Aerial detection task identification coordination system and method

Abstract

The application discloses an aviation detection task identification coordination system and method, relates to the technical field of program control, a task identification module performs format conversion, redundancy deduplication and feature extraction on an input information set, identifies a task type by using a Transformer model and constructs a task feature package, a scheme formulation module determines an optimal program execution group according to DRL decision and generates an optimal parameter vector through a domain adversarial network, integrates a detection execution scheme, and a monitoring and evaluation module monitors a detection progress in real time and feeds back and updates a knowledge graph after detection is completed, thereby improving aviation detection accuracy, efficiency and decision robustness.

Description

Technical Field

[0001] This invention relates to the field of program control technology, and specifically to an aviation inspection mission identification and coordination system and method. Background Technology

[0002] General aviation testing is a core component of ensuring the safe operation of general aviation equipment. It requires multi-dimensional testing of equipment performance, structure, and anti-interference capabilities to identify potential faults and is of critical importance to the safe development of the general aviation industry. Among these, temperature-altitude testing, explosive atmosphere testing, lightning-induced transient sensitivity testing, and aircraft power supply characteristic testing are all key steps in general aviation testing. They respectively undertake core testing functions such as simulated environment adaptability, extreme scenario tolerance, electromagnetic interference resistance, and power supply stability, and are important components to ensure the comprehensiveness of testing.

[0003] With the rapid development of the general aviation industry, the complexity of aviation inspection tasks has increased significantly, and the requirements for the coordination of multi-dimensional inspections are higher. The inspection environment needs to cope with real-time changes in factors such as wind speed, temperature and humidity, and electromagnetic interference. In addition, the concurrent scenarios of multiple inspection tasks are becoming more and more frequent, which further increases the demand for the flexibility and adaptability of inspection solutions.

[0004] Existing aviation inspection technologies have significant shortcomings: the execution sequence of inspection procedures such as temperature-altitude detection, explosive atmosphere testing, lightning-induced transient sensitivity detection, and aircraft power characteristics detection relies on human experience or preset fixed rules, and cannot be dynamically adjusted according to the characteristics of the inspected object and real-time environmental data for the current task, which easily leads to detection errors. The inspection procedure parameters mostly use default parameters or rely on human experience, which not only increases costs, but also results in low accuracy of inspection results due to insufficient parameter adaptation. Summary of the Invention

[0005] This invention addresses the shortcomings of existing technologies by proposing an aviation inspection task identification and coordination system and method. It processes the input information set and identifies the task type based on the Transformer model, constructs a task feature package, and integrates the results to obtain an inspection execution plan based on DRL decision-making and domain adversarial network correction. The system monitors the inspection in real time, calculates the inspection score, and updates the knowledge graph, effectively improving the accuracy, efficiency, and decision robustness of aviation inspection.

[0006] The technical solution to achieve the purpose of this invention is as follows:

[0007] The aviation inspection mission identification and coordination system includes a mission identification module, a scheme development module, and a monitoring and evaluation module.

[0008] The task identification module acquires the input information set and obtains a unified feature matrix through format conversion, redundancy removal and feature extraction in sequence. Based on the Transformer model, it identifies the task type and constructs a task feature package.

[0009] The scheme formulation module obtains the task feature package, calls the program number and default parameters of relevant detection programs based on the knowledge graph, defines the simulated detection scenario and arranges and combines relevant detection programs to construct an intelligent agent set, randomly selects intelligent agents, performs simulated detection based on the corresponding default parameter vector and calculates the simulated detection score, updates the value table through Q network learning and uses a greedy strategy to reselect intelligent agents, iteratively determines the optimal program execution group corresponding to the optimal intelligent agent, and determines the benchmark optimal parameter vector and benchmark task feature package based on the knowledge graph. Based on the domain adversarial network, the benchmark optimal parameter vector is corrected by the difference between the flattened result of the task feature package and the benchmark task feature package to obtain the optimal parameter vector. The detection execution scheme is obtained by integrating the task feature package, the optimal program execution group and the optimal parameter vector.

[0010] The monitoring and evaluation module constructs a program node table based on the detection execution plan, monitors the execution of relevant detection programs in real time, and marks the nodes in the program node table. When all nodes are marked, it obtains the detection result set and detection duration, calculates the detection score, and fills the knowledge graph with the task feature package, the optimal program execution group, the optimal parameter vector, the detection result set, and the detection score.

[0011] Furthermore, the task identification module includes an access processing unit, a feature generation unit, and a type identification unit;

[0012] The access processing unit obtains the input information set and obtains the input dataset through format conversion. It uses the SimHash algorithm to calculate the semantic similarity between any two input data and removes duplicates based on semantic similarity to obtain the standard dataset. Each input data in the input dataset includes requirement input data, object input data, and environment input data. Each standard data in the standard dataset includes standard requirement data, standard object data, and standard environment data.

