Evolutionary learning-based collaborative optimization method for determining aircraft maintenance timing

Abstract

This invention discloses an evolutionary learning-based collaborative optimization method for determining aircraft maintenance timing, belonging to the field of equipment health management and data processing. The invention includes: establishing an optimal maintenance timing prediction model, which outputs the optimal maintenance timing based on monitoring data; acquiring monitoring data for various indicators, performing sensitivity analysis on key model parameters, and extracting sensitive factors as prior knowledge and storing them in a knowledge base; constructing an evolutionary learning collaborative optimization algorithm to optimize the parameter vectors at each time point, with an exploration engine responsible for accumulating successful experiences and using sensitive factors to correct the optimization direction, and a learning engine learning the mapping relationship between parameter positions and optimization directions from the corrected experiences to obtain an optimized learning engine; and predicting the optimal maintenance timing for the test time based on the optimized learning engine. This invention realizes the transformation of parameter optimization from blind search to knowledge-guided optimization, effectively improving the accuracy and efficiency of maintenance timing determination.

Description

Technical Field

[0001] This invention relates to the field of equipment health management and data processing technology, specifically to an evolutionary learning-based collaborative optimization method for determining aircraft maintenance timing. Background Technology

[0002] Aircraft are characterized by complex system configurations, high timeliness of maintenance, and extremely high mission reliability requirements. Throughout the entire lifecycle of an aircraft, its various subsystems are subjected to complex loads over extended periods, inevitably leading to performance degradation in critical components. Therefore, determining the optimal maintenance timing for an aircraft is a core technical challenge in balancing operational reliability and maintenance costs, and in extending the aircraft's service life.

[0003] Currently, maintenance decisions for complex equipment primarily employ two main strategies in engineering practice: reliability-based scheduled maintenance and condition-based condition monitoring-based maintenance. Scheduled maintenance involves planned repairs based on fixed time cycles or operating durations. However, this strategy ignores the dynamic differences in the individual health status of aircraft. In harsh actual operating conditions, insufficient maintenance may lead to sudden failures, or overly conservative cycles may result in wasted resources. In contrast, condition-based maintenance determines maintenance timing by monitoring system status in real time, making it a more economical and reasonable maintenance strategy. However, aircraft operating environments are typically extremely harsh and their conditions highly variable. Monitoring data is easily affected by external environmental interference such as noise. This presents a significant challenge to accurately assessing system health status and determining the optimal maintenance timing, primarily in the following three aspects:

[0004] First, aircraft consist of numerous highly coupled subsystems with complex operating mechanisms. The relationships between various monitoring indicators and the overall system performance are often strongly nonlinear, making it difficult to establish accurate analytical mathematical models. Simultaneously, with continuous improvements in aircraft design and manufacturing processes, their reliability has significantly increased, resulting in extremely sparse abnormal state samples in the monitoring data. Traditional data-driven methods such as deep belief networks and support vector machines struggle to train effective predictive models due to a lack of sufficient fault samples. While knowledge-based methods such as expert systems and fuzzy reasoning can incorporate domain experience, their static rule bases are ill-suited to dynamically changing operating conditions.

[0005] Secondly, multi-source information fusion is insufficient under noise interference. Aircraft health monitoring typically involves information from multiple heterogeneous sensors, and the uncertainty of this information is more significant in noisy environments. Existing semi-quantitative information fusion methods based on Evidential Reasoning (ER) and Belief Rule Base (BRB) mostly ignore the potential correlations between different monitoring indicators during the fusion process, simply assuming that each piece of evidence is independent. This is inconsistent with the actual operating mechanism of a strongly coupled aircraft system, leading to biases in the information fusion results.

[0006] Third, under harsh operating conditions, the noise-affected monitoring data is coupled with the complex system mechanisms, making it difficult to characterize the true distribution space of model parameters. Most existing methods employ a single evolutionary algorithm for iterative optimization, repeatedly utilizing limited data for optimization. This lack of in-depth analysis of data characteristics and model mechanisms makes it difficult to form a universal and scientific understanding of successful optimization behavior, resulting in slow model convergence and a high likelihood of getting trapped in local optima, thus reducing the reliability of the final maintenance timing decision.

