Shipboard aircraft takeoff performance evaluation method and system based on digital twin enabling

Abstract

The invention discloses a carrier-based aircraft take-off performance evaluation method and system based on digital twin enabling, belongs to the technical field of performance evaluation, and aims to solve the problems that a traditional method is insufficient in full-working-condition coverage, low in multi-parameter coupling evaluation precision and poor in real-time performance. The core process of the method comprises the following steps: constructing an influence parameter space containing conventional and extreme working conditions, and generating a sampling parameter group through layered guided sampling; collecting data and constructing a health monitoring data set by relying on the simulation platform; the measuring point parameters are converted into mechanism parameters through a mechanism model of the engine core component; clustering and dividing typical clustering groups, dynamically matching weights, training a time sequence deep learning network, and constructing a clustering type reference performance digital twin model library; and during actual takeoff, a target cluster group model is matched to generate a predicted value, and performance grading is completed through residual analysis and multi-dimensional index calculation. According to the method, high-precision and real-time evaluation of the take-off performance of the shipboard aircraft under all working conditions is realized, and the evaluation suitability and the decision support reliability are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of performance evaluation, more particularly to a carrier-based aircraft take-off performance evaluation method and system based on digital twin empowerment. BACKGROUND

[0002] The carrier-based aircraft take-off performance directly determines the combat effectiveness and flight safety of the aircraft carrier, and the evaluation process needs to face multiple challenges such as high dynamicity of the marine environment, strong coupling of multiple parameters, and complexity of the working condition. The random fluctuations of temperature, wind speed, and sea wave parameters in the marine environment, combined with factors such as the movement of the aircraft carrier deck and the change of engine operating state, make the carrier-based aircraft take-off process involve the cross-influence of multiple dimensions of indicators such as environmental parameters, control parameters, and engine core parameters, and need to adapt to both conventional take-off and landing scenarios and critical working conditions such as extreme sea conditions and high temperature and humidity.

[0003] Currently, there are still many limitations in the related technology of carrier-based aircraft take-off performance evaluation: first, the working condition coverage has shortcomings, the existing evaluation methods focus on parameter matching under conventional working conditions, and the parameter boundary definition under extreme working conditions is not comprehensive, which makes it difficult to cover all kinds of critical safety scenarios, resulting in a lack of reference value for evaluation results under extreme conditions; second, the multi-parameter coupling processing capability is insufficient, the internal correlation and dynamic influence between different parameters cannot be fully considered, and only single-dimensional parameters or simple linear models are used for evaluation, which makes it difficult to accurately reflect the real take-off performance under complex coupling mechanism; third, the performance representation depth is insufficient, the existing technology is mostly based on engine surface measurement point parameters for analysis, and the internal operation mechanism of the core components cannot be deeply excavated, which makes it difficult to strip away environmental interference and external fluctuations, and accurately capture the real performance state of the engine; fourth, the model adaptability and real-time performance are poor, the evaluation models used are mostly general single models, which cannot adapt to the differentiated characteristics of different working conditions, have limited generalization ability, and have low efficiency in data processing and model matching, which makes it difficult to meet the real-time decision-making needs in the carrier-based aircraft take-off process; therefore, in order to overcome these limitations, the present application proposes a carrier-based aircraft take-off performance evaluation method and system based on digital twin empowerment. SUMMARY

[0004] In view of the deficiencies in the prior art, the present application aims to provide a carrier-based aircraft take-off performance evaluation method and system based on digital twin empowerment, which solves the problem of how to accurately, adaptively, and in real time, quantitatively evaluate the take-off performance in the scenario where conventional and extreme working conditions coexist and multiple parameters are coupled.

[0005] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0006] The carrier-based aircraft take-off performance evaluation method based on digital twin empowerment comprises:

[0007] An influence parameter space of the carrier-based aircraft is constructed, and a hierarchical guided sampling method is used to intelligently sample in the influence parameter space to construct a sampling parameter group in combination with priority weights of the influence parameters;

[0008] A carrier-based aircraft takeoff simulation training platform is configured according to the sampling parameter group, engine measurement point parameter data and performance parameter data in the takeoff stage under each influence parameter combination are collected to construct a health monitoring data set;

[0009] Based on the mechanism model of each core component of the carrier-based aircraft engine, the engine measurement point parameters of each sub-stage data processing node in the takeoff stage are converted into mechanism parameters;

[0010] Based on the performance parameter similarity and working condition type corresponding to each influence parameter combination, the sampling parameter group is clustered and divided to construct a typical cluster group and its associated training data set, the working condition features of each typical cluster group are extracted, and then the sub-stage adaptability weight and working condition sensitivity weight of each mechanism parameter are dynamically matched to build and train a time series deep learning network of each typical cluster group to construct a clustered reference performance digital twin model library;

[0011] In the actual takeoff process of the carrier-based aircraft, the target typical cluster group is selected based on the influence parameter data in the takeoff stage, and the predicted value sequence of the performance parameter is generated through the reference performance digital twin model of the target typical cluster group, and the performance parameter residual curve is constructed by combining the actual value sequence of the performance parameter, the multi-dimensional takeoff performance evaluation index is calculated, and the carrier-based aircraft takeoff performance level is divided.

[0012] Specifically, the steps of constructing the sampling parameter group include:

[0013] The conventional value interval of the influence parameter related to the carrier-based aircraft takeoff performance evaluation is obtained, and the extreme value interval of each influence parameter is set, and the influence parameters include environmental parameters and control parameters;

[0014] The conventional value interval and the extreme value interval of each influence parameter are fused to construct an influence parameter space;

[0015] The priority level of each influence parameter is set, and intelligent sampling is performed in the influence parameter space according to the priority level of each influence parameter by using a hierarchical guided sampling method, that is:

[0016] The sampling quota of each priority level is configured, the conventional value interval of each influence parameter is combined, and a conventional sampling point is set, wherein the sampling interval is dynamically adjusted according to the influence parameter change sensitivity;

[0017] The influence parameter change sensitivity refers to the degree of change of the carrier-based aircraft takeoff performance index caused by the unit change of the influence parameter;

[0018] For the extreme value interval of each influencing parameter, the boundary encryption sampling and key node locking are combined to set the extreme sampling points; and combined with the conventional sampling points, the preliminary sampling parameter set is constructed.

[0019] The rationality of each influencing parameter combination in the preliminary sampling parameter set is verified, and the influencing parameter combination that fails the rationality verification is removed, and finally the sampling parameter group is formed.

[0020] Specifically, the steps of constructing the health monitoring dataset include:

[0021] According to each influencing parameter combination of the sampling parameter group, the ship-borne aircraft take-off simulation training platform is configured, for the environmental parameters, the environmental parameter values in the influencing parameter combination are fixedly configured; for the control parameters, the control parameter values in the influencing parameter combination are taken as the target values of the take-off stage of the current simulation training, and the dynamic value sequence of each control parameter is generated;

[0022] The standardized take-off process of the ship-borne aircraft take-off simulation training platform is started, the engine measurement point parameter data and performance parameter data in the take-off stage are collected at a preset sampling frequency, and the collection time label is labeled, and the simulation time sequence data sequence is formed;

[0023] After performing the multi-dimensional data preprocessing operation on the simulation time sequence data sequence, the structured packaging is performed, and the identification label is added, the identification label includes the influencing parameter combination code, the training timestamp, and the working condition type; the working condition type includes the normal working condition and the extreme working condition;

[0024] According to the three-dimensional data structure of the influencing parameter combination, the parameter type and the collection time label, the association mapping relationship between the engine measurement point parameter data and the corresponding performance parameter data is established, and the standardized data unit is integrated; and all the standardized data units are integrated to form the health monitoring dataset.

[0025] Specifically, the step of converting the engine measurement point parameter of each sub-stage data processing node in the take-off stage into a mechanism parameter includes:

[0026] The sub-stage of the take-off stage of the standardized take-off simulation training under each influencing parameter combination is divided, and according to the time interval boundary of each sub-stage, combined with the preset parameter collection period, the data processing node of the simulation time sequence data sequence of each standardized data unit in the health monitoring dataset of each standardized take-off simulation training is divided;

[0027] The mapping relationship between the engine measurement point parameter and the mechanism model of the core component is established, the measurement point parameter of each data processing node is extracted according to the engine core component type, and the corresponding mechanism parameter is obtained through the mechanism model of the corresponding core component;

[0028] The mechanism model refers to a mathematical and physical model constructed based on the physical structure of the core component, aerodynamic thermodynamic principles and operation constraints, and capable of quantitatively describing the internal correlation between the input engine measurement point parameters and the output mechanism parameters;

[0029] The standard value range of each mechanism parameter is configured, each mechanism parameter is checked one by one, and the abnormal core component, abnormal sub-stage and abnormal data processing node number of the abnormal mechanism parameter exceeding the standard value range are identified and marked;

[0030] The engine measurement point parameter data of the non-abnormal data processing node in the abnormal sub-stage to which the abnormal mechanism parameter belongs is obtained; the abnormal data processing node where the abnormal mechanism parameter is located is corrected by interpolation according to the change trend of the mechanism parameter, and then the corresponding mechanism parameter is corrected;

[0031] The mechanism parameters of each core component after checking and correction are associated and bound with the corresponding data processing node, the identification of the sub-stage, the combination code of the influence parameter and the working condition type label, to form a mechanism parameter data set.