[0013] The feature generation unit uses the BERT model to transform each standard requirement data into a semantic feature vector and concatenates them to obtain a semantic feature matrix. It transforms the object type in the standard object data or the parameter type in the standard environment data into word vectors to obtain type codes. It then concatenates the normalized values ​​of the corresponding object parameters or parameter values ​​with the type codes to obtain the physical feature vector of each standard object data or the environment feature vector of each standard environment data. These are then concatenated to obtain the physical feature matrix and the environment feature matrix, respectively. Finally, it uses linear modulation to unify the dimensions of the semantic feature matrix, the physical feature matrix, and the environment feature matrix and concatenates them to obtain a unified feature matrix.

[0014] The type recognition unit inputs the unified feature matrix into the Transformer model. Based on the self-attention mechanism and feedforward neural network, it obtains an abstract feature representation and compares it with the cosine similarity of the typical feature vectors of all task types in the preset task type feature library. It outputs the adaptation probability distribution of each task type through the Softmax function, selects the task type with the highest adaptation probability and determines the corresponding result type set, extracts the standard object dataset and standard environment dataset from the standard dataset, and combines them with the task type and result type set to form a task feature package.

[0015] Furthermore, for the detection requirement text, the access processing unit uses OCR recognition and regular expression matching to extract the requirement descriptions corresponding to different requirement types and integrates them to obtain the corresponding requirement input data. For the object table being inspected, Pandas is used to read the object parameters corresponding to different object types in the table and the field mapping table is used to directly map and obtain all object input data. For the environment data, Spark Streaming is used to confirm the parameter values ​​and system time corresponding to different parameter types and integrate them into the corresponding environment input data.

[0016] Furthermore, the solution development module includes a program decision-making unit, a parameter determination unit, and an integration and delivery unit;

[0017] The program decision unit acquires the task feature package and executes DRL decision, calls the program number and default parameters of the relevant detection program that matches the task type, defines the simulated detection scenario based on the task feature package, defines any permutation and combination of the program numbers of the relevant detection programs as agents, randomly selects an agent, constructs a default parameter vector based on the default parameters of all relevant detection programs in the agent, performs simulated detection in the simulated detection scenario and calculates the simulated detection score, learns the reference simulated detection score to update the value table through the Q network, uses a greedy strategy combined with the value table to select the next agent, iterates to determine the optimal agent, and obtains the optimal program execution group. The simulated detection score is designed with reference to the completion of the result and the detection efficiency.

[0018] The parameter determination unit retrieves historical optimal parameter vectors and historical task feature packages that simultaneously satisfy the optimal program execution group and the highest detection score from the knowledge graph and uses them as the baseline optimal parameter vector and baseline task feature package, respectively. Based on the generator in the domain adversarial network, the baseline optimal parameter vector is corrected by the difference between the flattening results of the task feature package and the baseline task feature package to obtain the optimal parameter vector. The domain adversarial network also includes a discriminator to judge the authenticity and reliability of the optimal parameter vector. In the offline stage, the generator and discriminator are trained collaboratively using the adversarial win-win idea. In the online stage, the trained generator is used to generate the optimal parameter vector.

[0019] The integrated delivery unit integrates the task feature package, the optimal program execution group, and the optimal parameter vector to obtain the detection execution plan and sends it synchronously to the detection terminal and the monitoring and evaluation module.

[0020] Furthermore, DRL decision-making includes the following steps:

[0021] Identify the task type in the task feature package and call the program number and default parameters of each relevant detection program from the built-in program library. Construct a simulated detection scenario using the standard object dataset and standard environment dataset in the task feature package.

[0022] Arbitrarily arrange and combine the program numbers of all relevant detection programs to obtain all feasible detection program execution groups. Define the detection program execution group as an agent. Set each agent according to the corresponding default parameter vector. The default parameter vector includes the default parameters of all relevant detection programs in the agent.

[0023] Initialize the value table, learning rate, discount factor, exploration probability, and total number of iterations. Randomly select an agent to perform simulated detection, obtain the simulated detection result set and the simulated detection time. Divide the number of matching result types between the simulated detection result set and the result type set in the task feature package by the total number of result types in the result type set to obtain the result completion rate. Only set the sign function with a result completion rate of 1 to a value of 1. Normalize the negative of the simulated detection time using the Sigmoid function to obtain the efficiency score. Weight the sum of the sign function and the efficiency score to obtain the simulated detection score.

[0024] The agent's value is updated according to the value update formula learned by the Q network. The agent's value is equal to the original value plus the product of the learning rate and the simulated test score, plus the product of the learning rate, the discount factor and the maximum value in the value table, and then minus the product of the learning rate and the original value.

[0025] A random number between 0 and 1 is generated according to a greedy strategy and compared with the exploration probability to select an agent for the next simulation test. If the random number is less than or equal to the exploration probability, an agent that has not been selected is selected. When the random number is greater than the exploration probability, the agent with the highest value is selected from the value table. After each selection, the exploration probability is reduced by the exploration step size. The exploration probability is required to be greater than the exploration step size multiplied by the total number of iterations.

[0026] Repeat the iteration until the total number of iterations is reached, select the agent with the highest value in the value table as the optimal agent, and determine the optimal program execution group.

