Deep learning-based fuze initiation time calculation and target trajectory prediction method
By combining deep learning-based BiLSTM and multi-layer LSTM networks with a keyframe weighted loss function, the problem of trajectory prediction accuracy and detonation timing determination of fuzes in complex target maneuvering scenarios is solved, achieving high-precision multi-step prediction and adaptive detonation control, which is suitable for intelligent fuzes and unmanned combat platforms.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-18
- Publication Date
- 2026-03-31
AI Technical Summary
Existing fuze technologies lack sufficient accuracy in trajectory prediction under complex target maneuvering scenarios, have unreliable detonation timing determination, and have poor generalization capabilities, making it difficult to meet the requirements of actual combat safety and damage effectiveness.
A deep learning-based approach is adopted, using BiLSTM and multi-layer LSTM networks for target trajectory prediction. Combined with a custom keyframe weighted loss function, the prediction accuracy at the critical moment of detonation is optimized, and adaptive detonation control is achieved.
It achieves high-precision multi-step prediction of complex maneuvering targets, ensures the reliability of detonation timing and stable adaptability in multiple motion modes, and improves the intelligent decision-making capability of the fuse.
Smart Images
Figure CN121761718A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of fuze control technology, specifically relating to a method for calculating fuze detonation time and predicting target trajectory based on deep learning. Background Technology
[0002] The fuze is a critical component of an ammunition system, and its performance directly affects the destructive effectiveness and safety of the ammunition. Existing fuzes mainly employ preset or simple triggering methods such as timed detonation, contact detonation, and proximity detonation to control detonation. In highly dynamic and intense combat environments, targets often exhibit complex motion characteristics such as high-speed maneuvering, trajectory changes, and sudden speed changes. Traditional fuze methods based on fixed rules or simple filtering estimations struggle to dynamically adjust the detonation timing according to real-time sensing information, leading to a significant decrease in destructive effectiveness or an increased risk of accidental detonation.
[0003] To improve the intelligence level of fuzes, the industry has proposed the concept of "intelligent fuze / adaptive fuze," attempting to achieve dynamic detonation decisions based on multi-sensor data fusion, target recognition, and trajectory estimation. Traditional time-series estimation algorithms (such as Kalman filtering, extended Kalman filtering, and particle filtering) are effective in handling linear or slightly nonlinear targets, but when the target exhibits large maneuvers, strong nonlinearity, or complex noise, the prediction accuracy and robustness of these methods are significantly insufficient, making it difficult to meet the fuze's requirement for high-precision prediction of critical detonation moments.
[0004] In recent years, deep learning methods such as Recurrent Neural Networks (RNNs), Long Short-Term Memory Networks (LSTMs), and Bidirectional LSTMs (BiLSTMs) have made significant progress in trajectory prediction, track estimation, and temporal modeling, enabling them to better capture long-term and short-term temporal dependencies and improve multi-step prediction accuracy. Some studies have attempted to apply these deep models to ammunition damage probability assessment and intelligent fuze decision-making, but existing work still suffers from the following main shortcomings: First, most methods use the overall mean squared error (MSE) or mean absolute error (MAE) as the loss function, treating all moments in the sequence equally, without weighted optimization for the "most important detonation keyframe." Second, in multi-step prediction scenarios, existing models tend to focus on single-step or short-term predictions, lacking specific methods to improve accuracy at future critical moments (such as the end of the prediction window or the moment of maximum damage probability). Third, existing assessments are mostly based on data from limited or single motion modes, and have not yet been systematically validated in various complex maneuver scenarios (such as serpentine, sudden changes, acceleration / deceleration), resulting in insufficient generalization and practical adaptability of the methods.
[0005] In summary, existing technologies cannot provide a system solution that can perform high-precision multi-step trajectory prediction and adaptively determine the detonation timing based on the detonation key frame while meeting the requirements of combat safety and damage effectiveness. Therefore, there is an urgent need for a novel method that can: perform high-precision, multi-step prediction of the target's historical trajectory; optimize the prediction accuracy of the detonation key frame by focusing on the loss function and training targets; and verify its generalization ability in multi-motion mode scenarios, thereby providing a reliable adaptive detonation decision basis for intelligent fuses. The aforementioned problems and needs have been clearly identified in relevant research and serve as the starting point for this invention. Summary of the Invention
[0006] (a) Technical problems to be solved
[0007] The technical problem to be solved by this invention is how to provide a deep learning-based method for calculating the detonation time of a fuse and predicting the target trajectory, so as to solve the problems of insufficient trajectory prediction accuracy, unreliable detonation timing judgment, and poor generalization of existing fuse technologies in complex target maneuvering scenarios.
