Aircraft mechanism model parameter migration error evolution method under complex constraints
By constructing a long short-term memory network based on attention mechanism and batch normalization, and a hierarchical model switching strategy, combined with the least squares criterion, the problem of identifying and estimating parameter migration errors of aircraft mechanism models under complex constraints was solved, achieving high-precision and low-complexity state estimation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NANJING UNIV OF AERONAUTICS & ASTRONAUTICS
- Filing Date
- 2026-03-11
- Publication Date
- 2026-05-12
AI Technical Summary
Existing technologies struggle to accurately characterize the temporal evolution of parameter migration errors in aircraft mechanics models under complex constraints, leading to decreased estimation accuracy and increased computational complexity. Traditional methods are prone to estimation bias or divergence in nonlinear and non-Gaussian environments.
A mechanism model parameter transfer error identification model is constructed using a long short-term memory network based on attention mechanism and batch normalization. Through hierarchical structure and model switching strategy, combined with the least squares criterion, joint state estimation is performed to identify and screen high-confidence models for state estimation.
It improves the accuracy and training efficiency of identifying parameter migration errors in mechanistic models, reduces computational overhead, ensures the real-time performance and accuracy of aircraft state estimation under complex constraints, and avoids the computational explosion and reduced estimation accuracy problems of traditional methods.
Smart Images
Figure CN121808952B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of aerospace and information processing technology, specifically relating to a method for parameter migration error evolution of aircraft mechanism models under complex constraints. Background Technology
[0002] With the continuous development of aerospace technology, various aircraft exhibit diverse flight modes and complex flight environments during mission execution, resulting in uncertainties in their motion states in both time and space. Achieving continuous estimation and trajectory tracking of aircraft motion states under multiple constraints or complex operating environments is a crucial technological foundation for applications such as flight control, operational monitoring, and mission management. In these applications, aircraft state estimation and prediction typically rely on accurate descriptions of mechanistic models and their parameters.
[0003] In actual flight, aircraft are often affected by various factors such as environmental disturbances and control strategy adjustments, making it difficult for the pre-established mechanistic model parameters to remain consistent with the real system state over a long period of time. This results in the mechanistic model parameters shifting over time, producing mechanistic model parameter migration error.
[0004] To address the aforementioned issues, existing technologies typically employ multi-model estimation methods based on mechanistic models, constructing multiple candidate models to cover different parameter variation scenarios. However, when the parameter migration error of the mechanistic model changes significantly over time, a large number of models are often required to ensure estimation accuracy. This leads to a rapid increase in the size of the model set and a significant increase in computational complexity, making it difficult to meet real-time processing requirements. In practical applications, artificially reducing the number of models can easily introduce model mismatch problems, resulting in a decrease in state estimation accuracy.
[0005] With the development of data-driven methods, deep learning models such as Long Short-Term Memory (LSTM) networks have been introduced for time series feature extraction and pattern recognition, which has improved the ability to recognize complex motion patterns to some extent. However, when faced with the diverse maneuvering behaviors and parameter variation characteristics of aircraft, traditional LSTM networks still suffer from problems such as insufficient attention to key temporal features, low training efficiency, and susceptibility to overfitting, making it difficult to stably and accurately characterize the temporal evolution of parameter transfer errors in mechanistic models.
[0006] Furthermore, in the state estimation stage, existing methods are mostly based on the Gaussian noise assumption, such as extended Kalman filtering. However, under complex constraints, aircraft systems are often accompanied by state-related uncertainties. These uncertainties are coupled with the system state, and traditional filtering methods are prone to estimation bias or even divergence in such nonlinear, non-Gaussian environments. Although particle filtering can alleviate these problems to some extent, its computational cost increases significantly with the sampling scale, making it difficult to effectively promote in real-time online applications.
[0007] In summary, existing technologies still suffer from limitations in addressing parameter migration error problems in aircraft mechanism models under complex constraints. These limitations include limited ability to characterize the evolution of parameter migration errors, high redundancy in model set computation, and difficulty in balancing estimation accuracy and real-time performance in complex and uncertain environments. Summary of the Invention
[0008] The purpose of this invention is to provide a method for the evolution of parameter migration error in aircraft mechanics models under complex constraints, so as to solve the above-mentioned problems.
[0009] To achieve the above objectives, the present invention provides the following technical solution: a method for parameter migration error evolution of an aircraft mechanism model under complex constraints, the specific steps of which are as follows:
[0010] S1. Construct a parameter migration error dataset for the mechanism model using aircraft simulation data. Then, by performing differential processing and kinematic feature extraction on the three-dimensional position and attitude measurement information of the aircraft at discrete moments, construct a preprocessed feature sequence containing auxiliary feature terms such as velocity and curvature, forming a time-series feature sequence that can characterize the evolution characteristics of the parameter migration error of the aircraft under complex constraints.
[0011] S2. The temporal evolution characteristics of parameter transfer error in the mechanism model are modeled and trained by introducing an attention mechanism and batch normalization long short-term memory network to obtain a deep learning recognition model for subsequent parameter transfer error identification.
