Aviation maintenance skill simulation training system

Abstract

The invention discloses an aviation maintenance skill simulation training system. The system comprises an intelligent task simulation platform which carries out high-simulation-degree hardware operation, multi-mode fault injection and operation data acquisition; a three-dimensional virtual training platform: constructing a digital twin aircraft model; the dynamic evaluation engine adopts a neural symbol two-channel evaluation technology, a symbol layer verifies a safety operation specification through an IF-THEN rule engine, a neural layer analyzes an operation intention through a Transform encoder and calculates an innovation score in combination with a knowledge graph, and cross-modal cognitive analysis is fused; the self-adaptive training system is used for dynamically updating the weight of the knowledge graph through a meta-learning optimizer, and controlling task complexity parameters and neural architecture search to automatically optimize the structure of the knowledge graph by utilizing a Thompson sampling algorithm; the block chain is stored in a data table; and after desensitization, the data of the whole system is stored through a block chain and is stored through a PBFT consensus mechanism of a data-in-platform chain.

Description

Technical Field

[0001] This invention relates to the field of aircraft maintenance technology, and in particular to an aircraft maintenance skills simulation training system. Background Technology

[0002] Against the backdrop of the rapid introduction of new aircraft models and the accelerated iteration of new technologies, maintenance personnel in the engineering department face real challenges such as outdated knowledge structures and a disconnect between training programs and actual work. Traditional training models centered on time constraints are no longer adequate for the high standards required for complex system maintenance and the application of new technologies. The Competency-Based Training and Assessment (CBTA) model, promoted in the international civil aviation field, focuses on behavioral competency standards and achieves precise alignment between training and job requirements through dynamic assessment of trainees' knowledge, skills, and attitudes (KSA), providing an innovative methodology for addressing these issues. This study aims to construct a CBTA training system that meets the needs of the domestic aviation maintenance industry. Specifically, it encompasses the construction of a competency model combining the ICAO Document 10098 framework with domestic maintenance job task analysis, and the design of a dynamic assessment mechanism based on observable behavior (OB) assessment tools and a multi-dimensional assessment system. By clarifying behavioral competency standards at each level—basic training, type training, and retraining—and establishing a verification system combining formative and summative assessments, the study promotes a shift in training models from time-driven to competency-driven. Its significance lies not only in significantly improving the efficiency and relevance of aircraft maintenance personnel training, but also in forming a maintenance personnel training management system based on competency training and assessment that is suitable for the actual situation of the unit. This provides a solid human resource guarantee for aviation safety, while also providing valuable exploration and practical experience for the reform and innovation of civil aviation maintenance training in my country, ultimately achieving a comprehensive improvement in training quality, maintenance capabilities, and safety assurance levels.

[0003] In existing technologies, the IF-THEN rules of dynamic evaluation engines can only handle preset logic and cannot identify students' innovative solutions, such as temporarily using alternative tools to complete the operation. This results in insufficient depth of data analysis methods and weak dynamic adaptability. Therefore, an aircraft maintenance skills simulation training system is proposed. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing an aircraft maintenance skills simulation training system.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: An aircraft maintenance skills simulation training system includes: Intelligent mission simulator: It integrates replaceable avionics / hydraulics / engine components through modular hardware to build a physical simulation environment. Combined with multi-agent fault models and IoT sensor networks, it performs highly realistic hardware operation, multi-modal fault injection and operation data acquisition. The acquired physical operation data is synchronized to the three-dimensional virtual training platform in real time. 3D Virtual Training Platform: Based on Unity / Unreal Engine and integrating neural radiation field technology, a 1:1 digital twin aircraft model is constructed. Through AR collaborative spatial computing and dynamic occlusion management, it realizes the visualization of complex system principles, VR headset + gesture recognition immersive operation and multi-person collaborative training functions. The 3D virtual training platform records the students' operation trajectory and virtual environment data and inputs them into the dynamic evaluation engine. Dynamic evaluation engine: It adopts a dual-channel evaluation technology of neural symbols. The symbol layer verifies the safety operation specifications through the IF-THEN rule engine, while the neural layer parses the operation intention through the Transformer encoder and calculates the innovation score by combining the knowledge graph. It integrates cross-modal cognitive analysis to perform real-time capability evaluation, generate capability deficiency heat maps and output CBTA three-dimensional evaluation reports. The evaluation results drive the adaptive training system. Adaptive training system: Dynamically updates knowledge graph weights through meta-learning optimizer, and automatically optimizes knowledge graph structure by using Thompson sampling algorithm to control task complexity parameters and neural architecture search, completes personalized training task push and "evaluation-diagnosis-training" closed-loop evolution, and the generated personalized task parameters are synchronized to intelligent task simulation table and 3D virtual training platform to adjust hardware configuration and virtual scene difficulty. Blockchain-based evidence storage and data platform: After anonymization, all system data is stored on the blockchain through the PBFT consensus mechanism of the blockchain evidence storage and data platform. At the same time, QAR real flight data is integrated to optimize typical fault scenarios in the digital twin fault database.