[0008] (II) Technical Solution
[0009] To address the aforementioned technical problems, this invention proposes a method for calculating fuse detonation time and predicting target trajectory based on deep learning. This method includes the following steps:
[0010] S1. Data Acquisition and Preprocessing: Real-time acquisition of historical trajectory data of the target through sensors, including target position and velocity information; normalization, differential processing, and filtering preprocessing of the trajectory sequence;
[0011] S2, BiLSTM Encoding Layer: The Bidirectional Long Short-Term Memory (BiLSTM) network is used to encode the target's historical trajectory bidirectionally, extracting the temporal dependency information of the forward and backward directions to obtain rich trajectory feature representations.
[0012] S3, Multi-layer LSTM Decoding Layer: Using the output of BiLSTM as input, a multi-layer stacked LSTM network is used to achieve progressive prediction of future multi-step trajectories;
[0013] S4. Custom loss function: During training, the keyframe weighted mean square error loss function is used to give higher weights to specific keyframes in the future prediction sequence, thereby improving the model's prediction accuracy at the critical moment of detonation.
[0014] S5. Adaptive detonation decision: Based on the predicted future trajectory sequence, combined with the relative motion relationship between the ammunition and the target and the damage probability model, the optimal detonation time is automatically determined to achieve adaptive detonation control of the fuze.
[0015] (III) Beneficial Effects
[0016] This invention proposes a method for calculating fuse detonation time and predicting target trajectory based on deep learning. Compared with existing technologies, this invention has the following advantages:
[0017] 1. Higher prediction accuracy: By stacking BiLSTM and deep LSTM, the temporal dependencies of the target's historical trajectory can be fully explored, enabling high-precision multi-step prediction of complex maneuvering targets.
[0018] 2. More reliable detonation decision: By customizing the keyframe weighted loss function, the model focuses on optimizing the prediction accuracy at the critical moment of detonation, ensuring accurate decision-making at the critical time point.
[0019] 3. Strong robustness and adaptability: It can maintain stable prediction performance under various motion modes (such as constant speed, acceleration, serpentine, and sudden change trajectory), and is suitable for complex combat environments.
[0020] 4. High application value: The method of this invention can not only be applied to intelligent fuses, but also extended to target prediction and decision control in unmanned combat platforms and intelligent weapon systems, and has broad engineering application prospects. Attached Figure Description
[0021] Figure 1 This is a diagram of the network hierarchy of the present invention;
[0022] Figure 2 This is a diagram of the algorithm framework structure;
[0023] Figure 3 This is a comparison chart of predicted and actual values for the test samples. Detailed Implementation
[0024] To make the objectives, contents, and advantages of the present invention clearer, the specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples.
[0025] Existing fuze technologies still suffer from problems such as insufficient trajectory prediction accuracy, unreliable detonation timing determination, and poor generalization in complex target maneuvering scenarios, making it difficult to achieve real-time, intelligent adaptive detonation while meeting operational safety and damage effectiveness requirements.
[0026] The purpose of this invention is to provide an adaptive fuze detonation time calculation and target trajectory prediction method based on deep learning, so as to improve the trajectory prediction accuracy, optimize the prediction effect of detonation key frames, and maintain strong robustness in multiple motion modes, thereby realizing intelligent and adaptive decision-making of the fuze.
[0027] To achieve the above objectives, this invention proposes a method for calculating fuse detonation time and predicting target trajectory based on deep learning, the technical solution of which is as follows:
[0028] S1. Data Acquisition and Preprocessing: Real-time acquisition of historical trajectory data of the target through sensors, including information such as target position and velocity; preprocessing of the trajectory sequence such as normalization, differencing, and filtering to improve the robustness of the data and the generalization ability of the model.
