A neural network-based component intelligent identification method and system
The intelligent component identification method, which combines hybrid sampling and fully connected neural networks with genetic algorithms for optimization, solves the problems of reliability and accuracy in component identification in power information systems, and achieves efficient and low-cost component management.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- STATE GRID HUNAN ELECTRIC POWER COMPANY LIMITED
- Filing Date
- 2023-11-23
- Publication Date
- 2026-07-21
AI Technical Summary
In large-scale power information systems, existing component identification schemes are not reliable and accurate enough, making them difficult to apply to complex component management scenarios. Furthermore, manual identification and recording are time-consuming and prone to errors.
Data is acquired using a hybrid sampling method, a fully connected neural network model is constructed, and the network structure is optimized using a genetic algorithm. Combined with an adaptive selection operator and an adaptive iteration count method, intelligent component recognition is achieved.
It improves the reliability and accuracy of component identification, reduces data collection costs, is suitable for large-scale scenarios, and avoids the subjectivity of human intervention and getting trapped in local optima.
Smart Images
Figure CN117574237B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of computer technology, specifically relating to a component intelligent recognition method and system based on neural networks. Background Technology
[0002] With economic and technological development and the improvement of people's living standards, electricity has become an indispensable secondary energy source in people's production and daily life, bringing endless convenience. Therefore, ensuring a stable and reliable supply of electricity has become one of the most important tasks of the power system.
[0003] Currently, with the widespread application of IoT, computer, and power information system technologies in power systems, the stability, intelligence, and reliability of power systems have been greatly guaranteed. These applications serving power systems often rely on various hosts, databases, and middleware for support and operation. However, with the increasing variety of components and frequent version updates, accurately knowing the type and version of each component has become a thorny issue for power systems. For example, different types of hosts may require different operating system commands or configuration methods to achieve corresponding operations, and different versions of components may lead to compatibility issues.
[0004] However, in large-scale application environments, manually identifying and recording component types and versions is an extremely time-consuming and error-prone task. Furthermore, as the application scales and the number of components increases, manual component management becomes virtually impossible.
[0005] Therefore, some intelligent identification solutions already exist that can intelligently identify components in a system. The current mainstream approach uses tools like nmap and Fing to discover component assets within the system. However, these solutions determine component types based on the device's network characteristics, resulting in poor reliability and accuracy. Furthermore, these solutions are not suitable for today's large-scale scenarios. Summary of the Invention
[0006] One of the objectives of this invention is to provide a neural network-based intelligent component identification method that is applicable to large-scale scenarios and has high reliability and accuracy.
[0007] The second objective of this invention is to provide a system for implementing the neural network-based intelligent component recognition method.
[0008] The component intelligent recognition method based on neural networks provided by this invention includes the following steps:
[0009] S1. Identify known power information systems;
[0010] S2. In the power information system determined in step S1, a hybrid sampling method is used to obtain the data information of the power information system and construct a training dataset;
[0011] S3. Based on neural networks, construct an initial model for intelligent component recognition;
[0012] S4. Using the training dataset constructed in step S2, train the initial component intelligent recognition model constructed in step S3 to obtain the component intelligent recognition model;
[0013] S5. Using the component intelligent identification model obtained in step S4, perform intelligent identification of components in the target power information system.
[0014] Step S1, which involves determining a known power information system, specifically includes the following steps:
[0015] Identify known power information systems;
[0016] The known power information system specifically meets the following requirements:
[0017] The power information system has N servers; each server host has 0 to Num databases and middleware installed;
[0018] There are three types of power information systems: H-type mainframes, D-type databases, and M-type middleware.
[0019] The power information system has at least one server with global access permissions, which can access any device in the power information system;
[0020] All devices in the power information system have fixed IP addresses and have been centrally registered.
