Lightning protection system effectiveness evaluation method and system based on multi-dimensional research and judgment

Abstract

The invention provides a lightning protection system effectiveness evaluation method and system based on multi-dimensional research and judgment, and relates to the technical field of lightning protection, and the method comprises the steps: collecting system index data through a bidirectional feature extraction network, employing a separable convolution layer with an attention mechanism and multi-scale Fourier transform to extract features, and carrying out the analysis of the features; an index transfer function is constructed in combination with a causal reasoning network, evaluation grade division is performed by adopting kernel principal component analysis and density clustering, and system parameters are optimized through a hierarchical reinforcement learning framework, so that accurate evaluation and dynamic optimization of the state of the lightning protection system can be realized, and the reliability and the protection effect of the system are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of lightning protection, and in particular, to a method and system for evaluating the effectiveness of a lightning protection system based on multi-dimensional judgment. Background Art

[0002] A lightning protection system is an important infrastructure for ensuring the safety of power equipment and personnel. The evaluation of its operating status and protection effect is of great significance for system maintenance and optimization. Traditional lightning protection system evaluation methods mainly rely on manual experience judgment and single-index analysis. With the increase in system complexity and data dimensions, the accuracy and timeliness of evaluation face severe challenges.

[0003] Currently, the evaluation of lightning protection systems mainly uses static index detection and simple threshold judgment methods to evaluate the system status through regular inspections and offline analysis. Some studies have begun to introduce intelligent algorithms such as machine learning for system evaluation, but there are still problems such as single evaluation dimension and insufficient real-time performance. At the same time, existing evaluation methods often analyze each index separately and lack in-depth research on the overall performance of the system and the correlation between indexes.

[0004] Therefore, there is an urgent need for a solution to solve the problems existing in the prior art. Summary of the Invention

[0005] Embodiments of the present invention provide a method and system for evaluating the effectiveness of a lightning protection system based on multi-dimensional judgment, which can at least solve some problems existing in the prior art.

[0006] In the first aspect of the embodiments of the present invention, a method for evaluating the effectiveness of a lightning protection system based on multi-dimensional judgment is provided, including:

[0007] Construct a two-way feature extraction network to collect real-time monitoring data of system indexes, perform sliding sampling through a separable convolutional layer with an attention mechanism to obtain a basic feature sequence, perform multi-scale Fourier transform on the basic feature sequence to obtain periodic features in different frequency domains, fuse the periodic features with the basic feature sequence through a cross-frequency domain attention network to generate a time-series feature vector, construct a dynamic association graph based on the time-series feature vector, and combine with a graph neural network to iteratively update the node state and output a multi-modal feature matrix representing the overall state of the system.

[0008] Input the multi-modal feature matrix into a causal inference network, construct a transfer function between indexes based on a structural equation model, establish a long short-term memory module through a gated recurrent unit, perform time-series combination on the historical state information and the multi-modal feature matrix and construct a multi-head self-attention model, calculate the attention weight for the combined feature sequence, perform weighted aggregation on the attention weight and the corresponding feature, construct a loss function in combination with a contrast learning strategy, optimize the feature representation based on the loss function and output a high-dimensional evaluation vector.

[0009] Apply kernel principal component analysis to the high-dimensional evaluation vector for dimensionality reduction, input the dimensionality-reduced evaluation vector into a density-based spatial clustering algorithm, divide the evaluation levels according to the density reachability principle, construct a hierarchical reinforcement learning framework, where the high-level policy network plans an optimal target sequence based on the current system state, and the low-level execution network decomposes the target sequence into specific tuning parameters, samples and explores the parameter space through Monte Carlo tree search, and iteratively calculates the optimal parameter combination and outputs an improvement plan by combining Bayesian optimization.

[0010] In an alternative implementation,

[0011] Construct a bidirectional feature extraction network to collect real-time monitoring data of system metrics, obtain a basic feature sequence through sliding sampling by a separable convolutional layer with an attention mechanism, perform multi-scale Fourier transform on the basic feature sequence to obtain periodic features in different frequency domains, fuse the periodic features and the basic feature sequence through a cross-frequency domain attention network to generate a time series feature vector, construct a dynamic association graph based on the time series feature vector, and iteratively update the node states by combining a graph neural network to output a multi-modal feature matrix representing the overall state of the system, including:

[0012] Establish a bidirectional feature extraction network, the bidirectional feature extraction network includes a forward propagation branch and a backward propagation branch. The forward propagation branch inputs the real-time monitoring data into a dynamic buffer and divides the data stream into time segments through a sliding time window, processes the time segments by a separable convolutional layer that decomposes depth convolution and pointwise convolution to obtain multi-channel features, constructs a channel attention mechanism based on the multi-channel features, calculates weight coefficients according to the feature response intensity and channel dependence relationship to obtain weighted features, and the backward propagation branch constructs a mirror structure through a recurrent unit to perform long-range dependence correction on the weighted features, and combines the corrected bidirectional features through a feature selection gating mechanism to obtain a basic feature sequence;

[0013] Construct a pyramid hierarchical structure for the basic feature sequence and dynamically adjust the inter-layer scale based on the time series change characteristics of the basic feature sequence to obtain multi-layer features. Perform fast Fourier transform within each layer's time window to extract frequency components and phase information, and select the main frequency components according to the energy entropy distribution of the basic feature sequence. Input the main frequency components into a multi-head cross-attention network and perform feature fusion by combining the time series information of the basic feature sequence. Process the multi-layer features through a local feature extraction unit and a global feature fusion unit and optimize the feature combination path based on the time-varying characteristics of the basic feature sequence. Process the optimized features through residual connection and a feed-forward network to obtain a time series feature vector;

[0014] Calculate the correlation between nodes based on the temporal feature vectors to construct a dynamic association graph and determine the connection relationship. Use a hierarchical message passing framework to iteratively update the dynamic association graph. Initialize the node states with the temporal feature vectors and extract structural features through gated graph convolution. Calculate attention weights based on the similarity of node states to weighted aggregate and update the neighbor node information to obtain node representations. Preserve the multi-granularity features of the nodes through skip connections and fuse the updated node state features with the temporal feature vectors. Finally, obtain the multi-modal feature matrix representing the overall state of the system through the processing of the feature transformation network.

[0015] In an alternative embodiment,

[0016] Preserve the multi-granularity features of the nodes through skip connections and fuse the updated node state features with the temporal feature vectors. Finally, obtain the multi-modal feature matrix representing the overall state of the system through the processing of the feature transformation network, including:

[0017] Extract the original attributes and structural relationship information of the nodes through a multi-scale skip connection structure, obtain multi-scale features after weighting by a feature selection gating unit, map them to a unified feature space for attention weighting, obtain aligned features and interact with the temporal features, and combine with a multi-branch feature transformation network to generate a multi-modal feature matrix, specifically including:

[0018] Construct a multi-scale skip connection structure, which includes a shallow feature transfer path and a deep feature transfer path. Input the original attribute information of the nodes into the shallow feature transfer path, adjust the feature dimension to match the feature scale of the target layer through feature resampling. Input the structural relationship information of the nodes into the deep feature transfer path, reduce redundancy through feature compression to obtain high-order structural information. The outputs of the shallow feature transfer path and the deep feature transfer path respectively pass through a feature selection gating unit, and the feature selection gating unit calculates the feature importance score based on the historical state information of the nodes, and dynamically weights the outputs of the shallow feature transfer path and the deep feature transfer path to obtain multi-scale features;

[0019] Map the node features at different levels in the multi-scale features to a unified feature space through non-linear transformation, use an attention calculation unit to construct a feature correlation matrix, generate the importance weights of each feature based on the correlation matrix, fuse the mapped features according to the importance weights, establish a residual connection between the fusion result and the original features, and perform feature normalization to obtain aligned features;

[0020] Construct a bidirectional feature mapping network, where the bidirectional feature mapping network includes a node feature mapping path and a temporal feature mapping path. Input the aligned features into the node feature mapping path, extract the dependency relationship features between nodes through a multi-head attention module, input the temporal data into the temporal feature mapping path, calculate the temporal feature weights based on periodicity and continuity, and adaptively fuse the output features of the node feature mapping path and the temporal feature mapping path to obtain interaction features;

[0021] Input the interaction features into a multi-branch feature transformation network. The multi-branch feature transformation network includes a spatial feature processing unit, a temporal feature processing unit, and a frequency feature processing unit. The spatial feature processing unit extracts local features through depth convolution and pointwise convolution decomposition. The temporal feature processing unit extracts dynamic features through a recurrent network with a selection gate and an update gate. The frequency feature processing unit extracts periodic features through signal decomposition. The outputs of the three processing units are subjected to feature calibration and adaptive fusion to obtain a multi-modal feature matrix.

[0022] In an alternative embodiment,

[0023] Input the multi-modal feature matrix into a causal inference network, construct a transfer function between indicators based on a structural equation model, establish a long short-term memory module through a gated recurrent unit, perform a temporal combination of historical state information and the multi-modal feature matrix and construct a multi-head self-attention model, calculate attention weights for the combined feature sequence, perform weighted aggregation of the attention weights and the corresponding features, construct a loss function in combination with a contrastive learning strategy, and optimize the feature representation based on the loss function and output a high-dimensional evaluation vector, including:

[0024] Perform block preprocessing on the multi-modal feature matrix. Divide the features into multiple feature subsets according to the monitoring object type and physical attributes. Construct an independent transfer function between indicators for each feature subset. The transfer function between indicators includes a feature mapping layer and a non-linear transformation layer. The feature mapping layer projects the input features into a high-dimensional feature space to obtain projected features. The non-linear transformation layer performs activation transformation on the projected features to obtain initial transfer features. A feature distribution normalization unit and a skip connection path are set between adjacent transfer function layers. The initial transfer features are processed through the feature distribution normalization unit and the skip connection path to obtain optimized transfer features;

[0025] Input the optimized transfer feature into a gated recurrent unit, which includes a state calculation unit and a forgetting gate unit. The state calculation unit fuses the optimized transfer feature and historical state information to generate a candidate state, and the forgetting gate unit selectively forgets the candidate state to obtain a basic memory feature. The middle-layer memory module introduces a multi-head self-attention model to perform importance weighting on the basic memory feature to obtain an enhanced memory feature. The top-layer memory module adaptively modulates the enhanced memory feature based on task constraints to obtain a modulated memory feature;

[0026] Adopt a dynamic window mechanism to segment the modulated memory feature. The window length is adaptively adjusted according to the data change characteristics, and an overlapping area is set between adjacent windows to maintain sequence continuity to obtain a segmented feature sequence. Perform temporal combination on the segmented feature sequence. The temporal combination includes performing temporal alignment on the segmented feature sequence to obtain an aligned feature, and performing spatio-temporal fusion on the aligned feature to obtain a fused feature sequence;

[0027] Input the fused feature sequence into a multi-head self-attention model. The first layer decouples the fused feature sequence into multiple independent sub-features and calculates attention weights. The second layer adaptively combines the attention weights to obtain a global attention vector. The third layer reconstructs the fused feature sequence based on the global attention vector to obtain an attention feature;

[0028] Perform sequence truncation transformation, feature masking transformation, and random perturbation transformation on the attention feature to obtain an enhanced feature sequence. Input the enhanced feature sequence into a deep encoding network to extract a compressed representation. Based on the compressed representation, construct a loss function. The loss function generates positive and negative sample pairs and calculates a contrastive loss. Optimize the feature representation based on the contrastive loss and output a high-dimensional evaluation vector.