[0029] S2, BiLSTM Encoding Layer: The Bi-directional Long Short-Term Memory (BiLSTM) network is used to encode the target's historical trajectory bi-directionally, extracting forward and backward temporal dependency information to obtain rich trajectory feature representations.
[0030] S3, Multi-layer LSTM Decoding Layer: Using the output of BiLSTM as input, a multi-layer stacked LSTM network is used to progressively predict future multi-step trajectories. It can combine autoregressive prediction and parallel prediction methods.
[0031] S4. Custom loss function: During training, a keyframe weighted mean square error loss function is used to assign higher weights to specific keyframes in the future prediction sequence (such as the optimal detonation time or the end of the prediction window), thereby improving the model's prediction accuracy at the critical detonation moment.
[0032] S5. Adaptive detonation decision: Based on the predicted future trajectory sequence, combined with the relative motion relationship between the ammunition and the target and the damage probability model, the optimal detonation time is automatically determined to achieve adaptive detonation control of the fuze.
[0033] Example 1:
[0034] The present invention relates to a method for calculating the detonation time and predicting the target trajectory for an adaptive fuze. This method takes the target's historical trajectory sequence as input, and after preprocessing, deep temporal network prediction, keyframe weighted loss training, and decision module calculation, outputs the optimal detonation time point (or several candidate time points) for fuze control. The implementation steps and parameters are described in detail below with several embodiments.
[0035] Step S0, Overall Framework and Problem Description
[0036] This implementation method addresses the problem of multi-step trajectory prediction and adaptive detonation timing determination in smart fuses: given a target in a time sequence Historical position and status, predicting the next K steps The target location is determined, and the optimal detonation time is calculated and determined based on the predicted sequence. The overall process includes: dataset and simulation, data preprocessing, BiLSTM encoding, multi-layer LSTM decoding, custom keyframe weighted loss training, and prediction-based detonation determination.
[0037] Step S1: Dataset Design (Multi-Motion Mode Simulation Dataset)
[0038] To cover multiple target motion patterns, a target synthetic dataset was constructed, containing at least three typical motion patterns:
[0039] 1. Uniform linear motion;
[0040] 2. S-curve (snake-like) maneuver;
[0041] 3. Acceleration / deceleration speed change.
[0042] Each trajectory contains 20 historical frames and 10 future frames. Scale transformation, position offset, and Gaussian noise are added to the dataset to enhance generalization ability. Training / validation / testing are split in an 8:2 or 8:1:1 ratio to ensure that each motion pattern has representative samples in each subset.
[0043] Data preprocessing:
[0044] Collect raw trajectory data and preprocess it according to the following procedure:
[0045] S11. Historical target observation and collection: Real-time collection of observation data such as the target's spatiotemporal position and velocity to form a historical trajectory sequence of the target.
[0046] S12. Convert the original position into coordinates and differential features to form the vector input at each time step.
[0047] S13. Perform mean normalization on the trajectory sequence to facilitate training convergence.
[0048] S14. Apply Gaussian noise to the position signal to model it and use filtering to suppress spikes.
[0049] Step S2: BiLSTM coding layer (historical feature extraction)
[0050] Target historical state sequence ,in Includes target at time Features such as spatial coordinates and velocity.
[0051] Future multi-step prediction sequence Each step represents the spatial state (spatial coordinates, velocity, and other characteristics) of the target at a future time.
[0052] BiLSTM performs forward and backward encoding on the input sequence to obtain rich temporal context features. Its forward and backward hidden states are as follows:
[0053]
[0054]
[0055] In the BiLSTM coding layer For the forward LSTM at time The hidden state, For backward LSTM at time The hidden state.
[0056] Step S3: Multi-layer LSTM decoding layer (multi-step prediction)
[0057] Stacking multiple LSTM layers to receive BiLSTM outputs progressively generates future multi-step states:
[0058]
[0059]
[0060] Two prediction modes are used: autoregressive (prediction from the previous time step is used as input for the next time step) and fully parallel prediction. The network hierarchy is as follows: Figure 1 As shown.
[0061] in, for and The concatenated vector; It is the decoder LSTM at time... The hidden state or internal state vector is used to progressively predict the target position at the next moment and is the core internal variable for generating future sequences. This is the parameter matrix of the fully connected output layer (linear layer), used to store the hidden states of the LSTM. Mapping to the predicted output space (e.g., 3D coordinates) ). It is the bias vector of the output layer, used together with the weight matrix to form a linear mapping.