[0021] Step S2, which describes acquiring data from the power information system determined in step S1 using a hybrid sampling method to construct a training dataset, specifically includes the following steps:
[0022] Initial scan results were generated using scanning tools: several existing methods were employed to detect the host type, database type, and middleware type of servers in the power information system;
[0023] For the same server, if all the obtained components are of the same type, then that server is marked as a candidate server; assuming a total of R1 candidate servers are obtained, if... Then randomly select from R1 candidate servers Each candidate server is manually reviewed; if... Then, all candidate servers are manually reviewed, and a selection is made from the remaining servers. Each server undergoes manual review; this manual review is to verify the accuracy of the existing solutions used for evaluation. The accuracy of the benchmark data The scale of the baseline data;
[0024] Statistical accuracy based on benchmark data: Let the i-th existing scheme detect S... i Servers, of which The server was manually checked and found to be correct. If the detection results from the server are incorrect, then the accuracy Pa of the i-th existing solution will be reduced. i for:
[0025]
[0026] In the formula, ω is the penalty factor; The term reflects the server recognition rate and accuracy of the i-th existing scheme; The item is a penalty item;
[0027] Select accuracy Pa i For existing schemes where the threshold is exceeded, feature data and labels are collected; {x j ,y j} represents the j-th data item in the collected dataset, x j Let y be the feature vector of the j-th data point. j Let be the label vector of the j-th data item; where the label vector is a binary vector of dimension H+D+M, and a value of 1 for the kk-th bit indicates that the server has the component corresponding to the kk-th bit, and a value of 0 for the kk-th bit indicates that the server does not have the component corresponding to the kk-th bit.
[0028] The selected accuracy Pa i For existing schemes where the threshold is exceeded, feature data and labels are collected, specifically including the following steps:
[0029] Select accuracy Pa i For existing solutions where the number of devices exceeds a set threshold, feature data and tags are collected; if the number of collected servers is less than N, feature data and tags are collected separately for the uncollected devices.
[0030] If the selected existing solution cannot collect server feature data and tags that meet the requirements, then manual feature collection will be used.
[0031] Finally, feature terms are extracted from the collected feature data and labels.
[0032] The specific features include TTL value, whether the server is case-sensitive, whether the commonly used database port is open, whether the commonly used middleware port is open, the protocol used by the database, and the protocol used by the middleware.
[0033] Step S3, which involves constructing an initial intelligent identification model for components based on a neural network, specifically includes the following steps:
[0034] A fully connected neural network is used as the initial model for intelligent component recognition.
[0035] A fully connected neural network has L layers, and the set of neurons in each layer is represented as {π1, π2, ..., π}. L The number of neurons in the input and output layers is determined based on the dimensions of the input and label features.
[0036] A genetic algorithm was used to determine the number of layers L and the number of hidden neurons in a fully connected neural network.
[0037] The method of using a genetic algorithm to determine the number of layers L and the number of hidden layer neurons in a fully connected neural network specifically includes the following steps:
[0038] A. Chromosome encoding uses an integer encoding scheme: one chromosome is represented as... The chromosome length is L-1. Indicates the number of neurons in layer l and This represents the minimum number of neurons in each layer. Let be the maximum number of neurons in each layer; since the number of hidden layers is a variable, therefore only the front One element is activated;
[0039] B. Initialize the population size to N. p Each individual is randomly generated according to an integer encoding scheme; the population generation t is initialized to 0, and the variable p t Denotes the population of generation t;
[0040] C. Constructing Neural Networks: Build a neural network model for each individual. Let i represent the neural network model corresponding to individual i in generation t, and then train the model;
[0041] The following function is used as Loss function:
[0042]
[0043] In the formula for The loss function; Batch is a set of samples processed simultaneously during training, that is, a small batch of data randomly selected from the training data during training; w k As weight and For input y j Time model The value of the k-th element in the output vector; y j,k For input y j The value at the k-th position;
[0044] Using loss function Training a neural network model;
[0045] D. Evaluating model fitness: Model After training, the trained model is tested; the fitness of the model is measured using a loss function.
[0046] Assess the fitness of each individual in the population. This represents the fitness of the i-th individual;
[0047] E. Selection, Crossover, and Mutation: A roulette wheel selection operator is used for selection. For the parents, a two-point crossover operator is used to cross over all the parents to obtain two offspring. Then, a random single-point mutation operator is used on the two offspring to generate the final offspring.
