and regression tasks described in Example 7, including fully connected layers, softmax activation, and loss function calculation. The recognition unit supports multi-task learning and simultaneously outputs multiple prediction results such as arc type, intensity, location deviation, and duration.

[0574] The resource optimization unit implements the Bayesian online resource allocation method described in Example 8, including resource definition, value distribution establishment, Thompson sampling, resource allocation solution, and value update. The optimization unit dynamically adjusts the network's computing resources to balance accuracy and efficiency.

[0575] The fault diagnosis module 105 also includes a model management unit, which is responsible for loading, saving, version management, and deployment of models. The management unit supports hot model updates, updating models without interrupting service and enabling continuous optimization.

[0576] The early warning generation module 106 is used to generate an early warning signal based on the fault diagnosis result. The early warning signal includes the arc type, severity level, recommended handling measures, and expected impact range.

[0577] The early warning generation module 106 includes a type identification unit, a severity assessment unit, a measure generation unit, an impact analysis unit, and a signal output unit.

[0578] The type identification unit implements the arc type identification method described in Example 9, extracts classification results from the diagnostic results, and determines the arc type. The identification unit supports simultaneous identification of multiple types and handles complex arc cases.

[0579] The severity assessment unit implements the severity assessment method described in Example 9, comprehensively considering arc intensity, duration, location reliability, and location importance to calculate a severity score and classify levels. The assessment unit supports custom scoring formulas and thresholds to adapt to different application scenarios.

[0580] The measure generation unit implements the measure generation method described in Example 9, retrieving matching measures from the knowledge base. The generation unit supports online editing and updating of the knowledge base, allowing maintenance personnel to add new measures or modify existing ones.

[0581] The impact analysis unit implements the impact range prediction method described in Example 9, analyzing affected equipment based on topology. The analysis unit maintains a topology database of the substation's secondary circuits, supporting topology visualization and interactive querying.

[0582] The signal output unit implements the early warning signal output method described in Example 9, sending early warnings through multiple communication channels. The output unit supports various message formats and protocols, such as JSON, XML, MQTT, OPC UA, etc., to adapt to different monitoring systems.

[0583] The early warning generation module 106 also includes an early warning history record unit, which records complete information on all early warning events and supports historical query, statistical analysis and report generation.

[0584] The feedback learning module 107 is used to collect feedback information from maintenance personnel on the fault diagnosis results, evaluate model performance based on the feedback information, identify difficult samples, perform incremental learning training on the CNN-LSTM hybrid network, and adaptively adjust system parameters to achieve continuous optimization.

[0585] The feedback learning module 107 includes a feedback collection unit, a performance evaluation unit, a sample analysis unit, an incremental training unit, and a parameter adjustment unit.

[0586] The feedback collection unit implements the feedback information collection method described in Example 10, providing mobile application interfaces and web interfaces to support maintenance personnel in submitting structured feedback. The collection unit also supports photo and video uploads, automatically associating them with corresponding events.

[0587] The performance evaluation unit implements the model performance evaluation method described in Example 10, calculating indicators such as diagnostic accuracy, localization error, and severity deviation. The evaluation unit generates performance reports and trend charts, supporting multi-dimensional performance analysis.

[0588] The sample analysis unit implements the sample importance assessment method described in Example 10, analyzes erroneous samples, calculates importance scores, and identifies difficult samples. The analysis unit supports multiple error classification and importance calculation strategies and can be customized according to application requirements.

[0589] The incremental training unit implements the incremental learning training method described in Example 10, updating model parameters using difficult samples. The training unit employs techniques such as elastic weight consolidation to prevent catastrophic forgetting and supports both online and offline training modes.

[0590] The parameter adjustment unit implements the adaptive parameter adjustment method described in Example 10, optimizing feature extraction parameters, localization algorithm parameters, and classification thresholds based on performance evaluation results. The adjustment unit supports both automatic and manual adjustment modes and provides a visual interface for parameter adjustment.

[0591] The feedback learning module 107 also includes a knowledge base management unit, which maintains a knowledge base for handling measures, a database of equipment importance, a database of topology structures, etc., to support the continuous accumulation and updating of knowledge.

[0592] The device also includes a system management module 108, which is used for system configuration, user management, access control, log recording, system monitoring and other functions.