[0062] Step S4: Design of Keyframe Weighted Loss Function and Training Objective
[0063] To ensure high accuracy of the model at future keyframes (such as the optimal detonation time), this patent proposes a keyframe-weighted mean square error loss:
[0064]
[0065] in, For loss weights, if (Keyframe), then Take a larger value, and then appropriately reduce it at other times. For example, only apply a weighted increase in the last step. Compare this to the traditional mean squared error loss:
[0066]
[0067] This patented loss function is better suited to the high precision requirements of ammunition fuses at specific future moments.
[0068] in, The target predicted by the model in the future... Spatial location at any given moment With the goal of the future The true value of the spatial location at each moment.
[0069] The model employs a deep hybrid LSTM structure, including BiLSTM, Dropout, and multi-layer LSTM stacking, which can both capture the temporal dependencies of historical trajectories and improve the model's generalization ability. The output layer is a fully connected regression layer, ultimately predicting the position for the next 10 frames. The algorithm framework is as follows: Figure 2 As shown.
[0070] Multi-motion mode assessment and testing:
[0071] The model was validated and evaluated on a synthetic multi-motion mode dataset (uniform, S-shaped, and accelerating), and the MSE and MAE for the next K steps were calculated. The model's average MSE ≈ 0.0034 and MAE ≈ 0.0461 for the next 10 frames, indicating that the model can achieve high-precision predictions under various complex target motion scenarios. In high-maneuverability S-shaped and accelerating target scenarios, the model's prediction error slightly increased, but the overall trajectory trend remained well-matched, indicating that the designed deep LSTM structure possesses strong temporal modeling and generalization capabilities. By introducing data augmentation and regularization, the model maintained stable performance even in noisy environments.
[0072] Step S5: Determining the detonation timing based on prediction (decision module)
[0073] In real-time operation, detonation determination is made using the predicted sequence and the current ammunition state:
[0074] S51, For each prediction frame The system utilizes The relative distance, relative velocity, azimuth difference, and other geometric quantities are calculated based on the current state of the ammunition, and the corresponding damage probability is obtained through an empirical damage model to select the optimal detonation frame.
[0075] S52. The decision criteria can be set to maximize the probability of damage or minimize the risk of accidental damage, thereby selecting the optimal detonation frame and improving the effectiveness of the ammunition.
[0076] Beneficial effects:
[0077] Compared with the prior art, the present invention has the following advantages:
[0078] 1. Higher prediction accuracy: By stacking BiLSTM and deep LSTM, the temporal dependencies of the target's historical trajectory can be fully explored, enabling high-precision multi-step prediction of complex maneuvering targets.
[0079] 2. More reliable detonation decision: By customizing the keyframe weighted loss function, the model focuses on optimizing the prediction accuracy at the critical moment of detonation, ensuring accurate decision-making at the critical time point.
[0080] 3. Strong robustness and adaptability: It can maintain stable prediction performance under various motion modes (such as constant speed, acceleration, serpentine, and sudden change trajectory), and is suitable for complex combat environments.
[0081] 4. High application value: The method of this invention can not only be applied to intelligent fuses, but also extended to target prediction and decision control in unmanned combat platforms and intelligent weapon systems, and has broad engineering application prospects.
[0082] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A fuze initiation time calculation and target trajectory prediction method based on deep learning, characterized in that, The method comprises the following steps: S1, data acquisition and preprocessing: real-time acquisition of target historical trajectory data by sensors, including target position and speed information; normalization, difference and filtering preprocessing of the trajectory sequence; S2, BiLSTM encoding layer: bidirectional encoding of the target historical trajectory by using a bidirectional long short-term memory network BiLSTM, extracting forward and backward time sequence dependency information, and obtaining rich trajectory feature representation; S3, multi-layer LSTM decoding layer: taking the BiLSTM output as input, and using a multi-layer stacked LSTM network to realize progressive prediction of future multi-step trajectory; S4, self-defined loss function: in the training process, a key frame weighted mean square error loss function is used, and higher weight is given to specific key frames in the future prediction sequence, so as to improve the prediction accuracy of the model at the critical moment; S5, adaptive detonation decision: based on the predicted future trajectory sequence, the relative motion relationship between the ammunition and the target and the damage probability model are combined to automatically determine the optimal detonation time, so as to realize the adaptive detonation control of the fuze.
