Unmanned aerial vehicle course individualization ranking method based on historical training results

Abstract

The application discloses a kind of based on historical training results unmanned plane course individualization sequencing method, it is related to unmanned plane education technical field, including: the multidimensional historical results of unmanned plane training are collected, and are processed by forgetting curve model time weighting, generate the skill image reflecting the change of skill state;Establish the knowledge graph including course node and skill node;Skill image is input into deep knowledge tracking network modeling to obtain mastery probability prediction result, simultaneously, graph convolution coding is carried out to knowledge graph, and weak link of skill is identified in combination with prediction result;Based on the above prediction result and weak link of skill, a multi-objective optimization problem is constructed, and a personalized course sequence is generated by solving the problem using a multi-objective evolutionary algorithm;Finally, an iterative closed loop is realized according to the trigger condition in continuous monitoring.The application overcomes the defects of fixed course sequence, accurately tracks individual differences, preferentially strengthens weak links, significantly improves training resource utilization and shortens training period.

Description

Technical Field

[0001] This invention relates to the field of drone training and education technology, and more specifically, to a method for personalized ranking of drone courses based on historical training performance. Background Technology

[0002] With the widespread application of drone technology in fields such as aerial surveying, agricultural plant protection, logistics distribution, and power line inspection, the demand for professional operators is increasing daily. As a crucial link between theoretical learning and practical application, drone operation training directly impacts the efficiency of students' skill development and their final pass rate through its curriculum design and teaching sequence. A reasonable course sequence helps students gradually build their knowledge base, while an inappropriate course arrangement can lead to problems such as difficulty in addressing weaknesses and redundant training of already mastered skills, resulting in wasted resources. Therefore, the course sequencing method occupies a fundamental and important position in the drone training system.

[0003] Drone operation training courses typically cover multiple modules, including flight theory fundamentals, simulator operation training, live flight training, emergency response drills, and mission route planning. These modules have certain skill connections and prerequisite dependencies. Traditionally, drone training courses generally use a fixed linear sequence, conducting teaching activities in a predetermined order. In recent years, with the improvement of training management informatization, some training institutions have begun to try to introduce historical training performance data to record and track students' learning progress. However, existing technical solutions have extremely limited ways of utilizing historical training performance data, mostly only for simple pass / fail determination or performance statistics, that is, comparing students' interim test scores with preset passing thresholds to determine whether they are qualified to enter the next stage of learning.

[0004] The current simple pass / fail mechanism based on historical training scores has significant flaws. First, the rich information contained in historical training scores fails to be deeply explored and effectively utilized—complex correlation patterns exist between multi-dimensional data such as trainee control precision, error type distribution, training time consumption, and theoretical mastery. A single-dimensional score threshold cannot accurately reflect the trainee's true skill status and individual differences. Second, the fixed linear sequence of courses lacks sensitivity to the individualized learning needs of trainees. Regardless of a trainee's performance in a specific skill dimension, their subsequent course arrangement is determined by preset rules, resulting in repeated training of already mastered skills, wasting time, while weak skills cannot be prioritized for reinforcement and targeted improvement. Furthermore, existing technical solutions fail to fully consider the impact of forgetting curves on knowledge consolidation during human learning. A trainee's mastery of the same skill point at different times will decline over time, and a fixed course sequence cannot dynamically adjust review timing and training intensity according to the forgetting curve. These flaws combine to result in the current UAV training curriculum system having problems such as low efficiency in utilizing training resources, lengthy training cycles, and insufficient personalized training, failing to meet the needs of precise and efficient UAV operator training. Summary of the Invention

[0005] The present invention provides a method for personalized ranking of drone courses based on historical training results, in order to solve the above-mentioned technical problems.

[0006] To solve the above problems, the technical solution adopted by the present invention is as follows:

[0007] A method for personalized ranking of drone courses based on historical training performance, comprising:

[0008] Step 1: Collect multi-dimensional historical scores and construct a skill profile: Collect multi-dimensional historical scores during the drone training process, and perform time-weighted processing on the multi-dimensional historical scores through a forgetting curve model to generate a skill profile that reflects the temporal changes in skill status.

[0009] Step 2, establish a drone course knowledge graph: define course nodes and skill nodes in the drone training system, and establish the association edges between the course nodes and between the course nodes and the skill nodes to construct a drone course knowledge graph with a hybrid storage structure;

[0010] Step 3: Predict course mastery probability and identify skill gaps: Input the skill profile into a deep knowledge tracking network with self-attention mechanism for sequence modeling to obtain mastery probability prediction results; At the same time, perform graph convolutional encoding on the drone course knowledge graph, and combine the mastery probability prediction results to calculate the gap value of each skill dimension, and identify the key skill dimensions to be strengthened as skill gaps.

[0011] Step 4: Generate a personalized course sequence: Based on the mastery probability prediction results and the skill gaps, construct a multi-objective optimization problem. Under the premise of satisfying the prerequisite relationship constraints and skill mastery constraints in the associated edges, solve the problem using a multi-objective evolutionary algorithm to output the non-dominated optimal solution set and generate a personalized course sequence.

[0012] Step 5, achieve dynamic closed-loop optimization: During the continuous monitoring and collection of multi-dimensional historical scores, a judgment is made based on the preset fluctuation trigger conditions and the preset profile change trigger conditions. When the fluctuation trigger conditions or the profile change trigger conditions are met, the process returns to step 1 to update the skill profile, and then steps 3 and 4 are executed in sequence to generate an updated personalized course sequence and push it out, so as to form an iterative closed loop.

[0013] The technical effects of this technical solution are:

[0014] By collecting multidimensional historical scores and integrating them with a forgetting curve model for time-weighted processing to construct a skills profile, we break through the limitations of traditional simple score threshold judgment and can reflect each student's unique control precision, error distribution and learning trajectory in a deep and dynamic way.