[0048] F. Assess the fitness of each individual in the offspring population and select the next generation: o t Let t represent the offspring population generated from the t-th generation population. For each individual in the population, the following is performed: if the value range of a certain gene exceeds a set threshold, the value of that gene is replaced with the set threshold. Then, a neural network model corresponding to each individual in the offspring population is built and trained, and the fitness of each offspring individual is evaluated.
[0049] The parent and offspring populations are merged. The individual with the highest fitness is retained for the next generation. The remaining individuals are selected and retained according to their fitness values; the lower the fitness value, the lower the probability of being retained. Finally, N individuals are selected. p Individuals form the next generation population p t Then increment the population generation t by 1.
[0050] G. Repeat steps C to F until the population generation t is greater than the set value maxt; at this point, return the result of the optimal individual, where C* represents the chromosome encoding result of the optimal individual and M* represents the corresponding neural network model.
[0051] The aforementioned Specifically, it is calculated using the following formula:
[0052]
[0053] In the formula N p σ is the initial population size; ε is the scaling factor; t Let be the variance of fitness values in the t-th generation population.
[0054] The set value maxt is specifically calculated using the following formula:
[0055]
[0056] In the formula, min() is the minimum value operation; b1 is the upper bound of the set number of iterations; b1 is the baseline value of the set number of iterations; k1 is the first intermediate parameter, and k2 is a set real constant greater than 0. Let be the fitness value of the best individual in the t-th generation of the population. is the fitness value of the best individual in the (t-1)th generation population.
[0057] This invention also provides a system for implementing the neural network-based intelligent component identification method, comprising a system determination module, a data acquisition module, a model building module, a model training module, and an intelligent identification module; the system determination module, data acquisition module, model building module, model training module, and intelligent identification module are connected in series; the system determination module is used to determine a known power information system and upload the data to the data acquisition module; the data acquisition module is used to acquire data information of the power information system in the determined power information system using a hybrid sampling method based on the received data, construct a training dataset, and upload the data to the model building module; the model building module is used to construct an initial intelligent component identification model based on a neural network using the received data, and upload the data to the model training module; the model training module is used to train the constructed initial intelligent component identification model using the constructed training dataset using the received data, obtain an intelligent component identification model, and upload the data to the intelligent identification module; the intelligent identification module is used to perform intelligent component identification of the target power information system using the obtained intelligent component identification model based on the received data.
[0058] The component intelligent identification method and system based on neural networks provided by this invention adopts a hybrid data acquisition method to reduce the cost of data acquisition while ensuring the accuracy of data acquisition, and then uses an innovative neural network to realize the intelligent identification of components; therefore, this invention can be applied to large-scale scenarios, and has higher reliability and better accuracy. Attached Figure Description
[0059] Figure 1This is a schematic diagram of the method flow of the present invention.
[0060] Figure 2 This is a schematic diagram of the functional modules of the system of the present invention. Detailed Implementation
[0061] like Figure 1 The diagram shown is a flowchart of the method of the present invention: The component intelligent recognition method based on neural networks disclosed in this invention includes the following steps:
[0062] S1. Identify known power information systems; specifically including the following steps:
[0063] Identify known power information systems;
[0064] The known power information system specifically meets the following requirements:
[0065] The power information system has N servers; each server host has 0 to Num databases and middleware installed;
[0066] There are three types of power information systems: H-type mainframes, D-type databases, and M-type middleware.
[0067] The power information system has at least one server with global access permissions, which can access any device in the power information system;
[0068] All devices in the power information system have fixed IP addresses and have been centrally registered.
