Unmanned aerial vehicle subsystem health assessment method and system

Abstract

The invention discloses an unmanned aerial vehicle subsystem health assessment method and system, and relates to the technical field of unmanned aerial vehicle health management, and the method comprises the steps: obtaining the data of a multi-source sensor of an unmanned aerial vehicle, screening the features strongly correlated with a fault state through a Spearman rank correlation coefficient, and obtaining key monitoring parameters; according to the self-adaptive mutation mechanism, the virtual evolution strategy and the double-objective optimization strategy, an improved multi-objective genetic algorithm optimization model is constructed, then key monitoring parameters are input, and a visual report of a health state assessment result and confidence rating is obtained. According to the method, the problems that traditional single-target optimization neglects the misjudgment rate, the parameter search efficiency is low and the evaluation credibility is insufficient are solved, accurate evaluation is achieved, the evaluation accuracy is improved to 84.86%, the key misjudgment rate is reduced to 0.009%, the convergence speed is improved by 40% compared with a traditional method, the safety and real-time performance of unmanned aerial vehicle health state recognition are remarkably optimized, and the method is suitable for popularization and application. The method is suitable for embedded system deployment.

Description

Technical Field

[0001] The present invention relates to the technical field of drone health management, and in particular to a drone subsystem health assessment method and system. Background Art

[0002] With the rapid development of drone technology, its applications are becoming increasingly broad, encompassing a wide range of scenarios, including real-time monitoring, search and rescue, pipeline and power inspections, environmental monitoring, and agricultural plant protection. In this context, ensuring safe and reliable drone flight is crucial. From a reliability perspective, software or hardware failures or performance anomalies during drone flight can lead to mission interruptions, crashes, and even threaten the lives and property of personnel on the ground. Therefore, conducting health assessments of drone subsystems meets the development needs of the drone industry and market demands.

[0003] Traditional health assessment methods have the following limitations:

[0004] 1. Single-objective optimization flaws: Existing methods often use accuracy as a single optimization objective, ignoring key safety indicators such as false positive rate;

[0005] 2. Low parameter optimization efficiency: Traditional optimization methods such as grid search have exponentially increasing computational complexity as the hyperparameter space expands;

[0006] 3. Delayed state recognition: Insufficient sensitivity in detecting the sub-health status of key components such as gyroscopes;

[0007] 4. Low credibility of evaluation results: Lack of quantitative confidence indicators affects the reliability of maintenance decisions.

[0008] Therefore, a UAV subsystem health assessment method and system are urgently needed to solve these problems. Summary of the Invention

[0009] To solve the above problems, this application proposes a drone subsystem health assessment method and system, which solves the problems of traditional single-objective optimization ignoring the misjudgment rate, low parameter search efficiency and insufficient assessment credibility, and achieves accurate assessment.

[0010] The present invention provides a method for evaluating the health of a drone subsystem, comprising the following steps:

[0011] S1. Acquire multi-source sensor data of the drone; the multi-source sensor data of the drone includes: normal flight data, gyroscope fault data, accelerometer fault data, and GPS fault data;

[0012] S2. Based on the multi-source sensor data of the UAV, the Spearman rank correlation coefficient is used to screen the features that are strongly correlated with the fault state to obtain key monitoring parameters; the key monitoring parameters include: roll angular velocity, pitch angular velocity, yaw angular velocity and voltage parameters;

[0013] S3. Construct an improved multi-objective genetic algorithm optimization model based on the adaptive mutation mechanism, virtual evolution strategy, improved environmental selection mechanism, dual-objective optimization strategy, and parameter space and exploration strategy;

[0014] S4. Input the key monitoring parameters into the improved multi-objective genetic algorithm optimization model to obtain a visual report of the health status assessment results and confidence rating.

[0015] Preferably, in S3, an improved multi-objective genetic algorithm optimization model is constructed based on a dual-objective optimization strategy, an adaptive mutation mechanism, a virtual evolution strategy, an improved environment selection mechanism, and a parameter space and exploration strategy, specifically including:

[0016] S301, obtaining historical visualization reports of multi-source sensor data, health status historical assessment results, and confidence ratings of drones;

[0017] S302. Based on the historical data of the multi-source sensors of the UAV, the Spearman rank correlation coefficient is used to screen the features that are strongly correlated with the fault status to obtain the key monitoring historical parameters;

[0018] S303, constructing a training set and a test set based on historical visualization reports of key monitoring historical parameters, historical health status assessment results, and confidence ratings;

[0019] S304, based on the basic objective optimization strategy, a dual-objective optimization strategy is obtained by defining the fitness function as a dual-objective vector;

[0020] S305. Based on the basic adaptive mechanism, an adaptive mutation mechanism is obtained by dynamically adjusting the mutation probability and the mutation intensity;

[0021] S306. Based on the basic genetic algorithm, a virtual evolution strategy is obtained by embedding a local fine search in the global search process;

[0022] S307. Based on the basic environment selection mechanism, an improved environment selection mechanism is obtained by combining non-dominated sorting with optimized crowding distance calculation;

[0023] S308. Based on the basic parameter space and exploration strategy, a parameter space expansion and exploration strategy is obtained by expanding the search space boundary and adopting a logarithmic scale representation;

[0024] S309. Using a meta-learning algorithm, the adaptive mutation mechanism, virtual evolution strategy, improved environmental selection mechanism, dual-objective optimization strategy, and parameter space and exploration strategy are trained and tested in sequence based on the training set and the test set until the trained adaptive mutation mechanism, virtual evolution strategy, improved environmental selection mechanism, dual-objective optimization strategy, and parameter space and exploration strategy meet the set requirements. The trained adaptive mutation mechanism, virtual evolution strategy, improved environmental selection mechanism, dual-objective optimization strategy, and parameter space and exploration strategy are connected as an improved multi-objective genetic algorithm optimization model.