[0006] The above technical solution further includes: Furthermore, the key system is broken down into independent modules, each with an aerospace-grade interface. The module type is identified by RFID tags, and the corresponding simulation model is automatically loaded. Temperature, humidity, and vibration sensors are deployed to collect environmental data. Paris's law is used to simulate the propagation of metal fatigue cracks. The environmental data and component aging results are input into the rule engine to generate a composite scenario. The PTPv2 clock synchronization protocol is sampled to control the timestamp deviation of environmental sensors, component algorithms, and IoT sensor data. The coupled fault parameters are then input into the Unity / Unreal engine to dynamically adjust the state of the digital twin model.

[0007] Furthermore, the specific steps for constructing a 1:1 digital twin aircraft model using the aforementioned 3D virtual training platform are as follows: Data acquisition: Use laser scanners and SLR camera arrays to acquire aircraft entity data and generate dense point clouds; NeRF model training: Point cloud data is input into a neural radiation field network, and continuous view images are synthesized using a volume rendering equation. Where T(t) is the transmittance, Let c be the density field and c be the color field; Engine Integration: Export the trained NeRF model to a Unity / Unreal-recognizable format and render it in real time using a custom shader.

[0008] Furthermore, the specific steps of the three-dimensional virtual training platform to realize the visualization of complex system principles, immersive operation of VR headset + gesture recognition, and multi-person collaborative training functions through AR collaborative spatial computing and dynamic occlusion management are as follows: Distributed shared coordinates: Each student's AR device generates a local coordinate system using the SLAM algorithm, and then aligns it to the global coordinate system using the ICP algorithm. ,in, For local coordinate points, Let R be the global coordinate point, R be the rotation matrix, and t be the translation vector; Intent Prediction and Conflict Arbitration: A gesture intent prediction model is built based on federated learning (FedAvg algorithm) to predict operation paths in advance. Where w are model parameters. For learning rate, Let i be the loss function for the i-th device; Eye tracking and component importance assessment: The eye tracking module of the AR device records the student's gaze focus, and the rendering priority is calculated in combination with the functional importance of the components; Real-time rendering priority adjustment: Dynamically allocate rendering resources based on the calculated rendering priority, and prioritize the rendering of high-priority components.

[0009] Furthermore, the specific steps of the dynamic evaluation engine in performing dynamic evaluation are as follows: Data input and preprocessing: AR devices are used to record gesture paths and gaze focus, and component status and fault phenomena are recorded in real time. HRV data is recorded using a heart rate monitor, and time-domain and frequency-domain indicators are calculated. Symbol layer security verification: Based on CAAC regulations and maintenance manuals, define safe operation rules, input operation trajectory and virtual environment data into the IF-THEN rule engine, and verify the rule triggering conditions through forward reasoning; Neural layer intent parsing and innovation scoring: Input gesture path and virtual environment data into Transformer encoder, generate context embedding vector through self-attention mechanism, retrieve similar maintenance cases in knowledge graph, and calculate cosine similarity between intent embedding and case embedding; Cross-modal cognitive analysis: eye-tracking trajectories generate decision focus heatmaps, voice commands are analyzed using NLP sentiment analysis to output command ambiguity, and physiological signals are analyzed using an HRV stress model to generate stress resilience scores; Competency Shortcomings Heatmap and CBTA Report Generation: Integrating symbolic layer validation results, neural layer innovation scores, and cross-modal analysis data, a competency shortcomings heatmap is generated through weighted fusion, presenting the three-dimensional assessment results of knowledge, skills, and attitudes in a structured manner.

[0010] Furthermore, the adaptive training system dynamically updates the knowledge graph weights through a meta-learning optimizer, including the following steps: Data input and preprocessing: Historical training data collection technology is used to record trainees' past task parameters and results, and skill dimension data is extracted from CBTA reports by combining dynamic assessment results integration technology; Skills transfer trend prediction: Using LSTM model construction technology, through recursive calculations of input gate, forget gate, output gate and cell state, the historical skill data of trainees is analyzed and the future improvement trend of mechanical and avionics skills is predicted. Knowledge graph weight update: The weight of each node in the knowledge graph is dynamically adjusted by using weight update formula definition technology and node importance weighting technology. Effect verification and iteration: A / B testing technology is used to compare the recommendation effect of the knowledge graph before and after the update, and the parameters are adjusted by combining feedback iteration technology.

