Hydrogen energy system safety assessment method based on deep learning

Abstract

The invention discloses a hydrogen energy system safety assessment method based on deep learning, and the method comprises the following steps: S1, collecting and preprocessing the dynamic operation parameters of a hydrogen energy system, and generating an input feature matrix; s2, inputting the feature matrix into a multi-resolution feature sensing module to generate a high-dimensional feature map; s3, inputting the high-dimensional feature map into a dynamic time bending network to generate a parameter incidence matrix; s4, optimizing feature weight distribution based on an adaptive differential evolution algorithm; s5, abnormal distribution is captured through the dynamic graph convolutional network, and a dynamic abnormal graph is generated; s6, simulating an anomaly propagation path, and generating a cascade anomaly analysis result in combination with a dynamic traceability mechanism; s7, a cascade anomaly analysis result is optimized through a grey wolf-firefly hybrid optimization algorithm; and S8, dynamic operation parameters are adjusted based on a final cascade risk prediction result, and hydrogen energy system safety assessment is completed. According to the method, deep learning, a hybrid optimization algorithm and the like are combined, and safety evaluation of abnormal detection and risk prediction of the hydrogen energy system is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent security assessment, and particularly to a safety assessment method for a hydrogen energy system based on deep learning. Background Art

[0002] As a clean and efficient energy form, hydrogen energy is being widely used in fields such as transportation, industrial manufacturing, and energy storage globally due to its zero carbon emissions and sustainability. However, the physical and chemical properties of hydrogen energy determine that there are relatively high safety risks during its production, storage, transportation, and use. Hydrogen has characteristics such as being flammable, explosive, and having strong diffusivity. Once a leakage or other abnormal event occurs, it may lead to serious safety accidents. Therefore, how to effectively monitor the operating state of the hydrogen energy system, detect abnormal events in a timely manner, and implement risk mitigation measures has become a key issue in ensuring the safety of the hydrogen energy system.

[0003] Existing safety assessment methods for hydrogen energy systems mainly rely on traditional monitoring devices and single abnormal detection models, such as sensor data monitoring methods based on fixed thresholds, rule-driven abnormal detection models, etc. These methods can, to a certain extent, initially identify abnormal situations under normal operating conditions, but their detection capabilities for complex abnormal patterns with multi-variable dynamic associations are significantly insufficient. For example, when multiple parameters fluctuate simultaneously, traditional methods are difficult to capture the correlation relationships between parameters, resulting in low sensitivity and accuracy of abnormal detection. In addition, traditional monitoring systems mostly adopt an offline analysis mode and cannot achieve real-time data processing and abnormal warning, which clearly fails to meet the high dynamic and real-time requirements of the hydrogen energy system.

[0004] In recent years, with the development of artificial intelligence technology, especially the rise of deep learning and optimization algorithms, researchers have begun to attempt to introduce these technologies into the safety assessment of hydrogen energy systems. For example, convolutional neural networks (CNNs) and recurrent neural networks (RNNs) have been used to extract features from sensor data to identify potential abnormalities. However, these models mainly focus on single-variable feature extraction and static data analysis, and are insufficient in mining the spatio-temporal correlation characteristics in multi-variable dynamic time series, making it difficult to adapt to the complexity of parameter dynamic changes in hydrogen energy systems. In addition, some studies have introduced optimization algorithms to adjust the abnormal classification boundary and optimize the control strategy, but common optimization algorithms (such as genetic algorithms or particle swarm algorithms) have problems such as limited global search ability or insufficient local optimization accuracy, which makes the optimization results difficult to meet the high-precision requirements of hydrogen energy systems.

[0005] For the propagation and cascading analysis of abnormal events, existing technologies usually adopt a fixed model to model the abnormal propagation path. This method relies on predefined rules or static graph structures and cannot dynamically reflect the diffusion path and its influence range of abnormal events in actual operation. The propagation of abnormal events in the hydrogen energy system is highly dynamic and complex. Multiple abnormalities may occur simultaneously and produce a chain effect through associated paths. Existing methods are difficult to accurately capture this dynamic propagation characteristic. At the same time, existing technologies usually adopt a simple static scoring model in risk assessment, lacking the ability to evaluate the risk of a dynamic system in real time and unable to provide effective support for real-time safety control.

[0006] In terms of optimizing the abnormal classification boundary, although some studies have introduced hybrid optimization algorithms, their application in the safety assessment of hydrogen energy systems still has significant limitations. On the one hand, existing optimization algorithms are prone to local optimal solutions and have low computational efficiency when dealing with high-dimensional, multi-objective optimization problems. On the other hand, the synergistic effect between the optimization algorithm and the deep learning model has not been fully explored, making it difficult to achieve the deep integration of the algorithm and the model, resulting in the optimization results being difficult to fully guide abnormal classification and risk prediction.

[0007] Therefore, how to provide a safety assessment method for hydrogen energy systems based on deep learning is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0008] An object of the present invention is to propose a safety assessment method for hydrogen energy systems based on deep learning. The present invention combines deep learning, multi-resolution feature extraction, dynamic graph convolutional network and grey wolf-firefly hybrid optimization algorithm to realize dynamic abnormal detection, propagation path analysis and risk prediction of hydrogen energy systems. High-frequency fluctuation signals and low-frequency stable features are extracted through wavelet transform and deep convolutional network, nonlinear time series correlation is constructed by dynamic time warping analysis, spatio-temporal abnormal distribution is captured by combining dynamic graph convolutional network, and the optimization algorithm is used to improve the accuracy of the abnormal classification boundary. The present invention has the advantages of strong real-time performance, high accuracy and dynamic adaptability, can effectively ensure the safe operation of hydrogen energy systems, and reduce the risk of abnormal event diffusion.