[0007] Therefore, how to accurately assess the health status of aircraft and determine the optimal maintenance timing under complex and harsh operating conditions is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention provides an evolutionary learning-based collaborative optimization method for determining aircraft maintenance timing. This method constructs a health status assessment model based on evidence-based reasoning, integrating multi-source monitoring data and expert knowledge. Sensitivity analysis is used to extract prior knowledge of the model's mechanisms to guide parameter optimization. An evolutionary learning collaborative optimization algorithm is employed to iteratively optimize the model parameters, ensuring that the determined maintenance timing closely aligns with the actual health degradation trend of the equipment. This approach effectively reduces maintenance costs while maintaining the reliability of maintenance decisions.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] This invention proposes a method for determining aircraft maintenance timing based on evolutionary learning and collaborative optimization, comprising the following steps:

[0011] S1. Establish an optimal maintenance timing prediction model; the optimal maintenance timing prediction model is used to convert the monitoring data of each monitoring indicator of the aircraft into evidence distribution, fuse the evidence distribution corresponding to each monitoring indicator based on relevant evidence reasoning rules to obtain the confidence distribution of the aircraft health status, calculate the health status utility value according to the confidence distribution, and output the optimal maintenance timing prediction value based on the mapping relationship between the health status utility value and the optimal maintenance timing.

[0012] S2. Acquire monitoring data of various monitoring indicators of the aircraft; based on the monitoring data of various monitoring indicators, perform sensitivity analysis on the key parameters of the optimal maintenance timing prediction model, calculate the sensitivity factor of each key parameter to the optimal maintenance timing, and store the sensitivity factor in the knowledge base.

[0013] S3. Construct an evolutionary learning co-optimization algorithm consisting of an exploration engine, a learning engine, an experience base, and a knowledge base. Each key parameter at each time step forms a parameter vector. The evolutionary learning co-optimization algorithm is used to optimize the parameter vector at each time step. Specific methods include:

[0014] S301. For each time step, a random number label is generated. If the random number label is greater than the preset learning rate, the exploration engine is used to perform a global search optimization on the current parameter vector to obtain the first updated parameter vector. If the random number label is less than or equal to the preset learning rate, the current parameter vector is used as input to predict the second optimization direction using the current learning engine, and the current parameter vector is updated according to the second optimization direction to obtain the second updated parameter vector.

[0015] For each moment when the exploration engine performs optimization, if the fitness corresponding to the first updated parameter vector is better than the fitness corresponding to the current parameter vector, then the current parameter vector and the corresponding first optimization direction are stored in the experience base, wherein the first optimization direction is the difference between the first updated parameter vector and the current parameter vector; the fitness is the squared error between the predicted value of the optimal maintenance time and the reference value of the optimal maintenance time.

[0016] S302. Determine if the experience base is empty. If it is not empty, proceed to S303; if it is empty, proceed to S304.

[0017] S303. Use the sensitivity factors in the knowledge base to correct the first optimization direction in the experience base to obtain the corrected optimization direction; use the current parameter vector in the experience base as input and the corresponding corrected optimization direction as label to train the current learning engine to obtain the trained learning engine; after training, clear the experience base.

[0018] S304. Determine whether the current iteration round has reached the preset number of iterations. If not, use the parameter vector updated at each time step of the current iteration round as the current parameter vector for the next iteration round and return to S301. If it has reached the preset number of iterations, stop the iteration and output the learning engine trained in the current iteration round as the optimized learning engine.

[0019] S4. For the time to be measured, calculate the initial parameter vector based on the corresponding monitoring data, input the initial parameter vector into the optimized learning engine, output the predicted optimization direction, update the initial parameter vector according to the predicted optimization direction, obtain the optimized parameter vector, substitute the optimized parameter vector into the optimal maintenance timing prediction model, input the monitoring data into the optimal maintenance timing prediction model, and output the optimal maintenance timing for the time to be measured.

[0020] Furthermore, in S1, the triangular membership transformation method is used to transform the monitoring data of each monitoring indicator into an evidence distribution:

[0021]

[0022] In the formula, The first step in the health status of the aircraft Each assessment level ; The total number of assessment levels for the health status of the aircraft; For the first Each monitoring indicator at time belong Confidence level, For the first Each monitoring indicator at time Distribution of evidence.

[0023] Furthermore, in S1, the process of fusing the evidence distributions corresponding to each monitoring indicator includes:

[0024] The evidence distribution is transformed into a weighted confidence distribution with reliability, and the basic probability mass of each evidence distribution is calculated:

[0025]

[0026]

[0027] In the formula, For the first Each monitoring indicator at time Evidence distribution assigned The basic probability mass; The first step in the health status of the aircraft Each assessment level ; The total number of assessment levels for the health status of the aircraft; For the first Each monitoring indicator at time The correlation discount factor For the first Each monitoring indicator at time Weight of evidence For the first Each monitoring indicator at time Weighted evidence weights, For the first Each monitoring indicator at time belong Confidence level, A set of assessment levels for the health status of an aircraft; express The power set; For the first Each monitoring indicator at time The reliability of the evidence;

[0028] The basic probability quality of each evidence distribution is iteratively fused using evidence reasoning rules to obtain the confidence level of each health status assessment level after fusion. Based on the confidence levels of each health status assessment level after fusion, the health status confidence distribution is obtained.