[0011] Furthermore, the adaptive training system utilizes the Thompson sampling algorithm to control the specific steps of the task complexity parameters; Parameter initialization: Define the task parameter range: number of steps Concealment Initialize the Beta distribution parameters: α=1, β=1; Thompson sampling execution: Step count sampling: Samples the step count parameter from the Beta distribution; Stealth sampling: Samples the stealth parameter from the Beta distribution; Task parameter generation and push: Generate task parameters based on sampling results: number of steps and concealment level; Parameter feedback and Beta distribution update: Update the Beta distribution parameters based on the students' task completion status.

[0012] Furthermore, the adaptive training system utilizes neural architecture search to automatically optimize the knowledge graph structure through the following specific steps: Search space definition: Define possible structural variation operations for knowledge graphs: Add an edge: connect two nodes with low correlation. Delete edge: Remove low-weight connections; Adjust node weights: Increase the weight of key nodes; Reinforcement learning model construction: Define the state space S: the current knowledge graph structure and student ability data; Define action space A: structural mutation operation; Define the reward function R: the student's skill improvement rate after task recommendation; The loss function of the reinforcement learning model is expressed as: ,in, As a discount factor, For the current network parameters, For target network parameters; Structural mutation and performance evaluation: Randomly select a structural mutation operation from the search space, use a reinforcement learning model to evaluate the performance of the mutated knowledge graph, and calculate the reward R; Structure update and deployment: Update the reinforcement learning network parameters θ according to the reward R, and retain the structure mutation operations with a boost rate greater than 0.

[0013] Furthermore, the specific steps for data anonymization and on-chain storage in the blockchain-based evidence storage and data platform are as follows: Sensitive field identification: Define sensitive data fields; Application of desensitization algorithms: Adding noise using the Laplace mechanism. ,in, For sensitivity, Budget for privacy; Data masking and permutation: Hash encryption of ID-type fields; Node registration and permission management: Defining a blockchain node set Each node holds a public-private key pair. ; Three-phase consensus process: Preparatory phase: The master node broadcasts the message ⟨PRE-PREPARE,v,s,d,m>, where v is the view number, s is the sequence number, d is the digest, and m is the plaintext; Preparation phase: After node verification, broadcast ⟨PREPARE,v,s,d,i>, where i is the node ID; Commit phase: After a node has collected at least 2f+1 preparation messages, it broadcasts ⟨COMMIT,v,s,d,i> and performs a state update; Block generation and chained storage: After consensus is reached, block B={Header,Transactions} is generated, where the Header contains the previous block hash, timestamp, and Merkle root.

[0014] The present invention has the following beneficial effects: This invention employs a dual-channel neural-symbolic evaluation technique. The symbolic layer verifies safe operating procedures through the IF-THEN rule engine to ensure operational safety, while the neural layer analyzes the operational intent through the Transformer encoder and calculates innovation scores using a knowledge graph. The symbolic system ensures a safety baseline, while the neural network evaluates innovation, balancing standardization and flexibility. This approach avoids violations while encouraging student innovation, dynamically expanding the knowledge base, and enhancing the comprehensiveness and adaptability of the evaluation. Attached Figure Description