[0025] Preferably, in S305, based on the original adaptive mechanism, the adaptive mutation mechanism is obtained by dynamically adjusting the mutation rate and mutation intensity, which specifically includes:

[0026] The adaptive mutation rate is calculated based on the basic mutation rate and the diversity factor;

[0027] The adaptive mutation strength is calculated based on the basic mutation strength, diversity factor and boundary perception factor;

[0028] Based on the original adaptive mechanism, the adaptive mutation mechanism is obtained according to the adaptive mutation rate and adaptive mutation strength.

[0029] Preferably, the expression of the diversity factor is:

[0030]

[0031] Among them, S i is the number of individuals in the i-th frontier, k is the number of non-dominated frontiers, N pop is the initial population size;

[0032] The expression of the adaptive mutation rate is:

[0033]

[0034] Among them, P m is the basal mutation rate, is the adaptive mutation rate.

[0035] Preferably, the boundary perception factor E f The expression is:

[0036]

[0037] Among them, E f is the boundary perception factor, x is the relative position of the current solution within the feasible region of its corresponding decision variable (normalized value);

[0038] The expression of the adaptive mutation strength is:

[0039] σ adaptive =σ×(1-0.5D f )×E f ;

[0040] Among them, σ adaptive is the adaptive mutation strength, and σ is the basic mutation strength.

[0041] Preferably, in S306, based on the basic genetic algorithm, a virtual evolution strategy is obtained by embedding a local fine search in the global search process, which specifically includes:

[0042] Get the fixed number of generations that perform the deep search of elite individuals;

[0043] Within a fixed generation, a number of elite individuals are randomly selected from the Pareto frontier;

[0044] Perform Gaussian perturbations on several elite individuals to generate local variants;

[0045] Perform fitness evaluation on the local variant to obtain the fitness value;

[0046] The fitness value is used to determine whether to perform population update, thereby obtaining a virtual evolution strategy; the population update is to replace the worst individual if the fitness value of the best variant among the local variants is improved.

[0047] Preferably, in S307, based on the basic environment selection mechanism, the improved environment selection mechanism is obtained by combining non-dominated sorting with optimized crowding distance calculation, which specifically includes:

[0048] Based on the basic environment selection mechanism, the merged population is divided into several frontiers through non-dominated sorting, and several frontiers are added to the new population until the population size limit is reached;

[0049] If the last frontier cannot be added to the new population, the optimized crowding distance sorting is used to determine the required number of frontiers, resulting in an improved environment selection mechanism.

[0050] Preferably, if the last frontier cannot be added to the new population, the optimized crowding distance sorting is used to determine the required number of frontiers, and the specific content of the improved environment selection mechanism is:

[0051] If the last frontier cannot be added to the new population, the crowding distance of all frontiers is initialized to 0;

[0052] For each objective function, the frontier is sorted by the objective function's target and the crowding distance of the frontier is set to infinity;

[0053] The crowding distance increment of the intermediate frontier is calculated, and finally the frontiers are sorted in descending order of crowding distance to determine the required number of frontiers, thus obtaining an improved environment selection mechanism.

[0054] Preferably, in S308, based on the basic objective optimization strategy, a dual-objective optimization strategy is obtained by defining the fitness function as a dual-objective vector, specifically including:

[0055] Based on the basic objective optimization strategy, a dual-objective optimization strategy is obtained by simultaneously optimizing two mutually constrained objectives; the two mutually constrained objectives include: maximizing the F1 score and minimizing the false positive rate.

[0056] The present invention provides a UAV subsystem health assessment system, comprising:

[0057] An acquisition module is used to acquire multi-source sensor data of the UAV; the multi-source sensor data of the UAV includes: normal flight data, gyroscope fault data, accelerometer fault data and GPS fault data;

[0058] An extraction module is used to filter features strongly correlated with the fault state based on the multi-source sensor data of the UAV using the Spearman rank correlation coefficient to obtain key monitoring parameters; the key monitoring parameters include: roll angular velocity, pitch angular velocity, yaw angular velocity and voltage parameters;

[0059] Model building module, used to build an improved multi-objective genetic algorithm optimization model based on dual-objective optimization strategy, adaptive mutation mechanism, virtual evolution strategy, improved environment selection mechanism and parameter space and exploration strategy;

[0060] The evaluation module is used to input key monitoring parameters into the improved multi-objective genetic algorithm optimization model to obtain a visual report of the health status assessment results and confidence rating.

[0061] In summary, the drone subsystem health assessment method and system of the present invention have the following advantages over traditional technologies:

[0062] (1) The present invention solves the problems of traditional single-objective optimization, such as ignoring the misjudgment rate, low parameter search efficiency, and insufficient evaluation credibility.

[0063] (2) The present invention constructs a two-level evaluation system including normal and fault states, extracts key sensor parameters such as roll angular velocity, and uses an improved multi-objective genetic algorithm to achieve accurate evaluation.

[0064] (3) The present invention innovatively designs an adaptive mutation mechanism to enhance search efficiency by dynamically adjusting the mutation rate and boundary perception factor.