[0015] By deeply combining the mastery probability prediction results of deep learning model sequence modeling with the graph convolutional coding results based on knowledge graphs, the single-dimensional learning evaluation is elevated to the level of a systematic global cognitive network, accurately identifying the key skill gaps that restrict the pass rate of the assessment.

[0016] Based on the prediction results and skill gaps, a multi-objective optimization problem is constructed and solved under the premise of meeting prerequisites and skill constraints. The generated personalized course sequence eliminates the need for repetitive training of already proficient skills, prioritizes and targets improvement of weak links, and significantly shortens the training cycle.

[0017] A dynamic closed-loop triggering mechanism based on multidimensional performance fluctuations and profile changes is established, which enables the course sequence to be adaptively updated along with the actual skill development trajectory of trainees, thus maximizing the long-term training effect.

[0018] Further, in step three, the deep knowledge tracking network integrating the self-attention mechanism includes an embedding layer, a sequence modeling layer, a self-attention mechanism layer, and a fully connected output layer connected in sequence. The prediction process of the deep knowledge tracking network includes: the embedding layer receiving the skill profile and mapping the original features to a unified hidden vector space; the sequence modeling layer using a long short-term memory network, utilizing the collaborative operation of the forget gate, input gate, and output gate to capture long-term temporal dependencies in the performance sequence; the self-attention mechanism layer converting the output sequence of the sequence modeling layer into a query vector sequence, a key vector sequence, and a value vector sequence, calculating the attention weight matrix through transpose matrix multiplication and normalization, multiplying the attention weight matrix with the value vector sequence, and integrating the relative position encoding vector; and the fully connected output layer converting the aggregated result into a mastery probability prediction result for each course to be predicted through a normalized exponential function.

[0019] In step three, graph convolutional encoding is performed on the UAV course knowledge graph, and the gap values ​​of each skill dimension are calculated in conjunction with the mastery probability prediction results. Key skill dimensions to be strengthened are identified as skill weaknesses. Specifically, this includes: using a multi-hop neighbor aggregation strategy for graph convolutional encoding; for each node in the UAV course knowledge graph, its current layer feature vector is weighted and aggregated with the feature vectors of all neighboring nodes in the previous layer, where the weight coefficients are determined by the product of the square root of the degree of the neighboring nodes. After processing with a linear rectified activation function, the node feature matrix of the next layer is obtained, generating a node representation vector that integrates the knowledge of the graph topology; based on the completed courses and their mastery, the gap values ​​between each skill dimension in the current skill profile and the assessment standards are calculated; the gap values ​​are multiplied by the weights of the corresponding skill dimensions on the overall pass rate to obtain a comprehensive urgency score for each skill dimension; and the skill dimensions with the highest comprehensive urgency scores (preset number) are selected as the skill weaknesses.

[0020] The technical advantages of this solution are: the combination of sequence modeling layer and self-attention mechanism can efficiently capture long-term and short-term learning patterns, and multi-hop neighbor aggregation graph convolution can fully explore the complex cross-dependency network between courses and skills, ensuring extremely high accuracy of learning ability assessment from the algorithm's bottom layer.

[0021] Further, in step four, the objective function of the multi-objective optimization problem includes: a first objective function that minimizes the sum of the expected training durations of all courses to be taken, and a second objective function that maximizes the weighted sum of the expected improvement of each key skill dimension and its urgency weight; step four, in which a multi-objective evolutionary algorithm is used to solve and output the non-dominated optimal solution set, specifically includes: obtaining an initial population by randomly generating a course sequence that satisfies the prior relationship constraint; selecting individuals from the initial population to form a competition set using a tournament selection operator to select the best individual as the parent individual; generating offspring individuals by probabilistically exchanging the gene positions of the parent individuals using a simulated binary crossover operator; replacing the gene positions of the offspring individuals using a polynomial mutation operator; merging the generated parent and offspring individuals into a joint population; executing a fast non-dominated sorting algorithm to divide the joint population into multiple non-dominated front layers; selecting the top-ranked individuals to enter the next generation population using an elite retention strategy; and outputting all non-dominated solutions in the last generation population to form the non-dominated optimal solution set after evolution.

[0022] In step four, the method of solving for the non-dominated optimal solution set using a multi-objective evolutionary algorithm can be replaced by generating a personalized course sequence using a Monte Carlo tree search algorithm based on upper confidence intervals. The specific execution process includes: each simulation execution includes four stages: selection, expansion, simulation, and backpropagation; in the selection stage, the score of each optional node is calculated according to the upper confidence interval formula, and child nodes are recursively selected. The calculated value of the upper confidence interval formula is the average simulated reward of the corresponding child node plus the exploration term, where the exploration term is proportional to the square root of the natural logarithm of the total number of visits to the current parent node divided by the number of visits to the child node; in the simulation stage, a greedy strategy is used to sort the remaining courses, and the simulated reward value is backpropagated to update the number of visits and the cumulative reward value of nodes on the search path; after executing a preset number of search iterations, the complete path with the highest cumulative access frequency is output as the personalized course sequence.

[0023] The technical effect of this solution is that whether it is using an evolutionary algorithm that includes non-dominated sorting to find the Pareto optimal solution, or using Monte Carlo tree search with upper confidence intervals to find the highest-yield path, it can perfectly balance the two conflicting goals of "lowest training time cost" and "maximum skill improvement".

[0024] Furthermore, the personalized ranking method for UAV courses is deployed in a distributed computing architecture that includes a master control node, computing nodes, and storage nodes: all trainees are hash-partitioned according to their skill level and training progress stage and assigned to corresponding computing nodes. The computing nodes use a multi-threaded parallel processing mechanism to independently execute the ranking calculation task for each trainee. In step three, the graphics processor is used to package the course data to be predicted for multiple trainees into a batch input, and the calculation of the mastery probability prediction result is completed synchronously. The graph convolution encoding is performed using sparse matrix operations based on the sparsity of the graph structure.