[0012] S3. Construct a hierarchical mechanism model parameter transfer error evolution identification process using the trained deep learning recognition model. Decompose the complex parameter transfer error identification problem into multi-level sub-problems and output the mechanism model parameter transfer error model and the mechanism model parameter transfer error confidence characterization result based on the mechanism model parameter transfer error model. The mechanism model parameter transfer error model is a pre-constructed candidate mechanism model under different parameter offset conditions.
[0013] S4. Based on the confidence characterization results of the mechanism model parameter migration error obtained in S3, a model switching strategy is adopted to select a set of models for state estimation from the preset mechanism model parameter migration error model set.
[0014] A nonlinear state-space model is constructed, which includes multiplicative mechanistic stochastic parameters caused by the parameter migration error of the mechanistic model. The multiplicative mechanistic stochastic parameters are introduced by the parameter migration error of the corresponding model and are used to characterize the proportional offset of the mechanistic model parameters relative to the nominal value. Using the residual between the predicted output and the measured output of the nonlinear state-space model as the objective function, a joint state parameter estimation method is constructed using the least squares criterion to perform joint estimation of the system state and mechanistic model parameters, and output the state estimation result of the aircraft.
[0015] Preferably, S1 specifically includes the following steps: S11: Based on aircraft simulation data, an aircraft motion simulation model is established, generating a simulation dataset that simultaneously includes tracks for stable flight and maneuvering flight. This dataset contains multiple tracks covering various flight states, including turning maneuvers, attitude adjustment maneuvers, and periodic complex maneuvers, used to describe the aircraft's motion behavior characteristics under complex constraints. The dataset is then divided into training and validation sets according to a preset ratio for subsequent training and performance evaluation of the recognition network.
[0016] S12: Define discrete time intervals The following measurement information is The discrete time interval is The difference operation is performed on the measurement information at adjacent time points to obtain... The target velocity is calculated based on the changes in its three-dimensional position coordinates between adjacent time points.
[0017] ;
[0018] in, , , They represent the target at the 1st and 2nd positions, respectively. Spatial coordinates at a given time.
[0019] S13: To enhance the characterization capability of aircraft maneuvering characteristics and the evolution law of parameter migration errors in the mechanism model, attitude angle change information is introduced on the basis of velocity information to construct auxiliary feature terms:
[0020] ;
[0021] in, This indicates the change in yaw angle. This represents the change in track inclination, and combines the original measurement information, difference terms, velocity information, and auxiliary feature terms to form a preprocessed feature sequence. It is used as input for the subsequent mechanism model parameter transfer error identification network.
[0022] Preferably, step S2 specifically includes the following steps:
[0023] S21: To address the strong temporal correlation and nonlinear characteristics of the evolution of mechanism model parameter transfer error over time, a long short-term memory recognition network Att-LSTM+BN based on attention mechanism and batch normalization is constructed. The network consists of a fully connected layer, a batch normalization layer, a long short-term memory layer, an attention layer, and an output layer connected in sequence, which is used to learn and represent the characteristics of aircraft trajectory type and mechanism model parameter transfer error.
[0024] S22: Set network structure parameters according to different recognition tasks, including the track type recognition network RNNt and the mechanism model parameter recognition network RNN. 1, SFS RNN 2, SFS RNN 1, MFS RNN 2, MFS The number of neurons in the fully connected layer was set to 128 and 16, and the number of neurons in the long short-term memory layer was set to 128 and 64; the lateral maneuvering recognition network RNN d The number of neurons in the fully connected layer was set to 128 and 16, and the number of neurons in the long short-term memory layer was set to 512 and 256 to enhance the ability to model complex lateral maneuvering features.
[0025] S23: Set the network training parameters. The length N of the input feature sequence is 150 or 200. Use the cross-entropy loss function as the training objective function. Select RMSprop as the optimizer and set the learning rate to 0.001. The track type recognition network is trained for 20 iterations, and the other recognition networks are trained for 80 iterations.
[0026] S24: Use the training and validation sets constructed in S1 to train and validate each recognition network. Optimize the network parameters by monitoring the changes in the loss function and recognition accuracy to obtain a trained network model for identifying the mechanism model parameter transfer error evolution.
[0027] Preferably, the specific steps of S3 are as follows:
[0028] S31: The preprocessed feature sequence obtained in S1 Input the trajectory type recognition network RNNt to identify the flight status label L of the target at the current moment. t Simultaneously, the feature sequence is input into the lateral maneuver recognition network RNN. d Identify the target's lateral movement direction label L d This is to enable a preliminary determination of the overall motion pattern of the aircraft;
[0029] S32: Based on flight status label L tThe recognition results are used for branch judgment. When a stable flight state is determined, the corresponding mechanism model parameter recognition subnetwork RNN is invoked. 1, SFS and RNN 2, SFS The mechanism model parameter labels L1 and L2 are identified respectively; when a maneuvering flight state is determined, the corresponding mechanism model parameter identification subnetwork RNN is invoked. 1,MFS and RNN 2, MFS The mechanism model parameter labels L1 and L2 are identified respectively, thereby enabling the identification of differentiated parameter migration errors for different track types;
[0030] S33: Based on each recognition network, the measurement feature sequence is processed independently and the results are output. The confidence scores of the labels corresponding to the flight status label, lateral maneuver label and mechanism model parameter label are fused to determine the confidence score distribution of the target at the current time for each mechanism model parameter migration error model Mj, which is used for subsequent model switching and state estimation processes.