[0065] (4) The present invention combines the virtual evolution strategy to periodically conduct local depth search on elite individuals, thereby accelerating the convergence process.

[0066] (5) The present invention improves the evaluation accuracy to 84.86%, reduces the critical misjudgment rate to 0.009%, and increases the convergence speed by 40% compared with the traditional method. It significantly optimizes the security and real-time performance of drone health status identification and is suitable for embedded system deployment.

[0067] The technical method of the present invention is further described in detail below through the accompanying drawings and examples. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 This is a flow chart of a method for health assessment of a drone subsystem according to the present invention;

[0069] Figure 2 This is an architectural diagram of the drone subsystem health assessment system of the present invention;

[0070] Figure 3 This is a heat map of the correlation between the multi-source sensor data set of the UAV avionics system of the present invention;

[0071] Figure 4 Improved multi-objective genetic algorithm flow chart for the present invention;

[0072] Figure 5 Schematic diagram of the virtual evolution strategy in the present invention;

[0073] Figure 6 This is a comparison chart of the convergence curves of the evolutionary process of the standard NSGA-II and the improved multi-objective genetic algorithm in the present invention. DETAILED DESCRIPTION

[0074] The technical method of the present invention is further described below through the accompanying drawings and embodiments. It should be noted that unless otherwise specifically stated, the relative arrangement of components and steps, numerical expressions and values ​​described in these embodiments do not limit the scope of this application.

[0075] The following description of at least one exemplary embodiment is merely illustrative in nature and is in no way intended to limit the present disclosure, its application, or uses.

[0076] Technologies, systems, and devices known to those of ordinary skill in the relevant art may not be discussed in detail, but where appropriate, they should be considered part of the specification.

[0077] In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not limiting. Therefore, other examples of the exemplary embodiments may have different values.

[0078] Unless otherwise defined, technical or scientific terms used in the present invention shall have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention belongs.

[0079] The present invention provides a method for evaluating the health of a drone subsystem. Figure 1 As shown in the architecture diagram Figure 2 Shown, including:

[0080] Step S1: Acquire multi-source sensor data of the drone, wherein the multi-source sensor data of the drone includes normal flight data, gyroscope fault data, accelerometer fault data, and GPS fault data.

[0081] In step S1, data collection and preprocessing are first required, including:

[0082] Step S101: Collect and integrate multi-source sensor data from drones.

[0083] The present invention collects multi-source sensor data from the drone's avionics system, including real-time monitoring data from key subsystems such as the inertial measurement unit, global positioning system (GPS), and power management system. The data acquisition frequency is set to 5Hz, covering normal flight conditions and three typical failure modes of the drone's avionics system (accelerometer failure, gyroscope failure, and GPS failure). The collected multi-source sensor data from the drone's avionics system is stored in CSV format. Each file contains a timestamp sequence and 30-dimensional sensor parameters. The data structure (partial) is shown in Table 1:

[0084] Table 1 Data structure (partial)

[0085]

[0086]

[0087]

[0088] Step S102: performing outlier detection and data cleaning on the integrated drone multi-source sensor data.

[0089] Step S102 specifically uses the improved interquartile range (IQR) method to process outliers and establish dynamic thresholds for each sensor parameter. For each feature column x, its 25% quantile Q1 and 75% quantile Q3 are calculated. The outlier determination interval is:

[0090] Lower Bound=Q1-1.5×IQ R ;

[0091] Upper Bound=Q3+1.5×IQ R ;

[0092] Among them, Q R =Q3-Q1, Q R The interquartile range represents the distribution range of the middle 50% of the data in the dataset. By marking outliers instead of directly deleting them, the integrity of the original data is preserved.

[0093] Step S2: Based on the multi-source sensor data of the UAV, the Spearman rank correlation coefficient is used to screen features that are strongly correlated with the fault state to obtain key monitoring parameters. The key monitoring parameters include: roll angular velocity, pitch angular velocity, yaw angular velocity, and voltage parameters.

[0094] Step S2 specifically performs feature engineering and correlation analysis on the multi-source sensor data of the drone after outlier detection and data cleaning in step S102.

[0095] First, use the Spearman rank correlation coefficient to screen the features that are strongly correlated with the fault state. Calculate the correlation coefficient between each feature and the target variable fault_state, and select the top four key features, namely roll angular velocity, pitch angular velocity, yaw angular velocity, and voltage parameters; construct a correlation heat map, such as Figure 3 As shown, the calculation formula of the correlation coefficient is:

[0096] ρ=1-(6∑d i 2 ) / (n(n 2 -1);

[0097] Where ρ is the correlation coefficient, d i is the rank difference between the two variables, and n is the sample size (i.e. the total number of observations).

[0098] The feature selection module outputs the key_features.txt file for subsequent modeling.

[0099] Then the filtered and selected feature data is normalized and segmented. The Z-score standardization is used to process the feature data. The formula is:

[0100] z=(x-μ) / σ;

[0101] Among them, z is the standardized result value (indicating the number of standard deviations of the original data points from the mean), x is the original feature data value (i.e., the original sensor measurement value that needs to be standardized), μ is the mean value of the feature (indicating the arithmetic mean of all observations of the feature), and σ is the standard deviation of the feature (a statistic indicating the degree of dispersion of the feature data).