[0029] In an alternative embodiment,

[0030] Performing sequence truncation transformation, feature masking transformation, and random perturbation transformation on the attention feature to obtain an enhanced feature sequence. Inputting the enhanced feature sequence into a deep encoding network to extract a compressed representation includes:

[0031] Construct a three-channel parallel enhancement processing architecture for the attention feature, and perform sequence truncation transformation, feature masking transformation, and random perturbation transformation respectively. The sequence truncation transformation uses a three-level truncation window to extract multi-scale features. The feature masking transformation performs hierarchical masking based on feature importance. The random perturbation transformation performs differential perturbation according to feature correlation, and performs dynamic fusion based on information entropy and signal-to-noise ratio to obtain an enhanced feature sequence. Input the enhanced feature sequence into a deep encoding network to extract a compressed representation, specifically including:

[0032] Construct a parallel enhancement processing unit, which is provided with a sequence truncation transformation channel, a feature masking transformation channel and a random perturbation transformation channel. Synchronously input the attention features into the parallel enhancement processing unit. The sequence truncation transformation channel is provided with three levels of truncation windows. The first-level truncation window extracts the long-period variation information in the attention features to obtain the main structure feature sequence. The second-level truncation window extracts the medium-period variation information in the attention features to obtain the local feature sequence. The third-level truncation window extracts the short-period variation information in the attention features to obtain the transient feature sequence. Calculate the frequency-domain distribution characteristics of the main structure feature sequence, the local feature sequence and the transient feature sequence respectively. Determine the sequence fusion weights according to the frequency-domain distribution characteristics. Combine the three feature sequences based on the sequence fusion weights to obtain a multi-scale feature sequence;

[0033] The feature masking transformation channel calculates the information gain of the attention features to obtain the feature importance. Divide the attention features into three importance levels based on the feature importance. The first-level features adopt the minimum masking ratio and the shortest masking duration. The second-level features adopt the medium masking ratio and the medium masking duration. The third-level features adopt the maximum masking ratio and the longest masking duration. Perform masking processing on the three-level features to obtain a masked feature sequence;

[0034] The random perturbation transformation channel calculates the distances between pairs of the attention features to obtain a correlation matrix. Construct a minimum spanning tree based on the correlation matrix. Group the features with distances less than the threshold in the minimum spanning tree. Apply the same perturbation amplitude and perturbation frequency to the features in the same group, and apply different perturbation amplitudes and perturbation frequencies to the features in different groups. Calculate the first derivative of the attention features to obtain the change rate. Dynamically adjust the perturbation amplitude according to the change rate to obtain a perturbed feature sequence;

[0035] Construct a feature fusion unit, which calculates the probability distributions of the multi-scale feature sequence, the masked feature sequence and the perturbed feature sequence respectively to obtain the information entropy. Calculate the similarity between the multi-scale feature sequence, the masked feature sequence and the perturbed feature sequence and the original features to obtain the signal-to-noise ratio. Determine the feature fusion weights according to the information entropy and the signal-to-noise ratio. Combine the three feature sequences based on the feature fusion weights to obtain an enhanced feature sequence;

[0036] Construct a deep encoding network. A cross-layer feature transmission path is set between adjacent encoding layers of the deep encoding network. Each encoding layer is configured with a feature selection module. The feature selection module calculates response scores for the input features and filters out high-response features. A feature restoration module is set at the end of the deep encoding network. The feature restoration module maps the encoded features back to the original domain to obtain reconstructed features, calculates the deviation between the reconstructed features and the original features to obtain a reconstruction loss, and optimizes the parameters of the deep encoding network according to the reconstruction loss to obtain a compressed representation.

[0037] In an alternative implementation,

[0038] Apply kernel principal component analysis to the high-dimensional evaluation vector for dimensionality reduction, input the dimensionality-reduced evaluation vector into a density-based spatial clustering algorithm, divide the evaluation levels according to the density reachability principle, construct a hierarchical reinforcement learning framework, where the high-level policy network plans an optimal target sequence based on the current system state, the low-level execution network decomposes the target sequence into specific tuning parameters, samples and explores the parameter space through Monte Carlo tree search, and iteratively calculates the optimal parameter combination by combining Bayesian optimization and outputs an improvement plan, including:

[0039] Construct a kernel function for the high-dimensional evaluation vector, calculate a kernel similarity matrix through the kernel function, project the high-dimensional evaluation vector into a feature space using the kernel similarity matrix, calculate the sample covariance matrix in the feature space, perform eigenvalue decomposition on the sample covariance matrix to obtain an eigenvalue sequence and an eigenvector sequence, sort the eigenvalue sequence in descending order and calculate the cumulative contribution rate, select the main eigenvectors according to a preset contribution rate threshold to construct a dimensionality reduction projection matrix, and perform kernel principal component analysis dimensionality reduction on the high-dimensional evaluation vector using the dimensionality reduction projection matrix;

[0040] Calculate a distance matrix for all sample points in the dimensionality-reduced evaluation vector, statistically analyze the distance distribution characteristics between samples based on the distance matrix, determine the density calculation radius through the distance distribution characteristics, count the number of neighboring samples within the density calculation radius for each sample point to obtain a local density value, identify the density peak points as clustering centers according to the local density values, start from the clustering centers and sequentially add adjacent low-density samples to the clusters to form density-decreasing clustering chains, and stop clustering expansion when the density difference between adjacent samples exceeds a preset threshold, and divide the evaluation levels according to the density reachability principle;

[0041] Construct a hierarchical reinforcement learning framework based on the evaluation level. The high-level policy network performs multi-scale analysis on the current system state through the state feature extraction module to obtain a state feature representation. The state feature representation is input into the target planning module, and the target planning module generates an optimization target sequence. The low-level execution network receives the optimization target sequence and converts the optimization target sequence into tuning parameters using a hierarchical decomposition method. A value evaluation module is set between the two networks to achieve information interaction;

[0042] Construct a Monte Carlo search tree structure for the parameter space. The search tree nodes record candidate parameter groups and evaluation scores, and the search tree edges represent the parameter adjustment directions. Calculate the upper bound of the node value based on the historical scores and visit counts of the nodes, select the node with the maximum upper bound of the value for expansion, generate child nodes under the node with the maximum upper bound of the value, randomly sample the parameters of the child nodes and calculate the evaluation scores, and propagate the evaluation scores from the leaf nodes to the root node to update the path statistical information;

[0043] Establish a Bayesian optimization model. Input the historical parameter solutions and effect scores as training samples into the Bayesian optimization model. Use the Bayesian optimization model to predict the effects and uncertainty degrees of the parameter solutions to be evaluated. Design a combined acquisition function to comprehensively balance the predicted effects and uncertainty degrees, select the parameter solution corresponding to the maximum acquisition function value for actual evaluation, add the new evaluation results to the training samples to update the Bayesian optimization model, and output the optimal parameter combination and generate an improvement plan through iterative calculations.

[0044] In an alternative embodiment,

[0045] Using the Bayesian optimization model to predict the effects and uncertainty degrees of the parameter solutions to be evaluated, and designing a combined acquisition function to comprehensively balance the predicted effects and uncertainty degrees, and selecting the parameter solution corresponding to the maximum acquisition function value for actual evaluation includes:

[0046] Perform predictions using the Bayesian optimization model, partition the parameter space to perform local and global dual-scale predictions, construct an acquisition function through non-dominated ranks and crowding degrees, calculate the membership relationship of the parameter distribution family by combining the hierarchical clustering method, introduce a multi-scheme parallel evaluation mechanism, and perform local incremental updates on the evaluation data, specifically including:

[0047] Construct a Bayesian optimization model for the historical parameter solutions and evaluation data. The Bayesian optimization model automatically adjusts the kernel function parameters according to the data density distribution, divides the complete parameter space into multiple continuous sub-regions, performs local predictions in the multiple continuous sub-regions respectively to obtain local prediction results, merges the local prediction results based on the boundary constraint conditions to obtain global prediction results, calculates the local neighborhood prediction effects and global range prediction effects for each parameter solution to be evaluated respectively, and calculates the uncertainty degree by combining the sample distribution density;

[0048] Construct the prediction effect and the degree of uncertainty into a bi-objective optimization problem, calculate the non-dominated relationship of the parameter scheme on the two objectives to obtain the dominance level sequence, calculate the crowding degree index by statistically analyzing the distribution density of the parameter scheme in the objective space, construct a combined acquisition function by combining the dominance level sequence and the crowding degree index, record the coverage density of the historical evaluation scheme in the parameter space, and impose a scoring penalty on the high-frequency exploration area according to the coverage density;

[0049] Establish a similarity calculation criterion for historical parameter schemes, construct a distance matrix based on the similarity calculation criterion, perform hierarchical clustering division on the parameter schemes using the distance matrix to obtain a set of parameter distribution families, calculate the membership relationship between the scheme to be evaluated and each family in the set of parameter distribution families, and weightedly superimpose the membership relationship to the combined acquisition function;

[0050] Construct a measure index in the parameter space, calculate the distance difference and prediction feature difference between candidate schemes using the measure index, select multiple candidate schemes with the largest degree of difference to form an evaluation set based on the distance difference and the prediction feature difference, perform actual evaluations on the candidate schemes in the evaluation set synchronously, and analyze the correlation degree between the actual evaluation results;

[0051] Input the newly obtained evaluation data into the Bayesian optimization model, identify the model parameters within the influence range of the evaluation data, only perform local updates on the parameters within the influence range, establish a criterion for judging the effectiveness of the evaluation results, and screen out abnormal evaluation data according to the criterion;

[0052] Dynamically determine the number of parallel evaluation schemes according to the optimization stage, select multiple candidate schemes for parallel evaluation during the parameter space exploration stage, reduce the number of parallel evaluations during the parameter optimization stage, calculate the target weight adjustment step size based on the historical evaluation effect, and adaptively update the combined weight of the prediction effect and the degree of uncertainty using the adjustment step size to obtain new evaluation results.