[0029]

[0030] In the formula, For the aircraft at all times The confidence distribution of health status. For a moment The health status of the aircraft belongs to The confidence level.

[0031] Furthermore, in S2, the key parameters include three categories: evidence weight, evidence reliability, and relevance discount factor; the sensitivity factor is obtained by taking the partial derivative of the initial output of the optimal maintenance timing prediction model with respect to the key parameters.

[0032] Furthermore, in S3, the exploration engine uses a particle evolution algorithm to perform a global search optimization on the current parameter vector; the current learning engine uses a feedforward neural network to predict the second optimization direction.

[0033] Furthermore, in S303, the method for correcting the first optimization direction in the experience base is as follows:

[0034] Calculate the correction factors for each key parameter separately:

[0035]

[0036] In the formula, For the first Each monitoring indicator at time The Correction factors for class key parameters, For the first Each monitoring indicator at time The The amount of change corresponding to the key parameters of the class. The total number of categories for key parameters;

[0037] The first optimization direction is modified by adjusting the correction factor and the sensitivity factor in the knowledge base:

[0038]

[0039] In the formula, For the first Each monitoring indicator at time The The corrected changes in the class of key parameters. For a moment The fitness of the current parameter vector. For the first Each monitoring indicator at time The Sensitivity factors for class-critical parameters.

[0040] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0041] (1) This invention introduces a relevance discount factor into the evidence reasoning rules. By quantifying the degree of correlation between indicators through the distance correlation coefficient, the basic probability quality of the evidence is discounted, so that the fusion result more accurately reflects the true health status of the aircraft. At the same time, this invention uses sensitivity analysis to extract prior knowledge of the model mechanism and uses the prior knowledge to guide the subsequent parameter optimization process, ensuring that the key parameters of the model conform to engineering reality and providing a reliable data foundation for further determining the maintenance timing.

[0042] (2) This invention constructs an evolutionary learning collaborative optimization framework consisting of an exploration engine, a learning engine, an experience base, and a knowledge base. The exploration engine is responsible for global searching and accumulating successful experiences. The knowledge base uses prior knowledge to correct the optimization direction in successful experiences, ensuring it conforms to the model mechanism. The learning engine learns the mapping relationship between parameter positions and optimization directions from the corrected experiences, enabling the successful experiences accumulated in each iteration to be quantitatively extracted into parameter knowledge for the learning engine. The trained learning engine is then used to predict optimization directions in subsequent iterations, achieving a quantitative transformation of optimization experience into learning engine parameters. This solidifies successful experiences through neural network weights. The collaborative mechanism between the exploration engine and the learning engine in this invention avoids the inherent defects of traditional single evolutionary algorithms—such as reusing limited data and lacking experience accumulation—from an algorithmic perspective, realizing a shift in the parameter optimization process from blind search to knowledge-guided optimization.

[0043] (3) This invention obtains sensitivity factors through sensitivity analysis and stores these sensitivity factors as prior knowledge in a knowledge base. In each iteration, the sensitivity factors are used to quantitatively correct the optimization direction generated by the exploration engine. During the correction process, parameters with larger sensitivity factors have a greater impact on the model output per unit change, and the correction range is smaller, to prevent over-adjustment of key parameters from causing drastic fluctuations in output; parameters with smaller sensitivity factors allow for relatively larger adjustment ranges to improve optimization efficiency. This mechanism ensures the stability of the optimization process, so that the optimized parameter vector has both high model accuracy and conforms to the physical meaning constraints of the parameters themselves, avoiding the problem of optimization results deviating from engineering reality.

[0044] (4) This invention starts with evidence fusion from multi-source monitoring data, processes the correlation of indicators through a correlation discount factor, mines prior knowledge of the model through sensitivity analysis, and achieves efficient and accurate parameter optimization through an evolutionary learning collaborative optimization algorithm, ultimately establishing an optimal maintenance timing prediction model. The conclusions of sensitivity analysis guide the direction correction in the optimization process, the efficient optimization of the collaborative optimization algorithm ensures the generalization ability of the model, and the optimal maintenance timing output by the optimal maintenance timing prediction model can provide a quantitative reference for the formulation of engineering maintenance plans. Attached Figure Description

[0045] Figure 1 This is an overall flowchart of the aircraft maintenance timing determination method based on evolutionary learning collaborative optimization of the present invention.