[0102] Step S3: Based on the adaptive mutation mechanism, virtual evolution strategy, improved environment selection mechanism, dual-objective optimization strategy and parameter space and exploration strategy, an improved multi-objective genetic algorithm optimization model is constructed. The construction process is as follows: Figure 4 The specific contents include:

[0103] Step S301: Obtain historical data of multi-source sensors of the drone, historical health status assessment results, and historical visualization reports of confidence ratings.

[0104] Step S302: Based on the historical data of the multi-source sensors of the UAV, the Spearman rank correlation coefficient is used to screen the features that are strongly correlated with the fault status to obtain the key monitoring historical parameters.

[0105] Step S303: Construct training and test sets based on historical visualization reports of key monitoring parameters, historical health status assessment results, and confidence ratings. Split the dataset into training and test sets in a 7:3 ratio to maintain consistent category distribution.

[0106] Step S304: Based on the basic objective optimization strategy, a dual-objective optimization strategy is obtained by defining the fitness function as a dual-objective vector.

[0107] Step S304 specifically involves algorithm initialization and coding design. Chromosomes are represented by real number coding, and each individual corresponds to the var_smoothing parameter of the Gaussian Naive Bayes model. The population initialization formula is:

[0108] x j =-12+6×U(0,1), j=1,2,...,N pop ;

[0109] Among them, x j is the initial population, j is the individual index (indicates the jth individual in the population), N pop is the initial population size, and U(0,1) is a uniformly distributed random number.

[0110] The fitness function is defined as a dual objective vector:

[0111] maxF(x)=[f1(x),-f2(x)];

[0112] Among them, F(x) is the fitness function vector, x is the parameter, f1(x) is to maximize the F1 score, and f2(x) is to minimize the misjudgment rate.

[0113] The dual-objective optimization strategy is designed for the specific application of drone health assessments. It expands the single-objective optimization of classification performance into a multi-objective problem that balances accuracy and reliability. This dual-objective optimization strategy is directly reflected in the design of the fitness function. It simultaneously optimizes two mutually constrained objectives: maximizing the F1 score f1(x) and minimizing the false positive rate f2(x), with a particular focus on misclassifying faults as normal.

[0114] Mathematically, this is a dual-objective optimization problem.

[0115] The formula for maximizing the F1 score f1(x) is a standard classification evaluation metric:

[0116]

[0117] Precision is the accuracy rate, and recall is the recall rate. This indicator can comprehensively measure the classification performance of the model.

[0118] The formula for minimizing the misjudgment rate f2(x) is:

[0119]

[0120] Among them, I is the indicator function, n is the total number of test samples, i is the sample index, y i is the true label of the i-th sample, is the predicted label of the i-th sample. It takes the value 1 when the faulty sample is misclassified as normal, and 0 otherwise.

[0121] The dual-objective design of this invention incorporates a bias based on domain knowledge: in drone health assessment, misclassifying a fault as normal (a false negative) is more serious than misclassifying a normal as a fault (a false positive), as the former can lead to catastrophic failure. Therefore, by explicitly defining this bias and incorporating it into the optimization objective, the algorithm is able to learn a safer and more reliable decision boundary.

[0122] The Pareto optimization framework of NSGA-II in this invention offers a natural advantage for handling such multi-objective problems. It eliminates the need to pre-determine the weights of two objectives, instead generating a series of Pareto-optimal solutions that demonstrate the potential trade-offs between different objectives. Ultimately, the algorithm selects the solution with the highest F1 value from the Pareto frontier as the model parameter, while ensuring an acceptable error rate. This optimization strategy ensures that the resulting improved multi-objective genetic algorithm optimization model possesses sufficient fault detection capabilities while pursuing high accuracy, making it particularly suitable for safety-critical applications.

[0123] Step S305: Based on the basic adaptive mechanism, an adaptive mutation mechanism is obtained by dynamically adjusting the mutation probability and the mutation intensity.

[0124] Step S305 specifically achieves a balance between exploration and development by dynamically adjusting the probability and intensity of mutation. The adaptive mutation rate is calculated by first calculating the diversity factor, which is expressed as:

[0125]

[0126] Among them, D f is the diversity factor, i is the number of individuals in the frontier, k is the number of non-dominated frontiers, S i is the number of individuals in the i-th frontier.

[0127] Then, based on the basic mutation rate and diversity factor, the adaptive mutation rate is calculated, and the expression is:

[0128]

[0129] in, is the adaptive mutation rate, P m is the basic mutation rate. When the population diversity is high, D f The value is large, then increasing the mutation rate can accelerate the search process; on the contrary, when the population tends to converge, D f Smaller values ​​reduce the mutation rate and help maintain the high-quality solutions found.

[0130] In one embodiment of the present invention, when the population diversity D f When >0.7, the mutation rate increases to 1.5 times the baseline value.

[0131] The adaptive mutation strength is calculated based on the base mutation strength, diversity factor, and boundary perception factor, as follows:

[0132] By introducing the boundary perception factor to optimize Gaussian variation, the boundary perception factor E f The calculation formula is:

[0133]

[0134] Among them, x is the relative position of the current solution within the feasible range of its corresponding decision variable, N p is the normalized parameter position.

[0135] By calculating the boundary perception factor, the present invention enables genes close to the parameter boundary area to obtain a smaller mutation amplitude, avoiding invalid solutions, while maintaining sufficient exploration capabilities in the intermediate area, significantly improving the search efficiency and stability of the algorithm in complex parameter spaces.

[0136] Then the adaptive mutation strength is calculated. The calculation expression of the adaptive mutation strength is:

[0137] σ adaptive =σ×(1-0.5D f )×E f ;

[0138] Among them, σ adaptive is the adaptive mutation strength, and σ is the basic mutation strength.