[0032] Specifically, the step of constructing a typical cluster group comprises:

[0033] All influence parameter combinations in the sampling parameter group are extracted, and the performance parameters and working condition types corresponding to each influence parameter combination are associated one by one to construct a clustering original data set containing influence parameter combination codes, performance parameter sets and working condition types;

[0034] According to the working condition type, the clustering original data set is divided to generate clustering sub-data sets, including a regular clustering sub-data set and an extreme clustering sub-data set;

[0035] The performance parameters in each clustering sub-data set are standardized to map all performance parameters to the same numerical interval; for each clustering sub-data set, the silhouette coefficient method is used for clustering number optimization, and the silhouette coefficient of the clustering result corresponding to each clustering number is calculated to determine the optimal clustering number;

[0036] For each clustering sub-data set, the clustering center is initialized based on the optimal clustering number, and the clustering calculation method is set, and the iterative clustering operation is started to generate the clustering cluster corresponding to each clustering sub-data set;

[0037] The similarity of the performance parameters corresponding to all influence parameter combinations in each clustering cluster is calculated, and the similarity threshold is set to eliminate redundant influence parameter combination samples in the clustering cluster, and a typical cluster group is constructed.

[0038] Specifically, the step of constructing a clustered reference performance digital twin model library comprises:

[0039] For each typical clustering group, the corresponding mechanism parameter, performance parameter and influence parameter combination are associated to construct an associated training data set and normalized preprocessing is performed; the associated training data set of each typical clustering group is divided into a training set, a validation set and a test set by using a stratified sampling method;

[0040] Based on the influence parameter value interval distribution and performance parameter distribution in the associated training data set of the typical clustering group, the working condition characteristics of each typical clustering group are extracted; through a preset working condition characteristic mechanism parameter association matrix, the sub-stage adaptability weight and the working condition sensitivity weight of each mechanism parameter are dynamically matched;

[0041] The working condition characteristic mechanism parameter association matrix refers to a three-dimensional structured matrix constructed with the working condition characteristics of the typical clustering group as the row, the mechanism parameters of the engine core components as the column, and the sub-stage as the dimension, and the matrix element is the association strength coefficient of the corresponding working condition characteristics and mechanism parameters in the corresponding take-off sub-stage;

[0042] Based on the dynamically matched sub-stage adaptability weight and working condition sensitivity weight, combined with the influence parameter combination, a time series deep learning network is constructed for establishing a nonlinear mapping of the influence parameter combination and the mechanism parameter to the performance parameter in each typical clustering group;

[0043] For the time series deep learning network of each typical clustering group, independent training is carried out using the corresponding divided training set and validation set, and the test set is used to verify the accuracy of each trained time series deep learning network of the typical clustering group, and the relative error mean of the performance parameter predicted value and the true value is calculated as the performance parameter prediction error rate;

[0044] The time series deep learning network verified by accuracy is used as the benchmark performance digital twin model of its typical clustering group, and the benchmark performance digital twin model of all typical clustering groups is combined to construct a clustered benchmark performance digital twin model library.

[0045] Specifically, the time series deep learning network is composed of a time series feature extraction layer, a weight-enhanced attention layer, a multi-dimensional fusion layer and an output mapping layer;

[0046] The time series feature extraction layer is used to adapt different size convolution kernels to the change frequency of the weighted mechanism parameters and influence parameter combination in the typical clustering group, to capture the dynamic change characteristics of the mechanism parameters in each take-off sub-stage;

[0047] The weight-enhanced attention layer is used to strengthen the feature contribution of the mechanism parameters, influence parameters and their time series sequence fragments in the current take-off sub-stage based on the sub-stage adaptability weight and working condition sensitivity weight of the corresponding take-off sub-stage of the typical clustering group;

[0048] The multi-dimensional fusion layer is used for cross-dimension integration of feature vectors of each influence parameter and each mechanism parameter in a typical clustering group, mapping to a unified high-dimensional feature space to form a high-dimensional feature vector;

[0049] The output mapping layer is used for mapping the high-dimensional feature vector output by the multi-dimensional fusion layer to a carrier-based aircraft take-off performance parameter space through a multi-layer fully connected neural network structure, and outputting a continuous performance parameter prediction value by means of a linear activation function.

[0050] Specifically, the step of generating a prediction value sequence of the performance parameter comprises:

[0051] The pre-processing operation is performed on the real-time collected influence parameter data, engine measurement point parameter data and actual performance parameter data, and the pre-processed engine measurement point parameter data is converted into corresponding mechanism parameter data;

[0052] Based on the pre-processed influence parameter data, the real-time working condition feature is extracted, the real-time working condition feature is compared with the working condition feature model index mapping table of the clustered reference performance digital twin model library, and the similarity of the real-time working condition feature and the working condition feature boundary of each typical clustering group is calculated to screen a target typical clustering group;

[0053] If the target typical clustering group in the take-off stage is unique, the reference performance digital twin model of the target typical clustering group is obtained; the real-time influence parameter data and the mechanism parameter data are inputted to generate a prediction value sequence of the corresponding performance parameter;

[0054] If the target typical clustering group in the take-off stage is not unique, the current take-off stage is segmented according to a fixed time interval, the working condition feature coincidence degree of each target typical clustering group in each time period is counted, the effective existence time length proportion of each target typical clustering group is calculated, the dominant target typical clustering group is screened, the corresponding reference performance digital twin model thereof is obtained, and the real-time data is inputted to generate a dominant prediction value sequence of the performance parameter; meanwhile, the reference performance digital twin models of the remaining target typical clustering groups are obtained to generate auxiliary prediction value sequences of the performance parameter respectively; the effective existence time length proportion of each target typical clustering group is set as a weighting coefficient, and the dominant prediction value sequence and all auxiliary prediction value sequences are weighted and combined to obtain a performance parameter prediction value sequence.

[0055] Specifically, the step of dividing the carrier-based aircraft take-off performance level comprises:

[0056] The actual value sequence and the prediction value sequence of the performance parameter are calibrated in time dimension, the performance parameter residual value of each time node is calculated, and a performance parameter residual curve is constructed;

[0057] The time sequence variation characteristics of the residual curve are extracted by a sliding window method, and based on the residual curve and the time sequence variation characteristics thereof, multi-dimensional take-off performance evaluation indexes are calculated, including a time sequence coincidence degree index, a parameter deviation rate index and a sub-stage matching degree index.

[0058] The time sequence coincidence degree index is obtained by calculating the correlation coefficient of the actual value sequence and the predicted value sequence of the performance parameter; the parameter deviation rate index is obtained by calculating the relative error of the residual value of each performance parameter at each time node and the corresponding actual value, and taking the average of the relative errors of all performance parameters; the sub-stage matching degree is obtained by calculating the parameter deviation rate and the time sequence coincidence degree in the stage respectively according to the sub-stage, and then combining the preset sub-stage weight to perform weighted summation on the two indexes.

[0059] A quantitative performance level determination standard is set, and the time sequence coincidence degree index, the parameter deviation rate index and the sub-stage matching degree index are combined to divide the carrier aircraft take-off performance level.

[0060] The carrier aircraft take-off performance evaluation system based on digital twin empowerment includes a data acquisition module, a model scheduling module, a performance evaluation module and a model training module.

[0061] The data acquisition module is used to acquire influence parameter data, engine measurement point parameter data and actual performance parameter data during the actual take-off process of the carrier aircraft; the model scheduling module extracts real-time working condition characteristics based on the influence parameter data, screens a target typical clustering group, and generates a predicted value sequence of the performance parameter through a benchmark performance digital twin model of the target typical clustering group; the performance evaluation module is used to combine the actual value sequence of the performance parameter to construct a performance parameter residual curve, calculate multi-dimensional take-off performance evaluation indexes, and divide the carrier aircraft take-off performance level; and the model training module is used to construct a typical clustering group and its associated training data set based on a sampling parameter group, a health monitoring data set and a mechanism parameter data set, build and train a time sequence deep learning network of each typical clustering group, and construct a clustered benchmark performance digital twin model library.

[0062] The beneficial effects of the present application are as follows:

[0063] The application guarantees the rationality and comprehensiveness of the sampling parameter combination by constructing and layering intelligent sampling of the influence parameter space covering conventional and extreme working conditions comprehensively, and provides solid data support for evaluation; the data quality and structured degree are improved by constructing a standard health monitoring data set; the deep conversion of measurement point parameters to mechanism parameters is realized by means of the mechanism model of the engine core component, the external interference is accurately stripped, and the real running state of the core component is captured; the cluster type benchmark performance digital twin model library is constructed by clustering and dividing typical cluster groups, dynamically matching weights and training exclusive time sequence deep learning networks, and the precise adaptation to different working condition characteristics is realized; through target typical cluster group matching and prediction value sequence generation during actual take-off, combined with multi-dimensional evaluation indexes and performance level division, the problems of multi-parameter coupling, poor performance characterization, poor model adaptability and insufficient real-time performance under complex working conditions are effectively solved, and finally the comprehensive, accurate and real-time quantitative evaluation of the performance of the carrier-based aircraft during take-off is realized, which provides a reliable basis for take-off safety guarantee and operation and maintenance decision-making. BRIEF DESCRIPTION OF DRAWINGS

[0064] Figure 1 The flowchart of the carrier-based aircraft take-off performance evaluation method based on digital twin empowerment of the application;

[0065] Figure 2 The flowchart of constructing a typical cluster group of the application;

[0066] Figure 3 The flowchart of generating a prediction value sequence of performance parameters of the application;

[0067] Figure 4 The flowchart of dividing the performance level of the carrier-based aircraft during take-off of the application. DETAILED DESCRIPTION

[0068] Please refer to Figure 1 The embodiment introduces a carrier-based aircraft take-off performance evaluation method based on digital twin empowerment, which includes:

[0069] Step S1: Based on the characteristics of carrier-based takeoff scenarios, identify the influencing parameters related to carrier-based aircraft takeoff performance evaluation. Influencing parameters refer to various parameters that directly or indirectly affect the carrier-based aircraft takeoff process, impacting takeoff performance and safety. Specifically, these include environmental parameters reflecting the external environment characteristics of carrier-based takeoff and control parameters reflecting human operation and system control commands. Combining the changing patterns of the marine environment and the operational characteristics of the aircraft carrier platform, obtain the conventional value ranges of the influencing parameters for carrier-based aircraft. Simultaneously, based on the carrier-based aircraft takeoff safety boundary standards and the physical limits of engine operation, set the extreme ranges for each influencing parameter to ensure that extreme parameters comprehensively cover various critical operating conditions, thereby constructing a complete influencing parameter space encompassing both conventional and extreme operating conditions. Furthermore, combining the priority weights of the influencing parameters, a hierarchical guided sampling method is used for intelligent sampling in the influencing parameter space. During the sampling process, sampling weights are allocated according to parameter priority, prioritizing the sampling density of high-weight parameters. A physical rationality verification mechanism is introduced to verify the sampled parameter combinations, eliminating parameter combinations that violate thermodynamic laws and mechanical operating constraints. This constructs a sampled parameter set, ensuring that limited sampled data can comprehensively reflect the takeoff performance characteristics under various operating conditions, effectively reducing subsequent model fitting bias.