[0046] Figure 2 This is a graph showing the monitoring data of vibration amplitude;

[0047] Figure 3 A graph showing the monitoring data of vibration acceleration;

[0048] Figure 4 A graph showing monitoring data of ground tilt angle;

[0049] Figure 5 A comparison chart of the initial maintenance timing and the optimal maintenance timing reference values;

[0050] Figure 6 This is a comparison chart of the optimized maintenance timing and the optimal maintenance timing reference value. Detailed Implementation

[0051] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] Example

[0053] refer to Figure 1 This embodiment provides a method for determining aircraft maintenance timing through evolutionary learning-based collaborative optimization, which is implemented according to the following steps:

[0054] S1. Establish an optimal maintenance timing prediction model. The optimal maintenance timing prediction model is used to output the optimal maintenance timing based on the monitoring data of multiple monitoring indicators.

[0055] Because the monitoring data of various health indicators have different dimensions and units, the optimal maintenance timing prediction model first unifies the expression of each health indicator, transforming the monitoring data of each indicator into an evidence distribution:

[0056]

[0057] In the formula, The first step in the health status of the aircraft Each assessment level ; The total number of assessment levels for the health status of the aircraft; For the first Each monitoring indicator at time belong Confidence level, , , For the first Each monitoring indicator at time Distribution of evidence.

[0058] in The calculation formula is:

[0059]

[0060] In the formula, for Reference value, For the first Each monitoring indicator at time The monitoring data.

[0061] Based on relevant evidence reasoning rules, the evidence distribution of each monitoring indicator is fused to obtain the confidence distribution of the aircraft's health status:

[0062] The process of fusing the evidence distributions corresponding to each monitoring indicator specifically includes:

[0063] The evidence distribution is transformed into a weighted confidence distribution with reliability, and the basic probability mass of each evidence distribution is calculated:

[0064]

[0065]

[0066] In the formula, For the first Each monitoring indicator at time Evidence distribution assigned The basic probability mass; The first step in the health status of the aircraft Each assessment level ; The total number of assessment levels for the health status of the aircraft; For the first Each monitoring indicator at time The correlation discount factor For the first Each monitoring indicator at time Weight of evidence For the first Each monitoring indicator at time Weighted evidence weights, For the first Each monitoring indicator at time belong Confidence level, A set of assessment levels for the health status of an aircraft; express The power set; For the first Each monitoring indicator at time The reliability of the evidence;

[0067] Evidence-based reasoning rules are used to iteratively fuse the basic probability qualities of various monitoring indicators at the same time. The basic probability quality of the first monitoring indicator is used as the initial value for fusion. The basic probability qualities of two monitoring indicators are fused at a time, until the basic probability qualities of all monitoring indicators are fused. The formula for iterative fusion is:

[0068]

[0069]

[0070] In the formula, For the front Each monitoring indicator at time After iterative fusion of the basic probability mass, it is assigned to Part of ; B is the first Each monitoring indicator corresponds to a focal point, where C is the leading [value]. The fusion of monitoring indicators into a single element For the first Each monitoring indicator at time The portion of the basic probability mass assigned to B. For the front Each monitoring indicator at time The portion of the basic probability mass assigned to C after iterative fusion. For the first Each monitoring indicator at time Basic probability mass assigned Part of For the front Each monitoring indicator at time After iterative fusion of the basic probability mass, it is assigned to Part of For the front Each monitoring indicator at time After iterative fusion of the basic probability mass, it is assigned to Part of For the first Each monitoring indicator at time Basic probability mass assigned The part.

[0071] After integrating the basic probability quality of all monitoring indicators, the confidence level calculation formula for each assessment level is as follows:

[0072]

[0073] For a moment The health status of the aircraft belongs to The confidence level.

[0074] The confidence distribution of health status is as follows:

[0075]

[0076] In the formula, For the aircraft at all times The confidence distribution of health status.

[0077] Calculate the aircraft's health status utility value based on the confidence distribution of health status:

[0078]

[0079] For a moment The health status utility value of the aircraft. for The corresponding utility value is usually determined based on national standards or industrial statistics.

[0080] Based on the mapping relationship between health status utility value and optimal maintenance timing, the predicted value of optimal maintenance timing is output. In this embodiment, maintenance timing refers to the time interval between the acquisition of monitoring data and the recommended downtime for maintenance, in days. Maintenance timing reflects the urgency of maintaining the aircraft; the smaller the value, the worse the overall health level of the aircraft, and the higher the urgency of maintenance.

[0081] The mapping relationship between health status utility value and optimal maintenance timing is as follows:

[0082]

[0083] in, Indicates at time The optimal maintenance time; This indicates the maximum maintenance interval for the aircraft, which is set according to the equipment's factory parameters. This is the decay rate coefficient, used to control the rate at which the health status utility value decays as it maps to the maintenance opportunity. This is an exponential adjustment coefficient used to adjust the sensitivity of health status utility value to the timing of maintenance.

[0084] S2. Acquire monitoring data for various monitoring indicators of the aircraft. Based on the monitoring data, perform sensitivity analysis on the key parameters of the optimal maintenance timing prediction model. For the evidence reasoning rule, the key parameters include evidence weight, evidence reliability, and relevance discount factor. Therefore, this embodiment performs sensitivity analysis on evidence weight, evidence reliability, and relevance discount factor respectively.