[0139] The adaptive mutation intensity adjustment strategy complements the mutation rate: when diversity is high, the mutation intensity is reduced to avoid drastic mutations that destroy existing high-quality genes; when diversity is low, the mutation intensity is increased to help the population escape the local optimum.

[0140] In the traditional NSGA-II algorithm, the mutation rate and strength are typically fixed values, making them unable to adapt to the needs of different stages in the optimization process, resulting in low search efficiency. Therefore, the improved NSGA-II algorithm in this paper adopts a composite adaptive mechanism, namely an adaptive mutation mechanism. Based on the diversity factor, which quantifies the degree of diversity in the current population, the algorithm dynamically adjusts the mutation probability and strength, enabling it to automatically adapt to the population state during the search process.

[0141] S306, based on the evolution process of the basic genetic algorithm, by embedding local fine search in the global search process, a virtual evolution strategy is obtained, such as Figure 5 shown.

[0142] The specific contents of S306 are:

[0143] Obtain a fixed number of generations for performing the elite individual deep search. In one embodiment of the present invention, the elite individual deep search is performed every 5 generations.

[0144] Within a fixed number of generations, a number of elite individuals are randomly selected from the Pareto front. Preferably, three elite individuals are randomly selected from the Pareto front.

[0145] Gaussian perturbations are performed on several elite individuals to generate local variants. The expression of local variants is:

[0146]

[0147] Among them, x variant is the local variant of the elite individual (a new solution generated near the elite individual), x elite For elite individuals, is a Gaussian distributed random disturbance.

[0148] The fitness of the local variants is evaluated to obtain a fitness value, preferably the fitness values ​​of 10 variants are calculated.

[0149] The fitness value is used to determine whether to update the population and obtain a virtual evolution strategy. The population update is to replace the worst individual if the fitness value of the best variant in the local variant is improved. The expression is:

[0150] P t+1 =P t \{arg min F(x)}∪{x best};

[0151] Among them, t is the current generation, P t+1 is the updated population, P t is the current population, arg min F(x) is the individual with the worst performance in the current population, x best is the newly generated optimal individual.

[0152] The virtual evolution strategy in this paper is a periodic, local, and refined search mechanism that breaks with the conventional generational evolution model of traditional genetic algorithms. It triggers a special virtual evolution process every fixed number of generations, performing a deep, local search for elite individuals on the current Pareto front. While conventional evolution processes focus on the overall evolution of the population, virtual evolution focuses on fine-tuning the most promising solutions.

[0153] In mathematical expression, virtual evolution can be viewed as constructing a series of local subspaces in the original search space S:

[0154] S i =x|||xx e lite||2<ε;

[0155] Among them, i is the index of the local subspace, S i is the i-th local subspace, x is a point in the search space, e is the Euclidean distance function, x e lite is the selected elite individual, and ε is the local search radius. For each elite individual, the algorithm generates n variants around it. These variants are generated through small mutations:

[0156] x variant =x elite +ΔG;

[0157] Among them, ΔG is determined by the Gaussian distribution N(0,σ l ocal) generated, where σ l ocal is the disturbance variance.

[0158] After the local variants are evaluated, only the best performing local variant has the opportunity to replace an individual in the population. The replacement rule follows: If f(x bestvariant )>f(x elite ), where f(xbestvariant ) is the fitness of the best elite variant, x bestvatiant For the best elite variant, f(x elite ) is the fitness of the current elite variant, then randomly select a position j in the population and execute P j =x bestvariant , where P j is the individual at position j in the population. The virtual evolution strategy realizes the intelligent allocation of resources and only applies the results when the virtual evolution can produce performance improvement, thus avoiding the waste of computing resources.

[0159] The virtual evolution strategy provides a "learning" mechanism that can embed local refined search in the global search process without significantly increasing computational overhead. This "two-stage" search strategy accelerates algorithm convergence and is particularly suitable for parameter-sensitive tasks, enabling the algorithm to find more accurate solutions within a limited number of iterations. Numerical experiments show that this strategy can increase the convergence speed by 30%-50%.

[0160] S307. Based on the basic environment selection mechanism, the improved environment selection mechanism is obtained by combining non-dominated sorting with optimized crowding distance calculation.

[0161] The calculation process of non-dominated sorting and optimized crowding distance in S307 is as follows:

[0162] Non-dominated sorting: Calculate the dominance count n of each individual p， , where p is an individual, and the dominated set S p , the time complexity is optimized to O(NlogN), where O is the complexity calculation function and N is the population size.

[0163] Adaptive crowding distance is the optimized crowding distance: after sorting the i-th objective function value, the intermediate individual distance is calculated as:

[0164]

[0165] Among them, dj 为 The distance between the middle individuals, j is the position index of the individual after sorting by the i-th target, is the value of the next individual after the jth individual on the i-th target, is the value of the previous individual of the jth individual on the i-th target, is the maximum value of the i-th target, is the minimum value of the i-th target. For the boundary individuals, let d = +∞, where d is the individual distance.

[0166] Hybrid selection strategy: After merging the parent and offspring populations, the new population is selected according to the following priorities: individuals with higher non-dominated ranks are given priority; individuals with larger crowding distances within the same rank are given priority.

[0167] This improved environment selection mechanism builds on the standard NSGA-II framework by introducing a more sophisticated diversity-preserving strategy, addressing the traditional NSGA-II problem of declining diversity in high-dimensional target spaces. The core of this improved environment selection mechanism is a combination of non-dominated sorting and optimized crowding distance calculation to achieve a more balanced solution distribution.

