Fuse accurate height measurement method based on parameterized model and deep learning
By combining parametric models with deep learning, the problems of low ranging accuracy and poor scene adaptability of radio fuses under complex terrain conditions were solved, achieving accurate altitude measurement of the fuse, improving ranging accuracy and enhancing the adaptability of the model.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING INST OF TECH
- Filing Date
- 2026-01-20
- Publication Date
- 2026-05-12
AI Technical Summary
Existing empirical models have low ranging accuracy and poor scene adaptability under complex terrain conditions, making it difficult to achieve accurate altitude measurement.
A parametric model and deep learning-based approach is adopted to construct a parametric ground echo simulator for electromagnetic scattering. A deep fully connected network is used to extract echo signal features and estimate electromagnetic scattering parameters. Gradient boosting trees are combined to correct the error of height information, thereby achieving accurate height measurement of the fuze.
It improves the ranging accuracy of the fuze under complex terrain conditions, enhances scene adaptability, reduces altitude measurement error, and simplifies model deployment.
Smart Images
Figure CN122020371A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of radio fuze technology, specifically relating to a precise height measurement method for fuzes based on parametric models and deep learning. Background Technology
[0002] The electromagnetic scattering characteristics of complex terrain significantly affect the echo signal of radio fuzes, and obtaining typical ground parameters is crucial for achieving high ranging accuracy in complex terrain environments. Existing empirical models have weak generalization ability and poor interpretability, and their parameter values often need to be matched with different environmental conditions, fuze characteristics, and signal processing methods. To address the issues of reduced ranging accuracy and poor scene adaptability of radio fuzes in complex terrain conditions, a precise altimeter measurement method based on parametric models and deep learning is proposed.
[0003] Existing classic empirical models of ground scattering coefficients, such as the GIT model and the Wollaby model, are mostly based on measured data of various ground features obtained from commonly used radar bands. They derive corresponding empirical formulas through curve fitting, which is insufficient to cover all terrain types. Furthermore, empirical models are not adaptable enough to non-uniform terrains, and their parameter assumptions deviate from the real-world scenario. Summary of the Invention
[0004] In view of this, the present invention proposes a method for accurate height measurement of fuses based on parametric models and deep learning, which can achieve accurate height measurement of fuses.
[0005] The technical solution for implementing the present invention is as follows:
[0006] Firstly, this invention provides a precise height measurement method for fuses based on a parameterized model and deep learning, the specific process of which is as follows: Dataset Acquisition: First, the electromagnetic scattering parameter combinations used in the ground echo model establishment process are determined, including the Ulaby model fitting parameters, incident angle, and height. Sampling is performed in the multidimensional prediction parameter space to obtain N sets of electromagnetic scattering parameter combinations. Second, a controllable electromagnetic scattering parameterized ground echo simulator is constructed based on the improved Ulaby model. This ground echo simulator incorporates a correction factor. The echo spectrum characteristics of different landforms are simulated. Finally, the simulator is used to model the echo of N sets of parameters to obtain the original echo signal. The features of the original echo signal are extracted and combined into a high-dimensional feature vector as the input of the deep learning model. Network model training: Construct a deep fully connected network, using feature vectors as input and estimated values of electromagnetic scattering parameters as labels to train the network model; Measured data height prediction and correction: Collect actual fuze echo data, extract high-dimensional feature vectors, input them into a trained network model, and output the estimated value of the electromagnetic scattering parameter combination to be optimized; use the fuze echo data for inversion to optimize the estimated value of the electromagnetic scattering parameter combination to be optimized, and correct the height information in the optimized electromagnetic scattering parameters to achieve accurate fuze height measurement.
[0007] Optionally, the present invention utilizes the fuze echo data for inversion to optimize the estimated values of the electromagnetic scattering parameters to be optimized, and the optimization objective function is:
[0008] in: The first, representing the measured echo signal j 1 eigenvalue, The first electromagnetic scattering parameter calculated to be optimized is the first... j One eigenvalue; With q as the step size, the electromagnetic scattering parameters are selected within the set interval [Qmin, Qmax] as the optimization parameter interval, and the estimated electromagnetic scattering parameters corresponding to the minimum objective function are calculated. Repeat the above steps until the objective function meets the set requirements, and obtain the estimated values of the inverted electromagnetic scattering parameters.