[0009] A safety assessment method for hydrogen energy systems based on deep learning according to an embodiment of the present invention includes the following steps:

[0010] S1. Collect dynamic operation parameters of the hydrogen energy system through a multi-sensor network and perform preprocessing to generate an input feature matrix;

[0011] S2. Input the input feature matrix into the multi-resolution feature perception module, extract high-frequency fluctuation signals and low-frequency stable features based on wavelet transform and deep convolutional neural network, and generate a high-dimensional feature map;

[0012] S3. Input the high-dimensional feature map into the dynamic time warping network, analyze the multivariate time series correlation in the high-dimensional feature map, establish a dynamic mapping relationship, and generate a parameter correlation matrix;

[0013] S4. Optimize the feature weight distribution in the parameter correlation matrix based on the adaptive differential evolution algorithm, and adaptively adjust the differential weight and crossover probability to generate an optimized feature weight matrix;

[0014] S5. Input the optimized feature weight matrix into the dynamic anomaly map model, capture the spatio-temporal distribution relationship of anomalies through the dynamic graph convolutional network, and generate a dynamic anomaly map;

[0015] S6. Input the dynamic anomaly map into the anomaly cascade trigger model, simulate the propagation path of abnormal events in the hydrogen energy system, and combine the dynamic traceability mechanism to analyze the starting point, propagation path and influence range of the cascade anomaly, and generate a cascade anomaly analysis result;

[0016] S7. Optimize the cascade anomaly analysis result through the grey wolf-firefly hybrid optimization algorithm. The grey wolf algorithm is used for global search of the initial boundary of anomaly classification, and the firefly algorithm refines and optimizes the classification result to generate the final cascade risk prediction result;

[0017] S8. Adjust the dynamic operating parameters based on the final cascade risk prediction result, generate a safety risk mitigation strategy, and complete the safety status assessment of the hydrogen energy system.

[0018] Optionally, the dynamic operating parameters include pressure, temperature, flow rate, hydrogen concentration and vibration signal.

[0019] Optionally, the S2 specifically includes:

[0020] S21. The multi-resolution feature perception module includes an input layer, a wavelet decomposition layer, a high-frequency feature extraction branch, a low-frequency feature extraction branch, an attention mechanism fusion layer and an output layer; the input layer receives the input feature matrix, and the input feature matrix includes the time series data of the dynamic operating parameters of the hydrogen energy system;

[0021] S22. The wavelet decomposition layer performs multi-resolution decomposition on the input feature matrix, decomposes the dynamic operating parameters into high-frequency fluctuation signals and low-frequency stable features, where the high-frequency fluctuation signals are used to reflect rapid changes and short-term fluctuations, and the low-frequency stable features are used to reflect long-term trends and system stability;

[0022] S23. The high-frequency feature extraction branch receives the high-frequency fluctuation signals, extracts the short-term fluctuation features in the local time window by using multi-layer convolution operations, and combines the multi-scale filtering method to enhance the sensitivity to rapidly changing signals;

[0023] S24. The low-frequency feature extraction branch receives low-frequency stable features, uses a recursive convolutional structure and downsampling technology to extract long-term stable features, and improves the expression efficiency of low-frequency stable features by compressing redundant information;

[0024] S25. The attention mechanism fusion layer dynamically assigns weights, and performs weighted fusion on the outputs of the high-frequency feature extraction branch and the low-frequency feature extraction branch to generate a high-dimensional feature map containing high-frequency fluctuation signals and low-frequency stable features;

[0025] S26. The output layer is used to output the high-dimensional feature map to the dynamic time warping network.

[0026] Optionally, the S3 specifically includes:

[0027] S31. The dynamic time warping network receives the generated high-dimensional feature map and decomposes the high-dimensional feature map according to dynamic operating parameters to generate m groups of feature sequences {z 1 , z 2 , …, z m}, where each group of feature sequences represents the multi-resolution features of a dynamic operating parameter within the time range, and m represents the parameter dimension;

[0028] S32. Calculate the time matching degree between the feature sequences by the dynamic time warping method:

[0029]

[0030] where D(u, v) represents the minimum cumulative matching distance from the starting point of the sequence to the u-th time step and the v-th time step, D(u - 1, v - 1) represents the cumulative matching distance from the starting point of the sequence to the (u - 1)-th time step and the (v - 1)-th time step, D(u - 1, v) represents the cumulative matching distance from the starting point of the sequence to the (u - 1)-th time step and the v-th time step, D(u, v - 1) represents the cumulative matching distance from the starting point of the sequence to the u-th time step and the (v - 1)-th time step, and d(z u , z v ) represents the matching distance between the feature vectors at the u-th time step and the v-th time step, and α, β, and γ represent dynamic adjustment factors;

[0031] S33. Establish a dynamic mapping relationship according to the calculation result of the time matching degree to generate a time mapping relationship matrix M = [m ij m×m , where m ij represents the dynamic time correlation strength between the i-th parameter and the j-th parameter;

[0032] S34. Extract non-linear correlation features from the time mapping relationship matrix to generate a non-linear correlation feature matrix;

[0033] ​S35. Normalize the non - linear correlation feature matrix to finally generate a parameter correlation matrix, where each element of the parameter correlation matrix represents the dynamic time correlation strength between the normalized i - th parameter and j - th parameter.

[0034] Optionally, the specific steps of S4 are as follows:

[0035] S41. Optimize the feature weight distribution in the parameter correlation matrix based on the adaptive differential evolution algorithm. Initialize the population size as M, and each candidate solution represents the feature weight distribution, and m represents the parameter dimension;

[0036] S42. Construct a fitness function and evaluate the candidate solutions through a multivariate non - linear weighted model:

[0037]

[0038] where f(W k ) represents the fitness value of the candidate solution W k , a ij represents the correlation strength between the i - th parameter and j - th parameter in the parameter correlation matrix, represents the feature weight of the i - th parameter in the candidate solution W k , represents the feature weight of the j - th parameter in the candidate solution W k ;

[0039] S43. Generate a mutation vector through an adaptive mutation strategy:

[0040]

[0041] where, represents the weight value of the i - th parameter in the mutation vector, and represent the weight values of the i - th parameter in three randomly selected candidate solutions from the population, and F t represents the differential mutation coefficient:

[0042]

[0043] where F min represents the minimum value of the mutation coefficient, F max represents the maximum value of the mutation coefficient, t represents the current evolution generation, and T represents the maximum evolution generation;

[0044] Δ il represents the dynamic correlation increment between the i - th parameter and the l - th parameter:

[0045]

[0046] Among them, a il represents the association strength between the i-th parameter and the l-th parameter, and represents the weight value of the l-th parameter in the candidate solutions randomly selected from the population;

[0047] S44. Generate a trial vector using an adaptive crossover strategy:

[0048]

[0049] Among them, represents the weight value of the i-th parameter in the trial vector, represents the weight value of the i-th parameter in the mutation vector, represents the weight value of the i-th parameter in the candidate solution, r represents a randomly generated value between [0, 1], and C r,t represents the crossover probability:

[0050]

[0051] Among them, C r,min represents the minimum value of the crossover probability, and C r,max represents the maximum value of the crossover probability;

[0052] S45. Through fitness value comparison, select the vector with a higher fitness value from the candidate solutions of each generation to enter the next generation;

[0053] S46. After multiple generations of iterative optimization, select the candidate solution with the highest fitness value to generate an optimized feature weight matrix.