[0027] Furthermore, the parameter determination unit pre-collects multiple input information sets and detection result sets, processes the input information sets to obtain corresponding task feature packages and optimal program execution groups. For each task feature package, based on the optimal program execution group, it collaboratively retrieves all associated historical optimal parameter vectors, historical task feature packages, and detection scores in the knowledge graph. The historical optimal parameter vector and historical task feature package with the highest detection score are respectively used as the benchmark optimal parameter vector and benchmark task feature package corresponding to the optimal program execution group. Using the optimal program execution group as an index, and combining the benchmark optimal parameter vector, benchmark task feature package, and multiple corresponding task feature packages and detection result sets corresponding to each optimal program execution group, an index training dataset is constructed, and a domain adversarial network is trained based on the adversarial win-win concept.

[0028] Furthermore, the generator flattens the task feature package and the baseline task feature package with the same index into task feature vectors and baseline task feature vectors respectively, and calculates the difference to obtain the task deviation vector. The fully connected layer is used to map the task deviation vector into a parameter correction vector. The baseline parameter vector and the parameter correction vector are fused through the fully connected layer to obtain the optimal parameter vector.

[0029] Furthermore, the discriminator includes a classification discriminator and a simulation detector. The classification discriminator uses a fully connected layer to map the optimal parameter vector to the true probability, while the simulation detector performs simulation detection based on the task feature package, the optimal program execution group, and the optimal parameter vector to obtain a set of simulation detection results.

[0030] Furthermore, when the task feature packets in the indexed training dataset are input into the domain adversarial network for training, the cross-entropy loss of the classifier is calculated, and the mean squared error loss between the simulated detection result set and the actual detection result set output by the simulated detector is calculated. The generator updates its parameters using the cross-entropy loss, the mean squared error loss, and the gradient of the generator parameters. The classifier updates its parameters using the cross-entropy loss and the gradient of the classifier parameters, and feeds back to the generator through gradient inversion. The simulated detector updates its parameters using the mean squared error loss and the gradient of the detector parameters, and positively notifies the generator. When the cross-entropy loss converges to 0.693 and the mean squared error loss is minimized, the training of the domain adversarial network is complete.

[0031] An aviation inspection mission identification and coordination method, used to implement an aviation inspection mission identification and coordination system, includes the following steps:

[0032] The input information set is acquired and its format is converted and redundancy is removed to obtain a standard dataset, which includes standard requirement data, standard object data, and standard environment data.

[0033] Features are extracted from each standard data in the standard dataset and modulated and integrated into a unified feature matrix. The unified feature matrix is ​​then input into the Transformer model to determine the task type and result type set. This matrix is ​​then combined with the standard object dataset and standard environment dataset in the standard dataset to construct a task feature package.

[0034] Based on the knowledge graph, the program numbers and default parameters of all relevant detection programs that match the task type are called. Combined with the standard object dataset and standard environment dataset, a simulated detection scenario is defined. Each permutation and combination of program numbers is regarded as an agent. Agents are randomly selected and simulated detection is performed based on the corresponding default parameter vector. The simulated detection score is calculated and the value table is updated through Q network learning. The greedy strategy is used to iteratively select agents in combination with the value table until the total number of iterations is reached to obtain the optimal agent, i.e. the optimal program execution group.

[0035] The benchmark optimal parameter vector and benchmark task feature package that meet the same optimal program execution group and have the highest detection score are retrieved from the knowledge graph. Based on the domain adversarial network generator, the benchmark optimal parameter vector is corrected by the difference between the flattening results of the task feature package and the benchmark task feature package. The optimal parameter vector is then obtained and integrated with the task feature package and the optimal program execution group to form a detection execution scheme, which is then delivered to the detection terminal.

[0036] Based on the optimal program execution group, a program node table is constructed and the nodes corresponding to the relevant detection programs that have been executed are marked until all nodes are marked. The detection result set and detection time are obtained and the actual detection score is calculated. The task feature package, optimal program execution group, optimal parameter vector, detection result set, and detection score are filled into the knowledge graph.

[0037] Compared with existing technologies, the significant advantages of this invention are that it can accurately identify task types and construct task feature packages based on the Transformer model after format conversion, redundancy removal, and feature extraction of the input information set. Then, it uses DRL decision iteration to determine the optimal program execution group. At the same time, it uses domain adversarial network to correct the baseline optimal parameter vector by utilizing the difference between the task feature package and the baseline task feature package, thereby generating an accurate detection execution plan. Simultaneously, it monitors the execution of the detection program in real time, calculates the detection results and detection scores, and fills relevant information such as the task feature package and the optimal program execution group into the knowledge graph for updating, providing knowledge support for subsequent aviation detection, and effectively improving the accuracy, efficiency, and decision robustness of detection. Attached Figure Description

[0038] Figure 1 A schematic diagram of an aviation inspection mission identification and coordination system;

[0039] Figure 2 DRL decision-making flowchart;

[0040] Figure 3 This is a schematic diagram of domain adversarial network training.

[0041] Figure 4 Flowchart for identifying and coordinating methods for aviation inspection tasks. Detailed Implementation

[0042] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0043] Example 1

[0044] like Figure 1 As shown, a specific embodiment of the present invention discloses an aviation inspection mission identification and coordination system, including a mission identification module, a scheme formulation module, and a monitoring and evaluation module;

[0045] The task recognition module acquires the input information set and obtains a standard dataset through format conversion and redundancy removal. It extracts the semantic, physical, and environmental features of each standard data in the standard dataset and integrates them into a unified feature matrix. Based on the pre-trained Transformer model, it identifies the task type and constructs a task feature package.