[0593] The system management module 108 includes a configuration management unit, a user management unit, a log management unit, and a monitoring unit.

[0594] The configuration management unit provides a system parameter configuration interface, supporting the viewing and modification of parameters for each module. The configuration unit employs hierarchical access control, with different user roles having different configuration permissions to prevent accidental operations.

[0595] The user management unit is responsible for user account management, role assignment, and access control. The unit supports multiple authentication methods, such as username / password, digital certificates, and biometrics, to ensure system security.

[0596] The log management unit records all critical events during system operation, including data acquisition, signal processing, feature extraction, fault diagnosis, early warning sending, and feedback collection. The logs use a standard format and support log querying, filtering, and exporting, facilitating troubleshooting and auditing.

[0597] The monitoring unit monitors the system's operational status in real time, including the operational status of each module, resource usage, and performance metrics. The monitoring unit provides a dashboard interface that intuitively displays the system's health status and supports alarm settings, automatically notifying the administrator when system anomalies occur.

[0598] The device provided in this embodiment realizes a complete process for arc fault diagnosis of substation secondary circuits based on distributed optical fiber sensing through the coordinated work of various functional modules. It has high accuracy, high real-time performance, high reliability and continuous learning capability, and can effectively improve the safety monitoring level of substation secondary circuits.

Claims

1. A method for online monitoring of arc faults in the secondary circuit of an intelligent substation, characterized in that, include: A pulsed optical signal is sent to the redundant optical fiber in the secondary circuit optical cable of the substation using an optical time domain reflectometer, and the Rayleigh backscattered signal in the redundant optical fiber is collected to obtain the scattered signal data stream. Wavelet packet transform is performed on the scattered signal data stream to decompose the signal into multiple frequency bands. The energy characteristics and entropy characteristics of each frequency band are calculated to obtain a multidimensional feature vector. Based on the multidimensional feature vector, feature selection is modeled as a multi-objective optimization problem. The Pareto front-end optimization algorithm is applied to filter the non-dominated solution set to obtain the Pareto optimal solution set of the arc feature. Based on the Pareto optimal solution set, a directed graph model of the fiber optic sensor network is constructed. The directed minimum cut algorithm is applied to determine the candidate positions of the electric arc, and the location is determined by the tree filling algorithm to obtain the coordinates of the electric arc position and the location confidence. The Pareto optimal solution set, the arc position coordinates, and the position confidence are used to construct a time-series feature matrix. The position confidence is used as a weight coefficient for the position features to weight the feature matrix. The matrix is ​​then input into a CNN-LSTM hybrid network with Bayesian online resource allocation optimization for classification and recognition to obtain the fault diagnosis result. Based on the fault diagnosis results, the severity of the fault is assessed and a blocking decision strategy is generated. The blocking decision strategy is then encapsulated into an MMS message and sent to the station control layer to achieve protection linkage.

2. The method according to claim 1, characterized in that, The process involves sending pulsed optical signals to redundant optical fibers in the secondary circuit optical cable of the substation via an optical time-domain reflectometer, collecting Rayleigh backscattered signals from the redundant optical fibers, and obtaining a scattered signal data stream, including: A narrow-linewidth laser is used to generate continuous light, which is then modulated into a pulsed light signal with a pulse width of 10-100ns by an acousto-optic modulator. The pulsed optical signal is sent to the redundant optical fiber, and the Rayleigh backscattering signal in the redundant optical fiber is collected by a photodetector to obtain the original scattered signal. The original scattered signal is digitally filtered and then averaged multiple times to obtain the scattered signal data stream.

3. The method according to claim 1, characterized in that, The wavelet packet transform is performed on the scattered signal data stream to decompose the signal into multiple frequency bands. The energy characteristics and entropy characteristics of each frequency band are calculated to obtain a multi-dimensional feature vector, including: Applying db4 wavelet packet transform to the scattered signal data stream decomposes the signal into 5 layers, yielding wavelet packet coefficients in 32 frequency bands; The energy value and Shannon entropy value are calculated for the wavelet packet coefficients of each frequency band to obtain the energy characteristics and entropy characteristics of each frequency band; Calculate the rate of change of the energy feature and the entropy feature between adjacent time windows to obtain the rate of change of energy and the rate of change of entropy. The energy feature, the entropy feature, the energy change rate, and the entropy change rate are combined to obtain the multidimensional feature vector.