[0054] Optionally, the S5 specifically includes:

[0055] S51. Input the optimized feature weight matrix into the dynamic anomaly graph model, and construct an initial dynamic anomaly graph in combination with the parameter association matrix, where the nodes represent the dynamic operation parameters of the hydrogen energy system, and the weight of the edge represents the association strength between the nodes; the initial dynamic anomaly graph is used to characterize the static and initial time series relationships between the dynamic operation parameters;

[0056] S52. The dynamic anomaly graph model includes a graph construction module, a time update module, and a feature extraction module. Initialize the node attributes and edge weights into the dynamic anomaly graph through the graph construction module, and determine the initial association strength of the edges according to the weight values of the parameter association matrix;

[0057] S53. Dynamically adjust the node attributes and edge weights of the initial dynamic anomaly graph through the time update module, where the node attributes are updated according to the time change of the dynamic operation parameters, and the edge weights are updated according to the change of the dynamic association between the nodes to generate a time series dynamic anomaly graph;

[0058] S54. The feature extraction module performs feature normalization on the characteristics of the nodes and edges of the temporal dynamic anomaly graph;

[0059] S55. Input the temporal dynamic anomaly graph into the dynamic graph convolutional network, which is composed of multiple convolutional units. Each convolutional unit extracts and aggregates features of node characteristics and edge weights, and captures the anomaly distribution law in the initial dynamic anomaly graph layer by layer;

[0060] S56. Generate the final dynamic anomaly graph through the multi-layer aggregation output of the dynamic graph convolutional network.

[0061] Optionally, S6 specifically includes:

[0062] S61. Input the generated dynamic anomaly graph into the anomaly cascade trigger model, which is composed of an anomaly detection module, a propagation path modeling module, a dynamic traceability analysis module, and an impact scope evaluation module, and is used to analyze the propagation characteristics and impact scope of abnormal events in the hydrogen energy system;

[0063] S62. The anomaly detection module evaluates the anomaly intensity of the nodes of the dynamic anomaly graph, identifies the initial set of abnormal nodes based on the node attributes and the association characteristics of adjacent nodes, and uses it as the starting point of the propagation path;

[0064] S63. The propagation path modeling module takes the initial set of abnormal nodes as the starting point, constructs the propagation path of the abnormal event in a recursive expansion manner, dynamically updates the association relationship between nodes during the expansion process, and combines the anomaly characteristics and association strength of the nodes to mark the trigger nodes and target nodes of the propagation path;

[0065] S64. The dynamic traceability analysis module traces the propagation path, records the nodes and characteristics on each path, and marks the evolution law of the anomaly chain during the propagation process;

[0066] S65. The impact scope evaluation module analyzes the influence of the nodes and edges in the propagation path, identifies the high-risk areas within the propagation scope, evaluates the risk levels of each node according to the cumulative effect of the propagation path, and marks the potential impact scope of the nodes;

[0067] S66. Synthesize the results of the anomaly cascade trigger model to generate the cascade anomaly analysis results, which include the starting point of the abnormal event, the propagation path, the impact scope, and the risk level evaluation of the nodes.

[0068] Optionally, S7 specifically includes:

[0069] S71. Optimize the cascaded anomaly analysis results through the Grey Wolf-Firefly hybrid optimization algorithm. Initialize the population of the Grey Wolf algorithm, where each individual represents a multi-dimensional parameter vector of the anomaly classification boundary. Set the population size to N, and define the optimization objective function as the weighted comprehensive error of anomaly classification:

[0070]

[0071] Among them, L represents the optimization objective function, λ 1 and λ 2 represent the weight coefficients, Q represents the total number of samples, y q represents the actual classification label of sample q, represents the predicted value of the current classification boundary parameter X k , w q represents the weight of sample q, d q represents the distance from sample q to the classification boundary;

[0072] S72. The Grey Wolf algorithm optimizes the anomaly classification boundary through global search and uses the dynamic adaptive position update formula:

[0073]

[0074] Among them, represents the position of the k-th individual in the (t + 1)-th iteration, represents the position of the optimal individual in the current iteration, A t and B t represent the dynamic adjustment factors, represents the position of the sub-optimal individual in the current iteration, represents the position of the k-th individual in the t-th iteration;

[0075] S73. Take the anomaly classification boundary optimized by the Grey Wolf algorithm as the initial solution and input it into the Firefly algorithm for local refinement optimization:

[0076]

[0077] Among them, Φ(s,t) represents the light intensity of the s-th individual in the t-th iteration, Φ max represents the initial maximum light intensity, ρ represents the attenuation coefficient, represents the distance between the current individual and the target solution, represents the change amount between the current individual and the neighboring individuals, and σ represents the sensitivity adjustment factor of the neighborhood change;

[0078] S74. Dynamically adjust the individual positions according to the change of light intensity:

[0079]

[0080] Among them, represents the position of the s-th individual in the (t + 1)-th iteration, represents the position of the s-th individual in the t-th iteration, represents the position of the individual with the maximum light intensity in the t-th iteration;

[0081] S75. After optimizing through the Grey Wolf-Firefly hybrid optimization algorithm, the final cascaded risk prediction result is obtained.

[0082] The beneficial effects of the present invention are as follows:

[0083] First of all, aiming at the complexity and correlation of multi-variable dynamic operation parameters in the hydrogen energy system, the present invention adopts a multi-resolution feature perception module, which combines wavelet transform and deep convolutional neural network, and can simultaneously extract high-frequency fluctuation signals and low-frequency stable features, thereby realizing the accurate capture of rapid abnormal changes and long-term stable trends, and providing high-quality input for subsequent dynamic time series analysis.

[0084] Secondly, the present invention deeply mines the time series correlation of the high-dimensional feature map through a dynamic time warping network, establishes a dynamic mapping relationship, and generates a parameter correlation matrix, so as to reveal the non-linear time series relationship of each operation parameter in the hydrogen energy system, effectively solving the deficiencies of existing single-variable or simple linear analysis models in a multi-variable dynamic time series environment. At the same time, by optimizing the parameter correlation matrix based on the adaptive differential evolution algorithm, the dynamic adjustment ability of the adaptive adjustment strategy for the differential weight and crossover probability is fully exerted, further improving the rationality of the feature weight distribution and the model adaptability, and providing an optimized weight matrix for anomaly detection and propagation modeling.