[0015] Figure 1 This is a system block diagram of an aircraft maintenance skills simulation training system proposed in this invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Please see Figure 1 As shown, the present invention is an aircraft maintenance skills simulation training system, comprising: Intelligent Mission Simulation Platform: A physical simulation environment is constructed by integrating replaceable avionics / hydraulic / engine components through a modular hardware platform. It combines a multi-agent fault model (integrating environmental sensors and metal fatigue algorithms to generate complex fault scenarios such as "high altitude and low temperature + hydraulic oil emulsification + sensor offset") and an IoT sensor network (deploying torque sensors, thermal imaging cameras, and millimeter-wave radar) to perform highly realistic hardware operations, multi-modal fault injection, and 0.1mm-level precision operation data acquisition. The acquired physical operation data (torque values, step sequence) is synchronized to the three-dimensional virtual training platform in real time. 3D Virtual Training Platform: Based on the Unity / Unreal engine and integrating Neural Radiation Field (NeRF) technology, a 1:1 digital twin aircraft model is constructed. Through AR collaborative spatial computing (Distributed Shared Coordinate System DSCS supports multi-person viewpoint synchronization deviation <1°) and dynamic occlusion management (Real-time Rendering Priority Algorithm RPA ensures 100% visibility of key component labels), it realizes the visualization of complex system principles, VR headset + gesture recognition immersive operation, and multi-person collaborative training functions. The 3D virtual training platform records the student's operation trajectory (gesture path, gaze focus) and virtual environment data (component status, fault phenomena) and inputs them into the dynamic evaluation engine. Dynamic evaluation engine: It adopts a dual-channel evaluation technology of neural symbols. The symbolic layer verifies the safety operation specifications through the IF-THEN rule engine, while the neural layer parses the operation intention through the Transformer encoder and calculates the innovation score by combining the knowledge graph. It integrates cross-modal cognitive analysis to perform real-time capability evaluation, generate capability deficiency heat maps (such as "line measurement timeout rate 40%), and output CBTA three-dimensional evaluation reports (knowledge / skills / attitude dimensions). The evaluation results drive the adaptive training system. The adaptive training system dynamically updates the knowledge graph weights through a meta-learning optimizer (combined with LSTM capability transfer trend prediction for mechanical → avionics skills), and uses the Thompson sampling algorithm to control task complexity parameters (such as number of steps and concealment) and neural architecture search (NAS) to automatically optimize the knowledge graph structure. This completes personalized training task push (basic disassembly and assembly → advanced troubleshooting) and "evaluation-diagnosis-training" closed-loop evolution. The generated personalized task parameters are synchronized to the intelligent task simulator and 3D virtual training platform to adjust hardware configuration and virtual scene difficulty. Blockchain-based evidence storage and data platform: All system data (operation records, evaluation results, training logs) are anonymized and then stored on the blockchain through the lightweight PBFT consensus mechanism of the blockchain evidence storage and data platform. At the same time, QAR real flight data is integrated to optimize typical fault scenarios of the digital twin fault library.

[0018] In one embodiment, key systems such as avionics, hydraulics, and engines are disassembled into independent modules (such as "hydraulic pump module" and "avionics integrated display module"). Each module uses an aviation-grade interface (ARINC 404 / 600 standard) to achieve hot-swapping. The module type is identified by RFID tags, and the corresponding simulation model is automatically loaded (such as when replacing the engine module, the system automatically retrieves the CFM56-7B engine parameters). Temperature, humidity, and vibration sensors are deployed to collect environmental data. Paris's law is used to simulate the propagation of metal fatigue cracks. The environmental data and component aging results are input into the rule engine (IF-THEN) to generate a composite scenario. Paris's Law is as follows: Where a is the crack length, N is the number of cycles, and ΔK is the stress intensity factor range; The PTPv2 clock synchronization protocol is sampled to control the timestamp deviation of environmental sensors, component algorithms, and IoT sensor data. The coupled fault parameters (such as hydraulic oil viscosity and pipeline crack length) are input into the Unity / Unreal engine to dynamically adjust the state of the digital twin model.

[0019] In one embodiment, the specific steps for constructing a 1:1 digital twin aircraft model using the 3D virtual training platform are as follows: Data acquisition: Aircraft entity data was acquired using a 360° laser scanner (0.1mm accuracy) and an array of 12 SLR cameras (8K resolution) to generate dense point clouds (density > 10 points / mm²). NeRF model training: Point cloud data is input into a neural radiation field network, and continuous view images are synthesized using a volume rendering equation. Where T(t) is the transmittance, Let c be the density field and c be the color field; Engine Integration: Export the trained NeRF model to a Unity / Unreal-recognizable format and render it in real time using a custom shader.

[0020] In one embodiment, the specific steps of the three-dimensional virtual training platform to realize the visualization of complex system principles, immersive operation of VR headset + gesture recognition and multi-person collaborative training functions through AR collaborative spatial computing and dynamic occlusion management are as follows: Distributed shared coordinates: Each student's AR device generates a local coordinate system using the SLAM algorithm, and then aligns it to the global coordinate system using the ICP algorithm. ,in, For local coordinate points, Let R be the global coordinate point, R be the rotation matrix, and t be the translation vector; Intent Prediction and Conflict Arbitration: A gesture intent prediction model is built based on federated learning (FedAvg algorithm) to predict operation paths in advance. Where w are model parameters. For learning rate, Let i be the loss function for the i-th device; Eye tracking and component importance assessment: The eye tracking module of the AR device (sampling rate 120 Hz) records the student's gaze focus, and the rendering priority is calculated by combining the functional importance of the components (e.g., the weight of the core engine component > 0.8). Real-time Render Priority Adjustment (RPA): Dynamically allocates rendering resources based on the calculated rendering priority, prioritizing the rendering of high-priority components.

