A multi-dimensional filtering feature-based launch data correction method

Abstract

The application discloses a launch data correction method based on multi-dimensional filtering features, and the offline stage comprises the following steps: acquiring radar measurement data and high-precision real track data; performing interactive multiple model filtering based on the radar measurement data to obtain fused states, model probabilities, new information and covariance features; predicting a future position of a target based on a current filtering state and calculating a reference launch data, calculating real track reference launch data based on the real track, and taking a difference between the two as a launch data compensation label; taking the fused states, the model probabilities, the new information and the covariance features in a history window as input, and taking the compensation label as expected output to train a long short-term memory network. In the online stage, the radar measurement data is acquired in real time and interactive multiple model filtering is performed, the current reference launch data is calculated based on the current filtering state, the filtering features in the history window are input into the trained network to obtain predicted launch data compensation, and the compensation is superimposed on the reference launch data to generate corrected launch data. The application reduces the data deviation and improves the interception accuracy by using interactive multiple model filtering features.

Description

Technical Field

[0001] This invention relates to safety protection and launch control technology for low-altitude unmanned aerial vehicles (UAVs), specifically to a method for correcting launch parameters based on multi-dimensional filtering characteristics. Background Technology

[0002] As the application of low-altitude drones expands, drone targets entering airport airspace, important facilities, and major event venues without permission pose a continuous risk to low-altitude safety.

[0003] In low-altitude UAV security scenarios, the accuracy of launch parameter calculations for high-speed projectile interceptors directly impacts their effectiveness in intercepting maneuvering low-altitude targets. Traditional launch parameter calculations typically rely on target motion models and, to some extent, assume that the target maintains a constant speed, uniform acceleration, or other known motion patterns during the projectile's flight time. However, in scenarios such as airport airspace, perimeters of critical facilities, or low-altitude test ranges, small UAVs may exhibit complex motions such as acceleration / deceleration, turning, climbing, diving, and serpentine maneuvers, leading to a mismatch between the actual target motion and the extrapolation model used by the interceptor.

[0004] Interactive multi-model filtering (IMF) can jointly estimate the target state using multiple candidate motion models and reflect the degree of matching between different motion models and the current target motion state through model probabilities. Compared with single-model filtering, IMF has stronger adaptability under target maneuvering conditions. However, when the target's true motion model is not included in the candidate motion model set, although the position estimation at the current moment can maintain high accuracy, the extrapolation of future trajectories based on velocity, acceleration, or turning rate may still produce deviations, leading to errors in the calculation of launch parameters.

[0005] Existing launch parameter compensation methods based on bias estimation typically focus on establishing compensation relationships using historical launch parameter biases, while failing to adequately utilize the internal features such as model probabilities, innovations, and covariance generated during the interactive multi-model filtering process. This makes it difficult to fully characterize the impact of target motion model mismatch on launch parameter calculations. Therefore, it is necessary to propose a launch parameter compensation method that combines interactive multi-model filtering features with a sequence prediction model to improve the accuracy of launch parameter calculations under complex maneuvering target conditions. Summary of the Invention

[0006] The purpose of this invention is to propose a method for correcting emission parameters based on multidimensional filtering characteristics.

[0007] The technical solution to achieve the purpose of this invention is: a method for correcting emission parameters based on multi-dimensional filtering features, comprising an offline training stage and an online application stage, as detailed below:

[0008] Offline training phase:

[0009] Step 1: Acquire radar measurement data and high-precision real trajectory data of low-altitude moving targets; perform interactive multi-model filtering based on radar measurement data to obtain target fusion state, model probability, innovation, and covariance features; the fusion state is the weighted fusion result of the filtering states of each motion model, the model probability is the probability of each motion model estimated by interactive multi-model filtering, the innovation is the difference between the measurement prediction value and the actual measurement value of each motion model in interactive multi-model filtering, and the covariance features are used to characterize the uncertainty of state estimation and / or measurement prediction in the interactive multi-model filtering process;

[0010] Step 2: Predict the future trajectory of the target based on the current interactive multi-model filtering state and solve the reference emission parameters; solve the real trajectory reference emission parameters based on high-precision real trajectory data; generate emission parameter compensation labels based on the difference between the real trajectory reference emission parameters and the reference emission parameters.