[0085] In terms of the propagation and analysis of abnormal events, the present invention proposes a dynamic abnormal graph modeling method based on a dynamic graph convolutional network, which can capture the spatio-temporal distribution characteristics of dynamic operation parameters, and use an abnormal cascade trigger model to simulate the propagation path and influence range of abnormal events. Through a recursive expansion and dynamic traceability mechanism, key nodes and high-risk areas in the propagation path are identified, breaking through the bottleneck that traditional static propagation models cannot dynamically adapt to complex abnormal diffusion paths. Combined with a risk assessment module, it can quantitatively analyze the starting point, propagation path and cumulative impact of cascaded anomalies, providing accurate input for the generation of subsequent safety control strategies.

[0086] In addition, through the innovative application of the Grey Wolf-Firefly hybrid optimization algorithm, the present invention realizes the combination of global search and local optimization of the anomaly classification boundary. The Grey Wolf algorithm uses its global search ability to quickly converge to the optimal solution, while the Firefly algorithm refines and optimizes the initial solution. In particular, by introducing a dynamic light intensity adjustment mechanism, the accuracy and robustness of the classification boundary optimization are significantly improved. This collaborative optimization mechanism not only enhances the accuracy of anomaly classification but also effectively improves the adaptability of the model in complex scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0087] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention, but do not constitute a limitation to the present invention. In the drawings:

[0088] Figure 1 is the overall flowchart of a method for safety assessment of a hydrogen energy system based on deep learning proposed by the present invention;

[0089] Figure 2 is the structural schematic diagram of an anomaly cascade trigger model of a method for safety assessment of a hydrogen energy system based on deep learning proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0090] Now, the present invention will be further described in detail with reference to the drawings. These drawings are all simplified schematic diagrams, only showing the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0091] Refer to Figure 1 and Figure 2 , a method for safety assessment of a hydrogen energy system based on deep learning, includes the following steps:

[0092] S1. Collect the dynamic operation parameters of the hydrogen energy system through a multi-sensor network and perform preprocessing to generate an input feature matrix;

[0093] S2. Input the input feature matrix into a multi-resolution feature perception module, and extract high-frequency fluctuation signals and low-frequency stable features based on wavelet transform and deep convolutional neural network to generate a high-dimensional feature map;

[0094] S3. Input the high-dimensional feature map into a dynamic time warping network, analyze the multi-variable time series correlation in the high-dimensional feature map, establish a dynamic mapping relationship, and generate a parameter correlation matrix;

[0095] S4. Optimize the feature weight distribution in the parameter correlation matrix based on the adaptive differential evolution algorithm, and adaptively adjust the differential weight and crossover probability to generate an optimized feature weight matrix;

[0096] S5. Input the optimized feature weight matrix into the dynamic anomaly map model, capture the spatio-temporal distribution relationship of anomalies through the dynamic graph convolutional network, and generate a dynamic anomaly map;

[0097] S6. Input the dynamic anomaly map into the anomaly cascade trigger model, simulate the propagation path of abnormal events in the hydrogen energy system, combine the dynamic traceability mechanism to analyze the starting point, propagation path and influence range of cascade anomalies, and generate a cascade anomaly analysis result;

[0098] S7. Optimize the cascade anomaly analysis result through the grey wolf-firefly hybrid optimization algorithm. The grey wolf algorithm is used for global search of the initial boundary of anomaly classification, and the firefly algorithm refines and optimizes the classification result to generate the final cascade risk prediction result;

[0099] S8. Adjust the dynamic operating parameters based on the final cascade risk prediction result, generate a safety risk mitigation strategy, and complete the safety status assessment of the hydrogen energy system.

[0100] In this embodiment, the dynamic operating parameters include pressure, temperature, flow rate, hydrogen concentration and vibration signal.

[0101] In this embodiment, the specific steps of S2 are as follows:

[0102] S21. The multi-resolution feature perception module includes an input layer, a wavelet decomposition layer, a high-frequency feature extraction branch, a low-frequency feature extraction branch, an attention mechanism fusion layer and an output layer; the input layer receives the input feature matrix, and the input feature matrix includes the time series data of the dynamic operating parameters of the hydrogen energy system;

[0103] S22. The wavelet decomposition layer performs multi-resolution decomposition on the input feature matrix, decomposes the dynamic operating parameters into high-frequency fluctuation signals and low-frequency stable features, where the high-frequency fluctuation signals are used to reflect rapid changes and short-term fluctuations, and the low-frequency stable features are used to reflect long-term trends and system stability;

[0104] S23. The high-frequency feature extraction branch receives the high-frequency fluctuation signals, extracts the short-term fluctuation features in the local time window by using multi-layer convolution operations, and combines the multi-scale filtering method to enhance the sensitivity to rapidly changing signals;

[0105] S24. The low-frequency feature extraction branch receives the low-frequency stable features, extracts the long-term stable features by using a recursive convolution structure and downsampling technology, and improves the expression efficiency of the low-frequency stable features by compressing redundant information;

[0106] S25. The attention mechanism fusion layer dynamically assigns weights, performs weighted fusion on the outputs of the high-frequency feature extraction branch and the low-frequency feature extraction branch, and generates a high-dimensional feature map containing high-frequency fluctuation signals and low-frequency stable features;

[0107] S26. The output layer is used to output the high-dimensional feature map to the dynamic time warping network.

[0108] In this embodiment, S3 specifically includes:

[0109] S31. The dynamic time warping network receives the generated high-dimensional feature map and decomposes the high-dimensional feature map according to dynamic operation parameters to generate m groups of feature sequences {z 1 , z 2 , …, z m}, where each group of feature sequences represents the multi-resolution features of a dynamic operation parameter within the time range, and m represents the parameter dimension;

[0110] S32. Calculate the time matching degree between the feature sequences by the dynamic time warping method:

[0111]

[0112] where D(u, v) represents the minimum cumulative matching distance from the starting point of the sequence to the u-th time step and the v-th time step, D(u - 1, v - 1) represents the cumulative matching distance from the starting point of the sequence to the (u - 1)-th time step and the (v - 1)-th time step, D(u - 1, v) represents the cumulative matching distance from the starting point of the sequence to the (u - 1)-th time step and the v-th time step, D(u, v - 1) represents the cumulative matching distance from the starting point of the sequence to the u-th time step and the (v - 1)-th time step, d(z u , z v ) represents the matching distance between the feature vectors at the u-th time step and the v-th time step, and α, β, and γ represent dynamic adjustment factors;

[0113] S33. Establish a dynamic mapping relationship according to the calculation result of the time matching degree to generate a time mapping relationship matrix M = [m ij m×m , where m ij represents the dynamic time correlation strength between the i-th parameter and the j-th parameter;

[0114] S34. Extract non-linear correlation features from the time mapping relationship matrix to generate a non-linear correlation feature matrix;

[0115] S35. Perform normalization processing on the non-linear correlation feature matrix to finally generate a parameter correlation matrix, and each element of the parameter correlation matrix represents the normalized dynamic time correlation strength between the i-th parameter and the j-th parameter.