[0168] In the traditional NSGA-II, environmental selection stratifies the merged population through non-dominated sorting and uses crowding distance as a secondary sorting criterion. The improved algorithm retains this basic framework while optimizing the crowding distance calculation method.

[0169] The optimized crowding distance calculation method introduces a custom crowding distance ranking function. This is implemented through the following steps: First, the crowding distances of all individuals are initialized to 0. Then, for each objective function, individuals are sorted according to that objective, with the crowding distances of boundary individuals set to infinity. Next, the crowding distance increments of intermediate individuals are calculated. Finally, individuals are sorted in descending order of crowding distance, with the top k selected as the result. This optimized crowding distance calculation method avoids the sorting problem in the DEAP library and improves the algorithm's ability to uniformly cover the target space.

[0170] The precise selection process after the parent and child generations are merged is also optimized. The algorithm first divides the merged population into multiple frontiers F1, F2, ..., F n , where F1 is the first frontier, F2 is the second frontier, and F n The nth front is then added to the new population sequentially until the population size limit is reached. For the last front that cannot be fully included, the required number of individuals is selected using an optimized crowding distance sort. This precise selection maximizes the diversity of non-dominated solutions while maintaining the population size. It is particularly well-suited for multimodal optimization problems, enabling the simultaneous exploration of multiple optimal solution regions.

[0171] S308. Based on the basic parameter space and exploration strategy, the parameter space expansion and exploration strategy is obtained by expanding the search space boundary and adopting a logarithmic scale representation.

[0172] The parameter space expansion and exploration strategy significantly enhances the algorithm's ability to find the global optimal solution by redefining the search space boundaries and adopting a logarithmic scale. In health assessment problems, the var_smoothing parameter of the naive Bayes classifier has a significant impact on model performance, and its reasonable value range may span multiple orders of magnitude. Traditional optimization methods often restrict the search to a narrow range, such as 10^(-11) to 10^(-7). However, the improved algorithm expands the search space to 10^(-12) to 10^(-6), covering a wider range of potential optimal solutions.

[0173] This expansion is mathematically achieved through a logarithmic transformation: the actual search is performed in the range [-12, -6], while the parameter values ​​are obtained by 10^x transformation. This logarithmic representation has three key advantages: first, it makes the search space more numerically balanced, avoiding under-searching in the minimum region; second, it allows the mutation operation to produce more reasonable parameter changes, both in the large and small value regions; and finally, it simplifies the handling of parameter boundaries, making constraint enforcement more intuitive.

[0174] To handle the expanded parameter space, the algorithm adopts a multi-level exploration strategy. At the global level, the initial population is generated by uniformly sampling across the entire parameter range to ensure diversity in the search starting point. During the evolutionary process, large-scale mutation operations ensure full exploration of the entire parameter space, while small-scale mutations are responsible for local fine-grained search. This multi-scale exploration strategy enables the algorithm to focus on both global exploration and local development. A logarithmic scale transformation is used to achieve uniform sampling of the parameter space. The transformation formula is:

[0175] param_value=10log_min+x(log_max-log_min);

[0176] Among them, x∈[0,1] is the normalized sampling value.

[0177] Parameter constraint handling has also been optimized; when a mutation produces an out-of-bounds value, the algorithm maps it back to the valid range:

[0178] x'=max(lower_bound,min(upper_bound,x));

[0179] Among them, x' is the parameter value after mapping, max(·) is the maximum value function, lower_bound is the lower bound of the parameter, min(·) is the minimum value function, upper_bound is the upper bound of the parameter, and x is the original parameter value.

[0180] This soft constraint approach is more effective than simple truncation or regeneration because it preserves useful information about solutions near the boundary while ensuring that all solutions are within the valid parameter space. The parameter space expansion and exploration strategy enables the algorithm to discover optimal parameter combinations that are difficult to reach with traditional methods.

[0181] The termination condition and solution selection are preferably:

[0182] The dual termination criteria set include:

[0183] (1) Algebraic control: Maximum number of iterations Gen max =30;

[0184] (2) Convergence judgment: when the Pareto frontier improvement rate ΔF is less than 1% for five consecutive generations;

[0185] (3) Optimal solution selection: Select the solution with the largest F1 value from the final frontier:

[0186]

[0187] Among them, x best is the solution with the largest F1 value, and PF is the Pareto front.

[0188] S309. Using a meta-learning algorithm, the adaptive mutation mechanism, virtual evolution strategy, improved environmental selection mechanism, dual-objective optimization strategy, and parameter space and exploration strategy are trained and tested in sequence based on the training set and the test set until the trained adaptive mutation mechanism, virtual evolution strategy, improved environmental selection mechanism, dual-objective optimization strategy, and parameter space and exploration strategy meet the set requirements. The trained adaptive mutation mechanism, virtual evolution strategy, improved environmental selection mechanism, dual-objective optimization strategy, and parameter space and exploration strategy are connected as an improved multi-objective genetic algorithm optimization model.

[0189] A UAV subsystem health assessment method also requires system deployment and performance evaluation, as follows:

[0190] System deployment involves embedded system integration, specifically deploying the trained improved multi-objective genetic algorithm optimization model to the UAV flight control computer. The optimization strategies include:

[0191] (1) Model lightweighting: solidify the Gaussian Naive Bayes model parameters into a lookup table.