[0021] In one embodiment, the specific steps of the dynamic evaluation engine in performing dynamic evaluation are as follows: Data input and preprocessing: The AR device records the gesture path (coordinate sequence of 21 key points) and gaze focus (screen coordinate sequence) (operation trajectory), records the component status (such as hydraulic pump pressure, valve opening) and fault phenomena (such as pipeline leakage, sensor offset) in real time (virtual environment data), records HRV data (RR interval sequence) through a heart rate monitor, and calculates time domain indicators (SDNN) and frequency domain indicators (LF / HF ratio) (physiological signals). Symbol layer security verification: Based on CAAC regulations and maintenance manuals, define safe operation rules, input operation trajectory and virtual environment data into the IF-THEN rule engine, and verify the rule triggering conditions through forward chaining; Neural layer intent parsing and innovation scoring: Input gesture path and virtual environment data into Transformer encoder, generate context embedding vector through self-attention mechanism, retrieve similar maintenance cases in knowledge graph, and calculate cosine similarity between intent embedding and case embedding; Cross-modal cognitive analysis: eye-tracking trajectories generate decision focus heatmaps; voice commands are analyzed using NLP sentiment analysis to output command ambiguity; physiological signals are analyzed using an HRV stress model to generate stress resilience scores. Specifically, a gaze persistence heatmap is calculated using a spatiotemporal attention model (such as Gaussian-ST). ,in, As time weight, With a Gaussian kernel width, sentiment analysis and intent recognition are performed using a BERT model, instruction ambiguity is calculated, and resilience is evaluated by combining SDNN and LF / HF ratio. Competency Shortcomings Heatmap and CBTA Report Generation: Integrating symbolic layer validation results, neural layer innovation scores, and cross-modal analysis data, a competency shortcomings heatmap is generated through weighted fusion, presenting the three-dimensional assessment results of knowledge, skills, and attitudes in a structured manner.

[0022] Knowledge dimension: accuracy rate of theoretical question bank, AR response time.

[0023] Skills dimension: operational precision and procedural compliance.

[0024] Attitude dimension: communication efficiency, stress resistance.

[0025] In one embodiment, the adaptive training system dynamically updates the knowledge graph weights through a meta-learning optimizer, including the following steps: Data input and preprocessing: Historical training data collection technology is used to record trainees' past task parameters (such as number of steps, concealment) and results (success rate), and dynamic evaluation result integration technology is used to extract skill dimension data such as operational accuracy and step compliance from CBTA reports; Skills transfer trend prediction: Using LSTM model construction technology, through recursive calculations of input gate, forget gate, output gate and cell state, the historical skill data of trainees is analyzed and the future improvement trend of mechanical and avionics skills is predicted (e.g., avionics skills improve by 20%). The Forgotten Gate: Deciding which historical information to discard. ; in, For the Sigmoid function, This is the weight matrix. This is the hidden state from the previous moment. For the current input, For the bias term of the forget gate; Input gate: Determine which new information to update: ; ; in, The output of the input gate, Here is the weight matrix of the input gate. For the bias term of the input gate, The candidate cell states are generated using the tanh function, representing new information, and their range is between -1 and 1. To generate the weight matrix for candidate cell states, Bias terms for generating candidate cell states; Cell state update: fusing historical and new information: ; in, This represents the current cell state, and ⊙ represents element-wise multiplication. This represents the cell state at the previous moment; Output gate: Generates the current hidden state: ; ; in, For the output of the output gate, Here is the weight matrix of the output gate. This is the bias term for the output gate. The current hidden state; Knowledge graph weight update: Utilizing a weight update formula definition technique (combining the difference between learning rate and prediction - current improvement). , The amount by which the prediction skill of node n is improved. (This refers to the amount of skill improvement) and node importance weighting techniques (which assign higher weights to key nodes through a soft attention mechanism). The weights of each node in the knowledge graph (such as "hydraulic valve disassembly" and "avionics fault diagnosis") are dynamically adjusted to ensure that the recommended tasks match the trainees' ability development trends. Performance Validation and Iteration: A / B testing is used to compare the recommendation performance of the knowledge graph before and after the update, and feedback iteration is combined to adjust parameters such as the learning rate and continuously optimize model performance.

[0026] In one embodiment, the adaptive training system utilizes the Thompson sampling algorithm to control the specific steps of the task complexity parameters. Parameter initialization: Define the task parameter range: number of steps Concealment Initialize the Beta distribution parameters: α=1, β=1 (corresponding to a uniform distribution, indicating that there is no initial preference); Thompson Sampling Execution: Step-Number Sampling: Sampling the step-number parameter from the Beta distribution Concealment sampling: Sampling concealment parameters from the Beta distribution. ; Task parameter generation and push: Generate task parameters based on sampling results: number of steps and concealment level; Parameter feedback and Beta distribution update: Update the Beta distribution parameters based on the students' task completion status.