[0085] Calculate the partial derivatives of the initial output of the optimal maintenance timing prediction model with respect to the evidence weight, evidence reliability, and relevance discount factor, respectively. Use the absolute values ​​of these partial derivatives as the sensitivity factors for the corresponding key parameters. Taking the evidence weight as an example, the formula for calculating its sensitivity factor is as follows:

[0086]

[0087] In the formula, For the first Each monitoring indicator at time Evidence weighting sensitivity factor For the first Each monitoring indicator at time Weight of evidence For a moment The corresponding initial optimal maintenance timing will be at time Substitute the initial parameter vector into the optimal maintenance timing prediction model, and then input each monitoring index at time... The monitoring data is input into the optimal maintenance timing prediction model, and the output is... .

[0088] time The initial parameter vector is calculated using the monitoring data of each monitoring indicator at that moment. The evidence weight can be calculated from the monitoring data using the coefficient of variation method or the entropy weight method, the evidence reliability can be calculated using the perturbation coefficient method, and the correlation discount factor can be calculated using the distance correlation coefficient. The calculation results are then normalized to obtain the initial parameter vector.

[0089] The sensitivity factors of each key parameter obtained are stored in the knowledge base as prior knowledge.

[0090] Sensitivity factors can quantitatively reflect the degree of influence of changes in key parameters on the optimal maintenance timing. That is, the larger the sensitivity factor, the greater the influence of the corresponding key parameter on the output of the optimal maintenance timing prediction model.

[0091] S3. The key parameters at each time step form a parameter vector, and each key parameter satisfies the following constraints:

[0092]

[0093] In the formula, No. Each monitoring indicator at time Weight of evidence For the first Each monitoring indicator at time The reliability of the evidence For the first Each monitoring indicator at time The correlation discount factor.

[0094] An evolutionary learning co-optimization algorithm is constructed, consisting of an exploration engine, a learning engine, an experience base, and a knowledge base. This algorithm optimizes the parameter vector at each time step; that is, it updates the parameter vector at each time step in each iteration.

[0095] In this embodiment, an exploration engine is used to perform a global search of the model's key parameters, accumulating successful optimization experience. Sensitive factors stored in a knowledge base are used to correct the optimization direction of this successful experience, ensuring that the optimization behavior conforms to the model's mechanism. A learning engine is then used to learn from the corrected experience, establishing a mapping relationship between parameter positions and optimization directions. Through the collaborative work of the exploration and learning engines, efficient and accurate parameter optimization is achieved. The exploration engine is a particle evolution algorithm, and the learning engine is a feedforward neural network.

[0096] The parameter vector is represented as:

[0097]

[0098]

[0099]

[0100]

[0101] In the formula, For a moment The parameter vector, For all monitoring indicators at time The set of evidence weights For all monitoring indicators at time The set of evidence reliability, For all monitoring indicators at time The set of correlation discount factors.

[0102] The process of optimizing the parameter vector includes:

[0103] S301. Initialize the population size for the particle evolution algorithm. Inertia weight Individual cognitive coefficient Group cognition coefficient In this embodiment, the population size... =30, inertia weight =0.7, Individual Cognitive Coefficient =1.5, Group Cognition Coefficient =1.5.

[0104] Initialize the structural parameters of the feedforward neural network. In this embodiment, the input dimension of the feedforward neural network is... and output dimensions All ,in The total number of assessment levels for the aircraft's health status. This represents the total number of monitoring indicators; the number of hidden layers is set to 2, and the number of hidden units in each hidden layer is 16.

[0105] Initialize the collaborative control parameters of the evolutionary learning collaborative optimization algorithm, including the preset learning rate. and preset number of iterations In this embodiment, a preset learning rate is used. =0.2, used to control the proportion of the exploration engine and the learning engine playing a role in the optimization process; preset number of iterations. =10, used to control the number of training rounds for the evolutionary learning collaborative optimization algorithm.

[0106] S302. For each time t, generate a random number label. Based on the random number label at that moment. The comparison with the preset learning rate determines whether to use the exploration engine or the learning engine for the current parameter vector at that moment. Optimize:

[0107] like Then the optimization is performed by the exploration engine. The exploration engine optimizes the current parameter vector. A global search optimization is performed, with the goal of minimizing fitness, to obtain the first updated parameter vector. .

[0108] Calculate the first update parameter vector Corresponding fitness and the pre-parameter vector Corresponding fitness .