4. The method according to claim 1, characterized in that, Based on the multidimensional feature vector, feature selection is modeled as a multi-objective optimization problem. The Pareto front-end optimization algorithm is applied to filter the non-dominated solution set, obtaining the Pareto optimal solution set for the arc features, including: The information gain and Fisher score are calculated for each feature in the multidimensional feature vector to obtain the feature importance scoring matrix; Feature selection is modeled as a multi-objective optimization problem, with the optimization objectives defined as maximizing feature discriminative power and minimizing the number of features. Based on the multidimensional feature vector, a candidate feature subset is generated using binary encoding, wherein each candidate feature subset is represented by a binary vector with a length equal to the dimension of the multidimensional feature vector; Based on the feature importance scoring matrix, the Pareto front-end optimization algorithm is applied to traverse the candidate feature subsets and filter feature combinations that are not dominated by the optimization objective to obtain the Pareto optimal solution set.

5. The method according to claim 1, characterized in that, The directed graph model of the fiber optic sensor network constructed based on the Pareto optimal solution set is used to determine the candidate positions of the electric arc using the directed minimum cut algorithm, and the arc position coordinates and positioning confidence are obtained through the tree filling algorithm. This includes: The fiber optic sensor network is abstracted as a directed graph model, where nodes represent virtual sensing points on the fiber, edges represent the signal propagation relationship between adjacent sensing points, and edge weights are defined based on the correlation and intensity variation of the scattered signals. In the directed graph model, a source point and a sink point are set, and the directed minimum cut algorithm is applied to calculate the minimum cut set from the source point to each candidate point. The point with the smallest cut capacity is determined as the candidate position of the electric arc. A tree-like network structure is constructed based on the candidate locations of the electric arc. The signal strength residual of each node is calculated by applying the tree-filling algorithm, and the path with the minimum residual is determined to obtain the coordinates of the electric arc location and the location confidence.

6. The method according to claim 1, characterized in that, The directed graph model of the fiber optic sensor network constructed based on the Pareto optimal solution set is used to determine the candidate positions of the electric arc using the directed minimum cut algorithm, and the arc position coordinates and positioning confidence are obtained through the tree filling algorithm. This includes: Based on the Pareto optimal solution set, the anomaly score of each virtual sensing point on the optical fiber is calculated. The anomaly score is the weighted distance of each feature value from the mean of the normal state. The virtual sensing points are filtered based on the anomaly score and adaptive threshold, and consecutive anomaly points are clustered to form candidate anomaly regions. A local directed subgraph is constructed for the candidate anomaly region, and the directed minimum cut algorithm is applied to calculate the minimum cut set. The point with the smallest cut capacity is determined as the candidate arc location. A tree-like network structure is constructed based on the candidate locations of the electric arc. A filling function is defined to represent the signal strength of each node. The filling function value of the internal node is calculated by weighted averaging of the child nodes. Calculate the residual between the fill function value of each node and the measured signal, determine the path with the smallest residual, and perform interpolation positioning between adjacent sampling points to obtain the arc position coordinates and positioning confidence.

7. The method according to claim 1, characterized in that, The Pareto optimal solution set, the arc position coordinates, and the location confidence are constructed into a time-series feature matrix. The location confidence is used as a weighting coefficient for the location features, and the feature matrix is ​​weighted and input into a Bayesian online resource allocation optimization CNN-LSTM hybrid network for classification and recognition to obtain fault diagnosis results, including: Based on the features in the Pareto optimal solution set, the arc position coordinates, and the position confidence, a sliding window with a time window of 100ms is constructed to organize the continuous feature vectors into a time-series feature matrix. The time series feature matrix is ​​standardized to obtain a standardized time series feature matrix; The standardized temporal feature matrix is ​​input into a CNN network containing a multi-layer convolutional structure. Local feature patterns of the arc signal are extracted through convolution operations to obtain the CNN feature map. The CNN feature map is input into a bidirectional LSTM network to capture the temporal dependencies of the arc signal and obtain a temporal feature representation. The temporal feature representation is classified to obtain the fault diagnosis result, which includes arc intensity, location, and duration.