[0046] The scheme formulation module obtains the task feature package, calls the program number and default parameters of all relevant detection programs based on the knowledge graph, defines the simulated detection scenario, constructs an intelligent agent set by arranging and combining relevant detection programs, selects intelligent agents and performs simulated detection based on the corresponding default parameter vector, calculates the simulated detection score and updates the value table through Q-network learning, selects the next intelligent agent using a greedy strategy combined with the value table, iteratively determines the optimal program execution group corresponding to the optimal intelligent agent, and retrieves the corresponding benchmark optimal parameter vector and benchmark task feature package from the knowledge graph. Based on the domain adversarial network generator, the benchmark optimal parameter vector is corrected by the difference between the flattening result of the task feature package and the benchmark task feature package to obtain the optimal parameter vector. The task feature package, the optimal program execution group and the optimal parameter vector are integrated to obtain the detection execution scheme and are simultaneously sent to the monitoring and evaluation module and the detection terminal. The knowledge graph is pre-constructed and will not be described in detail in this application's technical solution.

[0047] The monitoring and evaluation module constructs a program node table based on the optimal program execution group in the detection execution plan. It monitors the execution of relevant detection programs on the detection terminal in real time through Socket communication. Whenever a relevant detection program completes execution, the corresponding node in the program node table is marked until all nodes in the program node table are marked, thus obtaining the detection result set and detection duration. The detection score is obtained by reusing the simulation detection score calculation method in the scheme formulation module. The task feature package, optimal program execution group, optimal parameter vector, detection result set, and detection score are filled into the knowledge graph according to a preset format to achieve synchronous updates of the knowledge graph.

[0048] Furthermore, the task identification module includes an access processing unit, a feature generation unit, and a type identification unit;

[0049] The access processing unit obtains the input information set from the access port. Each input information set includes the detection requirement text, the object to be inspected table, and environmental data. The input information set is converted into an input dataset through format conversion. The SimHash algorithm is used to calculate the semantic similarity between any two input data. When the semantic similarity between two input data is greater than the semantic similarity threshold, one of the input data is removed to achieve redundancy deduplication and obtain a standard dataset. The object to be inspected table includes a design parameter table and a maintenance record table. The environmental data includes meteorological data and electromagnetic interference data. Each input data in the input information set includes requirement input data, object input data, and environmental input data. Each standard data in the standard dataset is the input data retained after redundancy deduplication, including standard requirement data, standard object data, and standard environmental data.

[0050] The feature generation unit adaptively performs different feature extractions based on the type and fields of the standard data. For each standard requirement data, a semantic feature vector is generated by fine-tuning the BERT model. The semantic feature vectors of different standard requirement data are concatenated row by row to obtain a semantic feature matrix. For each standard object data or standard environment data, the object type in the standard object data or the parameter type in the standard environment data is converted into word vectors based on Word2Vec to obtain a type code. The normalized value of the corresponding object parameter or parameter value is concatenated column by column with the type code to obtain the physical feature vector of each standard object data or the environment feature vector of each standard environment data. The physical feature vectors of different object types or the environment feature vectors of different parameter types are concatenated row by row to obtain the physical feature matrix and the environment feature matrix. The dimensions of the semantic feature matrix, the physical feature matrix, and the environment feature matrix are unified by linear modulation and concatenated to obtain a unified feature matrix.

[0051] The type recognition unit inputs a unified feature matrix into the Transformer model. Based on the self-attention mechanism, it assigns attention weights to features of different dimensions in the unified feature matrix to enhance the response strength of key features. It uses the attention weights to perform semantic association calculations on features of each dimension in the unified feature matrix, captures the contextual dependencies between features of different dimensions, and inputs them into a feedforward neural network for nonlinear mapping, transforming them into an abstract feature representation with task discrimination. The abstract feature representation is compared with the typical feature vectors of all task types in the preset task type feature library using cosine similarity. The Softmax function outputs the adaptation probability distribution of each task type. The task type with the highest adaptation probability is selected, and the result type set corresponding to the task type is determined based on the built-in task result table. Each standard object data and standard environment data in the standard dataset is combined into a standard object dataset and a standard environment dataset, respectively. These are combined with the task type and result type set to form a task feature package and delivered to the scheme formulation module. The result type set includes all result types that should be obtained after executing the detection task corresponding to the task type.

[0052] Furthermore, the access processing unit defines the fields for the input data. The fields for requirement input data include requirement type and requirement description; the fields for object input data include object type and object parameters; and the fields for environment input data include parameter type, parameter value, and timestamp. For the detection requirement text, OCR recognition and regular expression matching are used to extract the requirement descriptions corresponding to different requirement types, and these are integrated to generate requirement input data equal to the total number of requirement types. For the inspected object table, Pandas is used to read the object parameters corresponding to different object types in the table and a field mapping table is used to directly map them to obtain object input data equal to the total number of object types. For environment data, Spark Streaming is used to confirm the parameter values ​​corresponding to different parameter types and record the corresponding system time, and these are integrated to obtain environment input data equal to the total number of parameter types. The input data is in JSON format.