[0192] (2) Real-time guarantee: Feature processing takes 1.2±0.3ms; classification reasoning takes 0.8±0.2ms.

[0193] (3) Memory usage optimization: Model parameter storage is 12.6KB; runtime memory is <8MB.

[0194] The experimental design and comparative analysis for performance evaluation are as follows:

[0195] Four sets of experiments were constructed: grid search, Bayesian optimization, standard NSGA-II and improved multi-objective genetic algorithm (improved MOGA) as shown in Table 2.

[0196] Table 2 Experiments on grid search, Bayesian optimization, standard NSGA-II and improved multi-objective genetic algorithm (improved MOGA)

[0197]

[0198]

[0199] The performance indicators (accuracy, F1 value, false positive rate, convergence number and training time) of the four groups of experiments are compared in Table 3.

[0200] Table 3 Comparison of performance indicators (accuracy, F1 value, false positive rate, convergence number and training time) of the four groups of experiments

[0201]

[0202] The results in Table 3 show that the improved MOGA algorithm performs best in all evaluation indicators. Compared with the traditional method, its accuracy is improved to 84.86%, the F1 value reaches 0.812, and the misjudgment rate is reduced to the lowest 0.009%. At the same time, the convergence speed is significantly accelerated (only 19 generations can converge, 9 generations less than the standard NSGA-II), and the training time is the shortest (33.2 seconds).

[0203] The results show that the improved MOGA algorithm successfully balances classification performance and misjudgment control through multi-objective optimization strategy and virtual evolution mechanism, which not only improves model accuracy but also reduces computational overhead. It has obvious advantages in applications such as drone health assessment that have high requirements for accuracy and efficiency.

[0204] Comparison of the convergence curves of the standard NSGA-II and improved MOGA evolution processes, such as Figure 6 shown.

[0205] Depend on Figure 6 As can be seen, the improved MOGA algorithm significantly outperforms the standard NSGA-II algorithm in both maximum and average F1 values. The improved MOGA algorithm exhibits a step-by-step increase, consistent with the periodic triggering mechanism of its virtual evolution strategy, with a significant performance jump at specific generations. Around generation 60, the improved MOGA's F1 value reaches a plateau, ultimately converging to a high value of approximately 0.814. In contrast, the standard NSGA-II algorithm exhibits a more gradual increase in F1 value, a slower convergence rate, and a lower final convergence value (approximately 0.809).

[0206] The present invention provides a UAV subsystem health assessment system, comprising:

[0207] The acquisition module is used to obtain multi-source sensor data from the drone. The multi-source sensor data includes normal flight data, gyroscope fault data, accelerometer fault data, and GPS fault data.

[0208] The extraction module is used to filter out features that are strongly correlated with the fault state based on the Spearman rank correlation coefficient based on the multi-source sensor data of the drone to obtain key monitoring parameters. Key monitoring parameters include: roll angular velocity, pitch angular velocity, yaw angular velocity, and voltage parameters.

[0209] The model building module is used to construct an improved multi-objective genetic algorithm optimization model based on the dual-objective optimization strategy, adaptive mutation mechanism, virtual evolution strategy, improved environment selection mechanism and parameter space and exploration strategy.

[0210] The evaluation module is used to input key monitoring parameters into the improved multi-objective genetic algorithm optimization model to obtain a visual report of the health status assessment results and confidence rating.

[0211] Finally, it should be noted that the above embodiments are only used to illustrate the technical method of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical method of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical method to deviate from the spirit and scope of the technical method of the present invention.

Claims

1. A method for evaluating the health of a drone subsystem, characterized in that: The following steps are involved: S1, obtain multi-source sensor data of UAV; The multi-source sensor data of the UAV includes: normal flight data, gyroscope fault data, accelerometer fault data and GPS fault data; S2. Based on the multi-source sensor data of the UAV, the Spearman rank correlation coefficient is used to screen the features that are strongly correlated with the fault state to obtain key monitoring parameters; the key monitoring parameters include: roll angular velocity, pitch angular velocity, yaw angular velocity and voltage parameters; S3. Construct an improved multi-objective genetic algorithm optimization model based on the dual-objective optimization strategy, adaptive mutation mechanism, virtual evolution strategy, improved environment selection mechanism, and parameter space and exploration strategy; S4. Input the key monitoring parameters into the improved multi-objective genetic algorithm optimization model to obtain a visual report of the health status assessment results and confidence rating.

2. A method for evaluating the health of a drone subsystem according to claim 1, characterized in that: In S3, an improved multi-objective genetic algorithm optimization model is constructed based on the dual-objective optimization strategy, adaptive mutation mechanism, virtual evolution strategy, improved environment selection mechanism, parameter space and exploration strategy. Specifically, it includes: S301, obtaining historical visualization reports of multi-source sensor data, health status historical assessment results, and confidence ratings of drones; S302. Based on the historical data of the multi-source sensors of the UAV, the Spearman rank correlation coefficient is used to screen the features that are strongly correlated with the fault status to obtain the key monitoring historical parameters; S303, constructing a training set and a test set based on historical visualization reports of key monitoring historical parameters, historical health status assessment results, and confidence ratings; S304, based on the basic objective optimization strategy, a dual-objective optimization strategy is obtained by defining the fitness function as a dual-objective vector; S305. Based on the basic adaptive mechanism, an adaptive mutation mechanism is obtained by dynamically adjusting the mutation probability and the mutation intensity; S306. Based on the basic genetic algorithm, a virtual evolution strategy is obtained by embedding a local fine search in the global search process; S307. Based on the basic environment selection mechanism, an improved environment selection mechanism is obtained by combining non-dominated sorting with optimized crowding distance calculation; S308. Based on the basic parameter space and exploration strategy, a parameter space expansion and exploration strategy is obtained by expanding the search space boundary and adopting a logarithmic scale representation; S309. Using a meta-learning algorithm, the adaptive mutation mechanism, virtual evolution strategy, improved environmental selection mechanism, dual-objective optimization strategy, and parameter space and exploration strategy are trained and tested in sequence based on the training set and the test set until the trained adaptive mutation mechanism, virtual evolution strategy, improved environmental selection mechanism, dual-objective optimization strategy, and parameter space and exploration strategy meet the set requirements. The trained adaptive mutation mechanism, virtual evolution strategy, improved environmental selection mechanism, dual-objective optimization strategy, and parameter space and exploration strategy are connected as an improved multi-objective genetic algorithm optimization model.