[0053] In the second aspect of the embodiments of the present invention, a lightning protection system effectiveness evaluation system based on multi-dimensional research and judgment is provided, including:

[0054] A first unit for constructing a two-way feature extraction network to collect real-time monitoring data of system indicators, obtaining a basic feature sequence through sliding sampling by a separable convolutional layer with an attention mechanism, performing multi-scale Fourier transform on the basic feature sequence to obtain periodic features in different frequency domains, fusing the periodic features and the basic feature sequence through a cross-frequency domain attention network to generate a time series feature vector, constructing a dynamic association graph based on the time series feature vector, and iteratively updating the node states in combination with a graph neural network to output a multi-modal feature matrix representing the overall state of the system;

[0055] The second unit is used to input the multi-modal feature matrix into the causal inference network, construct the transfer function between indicators based on the structural equation model, establish a long short-term memory module through a gated recurrent unit, perform temporal combination of historical state information and the multi-modal feature matrix, construct a multi-head self-attention model, calculate the attention weights for the combined feature sequence, perform weighted aggregation of the attention weights and the corresponding features, construct a loss function in combination with the contrast learning strategy, optimize the feature representation based on the loss function, and output a high-dimensional evaluation vector;

[0056] The third unit is used to perform dimensionality reduction on the high-dimensional evaluation vector by applying kernel principal component analysis, input the dimensionality-reduced evaluation vector into the density-based spatial clustering algorithm, divide the evaluation levels according to the density reachability principle, construct a hierarchical reinforcement learning framework, where the high-level policy network plans and optimizes the target sequence based on the current system state, the low-level execution network decomposes the target sequence into specific tuning parameters, samples and explores the parameter space through Monte Carlo tree search, iteratively calculates the optimal parameter combination in combination with Bayesian optimization, and outputs an improvement plan.

[0057] In the third aspect of the embodiments of the present invention,

[0058] A kind of electronic device is provided, including:

[0059] A processor;

[0060] A memory for storing instructions executable by the processor;

[0061] Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0062] In the fourth aspect of the embodiments of the present invention,

[0063] A computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0064] In the present invention, by constructing a bidirectional feature extraction network and a cross-frequency domain attention mechanism, the efficient acquisition and feature fusion of multi-dimensional indicators of the lightning protection system are realized, the dynamic change characteristics of the system state can be accurately captured, the comprehensiveness and accuracy of feature extraction are improved, a causal inference network and a multi-head self-attention model are adopted to establish the transfer relationship and temporal dependence analysis between indicators, the feature representation is optimized through contrast learning, the accurate evaluation of the system state is realized, the reliability and scientificity of the evaluation results are improved, combined with kernel principal component analysis and density clustering algorithm, a hierarchical reinforcement learning framework is constructed, parameter optimization is carried out through Monte Carlo tree search and Bayesian optimization, an adaptive system improvement plan can be given, and the practicality and operability of the evaluation results are improved. Description of the Drawings

[0065] Figure 1 This is a schematic flowchart of the method for evaluating the effectiveness of a lightning protection system based on multi-dimensional judgment in an embodiment of the present invention;

[0066] Figure 2 This is a schematic structural diagram of the system for evaluating the effectiveness of a lightning protection system based on multi-dimensional judgment in an embodiment of the present invention. Detailed implementation manners

[0067] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are only a part rather than all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0068] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0069] Figure 1 This is a schematic flowchart of the method for evaluating the effectiveness of a lightning protection system based on multi-dimensional judgment in an embodiment of the present invention, as Figure 1 shown, the method includes:

[0070] S1. Construct a bidirectional feature extraction network to collect real-time monitoring data of system indicators, perform sliding sampling on the obtained basic feature sequence through a separable convolutional layer with an attention mechanism, perform multi-scale Fourier transform on the basic feature sequence to obtain periodic features in different frequency domains, fuse the periodic features and the basic feature sequence through a cross-frequency domain attention network to generate a time series feature vector, construct a dynamic association graph based on the time series feature vector, and iteratively update the node states in combination with a graph neural network to output a multi-modal feature matrix representing the overall state of the system;

[0071] In an optional implementation manner,

[0072] Constructing a bidirectional feature extraction network to collect real-time monitoring data of system indicators, performing sliding sampling on the obtained basic feature sequence through a separable convolutional layer with an attention mechanism, performing multi-scale Fourier transform on the basic feature sequence to obtain periodic features in different frequency domains, fusing the periodic features and the basic feature sequence through a cross-frequency domain attention network to generate a time series feature vector, constructing a dynamic association graph based on the time series feature vector, and iteratively updating the node states in combination with a graph neural network to output a multi-modal feature matrix representing the overall state of the system includes:

[0073] Build a bidirectional feature extraction network, which includes a forward propagation branch and a backward propagation branch. The forward propagation branch inputs real-time monitoring data into a dynamic buffer and divides the data stream into time segments through a sliding time window. A separable convolutional layer that decomposes depth convolution and pointwise convolution is used to process the time segments to obtain multi-channel features. A channel attention mechanism is constructed based on the multi-channel features, and weight coefficients are calculated according to the feature response intensity and channel dependence relationship to obtain weighted features. The backward propagation branch constructs a mirror structure through a recurrent unit to perform long-range dependence correction on the weighted features, and the corrected bidirectional features are merged through a feature selection gating mechanism to obtain a basic feature sequence;

[0074] Construct a pyramid hierarchical structure for the basic feature sequence and dynamically adjust the inter-layer scale based on the temporal variation characteristics of the basic feature sequence to obtain multi-layer features. Perform fast Fourier transform within each layer's time window to extract frequency components and phase information, and select the main frequency components according to the energy entropy distribution of the basic feature sequence. Input the main frequency components into a multi-head cross-attention network and perform feature fusion in combination with the temporal information of the basic feature sequence. Process the multi-layer features through a local feature extraction unit and a global feature fusion unit and optimize the feature combination path based on the time-varying characteristics of the basic feature sequence. The optimized features are processed through residual connection and a feed-forward network to obtain a temporal feature vector;

[0075] Calculate the correlation between nodes based on the temporal feature vector to construct a dynamic association graph and determine the connection relationship. Use a hierarchical message passing framework to iteratively update the dynamic association graph. Initialize the node state with the temporal feature vector and extract the structural features through gated graph convolution. Calculate the attention weights based on the similarity of node states to perform weighted aggregation update on the neighbor node information to obtain the node representation. Keep the multi-granularity features of the nodes through skip connections and fuse the updated node state features with the temporal feature vector. Finally, obtain a multi-modal feature matrix representing the overall state of the system through the processing of a feature transformation network.

[0076] Build a bidirectional feature extraction network, which consists of a forward propagation branch and a backward propagation branch. The main task of the forward propagation branch is to input real-time monitoring data into a dynamic buffer and divide the data stream into time segments through a sliding time window. Each time segment is processed by a separable convolutional layer that decomposes depth convolution and pointwise convolution to obtain multi-channel features. These multi-channel features will be used to construct a channel attention mechanism. The channel attention mechanism calculates weight coefficients according to the feature response intensity and the dependence relationship between channels, thereby obtaining weighted features.

[0077] The backpropagation branch constructs a mirror structure through a recurrent unit to correct the long-range dependence of the weighted features. The corrected bidirectional features are merged through a feature selection gating mechanism, and finally a basic feature sequence is obtained.

[0078] A pyramid hierarchical structure is constructed for the basic feature sequence. This structure dynamically adjusts the scale between layers according to the temporal variation characteristics of the basic feature sequence to form multi-layer features. Within the time window of each layer, the fast Fourier transform is performed to extract frequency components and phase information. According to the energy entropy distribution of the basic feature sequence, the main frequency components are selected and input into the multi-head cross-attention network. Feature fusion is performed by combining the temporal information of the basic feature sequence, and the multi-layer features are processed by a local feature extraction unit and a global feature fusion unit. Based on the time-varying characteristics of the basic feature sequence, the feature combination path is optimized. The optimized features are processed through residual connection and a feed-forward network, and finally a temporal feature vector is obtained.

[0079] Based on the temporal feature vector, the correlation between nodes is calculated to construct a dynamic association graph and the connection relationship is determined. A hierarchical message passing framework is used to iteratively update the dynamic association graph. The node state is initialized with the temporal feature vector, and the structural features are extracted through gated graph convolution. The attention weights are calculated based on the similarity of the node states, and the neighbor node information is weighted and aggregated for update to obtain the node representation. The multi-granularity features of the nodes are maintained through skip connections, and the updated node state features are fused with the temporal feature vector. After being processed by the feature transformation network, a multi-modal feature matrix representing the overall state of the system is finally obtained.

[0080] In this embodiment, through the design of the bidirectional feature extraction network, the temporal features and frequency features in the data can be captured more comprehensively, thereby improving the monitoring accuracy of the system. Through the multi-scale Fourier transform and feature fusion technology, the periodic features in different frequency domains can be effectively extracted, enhancing the system's ability to analyze complex signals. The construction of the dynamic association graph and the iterative update of the node states enable the system to reflect the state changes of the monitoring object in real time, providing a more accurate multi-modal feature matrix and providing a reliable basis for subsequent decision-making.

[0081] In an alternative embodiment,

[0082] Maintaining the multi-granularity features of the nodes through skip connections and fusing the updated node state features with the temporal feature vector, and finally obtaining a multi-modal feature matrix representing the overall state of the system through the processing of the feature transformation network includes:

[0083] Extract the original attributes and structural relationship information of nodes through a multi-scale skip connection structure. After being weighted by a feature selection gating unit, multi-scale features are obtained and mapped to a unified feature space for attention weighting to obtain aligned features, which are then interacted with temporal features. Combine a multi-branch feature transformation network to generate a multi-modal feature matrix, specifically including:

[0084] Construct a multi-scale skip connection structure. The multi-scale skip connection structure includes a shallow feature transfer path and a deep feature transfer path. Input the original attribute information of the nodes into the shallow feature transfer path, adjust the feature dimension to match the feature scale of the target layer through feature resampling, input the structural relationship information of the nodes into the deep feature transfer path, and reduce redundancy through feature compression to obtain high-order structural information. The outputs of the shallow feature transfer path and the deep feature transfer path respectively pass through a feature selection gating unit. The feature selection gating unit calculates the feature importance score based on the historical state information of the nodes, and dynamically weights the outputs of the shallow feature transfer path and the deep feature transfer path to obtain multi-scale features;

[0085] Map the node features at different levels in the multi-scale features to a unified feature space through a non-linear transformation. Use an attention calculation unit to construct a correlation matrix between features, generate the importance weight of each feature based on the correlation matrix, fuse the mapped features according to the importance weight, establish a residual connection between the fusion result and the original feature, and perform feature normalization to obtain aligned features;

[0086] Construct a bidirectional feature mapping network. The bidirectional feature mapping network includes a node feature mapping path and a temporal feature mapping path. Input the aligned features into the node feature mapping path, and extract the dependency relationship features between nodes through a multi-head attention module. Input the temporal data into the temporal feature mapping path, calculate the temporal feature weight based on periodicity and continuity, and adaptively fuse the output features of the node feature mapping path and the temporal feature mapping path to obtain interaction features;

[0087] Input the interaction features into a multi-branch feature transformation network. The multi-branch feature transformation network includes a spatial feature processing unit, a temporal feature processing unit, and a frequency feature processing unit. The spatial feature processing unit extracts local features through depth convolution and pointwise convolution decomposition. The temporal feature processing unit extracts dynamic features through a recurrent network with a selection gate and an update gate. The frequency feature processing unit extracts periodic features through signal decomposition. Calibrate and adaptively fuse the outputs of the three processing units to obtain a multi-modal feature matrix.