[0070] In this embodiment, step S1 uses a hierarchical weighted sampling method combined with a physical rationality verification mechanism to achieve intelligent sampling within the influence parameter space covering both normal and extreme operating conditions. This ensures the sampling effectiveness of high-impact parameters through priority weight allocation and ensures the comprehensiveness and scientific nature of the data through limit parameter setting and rationality screening. Ultimately, it achieves complete coverage of takeoff performance characteristics under various operating conditions with limited sampled data, effectively reducing subsequent model fitting bias and providing accurate and comprehensive data support for the construction of a benchmark performance digital twin model.

[0071] Preferably, the specific steps for constructing the sampling parameter set include:

[0072] By combining the changing patterns of the marine environment with the operational characteristics of the aircraft carrier platform, we define the conventional value ranges for environmental and control parameters related to the takeoff performance evaluation of carrier-based aircraft. We determine the conventional value ranges for influencing parameters that conform to actual operating conditions, ensuring that the conventional value ranges can fully cover the parameter variation range under conventional takeoff scenarios, and providing basic range support for subsequent parameter space construction.

[0073] Based on the safety boundary standards for carrier-based aircraft takeoff and the physical limit requirements for engine operation, extreme value ranges for each influencing parameter are set. By analyzing various critical safety scenarios during carrier-based aircraft takeoff, the maximum allowable fluctuation boundaries of each parameter under extreme conditions are clarified. At the same time, combined with the physical operating constraints of the engine's core components, the rationality of the extreme value ranges is verified to ensure that the extreme parameters can fully cover various critical operating scenarios and do not exceed the equipment's safe operating threshold.

[0074] By merging the defined normal value ranges and the set extreme value ranges of each influencing parameter, an influencing parameter space containing both normal and extreme operating conditions is constructed. This influencing parameter space establishes the inherent logical relationship between each influencing parameter through parameter dimension association mapping technology, forming a structured parameter space system, and providing a comprehensive and orderly set of parameters for subsequent intelligent sampling.

[0075] Based on the mechanism of carrier-based aircraft takeoff performance and the statistical results of historical operation data, a performance impact quantification evaluation method is adopted to set the priority level of each influencing parameter. Specifically, a multi-dimensional evaluation index system including parameter sensitivity, coupling effect coefficient, and performance contribution is constructed. Combined with grey relational analysis and hierarchical analysis, the influence of each parameter on takeoff performance is quantified and priority levels are divided for subsequent intelligent sampling process sampling resource allocation. This ensures that high-priority parameters that have a significant impact on takeoff performance receive higher sampling density, accurately capture the impact characteristics of key parameters on takeoff performance with limited sampling data, and improve the information completeness and relevance of the sampling parameter group.

[0076] Based on the priority level of each influencing parameter, a hierarchical guided sampling method is used to perform intelligent sampling within the influencing parameter space, namely:

[0077] Sampling quotas are configured for each priority level. Combined with the typical value ranges of each influencing parameter, dynamic equidistant sampling technology is used to set up regular sampling points within these ranges. The sampling interval is dynamically adjusted based on the sensitivity of the influencing parameter changes. In areas sensitive to parameter changes, the sampling interval is reduced and the number of sampling points is increased; in areas with gradual parameter changes, the sampling interval is increased to ensure accurate capture of key characteristics of parameter changes. The sensitivity of influencing parameter changes refers to the degree of change in carrier-based aircraft takeoff performance indicators caused by a unit change in the influencing parameter. This is obtained by constructing a parameter performance correlation response model. Specifically, combining the carrier-based aircraft takeoff dynamics mechanism with historical takeoff performance data, a quantitative mapping relationship between each influencing parameter and takeoff performance indicators is established. The rate of change of performance indicators corresponding to parameter changes is calculated, thereby quantitatively characterizing the sensitivity of each influencing parameter to changes within different value ranges.

[0078] For the extreme value ranges of each influencing parameter, a combination of boundary encrypted sampling and key node locking is used to set extreme sampling points. That is, encrypted sampling points are set along the boundary line of the extreme value range at the minimum interval, while the influencing parameter values ​​corresponding to the performance critical nodes under extreme conditions are locked as core sampling points to ensure that key conditions and boundary scenarios within the extreme value range are effectively covered. During the sampling process, parameter combination correlation constraint rules are introduced simultaneously. Based on the parameter dimension correlation mapping relationship in the influencing parameter space, it is ensured that the parameter combination formed by regular sampling points and extreme sampling points can cover the cross-coupling scenarios of different dimension parameters, avoid information loss caused by isolated sampling, and form a preliminary sampling parameter set.

[0079] The rationality of each combination of influencing parameters in the preliminary sampling parameter set is verified to determine whether the takeoff process corresponding to the combination of influencing parameters conforms to objective physical laws and whether it exceeds the safe operation boundary of the equipment. The combination of influencing parameters that fails the rationality verification is eliminated, and finally a sampling parameter set covering all operating conditions, with reasonable parameter combinations and complete information characteristics is formed.

[0080] Step S2: Based on the constructed sampling parameter set, a carrier-based aircraft takeoff simulation training platform is built. A standardized takeoff simulation training process is executed sequentially according to each combination of influencing parameters in the sampling parameter set. During training, the corresponding combination of influencing parameters is loaded in real time through the platform parameter control module, and engine measurement point parameter data and performance parameter data are collected in real time during the simulated takeoff phase. Engine measurement point parameters refer to the basic physical parameters that reflect the real-time operating status of the engine, directly collected by sensors deployed at various core components and key operating locations of the engine. These parameters specifically include engine inlet pressure, inlet temperature, compressor outlet pressure and temperature, turbine inlet temperature, engine speed, fuel injection pressure, fuel flow rate, and nozzle outlet exhaust pressure. Performance parameters directly reflect the carrier-based aircraft's... The core indicators of takeoff performance level cover parameters related to engine thrust output, flight attitude, and energy consumption during takeoff, including engine exhaust temperature, thrust coefficient, takeoff distance, departure speed, climb rate, and fuel consumption rate. During data acquisition, a time-series synchronization calibration technique is employed. A high-precision timestamp synchronization mechanism is used to align data from each acquisition channel in real time. A dynamic deviation compensation algorithm corrects the time delay of data from different sensors, ensuring accurate temporal matching between engine measurement parameters and performance parameters. After acquisition, the raw data undergoes noise reduction preprocessing and structured encapsulation to construct a health monitoring dataset, providing high-quality, high-reliability foundational data support for the subsequent construction of a benchmark performance digital twin model.

[0081] In this embodiment, step S2 establishes a high-fidelity carrier-based aircraft takeoff simulation training platform. By combining standardized simulation training procedures with high-precision time-series synchronous acquisition and data preprocessing technology, it achieves accurate, complete, and synchronized acquisition of engine measurement point parameter data and performance parameters corresponding to each combination of sampling parameters under the takeoff scenario. This constructs a high-quality, highly reliable health monitoring dataset, effectively ensuring the data reliability and effectiveness of the subsequent construction of the benchmark performance digital twin model, and laying a solid data foundation for the model to accurately map the relationship between parameters and performance.

[0082] Preferably, the specific steps for constructing a health monitoring dataset include:

[0083] The carrier-based aircraft takeoff simulation training platform is configured sequentially according to the various combinations of influencing parameters in the sampling parameter group. For environmental parameters, the environmental parameter values ​​in the combination of influencing parameters are fixed to ensure the stability of the external environment throughout the simulation training. For control parameters, the control parameter values ​​in the combination of influencing parameters are used as the target values ​​for the takeoff phase of the current simulation training. Based on the dynamic characteristics of the carrier-based aircraft takeoff phase, the pilot's operating procedures, and the engine response characteristics, the pre-set segmented dynamic adjustment rules for the control parameters generate a dynamic value sequence for each control parameter. This sequence is used to precisely drive the carrier-based aircraft takeoff simulation training platform to execute the control actions of each phase according to the actual takeoff operation logic, reproduce the dynamic changes of control parameters with the takeoff process, and ensure that the collected health monitoring data is highly consistent with the control characteristics of the actual takeoff conditions.

[0084] The standardized takeoff procedure of the carrier-based aircraft takeoff simulation training platform is initiated, and the multi-dimensional high-precision data acquisition system is activated simultaneously. Engine measurement point parameter data and performance parameter data during the takeoff phase are collected at a preset sampling frequency. Time-series synchronous calibration technology is used for real-time processing, and the acquisition time labels are marked for each engine measurement point parameter data and performance parameter data to form a continuous and complete simulation time-series data sequence.

[0085] Multi-dimensional data preprocessing operations are performed on the simulated time-series data sequence. Wavelet threshold denoising algorithm is used to filter random noise and electromagnetic interference signals from sensors, and sliding window smoothing is used to eliminate instantaneous pulse fluctuations in the data. Based on the physical limits of engine operation, the reasonable range of takeoff performance parameters, and parameter coupling constraints, abnormal data exceeding reasonable thresholds are automatically identified and removed. For missing data regions, an interpolation completion method based on the trend of adjacent time-series data and the correlation law of parameters is used to repair them, ensuring the integrity, reliability, and rationality of the data.