[0009] Optionally, the present invention corrects the height information in the optimized electromagnetic scattering parameters by means of the following: Collect real echo data of the ground at known heights, and use a trained DNN to obtain estimated values of the inverted electromagnetic scattering parameters; use the inverted electromagnetic scattering parameters as input and the error between the real height and the measured height as output, and train a linear regression correction model based on gradient boosting tree. The linear regression model is used to correct the error in the height information of the optimized electromagnetic scattering parameters.
[0010] Optionally, the ground echo simulator model of the present invention is as follows:
[0011] in, N U It is the total number of scattering units within the ground echo region. P t It is the power of the transmitted signal. G i It is the first i Antenna gain of each scattering element s i It is the backscattering coefficient of the scattering unit. l It is the operating wavelength. Ri ( t ) is the current moment from the th i The distance from each scattering unit to the detector, It is the time delay of the target echo signal relative to the transmitted signal. The phase shift is caused by the reflection of the target. f 0 is the carrier frequency. It is the frequency modulation slope.
[0012] Optionally, the present invention uses measured data to fit the frequency domain of echo signals under different terrain conditions to obtain corresponding empirical correction factors. .
[0013] Optionally, when extracting the features of the original echo signal, this invention uses echo signals from several consecutive cycles, integrates them into a time-domain matrix, processes the multi-cycle echo signal using statistical data analysis, and extracts the mean of the time-domain features of each cycle echo signal; performs Fourier transform on the multi-cycle echo signal to extract the mean of the spectral features; uses short-time Fourier transform to obtain the joint time-frequency distribution of each cycle signal and extracts the mean of the time-spectral features; calculates the mean of the energy features of each cycle echo; and concatenates and splices the extracted features to form a high-dimensional echo signal feature vector.
[0014] Optionally, this invention utilizes a Latin hypercube sampling method to obtain a full-coverage dataset of the high-dimensional parameter space, thereby obtaining N sets of parameter combinations.
[0015] Optionally, the present invention performs normalization processing on the feature vector, mapping the values in the feature vector to the range (0,1).
[0016] Optionally, the present invention randomly divides the obtained dataset into a training set, a validation set, and a test set according to a set ratio, and trains, validates, and tests the network model.
[0017] Optionally, the neural network model of the present invention uses mean squared error as the loss function to make the model converge; and uses the gradient descent algorithm to update the gradients of the weights and biases according to the loss function.
[0018] Beneficial effects: First, this invention abstracts the complex surface into several key electromagnetic scattering parameters, uses a deep learning network to establish a mapping relationship between echo signal characteristics and electromagnetic scattering parameters, and inverts the characteristics of the actual collected fuze echo signal to obtain the estimated values of electromagnetic scattering parameters corresponding to the measured data.
[0019] Secondly, the present invention also utilizes fuze echo data for inversion to optimize the estimated value of the electromagnetic scattering parameter combination to be optimized, and performs error correction on the height information in the optimized electromagnetic scattering parameters to achieve accurate fuze height measurement.
[0020] Third, this invention overcomes the shortcomings of traditional single-parameter models by using multi-parameter joint inversion, reducing altimetry errors while having wider applicability. The model does not rely on a large amount of actual collected data, and the lightweight model is easy to deploy. Attached Figure Description
[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0022] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation
[0023] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
[0024] It should be noted that, in the absence of conflict, the following embodiments and features can be combined with each other; and, based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.
[0025] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this disclosure, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.
[0026] This application proposes a precise height measurement method for fuses based on parametric models and deep learning. The specific process is as follows: Step 1: Determine the core parameters The core parameters for establishing the echo model were determined, including the Ulaby model fitting parameters P1-P6, the incident angle θ, and the height h0. The Latin hypercube sampling method was used to sample in the multidimensional prediction parameter space, ensuring that reasonable parameter regions were uniformly covered, generating a total of N sets of parameters.