[0025] The technical effect of this solution is that by using hash partitioning scheduling, graphics processor batch inference acceleration, and sparse matrix graph convolution operations, the concurrent processing throughput of the system is greatly improved, enabling the method to maintain extremely high computational efficiency when facing the concurrent scheduling needs of a massive number of drone students. Attached Figure Description

[0026] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 This is a flowchart of the personalized ranking method for UAV courses based on historical training results as described in this embodiment of the invention.

[0028] Figure 2 This is a schematic diagram illustrating the core principle framework of the deep knowledge tracking network that integrates attention mechanism and knowledge graph convolutional coding in an embodiment of the present invention.

[0029] Figure 3 This is a flowchart illustrating the generation of personalized course sequences based on a multi-objective evolutionary algorithm under the constraint of prior relationship, as described in an embodiment of the present invention.

[0030] Figure 4 This is a flowchart illustrating the feedback control process for achieving dynamic closed-loop optimization in an embodiment of the present invention. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0032] Example 1

[0033] Please refer to Figure 1This invention provides a method for personalized ranking of drone courses based on historical training results, as detailed below:

[0034] Step 1 involves collecting multi-dimensional historical performance data and constructing a skills profile, specifically including the following workflow:

[0035] The flight attitude stability data acquisition subsystem collects data in real time using a three-axis gyroscope and a three-axis accelerometer mounted on the UAV, with a sampling frequency set to 50Hz. The system continuously monitors the changes in the UAV's roll, pitch, and yaw angles. At the data processing end, for each training mission's complete flight data, the system calculates the time-series standard deviation of the roll angle as a roll stability index, the time-series standard deviation of the pitch angle as a pitch stability index, and the time-series standard deviation of the yaw angle as a directional stability index. The overall score for flight attitude stability is obtained by weighted fusion of these three secondary indices, with the weight ratio determined based on the degree of influence of each index on the overall flight quality. When the standard deviations of the roll, pitch, and yaw angles are less than or equal to 5°, and 3°, the overall score is close to 100 points; when the standard deviations of the roll, pitch, and yaw angles exceed 20°, 18°, and 12°, the overall score drops below 60 points.

[0036] The flight path deviation data acquisition subsystem acquires the UAV's position coordinates in real time using the RTK positioning module mounted on the UAV, achieving a positioning accuracy better than 2cm. The system simultaneously acquires the reference coordinate sequence of the preset flight path, calculating the horizontal deviation distance and vertical altitude deviation between the actual flight path and the preset path. After each training task, the system summarizes the statistical characteristics of all waypoint position deviations from that task, including the mean deviation, maximum deviation, and the dispersion of the deviation distribution. The overall score for flight path deviation is evaluated based on the degree of conformity between the above statistical characteristics and preset standards; the smaller the mean deviation and the more concentrated the deviation distribution, the higher the score.

[0037] The emergency response delay data acquisition subsystem consists of an instructor control terminal and a trainee response terminal. The instructor control terminal issues a unified emergency signal, including voice prompts and visual warnings, at preset emergency scenario trigger points. The trainee response terminal records the time interval from the issuance of the emergency signal to the trainee completing the standardized operation. The system records response delay data for multiple emergency scenarios set up in each training session, including typical scenarios such as communication loss emergency handling, low battery return-to-home, and GPS signal loss / hovering. The comprehensive score for emergency response delay is determined based on the weighted average of the response delays for each scenario; the shorter the response delay and the higher the score for standardized action, the higher the comprehensive score.

[0038] The theoretical test score data collection subsystem automatically collects trainees' answer data after each theoretical training session through an online quiz system. The test questions cover multiple knowledge areas, including UAV flight principles, meteorological knowledge, civil aviation regulations, and equipment operation procedures. Each module contains a pre-set number of questions. The system records the trainees' accuracy rate in each module and uses a forgetting curve model to perform time-series weighted processing on historical test scores to generate a comprehensive theoretical score for each module.

[0039] When applying time-weighted processing to students' historical performance, the forgetting curve model employs the exponential decay principle of the Ebbinghaus forgetting curve. The forgetting function is expressed as:

[0040]

[0041] in, This represents the time interval from the current moment. This represents the specific time decay constant for the learner, the value of which is determined by a parameter identification algorithm based on the learner's historical learning data, typically ranging from 7 to 14 days. When new performance data is collected, the system first calculates the time interval between the corresponding moment of the performance data and the current moment, and then calculates the time decay coefficient of the historical performance based on the forgetting function. Performance data more recent to the current moment is assigned a higher weight value. Historical data older than 30 days from the current moment continues to have its contribution to the skill profile reduced according to the decay coefficient, and historical data older than 60 days from the current moment is directly removed to control the data size. When the skill profile is updated, the newly collected original performance data and the historical weighted performance data after forgetting processing are merged according to a preset ratio to ensure that the skill profile can reflect the learner's latest skill status in a timely manner while maintaining the smoothness and continuity of time-series changes.

[0042] The skills profile data structure includes a trainee identifier field, a skills score vector field, a confidence level field, and a timestamp field. The skills score vector consists of multiple primary dimensions, including flight stability control score, precise flight path control score, emergency response score, and theoretical knowledge score. Flight stability control score is further subdivided into three secondary dimensions: roll stability score, pitch stability score, and directional stability score. Precise flight path control score is further subdivided into two secondary dimensions: horizontal deviation score and altitude deviation score. Emergency response score is further subdivided into three secondary dimensions: reaction speed score, decision accuracy score, and operational standardization score. Theoretical knowledge score is further subdivided into three secondary dimensions: flight principles score, meteorological knowledge score, and regulatory knowledge score. The confidence level field is calculated based on the time freshness of the data in each dimension of the skills profile; the newer the data, the higher the confidence level.