[0053] Furthermore, the solution development module includes a program decision-making unit, a parameter determination unit, and an integration and delivery unit;

[0054] The program decision unit acquires the task feature package and executes DRL decision-making. It calls the program numbers and default parameters of all relevant detection programs matching the task type from the built-in program library. Based on the standard object dataset and standard environment dataset in the task feature package, it collaboratively defines a simulated detection scenario. It defines any permutation and combination of the program numbers of relevant detection programs as agents. It randomly selects an agent, constructs a default parameter vector based on the default parameters of all relevant detection programs in the agent, and performs simulated detection in the simulated detection scenario. It calculates the simulated detection score, learns from the simulated detection score through the Q network to update the value table, and uses a greedy strategy combined with the value table to select the next agent. It iteratively determines the optimal agent and obtains the optimal program execution group. The simulated detection score is designed with reference to the completion of the results and the detection efficiency.

[0055] The parameter determination unit searches the knowledge graph based on the optimal program execution group to determine the historical optimal parameter vector and historical task feature package with the highest detection score, and uses them as the benchmark optimal parameter vector and benchmark task feature package, respectively. Based on the domain adversarial network generator, the benchmark optimal parameter vector is corrected by the difference between the flattening results of the task feature package and the benchmark task feature package to obtain the optimal parameter vector. The domain adversarial network also includes a discriminator to judge the authenticity and reliability of the optimal parameter vector. In the offline stage, the generator and discriminator are trained collaboratively using the adversarial win-win idea. After training, the generator parameters are fixed and the optimal parameter vector is generated by the generator in the online stage.

[0056] The integrated delivery unit calls the task feature package, the optimal program execution group, and the optimal parameter vector, and according to a fixed structure, integrates them to obtain a detection execution plan, which is then simultaneously sent to the detection terminal and the monitoring and evaluation module to execute the aviation detection and start the detection monitoring.

[0057] like Figure 2 As shown, further, DRL decision-making includes the following steps:

[0058] The task type in the task feature package is identified, and the program number and default parameters of each relevant detection program are called from the built-in program library. The virtual object under inspection and the virtual detection environment are constructed using the standard object dataset and the standard environment dataset in the task feature package. The simulated detection scenario is defined collaboratively. The virtual object under inspection and the virtual detection environment can be constructed using existing digital twin methods, which will not be elaborated on in this application.

[0059] Arbitrarily arrange and combine the program numbers in the program information set to obtain all non-repeating detection program execution groups that are related to the task type. Define each program execution group as an agent and construct an agent set. Set each agent according to the corresponding default parameter vector. The program execution group indicates all related detection programs and their execution order. The default parameter vector includes the default parameters determined when writing all related detection programs in the agent.

[0060] A value table is constructed based on the task type and agent and initialized to 0. The learning rate, discount factor, exploration probability and total number of iterations for DRL decision are configured. The value table stores the value of all agents.

[0061] A simulated detection is performed by randomly selecting an agent from the set of agents. The simulated detection result set and the simulation detection time are statistically analyzed. Result type sets are extracted from the task feature package. The completion rate of the simulated detection result set and result type sets is calculated. The completion rate is equal to the number of matching result types in the simulated detection result set and result type set divided by the total number of result types in the result type set. A result type is considered complete when each simulated detection result set corresponds exactly to a result type in the result type set. The detection task can be completed with a completion rate of 1. A sign function is set, which is set to 1 only when the completion rate is 1, and 0 in all other cases. The inverse of the simulated detection time is normalized by the Sigmoid function to obtain the efficiency score. The sign function and the efficiency score are weighted and summed according to a preset scaling factor to obtain the agent's simulated detection score. The simulated detection score is used to comprehensively evaluate the agent's completion result and efficiency of the detection task. The closer the simulated detection score is to 1, the better the adaptation effect between the agent and the detection task.

[0062] The value of the agent is updated according to the value update formula learned by the Q network. The value of the agent is equal to the original value plus the product of the learning rate and the simulated test score, plus the product of the learning rate, the discount factor and the maximum value in the value table, and then the product of the learning rate and the original value is subtracted. The maximum value is introduced as the future expectation to improve the robustness of decision-making.

[0063] A random number between 0 and 1 is generated based on a greedy strategy. When the random number is less than or equal to the exploration probability, an unexplored agent is randomly selected from the agent set for the next simulation test. When the random number is greater than the exploration probability, the agent with the current maximum value is selected from the value table for the next simulation test. After each selection, the exploration probability is reduced by the exploration step size. The exploration probability is required to be greater than the exploration step size multiplied by the total number of iterations to prevent the exploration probability from dropping to 0. Based on the exploration probability, in the early stage of DRL decision-making, the focus is on selecting unexplored agents to prevent local convergence of DRL decision-making. As the number of iterations increases, the focus gradually shifts to selecting agents that have been explored and have the best detection results to prevent non-convergence of DRL decision-making.

[0064] Repeat the iteration until the total number of iterations is reached, then select the agent with the highest value in the value table as the optimal agent and determine it as the optimal program execution group.