[0011] Step 3: Train the emission parameter compensation prediction model with the fusion state, model probability, innovation and covariance features within the historical window as input and the emission parameter compensation amount label as the expected output.

[0012] Online application stage:

[0013] Step 4: Acquire radar measurement data of the target in real time, perform interactive multi-model filtering to obtain the current target fusion state, model probability, innovation and covariance features, predict the future trajectory of the target based on the current filtering state, solve the current reference launch parameters, input the fusion state, model probability, innovation and covariance features in the historical window into the trained launch parameter compensation prediction model to obtain the predicted launch parameter compensation amount, and superimpose the predicted launch parameter compensation amount onto the reference launch parameters to generate the corrected launch parameters of the high-speed projectile interceptor.

[0014] Furthermore, the filtering features include target fusion state, model probability, innovation, and covariance features; the covariance features are formed by compressing the state covariance matrix and / or innovation covariance matrix obtained by interactive multi-model filtering to form a one-dimensional covariance feature vector; the compression process includes extracting at least one of the diagonal elements, trace, and eigenvalues ​​of the covariance matrix.

[0015] Furthermore, the radar measurement data includes slant range, azimuth angle, and elevation angle; the radar measurement data is used for measurement updates of interactive multi-model filtering; the historical window consists of multiple consecutive sampling times.

[0016] Furthermore, the reference launch parameters include at least one of the reference azimuth angle, reference elevation angle, and reference projectile flight time; the launch parameter compensation amount includes at least one of the azimuth angle compensation amount, elevation angle compensation amount, and projectile flight time compensation amount.

[0017] Furthermore, the emission parameter compensation prediction model includes at least one long short-term memory network sub-model; the long short-term memory network sub-model is used to encode the filtering features within the historical window and output the emission parameter compensation amount through the output layer. When the emission parameter compensation prediction model includes multiple long short-term memory network sub-models, each sub-model outputs the candidate emission parameter compensation amount.

[0018] Furthermore, a mean squared error loss function is used when training the launch parameter compensation prediction model. This mean squared error loss function is used to measure the difference between the launch parameter compensation predicted by the network and the launch parameter compensation label.

[0019] Furthermore, the interactive multi-model filter includes at least two target motion models; the target motion models include at least one of a uniform speed model, a uniform acceleration model, a left coordinated turn model, a right coordinated turn model, and a coordinated turn model with different turning rates.

[0020] Furthermore, the launch parameter compensation prediction model includes one or more sub-predictors; when the compensation prediction model includes multiple sub-predictors, the multiple sub-predictors correspond to different target maneuver types, different model mismatch modes, or different training sample sets, and each sub-predictor outputs candidate launch parameter compensation quantities. The weight of each sub-predictor is determined according to the sub-predictor adaptability evaluation quantity, and the multiple candidate launch parameter compensation quantities are weighted and fused to obtain the final launch parameter compensation quantity.

[0021] Furthermore, the sub-predictor adaptive evaluation quantity characterizes the degree of deviation between the candidate corrected emission parameters and the filtered trajectory emission parameters. The smaller the deviation, the greater the weight of the corresponding sub-predictor. The candidate corrected emission parameters are obtained by superimposing the baseline emission parameters and the compensation amount of the candidate emission parameters output by the corresponding sub-predictor. The filtered trajectory emission parameters refer to the emission parameters obtained by extrapolating the target motion and solving the hit equation based on the interactive multi-model filtered trajectory before the current time.

[0022] A transmission parameter correction system based on multidimensional filtering features includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the transmission parameter correction method based on multidimensional filtering features.