[0116] In this embodiment, S4 specifically includes:

[0117] ​S41. Optimize the feature weight distribution in the parameter correlation matrix based on the adaptive differential evolution algorithm. Initialize the population size as M, where each candidate solution represents the feature weight distribution, and m represents the parameter dimension;

[0118] S42. Construct a fitness function and evaluate the candidate solutions through a multivariate non-linear weighted model:

[0119]

[0120] where f(W k ) represents the fitness value of the candidate solution W k , a ij represents the correlation strength between the i-th parameter and the j-th parameter in the parameter correlation matrix, represents the feature weight of the i-th parameter in the candidate solution W k , represents the feature weight of the j-th parameter in the candidate solution W k ;

[0121] S43. Generate a mutant vector through an adaptive mutation strategy:

[0122]

[0123] where, represents the weight value of the i-th parameter in the mutant vector, and represent the weight values of the i-th parameter in three randomly selected candidate solutions from the population, and F t represents the differential mutation coefficient:

[0124]

[0125] where F min represents the minimum value of the mutation coefficient, F max represents the maximum value of the mutation coefficient, t represents the current generation of evolution, and T represents the maximum number of generations of evolution;

[0126] Δ il represents the dynamic correlation increment between the i-th parameter and the l-th parameter:

[0127]

[0128] where a il represents the correlation strength between the i-th parameter and the l-th parameter, represents the weight value of the l-th parameter in a randomly selected candidate solution from the population;

[0129] S44. Generate a trial vector using an adaptive crossover strategy:

[0130]

[0131] Among them, represents the weight value of the i-th parameter in the test vector, represents the weight value of the i-th parameter in the mutation vector, represents the weight value of the i-th parameter in the candidate solution, and r represents a randomly generated value between [0, 1], C r,t represents the crossover probability:

[0132]

[0133] Among them, C r,min represents the minimum value of the crossover probability, C r,max represents the maximum value of the crossover probability;

[0134] S45. Through fitness value comparison, select the vector with a higher fitness value among the candidate solutions of each generation to enter the next generation;

[0135] S46. After multiple generations of iterative optimization, select the candidate solution with the highest fitness value to generate an optimized feature weight matrix.

[0136] In this embodiment, the specific steps of S5 include:

[0137] S51. Input the optimized feature weight matrix into the dynamic anomaly graph model, and construct an initial dynamic anomaly graph in combination with the parameter correlation matrix, where the nodes represent the dynamic operation parameters of the hydrogen energy system, and the weight of the edge represents the correlation strength between the nodes; the initial dynamic anomaly graph is used to characterize the static and initial time series relationships between the dynamic operation parameters;

[0138] S52. The dynamic anomaly graph model includes a graph construction module, a time update module, and a feature extraction module. Initialize the node attributes and edge weights into the dynamic anomaly graph through the graph construction module, and determine the initial correlation strength of the edges according to the weight values of the parameter correlation matrix;

[0139] S53. Dynamically adjust the node attributes and edge weights of the initial dynamic anomaly graph through the time update module, where the node attributes are updated according to the time change of the dynamic operation parameters, and the edge weights are updated according to the change of the dynamic correlation between the nodes to generate a time series dynamic anomaly graph;

[0140] S54. The feature extraction module performs feature normalization processing on the characteristics of the nodes and edges of the time series dynamic anomaly graph;

[0141] S55. Input the time-series dynamic anomaly map into the dynamic graph convolutional network, which is composed of multiple convolutional units. Each convolutional unit performs feature extraction and aggregation on node characteristics and edge weights, and captures the anomaly distribution law in the initial dynamic anomaly map layer by layer;

[0142] S56. Generate the final dynamic anomaly map through the multi-layer aggregation output of the dynamic graph convolutional network.

[0143] In this embodiment, S6 specifically includes:

[0144] S61. Input the generated dynamic anomaly map into the anomaly cascade trigger model, which is composed of an anomaly detection module, a propagation path modeling module, a dynamic traceability analysis module, and an impact scope evaluation module, and is used to analyze the propagation characteristics and impact scope of abnormal events in the hydrogen energy system;

[0145] S62. The anomaly detection module evaluates the anomaly intensity of the nodes in the dynamic anomaly map, identifies the initial anomaly node set based on the node attributes and the association characteristics of adjacent nodes, and uses it as the starting point of the propagation path;

[0146] S63. The propagation path modeling module takes the initial anomaly node set as the starting point, constructs the propagation path of the abnormal event in a recursive expansion manner, dynamically updates the association relationship between nodes during the expansion process, and combines the anomaly characteristics and association intensity of the nodes to mark the trigger nodes and target nodes of the propagation path;

[0147] S64. The dynamic traceability analysis module traces the propagation path, records the nodes and characteristics on each path, and marks the evolution law of the abnormal chain during the propagation process;

[0148] S65. The impact scope evaluation module analyzes the influence of the nodes and edges in the propagation path, identifies the high-risk areas within the propagation scope, evaluates the risk levels of each node according to the cumulative effect of the propagation path, and marks the potential impact scope of the nodes;

[0149] S66. Synthesize the results of the anomaly cascade trigger model to generate the cascade anomaly analysis result, which includes the starting point of the abnormal event, the propagation path, the impact scope, and the risk level evaluation of the nodes.