[0027] Step 2, Ground Echo Model Based on the relevant parameters of the detector and antenna, the boundary of the ground region at a fixed height is calculated. The ground echo region is then divided into sections with side length Δ using the square uniform segmentation method. l Multiple scattering units are defined, and the center coordinates of each scattering unit are calculated as follows: x i , y i ,0).
[0028] Under different parameter combinations, the backscattering coefficient can be expressed as:
[0029] Therefore, the radar cross-section σ of each scattering element relative to the detector i for:
[0030] Antenna gain of each scattering element G i If it is exponentially related to the local incident angle, then G i It can be represented as:
[0031] The echo region can be considered as a large number of approximately equal scattering units; therefore, random ground fluctuations can be considered as an exponential distribution.
[0032] Therefore, the echo power of each scattering unit P ri It can be represented as:
[0033] For a point target, the time-domain waveform of the transmitted signal modulated by the sawtooth wave frequency can be expressed as:
[0034] in, A 0,i It is the amplitude of the transmitted signal. It is the initial phase. f 0 is the carrier frequency. It is the frequency modulation slope. It is frequency modulation offset. T m It is the period of the modulated signal.
[0035] Assume the amplitude attenuation of the echo signal is A 1,i echo signal s r,i (t This can be represented as:
[0036] in, It is the time delay of the target echo signal relative to the transmitted signal. R 0 represents the initial distance from the target to the detector. v r It is the relative speed between the missile and the target. c It's the speed of light. The phase shift is caused by the reflection of the target.
[0037] echo signal s r,i ( t ) and transmitted signals s t,i ( t The mixture is then filtered through a low-pass filter to obtain the point target difference frequency signal. s b,i ( t Complex forms of )
[0038] Based on the radar equation and signal power P t With amplitude A 1,i The relationship between the point target echo difference frequency signals of each scattering unit and the superposition of the point target echo difference frequency signals yields the surface target echo difference frequency signal s. Eb ( t ):
[0039] in, N U It is the total number of scattering units within the ground echo region. P t It is the power of the transmitted signal. G i It is the first i Antenna gain of each scattering element s i It is the backscattering coefficient of the scattering unit. l It is the operating wavelength. R i ( t ) is the current moment from the th i The distance from each scattering unit to the detector.
[0040] As shown in the formulas above, the influence of different terrain parameters on the echo signal in the Ulaby model is only reflected in the time-domain signal amplitude. However, under normal circumstances, different terrains have a certain impact on the frequency offset and broadening of the echo signal in the frequency domain. Since this step aims for accurate ranging, spectral broadening will not directly cause ranging errors; therefore, its influence (spectral broadening) is combined with the frequency offset into a single correction factor ζ. The frequency domain of the echo signal under different terrain conditions is fitted using measured data to obtain the corresponding empirical correction factor. The difference frequency signal model of the opposite target echo (i.e., the echo simulator) is then changed as follows:
[0041] Step 3: Echo signal feature extraction Batch simulations were performed with N sets of parameters to obtain the corresponding original echo signals. To eliminate systematic errors, echo signals from several consecutive periods were used and integrated into a time-domain matrix. Statistical analysis was used to process the multi-period echo signals, extracting the mean (TS1, TS2, ..., TSm) of the time-domain characteristics of each period's echo signal.
[0042] Perform Fourier transform (FFT) on the multi-cycle echo signal to extract the mean spectral features (TP1, TP2, ..., TPn).
[0043] The joint time-frequency distribution of each period signal is obtained by using the short-time Fourier transform (STFT), and the mean of the time-spectrum features (SP1, SP2, ..., SPx) is extracted.
[0044] Calculate the mean energy characteristics (TE1, TE2, ..., TEy) of the echoes in each period.
[0045] All the above features are concatenated and combined to form a high-dimensional echo signal feature vector F=[TS1,TS2,…,TSm,TP1,TP2,…,TPn,SP1,SP2,…,SPi,TE1,TE2,…,TEj] Step 4, Deep Learning Model Training The feature vectors are normalized, mapping the values to the range (0,1) to accelerate model convergence. For any feature in the feature vector, its normalized value is Feature':
[0046] The entire dataset (feature vectors as labels and parameter combinations as labels) is randomly divided into training set (N×x1), validation set (N×x2) and test set (N×x3) according to the ratio of x1:x2:x3 (where x1+x2+x3=1).