[0088] Construct a dual-pathway skip connection structure. For the shallow feature transfer pathway, the original attribute information of the receiving node is used as the input, and the feature dimension is adjusted through feature resampling operations to match the feature scale of the target layer. For the deep feature transfer pathway, the structural relationship information of the receiving node is used as the input, and a feature compression operation is performed to reduce information redundancy and extract high-order structural information. Then, a feature selection gating unit is designed, which reads the historical state information of the node, calculates the feature importance score, and performs a dynamic weighting operation on the outputs of the two pathways based on this score to finally generate multi-scale features;

[0089] Perform a non-linear transformation operation on the node features at different levels in the multi-scale features and map them to a unified feature space. Construct an attention calculation unit, which calculates the correlation matrix between features. Based on this correlation matrix, importance weights are generated for each feature. The mapped features are weighted and fused according to these weights. A residual connection is established between the fused result and the original input features, and feature normalization is performed to finally obtain aligned features;

[0090] Design a bidirectional feature mapping network with two mapping pathways. The node feature mapping pathway receives the aligned features as the input and extracts the dependency relationship features between nodes through the multi-head attention module. The temporal feature mapping pathway receives the temporal data as the input, analyzes the periodic and continuous features of the data, and calculates the temporal feature weights accordingly. Finally, an adaptive fusion operation is performed on the output features of the two pathways to generate interaction features;

[0091] Construct a multi-branch feature transformation network with three processing units. The spatial feature processing unit uses a combination of depth convolution and pointwise convolution to decompose and extract local features from the input features. The temporal feature processing unit constructs a recurrent network structure with a selection gate and an update gate to extract dynamically changing features. The frequency feature processing unit extracts periodic features through signal decomposition methods. Feature dimension calibration is performed on the outputs of the three processing units, and they are combined into the final multi-modal feature matrix through an adaptive fusion mechanism.

[0092] In this embodiment, the multi-scale skip connection structure and the feature selection gating mechanism can maintain the key information of the nodes at different levels, avoid information loss caused by deep networks, and improve the feature expression ability. The unified feature space mapping and attention weighted fusion mechanism realizes the alignment and effective fusion of different modal features, enhances the consistency and robustness of feature expression. The multi-branch feature transformation network extracts features from three dimensions of space, time series, and frequency, and combines with the adaptive fusion strategy to comprehensively capture the multi-modal information of the system state, improving the integrity and accuracy of feature expression.

[0093] S2. Input the multi-modal feature matrix into the causal inference network, construct the transfer function between indicators based on the structural equation model, establish a long short-term memory module through a gated recurrent unit, perform a temporal combination of the historical state information and the multi-modal feature matrix and construct a multi-head self-attention model, calculate the attention weights for the combined feature sequence, perform weighted aggregation of the attention weights and the corresponding features, construct a loss function in combination with the contrastive learning strategy, and optimize the feature representation based on the loss function and output a high-dimensional evaluation vector;

[0094] In an alternative embodiment,

[0095] Inputting the multi-modal feature matrix into the causal inference network, constructing the transfer function between indicators based on the structural equation model, establishing a long short-term memory module through a gated recurrent unit, performing a temporal combination of the historical state information and the multi-modal feature matrix and constructing a multi-head self-attention model, calculating the attention weights for the combined feature sequence, performing weighted aggregation of the attention weights and the corresponding features, constructing a loss function in combination with the contrastive learning strategy, and optimizing the feature representation based on the loss function and outputting a high-dimensional evaluation vector includes:

[0096] Perform block preprocessing on the multi-modal feature matrix, divide the features into multiple feature subsets according to the monitoring object type and physical attributes, construct an independent transfer function between indicators for each feature subset, the transfer function between indicators includes a feature mapping layer and a non-linear transformation layer, the feature mapping layer projects the input features into a high-dimensional feature space to obtain projection features, and the non-linear transformation layer performs activation transformation on the projection features to obtain initial transfer features, set a feature distribution normalization unit and a skip connection path between adjacent transfer function layers, and process the initial transfer features through the feature distribution normalization unit and the skip connection path to obtain optimized transfer features;

[0097] Input the optimized transfer features into the gated recurrent unit, the gated recurrent unit includes a state calculation unit and a forgetting gate unit, the state calculation unit fuses the optimized transfer features and the historical state information to generate a candidate state, and the forgetting gate unit selectively forgets the candidate state to obtain a basic memory feature. The middle-layer memory module introduces a multi-head self-attention model to perform importance weighting on the basic memory feature to obtain an enhanced memory feature, and the top-layer memory module adaptively modulates the enhanced memory feature based on the task constraints to obtain a modulated memory feature;

[0098] Adopt a dynamic window mechanism to segment the modulated memory feature, the window length is adaptively adjusted according to the data change characteristics, an overlapping area is set between adjacent windows to maintain the sequence continuity to obtain a segmented feature sequence, perform a temporal combination on the segmented feature sequence, the temporal combination includes performing temporal alignment on the segmented feature sequence to obtain aligned features, and performing spatio-temporal fusion on the aligned features to obtain a fused feature sequence;

[0099] Input the fused feature sequence into the multi-head self-attention model. At the first level, decouple the fused feature sequence into multiple independent sub-features and calculate the attention weights. At the second level, adaptively combine the attention weights to obtain a global attention vector. At the third level, reconstruct the fused feature sequence based on the global attention vector to obtain the attention features;

[0100] Perform sequence truncation transformation, feature masking transformation, and random perturbation transformation on the attention features to obtain an enhanced feature sequence. Input the enhanced feature sequence into a deep encoding network to extract the compressed representation. Based on the compressed representation, construct a loss function. The loss function generates positive and negative sample pairs and calculates the contrastive loss. Optimize the feature representation based on the contrastive loss and output a high-dimensional evaluation vector.

[0101] Perform block preprocessing on the input multi-modal feature matrix. Divide the features into a structural feature subset, an environmental feature subset, and a dynamic feature subset according to the physical characteristics of the monitoring object. Construct an independent transfer function between indicators for each feature subset. The transfer function includes a feature mapping layer and a non-linear transformation layer. The feature mapping layer projects the input features into a 512-dimensional feature space using a fully connected network to obtain the projected features. The non-linear transformation layer uses the ReLU activation function to transform the projected features to obtain the initial transfer features. Set a batch normalization unit between adjacent transfer function layers to normalize the feature distribution, and retain the original feature information through a residual connection to obtain the optimized transfer features.

[0102] Input the optimized transfer features into a gated recurrent unit for temporal modeling. The state calculation unit uses 128 hidden layer neurons to fuse the current input features and historical state information to generate a candidate state. The forget gate unit calculates the forgetting coefficient through the sigmoid function and selectively forgets the candidate state to obtain the basic memory features. The middle memory module uses an 8-head attention mechanism to calculate the attention weights for the basic memory features and perform importance weighting to obtain the enhanced memory features. The top memory module adaptively modulates the enhanced memory features based on the task constraints.

[0103] Use a dynamic window mechanism to segment the modulated memory features. The window length is dynamically adjusted between 32 and 128, and the overlapping rate of adjacent windows is set to 0.5. Align the segmented feature sequence in time series, calculate the alignment matrix using the bidirectional dynamic time warping algorithm, and rearrange the features based on the alignment matrix to obtain the aligned features. Perform spatio-temporal fusion on the aligned features through a spatio-temporal convolutional network to obtain a fused feature sequence.

[0104] Input the fused feature sequence into the multi-head attention model. The first layer decouples the sequence into 8 sub-features of 64 dimensions, calculates the query-key-value attention to obtain the attention weights. The second layer performs weighted average on the attention weights to obtain the global attention vector. The third layer reconstructs the fused feature sequence based on the global attention vector to obtain the attention features.

[0105] Perform data augmentation on the attention features. The sequence truncation transformation randomly extracts subsequences with lengths 0.8 - 1.0 times that of the original sequence. The feature masking transformation randomly sets 30% of the features to zero. The random perturbation transformation adds Gaussian noise with a mean of 0 and a standard deviation of 0.1 to the features. Input the augmented feature sequence into a 5-layer encoding network to extract a 256-dimensional compressed representation. Based on the compressed representation, construct an InfoNCE loss function, generate positive and negative sample pairs, and calculate the contrastive loss. Optimize the feature representation through backpropagation and output a 512-dimensional evaluation vector.

[0106] In this embodiment, by constructing an inter-index transfer function to perform block modeling on multi-modal features, combining feature distribution normalization and skip connections to optimize the feature transfer process, the feature representation ability is improved. The gated recurrent unit and multi-layer memory module are used to capture temporal dependencies, and the multi-head attention mechanism is introduced to perform feature importance weighting, enhancing the model's ability to model long-term dependencies. Based on the dynamic window and temporal combination, flexible segmentation and alignment fusion of the feature sequence are realized, and the contrastive learning strategy is combined to optimize the feature representation, improving the generalization performance and robustness of the model.