[0069] S2. In the power information system determined in step S1, a hybrid sampling method is used to acquire data information from the power information system and construct a training dataset; specifically, this includes the following steps:
[0070] Initial scan results are generated using scanning tools: Several existing methods (such as nmap and Fing tools, a total of two methods) are used to detect the host type, database type and middleware type of the server in the power information system;
[0071] For the same server, if all the obtained components are of the same type, then that server is marked as a candidate server; assuming a total of R1 candidate servers are obtained, if... Then randomly select from R1 candidate servers Each candidate server is manually reviewed; if... Then, all candidate servers are manually reviewed, and a selection is made from the remaining servers. Each server undergoes manual review; this manual review is to verify the accuracy of the existing solutions used for evaluation. The accuracy of the benchmark data The scale of the baseline data;
[0072] Statistical accuracy based on benchmark data: Let the i-th existing scheme detect S... i Servers, of which The server was manually checked and found to be correct. If the detection results from the server are incorrect, then the accuracy Pa of the i-th existing solution will be reduced. i for:
[0073]
[0074] In the formula, ω is the penalty factor; The term reflects the server recognition rate and accuracy of the i-th existing scheme. Obviously, if the detected... All servers belong to the S detected by tool i. i Within the server, and with a high accuracy rate, then The value will be higher; conversely, the failure to meet any condition will cause the value to decrease. The term is a penalty term, which penalizes errors in detection, thereby allowing for a more accurate measurement of the accuracy of tool i.
[0075] Select accuracy Pa i For existing schemes where the threshold is exceeded, feature data and labels are collected; {x j ,y j} represents the j-th data item in the collected dataset, x j Let y be the feature vector of the j-th data point. j Let be the label vector of the j-th data; where the label vector is a binary vector of dimension H+D+M, and the value of the kk-th bit in the vector is 1, indicating that the server has the component corresponding to the kk-th bit, and the value of the kk-th bit in the vector is 0, indicating that the server does not have the component corresponding to the kk-th bit.
[0076] In practical implementation, the accuracy Pa is selected. i For existing solutions where the number of devices exceeds a set threshold, feature data and tags are collected; if the number of collected servers is less than N, feature data and tags are collected separately for the uncollected devices.
[0077] If the selected existing solution cannot collect server feature data and tags that meet the requirements, then manual feature collection will be used.
[0078] Finally, feature items are extracted from the collected feature data and labels; the feature items specifically include TTL value, whether the server is case sensitive, whether the commonly used database port is open, whether the commonly used middleware port is open, the protocol used by the database, and the protocol used by the middleware, etc.
[0079] S3. Based on neural networks, construct an initial model for intelligent component recognition; specifically including the following steps:
[0080] Since the feature dimension of the collected data is not very high, a fully connected neural network is used as the initial model for component intelligent recognition to reduce the complexity of the algorithm.
[0081] A fully connected neural network has L layers, and the set of neurons in each layer is represented as {π1, π2, ..., π}. L The number of neurons in the input and output layers is determined based on the dimensions of the input and label features.
[0082] Since the number of layers and neurons in a neural network significantly impacts model performance, determining these hyperparameters is a challenging problem. Therefore, the following genetic algorithm (an improved version) is used to determine the number of layers L and the number of hidden neurons in a fully connected neural network. The specific steps include:
[0083] A. Chromosome encoding uses an integer encoding scheme: one chromosome is represented as... The chromosome length is L-1. Indicates the number of neurons in layer l and This represents the minimum number of neurons in each layer. Let be the maximum number of neurons in each layer; since the number of hidden layers is a variable, therefore only the front One element is activated;
[0084] B. Initialize the population size to N. p Each individual is randomly generated according to an integer encoding scheme; the population generation t is initialized to 0, and the variable p t Denotes the population of generation t;
[0085] C. Constructing a neural network: Construct a neural network model for each individual. Since the number of layers and the number of neurons in each layer are already determined for each individual, the structure of the neural network model is unique. Let i represent the neural network model corresponding to individual i in generation t, and then train the model;
[0086] The following function is used as Loss function:
[0087]
[0088] In the formula for The loss function; Batch is a set of samples processed simultaneously during training, that is, a small batch of data randomly selected from the training data during training; w k As weight and For input y j Time model The value of the k-th element in the output vector; y j,k For input y j The value at the k-th position;
[0089] Unlike the traditional absolute value loss function, this invention uses a weighted approach to determine the importance of different elements in the label. For example, if an application considers correctly identifying the host to be the most important, followed by identifying the middleware, and finally the database, then it is tolerable for the neural network to fail to correctly identify the database. Therefore, this design allows the neural network to focus on specific aspects during iterative training, optimizes resource allocation during the training process, and improves the performance of the neural network in this task.