[0065] Furthermore, the parameter determination unit pre-collects multiple input information sets and detection result sets, processes the input information sets to obtain corresponding task feature packages and optimal program execution groups. For each task feature package, based on the optimal program execution group, it collaboratively retrieves all associated historical optimal parameter vectors, historical task feature packages, and detection scores in the knowledge graph. The historical optimal parameter vector and historical task feature package with the highest detection score are respectively used as the benchmark optimal parameter vector and benchmark task feature package corresponding to the optimal program execution group. Using the optimal program execution group as an index, and combining the benchmark optimal parameter vector, benchmark task feature package, and multiple corresponding task feature packages and detection result sets corresponding to each optimal program execution group, an index training dataset is constructed, and a domain adversarial network is trained based on the adversarial win-win concept.

[0066] Furthermore, the generator flattens the task feature package and the baseline task feature package with the same index into task feature vector and baseline task feature vector, respectively. The difference between the task feature vector and the baseline task feature vector is used to obtain the task deviation vector. The task deviation vector is adjusted to the same dimension as the baseline optimal parameter vector using a fully connected layer and mapped by the ReLU function to obtain the parameter correction vector. The baseline parameter vector and the parameter correction vector of the same dimension are weighted and summed again by a fully connected layer and mapped by the LeakyReLU function to obtain the optimal parameter vector. This achieves adaptive adjustment of the optimal parameter vector for different inspected objects and different detection environments under the premise of the same optimal program execution group.

[0067] Furthermore, the discriminator includes a classification discriminator and a simulation detector. The classification discriminator uses a fully connected layer to reduce the dimensionality of the optimal parameter vector to a scalar and normalizes it to a true probability using the Sigmoid function. The true probability reflects the probability that the discriminator determines the optimal parameter vector to be a true parameter. The higher the true probability, the more difficult it is for the discriminator to distinguish the optimal parameter vector generated by the generator. The simulation detector reuses the simulation detection in the program decision unit, constructs a virtual object to be inspected and a virtual detection environment based on the task feature package, replaces the default parameter vector with the optimal parameter vector to perform simulation detection, and obtains a set of simulation detection results.

[0068] like Figure 3 As shown, further, when the task feature packets in the indexed training dataset are input into the domain adversarial network for training, the cross-entropy loss of the classifier and the mean squared error loss between the simulated detection result set and the actual detection result set output by the simulated detector are calculated. The generator uses the gradients of the cross-entropy loss and the mean squared error loss with respect to the generator parameters to update the generator parameters, thereby improving the authenticity and reliability of the optimal parameter vector generated by the generator. The classifier uses the gradients of the cross-entropy loss with respect to the classifier parameters to update the classifier parameters, thereby improving the accuracy of the classifier in judging the authenticity of the optimal parameter vector, and inverts the gradients. The feedback generator reduces the realism of the optimal parameter vector generated by the generator, achieving mutual improvement through adversarial training with the generator. The simulated detector uses the gradient of the mean squared error loss with respect to the detector parameters to update the detector parameters, improving the detector's simulated detection accuracy and positively informing the generator, achieving a win-win improvement with the generator. Through multiple rounds of training and adjustment until the cross-entropy loss approaches 0.693 and the mean squared error loss is minimized, the domain adversarial network training is complete. At this point, the true probability output by the discriminator based on the optimal parameter vector converges to 0.5. In binary classification tasks, if the cross-entropy loss is 0.5, then... The value is approximately 0.693, indicating that the discriminator cannot distinguish whether the optimal parameter vector is real or not, and can only make a random judgment. The optimal parameter vector generated by the generator is indistinguishable from the real situation and fits the actual detection result set. The generator parameters are fixed and called in the online stage to generate the optimal parameter vector.

[0069] Example 2

[0070] like Figure 4 As shown, the aviation inspection mission identification and coordination method, used to implement the aviation inspection mission identification and coordination system disclosed in Embodiment 1, includes the following steps:

[0071] The input information set is acquired and its format is converted and redundancy is removed to obtain a standard dataset, which includes standard requirement data, standard object data, and standard environment data.

[0072] The semantic, physical, and environmental features of each standard data in the standard dataset are extracted and modulated into a unified feature matrix. The Transformer model is used to capture the contextual dependencies in the unified feature matrix to identify the task type and determine the result type set. The task type, result type set, standard object dataset, and standard environment dataset are combined to construct the task feature package.

[0073] Based on DRL decision-making, the program numbers and default parameters of all relevant detection programs matching the task type are called according to the knowledge graph. Combined with the standard object dataset and standard environment dataset, the simulation detection scenario is defined. Each permutation and combination of program numbers is regarded as an agent. The agent is randomly selected and simulated detection is performed based on the corresponding default parameter vector. The simulation detection score is calculated and the value table is updated through Q network learning. The greedy strategy is used to iteratively select agents in combination with the value table until the total number of iterations is reached to obtain the optimal agent, i.e. the optimal program execution group.

[0074] The benchmark optimal parameter vector and benchmark task feature package that meet the same optimal program execution group and have the highest detection score are retrieved from the knowledge graph. Based on the generator of the domain adversarial network determined by offline training, the benchmark optimal parameter vector is corrected by the difference between the flattening results of the task feature package and the benchmark task feature package. The optimal parameter vector is then obtained and integrated with the task feature package and the optimal program execution group to form a detection execution scheme, which is then delivered to the detection terminal.