[0043] Step 2, establish a knowledge graph for the drone course, which includes the following workflow:

[0044] The course node definition module defines all courses in the UAV training system as course nodes in the graph. Each course node includes attributes such as a unique course identifier, course name, course description, estimated training duration, course category, and course difficulty level. Course categories are divided into basic flight training, advanced operation training, mission application training, and comprehensive assessment training. Course difficulty levels are represented by integers from 1 to 5, with level 1 being beginner and level 5 being expert. In a complete UAV operator training system, typical course nodes include basic flight training courses, attitude control training courses, flight path planning training courses, emergency response training courses, mission payload operation training courses, formation coordination training courses, and advanced maneuver training courses.

[0045] The skill node definition module defines all skills required for UAV operation as skill nodes in the skill map. Each skill node includes attributes such as a unique skill identifier, skill name, skill description, skill level requirements, and skill evaluation indicators. The skill tagging system includes four primary dimensions: operational stability, flight path accuracy, emergency response capability, and theoretical knowledge level. The operational stability dimension is further subdivided into secondary skill tags such as roll control accuracy, pitch control accuracy, heading maintenance capability, and hovering stability. The flight path accuracy dimension is further subdivided into secondary skill tags such as straight-line flight accuracy, curved-course flight accuracy, altitude maintenance accuracy, and waypoint arrival accuracy. The emergency response capability dimension is further subdivided into secondary skill tags such as fault identification speed, emergency decision-making accuracy, special situation handling standardization, and return-to-home safety. The theoretical knowledge level dimension is further subdivided into secondary skill tags such as understanding of flight principles, mastery of meteorological knowledge, familiarity with regulations and provisions, and understanding of equipment structure.

[0046] The association edge definition module defines various associations between course nodes and between course nodes and skill nodes. Prerequisite dependency edges connect two course nodes with a prerequisite relationship, pointing from the prerequisite course to the subsequent course. Edge attributes include the minimum prerequisite interval in days and the prerequisite course completion requirement. For example, there is a prerequisite dependency edge between the attitude control training course and the basic flight training course, requiring students to achieve a score of 70 or higher in the basic flight training course before starting the attitude control training course. Skill mapping edges connect course nodes and skill nodes, pointing from the course node to the skill node trained by that course. Attributes include skill contribution weight and skill evaluation method. For example, there is a skill mapping edge between the flight path planning training course and the flight path accuracy skill, with a weight of 0.8, indicating that this course contributes 80% to the flight path accuracy skill. Skill dependency edges connect two skill nodes with a dependency relationship, indicating that mastering a higher-order skill requires a foundation in a lower-order skill. For example, hovering stability skill depends on two basic skills: roll control accuracy and pitch control accuracy.

[0047] The graph storage employs a hybrid storage structure combining adjacency lists and adjacency matrices. The adjacency list efficiently traverses the direct neighbors of each node, supporting neighbor information aggregation during graph convolutional encoding. The adjacency matrix quickly determines whether a relationship exists between two nodes, supporting topological sorting and prior constraint verification. The graph structure supports dynamic expansion with the addition of course and skill nodes. When new courses are added to the training system or skill requirements change, the system can dynamically add new nodes and their associated edges to the graph without rebuilding the entire graph structure.

[0048] The prerequisite relationship verification module rigorously validates prerequisite constraints before course sorting. The system executes a topological sorting algorithm on the set of courses to be sorted, checking for course sequences that violate prerequisite relationships. If the topological sorting fails, it indicates the presence of cyclic prerequisite dependencies in the course set. The system automatically detects and reports conflicting nodes, and the user must adjust the course structure before continuing the sorting operation. If the topological sorting succeeds, the system generates a topological sequence that satisfies the prerequisite relationships, serving as the boundary constraint for the feasible solution search space in subsequent optimization sorting.

[0049] Step 3: Predict the probability of mastering the course and identify skill gaps, combined with... Figure 2 As shown, the specific workflow includes the following:

[0050] The input layer of the deep knowledge tracing network receives an encoded representation of the trainee's historical performance sequence. Each training record contains performance data across multiple dimensions, which the system concatenates into a fixed-length input vector. This input vector first undergoes a dimensionality transformation through an embedding layer, mapping the original feature space to a unified hidden vector space. The embedding dimension is set to 64 dimensions. The output of the embedding layer is then fed into the sequence modeling layer.

[0051] The sequence modeling layer uses a Long Short-Term Memory (LSTM) network as the basic unit for sequence modeling. The network structure consists of two stacked LSTM layers, each containing 128 hidden units. The LSTM layers effectively capture long-term temporal dependencies in the trainee's performance sequence through the coordinated operation of forget gates, input gates, and output gates. The forget gate determines which historical information should be forgotten based on the hidden state of the previous time step and the input of the current time step. The input gate determines how much of the input information at the current time step should be added to the unit state. The output gate determines the output content at the current time step based on the current unit state and the hidden state. In this invention, the LSTM layer can learn the dynamic evolution of the trainee's skill state, such as the trend of a trainee's flight path control ability continuously improving after multiple training sessions, or the state of a trainee's stagnant emergency response skills.

[0052] The self-attention mechanism layer is embedded after the LSTM layer to calculate the correlation weights between different time steps in the student's historical performance sequence. The self-attention layer first converts the output sequence of the LSTM layer into three different vector representations: a query vector sequence Q, a key vector sequence K, and a value vector sequence V. The length of each vector sequence is equal to the number of time steps in the input sequence, and each vector has a 64-dimensional dimension. The computation process of the self-attention layer includes the following steps: First, the transpose of the query vector sequence and the key vector sequence is multiplied to obtain the attention score matrix, whose dimension is the number of time steps multiplied by the number of time steps; second, the attention score matrix is ​​scaled by a scaling factor set to the reciprocal of the square root of the key vector dimension; then, the scaled attention score matrix is ​​normalized using the softmax function to obtain the attention weight matrix; finally, the attention weight matrix is ​​multiplied by the value vector sequence V to obtain the output sequence of the self-attention layer. To enhance the model's ability to perceive sequence position information, the self-attention layer also integrates a relative position encoding mechanism, adding an encoding vector related to its relative position to each position in the input sequence.