[0107] In an alternative implementation

[0108] Performing the sequence truncation transformation, feature masking transformation, and random perturbation transformation on the attention features to obtain an augmented feature sequence, and inputting the augmented feature sequence into a deep encoding network to extract the compressed representation includes:

[0109] Construct a three-channel parallel enhancement processing architecture for the attention features, and perform the sequence truncation transformation, feature masking transformation, and random perturbation transformation respectively. The sequence truncation transformation uses a three-stage truncation window to extract multi-scale features. The feature masking transformation performs hierarchical masking based on the feature importance. The random perturbation transformation performs differential perturbation according to the feature correlation, and performs dynamic fusion based on information entropy and signal-to-noise ratio to obtain the augmented feature sequence. Input the augmented feature sequence into a deep encoding network to extract the compressed representation, specifically including:

[0110] Construct a parallel enhancement processing unit, which is provided with a sequence truncation transformation channel, a feature masking transformation channel, and a random perturbation transformation channel. Synchronously input the attention features into the parallel enhancement processing unit. The sequence truncation transformation channel is provided with three levels of truncation windows. The first-level truncation window extracts the long-period variation information in the attention features to obtain the main structure feature sequence. The second-level truncation window extracts the medium-period variation information in the attention features to obtain the local feature sequence. The third-level truncation window extracts the short-period variation information in the attention features to obtain the transient feature sequence. Calculate the frequency domain distribution characteristics of the main structure feature sequence, the local feature sequence, and the transient feature sequence respectively. Determine the sequence fusion weights according to the frequency domain distribution characteristics. Combine the three feature sequences based on the sequence fusion weights to obtain a multi-scale feature sequence;

[0111] The feature masking transformation channel calculates the information gain of the attention features to obtain the feature importance. Divide the attention features into three importance levels based on the feature importance. The first-level features use the minimum masking ratio and the shortest masking duration. The second-level features use the medium masking ratio and the medium masking duration. The third-level features use the maximum masking ratio and the longest masking duration. Perform masking processing on the three-level features to obtain a masked feature sequence;

[0112] The random perturbation transformation channel calculates the distance between pairwise attention features to obtain a correlation matrix. Construct a minimum spanning tree based on the correlation matrix. Group the features with distances less than the threshold in the minimum spanning tree. Apply the same perturbation amplitude and perturbation frequency to the features in the same group, and apply different perturbation amplitudes and perturbation frequencies to the features in different groups. Calculate the first derivative of the attention features to obtain the change rate. Dynamically adjust the perturbation amplitude according to the change rate to obtain a perturbed feature sequence;

[0113] Construct a feature fusion unit. The feature fusion unit calculates the probability distributions of the multi-scale feature sequence, the masked feature sequence, and the perturbed feature sequence respectively to obtain the information entropy. Calculate the similarity between the multi-scale feature sequence, the masked feature sequence, and the perturbed feature sequence and the original features to obtain the signal-to-noise ratio. Determine the feature fusion weights according to the information entropy and the signal-to-noise ratio. Combine the three feature sequences based on the feature fusion weights to obtain an enhanced feature sequence;

[0114] Construct a deep encoding network. A cross-layer feature transmission path is set between adjacent encoding layers. Each encoding layer is configured with a feature selection module. The feature selection module calculates response scores for the input features and filters out high-response features. A feature restoration module is set at the end of the deep encoding network. The feature restoration module maps the encoded features back to the original domain to obtain reconstructed features, calculates the deviation between the reconstructed features and the original features to obtain a reconstruction loss, and optimizes the parameters of the deep encoding network according to the reconstruction loss to obtain a compressed representation.

[0115] Construct a parallel enhancement processing unit, which includes a sequence truncation transformation channel, a feature masking transformation channel, and a random perturbation transformation channel. Process the input attention features, where the dimension of the attention features is 256 and the length of the time series is 1024.

[0116] The sequence truncation transformation channel sets three levels of truncation windows. The length of the first-level truncation window is 512, which is used to extract long-period change information to obtain a main structure feature sequence; the length of the second-level truncation window is 256, which extracts medium-period change information to obtain a local feature sequence; the length of the third-level truncation window is 128, which extracts short-period change information to obtain a transient feature sequence. Calculate the frequency-domain distribution characteristics of the three feature sequences respectively, determine the sequence fusion weights to be 0.5, 0.3, and 0.2 according to the frequency-domain energy distribution, and weighted-combine the three feature sequences to obtain a multi-scale feature sequence.

[0117] The feature masking transformation channel calculates the information gain of the attention features to obtain feature importance. The features are divided into three levels according to the importance: those with an importance greater than 0.8 are in the first level, with a masking ratio of 0.1 and a masking duration of 32; those with an importance between 0.5 and 0.8 are in the second level, with a masking ratio of 0.3 and a masking duration of 64; those with an importance less than 0.5 are in the third level, with a masking ratio of 0.5 and a masking duration of 128. Mask the features at the three levels to obtain a masked feature sequence.

[0118] The random perturbation transformation channel calculates the Euclidean distance between pairs of attention features to obtain a correlation matrix. Based on the correlation matrix, construct a minimum spanning tree, and group the features with a distance less than 0.3 into the same group. Apply the same perturbation to the features in the same group: the amplitude is 0.1 and the frequency is 0.01; apply a different perturbation to the features in different groups: the amplitude is between 0.05 and 0.15, and the frequency is between 0.005 and 0.015. Calculate the first-order derivative of the attention features to obtain a change rate, and dynamically adjust the perturbation amplitude according to the change rate to obtain a perturbed feature sequence.

[0119] The feature fusion unit calculates the information entropy and signal-to-noise ratio of the three feature sequences respectively. The larger the information entropy, the greater the amount of information, and the higher the signal-to-noise ratio, the better the feature quality. According to the information entropy and signal-to-noise ratio, the feature fusion weights are determined to be 0.4, 0.35, and 0.25 respectively, and the three feature sequences are weighted and combined to obtain an enhanced feature sequence.

[0120] The deep encoding network contains 4 encoding layers, and the number of neurons in each layer is 1024, 512, 256, and 128 in sequence. Cross-layer feature transmission paths are set between adjacent encoding layers. Each encoding layer is configured with a feature selection module to calculate the feature response scores and retain the top 50% of the features with the highest scores. The feature restoration module at the end of the network maps the 128-dimensional encoded features back to the 1024-dimensional original domain to obtain the reconstructed features, calculates the mean square error between the reconstructed features and the original features as the reconstruction loss, and optimizes the network parameters through backpropagation to obtain the compressed representation.

[0121] In this embodiment, through the three-channel parallel enhancement architecture and multi-scale feature extraction, the richness and robustness of feature representation are improved, the model's ability to capture information at different scales is enhanced, based on hierarchical masking of feature importance and correlation-driven differential perturbation, the adaptive adjustment of feature enhancement is realized, the pertinence and effectiveness of data enhancement are improved, and the deep encoding network and cross-layer feature transmission are adopted to achieve efficient compression while maintaining the feature expression ability, reducing the storage overhead and improving the calculation efficiency.

[0122] S3. Apply kernel principal component analysis to the high-dimensional evaluation vector for dimensionality reduction, input the dimensionality-reduced evaluation vector into the density-based spatial clustering algorithm, divide the evaluation levels according to the density reachability principle, construct a hierarchical reinforcement learning framework, the high-level policy network plans the optimization target sequence based on the current system state, the low-level execution network decomposes the target sequence into specific tuning parameters, samples and explores the parameter space through Monte Carlo tree search, and iteratively calculates the optimal parameter combination in combination with Bayesian optimization and outputs the improvement plan.

[0123] In an alternative embodiment,

[0124] Applying kernel principal component analysis to the high-dimensional evaluation vector for dimensionality reduction, inputting the dimensionality-reduced evaluation vector into the density-based spatial clustering algorithm, dividing the evaluation levels according to the density reachability principle, constructing a hierarchical reinforcement learning framework, the high-level policy network planning the optimization target sequence based on the current system state, the low-level execution network decomposing the target sequence into specific tuning parameters, sampling and exploring the parameter space through Monte Carlo tree search, and iteratively calculating the optimal parameter combination in combination with Bayesian optimization and outputting the improvement plan includes:

[0125] Construct a kernel function for the high-dimensional evaluation vector, calculate the kernel similarity matrix through the kernel function, project the high-dimensional evaluation vector into the feature space using the kernel similarity matrix, calculate the sample covariance matrix in the feature space, perform eigenvalue decomposition on the sample covariance matrix to obtain an eigenvalue sequence and an eigenvector sequence, sort the eigenvalue sequence in descending order and calculate the cumulative contribution rate, select the main eigenvectors according to the preset contribution rate threshold to construct a dimensionality reduction projection matrix, and perform kernel principal component analysis dimensionality reduction on the high-dimensional evaluation vector using the dimensionality reduction projection matrix;

[0126] Calculate the distance matrix for all sample points in the dimensionality-reduced evaluation vector, statistically analyze the distance distribution characteristics between samples based on the distance matrix, determine the density calculation radius through the distance distribution characteristics, count the number of neighborhood samples within the density calculation radius for each sample point to obtain the local density value, identify the density peak points based on the local density value and set them as cluster centers, start from the cluster centers and sequentially add adjacent low-density samples to the cluster to form a density-decreasing cluster chain, and stop cluster expansion when the density difference between adjacent samples exceeds the preset threshold, and divide the evaluation levels based on the principle of density reachability;

[0127] Construct a hierarchical reinforcement learning framework based on the evaluation levels. The high-level policy network performs multi-scale analysis on the current system state through the state feature extraction module to obtain the state feature representation. The state feature representation is input into the target planning module, and the target planning module generates an optimization target sequence. The low-level execution network receives the optimization target sequence, uses the hierarchical decomposition method to transform the optimization target sequence into tuning parameters, and sets a value evaluation module between the two networks to achieve information interaction;

[0128] Construct a Monte Carlo search tree structure for the parameter space. The search tree nodes record the candidate parameter groups and evaluation scores. The search tree edges represent the parameter adjustment directions. Calculate the upper bound of the node value based on the historical scores and visit counts of the nodes, select the node with the maximum upper bound of the value for expansion, generate child nodes under the node with the maximum upper bound of the value, randomly sample the parameters of the child nodes and calculate the evaluation scores, and propagate the evaluation scores from the leaf nodes to the root node to update the path statistical information;

[0129] Establish a Bayesian optimization model, input the historical parameter solutions and effect scores as training samples into the Bayesian optimization model, use the Bayesian optimization model to predict the effects and uncertainty levels of the parameter solutions to be evaluated, design a combined acquisition function to comprehensively balance the predicted effects and uncertainty levels, select the parameter solution corresponding to the maximum acquisition function value for actual evaluation, add the new evaluation results to the training samples to update the Bayesian optimization model, and output the optimal parameter combination and generate an improvement plan through iterative calculations.

[0130] Construct a kernel function to process the high-dimensional evaluation vectors. The selection of the kernel function should consider the characteristics of the data. Commonly used kernel functions include the Gaussian kernel and the polynomial kernel. Through the selected kernel function, a kernel similarity matrix is calculated, which reflects the similarity between samples. Then, the high-dimensional evaluation vectors are projected into the feature space using the kernel similarity matrix. In the feature space, the covariance matrix of the samples is calculated and eigen-decomposed to obtain eigenvalues and eigenvectors. The eigenvalues are sorted in descending order, and their cumulative contribution rates are calculated to determine the principal eigenvectors. According to a preset contribution rate threshold, the principal eigenvectors are selected to construct a dimensionality reduction projection matrix, and finally, the high-dimensional evaluation vectors are subjected to kernel principal component analysis using this projection matrix to complete dimensionality reduction.