3. A method for evaluating the health of a drone subsystem according to claim 2, characterized in that: In S305, based on the original adaptive mechanism, the adaptive mutation mechanism is obtained by dynamically adjusting the mutation rate and mutation intensity, which specifically includes: The adaptive mutation rate is calculated based on the basic mutation rate and the diversity factor; The adaptive mutation strength is calculated based on the basic mutation strength, diversity factor and boundary perception factor; Based on the original adaptive mechanism, the adaptive mutation mechanism is obtained according to the adaptive mutation rate and adaptive mutation strength.

4. A method for evaluating the health of a drone subsystem according to claim 3, characterized in that: The expression of the diversity factor is: Among them, S i is the number of individuals in the i-th frontier, k is the number of non-dominated frontiers, N pop is the initial population size; The expression of the adaptive mutation rate is: Among them, P m is the basal mutation rate, is the adaptive mutation rate.

5. The method for evaluating the health of a drone subsystem according to claim 3, wherein: The boundary perception factor E f The expression is: Among them, E f is the boundary perception factor, x is the relative position of the current solution within the feasible region of its corresponding decision variable (normalized value); The expression of the adaptive mutation strength is: s adaptive =σ×(1-0.5D f )×E f ; Among them, σ adaptive is the adaptive mutation strength, and σ is the basic mutation strength.

6. A method for evaluating the health of a drone subsystem according to claim 2, characterized in that: In S306, based on the basic genetic algorithm, a virtual evolution strategy is obtained by embedding a local fine search in the global search process, which specifically includes: Get the fixed number of generations that perform the deep search of elite individuals; Within a fixed generation, a number of elite individuals are randomly selected from the Pareto frontier; Perform Gaussian perturbations on several elite individuals to generate local variants; Perform fitness evaluation on the local variant to obtain the fitness value; Whether to perform population update is determined according to the fitness value to obtain a virtual evolution strategy; the population update is to replace the worst individual if the fitness value of the best variant in the local variants is improved.

7. A method for evaluating the health of a drone subsystem according to claim 2, characterized in that: In S307, based on the basic environment selection mechanism, the improved environment selection mechanism is obtained by combining non-dominated sorting and optimized crowding distance calculation, which specifically includes: Based on the basic environment selection mechanism, the merged population is divided into several frontiers through non-dominated sorting, and several frontiers are added to the new population until the population size limit is reached; If the last frontier cannot be added to the new population, the optimized crowding distance sorting is used to determine the required number of frontiers, resulting in an improved environment selection mechanism.

8. A method for evaluating the health of a drone subsystem according to claim 7, characterized in that: If the last frontier cannot be added to the new population, the optimized crowding distance sorting is used to determine the required number of frontiers, and the specific content of the improved environment selection mechanism is: If the last frontier cannot be added to the new population, the crowding distance of all frontiers is initialized to 0; For each objective function, the frontier is sorted by the objective function's target and the crowding distance of the frontier is set to infinity; The crowding distance increment of the intermediate frontier is calculated, and finally the frontiers are sorted in descending order of crowding distance to determine the required number of frontiers, thus obtaining an improved environment selection mechanism.

9. The method for evaluating the health of a drone subsystem according to claim 2, wherein: In S308, based on the basic objective optimization strategy, a dual-objective optimization strategy is obtained by defining the fitness function as a dual-objective vector, which specifically includes: Based on the basic objective optimization strategy, a dual-objective optimization strategy is obtained by simultaneously optimizing two mutually constrained objectives; the two mutually constrained objectives include: maximizing the F1 score and minimizing the false positive rate.

10. A drone subsystem health assessment system, characterized in that: include: Acquisition module, used to obtain multi-source sensor data of UAV; The multi-source sensor data of the UAV includes: normal flight data, gyroscope fault data, accelerometer fault data and GPS fault data; An extraction module is used to filter features strongly correlated with the fault state based on the multi-source sensor data of the UAV using the Spearman rank correlation coefficient to obtain key monitoring parameters; the key monitoring parameters include: roll angular velocity, pitch angular velocity, yaw angular velocity and voltage parameters; Model building module, used to build an improved multi-objective genetic algorithm optimization model based on dual-objective optimization strategy, adaptive mutation mechanism, virtual evolution strategy, improved environment selection mechanism and parameter space and exploration strategy; The evaluation module is used to input key monitoring parameters into the improved multi-objective genetic algorithm optimization model to obtain a visual report of the health status assessment results and confidence rating.

Images