[0109] The calculation formula is:

[0110]

[0111] In the formula, For a moment The corresponding optimal maintenance timing reference value is provided by experts and serves as a benchmark for measuring model accuracy. for The corresponding optimal maintenance time will Substitute the time into the optimal maintenance timing prediction model, and... Input the monitoring data into the optimal maintenance timing prediction model, and it will output... .

[0112] The calculation formula is:

[0113]

[0114] In the formula, for The corresponding optimal maintenance time will Substitute the time into the optimal maintenance timing prediction model, and... Input the monitoring data into the optimal maintenance timing prediction model, and it will output... .

[0115] like Less than Then the current parameter vector With the first optimization direction As a successful experience, it is stored in the experience base. Since fitness is the mean squared error, a smaller value is better. The first optimization direction... Defined as the difference between the first updated parameter vector and the current parameter vector, i.e. Its physical meaning is the direction and magnitude of the parameter vector's movement from the current position to the updated position, representing the optimization direction found by the exploration engine in this iteration.

[0116] like Then the learning engine will perform optimization. Using the current parameter vector... Using the current learning engine as input, predict the second optimization direction. And according to the second optimization direction Update the current parameter vector to obtain the second updated parameter vector. ,Right now .

[0117] After the parameter vectors at all times have been processed, check if the experience base is empty. If it is not empty, proceed to S303; if it is empty, proceed to S304.

[0118] S303. Use the sensitivity factors stored in the knowledge base to revise the first optimization direction of each successful experience in the experience base. The specific revision method is as follows:

[0119] Calculate the correction factors for each key parameter separately:

[0120]

[0121] In the formula, For the first Each monitoring indicator at time The The correction factor for the class of key parameters has the following physical meaning: the first The proportion of the change in a key parameter class to the total change in the three key parameters class reflects the activity level of that parameter in this iteration. Corresponding weight of evidence Corresponding to the reliability of evidence, Corresponding correlation discount factor, In this iteration, the first Each monitoring indicator at time The The change in the key parameter, i.e., the component of that parameter in the first optimization direction. This represents the total number of categories for key parameters.

[0122] The first optimization direction is corrected using the correction factor and the sensitivity factors in the knowledge base according to the following formula:

[0123]

[0124] In the formula, For the first Each monitoring indicator at time The The corrected changes for each key parameter are represented by the corrected changes for all key parameters, which together constitute the corrected optimization direction. For a moment Current parameter vector fitness For the first Each monitoring indicator at time The Sensitivity factors for key parameters. As can be seen from this formula, the greater the fitness of the current parameter vector, the greater the correction magnitude; the larger the correction factor, the greater the correction magnitude; the larger the sensitivity factor, the smaller the correction magnitude, thus preventing over-adjustment of key parameters from causing drastic output fluctuations.

[0125] After correcting the first optimization direction in the experience base, the current parameter vector of each successful experience in the experience base is... As input, the corresponding corrected optimization direction serves as the label, and the current learning engine is trained to learn the mapping relationship between parameter positions and optimization directions. After training, the experience base is cleared to prepare for the next iteration. The trained learning engine is then used for the next iteration of optimization.

[0126] S304. Determine whether the current iteration round has reached the preset number of iterations. If the target is not reached, the updated parameter vectors at each time step of the current iteration are used as the current parameter vectors for the next iteration, and the process returns to S302; if the target is reached, the iteration stops, and the learning engine trained in the current iteration is output as the optimized learning engine.

[0127] It should be noted that in the current iteration round, at each time step It may be optimized by either the exploration engine or the learning engine, so each time step corresponds to either the first updated parameter vector or the second updated parameter vector.

[0128] S4. Based on the optimized learning engine, predict the optimal maintenance time corresponding to the test time.

[0129] For any given time point to be tested, the initial parameter vector is first calculated based on the monitoring data at that time. This initial parameter vector is then input into the optimized learning engine, which outputs the optimization direction for that time point. The initial parameter vector is updated along this optimization direction to obtain the optimized parameter vector for that time point. The optimized parameter vector is then substituted into the optimal maintenance timing prediction model, and the monitoring data for that time point is also input into the model to output the optimal maintenance timing for that time point.

[0130] To better illustrate the beneficial effects of this invention, the maintenance timing determination method proposed in the above embodiments was verified using an aircraft health monitoring experimental platform. This experimental platform simulates the vibration response of the aircraft's main structure under external excitation, reproducing the health degradation characteristics of the aircraft during service. The experimental platform mainly consists of a main body and sensing devices. A base is located at the bottom of the main body, and a motor and electric cylinder are placed inside the base, capable of applying pseudo-random vibration signals to the main body and base. The sensing devices include three main types of sensors: six laser rangefinders, four impact vibration sensors, and one ground settlement sensor. The six laser rangefinders are deployed in the six quadrants of the main body's outer surface, measuring the tilt angle of the main body and calculating the vibration displacement at each measuring point. The average vibration displacement of the six measuring points is used as the vibration amplitude monitoring data. The four impact vibration sensors monitor the vibration acceleration caused by external impacts on the main body, and the root mean square of the four measured values ​​is used as the vibration acceleration monitoring data. The ground settlement sensor consists of six levels distributed on the base of the main body, measuring the tilt angle of the base (i.e., the ground tilt angle). That is, monitoring data of three monitoring indicators, namely vibration amplitude, vibration acceleration and ground tilt angle, are obtained by three types of sensors.