[0086] The simulated time-series data sequence after multi-dimensional data preprocessing is structurally encapsulated and labeled. The label includes the encoding of the influencing parameter combination, the training timestamp, and the operating condition type. The operating condition type refers to the takeoff scenario category divided according to the value range characteristics of the influencing parameter combination, including normal operating conditions and extreme operating conditions. The normal operating condition corresponds to the takeoff scenario where all influencing parameter combinations fall within the preset normal value range, and the extreme operating condition corresponds to the takeoff scenario where at least one parameter in the influencing parameter combination falls within the preset extreme value range. Based on the three-dimensional data structure of influencing parameter combination, parameter type, and acquisition timestamp, the association mapping relationship between engine measurement point parameter data and corresponding performance parameter data is established and integrated into independent standardized data units to ensure data traceability and correlation. Among them, parameter types refer to engine measurement point parameter data and performance parameter data. Engine measurement point parameter data are basic physical parameters that are directly collected by sensors deployed at various core components and key operating positions of the engine and are used to reflect the real-time operating status of the engine, such as engine intake pressure, intake temperature, compressor outlet pressure and temperature, etc. Performance parameter data are core indicator data that directly reflect the takeoff performance level of carrier-based aircraft, such as engine exhaust temperature, thrust coefficient, takeoff distance, departure speed, etc.

[0087] All standardized data units are integrated according to a unified data format and storage specifications. The format uniformity and logical correlation of each data unit are checked through a data consistency verification mechanism. Redundant data detection algorithms are used to remove duplicate data. Finally, a health monitoring dataset covering all working conditions, with accurate time-series matching and a standardized labeling system is constructed, providing high-quality and highly reliable data source support for the subsequent construction of a benchmark performance digital twin model.

[0088] Step S3: Based on the structural characteristics and aerodynamic-thermodynamic operating laws of the core components of the carrier-based aircraft engine, construct the mechanism models of each core component. This model is used to convert the engine's measurement parameters into the mechanism parameters of the core components, providing mechanistic data support for the benchmark performance digital twin model that can accurately characterize the engine's actual operating state. The core components refer to the key components that play a decisive role in the engine's thrust output, energy conversion efficiency, and operational safety, including the compressor, combustion chamber, turbine, and nozzle. The mechanism model is a mathematical-physical model constructed based on the physical structure, aerodynamic-thermodynamic principles, and operating constraints of the core components. It can quantitatively describe the intrinsic relationship between the input engine measurement parameters and the output mechanism parameters of the components. It is constructed based on the laws of conservation of mass, energy, and momentum, as well as the aerodynamic-thermodynamic characteristic equations of each core component, ensuring that the model accurately reflects the actual operating mechanism of the core components. For the simulated time-series data sequences of each standardized data unit in the health monitoring dataset, data processing nodes are divided according to the time-series characteristics of the carrier-based aircraft takeoff phase. Each data processing node corresponds to a parameter acquisition cycle of a fixed time interval within the takeoff phase. The engine measurement parameters of each data processing node are input into the mechanism model of the corresponding core component to obtain the corresponding mechanism parameters. For example, in the compressor mechanism model, the compressor efficiency and compressor boost ratio are obtained based on the intake air temperature, intake air pressure, and compressor outlet pressure. In the combustion chamber mechanism model, the combustion efficiency is calculated based on the fuel flow rate, intake air flow rate, and combustion chamber outlet temperature. In the turbine mechanism model, the turbine efficiency and turbine expansion ratio are derived by inverting the turbine inlet temperature, turbine outlet temperature, and engine speed. In the nozzle mechanism model, the fuel supply coefficient is derived by combining the nozzle outlet exhaust pressure and engine speed. Finally, the engine measurement parameter data is converted from external appearance parameters to internal performance parameters.

[0089] In this embodiment, step S3 constructs a mechanism model of the core components that conforms to the actual operating mechanism of the engine. Combined with time-series data processing during takeoff and multi-component parameter inversion calculation, it achieves accurate conversion of engine measurement point parameters into core mechanism parameters. This effectively removes the influence of environmental interference on performance characterization, accurately extracts the actual operating performance characteristics of the engine, and provides highly reliable and targeted mechanism layer data support for the subsequent construction of a digital twin model of benchmark performance. This ensures that the model can deeply map the intrinsic relationship between parameters and performance.

[0090] Preferably, the specific steps for converting engine measurement point parameters into mechanistic parameters of core components include:

[0091] Based on the objective flight state characteristics and preset judgment thresholds during carrier-based aircraft takeoff, the standardized takeoff simulation training under various combinations of influencing parameters is divided into sub-stages of the takeoff phase, including the takeoff roll sub-stage, the departure sub-stage, and the climb sub-stage. The judgment criteria are selected from objective indicators unrelated to control parameters, specifically landing gear pressure, flight speed, and flight altitude. The preset judgment thresholds are: for the takeoff roll sub-stage, the landing gear pressure is greater than the preset pressure threshold and the flight speed is less than the preset departure speed; for the departure sub-stage, the landing gear pressure is less than or equal to the preset pressure threshold and the flight altitude is less than the preset climb altitude; for the climb sub-stage, the flight altitude is greater than or equal to the preset climb altitude and the flight speed is greater than or equal to the preset departure speed. The time interval boundaries of each sub-stage are defined by combining the acquisition time stamps of the simulated time-series data.

[0092] Based on the time interval boundaries of each sub-stage and combined with the preset parameter acquisition cycle, the simulated time series data sequence of each standardized data unit in the health monitoring dataset of each standardized takeoff simulation training is divided into several data processing nodes with equal time intervals. Each node corresponds to a set of synchronously acquired multi-dimensional engine measurement point parameters, and each data processing node is bound to its respective sub-stage identifier and acquisition time.

[0093] A mapping relationship is established between engine measurement point parameters and the mechanistic models of core components. Measurement point parameters of each data processing node are extracted according to the type of engine core components. The corresponding mechanistic parameters are obtained through the mechanistic models of the corresponding core components. For example, parameters such as intake manifold pressure, intake air temperature, and compressor outlet pressure and temperature are matched to the compressor mechanistic model to solve for compressor efficiency and boost ratio; parameters such as fuel flow rate, intake air flow rate, and combustion chamber outlet temperature are matched to the combustion chamber mechanistic model to calculate combustion efficiency; parameters such as turbine inlet temperature, turbine outlet temperature, and engine speed are matched to the turbine mechanistic model to calculate turbine efficiency and expansion ratio; and parameters such as nozzle outlet exhaust pressure and engine speed are matched to the nozzle mechanistic model to calculate fuel supply coefficient.

[0094] Based on the physical operating limits of each core component, parameter coupling constraints, and historical calibration data, the standard value range of each mechanism parameter is configured, and each mechanism parameter is verified one by one. The abnormal mechanism parameter that exceeds the standard value range is automatically identified and marked to the abnormal core component, abnormal sub-stage, and abnormal data processing node sequence number.

[0095] Obtain engine measurement point parameter data of the non-abnormal data processing node within the abnormal sub-stage to which the abnormal mechanism parameter belongs; based on the changing trend of the mechanism parameter, perform interpolation correction on the abnormal data processing node where the abnormal mechanism parameter is located, and then correct the corresponding mechanism parameter.

[0096] The verified and corrected mechanism parameters of each core component are associated and bound with the corresponding data processing node, the sub-stage identifier, the combination code of the influencing parameter, and the operating condition type label to form a complete mechanism parameter dataset, thereby realizing the accurate conversion and standardized storage of engine measurement point parameters from external appearance to internal performance parameters.

[0097] Step S4: The sampled parameter groups are divided using a clustering algorithm to construct a clustered benchmark performance digital twin model, improving the accuracy of takeoff performance evaluation under multiple operating conditions. Based on the similarity of performance parameters and operating condition type corresponding to each combination of influencing parameters, the sampled parameter groups are clustered, redundant samples are removed, and the effectiveness of clustering is verified. Multiple typical clusters covering the entire operating condition range and with significant operating condition characteristics are obtained, achieving accurate operating condition classification and training sample optimization for the sampled parameter groups. For each typical cluster, its corresponding influencing parameter combination, sub-stage identifier, verified and corrected mechanism parameter, and standardized performance parameter are associated to construct a dedicated training dataset for influencing parameter combinations, sub-stages, mechanism parameters, and performance parameters. For the associated training dataset of each typical cluster, mechanism-driven and data-driven approaches are used. The system employs a dynamic dual-mode fusion architecture for independent model fitting and training. The data-driven part builds a customized deep learning network, which includes temporal convolutional layers, attention mechanism layers, and fully connected layers. The temporal convolutional layers are used to extract the temporal evolution features of parameters, the attention mechanism layers are used to strengthen the feature weights of sub-stages and performance parameters within clusters, and the fully connected layers are used to achieve nonlinear mapping between multi-dimensional parameters. During training, a K-fold cross-validation mechanism is introduced to iteratively optimize the model hyperparameters. Independent test sets are used to verify the accuracy of the models in each cluster group. Finally, a benchmark performance digital twin model adapted to the specific working conditions of each typical cluster group is output. Through clustered modeling, the model achieves accurate matching between different working conditions and scenarios, significantly improving the accuracy, adaptability, and generalization stability of the model for evaluating the takeoff performance of carrier-based aircraft.

[0098] In this embodiment, step S4 involves clustering and optimizing the sample parameter group, and using a dual-mode fusion architecture driven by mechanism and data to construct customized digital twin models for each typical cluster group. This achieves accurate matching between different working conditions and the model, significantly improving the accuracy, adaptability, and generalization stability of the model in evaluating the takeoff performance of carrier-based aircraft, and effectively solving the problem of insufficient evaluation accuracy of a single model under all working conditions.