[0150] In this embodiment, S7 specifically includes:

[0151] S71. Optimize the cascade anomaly analysis result through the Grey Wolf-Firefly hybrid optimization algorithm. The Grey Wolf algorithm initializes the population, each individual represents a multi-dimensional parameter vector of the anomaly classification boundary, sets the population size to N, and defines the optimization objective function as the weighted comprehensive error of anomaly classification:

[0152]

[0153] Among them, L represents the optimization objective function, λ 1 and λ 2 represent the weight coefficients, Q represents the total number of samples, y q represents the actual classification label of sample q, represents the predicted value of the current classification boundary parameter X k of, w q represents the weight of sample q, d q represents the distance from sample q to the classification boundary;

[0154] S72. The grey wolf algorithm optimizes the abnormal classification boundary through global search and uses the dynamic adaptive position update formula:

[0155]

[0156] Among them, represents the position of the k-th individual in the (t + 1)-th iteration, represents the position of the optimal individual in the current iteration, A t and B t represent the dynamic adjustment factors, represents the position of the sub-optimal individual in the current iteration, represents the position of the k-th individual in the t-th iteration;

[0157] S73. Take the abnormal classification boundary optimized by the grey wolf algorithm as the initial solution and input it into the firefly algorithm for local refinement optimization:

[0158]

[0159] Among them, Φ(s,t) represents the light intensity of the s-th individual in the t-th iteration, Φ max represents the initial maximum light intensity, ρ represents the attenuation coefficient, represents the distance between the current individual and the target solution, represents the change amount between the current individual and the neighboring individual, and σ represents the sensitivity adjustment factor of the neighborhood change;

[0160] S74. Dynamically adjust the individual position according to the change of light intensity:

[0161]

[0162] Among them, represents the position of the s-th individual in the (t + 1)-th iteration, represents the position of the s-th individual in the t-th iteration, represents the position of the individual with the maximum light intensity in the t-th iteration;

[0163] After optimization by the Grey Wolf-Firefly hybrid optimization algorithm, the final cascaded risk prediction result is obtained.

[0164] Example 1:

[0165] To verify the feasibility of the present invention in implementation, the present invention is applied to the hydrogen energy storage and transportation system of a large industrial park. The industrial park is located in the eastern coastal area of China, with a total area of more than 2,000 hectares, more than 20 hydrogen storage tanks, 8 high-pressure transmission pipelines and various key equipment. Due to the characteristics of hydrogen such as flammability, explosiveness and strong diffusibility, any abnormality in the storage and transportation process will pose a major threat to the safety production of the industrial park. For a long time, the park has adopted a traditional sensor monitoring and fixed threshold alarm system, but there are still many problems in the operation of the system, such as the inability to effectively capture the dynamic correlation between multi-variable parameters, the inability to predict the abnormal propagation path in real time, and the lack of an accurate risk assessment mechanism. Therefore, the park decides to introduce the deep learning-based hydrogen energy system safety assessment method described in the present invention to improve its safety monitoring ability.

[0166] In practical applications, the park first arranges a monitoring system composed of a multi-sensor network, including pressure sensors, temperature sensors, flow sensors, hydrogen concentration detectors and vibration monitoring devices, with a total of more than 300 sensing nodes deployed. These sensors can collect the dynamic operation parameters of the hydrogen energy storage and transportation system in real time and transmit the data to the safety assessment platform of the present invention through a wireless network. During the data transmission process, the system uses an interpolation algorithm to repair data loss and removes sensor noise through the Kalman filtering method to ensure the high quality of the input data.

[0167] In this scenario, the multi-resolution feature perception module of the present invention first processes the collected dynamic operation parameters. Through wavelet decomposition, high-frequency fluctuation signals are extracted to reflect short-term abnormalities (such as instantaneous pressure fluctuations and hydrogen leaks), and low-frequency stable features are extracted to reflect the long-term operation trend of the system (such as temperature rise caused by equipment aging). Subsequently, the dynamic time warping network analyzes the multi-variable time series correlation in the high-dimensional feature map and generates a parameter correlation matrix, thereby revealing the dynamic relationship between parameters such as pressure, temperature and flow.

[0168] In a simulation test of a pressure anomaly event, the present invention successfully captured the triggering node of the anomaly event and analyzed the propagation path of the anomaly in real time through a dynamic graph convolutional network. Specifically, when the internal pressure of storage tank A suddenly increased, the anomaly triggering module detected the abnormal intensity of the pressure parameter and identified storage tank A as the anomaly starting point. Subsequently, the propagation path modeling module simulated the diffusion process of the anomaly, and determined the key nodes in the anomaly path through the dynamic traceability analysis module, including storage tanks B, C, and pipeline D. The impact scope assessment module further quantified the risk level of the anomaly event and predicted the potential risks in the diffusion area, and finally generated a cascaded anomaly analysis result.

[0169] To verify the effectiveness of the present invention, a comparative experiment was conducted with a traditional threshold alarm system. Five common anomaly scenarios were simulated in the experimental setup, including hydrogen leakage, pipeline blockage, overheating caused by equipment aging, excessive vibration, and pressure imbalance. In each anomaly scenario, the anomaly detection time, the accuracy rate of key node identification, the prediction accuracy of the propagation path, and the comprehensive score of risk assessment were recorded. The experimental data showed that the present invention was significantly superior to the traditional method in all indicators. Especially in the multi-variable complex anomaly scenario, its detection accuracy was improved by more than 45%, and the response time was shortened by about 20 seconds.

[0170] Table 1 Performance comparison table of the present invention and the traditional method in different anomaly scenarios

[0171]

[0172]

[0173] The experimental results show that the performance of the present invention in the anomaly detection and risk assessment of the hydrogen energy system is significantly superior to the traditional method. Taking hydrogen leakage as an example, the anomaly detection time of the traditional method is 15 seconds, while the present invention only needs 5 seconds, and the detection efficiency is improved by 66.7%. In terms of the accuracy rate of key node identification, the traditional method is 78%, while the present invention reaches 96%, and the accuracy is improved by 18 percentage points. At the same time, the prediction accuracy of the propagation path is increased from 65% of the traditional method to 92%, showing the excellent performance of the present invention in dynamic complex scenarios.

[0174] In the scenario of pipeline blockage, the anomaly detection time of the traditional method is 20 seconds, the accuracy rate of key node identification is 82%, and the prediction accuracy of the propagation path is 70%. In contrast, the present invention reaches 8 seconds, 94%, and 90% in the corresponding indicators respectively, not only greatly shortening the response time, but also significantly improving the accurate prediction ability of the anomaly propagation. For scenarios such as equipment overheating and vibration anomalies, the present invention shows superiority in all indicators, among which the accuracy rate of key node identification and the prediction accuracy of the propagation path are increased by 15 - 25 percentage points respectively, and the anomaly detection time is shortened to one-third of the original.

[0175] Especially in the scenario of pressure imbalance, the present invention demonstrates comprehensive advantages. The abnormal detection time of the traditional method is 19 seconds, while the present invention only requires 6 seconds. The accuracy of key node identification has increased from 81% to 97%, and the prediction accuracy of the propagation path has been improved from 69% to 93%. This result shows that the multi-resolution feature extraction and dynamic time warping analysis method of the present invention can accurately capture the characteristics of pressure changes, and the optimized abnormal classification boundary significantly improves the performance of classification and prediction.