[0090] Using loss function Neural network models can be trained using methods such as stochastic gradient descent;
[0091] D. Evaluating model fitness: Model After training, the trained model is tested; the fitness of the model is measured using a loss function.
[0092] Assess the fitness of each individual in the population. This represents the fitness of the i-th individual;
[0093] E. Selection, Crossover, and Mutation: A roulette wheel selection operator is used for selection. For the parents, a two-point crossover operator is used to cross over all the parents to obtain two offspring. Then, a random single-point mutation operator is used on the two offspring to generate the final offspring.
[0094] In practice, the following formula is used to calculate the result.
[0095]
[0096] In the formula N p σ is the initial population size; ε is the scaling factor; tLet be the variance of the fitness values in the t-th generation population. A smaller variance indicates less data dispersion, meaning smaller differences between individuals and poorer population diversity. In this case, more parent pairs should be selected to increase individual diversity and improve the algorithm's ability to search for the optimal solution. Conversely, a larger variance indicates greater differences between individuals and higher population diversity. Therefore, fewer parent pairs should be selected to reduce the algorithm's complexity. This design of the present invention can achieve the above functions, reducing the algorithm's complexity and helping to maintain population diversity, thus avoiding getting trapped in local optima.
[0097] F. Assess the fitness of each individual in the offspring population and select the next generation: o t Let t represent the offspring population generated from the t-th generation population. For each individual in the population, the following is performed: if the value range of a certain gene exceeds a set threshold, the value of that gene is replaced with the set threshold. Then, a neural network model corresponding to each individual in the offspring population is built and trained, and the fitness of each offspring individual is evaluated.
[0098] The parent and offspring populations are merged. The individual with the highest fitness is retained for the next generation. The remaining individuals are selected and retained according to their fitness values; the lower the fitness value, the lower the probability of being retained. Finally, N individuals are selected. p Individuals form the next generation population p t Then increment the population generation t by 1.
[0099] G. Repeat steps C to F until the population generation t is greater than the set value maxt; at this point, return the result of the best individual, where C* represents the chromosome encoding result of the best individual and M* represents the corresponding neural network model;
[0100] In practice, the following formula is used to calculate maxt:
[0101]
[0102] In the formula, min() is the minimum value operation; b1 is the upper bound of the set number of iterations; b1 is the baseline value of the set number of iterations; k1 is the first intermediate parameter, and k2 is a set real constant greater than 0. Let be the fitness value of the best individual in the t-th generation of the population. is the fitness value of the best individual in the (t-1)th generation of the population;
[0103] The formula above is used to calculate maxt, which is essentially an adaptive iteration method. Its principle is as follows: if the difference between the fitness value of the best individual and the fitness value of the best individual in the previous generation is large, it indicates that the algorithm has not yet converged. In this case, the number of iterations needs to be increased to increase the chance of finding a better solution. Conversely, if the difference in fitness between the best individuals in the two generations is small, it indicates that the algorithm is close to convergence. In this case, continuing with a large number of iterations is not very meaningful, so the number of iterations needs to be reduced. In the formula for maxt, weighting k1 further ensures that the larger the difference, the larger the adaptive maxt value; otherwise, the smaller the adaptive maxt value.
[0104] S4. Using the training dataset constructed in step S2, train the initial component intelligent recognition model constructed in step S3 to obtain the component intelligent recognition model;
[0105] S5. Using the component intelligent identification model obtained in step S4, perform intelligent identification of components in the target power information system.
[0106] This invention addresses the problem of intelligent component identification in power information systems by designing a component identification method based on neural networks. The method employs a hybrid data acquisition approach, reducing data acquisition costs while maintaining a certain level of accuracy. To determine the optimal neural network structure, an improved genetic algorithm is proposed, enhancing final network performance while avoiding the subjectivity of manual structure determination. Furthermore, the improved genetic algorithm incorporates an adaptive selection operator to balance algorithm performance and time complexity. Additionally, an adaptive maximum iteration count method is designed to adaptively control the iteration count, preventing it from becoming trapped in local optima or terminating iterations before convergence.