[0099] Preferably, the specific components of constructing a clustered benchmark performance digital twin model include:

[0100] Please see Figure 2The process involves extracting all influencing parameter combinations from the sampling parameter group, associating each combination with its corresponding performance parameters and operating conditions, and constructing a clustered original dataset containing influencing parameter combination codes, performance parameter sets, and operating condition types. This ensures that each sample in the dataset corresponds one-to-one with an influencing parameter combination in the sampling parameter group. Furthermore, based on the operating condition type, the clustered original dataset is divided into clustered sub-datasets, including regular clustered sub-datasets and extreme clustered sub-datasets.

[0101] The performance parameters in each cluster subset are standardized to map all performance parameters to the same numerical range. A unified standardization method is used to eliminate the differences in the dimensions of different performance parameters. At the same time, an outlier detection algorithm is used to identify and remove outlier samples in the performance parameters to ensure the reliability and consistency of the clustering data.

[0102] For each clustered subset, the silhouette coefficient method is used to optimize the number of clusters. The preset range of cluster sizes is traversed, and the silhouette coefficient is calculated for each cluster size. The cluster size corresponding to the maximum silhouette coefficient is selected as the optimal number of clusters, determining the final clustering dimension. The preset range of cluster sizes is based on in-depth analysis and experience summarizing carrier-based aircraft takeoff conditions. Under normal conditions, although carrier-based aircraft takeoff performance fluctuates to some extent, parameter changes are relatively stable and concentrated within a certain range. Therefore, the number of clusters should not be too large, otherwise it will lead to model overfitting and poor adaptability to new data. Under extreme conditions, although parameter changes are drastic and diverse, the number of clusters cannot be increased indefinitely considering the computational complexity of the model and the needs of practical applications. This range balances the gradual clustering under normal conditions with the boundary scenario coverage under extreme conditions, avoiding insufficient model generalization due to overly coarse clustering or overfitting due to overly fine clustering. It can be dynamically optimized using the silhouette coefficient method.

[0103] For each clustered subset, cluster centers are initialized based on its optimal number of clusters: one sample is randomly selected from the samples in the clustered subset as the first center, and subsequently, the sample with the largest Euclidean distance to the selected center is selected as the new center, until the optimal number of clusters is reached. The clustering calculation method is set to Euclidean distance, and iterative clustering is initiated. In each iteration, the Euclidean distance between each sample and all cluster centers is calculated, and the sample is assigned to the nearest cluster. The center coordinates of each cluster are iteratively updated, and the iteration is repeated until the change in cluster center coordinates is less than a preset threshold, or the preset maximum number of iterations is reached, generating the clusters corresponding to each clustered subset.

[0104] Calculate the similarity of performance parameters corresponding to all combinations of influencing parameters within each cluster, set a similarity threshold, remove redundant combinations of influencing parameters within the cluster, and construct typical cluster groups. That is, compare the combinations of influencing parameters within the cluster pairwise. If the similarity between samples is higher than the preset similarity threshold, retain one core sample and delete the rest of the redundant samples. This reduces the number of samples within the cluster while fully preserving the clustering features.

[0105] For each typical cluster group, a dedicated association training dataset is constructed by associating corresponding validated and corrected mechanistic parameters, performance parameters, and combinations of influencing parameters based on the combination of influencing parameters, sub-stages, and collection time. For typical cluster groups with insufficient sample size, the SMOTE data augmentation algorithm is used to generate effective virtual samples to ensure the sample balance of the dataset. All association training datasets undergo normalization preprocessing, and the Min-Max normalization method is used to map each parameter to a uniform numerical range to eliminate the interference of dimensional differences on model training.

[0106] The associated training datasets for each typical cluster group are divided into training, validation, and test sets using stratified sampling. During stratified sampling, samples are partitioned according to the proportion of sub-stage identifiers and working condition features to ensure that the parameter distribution and sub-stage proportions of the training, validation, and test sets are consistent with the original dataset, avoiding model training bias caused by uneven sample distribution. After partitioning, labels are added to each dataset to clearly identify the corresponding typical cluster group, dataset type, and partitioning time.

[0107] Based on the distribution of the range of values ​​of the parameters and the distribution of the performance parameters in the association training dataset of typical clusters, the working condition characteristics of each typical cluster are extracted. The working condition characteristics refer to the set of key parameters and distribution patterns that can characterize the essential attributes of the working condition of the typical cluster, including the range of environmental parameters, the setting range of control parameters, the statistical characteristics of performance parameters, and the trend characteristics of sub-stage parameters.

[0108] Based on the operating condition characteristics of typical clusters, the model dynamically matches the sub-stage adaptability weights and operating condition sensitivity weights of each mechanistic parameter through a pre-set operating condition characteristic mechanism parameter correlation matrix. The sub-stage adaptability weights are weights assigned to different takeoff sub-stages within a typical cluster, based on the degree of influence of the mechanistic parameters on the takeoff performance of that sub-stage. This prioritizes mechanistic parameters that play a key role in performance at the corresponding sub-stage, such as turbine efficiency in the departure sub-stage and thrust coefficient correlation mechanism parameters in the climb sub-stage, improving the accuracy of the model's performance mapping for each sub-stage. The operating condition sensitivity weights are weights assigned based on the operating condition characteristics of typical clusters, according to the sensitivity of the mechanistic parameters to performance fluctuations under that operating condition. This strengthens the model's adaptability to operating condition differences, such as increasing the sensitivity weight of combustion chamber efficiency under extreme high-temperature conditions and increasing the sensitivity weight of aerodynamic drag correlation mechanism parameters under high wind speed conditions, ensuring that the model can accurately reflect the actual operating mechanism of the engine under different operating conditions.

[0109] The correlation matrix of operating condition characteristics and mechanism parameters refers to a three-dimensional structured matrix constructed with operating condition characteristics of typical clusters as rows, mechanism parameters of engine core components as columns, and takeoff sub-stages as dimensions. The matrix elements are the correlation strength coefficients of corresponding operating condition characteristics and mechanism parameters under the corresponding takeoff sub-stages. These correlation strength coefficients quantify the linear or nonlinear correlation between operating condition characteristics and mechanism parameters, and also include the benchmark weight information of sub-stage adaptability and operating condition sensitivity. By extracting operating condition characteristic data and corresponding mechanism parameter data of typical clusters, the linear correlation strength is calculated using Pearson correlation coefficient, and nonlinear correlation relationships are captured using Spearman rank correlation coefficient or mutual information method. Combined with the engine thermodynamic mechanism rule base, physical rationality is verified, weak correlation and contradictory correlation terms are eliminated, and then the correlation strength threshold is set based on historical operation data statistics and domain expert experience to determine the core elements of the matrix. Finally, the matrix coefficient settings are iteratively optimized through multiple rounds of model training and verification.

[0110] Based on the sub-stage adaptability weights and operating condition sensitivity weights after dynamic matching, combined with the combination of influencing parameters, a time-series deep learning network is constructed to establish a nonlinear mapping from the combination of influencing parameters and mechanism parameters to performance parameters in each typical cluster group.

[0111] Temporal deep learning networks refer to deep neural network models adapted to the operating conditions of typical cluster groups. They take the combination of engine core component mechanism parameters and multi-dimensional influence parameters after sub-stage adaptability weights and operating condition sensitivity weights as input, and capture the evolution of the above parameters in the time series and the nonlinear correlation between multi-dimensional parameters.

[0112] The system consists of a temporal feature extraction layer, a weight-enhanced attention layer, a multi-dimensional fusion layer, and an output mapping layer. The temporal feature extraction layer is used to capture the dynamic changes of weighted mechanistic parameters within each takeoff sub-stage by adapting convolutional kernels of different sizes to the frequency of parameter changes in the weighted combination of mechanistic and influencing parameters within typical clusters. For example, in the taxiing sub-stage, it extracts the temporal features of the gradual increase in engine speed over time using adapted convolutional kernels; in the departure sub-stage, it extracts the stable instantaneous changes in thrust coefficient after a sudden increase. The weight-enhanced attention layer, based on the sub-stage adaptability weights corresponding to the takeoff sub-stages of typical clusters and the condition sensitivity weights corresponding to the operating condition characteristics of that cluster, specifically strengthens the feature contribution of the mechanistic and influencing parameters of the typical clusters in the current takeoff sub-stage, as well as the temporal sequence fragments formed by the changes of these parameters over time within that takeoff sub-stage. Simultaneously, it weakens the interference of non-critical parameters and temporal fragments, ensuring that the model focuses on parameter information that plays a decisive role in takeoff performance. The multi-dimensional fusion layer integrates the feature vectors of influence parameters of different dimensions and weighted mechanism parameters of different core components within typical clusters across dimensions. Through operations such as feature channel concatenation and weighted summation, it maps the parameter features originally scattered across different physical dimensions to a unified high-dimensional feature space, fully exploring the coupling correlation between parameters and forming a fusion feature that can comprehensively represent the operating condition. The output mapping layer, through a multi-layer fully connected neural network structure, accurately maps the high-dimensional feature vectors output by the multi-dimensional fusion layer to the carrier-based aircraft takeoff performance parameter space, and outputs continuous performance parameter prediction values ​​through a linear activation function, ultimately completing the nonlinear mapping transformation from the combination of influence parameters and weighted mechanism parameters to performance parameters.

[0113] This temporal deep learning network is primarily used to accurately establish nonlinear correlation models between the combination of influencing parameters, weighted mechanism parameters, and takeoff performance parameters in each typical cluster. It fully utilizes the temporal evolution information of parameters through a temporal feature extraction layer and leverages the guiding role of weight optimization results through a weight-enhanced attention layer, thereby significantly improving the prediction accuracy of performance parameters. At the same time, the parameter configurations of each layer of the network are adapted to the operating characteristics of the corresponding typical cluster, ensuring that the model output can fit the actual operating law of the engine and the real takeoff performance characteristics of the carrier-based aircraft under the operating conditions, providing core nonlinear mapping capabilities to support the clustered benchmark performance digital twin model.