[0023] Compared with existing technologies, the significant advantages of this invention are: 1) It can fully characterize the impact of target motion model mismatch on the calculation of launch parameters by utilizing the internal features such as fusion state, model probability, innovation, and covariance generated during the interactive multi-model filtering process, thereby reducing launch parameter bias. 2) Using the fusion state, model probability, innovation, and covariance features within the historical window as input and the launch parameter compensation amount label as the expected output, the launch parameter compensation amount prediction model is trained, enabling the long short-term memory network to learn the relationship between the target maneuver state and the launch parameter bias. 3) In the online application stage, only the filtering features within the historical window need to be input into the trained network to obtain the predicted launch parameter compensation amount, resulting in low computational overhead and meeting the requirements of real-time interception. 4) In scenarios with significant model mismatch, such as turning, diving, and serpentine maneuvers, it can reduce launch parameter bias and interception bias, and achieves positive improvement under different geometric orientations (crossing, oblique approach, and head-on radial approach). 5) The multi-subpredictor architecture enhances the system's generalization ability and robustness to different maneuver modes by training separately for different maneuver types or model mismatch modes and adaptively weighting and fusing them according to the filtered trajectory evaluation. Attached Figure Description

[0024] Figure 1 This is a flowchart of a method for predicting and correcting emission parameters based on interactive multi-model filtering features.

[0025] Figure 2 This is a comparison chart showing the actual compensation required in a serpentine maneuver scenario and the output of various methods.

[0026] Figure 3 A 3D comparison of interception points using different methods in a serpentine maneuver scenario.

[0027] Figure 4 This is a graph showing the variation of interception deviations under different methods in a serpentine maneuver scenario. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0029] The proposed method for correcting launch parameters based on multi-dimensional filtering features inputs temporal features such as interactive multi-model filtering fusion state, model probability, innovation, and covariance features into a long short-term memory network to predict the compensation amount of launch parameters at the current moment, thereby achieving online compensation of the reference launch parameters. This method is particularly suitable for online launch parameter correction in authorized scenarios such as airport airspace, perimeters of important facilities, and low-altitude test ranges, for high-speed projectile interception of maneuvering small UAVs or similar low-altitude moving targets.

[0030] For ease of description, the following terms are first defined:

[0031] High-speed projectile interceptor devices refer to devices that can launch high-speed projectiles toward a target and intercept the target by means of azimuth angle, elevation angle, flight time or corresponding launch control parameters, including but not limited to high-speed interceptor missile launchers, kinetic energy projectile devices, close-range low-altitude target interceptors and their equivalent launch control platforms.

[0032] Reference launch parameters: Based on the current interactive multi-model filtering state, the future trajectory of the target is extrapolated, and the hit equation is solved on the extrapolated trajectory to obtain the reference launch parameters. In this invention, the reference launch parameters include azimuth angle, elevation angle, and projectile flight time, etc.

[0033] Real trajectory reference launch parameters: Launch parameters obtained based on high-precision real trajectory data are used as a reference standard and source of expected output labels during the offline training phase.

[0034] Filtered trajectory emission parameters: Based on the interactive multi-model filtered trajectory formed before the current time step, the hit equation is solved on the filtered trajectory to obtain the filtered trajectory emission parameters. This evaluation quantity is used in the optional multi-predictor extension implementation.

[0035] Launch parameter compensation: The difference between the actual trajectory reference launch parameters and the baseline launch parameters, representing the launch parameter deviation caused by the mismatch of the target motion model.

[0036] Interactive multi-model filter features: the fusion state, probabilities of each model, innovation, and covariance features output by the interactive multi-model filter.

[0037] The method of this invention includes an offline training phase and an online application phase, as detailed below:

[0038] Offline training phase: acquire radar measurement data and high-precision real trajectory data; perform interactive multi-model filtering on radar measurement data; extrapolate the future position of the target based on the current interactive multi-model filtering state and calculate the reference emission parameters; calculate the reference emission parameters of the real trajectory based on the high-precision real trajectory; obtain the emission parameter compensation label by subtracting the reference emission parameters of the real trajectory from the reference emission parameters; train the long short-term memory network with the interactive multi-model filtering features (fusion state, model probability, innovation, covariance features) within the historical window as input and the emission parameter compensation label as the expected output.