2. The deep learning-based fuze initiation time calculation and target trajectory prediction method of claim 1, wherein, In the S1, the real-time acquisition of target historical trajectory data by sensors comprises: constructing a target synthetic data set, which contains at least three typical motion modes: uniform linear motion, S-shaped curve maneuver and acceleration / deceleration variable speed motion; each trajectory contains a historical sequence and a future sequence, and scale transformation, position offset and Gaussian noise are added to the data set to enhance the generalization ability. 3.The fuze burst time calculation and target trajectory prediction method based on deep learning of claim 2, wherein, In the S1, the normalization, difference and filtering preprocessing of the trajectory sequence comprises: S11, historical target observation collection: real-time collection of target space-time position and speed observation data to form a target historical trajectory sequence; S12, converting the original position into coordinates and difference features to form a vector input at each time; S13, mean normalization processing of the trajectory sequence to facilitate training convergence; S14, applying Gaussian noise modeling to the position signal and using filtering to suppress spikes. 4.The fuze burst time calculation and target trajectory prediction method based on deep learning according to claim 3, wherein, The S2 comprises: Target history state sequence wherein comprises spatial coordinates, velocity characteristics of the target at time points Future multi-step prediction sequence Each step is a target spatial state at a future time, including spatial coordinates, velocity characteristics BiLSTM encodes the input sequence forward and backward to obtain rich time sequence context features, and the forward and backward hidden states are respectively: wherein, in the BiLSTM encoding layer is the hidden state of the forward LSTM at time is the hidden state of the backward LSTM at time is the hidden state of the backward LSTM at time is the hidden state of the backward LSTM at time 5.The fuze initiation time calculation and target trajectory prediction method based on deep learning according to claim 4, wherein, The S3 comprises: Stacked multi-layer LSTM receives the BiLSTM output to progressively generate future multi-step states: where, is the concatenation of and ; is the hidden state or internal state vector of the decoder LSTM at time , which is the core internal variable for the generation of future sequences, is used to recursively predict the target position at the next time step; is the parameter matrix of the fully connected output layer, which maps the hidden state of the LSTM to the output space of the prediction, is the bias vector of the output layer, which is used together with the weight matrix to form a linear mapping. 6.The fuze initiation time calculation and target trajectory prediction method based on deep learning according to claim 5, wherein, In the S3, two modes of autoregression and full parallel prediction are used.
7. The deep learning-based fuze blast time calculation and target trajectory prediction method of claim 1, wherein, In the S4, in order to ensure high accuracy of the model at the future key frame, a key frame weighted mean square error loss is used: wherein, is the loss weight, if , corresponding to the key frame, then take the larger value, and the rest of the time appropriately reduce; wherein, is the model predicted spatial position of the target at a future time step is the true spatial position of the target at a future time step is the true spatial position of the target at a future time step 8.The fuze initiation time calculation and target trajectory prediction method based on deep learning according to claim 7, wherein, In the S4, the key frame corresponds to the optimal detonation time or the end of the prediction window. 9.The fuze initiation time calculation and target trajectory prediction method based on deep learning of claim 7, wherein, In the S4, the model is evaluated on the synthetic multi-motion mode data set, and the MSE and MAE of the future K steps are calculated, and the multi-motion mode includes: uniform linear motion, S-shaped curve maneuver and acceleration / deceleration variable speed motion. 10.The fuze burst time calculation and target trajectory prediction method based on deep learning according to claim 1, wherein, The S5 comprises: Using the prediction sequence and the current ammunition state to determine the detonation: S51, For each prediction frame The system utilizes The relative distance, relative velocity, and azimuth difference geometry are calculated relative to the current state of the ammunition, and the corresponding damage probability is obtained through an empirical damage model to select the optimal detonation frame. S52, setting the decision criterion as maximizing the damage probability or minimizing the risk of injury, so as to select the optimal detonation frame and improve the ammunition effectiveness.