[0031] Preferably, the specific steps of S4 are as follows:
[0032] S41: Establish a system state variable Multiplicative mechanism model parameters The nonlinear discrete-time state-space model integrates the system state and mechanism model parameters into the state-space modeling framework, and describes the influence of mechanism model parameter migration error on system state evolution by introducing parameter deviation terms.
[0033] S42: For the above state-space model, a least-squares cost function based on measurement residuals is constructed. By minimizing the weighted sum of squared residuals between the system's predicted output and the actual measurement data, the joint estimation of the system state and mechanism model parameters is achieved.
[0034] S43: Using an alternating least squares optimization strategy, the system state is estimated under the condition of fixed mechanism model parameters, and the mechanism model parameters are updated under the condition of fixed system state. The least squares cost function is iteratively optimized until the preset convergence condition is met, and the system state estimation result at the current time is obtained.
[0035] Preferably, it also includes S5:
[0036] S51. Based on the confidence distribution results of the migration error models Mj of each mechanism model obtained in step S3, the models are sorted according to their confidence levels, and a model switching strategy is adopted to determine the set of models participating in state estimation at the current time. The model switching strategy is adjusted based on the confidence weight of each model to suppress the participation of low-confidence models, or adaptively determines the set of models participating in state estimation based on the cumulative confidence weight criterion.
[0037] S52: For the set of models determined in S51, the joint state parameter estimation method described in S4 is executed respectively. At each time step, the system state and mechanism model parameters are estimated in a coordinated manner to obtain the aircraft state estimation results under different model assumptions. The state estimation process of each model is independent of each other and can be calculated in parallel.
[0038] S53: Based on the confidence weights corresponding to the migration error models of each mechanism model parameters, the state estimation results obtained under different models are fused by weighted least squares to form the comprehensive state estimation result at the current moment.
[0039] The technical effects and advantages of this invention are as follows:
[0040] 1. This invention designs a long short-term memory recognition network based on attention mechanism and batch normalization. The key features on the time series are dynamically weighted by the attention mechanism, and the neuron input distribution is standardized by the batch normalization layer. This effectively solves the problems of slow convergence speed and easy overfitting of traditional recurrent neural networks when dealing with complex strategic maneuvers of aircraft, and significantly improves the accuracy and training efficiency of the identification of mechanism model parameter transfer error.
[0041] 2. The present invention discloses a method for the evolution of parameter migration error of aircraft mechanism model under complex constraints. Combining the identification results of parameter migration error of mechanism model, a dynamic model switching strategy is proposed. Based on the high confidence parameter migration error results identified in real time, a few models can be selected from a large set of models for state estimation. Compared with the traditional multi-model algorithm, the computational cost is greatly reduced while ensuring the estimation accuracy, and the computational explosion problem caused by large-scale model set is effectively solved.
[0042] 3. This invention discloses a method for parameter migration error evolution of an aircraft mechanics model under complex constraints.
[0043] By identifying and modeling the parameter migration error of the aircraft mechanism model, a nonlinear state-space model incorporating multiplicative mechanism model parameters is constructed based on model switching. A joint state estimation method based on the least squares criterion is then employed, with the error between the system's predicted output and actual observation data as the optimization objective, to achieve stable estimation of the aircraft system state. This method effectively suppresses the impact of uncertainties in multiplicative mechanism model parameters and non-Gaussian disturbances on state estimation accuracy, avoiding the problem of reduced or even divergent estimation accuracy in traditional Kalman filtering methods under strongly nonlinear and complex noise environments. Furthermore, compared to state estimation algorithms based on random sampling, such as particle filtering, this invention significantly reduces computational complexity and processing time while maintaining estimation accuracy, exhibiting good real-time performance and making it suitable for applications such as online state estimation and state monitoring of aircraft.
[0044] 4. The present invention discloses a method for the evolution of parameter migration error of aircraft mechanism model under complex constraints. It constructs a data preprocessing mechanism that includes auxiliary feature terms such as velocity, curvature and track angle change rate, and adopts a hierarchical recognition structure to identify track type and motion feature parameters respectively. By enhancing the representation ability of input features and decoupling the recognition task, it further improves the ability to capture the evolution law of parameter migration error of aircraft mechanism model under complex constraints. Attached Figure Description
[0045] Figure 1 This is a flowchart of the parameter migration error evolution method for aircraft mechanism model under complex constraints according to the present invention;
[0046] Figure 2 This is a schematic diagram of the internal structure of the long short-term memory neural network unit of the present invention;
[0047] Figure 3 The figure shows a comparison of the state estimation performance of the method of the present invention and the particle filter algorithm. Detailed Implementation
[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0049] This invention provides, for example Figures 1-3 The method for the evolution of parameter migration error of aircraft mechanism model under complex constraints is shown in the figure. Step 1: Construct a parameter migration error dataset of mechanism model using aircraft simulation data, and form a time-series feature sequence that can characterize the evolution characteristics of aircraft parameter migration error under complex constraints by performing differential and feature enhancement processing on the measurement information.