[0027] Success: α+=1 (e.g., if success occurs when the number of steps is 6, then αs+=1). Failure: β+=1 (e.g., if the concealment level is 0.7, then βh+=1). In this embodiment: Initial state: Trainee B has completed the basic disassembly and assembly task. Historical data shows that when the number of steps is 5, there are 4 successes and 1 failure. When the concealment level is 0.5, there are 3 successes and 2 failures.

[0028] Thompson sampling: Number of steps for sampling: s-sampling ∼ Beta(4+1,1+1)=Beta(5,2)⇒s-sampling≈7.

[0029] Concealment sampling: h-samples∼Beta(3+1,2+1)=Beta(4,3)⇒h-samples≈0.6.

[0030] Task notification: Trainee B received the task parameters "Number of steps = 7, Concealment level = 0.6".

[0031] Feedback: After completing the task, student B was successful when the number of steps was 7 and failed when the concealment level was 0.6.

[0032] Parameter update: The number of steps in the Beta distribution is updated to αs = 5 + 1 = 6, βs = 2 + 0 = 2.

[0033] The concealment Beta distribution is updated to αh=4+0=4, βh=3+1=4.

[0034] In one embodiment, the adaptive training system utilizes neural architecture search to automatically optimize the knowledge graph structure through the following specific steps: Search space definition: Define possible structural variation operations for knowledge graphs: Add an edge: connect two low-association nodes (e.g., "Hydraulic valve disassembly" → "Avionics fault diagnosis"); Delete edge: Remove low-weight connections (such as edges with a weight < 0.2); Adjust node weights: Increase the weight of key nodes (e.g., increase the weight of the "Sensor Calibration" node by 0.1). Reinforcement Learning Model Building (DQN): Define the state space S: the current knowledge graph structure and student ability data; Define action space A: structural mutation operations (add edge, delete edge, adjust weight); Define the reward function R: the student's skill improvement rate after task recommendation; The loss function of the reinforcement learning model is expressed as: ,in, As a discount factor, For the current network parameters, For target network parameters; Structural mutation and performance evaluation: Randomly select a structural mutation operation from the search space (e.g., add the edge "hydraulic valve disassembly" → "avionics fault diagnosis"), use a reinforcement learning model to evaluate the performance of the knowledge graph after mutation, and calculate the reward R; Structure update and deployment: Update the reinforcement learning network parameters θ according to the reward R, and retain the structure mutation operations with a boost rate greater than 0.

[0035] In this embodiment: Initial state: There is no direct connection between "hydraulic valve disassembly" and "avionics fault diagnosis" in the knowledge graph, and trainee D's avionics skill improvement rate is low (10%).

[0036] Structural variation: The NAS algorithm randomly selects the "Add Edge" operation, connecting "Hydraulic Valve Disassembly" → "Avionics Fault Troubleshooting".

[0037] Performance evaluation: After using the new structured knowledge graph, trainee D's avionics skill improvement rate increased to 20%, and the reward R=(0.2−0.1) / 0.1=100%.

[0038] Parameter update: The DQN network parameter θ is updated based on the high reward R, which strengthens the positive association of the "add edge" operation.

[0039] Structure deployment: The "hydraulic valve disassembly" → "avionics fault diagnosis" side is permanently retained in the knowledge graph.

[0040] In one embodiment, the specific steps for data anonymization and on-chain notarization by the blockchain-based notarization and data platform are as follows: Sensitive field identification: Define sensitive data fields (such as student ID, operation timestamp, raw values ​​of physiological signals); Application of desensitization algorithms: Laplace mechanism adds noise (differential privacy): ,in, For sensitivity, Budget for privacy; Data masking and permutation: Hash encryption of ID-type fields; Node registration and permission management: Defining a blockchain node set Each node holds a public-private key pair. ; Three-phase consensus process: Pre-Prepare phase: The master node broadcasts the message ⟨PRE-PREPARE,v,s,d,m>, where v is the view number, s is the sequence number, d is the digest, and m is the plaintext; Preparation phase: After node verification, broadcast ⟨PREPARE,v,s,d,i>, where i is the node ID; Commit phase: After a node has collected at least 2f+1 preparation messages, it broadcasts ⟨COMMIT,v,s,d,i> and performs a state update; Block generation and chained storage: After consensus is reached, block B={Header,Transactions} is generated, where the Header contains the previous block hash, timestamp, and Merkle root.