[0131] Calculate the distance matrix for all sample points in the dimensionality-reduced evaluation vectors. The distance matrix is used to statistically analyze the distance distribution characteristics between samples, and then determine the density calculation radius. By counting the number of neighboring samples within the density calculation radius for each sample point, the local density value is obtained. Based on the local density value, density peak points are identified and set as clustering centers. Starting from the clustering centers, adjacent low-density samples are successively added to the cluster to form a clustering chain with decreasing density. When the density difference between adjacent samples exceeds a preset threshold, the clustering expansion stops. Finally, the evaluation levels are divided based on the principle of density reachability.

[0132] Construct a hierarchical reinforcement learning framework based on the evaluation levels. The high-level policy network performs multi-scale analysis on the current system state through the state feature extraction module to obtain a state feature representation. This state feature representation is input into the target planning module to generate an optimization target sequence. The low-level execution network receives the optimization target sequence and uses a hierarchical decomposition method to transform it into specific tuning parameters. A value evaluation module is set between the two networks to achieve effective information interaction.

[0133] Construct a Monte Carlo search tree structure for the parameter space. The nodes of the search tree record the candidate parameter groups and their evaluation scores, and the edges represent the parameter adjustment directions. Based on the historical scores and visit counts of the nodes, the upper bound of the value of the nodes is calculated, and the node with the maximum upper bound of the value is selected for expansion. Sub-nodes are generated under this node, and random sampling is performed on the parameters of the sub-nodes and their evaluation scores are calculated. The evaluation scores are propagated from the leaf nodes to the root nodes to update the statistical information of the paths.

[0134] Establish a Bayesian optimization model and input the historical parameter solutions and effect scores as training samples into the model. Use the Bayesian optimization model to predict the effects and uncertainty levels of the parameter solutions to be evaluated. Design a combined acquisition function to comprehensively balance the predicted effects and uncertainty levels, and select the parameter solution corresponding to the maximum acquisition function value for actual evaluation. Add the new evaluation results to the training samples to update the Bayesian optimization model. Through repeated iterative calculations, the optimal parameter combination is output and an improved solution is generated.

[0135] In this embodiment, effective dimensionality reduction of high-dimensional data is achieved through kernel principal component analysis, important features are retained, the computational complexity is reduced, and the efficiency of subsequent analysis is improved. The density-based clustering method can effectively identify the natural clustering structure in the data, improving the accuracy and reliability of the evaluation level. The hierarchical reinforcement learning framework combines Monte Carlo tree search and Bayesian optimization, which can adaptively adjust parameters in a dynamic environment, optimize the decision-making process, and improve the overall performance of the system.

[0136] In an alternative embodiment,

[0137] Using the Bayesian optimization model to predict the effects and uncertainties of the parameter schemes to be evaluated, designing a combined acquisition function to comprehensively balance the prediction effects and uncertainties, and selecting the parameter scheme corresponding to the maximum acquisition function value for actual evaluation includes:

[0138] Using the Bayesian optimization model for prediction, partitioning the parameter space to perform local and global dual-scale predictions, constructing an acquisition function through non-dominated ranking and crowding degree, calculating the membership relationship of the parameter distribution family by combining the hierarchical clustering method, introducing a multi-scheme parallel evaluation mechanism, and implementing local incremental updates on the evaluation data, specifically including:

[0139] Constructing a Bayesian optimization model for historical parameter schemes and evaluation data. The Bayesian optimization model automatically adjusts the kernel function parameters according to the data density distribution, divides the complete parameter space into multiple continuous sub-regions, performs local predictions respectively within the multiple continuous sub-regions to obtain local prediction results, merges the local prediction results based on boundary constraint conditions to obtain global prediction results, calculates the local neighborhood prediction effects and global range prediction effects for each parameter scheme to be evaluated respectively, and calculates the uncertainty degree by combining the sample distribution density;

[0140] Constructing the prediction effects and the uncertainty degree as a bi-objective optimization problem, calculating the non-dominated relationship of the parameter schemes on the two objectives to obtain the dominant rank sequence, statistically obtaining the crowding degree index of the parameter schemes in the objective space, combining the dominant rank sequence and the crowding degree index to construct a combined acquisition function, recording the coverage density of historical evaluation schemes in the parameter space, and imposing a scoring penalty on the high-frequency exploration area according to the coverage density;

[0141] Establishing a similarity calculation criterion for historical parameter schemes, constructing a distance matrix based on the similarity calculation criterion, using the distance matrix to perform hierarchical clustering division on the parameter schemes to obtain a set of parameter distribution families, calculating the membership relationship between the scheme to be evaluated and each family in the parameter distribution family set, and weighted superimposing the membership relationship to the combined acquisition function;

[0142] Construct a measure index in the parameter space, calculate the distance difference and prediction feature difference between candidate solutions using the measure index, select multiple candidate solutions with the largest difference degrees based on the distance difference and the prediction feature difference to form an evaluation set, perform actual evaluations on the candidate solutions in the evaluation set synchronously, and analyze the correlation degree between the actual evaluation results;

[0143] Input the newly obtained evaluation data into the Bayesian optimization model, identify the model parameters within the influence range of the evaluation data, only perform local updates on the parameters within the influence range, establish a criterion for judging the effectiveness of the evaluation results, and screen out abnormal evaluation data according to the criterion;

[0144] Dynamically determine the number of parallel evaluation solutions according to the optimization stage, select multiple candidate solutions for parallel evaluation in the parameter space exploration stage, reduce the number of parallel evaluations in the parameter optimization stage, calculate the adjustment step size of the target weight based on the historical evaluation effect, and adaptively update the combined weight of the prediction effect and the uncertainty degree using the adjustment step size to obtain new evaluation results.

[0145] Process the historical parameter solutions and evaluation data. By analyzing the data distribution characteristics, use the radial basis function as the kernel function, and the kernel function parameters are adaptively adjusted according to the local data density. Divide the complete parameter space into multiple continuous sub-regions, and each sub-region contains similar parameter solutions. Perform local predictions within each sub-region to obtain local prediction results. Based on the boundary continuity constraint, merge the local prediction results into a global prediction result. For the solution to be evaluated, calculate the prediction effects in the local neighborhood and the global scope respectively, and calculate the uncertainty degree in combination with the sample distribution density.

[0146] Construct a two-objective optimization problem, where the objectives include the prediction effect and the uncertainty degree. Obtain the dominance rank sequence of the solutions through non-dominated sorting, and calculate the crowding degree index of the solutions in the objective space. Combine the dominance rank and the crowding degree to construct an acquisition function. Statistically analyze the coverage density of the historical solutions in the parameter space, and impose a scoring penalty on the high-frequency exploration areas.

[0147] Establish a criterion for calculating the similarity of parameter solutions, and use the Euclidean distance to construct a distance matrix. Perform hierarchical clustering based on the distance matrix to obtain a set of parameter distribution families. Calculate the membership relationship between the solution to be evaluated and each distribution family, and weightedly superimpose the membership relationship onto the acquisition function. For example, if the membership degrees of a solution to be evaluated with three distribution families are 0.3, 0.5, and 0.2 respectively, then the acquisition function value needs to be multiplied by the corresponding weights.

[0148] Construct a measure index in the parameter space, calculate the distance difference and prediction feature difference between candidate solutions. Select multiple candidate solutions with the largest difference degrees to form an evaluation set, and perform actual evaluations synchronously. Analyze the correlation degree of the evaluation results, and a correlation coefficient lower than 0.3 indicates that the solutions have good differences.

[0149] For the newly acquired evaluation data, identify its influence scope in the model. Only perform local updates on the model parameters within the influence scope to improve the calculation efficiency. Establish a criterion for judging the validity of the evaluation results and eliminate abnormal data. For example, if the evaluation result exceeds twice the range of the historical data, it is regarded as abnormal.

[0150] Select five to ten candidate solutions for parallel evaluation in the parameter space exploration stage, and reduce it to two to three in the optimization stage. Calculate the adjustment step size of the target weight based on the historical evaluation effect, adaptively update the combined weight of the prediction effect and the degree of uncertainty, and obtain the new evaluation result.

[0151] In this embodiment, a local and global dual-scale prediction strategy is adopted, which improves the prediction accuracy and reduces the calculation overhead. The acquisition function is constructed through non-dominated sorting and crowding degree calculation, realizing an effective trade-off between the prediction effect and the degree of uncertainty. The evaluation scheme is selected based on the distance difference and the prediction feature difference, ensuring the difference of the parallel evaluation schemes. By dynamically adjusting the number of parallel evaluations and the target weight, the reasonable allocation of evaluation resources is realized, and the optimization convergence speed is accelerated.

[0152] Figure 2 It is a schematic structural diagram of the lightning protection system effectiveness evaluation system based on multi-dimensional research and judgment according to the embodiment of the present invention, as Figure 2 shown, the system includes:

[0153] The first unit is used to construct a bidirectional feature extraction network to collect real-time monitoring data of system indicators, obtain the basic feature sequence through sliding sampling by a separable convolutional layer with an attention mechanism, perform multi-scale Fourier transform on the basic feature sequence to obtain periodic features in different frequency domains, fuse the periodic features and the basic feature sequence through a cross-frequency domain attention network to generate a time series feature vector, construct a dynamic association graph based on the time series feature vector, and combine the graph neural network to iteratively update the node state and output a multi-modal feature matrix representing the overall state of the system;

[0154] The second unit is used to input the multi-modal feature matrix into the causal inference network, construct a transfer function between indicators based on the structural equation model, establish a long-term and short-term memory module through a gated recurrent unit, perform time series combination on the historical state information and the multi-modal feature matrix and construct a multi-head self-attention model, calculate the attention weight for the combined feature sequence, perform weighted aggregation on the attention weight and the corresponding features, construct a loss function in combination with the contrast learning strategy, optimize the feature representation based on the loss function and output a high-dimensional evaluation vector;

[0155] The third unit is used to apply kernel principal component analysis to the high-dimensional evaluation vector for dimensionality reduction, input the dimensionality-reduced evaluation vector into a density-based spatial clustering algorithm, divide the evaluation levels according to the density reachability principle, construct a hierarchical reinforcement learning framework, where the high-level policy network plans an optimal target sequence based on the current system state, the low-level execution network decomposes the target sequence into specific tuning parameters, samples and explores the parameter space through Monte Carlo tree search, combines Bayesian optimization to iteratively calculate the optimal parameter combination and outputs an improvement plan.