[0107] like Figure 2The diagram shows the functional modules of the system of this invention: The system for implementing the neural network-based intelligent component identification method disclosed in this invention includes a system determination module, a data acquisition module, a model construction module, a model training module, and an intelligent identification module; the system determination module, data acquisition module, model construction module, model training module, and intelligent identification module are connected in series; the system determination module is used to determine a known power information system and upload the data to the data acquisition module; the data acquisition module is used to acquire data information of the power information system in the determined power information system using a hybrid sampling method based on the received data, construct a training dataset, and upload the data to the model construction module; the model construction module is used to construct an initial intelligent component identification model based on a neural network based on the received data and upload the data to the model training module; the model training module is used to train the constructed initial intelligent component identification model using the constructed training dataset based on the received data to obtain an intelligent component identification model and upload the data to the intelligent identification module; the intelligent identification module is used to perform intelligent component identification of the target power information system based on the received data and the obtained intelligent component identification model.
Claims
1. A component intelligent recognition method based on neural networks, comprising the following steps: S1. Identify known power information systems; S2. In the power information system determined in step S1, a hybrid sampling method is used to acquire data information from the power information system and construct a training dataset; specifically, this includes the following steps: There are three types of power information systems: H-type mainframes, D-type databases, and M-type middleware. Initial scan results were generated using scanning tools: several existing methods were employed to detect the host type, database type, and middleware type of servers in the power information system; For the same server, if all the components obtained are of the same type, then that server is marked as a candidate server; the total number of obtained components is set to... If there are candidate servers, Then for Randomly selected from candidate servers Each candidate server is manually reviewed; if... Then, all candidate servers are manually reviewed, and a selection is made from the remaining servers. Each server undergoes manual review to verify the accuracy of the existing solutions used for evaluation. The accuracy of the benchmark data The scale of the baseline data; Statistical accuracy based on benchmark data: The accuracy of the i-th existing scheme is defined as follows: Servers, of which The server was manually checked and found to be correct. If the test results from the server are incorrect, then the accuracy of the i-th existing solution will be... for: In the formula As a penalty factor; The term reflects the server recognition rate and accuracy of the i-th existing scheme; The item is a penalty item; Select accuracy For existing schemes where the threshold is exceeded, feature data and labels are collected. This represents the j-th data item in the collected dataset. Let j be the feature vector of the j-th data. Let be the label vector of the j-th data item; where the label vector is a vector with dimension . A binary vector, where a value of 1 for the kk-th bit indicates that the server has the component corresponding to the kk-th bit, and a value of 0 for the kk-th bit indicates that the server does not have the component corresponding to the kk-th bit. S3. Construct an initial model for intelligent component recognition based on neural networks; specifically including the following steps: A fully connected neural network is used as the initial model for intelligent component recognition. The number of layers in a fully connected neural network is The set of neurons in each layer is represented as The number of neurons in the input and output layers is determined based on the dimensions of the input and label features. Genetic algorithms are used to determine the number of layers in a fully connected neural network. and the number of neurons in the hidden layer; S4. Using the training dataset constructed in step S2, train the initial component intelligent recognition model constructed in step S3 to obtain the component intelligent recognition model; S5. Using the component intelligent identification model obtained in step S4, perform intelligent identification of components in the target power information system.
2. The component intelligent recognition method based on neural networks according to claim 1, characterized in that... Step S1, which involves determining a known power information system, specifically includes the following steps: Identify known power information systems; The known power information system specifically meets the following requirements: The power information system has N servers; each server host has 0 to Num databases and middleware installed. There are three types of power information systems: H-type mainframes, D-type databases, and M-type middleware. The power information system must have at least one server with global access permissions, enabling it to access any device within the power information system; All devices in the power information system have fixed IP addresses and have been centrally registered.
3. The component intelligent identification method based on neural networks according to claim 2, characterized in that... The accuracy of the selection For existing schemes where the threshold is exceeded, feature data and labels are collected, specifically including the following steps: Select accuracy For existing solutions where the number of devices exceeds a set threshold, feature data and tags are collected; if the number of collected servers is less than N, feature data and tags are collected separately for the uncollected devices. If the selected existing solution cannot collect server feature data and tags that meet the requirements, then manual feature collection will be used. Finally, feature terms are extracted from the collected feature data and labels.