[0131] Continuous monitoring was conducted on the response of the main body and base under external vibration excitation. Data was collected at 640 time points. The monitoring data for each indicator are as follows: Figures 2 to 4 As shown. Monitoring data from the first 400 time points was used for training, and monitoring data from the last 240 time points was used for validation. The maximum maintenance interval time is also included. =10 days, decay rate coefficient Exponential adjustment coefficient The particle evolution algorithm parameters were initialized as follows: population size 30, inertia weight 0.7, individual cognitive coefficient 1.5, and group cognitive coefficient 1.5. The collaborative control parameters of the evolutionary learning collaborative optimization algorithm were initialized as follows: learning rate 0.2 and iteration count 10. Based on the evaluation systems in GB / T6075.1-2012 and GB / T38036-2019, the health status of the aircraft was divided into three assessment levels: healthy, sub-healthy, and unhealthy. The reference values ​​of each monitoring indicator at each assessment level are shown in Table 1.

[0132] Table 1. Reference values ​​for the assessment levels of each monitoring indicator

[0133]

[0134] The utility values ​​for each assessment level of the aircraft's health status are shown in Table 2.

[0135] Table 2 Utility values ​​for each assessment level

[0136]

[0137] For any one of the subsequent 240 time points, calculate the corresponding initial parameter vector based on the monitoring data at that time point. Substitute the initial parameter vector into the optimal maintenance timing prediction model, and input the monitoring data at that time point into the optimal maintenance timing prediction model. Output the optimal maintenance timing, and use the optimal maintenance timing corresponding to the initial parameter vector as the initial maintenance timing. Figure 5 The initial maintenance timing and optimal maintenance timing reference values ​​are displayed at various time points. The root mean square error between the initial maintenance timing and the optimal maintenance timing reference value is calculated to be 0.708.

[0138] For any one of the subsequent 240 time points, the initial parameter vector for that time point is input into the optimized learning engine, which outputs an optimization direction. The initial parameter vector is then updated along this optimization direction to obtain the optimized parameter vector for that time point. This optimized parameter vector is then substituted into the optimal maintenance timing prediction model, and the monitoring data for that time point is also input into the optimal maintenance timing prediction model to output the optimal maintenance timing. The optimal maintenance timing corresponding to the optimized parameter vector is then taken as the optimized maintenance timing. Figure 6 The optimized maintenance timing and the reference value for the optimal maintenance timing are shown at various time points. The root mean square error between the optimized maintenance timing and the reference value for the optimal maintenance timing is calculated to be 0.234. This indicates that, compared to the initial parameter vector, substituting the optimized parameter vector into the optimal maintenance timing prediction model results in a more accurate output of the optimal maintenance timing.

[0139] The specific embodiments of the present invention are provided to enable those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention.

[0140] It should be understood that the present invention is not limited to the content already described above, and various modifications and changes can be made without departing from its scope. The scope of the present invention is limited only by the appended claims.