[0053] The fully connected output layer aggregates the output sequence of the self-attention layer, specifically by performing a linear transformation on the hidden state at the last time step of the sequence. The output dimension of the fully connected output layer is equal to the number of courses to be predicted. For each course to be predicted, the output layer uses the softmax function to convert the result of the linear transformation into a predicted value of the student's mastery probability of that course. The output of the softmax function ranges from 0 to 1, and the sum of the predicted probabilities of all courses to be predicted is 1.

[0054] The deep knowledge tracing network is trained using a backpropagation algorithm combined with the Adam optimizer. The training data is labeled with students' actual assessment scores after completing the courses. The cross-entropy loss function is used to measure the difference between the network's predicted mastery probability distribution and the actual mastery distribution. During network training, a batch gradient descent strategy is employed, with a batch size of 32 and an initial learning rate of 0.001. An exponential decay strategy is used to reduce the learning rate to 0.9 times its original value every 10 training epochs. Once trained, the network can be used to predict students' mastery probability of uncompleted courses.

[0055] The graph convolutional encoding of the knowledge graph employs a multi-hop neighbor aggregation strategy, mapping course nodes and skill nodes in the knowledge graph to a unified low-dimensional vector space. The input to the graph convolutional network includes the graph's adjacency matrix, node feature matrix, and node degree matrix. The adjacency matrix has the same number of rows and columns as the total number of nodes in the graph, with elements being either 0 or 1, indicating a direct connection between two nodes. Each row of the node feature matrix corresponds to the initial feature vector of a node; the initial features of course nodes are encoded from course attribute information, and the initial features of skill nodes are encoded from skill attribute information. The feature dimension is set to 32 dimensions. The node degree matrix is ​​a diagonal matrix, with diagonal elements representing the degree values ​​of the corresponding nodes.

[0056] The graph convolutional network has 3 layers. Each layer performs the following calculation: For each node in the graph, its current layer's feature vector is weighted and aggregated with the feature vectors of all its neighbors in the previous layer. The weight coefficients are determined by the product of the square root of the degree of each neighbor node. The specific calculation formula is as follows:

[0057]

[0058] in, Indicates the first The node feature matrix of the layer The initial input is the node feature matrix; , It is an identity matrix, used to preserve the features of a node itself during feature aggregation. It is an adjacency matrix; for A degree matrix whose diagonal elements are equal to The sum of the elements in the corresponding row; For the first The learnable weight matrix of the layer, This is the ReLU activation function.

[0059] Each graph convolutional operation is followed by batch normalization to stabilize the network training process and accelerate convergence. After three layers of graph convolutional operations, the model can aggregate the structural information of nodes within their three-hop neighbors, generating node representation vectors that incorporate knowledge of the graph topology.

[0060] The skill gap identification module integrates the mastery probability prediction results from deep knowledge tracing networks and the graph convolutional coding results from knowledge graphs to automatically identify the skill gaps with the largest gaps compared to the assessment standards. The specific process is as follows: First, based on the student's current course completion and mastery, the student's existing skill set and proficiency level are determined. Second, based on the mapping relationship between course nodes and skill nodes in the course knowledge graph, the gap between each skill dimension in the student's current skill profile and the assessment standards is calculated. Third, the gap value is multiplied by the weight of the skill dimension's impact on the overall pass rate to obtain a comprehensive urgency score for each skill gap. Finally, a preset number of skill dimensions with the highest comprehensive urgency scores are selected as key targets for improvement, and courses that can improve these skill gaps are prioritized in subsequent course ranking optimization.

[0061] Step 4: Generate a personalized course sequence.

[0062] like Figure 3 As shown, this embodiment of the invention uses a multi-objective evolutionary algorithm to generate personalized course sequences, specifically including the following workflow:

[0063] The mathematical modeling process for multi-objective optimization problems includes two parts: defining the objective function and defining the constraints. The objective function comprises a first objective function and a second objective function. The first objective function minimizes the total training time, which is expressed as minimizing the total time cost required for the learner to complete all pending courses. Mathematically, it is the sum of the expected training times for all pending courses. The second objective function maximizes the efficiency of skill gap improvement, which is expressed as maximizing the expected comprehensive improvement in skill gaps after the learner completes the current optimization sequence. Mathematically, it is the weighted sum of the expected improvement in each key skill enhancement dimension and its urgency weight.

[0064] The constraints include prerequisite constraints and skill mastery constraints. The prerequisite constraint requires that each course node can only be scheduled into the sequence after all its predecessor course nodes have been completed; this constraint is implemented through a topological sorting sequence validity check. The skill mastery constraint requires that when each course node is scheduled, the student's current skill profile must meet the minimum skill admission requirements for that course; this constraint is implemented through a matching check between the skill profile and the course's skill requirements.

[0065] The multi-objective evolutionary algorithm employs a genetic algorithm based on non-dominated sorting as the solution method, using operators including tournament selection, simulated binary crossover, and polynomial mutation. The population size is set to 100, and the number of iterations is set to 500 generations. The initial population is obtained by randomly generating a sequence of courses that satisfies the prerequisite constraints, ensuring that each sequence passes topological sorting validity verification during the generation process. Fitness evaluation calculates the first and second objective function values ​​for each individual in the population.

[0066] In each generation of evolution, the tournament selection operator randomly selects a predetermined number of individuals from the population to form a tournament participant set, choosing the individual with the best fitness as the parent individual to enter the mating pool. The simulated binary crossover operator pairs the parent individuals in the mating pool, performing a crossover operation on each pair to generate two offspring individuals. The crossover operation uses a simulated binary encoding method to probabilistically exchange gene positions in the parent individuals. The multinomial mutation operator performs gene position mutation operations on the offspring individuals generated by crossover, randomly selecting a portion of the offspring individuals' gene positions with a predetermined mutation probability and replacing them with other legitimate course nodes. The mutation amplitude follows a multinomial distribution.