[0047] Since the input is a feature vector, a deep fully connected network (DNN) is constructed. The DNN consists of an input layer, multiple hidden layers, and an output layer, with each layer fully connected to the layer below it. The DNN structure is as follows: Forward propagation is the process of data flowing from the input layer to the output layer, with the purpose of calculating predicted values.
[0048] Linear computation: The input value of each node is the weighted sum of the output values of all nodes in the previous layer, plus a bias term.
[0049]
[0050] in, It is the first l The linear output of the layer, It is the connection of the first l -1st floor and the l Weight matrix between layers It is the first l The output of layer -1 (i.e., the -1st layer) l (input of the layer) It is the first l The bias vector of the layer.
[0051] Activation function: Passes the linear output z through a non-linear activation function, ensuring that the output value is any function.
[0052]
[0053] The model is trained by taking the echo model feature vector as input and the modeling parameters as output, and the mean squared error (MSE) is used as the loss function to make the model converge.
[0054]
[0055] in, For the first i Predicted values of each feature value For the first i The actual value of each eigenvalue. m The number of features in the feature vector.
[0056] The gradient descent algorithm is used to update the gradients of the weights and biases based on the loss function.
[0057]
[0058]
[0059] in, The learning rate controls the step size for each update.
[0060] The model is trained using the training set, and its performance is validated using the validation set to prevent overfitting. The network hyperparameters (learning rate, number of network layers, number of nodes, etc.) are then adjusted.
[0061] The mean squared error (MSE) of the final model is evaluated on an independent test set to ensure that the model has good generalization ability.
[0062] Step 5: Application of measured data and height correction Under fixed fuze operating parameters, actual fuze echo data were collected. These measured data were processed using the same procedure as the simulation data for feature extraction and standardization. The data were then input into the trained model, which outputs estimated values of the electromagnetic scattering parameters (P1-P6, incident angle θ, height h0) corresponding to the fuze echo.
[0063] The characteristics of the measured fuze echo signal are inverted using an interval sampling parameter update method to optimize the estimated electromagnetic scattering parameters corresponding to the measured data. The objective function is defined as follows:
[0064] in: The first, representing the measured echo signal j 1 eigenvalue, f p [ j [The first] is the electromagnetic scattering parameter calculated to be optimized. j Each feature value.
[0065] Using q as the step size, the electromagnetic scattering parameters are selected within the interval [Qmin, Qmax] as the optimization parameter interval, and the estimated electromagnetic scattering parameters corresponding to the minimum objective function are calculated. This process is repeated until the objective function meets the initial requirement (less than ε), yielding the inverted estimated electromagnetic scattering parameters.
[0066] Collect a batch of real ground echo data at known altitudes, repeat the above steps, and obtain the estimated values of the electromagnetic scattering parameters retrieved from them.
[0067] Using the inverted electromagnetic scattering parameters as input and the error between the true height and the measured height as output, a linear regression correction model is trained based on gradient boosting tree (GBDT).
[0068] The inverted electromagnetic scattering parameter vector F As input, the error between the actual height and the model-inverted height is used as output. y = H true - H measured A regression model based on gradient boosting tree (GBDT) was established.
[0069]
[0070] Use the absolute value between the prediction error and the actual error The corrected height is used as the loss function to bring the model to convergence. H corrected for:
[0071] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A precise height measurement method for fuses based on parametric models and deep learning, characterized in that, The specific process is as follows: Dataset Acquisition: First, the electromagnetic scattering parameter combinations used in the ground echo model establishment process are determined, including the Ulaby model fitting parameters, incident angle, and height. Sampling is performed in the multidimensional prediction parameter space to obtain N sets of electromagnetic scattering parameter combinations. Second, a controllable electromagnetic scattering parameterized ground echo simulator is constructed based on the improved Ulaby model. This ground echo simulator incorporates a correction factor. The echo spectrum characteristics of different landforms are simulated. Finally, the simulator is used to model the echo of N sets of parameters to obtain the original echo signal. The features of the original echo signal are extracted and combined into a high-dimensional feature vector as the input of the deep learning model. Network model training: Construct a deep fully connected network, using feature vectors as input and estimated values of electromagnetic scattering parameters as labels to train the network model; Measured data height prediction and correction: Collect actual fuze echo data, extract high-dimensional feature vectors, input them into a trained network model, and output the estimated value of the electromagnetic scattering parameter combination to be optimized; use the fuze echo data for inversion to optimize the estimated value of the electromagnetic scattering parameter combination to be optimized, and correct the height information in the optimized electromagnetic scattering parameters to achieve accurate fuze height measurement.