8. The method according to claim 7, characterized in that, The CNN-LSTM hybrid network is optimized through Bayesian online resource allocation, including: Define a set of reusable resources in the CNN-LSTM hybrid network, which includes attention weights, inter-layer connection strengths, and feature channel activation strengths. A value probability distribution model is established for each resource in the reusable resource set, and the parameters of the value probability distribution model include the prior mean and the prior variance. In each identification task, based on the current input temporal feature matrix, the value estimate of each resource is sampled from the value probability distribution model based on Thompson sampling; Based on the estimated value and resource usage cost, a resource allocation scheme is solved under budget constraints to determine the usage status of each resource. Based on the usage status of the resources, classification and identification are performed to obtain the current fault diagnosis result. The loss function value is calculated based on the comparison between the current fault diagnosis result and the actual fault label. The loss function value is used as a feedback signal, and the parameters of the value probability distribution model are updated by applying the Bayesian update rule to realize the online optimization of the resource allocation strategy.

9. The method according to claim 1, characterized in that, The process of assessing the severity of the fault based on the fault diagnosis results and generating a blocking decision strategy, encapsulating the blocking decision strategy into an MMS message and sending it to the station control layer to achieve protection linkage, includes: Based on the arc intensity, duration and location information in the fault diagnosis results, the potential risk value of the arc fault to the secondary system is calculated, and the fault severity assessment result is obtained. Based on the severity assessment results of the fault, a multi-objective decision-making method is applied to find the balance between system security and power supply reliability, establish a blocking decision model that includes risk thresholds and decision rules, and generate a hierarchical blocking decision strategy. The fault diagnosis results and the hierarchical interlocking decision strategy are encapsulated into MMS messages according to the IEC61850 standard, and the message priority and target address are configured. The MMS message is sent to the protection device through the station control layer network. The protection device executes the corresponding protection action according to the hierarchical blocking decision strategy and returns the execution response result. The effectiveness of the hierarchical locking decision strategy is evaluated based on the execution response results and subsequent monitoring data. The risk threshold and decision rules in the locking decision model are adjusted according to the evaluation results to achieve adaptive optimization of the locking decision.

10. The method according to claim 1, characterized in that, Before transmitting pulsed optical signals to the redundant optical fibers in the secondary circuit optical cable of the substation via an optical time-domain reflectometer, the process also includes: The redundant fiber cores in the secondary circuit optical cable of the substation are fused together with the optical time domain reflectometer through the optical cable splice box to form a sensing optical fiber network. The pulse repetition frequency of the optical time domain reflectometer is configured to be 1-10kHz, and the sampling rate is above 100MHz.

11. An online monitoring device for arc faults in the secondary circuit of an intelligent substation, characterized in that, include: The fiber optic scattering signal acquisition module is used to send pulsed optical signals to the redundant optical fibers in the secondary circuit optical cable of the substation via an optical time domain reflectometer, acquire the Rayleigh backscattering signals in the redundant optical fibers, and obtain the scattering signal data stream. The feature extraction module is used to perform wavelet packet transform on the scattered signal data stream, decompose the signal into multiple frequency bands, calculate the energy features and entropy features of each frequency band, and obtain a multi-dimensional feature vector. The Pareto optimization module is used to model feature selection as a multi-objective optimization problem based on the multi-dimensional feature vector, and apply the Pareto front-end optimization algorithm to filter the non-dominated solution set to obtain the Pareto optimal solution set of the arc feature. The arc positioning module is used to construct a directed graph model of the fiber optic sensor network based on the Pareto optimal solution set, apply the directed minimum cut algorithm to determine the candidate position of the arc, and perform positioning through the tree filling algorithm to obtain the arc position coordinates and positioning confidence. The fault identification module is used to construct a time-series feature matrix from the Pareto optimal solution set, the arc position coordinates, and the position confidence, wherein the position confidence is used as a weight coefficient of the position feature to weight the feature matrix, and input it into a CNN-LSTM hybrid network with Bayesian online resource allocation optimization for classification and identification, so as to obtain the fault diagnosis result. The interlocking decision and communication module is used to assess the severity of the fault based on the fault diagnosis results and generate an interlocking decision strategy. The interlocking decision strategy is then encapsulated into an MMS message and sent to the station control layer to achieve protection linkage.

Images