[0039] In the online application phase: real-time acquisition of target radar measurement data; input of interactive multi-model filter to obtain target state estimation and filtering features; solution of reference launch parameters based on target state estimation and target motion extrapolation model; input of interactive multi-model filter features within the historical window into the trained long short-term memory network to obtain the predicted launch parameter compensation amount; superimposition of the predicted launch parameter compensation amount onto the reference launch parameters to obtain the corrected launch parameters.

[0040] The steps of the present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0041] Step 1: Obtain training sample data

[0042] Under typical low-altitude target maneuvering conditions, such as straight routes, climb routes, dive routes, and serpentine maneuvers, acquire the target's true trajectory data and radar measurement data. The true trajectory data can be obtained from a high-precision positioning or simulation system and includes the time and the target's three-dimensional position coordinates; the radar measurement data includes target measurement information such as azimuth, elevation, and slant range.

[0043] The target state can be represented as ,in For the target position component, For the target velocity component, Let be the target acceleration component. Radar measurements can be expressed as... ,in Slope distance It is the azimuth angle. For elevation and elevation angles.

[0044] Step 2: Construct interactive multi-model filtering features

[0045] Radar measurement data is input into an interactive multi-model filter (EMF), which can use EKF, UKF, or other nonlinear filtering methods to handle the nonlinearity of the measurement equations. The EMF includes multiple target motion models, such as uniform velocity models, uniform acceleration models, and coordinated turning models. Different models have their own state transition equations and process noise parameters, and the fused state at the current moment is obtained through EMF filtering. Model probability New information and covariance characteristics .

[0046] Interactive multi-model filtering can employ existing standard IMM computation procedures, including input interaction, model conditional filtering, likelihood calculation, model probability update, and state estimation fusion. This invention does not limit the specific filter form, only requiring that the online output can characterize the target state estimation, model matching degree, and measurement residual information.

[0047] The following is an optional IMM calculation procedure, where each model in the IMM has state equations and measurement equations:

[0048]

[0049]

[0050] In the formula: Indicates the sampling time; Model number; for time The state of the model; Represents the state transition matrix; For observation vectors; This is process noise; To observe noise; and Both are Gaussian noise with a mean of 0, and their corresponding covariance matrices are respectively and .

[0051] IMM comprises four steps: model interaction, filter update, model probability update, and fusion output. A commonly used formula for input interaction is:

[0052]

[0053]

[0054]

[0055]

[0056] In the formula, The first The state value and covariance of the matched filter of each model at the next time step. For each model filter The filtered state value and covariance at time t. It is a model Transform to model The transition probability, It is the first Time-of-flight model The model probability.

[0057] Commonly used formulas for model probability updates:

[0058]

[0059]

[0060]

[0061]

[0062] Commonly used formulas for state estimation fusion:

[0063]

[0064]

[0065] Step 3: Calculate the reference emission parameters

[0066] Based on the target state obtained from interactive multi-model filtering, the future motion state of the target is extrapolated, and the predicted interception point is solved by combining the projectile flight time. According to the predicted interception point position, the reference launch parameters at the current moment are solved using the launch control solution method. The reference launch parameters include azimuth angle, elevation angle, and projectile flight time.

[0067] Record based on the first The future predicted position of the target obtained by extrapolating the filtered state at any time is: , Time of flight of the projectile. Reference launch parameters. Solving function from launch control Give, that is

[0068] Step 4: Calculate the reference launch parameters of the actual trajectory and generate compensation labels.

[0069] Based on the high-precision real trajectory, the real trajectory reference launch parameters at the corresponding time are solved; the difference between the real trajectory reference launch parameters and the reference launch parameters is calculated to obtain the launch parameter compensation amount label, which is used as the expected output in the training process of the launch parameter compensation amount prediction model.

[0070] Let the reference interception location obtained based on the actual trajectory be denoted as . The actual trajectory reference launch parameters are The launch parameter compensation amount label is This compensation amount may include azimuth compensation. Elevation / Elevation Compensation Amount Optionally, it may include projectile time-of-flight compensation. .