[0114] For each typical cluster group, the temporal deep learning network is trained independently using corresponding training and validation sets. During training, an optimizer adapted to complex parameter mappings is selected, and the initial learning rate is set based on operating conditions. Overfitting is suppressed through weight decay. The loss function is a combination of mean squared error and mechanistic constraint terms. The mean squared error quantifies the prediction bias of performance parameters, while the mechanistic constraint term penalizes predictions that violate engine thermodynamics. The weights of the constraint terms are dynamically adjusted according to the operating condition type to enhance physical rationality. A multi-fold cross-validation mechanism is introduced during training, dividing the training set into multiple subsets for iterative validation and hyperparameter optimization. An early stopping mechanism is also enabled; when the validation set loss value shows no decrease and fluctuates steadily for several consecutive rounds, training is automatically terminated and a snapshot of the optimal model parameters is saved to prevent overtraining from affecting generalization ability. For models in extreme operating condition cluster groups, an additional regularization layer is added to further improve their adaptability to extreme parameter fluctuations.

[0115] The accuracy of the temporal deep learning network for each typical cluster group after training is verified using the test set. The mean relative error between the predicted and actual values ​​of the performance parameters is calculated as the performance parameter prediction error rate to evaluate the accuracy of the model in predicting key takeoff performance. If it exceeds the preset accuracy threshold, the root cause of the problem is analyzed and optimized: if the data quality is insufficient, real samples of the corresponding working conditions are added and the dataset is reprocessed; if the network structure has poor adaptability, the configuration of the temporal feature extraction layer or attention layer is adjusted; if the mechanism constraint is insufficient, the constraint terms are strengthened or the mechanism equation is added. After optimization, the network is retrained until the indicators meet the standards.

[0116] Otherwise, a time-series deep learning network with verified accuracy will be used as the benchmark performance digital twin model for its typical cluster groups. By combining the benchmark performance digital twin models of all typical cluster groups, a clustered benchmark performance digital twin model library will be constructed. At the same time, a working condition feature model index mapping table will be established, which records the working condition feature boundaries of each typical cluster group, the storage path of the corresponding model, the performance indicators, and the applicable sub-stage range.

[0117] Step S5: During the actual takeoff of the carrier-based aircraft, data on influencing parameters, engine measurement point parameters, and actual performance parameters are collected. The engine measurement point parameter data is converted into mechanistic parameter data, and typical clusters are matched based on the influencing parameter data to locate the baseline performance digital twin model and generate predicted values ​​for the performance parameters. A time series alignment algorithm is used to calibrate the actual and predicted values ​​of the performance parameters over time, constructing a complete performance parameter residual curve. The time series variation characteristics of the residuals are analyzed using a residual trend extraction algorithm to establish a multi-dimensional takeoff performance evaluation index system. Based on the preset performance evaluation thresholds and anomaly judgment rules, combined with the results of extreme operating condition identification, the residual curves and evaluation indicators are comprehensively analyzed to quantitatively evaluate the takeoff performance of the carrier-based aircraft.

[0118] In this embodiment, step S5 collects multi-dimensional operational data in real time and converts it into mechanistic parameters. It relies on typical cluster group matching to achieve accurate positioning of the benchmark performance digital twin model. Combined with time series alignment algorithm, it ensures the temporal consistency between the actual and predicted values ​​of performance parameters. Through residual curve construction, multi-dimensional statistical feature extraction and extreme condition correlation analysis, it not only realizes the quantitative and high-precision evaluation of the takeoff performance level of carrier-based aircraft, but also effectively identifies performance anomalies and degradation trends, providing accurate data support for takeoff safety assurance and maintenance decisions.

[0119] Preferably, the specific steps for quantitatively evaluating the takeoff performance of carrier-based aircraft include:

[0120] Please see Figure 3 Preprocessing operations are performed on the real-time acquired impact parameter data, engine measurement point parameter data, and actual performance parameter data. The wavelet threshold denoising algorithm, consistent with step S2, is used to filter out random noise and electromagnetic interference from the sensors. Abnormal data is eliminated based on the engine's physical limits and the reasonable range of performance parameters. Data gaps are filled using interpolation. Simultaneously, the mechanism model of the core component from step S3 is used to convert the preprocessed engine measurement point parameter data into corresponding mechanism parameter data. This ensures that the real-time mechanism parameters are consistent with the parameter format and physical meaning during the model training phase, laying a data foundation for model calling and performance comparison.

[0121] Based on the preprocessed impact parameter data, real-time operating condition features are extracted, including real-time values ​​of environmental parameters, dynamic range of control parameters, and current takeoff sub-stage identifier. The real-time operating condition features are compared with the operating condition feature model index mapping table of the clustered benchmark performance digital twin model library. The Euclidean distance algorithm is used to calculate the similarity between the real-time operating condition features and the operating condition feature boundaries of each typical cluster group in order to screen the target typical cluster group.

[0122] If the target typical cluster group is unique during the takeoff phase, obtain the baseline performance digital twin model of the target typical cluster group; input real-time impact parameter data and mechanism parameter data to generate a sequence of predicted values ​​for the corresponding performance parameters.

[0123] If the target typical clusters during takeoff are not unique, a sliding window statistical mechanism is introduced. The current takeoff phase is segmented at fixed time intervals, and the degree of conformity of the operating condition characteristics of each target typical cluster in each time period is statistically analyzed. The time periods in which the conformity meets the preset threshold are accumulated as the effective existence duration, and the proportion of effective existence duration of each target typical cluster is calculated. Based on the proportion of effective existence duration, the dominant target typical cluster with the highest proportion is selected, and its corresponding benchmark performance digital twin model is obtained. Real-time data is input to generate the dominant predicted value sequence of performance parameters. At the same time, the benchmark performance digital twin models of the remaining target typical clusters are obtained, and real-time data is input to generate the auxiliary predicted value sequences of performance parameters. A weighting coefficient is set based on the proportion of effective existence duration of each target typical cluster, and the dominant predicted value sequence and all auxiliary predicted value sequences are weighted and merged. During the merging process, a dynamic time warping algorithm is used to correct the temporal deviation of different sequences to ensure that the fused predicted value sequence is smooth and continuous. Finally, a performance parameter predicted value sequence that can comprehensively reflect the characteristics of the current complex operating conditions is obtained.

[0124] Please see Figure 4 The system employs a dynamic time warping algorithm to calibrate the actual and predicted performance parameter sequences in the time dimension, correcting time-series deviations caused by acquisition delays and dynamic changes in operating conditions, ensuring that the two are accurately aligned at the same time node. Based on the calibrated sequence, the system calculates the performance parameter residual value at each time node, constructs a complete performance parameter residual curve, and extracts the time-series change characteristics of the residual curve using the sliding window method, including residual trend characteristics, abrupt change characteristics, and fluctuation characteristics, to identify key time nodes of performance degradation or abnormal fluctuations.

[0125] Based on the residual curve and its temporal variation characteristics, multi-dimensional takeoff performance evaluation indicators are calculated, including: temporal consistency index, parameter deviation rate index, and sub-stage matching index. The temporal consistency index quantifies the consistency of the overall trend of the actual and predicted value sequences of performance parameters over time. It is obtained by calculating the correlation coefficient between the two, with the correlation coefficient reflecting the degree of trend consistency; the closer the value is to the positive extreme value, the more consistent the temporal variation patterns of the two. The parameter deviation rate index comprehensively measures the deviation between the actual and predicted values ​​of performance parameters. It is obtained by first calculating the relative error between the residual value and the corresponding actual value of each performance parameter at each time node, and then taking the average of the relative errors of all performance parameters, intuitively reflecting the overall deviation level between the predicted and actual values. The sub-stage matching index evaluates the matching effect of performance parameters within each takeoff sub-stage. It is obtained by calculating the parameter deviation rate and temporal consistency within each sub-stage, and then weighting and summing the two indicators based on preset sub-stage weights, highlighting the performance matching performance of key sub-stages.

[0126] Based on the requirements for safe takeoff operation of carrier-based aircraft, engine performance design specifications, and historical operational data statistics, a quantitative performance level judgment standard is established. This standard is used to classify the takeoff performance level of carrier-based aircraft according to the timing conformity index, parameter deviation rate index, and sub-stage matching degree index. For example, it is divided into four levels: Excellent, Good, Average, and Poor. The Excellent level has a high level of timing conformity, a low level of parameter deviation rate, and excellent matching degree scores in each sub-stage, indicating that the takeoff performance is highly consistent with the benchmark model and there are no potential risks. The Good level has good timing conformity, parameter deviation rate within a reasonable range, and matching degree scores in each sub-stage meet the standards, indicating that the takeoff performance is stable and reliable with only minor and controllable fluctuations. The Average level has average timing conformity, parameter deviation rate exceeds the lower level but does not break through the safety threshold, and the matching degree scores in some sub-stages are close to the standard line, indicating that the takeoff performance has some fluctuations and the trend of key parameter changes needs to be closely monitored. The Poor level does not meet the requirements of the Average level, or any core evaluation index exceeds the safety warning threshold, indicating that there is a significant risk of abnormality or degradation in takeoff performance.

[0127] This embodiment also introduces a carrier-based aircraft takeoff performance evaluation system based on digital twin empowerment, including a data acquisition module, a model scheduling module, a performance evaluation module, and a model training module;

[0128] The data acquisition module is responsible for collecting real-time data on influencing parameters, engine measurement point parameters, and actual performance parameters during the actual takeoff process of the carrier-based aircraft. It uses the same wavelet threshold denoising algorithm as the model training phase to filter out random noise and electromagnetic interference from sensors. Abnormal data is eliminated based on the engine's physical limits and reasonable performance parameter ranges. Data gaps are filled using interpolation. Simultaneously, the module uses the mechanism model of the engine's core components to convert the preprocessed engine measurement point parameter data into corresponding mechanism parameter data. This ensures that the real-time mechanism parameters are consistent with the parameter format and physical meaning in the model training phase, laying a standardized and high-quality data foundation for subsequent operating condition feature matching, benchmark performance digital twin model invocation, and performance comparison.