[0176] Based on the experimental data of various scenarios, the present invention is significantly superior to the traditional method in three key indicators: abnormal detection time, accuracy of key node identification, and prediction accuracy of the propagation path. The traditional method is limited by single-threshold monitoring and static analysis models and cannot effectively handle complex dynamic scenarios. However, through the collaborative application of multi-resolution feature extraction, dynamic graph convolutional network, and optimization algorithms, the present invention not only improves the detection accuracy and response speed of the system but also enhances the prediction ability of the propagation of abnormal events.

[0177] The above is only the preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and all should be covered by the protection scope of the present invention.

Claims

1. A hydrogen energy system safety assessment method based on deep learning, characterized in that: The steps include: S1. Collect dynamic operating parameters of the hydrogen energy system through a multi-sensor network and perform preprocessing to generate an input feature matrix; S2, input the input feature matrix into the multi-resolution feature perception module, extract high-frequency fluctuation signals and low-frequency stable features based on wavelet transform and deep convolutional neural network, and generate a high-dimensional feature map; S3, inputting the high-dimensional feature graph into the dynamic time warping network, analyzing the multivariate temporal correlation in the high-dimensional feature graph, establishing a dynamic mapping relationship, and generating a parameter correlation matrix; S4, optimizing the feature weight distribution in the parameter association matrix based on the adaptive differential evolution algorithm, and adaptively adjusting the differential weight and crossover probability to generate an optimized feature weight matrix; S5. Input the optimized feature weight matrix into the dynamic anomaly graph model, capture the spatiotemporal distribution relationship of the anomaly through the dynamic graph convolutional network, and generate a dynamic anomaly graph; S6. Input the dynamic anomaly graph into the anomaly cascade trigger model, simulate the propagation path of abnormal events in the hydrogen energy system, analyze the starting point, propagation path and impact range of the cascade anomaly in combination with the dynamic tracing mechanism, and generate the cascade anomaly analysis results; S7. The cascade anomaly analysis results are optimized by using the gray wolf-firefly hybrid optimization algorithm. The gray wolf algorithm is used to globally search for the initial boundary of anomaly classification, and the firefly algorithm refines and optimizes the classification results to generate the final cascade risk prediction results. S8. Adjust the dynamic operating parameters based on the final cascade risk prediction results, generate a safety risk mitigation strategy, and complete the safety status assessment of the hydrogen energy system.

2. According to a deep learning-based hydrogen energy system safety assessment method according to claim 1, it is characterized in that: The dynamic operating parameters include pressure, temperature, flow rate, hydrogen concentration and vibration signal.

3. According to a deep learning-based hydrogen energy system safety assessment method according to claim 1, it is characterized in that: The S2 specifically includes: S21, the multi-resolution feature perception module includes an input layer, a wavelet decomposition layer, a high-frequency feature extraction branch, a low-frequency feature extraction branch, an attention mechanism fusion layer and an output layer; the input layer receives an input feature matrix, and the input feature matrix includes time series data of dynamic operating parameters of the hydrogen energy system; S22, the wavelet decomposition layer performs multi-resolution decomposition on the input feature matrix, decomposing the dynamic operation parameters into high-frequency fluctuation signals and low-frequency stable features, wherein the high-frequency fluctuation signals are used to reflect rapid changes and short-term fluctuations, and the low-frequency stable features are used to reflect long-term trends and system stability; S23, the high-frequency feature extraction branch receives the high-frequency fluctuation signal, uses a multi-layer convolution operation to extract the short-term fluctuation features in the local time window, and combines a multi-scale filtering method to enhance the sensitivity to the rapidly changing signal; S24, the low-frequency feature extraction branch receives low-frequency stable features, uses a recursive convolution structure and downsampling technology to extract long-term stable features, and improves the expression efficiency of low-frequency stable features by compressing redundant information; S25, the attention mechanism fusion layer dynamically assigns weights, performs weighted fusion on the outputs of the high-frequency feature extraction branch and the low-frequency feature extraction branch, and generates a high-dimensional feature map containing high-frequency fluctuation signals and low-frequency stable features; S26. The output layer is used to output the high-dimensional feature map to the dynamic time warping network.

4. The hydrogen energy system safety assessment method based on deep learning according to claim 1 is characterized in that: The S3 specifically includes: S31, the dynamic time warping network receives the generated high-dimensional feature graph, and decomposes the high-dimensional feature graph according to the dynamic operation parameters to generate m groups of feature sequences {z1, z2, ..., z m }, where each set of feature sequences represents a multi-resolution feature of a dynamic operation parameter within a time range, and m represents the parameter dimension; S32. Calculate the time matching degree between feature sequences by dynamic time warping method: Where D(u,v) represents the minimum cumulative matching distance from the starting point of the sequence to the u-th time step and the v-th time step, D(u-1,v-1) represents the cumulative matching distance from the starting point of the sequence to the u-1th time step and the v-1th time step, D(u-1,v) represents the cumulative matching distance from the starting point of the sequence to the u-1th time step and the v-1th time step, D(u,v-1) represents the cumulative matching distance from the starting point of the sequence to the u-1th time step and the v-1th time step, and d(z u ,z v ) represents the matching distance between the feature vectors at the uth time step and the vth time step, α, β and γ represent dynamic adjustment factors; S33, establish a dynamic mapping relationship based on the calculation result of the time matching degree, and generate a time mapping relationship matrix M=[m ij ] m×m , where m ij represents the dynamic time correlation strength between the i-th parameter and the j-th parameter; S34, extracting nonlinear correlation features from the time mapping relationship matrix to generate a nonlinear correlation feature matrix; S35. Normalize the nonlinear correlation feature matrix to finally generate a parameter correlation matrix, wherein each element of the parameter correlation matrix represents the normalized dynamic time correlation strength between the i-th parameter and the j-th parameter.