[0071] Step 5: Construct the training dataset and train the compensation amount prediction model

[0072] The fusion state, model probability, innovation, and covariance features within the historical window are used to form a filtered feature vector, and the emission parameter compensation is used as the output label to construct a training dataset. A long short-term memory network is used to build an emission parameter compensation prediction model, and the model parameters are trained and optimized using the training and validation sets.

[0073]

[0074]

[0075]

[0076]

[0077] in, For the first The filtered feature vector at time step 1. For length is The history window input sequence, Represents Long Short-Term Memory Network, For network parameters, This is the compensation amount for the launch parameters predicted by the network. The mean squared error loss function is minimized during the training phase. Optimize network parameters; the root mean square error of the compensation amount can be used as a post-training evaluation indicator.

[0078] Step 6: Acquire target measurements online and perform interactive multi-model filtering.

[0079] During the online application phase, target radar measurement information is acquired in real time and input into an interactive multi-model filter to obtain the target fusion state, model probability, innovation, and covariance features at the current moment.

[0080] Step 7: Solve the reference emission parameters online

[0081] Extrapolate the target's future motion state based on the current filtering results, and solve for the reference emission parameters at the current moment.

[0082] Step 8: Online prediction of launch parameter compensation

[0083] Input the filtered features (fusion state, model probability, innovation, covariance features) within the historical window into the trained Long Short-Term Memory network, and output the emission parameter compensation amount at the current moment.

[0084]

[0085] Step 9: Generate the corrected emission parameters

[0086] The predicted launch parameters are superimposed onto the baseline launch parameters to obtain the corrected launch parameters, which are then used as the launch parameter parameters of the high-speed projectile interceptor.

[0087]

[0088] in, For the first Launch parameters after time correction This is the compensation amount for the transmission parameters predicted by the Long Short-Term Memory Network. If the compensation amount only includes azimuth and elevation angles, then the corrected azimuth and elevation angles are obtained by superimposing the corresponding angle compensation amounts.

[0089] Step 10: Optional multi-predictor weighted fusion

[0090] In another alternative implementation, the compensation quantity prediction model includes multiple sub-predictors, each trained for different target maneuver types, different model mismatch modes, or different training sample sets. During the offline phase, a total of [number of sub-predictors] are trained. Individual predictors, They are the first The variance of the prediction error of each sub-predictor For the first The compensation amount of the emission parameters output by the sub-predictor This represents the number of training samples.

[0091]

[0092]

[0093]

[0094] During the online phase, the The sub-predictor outputs candidate emission parameter compensation. Candidate revised launch parameters The adaptive evaluation value of each sub-predictor can be calculated based on the deviation between the candidate corrected emission parameters and the filtered trajectory emission parameters. The multiple candidate compensation quantities are then weighted and fused to obtain the final launch parameter compensation quantity. .

[0095] This section provides the parameters for adjusting the launch parameters based on the candidate launch. With filter trajectory emission parameters Adaptive evaluation of the constructor predictor The calculation formula and weights can be adopted. The calculation is formal, therefore the calculation expression is:

[0096]

[0097]

[0098]

[0099]

[0100] The main implementation of this invention uses a single long short-term memory network to predict the transmit parameter compensation amount; the aforementioned multi-predictor weighted fusion is an optional extended implementation and does not constitute a necessary step in the main process. This embodiment uses a single IMM+LSTM compensation amount prediction model for simulation verification, and does not use the multi-predictor fusion result as an experimental conclusion.

[0101] Example

[0102] To verify the applicability of this method under different geometric configurations and different target motion types, the following experiments were conducted.