[0067] The non-dominated sorting process merges parent and offspring individuals generated during evolution into a joint population. Then, a fast non-dominated sorting algorithm is applied to this joint population, dividing it into multiple non-dominated levels. The fast non-dominated sorting algorithm iterates through each individual in the population, calculating the number of times that individual is dominated by other individuals and the set of other individuals dominated by that individual. These two indicators are used to divide the population into different non-dominated front layers. An elite preservation strategy selects individuals with high rankings in the non-dominated levels from the joint population to enter the next generation, ensuring that excellent non-dominated solutions are not lost.

[0068] After evolution, the algorithm outputs a Pareto optimal solution set consisting of all non-dominated solutions in the last generation of the population. Each solution in the Pareto optimal solution set is one that cannot be further improved on both objective functions simultaneously, relative to other solutions. The system then displays all course sequence schemes in the Pareto optimal solution set to trainees or training managers, allowing them to select the most suitable course sequence scheme based on actual training time constraints and skill enhancement priorities.

[0069] Monte Carlo Tree Search (MCS), as an alternative to multi-objective evolutionary algorithms, employs an upper confidence interval algorithm for node selection. Each simulation execution consists of four phases: selection, expansion, simulation, and backpropagation. The selection phase starts from the root node and recursively selects child nodes until a leaf node or a node that is not fully expanded is reached. Node selection calculates a score for each potential node based on the upper confidence interval formula, which considers both the node's average reward and the number of visits, balancing exploration and utilization.

[0070] Specifically, the formula for calculating the upper confidence interval is as follows:

[0071]

[0072] in, For the first The score of each child node, For the first The average simulated profit of each child node This represents the total number of visits to the current parent node. For the first The number of visits to each child node. To adjust the hyperparameters (preset constants) that balance exploration and utilization.

[0073] The expansion phase adds new child nodes to the selected nodes, with each new child node corresponding to a previously unselected course. The simulation phase starts from the newly expanded nodes and uses a greedy strategy to quickly sort the remaining courses, calculating the total revenue of the simulated sequence based on the objective function. The backpropagation phase transmits the simulated revenue values ​​back along the search path, updating the statistics of all nodes on the path, including the total number of visits and cumulative revenue values.

[0074] The Monte Carlo tree search algorithm continuously executes a preset number of search iterations, set to 10,000. The tree search depth is set to the total number of courses to be sorted. After each iteration, the system selects the course with the highest access frequency as the first course in the sequence based on the access statistics of all nodes. Then, it continues the search process with this course as the new root node until all courses have been selected. The final output course sequence is the complete path with the highest cumulative access frequency during the Monte Carlo tree search process.

[0075] Step 5, implement dynamic closed-loop optimization, which specifically includes the following workflow:

[0076] The adaptive mechanism sets re-ranking trigger conditions, including fluctuation trigger conditions and profile change trigger conditions. The fluctuation trigger condition is calculated as follows: the system continuously monitors the fluctuation range of the student's most recent 5 training scores. Re-ranking is triggered when the change in the latest training score relative to the previous score exceeds a preset fluctuation threshold. The fluctuation threshold is set to 20%, meaning that when the change rate between the latest score and the previous score exceeds ±20%, the system determines that the student's skill status has changed significantly. The profile change trigger condition is calculated as follows: the system updates the skill profile immediately after each student completes the current course. Re-ranking is triggered when the change in the score of any skill dimension in the skill profile exceeds a preset profile change threshold. The profile change threshold is set to 15 points, meaning that when the change in the score of any skill dimension exceeds 15 points, the system determines that the student's skill structure has changed significantly.

[0077] Once the reordering is triggered, the system automatically executes the complete reordering logic: First, it collects all training performance data of the student up to the current moment and updates the student's skill profile; second, it calls the deep knowledge tracking network to predict the student's mastery probability of all uncompleted courses; third, it calls the knowledge graph convolutional encoding to analyze the skill dependencies between courses; then, it inputs the updated skill profile, mastery probability prediction, and skill dependencies into the multi-objective optimization module to generate a new course sequence scheme; finally, it compares the newly generated course sequence scheme with the currently executed course sequence. If there are differences, it pushes updated course recommendation information to the student; if there are no differences, it maintains the current course sequence.

[0078] like Figure 4 As shown, the dynamic closed-loop optimization process includes six stages: performance collection, skill profile update, weakness identification, reordering execution, course delivery, and effect evaluation. The performance collection stage uses sensors, positioning modules, timers, and an online quiz system to collect multi-dimensional training performance data from trainees in real time. The data collection frequency is dynamically adjusted according to the training task type: 10Hz for basic flight training and 50Hz for task application training. The skill profile update stage executes a forgetting curve weighted fusion update algorithm based on the newly collected performance data to generate a skill profile vector reflecting the trainee's latest skill status. The weakness identification stage calls the weakness identification module based on the updated skill profile to calculate the gap between each skill dimension and the assessment standards and determine key areas for improvement. The reordering execution stage calls the multi-objective optimization module based on the weakness identification results and course knowledge graph constraints to generate the optimal course sequence that satisfies prerequisites and skill access conditions. The course delivery stage displays the optimized course sequence to trainees through the user interface, which also displays the name of the next recommended course, the estimated training duration, the skill improvement goals, and prerequisite requirements. The effectiveness evaluation process automatically records the course completion status, assessment scores, and skill improvement effects after students complete each course. The evaluation results serve as feedback data for subsequent course ranking optimization and enter the next round of closed-loop iteration.

[0079] The course review timing determination module uses a forgetting curve model to predict the point at which a student's mastery of a completed course will decline to a preset critical value, and automatically inserts corresponding review courses near this point. The forgetting curve parameters are dynamically determined based on each student's historical forgetting data using a parameter identification algorithm that fits the shape of the student's forgetting curve using the least squares method. The critical value is set to 0.4, meaning that when a student's predicted mastery of a completed course drops below 40% of the assessment standard, the system determines that the course content has entered the rapid forgetting zone and requires review. Each time it is predicted that a student's mastery of a completed course will decline below the critical value within a preset time window, the system automatically inserts the course into the appropriate position in the course sequence at the beginning of that time window, prioritizing its knowledge relevance to subsequent courses.