[0050] Step 1.1: Establish an aircraft motion simulation model based on aircraft simulation data, and generate a simulation dataset that includes both steady flight state tracks and maneuvering flight state tracks. The dataset contains a total of 11,520 tracks, covering various flight states such as turning maneuvers, attitude adjustment maneuvers, and periodic complex maneuvers, and is used to describe the motion behavior characteristics of the aircraft under complex constraints.
[0051] Step 1.2: Define discrete time intervals The following measurement information is The discrete time interval is The difference operation is performed on the measurement information at adjacent time points to obtain... The target velocity is calculated based on the changes in its three-dimensional position coordinates between adjacent time points.
[0052] ;
[0053] in, , , They represent the target at the 1st and 2nd positions, respectively. Spatial coordinates at a given time.
[0054] Step 1.3: To enhance the characterization ability of the aircraft's maneuvering characteristics and the evolution law of parameter migration errors in the mechanism model, attitude angle change information is introduced on the basis of velocity information to construct auxiliary feature terms:
[0055] ;
[0056] in, This indicates the change in yaw angle. This represents the change in track inclination, and combines the original measurement information, difference terms, velocity information, and auxiliary feature terms to form a preprocessed feature sequence. It is used as input for the subsequent mechanism model parameter transfer error identification network.
[0057] Step 2: The temporal evolution characteristics of parameter transfer error in the mechanistic model are modeled and trained using an attention mechanism and a batch-normalized long short-term memory network to obtain a deep learning recognition model for subsequent parameter transfer error identification.
[0058] Step 2.1: To address the strong temporal correlation and nonlinear characteristics of the evolution of mechanism model parameter transfer errors over time, a long short-term memory recognition network Att-LSTM+BN based on attention mechanism and batch normalization is constructed. The network consists of a fully connected layer, a batch normalization layer, a long short-term memory layer, an attention layer, and an output layer connected in sequence, which is used to learn and represent the characteristics of aircraft trajectory types and mechanism model parameter transfer errors.
[0059] Step 2.2: Set the network structure parameters according to different recognition tasks, including the track type recognition network RNN. t and Mechanism Model Parameter Recognition Network RNN 1, SFS RNN 2, SFS RNN 1, MFS RNN 2, MFS The number of neurons in the fully connected layer was set to 128 and 16, and the number of neurons in the long short-term memory layer was set to 128 and 64; the lateral maneuvering recognition network RNN d The number of neurons in the fully connected layer was set to 128 and 16, and the number of neurons in the long short-term memory layer was set to 512 and 256 to enhance the ability to model complex lateral maneuvering features.
[0060] Step 2.3: Set the network training parameters. The length N of the input feature sequence is 150 or 200. The cross-entropy loss function is used as the training objective function. The optimizer is RMSprop and the learning rate is set to 0.001. The track type recognition network is trained for 20 iterations and the other recognition networks are trained for 80 iterations.
[0061] Step 2.4: Use the training and validation sets constructed in Step 1 to train and validate each recognition network. Optimize the network parameters by monitoring the changes in the loss function and recognition accuracy to obtain a trained network model for the recognition of mechanistic model parameter transfer error evolution.
[0062] Step 3: Utilize the trained recognition network to construct a hierarchical mechanism model parameter transfer error evolution recognition process, decompose the complex parameter transfer error recognition problem into multi-level sub-problems, and output the confidence characterization results of the mechanism model parameter transfer error.
[0063] Step 3.1: Process the preprocessed feature sequence obtained in Step 1. The target's flight status label Lt is identified by inputting the trajectory type identification network RNNt. At the same time, the feature sequence is input into the lateral maneuver identification network RNNd to identify the target's lateral maneuver direction label Ld, so as to achieve a preliminary judgment of the overall motion mode of the aircraft.
[0064] Step 3.2: Based on the flight status label L t The recognition results are used for branch judgment. When a stable flight state is determined, the corresponding mechanism model parameter recognition subnetwork RNN is invoked. 1, SFS and RNN 2, SFS The mechanism model parameter labels L1 and L2 are identified respectively; when a maneuvering flight state is determined, the corresponding mechanism model parameter identification subnetwork RNN is invoked. 1,MFS and RNN 2, MFS The mechanism model parameter labels L1 and L2 are identified respectively, thereby enabling the identification of differentiated parameter migration errors for different track types.