[0103] A multi-scenario simulation dataset under high-speed projectile conditions was constructed. The initial velocity of the projectile was set to 1080 m / s, the mass to 0.365 kg, the equivalent diameter to 0.030 m, and the drag coefficient to 0.30. Target motion scenarios included uniform linear motion, uniform acceleration, turning, climbing, diving, and serpentine maneuvers; the geometric configurations between the target and the interceptor included cross-traversal, oblique approach, and head-on approach. The above parameters are only used to illustrate the simulation verification process of the method of this invention and do not limit the specific engineering form and parameter range of the high-speed projectile interceptor.

[0104] In the sample construction stage, each target trajectory is divided into multiple samples according to time sequence, and the historical window sequence features and emission parameter compensation amount labels corresponding to each sample are generated according to the aforementioned steps.

[0105] Training, validation, and testing are grouped according to the target trajectory to ensure no overlap between training, validation, and testing trajectories. The testing phase evaluates only the testing trajectory to avoid the same trajectory sample appearing in both the training and testing sets.

[0106] The comparison methods include the emission parameter calculation method based on the CV model (CV uncompensated), the emission parameter correction method based on CV model ARMA compensation (CV+ARMA), the emission parameter calculation method based on interactive multi-model filtering (IMM uncompensated), the emission parameter correction method based on interactive multi-model filtering ARMA compensation (IMM+ARMA), and the emission parameter correction method based on multi-dimensional filtering features of the present invention (IMM+LSTM).

[0107] Evaluation metrics include average interception deviation, root mean square interception deviation, and the improvement rate of average interception deviation relative to the transmission parameter calculation method based on interactive multi-model filtering.

[0108] Simulation results show that the method of this invention can learn the relationship between the target maneuver state and the launch parameter deviation by utilizing interactive multi-model filtering fusion state, model probability, innovation, and covariance features, thereby reducing launch parameter deviation and interception deviation in scenarios such as turning, diving, and serpentine maneuvers. All methods were evaluated under the same test trajectory and the same fire control calculation caliber.

[0109] Table 1 Comparison of overall interception deviations of different launch control methods on the test set.

[0110]

[0111] As shown in Table 1, across 5250 test samples, the average interception error of the IMM+LSTM method decreased from 0.820m with uncompensated IMM launch control to 0.480m, representing an average improvement of approximately 41.4%; the interception error RMSE decreased from 0.926m to 0.598m. These results demonstrate that using the timing characteristics of IMM filtering to predict launch parameter compensation can still effectively reduce launch parameter errors caused by model mismatch.

[0112] Table 2. Average interception deviation improvement effect of the IMM+LSTM method under different maneuver types.

[0113]

[0114] As shown in Table 2, IMM+LSTM achieves positive improvements across all maneuver types, with more significant improvements observed in scenarios where model mismatch is pronounced, such as serpentine maneuvers, dives, climbs, and coordinated turns. Simulation results demonstrate that this invention does not simply improve the launch pointing accuracy of a specific flight path, but rather utilizes the fusion state, model probability, innovation, and covariance features of IMM to learn the relationship between the target maneuver state and launch parameter deviations, thereby improving the accuracy of launch parameter calculations across various maneuver scenarios.

[0115] Table 3 Comparison of average interception deviations of various methods under different geometries

[0116]

[0117] As shown in Table 3, IMM+LSTM achieves positive improvements under different target orientations, including lateral, diagonal, and oncoming radial approaches. Specifically, the improvement rates for left-front diagonal and right-front diagonal approaches are approximately 41.2% and 47.3%, respectively, while the improvement rate for oncoming radial approaches is approximately 39.5%. CV+ARMA and IMM+ARMA show some improvement compared to the uncompensated baseline under most geometric orientations, indicating that the statistical extrapolation method based on online RELS-ARMA can achieve a small compensation benefit under multiple geometric configurations, but its improvement is still significantly lower than that of the LSTM method based on multi-dimensional filtering features.

[0118] Figure 2 , Figure 3 , Figure 4 The results are presented from three dimensions: comparison of compensation output, spatial distribution of interception points, and temporal variation of interception deviation.