[0156] In the third aspect of the embodiments of the present invention,

[0157] A provided electronic device includes:

[0158] A processor;

[0159] A memory for storing instructions executable by the processor;

[0160] Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0161] In the fourth aspect of the embodiments of the present invention,

[0162] A provided computer-readable storage medium stores computer program instructions, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0163] The present invention may be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions for performing various aspects of the present invention loaded thereon.

[0164] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than limiting them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An effectiveness evaluation method for a lightning protection system based on multi-dimensional research and judgment, characterized in that Including: Construct a bidirectional feature extraction network to collect real-time monitoring data of system metrics. Perform sliding sampling on the obtained basic feature sequence through a separable convolutional layer with an attention mechanism, perform multi-scale Fourier transform on the basic feature sequence to obtain periodic features in different frequency domains, fuse the periodic features and the basic feature sequence through a cross-frequency domain attention network to generate a time-series feature vector, construct a dynamic association graph based on the time-series feature vector, and iteratively update the node states in combination with a graph neural network to output a multi-modal feature matrix representing the overall state of the system; Input the multi-modal feature matrix into a causal inference network, construct a transfer function between metrics based on a structural equation model, establish a long short-term memory module through a gated recurrent unit, perform temporal combination on the historical state information and the multi-modal feature matrix and construct a multi-head self-attention model, calculate attention weights for the combined feature sequence, perform weighted aggregation on the attention weights and the corresponding features, construct a loss function in combination with a contrastive learning strategy, and optimize the feature representation based on the loss function and output a high-dimensional evaluation vector; Apply kernel principal component analysis to the high-dimensional evaluation vector for dimensionality reduction, input the dimensionality-reduced evaluation vector into a density-based spatial clustering algorithm, divide the evaluation levels according to the density reachability principle, construct a hierarchical reinforcement learning framework, where the high-level policy network plans an optimization target sequence based on the current system state, the low-level execution network decomposes the target sequence into specific tuning parameters, samples and explores the parameter space through Monte Carlo tree search, iteratively calculates the optimal parameter combination in combination with Bayesian optimization and outputs an improvement plan.

2. The method according to claim 1, wherein Construct a bidirectional feature extraction network to collect real-time monitoring data of system metrics. Perform sliding sampling on the obtained basic feature sequence through a separable convolutional layer with an attention mechanism, perform multi-scale Fourier transform on the basic feature sequence to obtain periodic features in different frequency domains, fuse the periodic features and the basic feature sequence through a cross-frequency domain attention network to generate a time-series feature vector, construct a dynamic association graph based on the time-series feature vector, and iteratively update the node states in combination with a graph neural network to output a multi-modal feature matrix representing the overall state of the system, including: Establish a bidirectional feature extraction network, where the bidirectional feature extraction network includes a forward propagation branch and a backward propagation branch. The forward propagation branch inputs the real-time monitoring data into a dynamic buffer and divides the data stream into time segments through a sliding time window, processes the time segments with a separable convolutional layer that decomposes depth convolution and pointwise convolution to obtain multi-channel features, constructs a channel attention mechanism based on the multi-channel features, calculates weight coefficients according to the feature response intensity and channel dependence relationship and obtains weighted features. The backward propagation branch constructs a mirror structure through a recurrent unit to perform long-range dependence correction on the weighted features, and merges the corrected bidirectional features through a feature selection gating mechanism to obtain a basic feature sequence; Construct a pyramid hierarchical structure for the basic feature sequence and dynamically adjust the inter-layer scale based on the temporal variation characteristics of the basic feature sequence to obtain multi-layer features. Perform fast Fourier transform within each layer's time window to extract frequency components and phase information, and select the main frequency components according to the energy entropy distribution of the basic feature sequence. Input the main frequency components into a multi-head cross-attention network and perform feature fusion by combining the temporal information of the basic feature sequence. Process the multi-layer features through a local feature extraction unit and a global feature fusion unit, and optimize the feature combination path based on the time-varying characteristics of the basic feature sequence. Process the optimized features through residual connection and a feed-forward network to obtain a temporal feature vector; Calculate the correlation between nodes based on the temporal feature vector to construct a dynamic association graph and determine the connection relationship. Use a hierarchical message passing framework to iteratively update the dynamic association graph. Initialize the node state with the temporal feature vector and extract structural features through gated graph convolution. Calculate the attention weights based on the similarity of node states to perform weighted aggregation and update of neighbor node information to obtain node representations. Maintain the multi-granularity features of the nodes through skip connections and fuse the updated node state features with the temporal feature vector. Finally, obtain a multi-modal feature matrix representing the overall state of the system through the processing of a feature transformation network.

3. The method according to claim 2, characterized in that Maintain the multi-granularity features of the nodes through skip connections and fuse the updated node state features with the temporal feature vector. Finally, obtain a multi-modal feature matrix representing the overall state of the system through the processing of a feature transformation network, including: Extract the original attributes and structural relationship information of the nodes through a multi-scale skip connection structure. After being weighted by a feature selection gating unit, obtain multi-scale features and map them to a unified feature space for attention weighting to obtain aligned features. Interact the aligned features with the temporal features and generate a multi-modal feature matrix in combination with a multi-branch feature transformation network, specifically including: Construct a multi-scale skip connection structure, which includes a shallow feature transfer path and a deep feature transfer path. Input the original attribute information of the nodes into the shallow feature transfer path, adjust the feature dimension through feature resampling to match the feature scale of the target layer. Input the structural relationship information of the nodes into the deep feature transfer path, reduce redundancy through feature compression to obtain high-order structural information. The outputs of the shallow feature transfer path and the deep feature transfer path respectively pass through a feature selection gating unit, and the feature selection gating unit calculates the feature importance score based on the historical state information of the nodes, and dynamically weights the outputs of the shallow feature transfer path and the deep feature transfer path to obtain multi-scale features; Map the node features at different levels in the multi-scale features to a unified feature space through non-linear transformation. Use an attention calculation unit to construct a correlation matrix between features, generate the importance weights of each feature based on the correlation matrix, fuse the mapped features according to the importance weights, establish a residual connection between the fusion result and the original features, and perform feature normalization to obtain aligned features; Construct a bidirectional feature mapping network, where the bidirectional feature mapping network includes a node feature mapping path and a temporal feature mapping path. Input the aligned features into the node feature mapping path, extract the dependency relationship features between nodes through a multi-head attention module, input the temporal data into the temporal feature mapping path, calculate the temporal feature weights based on periodicity and continuity, and adaptively fuse the output features of the node feature mapping path and the temporal feature mapping path to obtain interaction features; Input the interaction features into a multi-branch feature transformation network. The multi-branch feature transformation network includes a spatial feature processing unit, a temporal feature processing unit, and a frequency feature processing unit. The spatial feature processing unit extracts local features through depthwise convolution and pointwise convolution decomposition. The temporal feature processing unit extracts dynamic features through a recurrent network with a selection gate and an update gate. The frequency feature processing unit extracts periodic features through signal decomposition. Feature calibration and adaptive fusion are performed on the outputs of the three processing units to obtain a multi-modal feature matrix.

4. The method according to claim 1, wherein Input the multi-modal feature matrix into a causal inference network. Based on the structural equation model, construct a transfer function between indicators. Establish a long short-term memory module through a gated recurrent unit. Perform temporal combination of the historical state information and the multi-modal feature matrix and construct a multi-head self-attention model. Calculate the attention weights for the combined feature sequence, perform weighted aggregation of the attention weights and the corresponding features, and construct a loss function in combination with the contrastive learning strategy. Optimize the feature representation based on the loss function and output a high-dimensional evaluation vector, including: Perform block preprocessing on the multi-modal feature matrix. Divide the features into multiple feature subsets according to the monitoring object type and physical attributes. Construct an independent transfer function between indicators for each feature subset. The transfer function between indicators includes a feature mapping layer and a non-linear transformation layer. The feature mapping layer projects the input features into a high-dimensional feature space to obtain projected features. The non-linear transformation layer performs activation transformation on the projected features to obtain initial transfer features. A feature distribution normalization unit and a skip connection path are set between adjacent transfer function layers. The initial transfer features are processed through the feature distribution normalization unit and the skip connection path to obtain optimized transfer features; Input the optimized transfer features into a gated recurrent unit. The gated recurrent unit includes a state calculation unit and a forgetting gate unit. The state calculation unit fuses the optimized transfer features and the historical state information to generate a candidate state. The forgetting gate unit selectively forgets the candidate state to obtain a basic memory feature. A middle-level memory module introduces a multi-head self-attention model to perform importance weighting on the basic memory feature to obtain an enhanced memory feature. A top-level memory module adaptively modulates the enhanced memory feature based on task constraints to obtain a modulated memory feature; The modulation memory features are segmented using a dynamic window mechanism, where the window length is adaptively adjusted according to the data change characteristics. An overlapping area is set between adjacent windows to maintain sequence continuity, resulting in a segmented feature sequence. The segmented feature sequence is subjected to temporal combination, which includes temporal alignment of the segmented feature sequence to obtain aligned features, and spatio-temporal fusion of the aligned features to obtain a fused feature sequence; The fused feature sequence is input into a multi-head self-attention model. At the first level, the fused feature sequence is decoupled into multiple independent sub-features and attention weights are calculated. At the second level, the attention weights are adaptively combined to obtain a global attention vector. At the third level, the fused feature sequence is reconstructed based on the global attention vector to obtain attention features; Sequence truncation transformation, feature masking transformation, and random perturbation transformation are applied to the attention features to obtain an enhanced feature sequence. The enhanced feature sequence is input into a deep encoding network to extract a compressed representation. Based on the compressed representation, a loss function is constructed. The loss function generates positive and negative sample pairs and calculates a contrastive loss. The feature representation is optimized based on the contrastive loss, and a high-dimensional evaluation vector is output.