2. The precise height measurement method for fuses based on parametric models and deep learning according to claim 1, characterized in that, The inversion using the fuze echo data is used to optimize the estimated values of the electromagnetic scattering parameters to be optimized. The objective function for optimization is: in: The first, representing the measured echo signal j 1 eigenvalue, The first electromagnetic scattering parameter calculated to be optimized is the first... j One eigenvalue; With q as the step size, the electromagnetic scattering parameters are selected within the set interval [Qmin, Qmax] as the optimization parameter interval, and the estimated electromagnetic scattering parameters corresponding to the minimum objective function are calculated. Repeat the above steps until the objective function meets the set requirements, and obtain the estimated values of the inverted electromagnetic scattering parameters.
3. The precise height measurement method for fuses based on parametric models and deep learning according to claim 2, characterized in that, The error correction of the height information in the optimized electromagnetic scattering parameters is specifically as follows: Collect real echo data of the ground at known heights, and use a trained DNN to obtain estimated values of the inverted electromagnetic scattering parameters; use the inverted electromagnetic scattering parameters as input and the error between the real height and the measured height as output, and train a linear regression correction model based on gradient boosting tree. The linear regression model is used to correct the error in the height information of the optimized electromagnetic scattering parameters.
4. The precise height measurement method for fuses based on parametric models and deep learning according to claim 2, characterized in that, The ground echo simulator model is as follows: in, N U It is the total number of scattering units within the ground echo region. P t It is the power of the transmitted signal. G i It is the first i Antenna gain of each scattering element σ i It is the backscattering coefficient of the scattering unit. λ It is the operating wavelength. R i ( t ) is the current moment from the th i The distance from each scattering unit to the detector, It is the time delay of the target echo signal relative to the transmitted signal. The phase shift is caused by the reflection of the target. f 0 is the carrier frequency. It is the frequency modulation slope.
5. The precise height measurement method for fuses based on parametric models and deep learning according to claim 4, characterized in that, The method involves fitting the frequency domain of echo signals under different terrain conditions using measured data to obtain corresponding empirical correction factors. .
6. The precise height measurement method for fuses based on parametric models and deep learning according to claim 1, characterized in that, When extracting the features of the original echo signal, several consecutive cycles of echo signals are used and integrated into a time-domain matrix; statistical analysis of the data is used to process the multi-cycle echo signal and extract the mean of the time-domain features of each cycle of echo signal. Perform Fourier transform on the multi-cycle echo signal to extract the mean spectral features; use short-time Fourier transform to obtain the joint time-frequency distribution of each cycle signal and extract the mean time-frequency features; calculate the mean energy features of each cycle echo; concatenate and splice the extracted features to form a high-dimensional echo signal feature vector.
7. The precise height measurement method for fuses based on parametric models and deep learning according to claim 6, characterized in that, The feature vector is normalized to map the values in the feature vector to the range (0,1).
8. The precise height measurement method for fuses based on parametric models and deep learning according to claim 7, characterized in that, The obtained dataset is randomly divided into training, validation, and test sets according to a set ratio, and the network model is trained, validated, and tested.
9. The precise height measurement method for fuses based on parametric models and deep learning according to claim 1, characterized in that, The neural network model uses mean squared error as the loss function to achieve model convergence; the gradient descent algorithm is used to update the gradients of the weights and biases according to the loss function.
10. The precise height measurement method for fuses based on parametric models and deep learning according to claim 1, characterized in that, By using Latin hypercube sampling to sample a full-coverage dataset in the high-dimensional parameter space, N sets of parameter combinations are obtained.