[0041] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An aircraft maintenance skills simulation training system, characterized in that, include: Intelligent mission simulator: It integrates replaceable avionics / hydraulics / engine components through modular hardware to build a physical simulation environment. Combined with multi-agent fault models and IoT sensor networks, it performs highly realistic hardware operation, multi-modal fault injection and operation data acquisition. The acquired physical operation data is synchronized to the three-dimensional virtual training platform in real time. 3D Virtual Training Platform: Based on Unity / Unreal Engine and integrating neural radiation field technology, a 1:1 digital twin aircraft model is constructed. Through AR collaborative spatial computing and dynamic occlusion management, it realizes the visualization of complex system principles, VR headset + gesture recognition immersive operation and multi-person collaborative training functions. The 3D virtual training platform records the students' operation trajectory and virtual environment data and inputs them into the dynamic evaluation engine. Dynamic evaluation engine: It adopts a dual-channel evaluation technology of neural symbols. The symbol layer verifies the safety operation specifications through the IF-THEN rule engine, while the neural layer parses the operation intention through the Transformer encoder and calculates the innovation score by combining the knowledge graph. It integrates cross-modal cognitive analysis to perform real-time capability evaluation, generate capability deficiency heat maps and output CBTA three-dimensional evaluation reports. The evaluation results drive the adaptive training system. Adaptive training system: Dynamically updates knowledge graph weights through meta-learning optimizer, and automatically optimizes knowledge graph structure by using Thompson sampling algorithm to control task complexity parameters and neural architecture search, completes personalized training task push and "evaluation-diagnosis-training" closed-loop evolution, and the generated personalized task parameters are synchronized to intelligent task simulation table and 3D virtual training platform to adjust hardware configuration and virtual scene difficulty. Blockchain-based evidence storage and data platform; After being anonymized, all system data is stored on the blockchain and the data platform's PBFT consensus mechanism. At the same time, QAR real flight data is integrated to optimize typical fault scenarios in the digital twin fault database.

2. The aircraft maintenance skills simulation training system according to claim 1, characterized in that, The key system is broken down into independent modules, each with an aerospace-grade interface. The module type is identified by RFID tags, and the corresponding simulation model is automatically loaded. Temperature, humidity, and vibration sensors are deployed to collect environmental data. Paris's law is used to simulate the propagation of metal fatigue cracks. The environmental data and component aging results are input into the rule engine to generate a composite scenario. The PTPv2 clock synchronization protocol is sampled to control the timestamp deviation of environmental sensors, component algorithms, and IoT sensor data. The coupled fault parameters are then input into the Unity / Unreal engine to dynamically adjust the state of the digital twin model.

3. The aircraft maintenance skills simulation training system according to claim 1, characterized in that, The specific steps for constructing a 1:1 digital twin aircraft model using the 3D virtual training platform are as follows: Data acquisition: Use laser scanners and SLR camera arrays to acquire aircraft entity data and generate dense point clouds; NeRF model training: Point cloud data is input into a neural radiation field network, and continuous view images are synthesized using a volume rendering equation. Where T(t) is the transmittance, Let c be the density field and c be the color field; Engine Integration: Export the trained NeRF model to a Unity / Unreal-recognizable format and render it in real time using a custom shader.

4. The aircraft maintenance skills simulation training system according to claim 3, characterized in that, The specific steps of the three-dimensional virtual training platform to achieve visualization of complex system principles, immersive operation of VR headset + gesture recognition and multi-person collaborative training functions through AR collaborative spatial computing and dynamic occlusion management are as follows: Distributed shared coordinates: Each student's AR device generates a local coordinate system using the SLAM algorithm, and then aligns it to the global coordinate system using the ICP algorithm. ,in, For local coordinate points, Let R be the global coordinate point, R be the rotation matrix, and t be the translation vector; Intent Prediction and Conflict Arbitration: A gesture intent prediction model is built based on federated learning (FedAvg algorithm) to predict operation paths in advance. Where w are model parameters. For learning rate, Let i be the loss function for the i-th device; Eye tracking and component importance assessment: The eye tracking module of the AR device records the student's gaze focus, and the rendering priority is calculated in combination with the functional importance of the components; Real-time rendering priority adjustment: Dynamically allocate rendering resources based on the calculated rendering priority, and prioritize the rendering of high-priority components.