[0119] Depend on Figure 2 As can be seen, in the serpentine maneuver scenario, the target trajectory continuously oscillates laterally, and the required launch parameter compensation amount also changes continuously over time. Specifically, the azimuth and elevation angle compensation amounts exhibit alternating positive and negative values ​​and local peak variations, while the projectile flight time compensation amount also shows significant deviations at certain moments. The IMM+ARMA method can respond to some slowly changing compensation trends, but its overall output is relatively flat, especially for elevation angle and projectile flight time compensation, which are close to zero, making it difficult to fully follow the rapid changes caused by the serpentine maneuver. In contrast, the output of the IMM+LSTM method is closer to the actual required compensation amount in terms of both direction and magnitude of change, and it can still better reflect the actual compensation requirements during periods of rapid increase, decrease, or reverse change in compensation amount. This indicates that by introducing temporal features such as fusion state, model probability, innovation, and covariance in the IMM filtering process, the compensation amount prediction model can more accurately reflect the launch parameter deviations caused by target maneuvering, thus providing more effective compensation input for subsequent launch corrections.

[0120] Depend on Figure 3As can be seen, in the serpentine maneuver test trajectory, the target's actual trajectory exhibits continuous curvature in three-dimensional space. A significant deviation exists between the intercept point obtained by the uncompensated IMM method and the target position at the moment of impact, indicating that when extrapolating based solely on the filtered state, the serpentine maneuver causes the predicted intercept point to deviate from the actual target position. The IMM+ARMA method provides some correction to the intercept point position, bringing some points closer to the target location, but the overall distribution remains relatively dispersed, and the distance to the target position can still be seen within the magnified local area. In contrast, the intercept points obtained by the IMM+LSTM method are more concentrated near the target position at the moment of impact, with a significantly reduced spatial deviation. This result demonstrates that after using IMM filtering features to predict the launch parameter compensation, the corrected launch direction can better adapt to the serpentine maneuver trajectory, thereby reducing the spatial deviation between the three-dimensional intercept point and the actual target position.

[0121] Depend on Figure 4 It is evident that in the serpentine maneuver scenario, all methods exhibit some interception deviation fluctuations in the initial stage. The IMM+LSTM method also shows significant deviations at certain initial sampling points, primarily due to insufficient historical window information and the incomplete stabilization of the IMM filter fusion state and model probability. As the filtering process gradually converges, the overall interception deviation of each method tends to decrease, but the uncompensated IMM and IMM+ARMA curves still show some fluctuations. In contrast, the IMM+LSTM curve maintains a lower level at most sampling points, especially in the later stages where error fluctuations are more pronounced, generally lower than the other two methods. Statistical results are consistent with the curve trend: among the 50 test sampling points in this scenario, the average interception deviation of uncompensated IMM is approximately 0.877m, IMM+ARMA is approximately 0.819m, while IMM+LSTM reduces it to approximately 0.510m, representing an average improvement rate of approximately 41.8% compared to uncompensated IMM. This indicates that using IMM filter features for compensation prediction can more effectively reduce the interception point deviation caused by serpentine maneuvers and improve the launch parameter correction effect. The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0122] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these modifications and improvements all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for correcting emission parameters based on multidimensional filtering features, characterized in that, It includes an offline training phase and an online application phase, as detailed below: Offline training phase: Step 1: Acquire radar measurement data and high-precision real trajectory data of low-altitude moving targets; perform interactive multi-model filtering based on radar measurement data to obtain target fusion state, model probability, innovation and covariance features; The fusion state is the weighted fusion result of the filtering states of each motion model, the model probability is the probability of each motion model estimated by the interactive multi-model filtering, the innovation is the difference between the measurement prediction value and the actual measurement value of each motion model in the interactive multi-model filtering, and the covariance feature is used to characterize the uncertainty of state estimation and / or measurement prediction in the interactive multi-model filtering process. Step 2: Predict the future trajectory of the target based on the current interactive multi-model filtering state, and solve for the reference emission parameters; The reference launch parameters of the real trajectory are calculated based on high-precision real trajectory data; Launch parameter compensation labels are generated based on the difference between the actual trajectory reference launch parameters and the baseline launch parameters. Step 3: Train the emission parameter compensation prediction model with the fusion state, model probability, innovation and covariance features within the historical window as input and the emission parameter compensation amount label as the expected output. Online application stage: Step 4: Acquire radar measurement data of the target in real time, perform interactive multi-model filtering to obtain the current target fusion state, model probability, innovation and covariance features, predict the future trajectory of the target based on the current filtering state, solve the current reference launch parameters, input the fusion state, model probability, innovation and covariance features in the historical window into the trained launch parameter compensation prediction model to obtain the predicted launch parameter compensation amount, and superimpose the predicted launch parameter compensation amount onto the reference launch parameters to generate the corrected launch parameters of the high-speed projectile interceptor.