5. The hydrogen energy system safety assessment method based on deep learning according to claim 1 is characterized in that: The S4 specifically includes: S41, based on the adaptive differential evolution algorithm, optimize the feature weight distribution in the parameter association matrix, initialize the population size to M, and each candidate solution represents the feature weight distribution, and m represents the parameter dimension; S42. Construct a fitness function and evaluate candidate solutions through a multivariate nonlinear weighted model: Among them, f(W k ) represents the candidate solution W k The fitness value, a ij represents the correlation strength between the i-th parameter and the j-th parameter in the parameter correlation matrix, Indicates that the i-th parameter in the candidate solution W k The feature weights in Indicates that the jth parameter in the candidate solution W k The feature weights in ; S43, generating mutation vectors through adaptive mutation strategy: in, represents the weight value of the i-th parameter in the mutation vector, and represents the weight value of the i-th parameter among the three candidate solutions randomly selected from the population, F t Express the coefficient of variation of the difference: Among them, F min Indicates the minimum value of the coefficient of variation, F max represents the maximum value of the coefficient of variation, t represents the current evolutionary generation, and T represents the maximum evolutionary generation; Δ il Represents the dynamic association increment between the i-th parameter and the l-th parameter: Among them, a il represents the correlation strength between the ith parameter and the lth parameter, represents the weight value of the lth parameter in the candidate solution randomly selected from the population; S44, using adaptive crossover strategy to generate test vectors: in, represents the weight value of the i-th parameter in the test vector, represents the weight value of the i-th parameter in the mutation vector, represents the weight value of the i-th parameter in the candidate solution, r represents a randomly generated value between [0,1], C r,t Represents the crossover probability: Among them, C r,min represents the minimum value of the crossover probability, C r,max represents the maximum value of the crossover probability; S45, by comparing the fitness values, selecting the vector with the highest fitness value from the candidate solutions of each generation to enter the next generation; S46. After multiple generations of iterative optimization, the candidate solution with the highest fitness value is selected to generate an optimized feature weight matrix.

6. The hydrogen energy system safety assessment method based on deep learning according to claim 1 is characterized in that: The S5 specifically includes: S51, inputting the optimized feature weight matrix into the dynamic anomaly graph model, and constructing an initial dynamic anomaly graph in combination with the parameter association matrix, wherein the nodes represent the dynamic operation parameters of the hydrogen energy system, and the edge weights represent the association strength between the nodes; the initial dynamic anomaly graph is used to characterize the static and initial temporal relationships between the dynamic operation parameters; S52, the dynamic anomaly graph model includes a graph construction module, a time update module and a feature extraction module, the node attributes and edge weights are initialized into the dynamic anomaly graph through the graph construction module, and the initial association strength of the edge is determined according to the weight value of the parameter association matrix; S53, dynamically adjusting the node attributes and edge weights of the initial dynamic anomaly graph through the time update module, wherein the node attributes are updated according to the time changes of the dynamic operation parameters, and the edge weights are updated according to the changes of the dynamic associations between the nodes, to generate a time series dynamic anomaly graph; S54, the feature extraction module performs feature normalization processing on the characteristics of the nodes and edges of the time series dynamic anomaly graph; S55, inputting the time series dynamic anomaly graph into a dynamic graph convolutional network, wherein the dynamic graph convolutional network is composed of multiple layers of convolutional units, each of which performs feature extraction and aggregation on node characteristics and edge weights, and captures the anomaly distribution law in the initial dynamic anomaly graph layer by layer; S56. Generate the final dynamic anomaly graph through the multi-layer aggregation output of the dynamic graph convolutional network.

7. The hydrogen energy system safety assessment method based on deep learning according to claim 1 is characterized in that: The S6 specifically includes: S61, inputting the generated dynamic anomaly graph into an anomaly cascade trigger model, wherein the anomaly cascade trigger model is composed of an anomaly detection module, a propagation path modeling module, a dynamic source tracing analysis module and an impact range assessment module, and is used to analyze the propagation characteristics and impact range of abnormal events in the hydrogen energy system; S62, the anomaly detection module performs an anomaly strength assessment on the nodes of the dynamic anomaly graph, identifies an initial set of abnormal nodes based on node attributes and correlation characteristics of adjacent nodes, and uses the initial set of abnormal nodes as the starting point of the propagation path; S63, the propagation path modeling module takes the initial abnormal node set as the starting point, constructs the propagation path of the abnormal event by recursive expansion, dynamically updates the association relationship between the nodes during the expansion process, and marks the trigger node and the target node of the propagation path in combination with the abnormal characteristics and association strength of the nodes; S64, the dynamic tracing analysis module traces the propagation path, records the nodes and characteristics on each path, and marks the evolution law of the abnormal chain during the propagation process; S65, the impact range assessment module performs influence analysis on the nodes and edges in the propagation path, identifies high-risk areas within the propagation range, assesses the risk level of each node based on the cumulative effect of the propagation path, and marks the potential impact range of the node; S66. The results of the integrated abnormal cascade trigger model are used to generate a cascade abnormality analysis result, wherein the cascade abnormality analysis result includes the starting point, propagation path, impact range and risk level assessment of the node of the abnormal event.

8. The hydrogen energy system safety assessment method based on deep learning according to claim 1 is characterized in that: The S7 specifically includes: S71. The cascade anomaly analysis results are optimized by the gray wolf-firefly hybrid optimization algorithm. The gray wolf algorithm initializes the population, each individual represents a multidimensional parameter vector of the anomaly classification boundary, sets the population size to N, and defines the optimization objective function as the weighted comprehensive error of the anomaly classification: Among them, L represents the optimization objective function, λ1 and λ2 represent weight coefficients, Q represents the total number of samples, and y q represents the actual classification label of sample q, Represents the current classification boundary parameter X k The predicted value, w q represents the weight of sample q, d q Represents the distance from sample q to the classification boundary; S72, the Gray Wolf Algorithm optimizes the abnormal classification boundary through global search and uses the dynamic adaptive position update formula: in, represents the position of the kth individual in the t+1th iteration, Indicates the position of the best individual in the current iteration, A t and B t represents the dynamic adjustment factor, represents the position of the suboptimal individual in the current iteration, represents the position of the kth individual in the tth iteration; S73. The abnormal classification boundary obtained by optimizing the gray wolf algorithm is used as the initial solution and input into the firefly algorithm for local refinement optimization: Among them, Φ(s,t) represents the light intensity of the sth individual in the tth iteration, Φ max represents the initial maximum light intensity, ρ represents the attenuation coefficient, represents the distance between the current individual and the target solution, represents the amount of change between the current individual and the neighboring individuals, and σ represents the sensitivity adjustment factor of neighborhood changes; S74. Dynamically adjust individual positions according to changes in light intensity: in, represents the position of the sth individual in the t+1th iteration, represents the position of the sth individual in the tth iteration, represents the individual position with the maximum light intensity in the tth iteration; S75. After optimization by the gray wolf-firefly hybrid optimization algorithm, the final cascade risk prediction result is obtained.

Images