[0080] Example 2

[0081] To meet the concurrent sorting requirements in large-scale UAV training scenarios, this embodiment provides a distributed parallelized personalized course sorting method. Based on the technical framework of Embodiment 1, this solution introduces a distributed computing architecture to improve sorting efficiency.

[0082] The distributed architecture comprises three roles: master node, compute nodes, and storage nodes. The master node is responsible for task scheduling and result aggregation; the compute nodes are responsible for executing deep knowledge tracing network inference and graph convolutional coding computations; and the storage nodes are responsible for managing the course knowledge graph and student skill profile data. The master node and compute nodes communicate asynchronously via message queues, while the compute nodes and storage nodes exchange data via a caching protocol.

[0083] The student data partitioning strategy divides all students into hash partitions based on their skill level and training progress stage. Each computing node is responsible for processing a fixed number of student data partitions. Internally, each computing node employs a multi-threaded parallel processing mechanism, with each thread independently executing the sorting calculation task for one student. Threads are scheduled and load-balanced through a thread pool manager.

[0084] The accelerated course ranking computation method comprises two levels: model inference acceleration and optimization algorithm acceleration. Model inference acceleration employs a batch inference strategy, packaging the course data to be predicted for multiple students into a batch input, and utilizing the parallel computing capabilities of GPUs to simultaneously complete the mastery probability prediction calculations for multiple students. Graph convolutional encoding computation uses a sparse matrix operation optimization strategy, leveraging the sparsity of the graph structure to reduce unnecessary computation. Optimization algorithm acceleration employs an adaptive population initialization strategy, generating a high-quality initial population based on the topological features of the course knowledge graph, reducing the number of iterations required for the evolutionary algorithm to converge.

[0085] The output of the distributed ranking system is integrated through a unified result aggregation module. The aggregation module summarizes the ranking results of each computing node into a complete student course recommendation table according to the student identifier. The recommendation table includes information such as student identifier, recommended course sequence, predicted total training time, and expected skill improvement effect.

[0086] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for personalized ranking of drone courses based on historical training performance, characterized in that, include: Step 1: Collect multi-dimensional historical scores and construct a skill profile: Collect multi-dimensional historical scores during the drone training process, and perform time-weighted processing on the multi-dimensional historical scores through a forgetting curve model to generate a skill profile that reflects the temporal changes in skill status. Step 2, establish a drone course knowledge graph: define course nodes and skill nodes in the drone training system, and establish the association edges between the course nodes and between the course nodes and the skill nodes to construct a drone course knowledge graph with a hybrid storage structure; Step 3: Predict course mastery probability and identify skill gaps: Input the skill profile into a deep knowledge tracking network with self-attention mechanism for sequence modeling to obtain mastery probability prediction results; At the same time, perform graph convolutional encoding on the drone course knowledge graph, and combine the mastery probability prediction results to calculate the gap value of each skill dimension, and identify the key skill dimensions to be strengthened as skill gaps. Step 4: Generate a personalized course sequence: Based on the mastery probability prediction results and the skill gaps, construct a multi-objective optimization problem. Under the premise of satisfying the prerequisite relationship constraints and skill mastery constraints in the associated edges, solve the problem using a multi-objective evolutionary algorithm to output the non-dominated optimal solution set and generate a personalized course sequence. Step 5, achieve dynamic closed-loop optimization: During the continuous monitoring and collection of multi-dimensional historical scores, a judgment is made based on the preset fluctuation trigger conditions and the preset profile change trigger conditions. When the fluctuation trigger conditions or the profile change trigger conditions are met, the process returns to step 1 to update the skill profile, and then steps 3 and 4 are executed in sequence to generate an updated personalized course sequence and push it out, so as to form an iterative closed loop.

2. The method for personalized ranking of UAV courses based on historical training results according to claim 1, characterized in that, The multidimensional historical scores include flight attitude stability, flight path deviation, emergency response delay, and theoretical test scores; the process of collecting multidimensional historical scores in step one specifically includes: The roll angle, pitch angle and yaw angle variation data of the UAV are collected by a three-axis gyroscope and a three-axis accelerometer. The corresponding time series standard deviations are calculated as stability indicators, and the flight attitude stability is obtained by weighted fusion. The position coordinates of the actual flight path of the UAV are obtained by the real-time dynamic differential positioning module. The deviation distance in the horizontal direction and the height deviation in the vertical direction between the actual flight path and the preset flight path are calculated. The flight path deviation is obtained by summarizing the statistical features. The emergency response delay is obtained by sending an emergency signal through the instructor's control terminal and recording the time interval from when the trainee's response terminal receives the emergency signal to when the trainee completes the standard operation. The theoretical test score is generated by collecting the correct answer rates for each knowledge area through an online quiz system.

3. The method for personalized ranking of UAV courses based on historical training results according to claim 1, characterized in that, In step one, the method for time-weighting the multidimensional historical scores specifically includes: calculating the time interval between the newly collected historical scores and the current time; calculating the time decay coefficient of the multidimensional historical scores based on the principle of exponential decay, wherein the negative exponent of the natural constant is used as the decay result, where the numerator of the negative exponent is the time interval and the denominator is the time decay constant; the time decay constant is determined based on historical learning data through a parameter identification algorithm; performing time-series weighting on the multidimensional historical scores using the time decay coefficient, removing historical data that is more than a preset number of days away from the current time, and merging the newly collected original scores with the weighted scores according to a preset ratio to update the skill profile.