[0075] Based on the optimal program execution group, a program node table is constructed and the nodes corresponding to the relevant detection programs that have been executed are marked until all nodes are marked. The detection result set and detection time are obtained and the actual detection score is calculated. The task feature package, optimal program execution group, optimal parameter vector, detection result set, and detection score are filled into the knowledge graph.

[0076] This invention discloses an aviation inspection task identification and coordination system and method. The system performs format conversion on the input information set, uses the SimHash algorithm to calculate semantic similarity to achieve redundancy removal and obtain a standard dataset. The standard requirement data is transformed into semantic feature vectors using the BERT model. Standard object data and standard environment data are respectively converted into word vectors and their corresponding parameter normalization values ​​are concatenated to obtain physical feature vectors and environmental feature vectors. After linear modulation to unify the dimensions of each feature matrix, they are concatenated into a unified feature matrix. This unified feature matrix is ​​input into a Transformer model to determine the task type and construct a task feature package. The system executes a DRL decision to call the detection program number matching the task type and its default parameters, defining... The system simulates detection scenarios and arranges program numbers into intelligent agents. It iteratively determines the optimal program execution group. Using a domain adversarial network, it corrects the baseline optimal parameter vector by utilizing the difference between the task feature package and the baseline task feature package to obtain the optimal parameter vector. It integrates the task feature package, the optimal program execution group, and the optimal parameter vector to generate a detection execution plan. Based on the detection execution plan, it constructs a program node table to monitor the execution of the detection program in real time and marks the corresponding nodes. After all nodes are marked, it obtains the detection result set and detection duration and calculates the detection score. It then fills the knowledge graph with the task feature package, the optimal program execution group, the optimal parameter vector, the detection result set, and the detection score to achieve updates, thereby improving the accuracy, efficiency, and decision robustness of general aviation detection.

[0077] The above description is merely a preferred embodiment of the present invention, and the scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for users of ordinary technical skills, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. An aerial detection task identification coordination system, characterized in that, The task recognition module, the scheme formulation module and the monitoring and evaluation module are included. The task recognition module obtains an input information set and obtains a unified feature matrix through format conversion, redundancy removal and feature extraction, identifies a task type based on a Transformer model and constructs a task feature package; The scheme formulation module obtains the task feature package, calls the program number and default parameters of the related detection program according to the knowledge graph, defines a simulation detection scene, and arranges and combines the related detection program as an intelligent agent, randomly selects the intelligent agent and performs simulation detection according to the corresponding default parameter vector, calculates the simulation detection score and updates the value table through Q network learning, reselects the intelligent agent by using the greedy strategy, iteratively determines the optimal program execution group corresponding to the optimal intelligent agent, and determines the benchmark optimal parameter vector and the benchmark task feature package according to the knowledge graph, modifies the benchmark optimal parameter vector based on the difference between the flattened results of the task feature package and the benchmark task feature package by using the domain adversarial network, obtains the optimal parameter vector, integrates and constructs a detection execution scheme, and delivers the detection terminal; The monitoring and evaluation module monitors the program execution progress and calculates the detection score after the detection is completed, and fills the task feature package, the optimal program execution group, the optimal parameter vector, the detection result set and the detection score into the knowledge graph; The scheme formulation module includes a program decision unit, a parameter determination unit and an integration and delivery unit. The program decision unit executes DRL decision based on the task feature package, calls the program number and default parameters of the related detection program, defines a simulation detection scene, arranges and combines the program number as an intelligent agent, calculates the detection score after simulation detection, and updates the value table by using Q network learning, and iteratively determines the optimal program execution group corresponding to the optimal intelligent agent; The parameter determination unit retrieves the benchmark optimal parameter vector and the benchmark task feature package in the knowledge graph while meeting the optimal program execution group, modifies the optimal parameter vector by using the generator of the domain adversarial network; The integration and delivery unit integrates the task feature package, the optimal program execution group and the optimal parameter vector to obtain a detection execution scheme, and sends the detection execution scheme to the detection terminal and the monitoring and evaluation module.

2. The aerial detection tasking coordination system of claim 1, wherein, The task recognition module includes an access processing unit, a feature generation unit and a type identification unit. The access processing unit obtains an input information set and converts the format to obtain an input data set, calculates the semantic similarity of any two input data by using the SimHash algorithm, and removes the redundancy based on the similarity to obtain a standard data set; The feature generation unit converts the standard demand data into a semantic feature vector by using the BERT model and splices the semantic feature matrix, converts the type of the standard object or environmental data into a type code by using a word vector, splices the corresponding normalized parameter value to obtain a physical feature matrix and an environmental feature matrix, and modulates and splices the semantic feature matrix, the physical feature matrix and the environmental feature matrix to obtain a unified feature matrix; The type identification unit inputs the unified feature matrix into the Transformer model to determine the task type and the result type set, combines the task type, the result type set, the standard object data set and the standard environmental data set in the standard data set to obtain the task feature package.