5. The method according to claim 4, wherein Sequence truncation transformation, feature masking transformation, and random perturbation transformation are applied to the attention features to obtain an enhanced feature sequence. Inputting the enhanced feature sequence into a deep encoding network to extract a compressed representation includes: Construct a three-channel parallel enhancement processing architecture for the attention features, and perform sequence truncation transformation, feature masking transformation, and random perturbation transformation respectively. The sequence truncation transformation uses a three-level truncation window to extract multi-scale features. The feature masking transformation performs hierarchical masking based on feature importance. The random perturbation transformation implements differential perturbation according to feature correlation, and performs dynamic fusion based on information entropy and signal-to-noise ratio to obtain an enhanced feature sequence. The enhanced feature sequence is input into a deep encoding network to extract a compressed representation, which specifically includes: Construct a parallel enhancement processing unit. The parallel enhancement processing unit is provided with a sequence truncation transformation channel, a feature masking transformation channel, and a random perturbation transformation channel. The attention features are synchronously input into the parallel enhancement processing unit. The sequence truncation transformation channel is provided with a three-level truncation window. The first-level truncation window extracts the long-period change information in the attention features to obtain a main structure feature sequence. The second-level truncation window extracts the medium-period change information in the attention features to obtain a local feature sequence. The third-level truncation window extracts the short-period change information in the attention features to obtain a transient feature sequence. The frequency domain distribution characteristics of the main structure feature sequence, the local feature sequence, and the transient feature sequence are calculated respectively. The sequence fusion weights are determined according to the frequency domain distribution characteristics. The three feature sequences are combined based on the sequence fusion weights to obtain a multi-scale feature sequence; The feature masking transformation channel calculates the information gain of the attention features to obtain the feature importance. Based on the feature importance, the attention features are divided into three importance levels. The first-level features use the minimum masking ratio and the shortest masking duration, the second-level features use the medium masking ratio and the medium masking duration, and the third-level features use the maximum masking ratio and the longest masking duration. Masking processing is performed on the three-level features to obtain a masked feature sequence; The random perturbation transformation channel calculates the distance between pairwise attention features to obtain a correlation matrix. Based on the correlation matrix, a minimum spanning tree is constructed. Features with distances less than a threshold in the minimum spanning tree are grouped into the same group. The same perturbation amplitude and perturbation frequency are applied to the features in the same group, and different perturbation amplitudes and perturbation frequencies are applied to the features in different groups. The first derivative of the attention features is calculated to obtain the change rate, and the perturbation amplitude is dynamically adjusted according to the change rate to obtain a perturbed feature sequence; A feature fusion unit is constructed. The feature fusion unit calculates the probability distribution of the multi-scale feature sequence, the masked feature sequence, and the perturbed feature sequence respectively to obtain the information entropy, calculates the similarity between the multi-scale feature sequence, the masked feature sequence, and the perturbed feature sequence and the original features to obtain the signal-to-noise ratio. The feature fusion weights are determined according to the information entropy and the signal-to-noise ratio, and the three feature sequences are combined based on the feature fusion weights to obtain an enhanced feature sequence; A deep encoding network is constructed. The deep encoding network sets a cross-layer feature transmission path between adjacent encoding layers. Each encoding layer is configured with a feature selection module. The feature selection module calculates the response score of the input features and filters the high-response features. A feature restoration module is set at the end of the deep encoding network. The feature restoration module maps the encoded features back to the original domain to obtain the reconstructed features, calculates the deviation between the reconstructed features and the original features to obtain the reconstruction loss, and optimizes the parameters of the deep encoding network according to the reconstruction loss to obtain the compressed representation.

6. The method according to claim 1, characterized in that, Apply kernel principal component analysis to the high-dimensional evaluation vector for dimensionality reduction. Input the dimensionality-reduced evaluation vector into a density-based spatial clustering algorithm, divide the evaluation levels according to the density reachability principle, construct a hierarchical reinforcement learning framework. The high-level policy network plans the optimal target sequence based on the current system state, and the low-level execution network decomposes the target sequence into specific tuning parameters, samples and explores the parameter space through Monte Carlo tree search, and iteratively calculates the optimal parameter combination in combination with Bayesian optimization and outputs the improvement plan, including: Construct a kernel function for the high-dimensional evaluation vector, calculate the kernel similarity matrix through the kernel function, project the high-dimensional evaluation vector into the feature space using the kernel similarity matrix, calculate the sample covariance matrix in the feature space, perform eigenvalue decomposition on the sample covariance matrix to obtain the eigenvalue sequence and the eigenvector sequence, sort the eigenvalue sequence in descending order and calculate the cumulative contribution rate, select the main eigenvectors according to the preset contribution rate threshold to construct a dimensionality reduction projection matrix, and perform kernel principal component analysis dimensionality reduction on the high-dimensional evaluation vector using the dimensionality reduction projection matrix; Calculate the distance matrix for all sample points in the dimension-reduced evaluation vector, statistically analyze the distance distribution characteristics between samples based on the distance matrix, determine the density calculation radius through the distance distribution characteristics, count the number of neighboring samples within the density calculation radius for each sample point to obtain the local density value, identify the density peak points based on the local density value as the clustering centers, start from the clustering centers and sequentially add adjacent low-density samples to the clusters to form density-decreasing clustering chains, stop the clustering expansion when the density difference between adjacent samples exceeds a preset threshold, and divide the evaluation levels based on the density reachability principle; Construct a hierarchical reinforcement learning framework based on the evaluation levels. The high-level policy network performs multi-scale analysis on the current system state through the state feature extraction module to obtain the state feature representation. The state feature representation is input into the target planning module, and the target planning module generates an optimization target sequence. The low-level execution network receives the optimization target sequence, and uses the hierarchical decomposition method to transform the optimization target sequence into tuning parameters. A value evaluation module is set between the two networks to achieve information interaction; Construct a Monte Carlo search tree structure for the parameter space. The search tree nodes record the candidate parameter groups and evaluation scores. The search tree edges represent the parameter adjustment directions. Calculate the upper bound of the node value based on the historical scores and visit counts of the nodes. Select the node with the maximum upper bound of the value for expansion. Generate child nodes under the node with the maximum upper bound of the value, randomly sample the parameters of the child nodes and calculate the evaluation scores, and propagate the evaluation scores from the leaf nodes to the root node to update the path statistical information; Establish a Bayesian optimization model. Input the historical parameter solutions and effect scores as training samples into the Bayesian optimization model. Use the Bayesian optimization model to predict the effects and uncertainty levels of the parameter solutions to be evaluated. Design a combined acquisition function to comprehensively balance the predicted effects and uncertainty levels. Select the parameter solution corresponding to the maximum acquisition function value for actual evaluation. Add the new evaluation results to the training samples to update the Bayesian optimization model. Output the optimal parameter combination and generate an improvement plan through iterative calculations; 7. The method according to claim 6, characterized in that Use the Bayesian optimization model to predict the effects and uncertainty levels of the parameter solutions to be evaluated. Design a combined acquisition function to comprehensively balance the predicted effects and uncertainty levels. Select the parameter solution corresponding to the maximum acquisition function value for actual evaluation, including: Use the Bayesian optimization model for prediction. Partition the parameter space to perform local and global dual-scale predictions. Construct an acquisition function through non-dominated ranks and crowding degrees. Calculate the membership relationship of the parameter distribution family in combination with the hierarchical clustering method. Introduce a multi-scheme parallel evaluation mechanism and perform local incremental updates on the evaluation data, specifically including: Construct a Bayesian optimization model for historical parameter solutions and evaluation data. The Bayesian optimization model automatically adjusts the kernel function parameters according to the data density distribution, divides the complete parameter space into multiple continuous sub-regions, performs local predictions in the multiple continuous sub-regions respectively to obtain local prediction results, combines the local prediction results based on boundary constraint conditions to obtain global prediction results, calculates the local neighborhood prediction effect and the global range prediction effect for each parameter solution to be evaluated respectively, and calculates the degree of uncertainty in combination with the sample distribution density; Construct the prediction effect and the degree of uncertainty into a two-objective optimization problem, calculate the non-dominated relationship of the parameter solution on the two objectives to obtain the dominance rank sequence, statistically obtain the crowding degree index of the distribution density of the parameter solution in the objective space, combine the dominance rank sequence and the crowding degree index to construct a combined acquisition function, record the coverage density of the historical evaluation solution in the parameter space, and impose a scoring penalty on the high-frequency exploration area according to the coverage density; Establish a similarity calculation criterion for historical parameter solutions, construct a distance matrix based on the similarity calculation criterion, use the distance matrix to perform hierarchical clustering on the parameter solutions to obtain a set of parameter distribution families, calculate the membership relationship between the solution to be evaluated and each family in the set of parameter distribution families, and weighted sum the membership relationship to the combined acquisition function; Construct a measure index in the parameter space, use the measure index to calculate the distance difference and prediction feature difference between candidate solutions, select multiple candidate solutions with the largest degree of difference to form an evaluation set based on the distance difference and the prediction feature difference, synchronously perform actual evaluations on the candidate solutions in the evaluation set, and analyze the correlation degree between the actual evaluation results; Input the newly obtained evaluation data into the Bayesian optimization model, identify the model parameters within the influence range of the evaluation data, only perform local updates on the parameters within the influence range, establish a criterion for judging the validity of the evaluation results, and screen out abnormal evaluation data according to the criterion; Dynamically determine the number of parallel evaluation solutions according to the optimization stage, select multiple candidate solutions for parallel evaluation in the parameter space exploration stage, reduce the number of parallel evaluations in the parameter optimization stage, calculate the target weight adjustment step based on the historical evaluation effect, and adaptively update the combined weight of the prediction effect and the degree of uncertainty using the adjustment step to obtain new evaluation results.

8. A lightning protection system effectiveness evaluation system based on multi-dimensional judgment is used to implement the method described in any one of the foregoing claims 1-7, and is characterized in that Including: The first unit is used to construct a two-way feature extraction network to collect real-time monitoring data of system indicators, perform sliding sampling through a separable convolutional layer with an attention mechanism to obtain a basic feature sequence, perform multi-scale Fourier transform on the basic feature sequence to obtain periodic features in different frequency domains, fuse the periodic features and the basic feature sequence through a cross-frequency domain attention network to generate a time series feature vector, construct a dynamic association graph based on the time series feature vector, and iteratively update the node states in combination with a graph neural network to output a multi-modal feature matrix representing the overall state of the system; The second unit is used to input the multi-modal feature matrix into the causal inference network, construct the transfer function between indicators based on the structural equation model, establish a long short-term memory module through the gated recurrent unit, perform temporal combination of the historical state information and the multi-modal feature matrix and construct a multi-head self-attention model, calculate the attention weights for the combined feature sequence, perform weighted aggregation of the attention weights and the corresponding features, construct a loss function in combination with the contrastive learning strategy, optimize the feature representation based on the loss function and output a high-dimensional evaluation vector; The third unit is used to apply kernel principal component analysis to the high-dimensional evaluation vector for dimensionality reduction, input the dimension-reduced evaluation vector into the density-based spatial clustering algorithm, divide the evaluation levels according to the density reachability principle, construct a hierarchical reinforcement learning framework, the high-level policy network plans the optimization target sequence based on the current system state, the low-level execution network decomposes the target sequence into specific tuning parameters, samples and explores the parameter space through Monte Carlo tree search, iteratively calculates the optimal parameter combination in combination with Bayesian optimization and outputs an improvement plan.

9. An electronic device, characterized in that, Comprising: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, the method according to any one of claims 1 to 7 is implemented.

Images