4. The component intelligent identification method based on neural networks according to claim 3, characterized in that... The specific features include TTL value, whether the server is case-sensitive, whether the commonly used database port is open, whether the commonly used middleware port is open, the protocol used by the database, and the protocol used by the middleware.
5. The component intelligent identification method based on neural networks according to claim 4, characterized in that... The method described above uses a genetic algorithm to determine the number of layers in a fully connected neural network. The number of neurons in the hidden layer is determined by the following steps: A. Chromosome encoding uses an integer encoding scheme: one chromosome is represented as... The chromosome length is , , Indicates the number of neurons in layer l and , This represents the minimum number of neurons in each layer. This represents the maximum number of neurons in each layer; since the number of hidden layers is a variable, therefore... only the front One element is activated; B. Initialize the population size to Each individual is randomly generated according to an integer encoding scheme; the population generation is initialized. The variable is 0. Denotes the population of generation t; C. Constructing Neural Networks: Build a neural network model for each individual. Let i represent the neural network model corresponding to individual i in generation t, and then train the model; The following function is used as Loss function: In the formula for The loss function; This refers to a set of samples processed simultaneously during the training process. As weight and ; For input Time model The value of the k-th element in the output vector; For input The value at the k-th position; Using loss function Train the neural network model; D. Evaluate model fitness: Model After training, the trained model is tested; the fitness of the model is measured using a loss function. Assess the fitness of each individual in the population. This represents the fitness of the i-th individual; E. Selection, Crossover, and Mutation: A roulette wheel selection operator is used for selection. For the parents, a two-point crossover operator is used to cross over all the parents to obtain two offspring. Then, a random single-point mutation operator is used on the two offspring to generate the final offspring. F. Assess the fitness of each individual in the offspring population and select the next generation: This represents the offspring population generated by the t-th generation population. For each individual in the population, the following is performed: if the value range of a certain gene exceeds a set threshold, then the set threshold is used to replace the value of that gene. Then, a neural network model corresponding to each individual in the offspring population is built and trained, and the fitness of each offspring individual is evaluated. The parent and offspring populations are merged. The individual with the highest fitness is retained for the next generation. The remaining individuals are selected and retained according to their fitness values; the lower the fitness value, the lower the probability of being retained. The final selection... Individuals form the next generation population. Then let the population number of generations The value increases by 1; G. Repeat steps C to F until the population reaches the required number of generations. The value is greater than the set value At this point, return the result of the optimal individual. This represents the chromosome coding result of the optimal individual. This represents the corresponding neural network model.
6. The component intelligent recognition method based on neural networks according to claim 5, characterized in that... The aforementioned Specifically, it is calculated using the following formula: In the formula The initial population size; This is the scaling factor; Let be the variance of fitness values in the t-th generation population.
7. The component intelligent recognition method based on neural networks according to claim 6, characterized in that... The set value Specifically, it is calculated using the following formula: In the formula This is an operation to retrieve the minimum value; This is the upper bound of the set number of iterations; The baseline value for the set number of iterations; It is the first intermediate parameter, and , For a set real constant greater than 0, Let be the fitness value of the best individual in the t-th generation of the population. is the fitness value of the best individual in the (t-1)th generation population.
8. A system for implementing the neural network-based intelligent component recognition method according to any one of claims 1 to 7, characterized in that... It includes a system identification module, a data acquisition module, a model building module, a model training module, and an intelligent recognition module; these modules are connected in series. The system identification module is used to identify known power information systems and upload the data to the data acquisition module. The data acquisition module is used to acquire data information of the power information systems in the identified power information systems using a hybrid sampling method based on the received data, construct a training dataset, and upload the data to the model building module. The model building module is used to build an initial model based on the received data and a neural network, and then upload the data to the model training module. The model training module is used to train the initial component intelligent recognition model based on the received data and the constructed training dataset, thereby obtaining the component intelligent recognition model, and then uploading the data to the intelligent recognition module. The intelligent identification module is used to perform intelligent identification of components of the target power information system based on the received data and the obtained intelligent identification model of the components.