[0065] Step 3.3: Based on each identification network, the measurement feature sequence is processed independently and the results are output. The confidence scores of the labels corresponding to the flight status label, lateral maneuver label and mechanism model parameter label are fused to determine the confidence score distribution of the target at the current time for each mechanism model parameter migration error model Mj, which is used for subsequent model switching and state estimation processes.
[0066] Step 4: Construct a nonlinear aircraft state estimation method based on the least squares criterion to achieve joint estimation of system state and model parameters under the conditions of mechanistic model parameter migration error and multiplicative stochastic uncertainty.
[0067] Step 4.1: Establish a system containing state variables Multiplicative mechanism model parameters The nonlinear discrete-time state-space model integrates the system state and mechanism model parameters into the state-space modeling framework, and introduces parameter deviation terms to describe the influence of mechanism model parameter migration error on the system state evolution.
[0068] Step 4.2: For the above state-space model, construct a least-squares cost function based on measurement residuals. By minimizing the weighted sum of squared residuals between the system's predicted output and the actual measurement data, the joint estimation of the system state and mechanism model parameters can be achieved.
[0069] Step 4.3: Using an alternating least squares optimization strategy, the system state is estimated under the condition of fixed mechanism model parameters, and the mechanism model parameters are updated under the condition of fixed system state. The least squares cost function is iteratively optimized until the preset convergence condition is met, and the system state estimation result at the current time is obtained.
[0070] Step 5: Integrate the parameter migration error identification results of the fusion mechanism model with the iterative state estimation method based on residual minimization, and construct an aircraft state estimation and result fusion process based on the model switching strategy to reduce computational complexity while ensuring estimation accuracy.
[0071] Step 5.1: Based on the confidence distribution results of the migration error model Mj of each mechanism model obtained in Step 3, the candidate models are sorted according to their confidence levels, and a model switching strategy is adopted to determine the set of models participating in state estimation at the current time. The model switching strategy is to adjust the fusion weights corresponding to the model confidence levels to suppress the participation of low-confidence models, or to adaptively determine the set of models that make effective contributions to state estimation based on the cumulative confidence weight criterion.
[0072] Step 5.2: For the set of models determined in Step 5.1, execute the joint state-parameter estimation method based on weighted residual minimization described in Step 4. At each time step, perform joint estimation of the system state and mechanism model parameters to obtain the aircraft state estimation results under different model assumptions. The state estimation processes of each model are independent of each other and can be computed in parallel.
[0073] Step 5.3: Based on the confidence weights corresponding to the migration error models of each mechanism model, perform weighted least squares fusion on the state estimation results obtained under different models to form the comprehensive state estimation result at the current moment.
[0074] Step 6: Verify the state estimation accuracy and computational efficiency of the method under complex constraints and non-Gaussian multiplicative noise conditions through numerical simulation, and evaluate the effectiveness and practicality of the method.
[0075] Step 6.1: Construct an aircraft simulation environment incorporating nonlinear dynamics and non-Gaussian multiplicative noise, and perform Monte Carlo simulations to verify the method. By setting different simulation runs, the stability and robustness of the algorithm under different stochastic conditions are evaluated, and the particle filtering method (PF) is selected as a comparison algorithm. The particle filtering method approximates the posterior probability distribution of the system state using a finite number of particles, and utilizes weight update and resampling mechanisms to achieve state estimation.
[0076] Step 6.2: Statistical analysis is performed on the aircraft position and velocity estimation results obtained by each algorithm during the simulation process. The root mean square error (RMSE) is calculated as a performance evaluation index. Simulation results show that under 100 Monte Carlo runs, the method described in this invention outperforms the particle filter method in both position and velocity estimation accuracy. Furthermore, with a particle count of 200, its estimation performance is close to that of the ideal filter. The ideal filter method refers to the state estimation result obtained by the optimal filter under the assumption that the system mechanism model parameters and noise statistical characteristics are completely known under simulation conditions. Its performance can be regarded as the theoretical upper bound achievable by the algorithm. As shown in Table 1:
[0077] Table 1. Comparison of RMSE experimental results for the parameter migration error evolution method.
[0078] method Number of particles Location RMSE (m) Speed RMSE (m / s) Particle filtering method 200 0.46 0.29 This method 200 0.41 0.26 Ideal filtering method 200 0.39 0.25
[0079] Step 6.3: Under the same simulation conditions and computing platform, the computation time of each algorithm is compared and analyzed. The results show that when the number of iterations of the least squares optimization process is greater than 3, the estimation performance of the method tends to stabilize, and its single simulation computation time is controlled within 2.5s, which is significantly lower than the computation time of the particle filter method. This verifies that the present invention has low computational cost while ensuring high-precision state estimation.
[0080] Example: As a preferred embodiment of the present invention, the verification process includes the following steps:
[0081] S21, Definition of the The measurement information at each sampling time is ,in For discrete-time indexing, The time interval between adjacent sampling times;
[0082] S22. Calculate the target's position in the third dimension based on the change in the aircraft's position in three-dimensional space. Speed of time The calculation formula is as follows:
[0083] ;
[0084] in, , , For the first Position coordinates at that moment , , For the previous moment The location coordinates.