3. The aerial detection tasking coordination system of claim 2, wherein, The access processing unit adopts OCR recognition and regular expression to match and extract different requirement types of requirement description from the detected requirement text, integrates to obtain requirement input data, reads object parameters in the detected object table through Pandas and maps to obtain object input data through a field mapping table, uses SparkStreaming stream processing to confirm parameter values in the environment data and system time, and integrates environment input data.

4. The aerial detection tasking coordination system of claim 1, wherein, The DRL decision includes: Confirming the task type and calling the program number and default parameters of each related detection program to build a simulated detection scene; Randomly arranging and combining the program numbers of all related detection programs to obtain all detection program execution groups and define them as agents, respectively, and setting each agent according to the corresponding default parameter vector; Randomly selecting an agent to perform simulated detection to obtain a simulated detection result set and a simulated detection time, dividing the matching result type number of the simulated detection result set and the result type set in the task feature package by the total number of result types in the result type set to obtain a result completion rate, setting the value of the sign function only when the result completion rate is equal to 1 to be 1, normalizing the opposite number of the simulated detection time through a Sigmoid function to obtain an efficiency score, and weighting and summing the sign function and the efficiency score to obtain a simulated detection score; Updating the value of the agent according to the value update formula of the Q network learning, which is equal to the original value plus the product of the learning rate and the simulated detection score, plus the product of the learning rate, the discount factor and the maximum value in the value table, minus the product of the learning rate and the original value; Generating a random number between 0 and 1 according to the greedy strategy and comparing it with the exploration probability to select an agent for the next simulated detection, if the random number is less than or equal to the exploration probability, selecting an agent that has not been selected, when the random number is greater than the exploration probability, selecting the agent with the maximum value from the value table, and each time after selection, reducing the exploration probability by the exploration step; Repeat the iteration until the total number of iterations is reached, select the agent with the maximum value in the value table as the optimal agent and determine the optimal program execution group.

5. The aerial detection tasking coordination system of claim 1, wherein, The domain adversarial network includes a generator, which flattens the task feature package and the benchmark task feature package with the same index into a task feature vector and a benchmark task feature vector respectively, and obtains a task deviation vector by difference, maps the task deviation vector into a parameter correction vector through a fully connected layer, and obtains an optimal parameter vector by fusing the benchmark parameter vector and the parameter correction vector through a fully connected layer.

6. The aerial detection tasking coordination system of claim 1, wherein, The domain adversarial network includes a discriminator, which includes a classification discriminator and a simulated detector, the classification discriminator maps the optimal parameter vector into a real probability through a fully connected layer, and the simulated detector performs simulated detection based on the task feature package, the optimal program execution group and the optimal parameter vector to obtain a simulated detection result set.

7. The aerial detection tasking coordination system of claim 1, wherein, When training the domain adversarial network, the cross-entropy loss of the classification discriminator is calculated, and the mean square error loss of the simulated detection result set output by the simulation detector and the detection result set is calculated, the generator updates the generator parameters using the cross-entropy loss and the mean square error loss and the gradient of the generator parameters, the classification discriminator updates the classifier parameters using the cross-entropy loss and the gradient of the classifier parameters, and the generator is fed back through gradient reversal, the simulation detector updates the detector parameters using the mean square error loss and the gradient of the detector parameters, and the generator is fed forward, when the approximation is 0.693 and the mean square error loss is the minimum, the domain adversarial network training is completed.

8. The aerial detection tasking coordination system of claim 1, wherein, The monitoring evaluation module builds a program node table based on the optimal program execution group, monitors the detection program execution in real time through Socket communication, and marks the corresponding node whenever a detection program is completed; after all nodes are marked, the detection is completed.

9. An aerial detection task identification coordination method for implementing the aerial detection task identification coordination system according to any one of claims 1-8, characterized in that, The method comprises the following steps: An input information set is obtained, a standard data set is obtained through format conversion and redundancy deduplication, and a unified feature matrix is obtained through feature extraction and integration, the unified feature matrix is input into a Transformer model to determine a task type and a result type set, and a task feature package is constructed by combining a standard object data set and a standard environment data set in the standard data set with the program number and the default parameters of the related detection program matched with the task type according to the knowledge graph; Each permutation and combination of the program number is regarded as an agent, an agent is randomly selected and simulated detection is performed based on the corresponding default parameter vector, a simulated detection score is calculated, a value table is updated through Q network learning, an agent is iteratively selected by combining the value table with the greedy strategy, and the optimal agent corresponding to the optimal program execution group is obtained when the total number of iterations is reached; A benchmark optimal parameter vector and a benchmark task feature package that meet the same optimal program execution group and have the highest detection score are retrieved in the knowledge graph, the difference between the flattened results of the task feature package and the benchmark task feature package is used to correct the benchmark optimal parameter vector based on the generator of the domain adversarial network, an optimal parameter vector is obtained, and the detection execution scheme is integrated by combining the optimal parameter vector with the task feature package, the optimal program execution group, and delivered to the detection terminal; The program execution progress is monitored, and a detection score is calculated after the detection is completed, and the task feature package, the optimal program execution group, the optimal parameter vector, the detection result set and the detection score are filled into the knowledge graph.

Images