4. The method for personalized ranking of UAV courses based on historical training results according to claim 1, characterized in that, In step two, the method for constructing a UAV course knowledge graph with a hybrid storage structure specifically includes: encapsulating the course's unique identifier and expected training duration attribute as the course node, and encapsulating the skill evaluation index attribute as the skill node; defining prerequisite dependency edges pointing from prerequisite courses to subsequent courses, defining skill mapping edges pointing from course nodes to skill nodes that cultivate skills, and defining skill dependency edges between high-level skill nodes and low-level basic skill nodes to form the association edges; storing the UAV course knowledge graph using a hybrid storage structure combining adjacency lists and adjacency matrices, using adjacency lists for neighbor information aggregation operations, and using adjacency matrices for topological sorting and prerequisite constraint verification operations.

5. The method for personalized ranking of UAV courses based on historical training results according to claim 1, characterized in that, In step three, the deep knowledge tracking network integrating the self-attention mechanism includes an embedding layer, a sequence modeling layer, a self-attention mechanism layer, and a fully connected output layer connected in sequence. The prediction process of the deep knowledge tracking network includes: the embedding layer receives the skill profile and maps the original features to a unified hidden vector space; the sequence modeling layer uses a long short-term memory network and utilizes the collaborative operation of the forget gate, input gate, and output gate to capture long-term temporal dependencies in the performance sequence; the self-attention mechanism layer converts the output sequence of the sequence modeling layer into a query vector sequence, a key vector sequence, and a value vector sequence, calculates the attention weight matrix through transpose matrix multiplication and normalization, multiplies the attention weight matrix with the value vector sequence, and integrates the relative position encoding vector; the fully connected output layer converts the aggregated result into a mastery probability prediction result for each course to be predicted through a normalized exponential function.

6. The method for personalized ranking of UAV courses based on historical training results according to claim 1, characterized in that, In step three, graph convolutional encoding is performed on the UAV course knowledge graph, and the gap values ​​of each skill dimension are calculated in conjunction with the mastery probability prediction results. Key skill dimensions to be strengthened are identified as skill weaknesses. Specifically, this includes: using a multi-hop neighbor aggregation strategy for graph convolutional encoding; for each node in the UAV course knowledge graph, its current layer feature vector is weighted and aggregated with the feature vectors of all neighboring nodes in the previous layer, where the weight coefficients are determined by the product of the square root of the degree of the neighboring nodes. After processing with a linear rectified activation function, the node feature matrix of the next layer is obtained, generating a node representation vector that integrates the knowledge of the graph topology; based on the completed courses and their mastery, the gap values ​​between each skill dimension in the current skill profile and the assessment standards are calculated; the gap values ​​are multiplied by the weights of the corresponding skill dimensions on the overall pass rate to obtain a comprehensive urgency score for each skill dimension; and the skill dimensions with the highest comprehensive urgency scores (preset number) are selected as the skill weaknesses.

7. The method for personalized ranking of UAV courses based on historical training results according to claim 1, characterized in that, In step four, the objective function of the multi-objective optimization problem includes: a first objective function that minimizes the sum of the expected training durations of all courses to be taken, and a second objective function that maximizes the weighted sum of the expected improvement of each key skill enhancement dimension and its urgency weight. Step four involves solving and outputting the non-dominated optimal solution set using a multi-objective evolutionary algorithm, specifically including: obtaining an initial population by randomly generating a course sequence that satisfies the prior relationship constraints; selecting individuals from the initial population using a tournament selection operator to form a competition set to select the best individual as the parent individual; generating offspring individuals by probabilistically exchanging the gene positions of the parent individuals using a simulated binary crossover operator; replacing gene positions in the offspring individuals using a polynomial mutation operator; merging the generated parent and offspring individuals into a joint population; executing a fast non-dominated sorting algorithm to divide the joint population into multiple non-dominated front layers; selecting top-ranked individuals to enter the next generation population using an elite retention strategy; and outputting all non-dominated solutions in the last generation population to form the non-dominated optimal solution set after evolution.

8. The method for personalized ranking of UAV courses based on historical training results according to claim 1, characterized in that, In step four, the method of solving for the non-dominated optimal solution set using a multi-objective evolutionary algorithm can be replaced by generating a personalized course sequence using a Monte Carlo tree search algorithm based on upper confidence intervals. The specific execution process includes: each simulation execution includes four stages: selection, expansion, simulation, and backpropagation; in the selection stage, the score of each optional node is calculated according to the upper confidence interval formula, and child nodes are recursively selected. The calculated value of the upper confidence interval formula is the average simulated reward of the corresponding child node plus the exploration term, where the exploration term is proportional to the square root of the natural logarithm of the total number of visits to the current parent node divided by the number of visits to the child node; in the simulation stage, a greedy strategy is used to sort the remaining courses, and the simulated reward value is backpropagated to update the number of visits and the cumulative reward value of nodes on the search path; after executing a preset number of search iterations, the complete path with the highest cumulative access frequency is output as the personalized course sequence.

9. The method for personalized ranking of UAV courses based on historical training results according to claim 1, characterized in that, In step five, the fluctuation trigger condition is that the rate of change of the latest training score relative to the previous score exceeds a preset fluctuation threshold; the profile change trigger condition is that the score change of any skill dimension in the skill profile exceeds a preset profile change threshold. The dynamic closed-loop optimization also includes a course review timing determination process: based on the forgetting curve model, the predicted mastery level of the completed courses is predicted. When the predicted mastery level drops below a preset threshold, the corresponding review course is automatically inserted into the course sequence to be completed at the beginning of a preset time window.

10. The method for personalized ranking of UAV courses based on historical training results according to any one of claims 1 to 9, characterized in that, The personalized ranking method for UAV courses is deployed in a distributed computing architecture that includes a master control node, computing nodes, and storage nodes. All trainees are hash-partitioned according to their skill level and training progress stage and assigned to corresponding computing nodes. Each computing node independently executes the ranking calculation task for each trainee using a multi-threaded parallel processing mechanism. In step three, the graphics processor is used to package the course data to be predicted for multiple trainees into a batch input, and the calculation of the mastery probability prediction result is completed synchronously. The graph convolution encoding is performed using sparse matrix operations based on the sparsity of the graph structure.

Images