[0085] S23. Constructing auxiliary feature parameters based on velocity and attitude changes and The calculation formula is as follows:
[0086] ;
[0087] in, This indicates the change in yaw angle. It represents the change in trajectory tilt angle, used to reflect the aircraft's turning and pitch characteristics.
[0088] S24. Transfer the original measurement information Difference term Speed characteristics and auxiliary feature parameters and The data are uniformly organized and spliced to form a preprocessed measurement feature vector containing time-series related information, and then arranged in chronological order to form a preprocessed measurement information sequence. The measurement information sequence is used as input to the mechanism model parameter migration error identification network and as input data for the mechanism model parameter migration error identification step.
[0089] S31. Based on the temporal features output by the preceding Long Short-Term Memory network, normalize the features at each time step using the Softmax function, and calculate the corresponding attention weighting coefficients. The calculation formula is as follows:
[0090] ;
[0091] in, Indicates the preceding LSTM layer at the 1st... The output feature vector at each time step and These are the trainable weight parameters and bias parameters of the attention layer, respectively; the attention weighting coefficients are used to reflect the importance of features at different times to the current mechanism model parameter transfer error identification results.
[0092] S32. Based on the attention weighting coefficients, the output features of the preceding LSTM layer at each time step are weighted and summed to obtain the feature representation fused with temporal information. The calculation formula is as follows:
[0093] ;
[0094] in, The length of the input feature sequence is represented; the feature representation By comprehensively considering the contribution of features at different time steps to the recognition results, the output of the recognition network is not only dependent on the feature information at a single moment, but is related to the key moment features with discriminative value in the entire time series, thereby improving the accuracy and stability of the temporal evolution recognition of the mechanism model parameter migration error.
[0095] S41. The input value of any neuron in the neural network during the training process. Standardization is performed to adjust its distribution to a standardized distribution with a mean of 0 and a variance of 1. The calculation formula is as follows:
[0096] ;
[0097] in, This represents the average value of the input values in the current batch. Indicates the corresponding variance. To prevent the introduction of tiny positive numbers with a denominator of zero; through the aforementioned standardization process, the impact of differences in the distribution of data from different batches on the network training process can be reduced, thereby improving the stability of network training.
[0098] S42. After completing the standardization process, introduce trainable scaling parameters. Translation parameters The standardized data is then subjected to a linear reconstruction transformation to obtain the output of the batch normalization layer. The calculation formula is as follows:
[0099] ;
[0100] The reconstruction transformation is used to restore the nonlinear expressive ability of the network while maintaining the advantages of standardization, so that the batch normalization layer can suppress gradient vanishing and overfitting problems without affecting the effective representation of the migration error characteristics of the mechanism model parameters.
[0101] S51. Input the preprocessed measurement information sequence into the recognition network RNN. t To obtain the target's flight status tag Lt, the flight status tag L tThis is used to characterize whether the target is in a stable flight state or a maneuvering flight state within the current time window; simultaneously, the measurement information sequence is input into the recognition network RNN. d To obtain the target's lateral maneuver direction label L d The flight status label and lateral maneuver direction label are used to characterize the overall motion pattern of the aircraft during the current time period;
[0102] S52, based on the flight status label L t The recognition results are used for branch judgment. When the flight status label is determined to be a stable flight state, the recognition network RNN is invoked. 1,SFS and RNN 2, SFS The corresponding motion feature parameter labels L1 and L2 are identified respectively; when the flight state label is determined to be a maneuvering flight state, the identification network RNN is invoked. 1, MFS and RNN 2, MFS Identify the corresponding motion feature parameter labels L1 and L2. L1 represents energy-related parameter transfer features, and L2 represents attitude-coupled parameter transfer features.
[0103] S53. Based on the independent processing of the measurement information sequence by each recognition network and the output of the results, the label confidence scores output by each recognition network are fused, and the target mechanism model parameter migration error is calculated. The confidence score at the current moment is calculated as follows:
[0104] ;
[0105] in, Indicates the most recent A sequence of preprocessed measurement information at various time points. This represents the label output by the first recognition network at time k. Given a sequence of preprocessed measurement information from the most recent N time moments Under the given conditions, the parameter transfer error of the target mechanism model is the posterior confidence level of the j-th mode Mj.
[0106] S61. Based on the confidence distribution of the mechanism model parameter migration error obtained in step S3, determine the set of mechanism model parameter migration error models that participate in the state estimation at the current moment, and complete the state prediction and estimation based on the selected models, so as to achieve a trade-off between estimation accuracy and computational complexity.
[0107] S62. The total number of candidate mechanism model parameter transfer error models is N. The fusion weight of each model is determined according to its parameter transfer error confidence level, and they are sorted according to the weight. Models with smaller weights are suppressed during the fusion process, thereby adaptively determining the set of models that effectively contribute to the state estimation without explicitly limiting the number of models. N represents the number of mechanism models that effectively contribute to the state estimation result at the current time, which is adaptively determined by the model confidence weight distribution.