Claims

1. A method for determining aircraft maintenance timing through evolutionary learning and collaborative optimization, characterized in that, Includes the following steps: S1. Establish an optimal maintenance timing prediction model; the optimal maintenance timing prediction model is used to convert the monitoring data of each monitoring indicator of the aircraft into evidence distribution, fuse the evidence distribution corresponding to each monitoring indicator based on relevant evidence reasoning rules to obtain the confidence distribution of the aircraft health status, calculate the health status utility value according to the confidence distribution, and output the optimal maintenance timing prediction value based on the mapping relationship between the health status utility value and the optimal maintenance timing. S2. Acquire monitoring data of various monitoring indicators of the aircraft; based on the monitoring data of various monitoring indicators, perform sensitivity analysis on the key parameters of the optimal maintenance timing prediction model, calculate the sensitivity factor of each key parameter to the optimal maintenance timing, and store the sensitivity factor in the knowledge base. S3. Construct an evolutionary learning co-optimization algorithm consisting of an exploration engine, a learning engine, an experience base, and a knowledge base. Each key parameter at each time step forms a parameter vector. The evolutionary learning co-optimization algorithm is used to optimize the parameter vector at each time step. Specific methods include: S301. For each time step, a random number label is generated. If the random number label is greater than the preset learning rate, the exploration engine is used to perform a global search optimization on the current parameter vector to obtain the first updated parameter vector. If the random number label is less than or equal to the preset learning rate, the current parameter vector is used as input to predict the second optimization direction using the current learning engine, and the current parameter vector is updated according to the second optimization direction to obtain the second updated parameter vector. For each moment when the exploration engine performs optimization, if the fitness corresponding to the first updated parameter vector is better than the fitness corresponding to the current parameter vector, then the current parameter vector and the corresponding first optimization direction are stored in the experience base, wherein the first optimization direction is the difference between the first updated parameter vector and the current parameter vector; the fitness is the squared error between the predicted value of the optimal maintenance time and the reference value of the optimal maintenance time. S302. Determine if the experience base is empty. If it is not empty, proceed to S303; if it is empty, proceed to S304. S303. Use the sensitivity factors in the knowledge base to correct the first optimization direction in the experience base to obtain the corrected optimization direction; use the current parameter vector in the experience base as input and the corresponding corrected optimization direction as label to train the current learning engine to obtain the trained learning engine; after training, clear the experience base. S304. Determine whether the current iteration round has reached the preset number of iterations. If not, use the parameter vector updated at each time step of the current iteration round as the current parameter vector for the next iteration round and return to S301. If it has reached the preset number of iterations, stop the iteration and output the learning engine trained in the current iteration round as the optimized learning engine. S4. For the time to be measured, calculate the initial parameter vector based on the corresponding monitoring data, input the initial parameter vector into the optimized learning engine, output the predicted optimization direction, update the initial parameter vector according to the predicted optimization direction, obtain the optimized parameter vector, substitute the optimized parameter vector into the optimal maintenance timing prediction model, input the monitoring data into the optimal maintenance timing prediction model, and output the optimal maintenance timing for the time to be measured.

2. The method for determining aircraft maintenance timing based on evolutionary learning co-optimization according to claim 1, characterized in that, In S1, the triangular membership transformation method is used to transform the monitoring data of each monitoring indicator into an evidence distribution: In the formula, The first step in the health status of the aircraft Each assessment level ; The total number of assessment levels for the health status of the aircraft; For the first Each monitoring indicator at time belong Confidence level, For the first Each monitoring indicator at time Distribution of evidence.

3. The method for determining aircraft maintenance timing based on evolutionary learning co-optimization according to claim 1, characterized in that, In S1, the process of fusing the evidence distributions corresponding to each monitoring indicator includes: The evidence distribution is transformed into a weighted confidence distribution with reliability, and the basic probability mass of each evidence distribution is calculated: In the formula, For the first Each monitoring indicator at time Evidence distribution assigned The basic probability mass; The first step in the health status of the aircraft Each assessment level ; The total number of assessment levels for the health status of the aircraft; For the first Each monitoring indicator at time The correlation discount factor For the first Each monitoring indicator at time Weight of evidence For the first Each monitoring indicator at time Weighted evidence weights, For the first Each monitoring indicator at time belong Confidence level, A set of assessment levels for the health status of an aircraft; express The power set; For the first Each monitoring indicator at time The reliability of the evidence; The basic probability quality of each evidence distribution is iteratively fused using evidence reasoning rules to obtain the confidence level of each health status assessment level after fusion. Based on the confidence levels of each health status assessment level after fusion, the health status confidence distribution is obtained. In the formula, For the aircraft at all times The confidence distribution of health status. For a moment The health status of the aircraft belongs to The confidence level.

4. The method for determining aircraft maintenance timing through evolutionary learning co-optimization according to claim 1, characterized in that, In S2, the key parameters include three categories: evidence weight, evidence reliability, and relevance discount factor; the sensitivity factor is obtained by taking the partial derivative of the initial output of the optimal maintenance timing prediction model with respect to the key parameters.

5. The method for determining aircraft maintenance timing based on evolutionary learning co-optimization according to claim 1, characterized in that, In S3, the exploration engine uses a particle evolution algorithm to perform a global search optimization on the current parameter vector; the current learning engine uses a feedforward neural network to predict the second optimization direction.

6. The method for determining aircraft maintenance timing through evolutionary learning co-optimization according to claim 1, characterized in that, In S303, the method for correcting the first optimization direction in the experience base is as follows: Calculate the correction factors for each key parameter separately: In the formula, For the first Each monitoring indicator at time The Correction factors for class key parameters, For the first Each monitoring indicator at time The The amount of change corresponding to the key parameters of the class. The total number of categories for key parameters; The first optimization direction is modified by adjusting the correction factor and the sensitivity factor in the knowledge base: In the formula, For the first Each monitoring indicator at time The The corrected changes in the class of key parameters. For a moment The fitness of the current parameter vector. For the first Each monitoring indicator at time The Sensitivity factors for class-critical parameters.

Images