[0129] The model scheduling module extracts real-time operating condition features based on preprocessed impact parameter data, including real-time environmental parameter values, dynamic range of control parameters, and current takeoff sub-stage identifiers. These real-time operating condition features are compared with the operating condition feature model index mapping table in the clustered benchmark performance digital twin model library. The Euclidean distance algorithm is used to calculate the similarity between the real-time operating condition features and the operating condition feature boundaries of each typical cluster group, and target typical cluster groups are selected. If a target typical cluster group is unique, the corresponding benchmark performance digital twin model is directly obtained, and real-time data is input to generate a performance parameter prediction value sequence. If multiple target typical cluster groups are not unique, a sliding window statistical mechanism is used to calculate the effective existence time percentage of each target typical cluster group, selecting the dominant target typical cluster group and generating a dominant prediction value sequence. Simultaneously, prediction value sequences of the remaining auxiliary target typical cluster groups are obtained. These are combined with the effective existence time percentages to generate a weighted comprehensive prediction value sequence, achieving accurate scheduling of the benchmark performance digital twin model and reliable generation of prediction value sequences under complex operating conditions.

[0130] The performance evaluation module employs a dynamic time warping algorithm to calibrate the actual and predicted value sequences of performance parameters over time, correcting temporal deviations and constructing complete residual curves for the performance parameters. It then extracts residual trend characteristics, abrupt change characteristics, and fluctuation characteristics using a sliding window method. Based on the residual curves and their temporal variation characteristics, it calculates temporal fit indices, parameter deviation rate indices, and sub-stage matching indices. Combining the requirements for safe takeoff operation of carrier-based aircraft, engine performance design specifications, and historical operational data statistics, it evaluates the takeoff performance of carrier-based aircraft according to preset quantitative performance level judgment standards. Simultaneously, by analyzing the results of extreme condition identification and determining whether residual characteristics and evaluation indices exceed safety thresholds, it triggers corresponding level of anomaly warnings, identifies abnormal sub-stages, associated parameters, and risk consequences, and outputs a structured analysis report containing performance level, evaluation index details, anomaly warning information, and handling suggestions, providing data support for takeoff safety assurance and maintenance decisions.

[0131] The model training module, based on sampling parameter sets, health monitoring datasets, and mechanistic parameter datasets, constructs dedicated training datasets for each typical cluster group, including combinations of influence parameters, sub-stages, mechanistic parameters, and performance parameters. It employs a dual-mode fusion architecture of mechanism-driven and data-driven approaches, building a temporal deep learning network comprising a temporal feature extraction layer, a weight-enhanced attention layer, a multi-dimensional fusion layer, and an output mapping layer. Independent model training is performed on the associated training datasets for each typical cluster group. During training, a K-fold cross-validation mechanism is introduced to iteratively optimize hyperparameters. Overfitting is suppressed through weight decay and early stopping mechanisms. An additional regularization layer is added to the model for extreme operating condition cluster groups to improve adaptability. Model accuracy is verified using an independent test set. If the performance parameter prediction error rate exceeds a preset threshold, the dataset, network structure, or mechanistic constraints are backtracked and optimized until the model meets the target. All compliant models are integrated into a clustered benchmark performance digital twin model library, with a corresponding operating condition feature model index mapping table. The library is also supported for continuous iterative optimization based on new operating data, improving the model's adaptability and evaluation accuracy for various operating conditions.

[0132] Working principle and its effects:

[0133] This invention uses digital twin technology as the core link to construct a full-chain carrier-based aircraft takeoff performance evaluation system. By systematically integrating parameter space construction, high-fidelity data acquisition, mechanism parameter transformation, cluster model training, and real-time performance analysis, it solves the problems of insufficient accuracy, adaptability, and real-time performance evaluation of carrier-based aircraft under complex working conditions, and achieves closed-loop support from data to decision-making.

[0134] In the data and model foundation construction phase, this invention first integrates the normal and extreme value ranges of influencing parameters to form a complete parameter space. Combined with parameter priority, hierarchical guided sampling is adopted and its rationality is verified. This ensures that the sampled parameter group covers all operating conditions while eliminating invalid combinations, providing a high-quality data foundation for evaluation. Then, time-synchronized measurement point and performance data are collected through a simulation platform. After preprocessing and structured encapsulation, a health monitoring dataset is formed. At the same time, relying on the mechanism model of the engine core components, the measurement point parameters are transformed into mechanism parameters, effectively removing environmental interference and accurately capturing the real operating status of the core components, solving the problems of poor data quality and shallow performance characterization in traditional evaluation. In the cluster model construction and real-time evaluation stage, this invention generates typical cluster groups based on performance similarity and operating condition type. After extracting operating condition features, it dynamically matches the sub-stages of mechanism parameters and operating condition weights, and builds a dedicated deep learning network with layers such as temporal feature extraction and attention enhancement to form a clustered digital twin model library, ensuring model adaptability under different operating conditions. During actual takeoff, the target cluster group is matched with real-time operating condition features, and the performance level is divided by combining residual curves and multi-dimensional indicators. This not only improves the model's generalization ability but also achieves real-time accurate evaluation, solving the problems of poor adaptability and single evaluation dimensions of traditional single models.

[0135] In summary, this invention not only overcomes the shortcomings of traditional assessments in terms of extreme condition coverage, multi-parameter coupling analysis, and core performance characterization, but also significantly improves the accuracy, adaptability, and real-time performance of the assessment, providing scientific and reliable technical support for ensuring the safety of carrier-based aircraft takeoff and for operational decision-making.

[0136] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for evaluating carrier-based aircraft take-off performance based on digital twin empowerment, characterized in that, The application relates to a method for constructing a carrier-based aircraft engine health monitoring and performance prediction model. The method comprises the following steps: An influence parameter space of a carrier-based aircraft is constructed, and intelligent sampling is performed in the influence parameter space by using a hierarchical directed sampling method in combination with priority weights of the influence parameters to construct a sampling parameter group; A carrier-based aircraft takeoff simulation training platform is configured according to the sampling parameter group, engine measurement point parameter data and performance parameter data in the takeoff stage under each influence parameter combination are collected, and a health monitoring data set is constructed; Engine measurement point parameters of each data processing node in each sub-stage of the takeoff stage are converted into mechanism parameters based on mechanism models of each core component of the carrier-based aircraft engine; The sampling parameter group is clustered and divided based on the performance parameter similarity and the working condition type corresponding to each influence parameter combination, a typical cluster group and its associated training data set are constructed, working condition features of each typical cluster group are extracted, and then the sub-stage adaptability weight and the working condition sensitivity weight of each mechanism parameter are dynamically matched to build and train a time sequence deep learning network of each typical cluster group, and a clustered benchmark performance digital twin model library is constructed; 2. The digital twin enabled carrier aircraft launch performance evaluation method of claim 1, wherein, In the actual takeoff process of the carrier-based aircraft, a target typical cluster group is selected based on the influence parameter data in the takeoff stage, a performance parameter prediction value sequence is generated by using the benchmark performance digital twin model of the target typical cluster group, a performance parameter residual curve is constructed by combining an actual performance parameter value sequence, a multi-dimensional takeoff performance evaluation index is calculated, and a carrier-based aircraft takeoff performance level is divided. The step of constructing the sampling parameter group comprises the following steps: Regular value intervals of influence parameters related to carrier-based aircraft takeoff performance evaluation are obtained, and extreme value intervals of each influence parameter are set, wherein the influence parameters include environmental parameters and control parameters; The regular value intervals and the extreme value intervals of each influence parameter are fused to construct an influence parameter space; Priority levels of each influence parameter are set, and intelligent sampling is performed in the influence parameter space by using a hierarchical directed sampling method according to the priority levels of each influence parameter, namely: Sampling quotas of each priority level are configured, regular sampling points are set in combination with the regular value intervals of each influence parameter, and a sampling interval is dynamically adjusted according to the influence parameter change sensitivity; The influence parameter change sensitivity refers to the change degree of the carrier-based aircraft takeoff performance index caused by the unit change of the influence parameter; Extreme sampling points are set by using a combination of boundary encryption sampling and key node locking for the extreme value intervals of each influence parameter, and a preliminary sampling parameter set is constructed in combination with the regular sampling points; 3. The digital twin enabled carrier aircraft launch performance evaluation method of claim 1, wherein, Each influence parameter combination in the preliminary sampling parameter set is verified for rationality, and the influence parameter combinations that fail to pass the rationality verification are removed, and finally the sampling parameter group is formed. The step of constructing the health monitoring data set comprises the following steps: The carrier-based aircraft takeoff simulation training platform is configured according to each influence parameter combination of the sampling parameter group, the environmental parameter values in the influence parameter combination are used for fixed configuration for the environmental parameters, and dynamic value sequences of each control parameter are generated by taking the control parameter values in the influence parameter combination as target values of the takeoff stage of the current simulation training. The standard takeoff process of the carrier-based aircraft takeoff simulation training platform is started, engine measurement point parameter data and performance parameter data in the takeoff phase are collected at a preset sampling frequency, and a time tag is labeled to form a simulation time series data sequence; After performing multi-dimensional data preprocessing operations on the simulation time series data sequence, the data is structured and an identification tag is added, which includes influence parameter combination encoding, training timestamp, and working condition type; the working condition type includes normal working condition and extreme working condition; According to the three-dimensional data structure of the influence parameter combination, parameter type, and collection time tag, the association mapping relationship between the engine measurement point parameter data and the corresponding performance parameter data is established, and the standardization data unit is integrated; and all the standardization data units are integrated to form a health monitoring data set.