[0108] S63. Select models in descending order of confidence level according to the mechanism model parameter migration error model until the cumulative confidence weight of the selected models is greater than the preset threshold, thereby dynamically determining the number of models participating in state estimation to adapt to the change of mechanism model parameter migration error over time; and complete the state prediction for the next moment based on the determined model set.
[0109] S71. For a nonlinear state-space model that includes system state variables and multiplicative mechanism model parameters, construct an error model between the system's predicted output and the actual observed data, and use the error as an evaluation index for state estimation:
[0110] ;
[0111] in, Indicates the first The system state variables at each moment, This represents the parameter perturbation term of the model. Represents observation data, and These are the nonlinear state transition function and the observation function, respectively. and Indicates the disturbance term;
[0112] S72. Based on the aforementioned error model, an optimization model is constructed with the objective of minimizing the error between the system's predicted output and the observed data. The system state variables and the parameters of the multiplicative mechanism model are used as variables to be estimated, and the least squares criterion is employed to solve the optimization model.
[0113] ;
[0114] in, Indicates the first The estimated values of the system state variables at each sampling time. Indicates the first The estimated values of the parameters of the multiplicative mechanism model at each sampling time. Indicates the first The system state variables at each moment, This represents the parameter perturbation term of the model;
[0115] S73. By iteratively solving the least squares optimization problem, the estimated values of the system state variables and multiplicative mechanism model parameters are updated at each sampling time until the sum of squared residuals or the parameter update amount meets the preset convergence condition, thereby obtaining the aircraft state estimation result under the condition of uncertainty of multiplicative mechanism model parameters.
[0116] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for parameter migration error evolution of an aircraft mechanics model under complex constraints, characterized in that: The specific steps are as follows: S1. Construct a parameter migration error dataset for the mechanism model using aircraft simulation data. Then, by performing differential processing and kinematic feature extraction on the three-dimensional position and attitude measurement information of the aircraft at discrete moments, construct a preprocessed feature sequence containing auxiliary feature terms such as velocity and curvature, forming a time-series feature sequence that can characterize the evolution characteristics of the parameter migration error of the aircraft under complex constraints. S2. The temporal evolution characteristics of parameter transfer error in the mechanism model are modeled and trained by introducing an attention mechanism and batch normalization long short-term memory network to obtain a deep learning recognition model for subsequent parameter transfer error identification. S3. Construct a hierarchical mechanism model parameter transfer error evolution identification process using the trained deep learning recognition model. Decompose the complex parameter transfer error identification problem into multi-level sub-problems, and output the mechanism model parameter transfer error model and the mechanism model parameter transfer error confidence characterization result obtained based on the mechanism model parameter transfer error model. The mechanism model parameter transfer error model is a pre-constructed candidate mechanism model under different parameter offset conditions. S4. Based on the confidence characterization results of the mechanism model parameter migration error obtained in S3, a model switching strategy is adopted to select a set of models for state estimation from the preset mechanism model parameter migration error model set. A nonlinear state-space model is constructed that includes multiplicative mechanistic stochastic parameters caused by the parameter migration error of the mechanistic model. The multiplicative mechanistic stochastic parameters are introduced by the parameter migration error of the corresponding model and are used to characterize the proportional offset of the mechanistic model parameters relative to the nominal value. Using the residual between the predicted output and the measured output of the nonlinear state-space model as the objective function, a joint state parameter estimation method is constructed using the least squares criterion to jointly estimate the system state and mechanism model parameters, and output the state estimation result of the aircraft.
2. The method for parameter migration error evolution of an aircraft mechanism model under complex constraints according to claim 1, characterized in that: S1 specifically includes the following steps: S11: Based on the aircraft simulation data, establish an aircraft motion simulation model and generate a simulation dataset that simultaneously contains both stable flight state tracks and maneuvering flight state tracks. The dataset contains multiple tracks, covering various flight states including turning maneuvers, attitude adjustment maneuvers, and periodic complex maneuvers, and is used to describe the motion behavior characteristics of the aircraft under complex constraints. The dataset is then divided into a training set and a validation set according to a preset ratio for subsequent training and performance evaluation of the recognition network. S12: Define discrete time intervals The following measurement information is The discrete time interval is The difference operation is performed on the measurement information at adjacent time points to obtain... The target velocity is calculated based on the changes in its three-dimensional position coordinates between adjacent time points. ； in, , , They represent the target at the 1st and 2nd positions, respectively. Spatial coordinates at a given time; S13: To enhance the characterization capability of aircraft maneuvering characteristics and the evolution law of parameter migration errors in the mechanism model, attitude angle change information is introduced on the basis of velocity information to construct auxiliary feature terms: ； in, This indicates the change in yaw angle. This represents the change in track inclination, and combines the original measurement information, difference terms, velocity information, and auxiliary feature terms to form a preprocessed feature sequence. It is used as input for the subsequent mechanism model parameter transfer error identification network.