5. The aircraft maintenance skills simulation training system according to claim 1, characterized in that, The specific steps of the dynamic evaluation engine in performing dynamic evaluation are as follows: Data input and preprocessing: AR devices are used to record gesture paths and gaze focus, and component status and fault phenomena are recorded in real time. HRV data is recorded using a heart rate monitor, and time-domain and frequency-domain indicators are calculated. Symbol layer security verification: Based on CAAC regulations and maintenance manuals, define safe operation rules, input operation trajectory and virtual environment data into the IF-THEN rule engine, and verify the rule triggering conditions through forward reasoning; Neural layer intent parsing and innovation scoring: Input gesture path and virtual environment data into Transformer encoder, generate context embedding vector through self-attention mechanism, retrieve similar maintenance cases in knowledge graph, and calculate cosine similarity between intent embedding and case embedding; Cross-modal cognitive analysis: eye-tracking trajectories generate decision focus heatmaps, voice commands are analyzed using NLP sentiment analysis to output command ambiguity, and physiological signals are analyzed using an HRV stress model to generate stress resilience scores; Competency Shortcomings Heatmap and CBTA Report Generation: Integrating symbolic layer validation results, neural layer innovation scores, and cross-modal analysis data, a competency shortcomings heatmap is generated through weighted fusion, presenting the three-dimensional assessment results of knowledge, skills, and attitudes in a structured manner.

6. The aircraft maintenance skills simulation training system according to claim 1, characterized in that, The adaptive training system dynamically updates the knowledge graph weights through a meta-learning optimizer, including the following steps: Data input and preprocessing: Historical training data collection technology is used to record trainees' past task parameters and results, and skill dimension data is extracted from CBTA reports by combining dynamic assessment results integration technology; Skills transfer trend prediction: Using LSTM model construction technology, through recursive calculations of input gate, forget gate, output gate and cell state, the historical skill data of trainees is analyzed and the future improvement trend of mechanical and avionics skills is predicted. Knowledge graph weight update: The weight of each node in the knowledge graph is dynamically adjusted by using weight update formula definition technology and node importance weighting technology. Effect verification and iteration: A / B testing technology is used to compare the recommendation effect of the knowledge graph before and after the update, and the parameters are adjusted by combining feedback iteration technology.

7. The aircraft maintenance skills simulation training system according to claim 6, characterized in that, The adaptive training system utilizes the Thompson sampling algorithm to control the specific steps of task complexity parameters. Parameter initialization: Define the task parameter range: number of steps Concealment Initialize the Beta distribution parameters: α=1, β=1; Thompson sampling execution: Step count sampling: Samples the step count parameter from the Beta distribution; Stealth sampling: Samples the stealth parameter from the Beta distribution; Task parameter generation and push: Generate task parameters based on sampling results: number of steps and concealment level; Parameter feedback and Beta distribution update: Update the Beta distribution parameters based on the students' task completion status.

8. The aircraft maintenance skills simulation training system according to claim 7, characterized in that, The specific steps of the adaptive training system in automatically optimizing the knowledge graph structure using neural architecture search are as follows: Search space definition: Define possible structural variation operations for knowledge graphs: Add an edge: connect two nodes with low correlation. Delete edge: Remove low-weight connections; Adjust node weights: Increase the weight of key nodes; Reinforcement learning model construction: Define the state space S: the current knowledge graph structure and student ability data; Define action space A: structural mutation operation; Define the reward function R: the student's skill improvement rate after task recommendation; The loss function of the reinforcement learning model is expressed as: ,in, As a discount factor, For the current network parameters, For target network parameters; Structural mutation and performance evaluation: Randomly select a structural mutation operation from the search space, use a reinforcement learning model to evaluate the performance of the mutated knowledge graph, and calculate the reward R; Structure update and deployment: Update the reinforcement learning network parameters θ according to the reward R, and retain the structure mutation operations with a boost rate greater than 0.

9. The aircraft maintenance skills simulation training system according to claim 1, characterized in that, The specific steps for data anonymization and on-chain storage in the blockchain-based evidence storage and data platform are as follows: Sensitive field identification: Define sensitive data fields; Application of desensitization algorithms: Adding noise using the Laplace mechanism. ,in, For sensitivity, Budget for privacy; Data masking and permutation: Hash encryption of ID-type fields; Node registration and permission management: Defining a blockchain node set Each node holds a public-private key pair. ; Three-phase consensus process: Preparatory phase: The master node broadcasts the message ⟨PRE-PREPARE,v,s,d,m>, where v is the view number, s is the sequence number, d is the digest, and m is the plaintext; Preparation phase: After node verification, broadcast ⟨PREPARE,v,s,d,i>, where i is the node ID; Commit phase: After a node has collected at least 2f+1 preparation messages, it broadcasts ⟨COMMIT,v,s,d,i> and performs a state update; Block generation and chained storage: After consensus is reached, block B={Header,Transactions} is generated, where the Header contains the previous block hash, timestamp, and Merkle root.

Images