2. The emission parameter correction method based on multidimensional filtering features according to claim 1, characterized in that, The filtering features include target fusion state, model probability, innovation, and covariance features; the covariance features are formed by compressing the state covariance matrix and / or innovation covariance matrix obtained by interactive multi-model filtering to form a one-dimensional covariance feature vector; the compression process includes extracting at least one of the diagonal elements, trace, and eigenvalues ​​of the covariance matrix.

3. The emission parameter correction method based on multidimensional filtering features according to claim 1, characterized in that, The radar measurement data includes slant range, azimuth angle, and elevation angle; the radar measurement data is used for measurement updates via interactive multi-model filtering; the historical window consists of multiple consecutive sampling times.

4. The emission parameter correction method based on multidimensional filtering features according to claim 1, characterized in that, The reference launch parameters include at least one of the reference azimuth angle, reference elevation angle, and reference projectile flight time; the launch parameter compensation includes at least one of the azimuth angle compensation, elevation angle compensation, and projectile flight time compensation.

5. The emission parameter correction method based on multidimensional filtering features according to claim 1, characterized in that, The emission parameter compensation prediction model includes at least one long short-term memory network sub-model; the long short-term memory network sub-model is used to encode the filtering features within the historical window and output the emission parameter compensation amount through the output layer. When the emission parameter compensation prediction model includes multiple long short-term memory network sub-models, each sub-model outputs the candidate emission parameter compensation amount.

6. The emission parameter correction method based on multidimensional filtering features according to claim 1, characterized in that, When training the launch parameter compensation prediction model, a mean squared error loss function is used, which is used to measure the difference between the launch parameter compensation predicted by the network and the launch parameter compensation label.

7. The emission parameter correction method based on multidimensional filtering features according to claim 1, characterized in that, The interactive multi-model filter includes at least two target motion models; the target motion models include at least one of a uniform speed model, a uniform acceleration model, a left coordinated turn model, a right coordinated turn model, and a coordinated turn model with different turning rates.

8. The emission parameter correction method based on multidimensional filtering features according to claim 1, characterized in that, The launch parameter compensation prediction model includes one or more sub-predictors. When the compensation prediction model includes multiple sub-predictors, the multiple sub-predictors correspond to different target maneuver types, different model mismatch modes, or different training sample sets. Each sub-predictor outputs candidate launch parameter compensation quantities, and the weight of each sub-predictor is determined according to the sub-predictor adaptability evaluation quantity. The multiple candidate launch parameter compensation quantities are weighted and fused to obtain the final launch parameter compensation quantity.

9. The emission parameter correction method based on multidimensional filtering features according to claim 8, characterized in that, The sub-predictor adaptive evaluation quantity characterizes the degree of deviation between the candidate corrected emission parameters and the filtered trajectory emission parameters. The smaller the deviation, the greater the weight of the sub-predictor. The candidate corrected emission parameters are obtained by superimposing the baseline emission parameters and the compensation amount of the candidate emission parameters output by the corresponding sub-predictor. The filtered trajectory emission parameters refer to the emission parameters obtained by extrapolating the target motion and solving the hit equation based on the interactive multi-model filtered trajectory before the current time.

10. A system for correcting emission parameters based on multidimensional filtering features, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the emission parameter correction method based on multidimensional filtering features as described in any one of claims 1-9.

Images