4. The digital twin based empowered carrier aircraft launch performance evaluation method as claimed in claim 1, wherein, The step of converting the engine measurement point parameters of each sub-phase data processing node in the takeoff phase into mechanism parameters includes: Divide the sub-phases of the takeoff phase of the standard takeoff simulation training under each influence parameter combination, and divide the data processing nodes of the simulation time series data sequence of each standardization data unit in the health monitoring data set of each standardization takeoff simulation training according to the time interval boundaries of each sub-phase and the preset parameter collection period; Establish the mapping relationship between the engine measurement point parameters and the mechanism model of the core component, extract the measurement point parameters of each data processing node according to the engine core component type, and obtain the corresponding mechanism parameters through the mechanism model of the corresponding core component; The mechanism model refers to a mathematical and physical model that can quantitatively describe the internal relationship between the input engine measurement point parameters and the output mechanism parameters, which is constructed based on the physical structure of the core component, the aerodynamic thermodynamic principle and the operating constraint condition; Configure the standard value range of each mechanism parameter, and check each mechanism parameter one by one to identify and mark the abnormal mechanism parameters of the abnormal core components, abnormal sub-phases, and abnormal data processing node numbers that exceed the standard value range; Obtain the engine measurement point parameter data of the non-abnormal data processing nodes in the abnormal sub-phase of the abnormal mechanism parameter; according to the mechanism parameter change trend, the abnormal data processing node of the abnormal mechanism parameter is interpolated and corrected, and then the corresponding mechanism parameter is corrected; Bind the corrected mechanism parameters of each core component with the corresponding data processing node, the identification of the sub-phase, the influence parameter combination encoding, and the working condition type tag to form a mechanism parameter data set.

5. The digital twin enabled carrier aircraft launch performance evaluation method of claim 1, wherein, The step of constructing a typical cluster group includes: Extract all influence parameter combinations in the sampling parameter group, and associate the corresponding performance parameters and working condition types of each influence parameter combination one by one to construct a clustering original data set containing influence parameter combination encoding, performance parameter set, and working condition type; According to the working condition type, the clustering original data set is divided to generate a clustering sub-data set, including a normal clustering sub-data set and an extreme clustering sub-data set; The performance parameters in each clustering sub-data set are standardized to map all performance parameters to the same numerical interval; for each clustering sub-data set, the number of clusters is optimized by using the silhouette coefficient method, and the silhouette coefficient of the clustering result corresponding to each cluster number is calculated to determine the optimal cluster number; For each clustering sub-data set, the clustering center is initialized based on the optimal cluster number, and the clustering calculation method is set, and the iterative clustering operation is started to generate the clustering cluster corresponding to each clustering sub-data set; The similarity of the performance parameters corresponding to each combination of influence parameters in each clustering cluster is calculated, and a similarity threshold is set to eliminate redundant influence parameter combination samples in the clustering cluster, and a typical clustering group is constructed.

6. The digital twin enabled carrier aircraft launch performance assessment method of claim 1, wherein, The step of constructing the library of clustered benchmark performance digital twin models comprises: For each typical clustering group, the corresponding mechanism parameters, performance parameters and influence parameter combinations are associated to construct an associated training data set and perform normalization preprocessing; the associated training data set of each typical clustering group is divided into a training set, a validation set and a test set by using a stratified sampling method; Based on the distribution of the influence parameter value range and the performance parameter distribution in the associated training data set of the typical clustering group, the working condition characteristics of each typical clustering group are extracted; the sub-stage adaptability weight and the working condition sensitivity weight of each mechanism parameter are dynamically matched through a preset working condition characteristic mechanism parameter association matrix; The working condition characteristic mechanism parameter association matrix refers to a three-dimensional structured matrix constructed with the working condition characteristics of the typical clustering group as rows, the mechanism parameters of the engine core components as columns, and the sub-stages as dimensions, and the matrix elements are the association strength coefficients of the corresponding working condition characteristics and mechanism parameters in the corresponding take-off sub-stage; Based on the dynamically matched sub-stage adaptability weight and working condition sensitivity weight, a time series deep learning network is constructed for establishing a nonlinear mapping from the influence parameter combination and the mechanism parameter to the performance parameter in each typical clustering group; For each typical clustering group, the training set and the validation set are used to carry out independent training, and the test set is used to verify the accuracy of each trained typical clustering group, and the relative error mean of the performance parameter prediction value and the true value is calculated as the performance parameter prediction error rate; The time series deep learning network that passes the accuracy verification is used as the benchmark performance digital twin model of the typical clustering group, and all the benchmark performance digital twin models of the typical clustering groups are combined to construct a library of clustered benchmark performance digital twin models.

7. The digital twin based empowered carrier aircraft launch performance evaluation method as claimed in claim 6, wherein, The time series deep learning network comprises a time series feature extraction layer, a weight-enhanced attention layer, a multi-dimensional fusion layer and an output mapping layer; The time series feature extraction layer is used to adapt different size convolution kernels to the change frequency of the weighted mechanism parameters and influence parameter combinations in the typical clustering group to capture the dynamic change characteristics of the mechanism parameters in each take-off sub-stage; The weight-enhanced attention layer is used to strengthen the feature contribution of the mechanism parameters, influence parameters and their time series fragments in the current take-off sub-stage based on the sub-stage adaptability weight and the working condition sensitivity weight of the corresponding take-off sub-stage of the typical clustering group. The multi-dimensional fusion layer is configured to integrate the feature vectors of the influence parameters and the mechanism parameters in each typical clustering group across dimensions, map to a unified high-dimensional feature space, and form a high-dimensional feature vector. The output mapping layer is configured to map the high-dimensional feature vector output by the multi-dimensional fusion layer to a carrier aircraft takeoff performance parameter space through a multi-layer fully connected neural network structure, and output a continuous performance parameter prediction value by means of a linear activation function.

8. The digital twin enabled carrier aircraft launch performance evaluation method of claim 1, wherein, The step of generating the prediction value sequence of the performance parameter includes: Performing preprocessing operations on the real-time collected influence parameter data, engine measurement point parameter data and actual performance parameter data, and converting the preprocessed engine measurement point parameter data into corresponding mechanism parameter data; Based on the preprocessed influence parameter data, extracting real-time working condition features, comparing the real-time working condition features with the working condition feature model index mapping table of the clustered reference performance digital twin model library, calculating the similarity of the real-time working condition features and the working condition feature boundaries of each typical clustering group, and screening the target typical clustering group; If the target typical clustering group in the takeoff stage is unique, the reference performance digital twin model of the target typical clustering group is obtained; the real-time influence parameter data and the mechanism parameter data are inputted to generate the prediction value sequence of the corresponding performance parameter; If the target typical clustering group in the takeoff stage is not unique, the current takeoff stage is segmented at fixed time intervals, the working condition feature coincidence degree of each target typical clustering group in each time period is counted, the effective existence time length proportion of each target typical clustering group is calculated, the dominant target typical clustering group is screened out, the corresponding reference performance digital twin model of the dominant target typical clustering group is obtained, and the real-time data are inputted to generate the dominant prediction value sequence of the performance parameter; meanwhile, the reference performance digital twin models of the remaining target typical clustering groups are obtained, and the auxiliary prediction value sequences of the performance parameter are generated respectively; the effective existence time length proportion of each target typical clustering group is set as a weighting coefficient, and the dominant prediction value sequence and all auxiliary prediction value sequences are weighted and combined to obtain the performance parameter prediction value sequence.

9. The digital twin enabled carrier aircraft launch performance evaluation method of claim 1, wherein, The step of dividing the carrier aircraft takeoff performance level includes: Calibrating the performance parameter actual value sequence and the prediction value sequence in the time dimension, calculating the performance parameter residual value of each time node, and constructing a performance parameter residual curve; Extracting the time sequence variation characteristics of the residual curve by the sliding window method, calculating the multi-dimensional takeoff performance evaluation indexes based on the residual curve and the time sequence variation characteristics, including: the time sequence coincidence degree index, the parameter deviation rate index and the sub-stage matching degree index; The time sequence coincidence degree index is obtained by calculating the correlation coefficient of the performance parameter actual value sequence and the prediction value sequence; the parameter deviation rate index is obtained by calculating the relative error of each performance parameter at each time node and the corresponding actual value, and taking the mean value of the relative errors of all performance parameters; the sub-stage matching degree is obtained by calculating the parameter deviation rate and the time sequence coincidence degree in each sub-stage, and then combining the preset sub-stage weight to weight and sum the two indexes; The quantitative performance level judgment standard is set, and the carrier aircraft takeoff performance level is divided in combination with the time sequence coincidence degree index, the parameter deviation rate index and the sub-stage matching degree index.

10. A digital twin enabled aircraft launch performance evaluation system for implementing the digital twin enabled aircraft launch performance evaluation method of any one of claims 1-9, wherein, The application relates to a carrier-based aircraft performance prediction and evaluation system, which comprises a data acquisition module, a model scheduling module, a performance evaluation module and a model training module. The data acquisition module is used for collecting influence parameter data, engine measuring point parameter data and actual performance parameter data during an actual takeoff process of a carrier-based aircraft; the model scheduling module extracts real-time working condition characteristics based on the influence parameter data, screens a target typical clustering group, and generates a prediction value sequence of the performance parameter through a benchmark performance digital twin model of the target typical clustering group; The performance evaluation module is used for combining an actual value sequence of the performance parameter, constructing a performance parameter residual error curve, calculating a multi-dimensional takeoff performance evaluation index, and dividing a carrier-based aircraft takeoff performance level; The model training module is used for constructing a typical clustering group and an associated training data set based on a sampling parameter group, a health monitoring data set and a mechanism parameter data set, building and training a time sequence deep learning network of each typical clustering group, and constructing a clustered benchmark performance digital twin model library.

Images