3. The method for parameter migration error evolution of an aircraft mechanism model under complex constraints according to claim 1, characterized in that: S2 specifically includes the following steps: S21: To address the strong temporal correlation and nonlinear characteristics of the evolution of mechanism model parameter transfer error over time, a long short-term memory recognition network Att-LSTM+BN based on attention mechanism and batch normalization is constructed. The network consists of a fully connected layer, a batch normalization layer, a long short-term memory layer, an attention layer, and an output layer connected in sequence, which is used to learn and represent the characteristics of aircraft trajectory type and mechanism model parameter transfer error. S22: Set network structure parameters according to different recognition tasks, including the track type recognition network RNNt and the mechanism model parameter recognition network RNN. 1, SFS RNN 2, SFS RNN 1, MFS RNN 2, MFS The number of neurons in the fully connected layer was set to 128 and 16, and the number of neurons in the long short-term memory layer was set to 128 and 64; the lateral maneuvering recognition network RNN d The number of neurons in the fully connected layer was set to 128 and 16, and the number of neurons in the long short-term memory layer was set to 512 and 256 to enhance the ability to model complex lateral maneuvering features. S23: Set the network training parameters. The length N of the input feature sequence is 150 or 200. Use the cross-entropy loss function as the training objective function. Select RMSprop as the optimizer and set the learning rate to 0.
001. The track type recognition network is trained for 20 iteration cycles, and the other recognition networks are trained for 80 iteration cycles. S24: Use the training and validation sets constructed in S11 to train and validate each recognition network. Optimize the network parameters by monitoring the changes in the loss function and recognition accuracy to obtain a trained network model for identifying the mechanism model parameter transfer error evolution.
4. The method for parameter migration error evolution of an aircraft mechanism model under complex constraints according to claim 1, characterized in that: The specific steps of S3 are as follows: S31: The preprocessed feature sequence obtained in S1 Input the trajectory type recognition network RNNt to identify the target's flight status label L at the current moment. t Simultaneously, the feature sequence is input into the lateral maneuver recognition network RNN. d Identify the target's lateral movement direction label L d This is to enable a preliminary determination of the overall motion pattern of the aircraft; S32: Based on flight status label L t The recognition results are used for branch judgment. When a stable flight state is determined, the corresponding mechanism model parameter recognition subnetwork RNN is invoked. 1, SFS and RNN 2, SFS The mechanism model parameter labels L1 and L2 are identified respectively; when a maneuvering flight state is determined, the corresponding mechanism model parameter identification subnetwork RNN is invoked. 1,MFS and RNN 2, MFS The mechanism model parameter labels L1 and L2 are identified respectively, thereby enabling the identification of differentiated parameter migration errors for different track types; S33: Based on each recognition network, the measurement feature sequence is processed independently and the results are output. The confidence scores of the labels corresponding to the flight status label, lateral maneuver label and mechanism model parameter label are fused to determine the confidence score distribution of the target at the current time for each mechanism model parameter migration error model Mj, which is used for subsequent model switching and state estimation processes.
5. The method for parameter migration error evolution of an aircraft mechanism model under complex constraints according to claim 1, characterized in that: The specific steps of S4 are as follows: S41: Establish a system state variable Multiplicative mechanism model parameters The nonlinear discrete-time state-space model integrates the system state and mechanism model parameters into the state-space modeling framework, and describes the influence of mechanism model parameter migration error on system state evolution by introducing parameter deviation terms. S42: For the above state-space model, a least-squares cost function based on measurement residuals is constructed. By minimizing the weighted sum of squared residuals between the system's predicted output and the actual measurement data, the joint estimation of the system state and mechanism model parameters is achieved. S43: Using an alternating least squares optimization strategy, the system state is estimated under the condition of fixed mechanism model parameters, and the mechanism model parameters are updated under the condition of fixed system state. The least squares cost function is iteratively optimized until the preset convergence condition is met, and the system state estimation result at the current time is obtained.
6. The method for parameter migration error evolution of an aircraft mechanism model under complex constraints according to claim 1, characterized in that: Also includes S5: S51. Based on the confidence distribution results of the migration error models Mj of each mechanism model obtained in step S3, the models are sorted according to their confidence levels, and a model switching strategy is adopted to determine the set of models participating in state estimation at the current time. The model switching strategy is adjusted based on the confidence weight of each model to suppress the participation of low-confidence models, or adaptively determines the set of models participating in state estimation based on the cumulative confidence weight criterion. S52: For the set of models determined in S51, the joint state parameter estimation method described in S4 is executed respectively. At each time step, the system state and mechanism model parameters are estimated in a coordinated manner to obtain the aircraft state estimation results under different model assumptions. The state estimation process of each model is independent of each other and can be calculated in parallel. S53: Based on the confidence weights corresponding to the migration error models of each mechanism model parameters, the state estimation results obtained under different models are fused by weighted least squares to form the comprehensive state estimation result at the current moment.