Method for improving laser fuze target detection precision
Through the methods of self-supervised learning and multi-source data fusion, the target detection accuracy of the laser fuze in complex environments is improved, the problem of insufficient detection accuracy in existing technologies is solved, and efficient and stable target recognition and tracking are achieved.
Patent Information
- Application Number
- CN202510977003.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-16
- Publication Date
- 2025-10-10
AI Technical Summary
Existing laser fuzes have insufficient target detection accuracy in complex environments, especially under conditions of multiple targets, high-speed motion and strong interference, where errors are large and traditional algorithms are ineffective.
By optimizing laser reflection signal processing, adopting self-supervised learning mechanism to train convolutional neural network, combining multi-source heterogeneous data fusion and adaptive Kalman filtering algorithm, the target detection accuracy is improved.
It significantly improves target recognition and tracking capabilities in complex environments, reduces the probability of false or missed triggers, ensures system stability and real-time performance, and can run efficiently in resource-constrained hardware.
Smart Images
Figure CN120760550A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of laser fuzes, and in particular relates to a method for improving the target detection accuracy of laser fuzes. Background Art
[0002] Laser fuzes use laser beams to detect targets and trigger fuzes, enabling them to detect and track high-speed, long-range targets. The basic principle is to emit a laser beam and receive the laser signal reflected by the target. Based on information such as the signal's time delay and intensity, the laser fuze is then triggered at the appropriate time to effectively strike the target.
[0003] However, existing laser fuzes often face problems with insufficient target detection accuracy and large errors in complex environments (such as smoke, rain, snow, strong light interference, etc.). In addition, factors such as target reflection characteristics, laser emission power fluctuations, and received signal noise can also affect detection accuracy. Traditional laser signal processing algorithms mostly rely on simple mathematical models, such as calculations based on time difference or frequency offset. Although this method is effective, it may result in large errors in complex environments, especially when there are multiple targets and high-speed movement. Therefore, how to improve the target detection accuracy of laser fuzes and enhance target recognition and tracking capabilities in complex environments is becoming increasingly important in the field of laser fuzes. Summary of the Invention
[0004] The purpose of the present invention is to provide a method for improving the target detection accuracy of a laser fuze, by optimizing the processing process of laser reflection signals and improving target detection and tracking algorithms, so as to improve the target detection accuracy of the system in complex dynamic environments.
[0005] The technical solution adopted by the present invention to achieve the above-mentioned object is: a method for improving the target detection accuracy of a laser fuze, comprising the following steps:
[0006] Step S1: obtaining a laser reflection signal through a laser transmitter and receiver module, and preprocessing the laser reflection signal to obtain a preprocessed laser reflection signal;
[0007] Step S2: Train the convolutional neural network through a self-supervised learning mechanism, and construct a training data set by fusing multi-source heterogeneous data to complete pre-training;
[0008] Step S3: Extract multi-scale features from the preprocessed signal through the trained convolutional neural network, generate candidate target boxes based on the region proposal network, and obtain the target classification result;
[0009] Step S4: Based on the target classification result, the 3D size template and reflection characteristics of the corresponding category in the local target model library are called, and the precise position of the target is located by combining the Bayesian uncertainty estimation;
[0010] Step S5: Adopting the adaptive Kalman filter algorithm regulated by the neural network to further optimize the positioning accuracy of the target and realize the target motion state tracking.
[0011] The pre-processing in step S1 includes: denoising, normalization and adaptive gain adjustment;
[0012] Wherein, the denoising process uses wavelet transform to achieve frequency domain noise removal;
[0013] The adaptive gain adjustment optimizes the signal gain in real time according to the ambient noise intensity.
[0014] The convolutional neural network uses MobileNetV3-Small as the backbone network. After self-supervised pre-training, it uses the measured data of the fuse to perform small sample fine-tuning.
[0015] The small sample fine-tuning adopts freezing the feature layer to train the Softmax classification head, or unfreezing the layer by layer to perform fine-tuning of the entire network.
[0016] In step S2, the self-supervised learning mechanism adopts the SimCLR contrastive learning framework, which specifically includes:
[0017] S2-1: Apply random enhancement transformations to the laser echo image to generate positive sample pairs; the enhancement operations include: adding noise, brightness adjustment, local occlusion, affine transformation, geometric rotation, and frequency domain perturbation;
[0018] S2-2: Features are extracted through a lightweight CNN backbone network and converted into embedding vectors through the projection head;
[0019] S2-3: The contrastive loss function is used to optimize the network to minimize the embedding distance of enhanced views of the same target, that is:
[0020]
[0021] Among them, z i ,z j The enhanced view embedding vector representing the same object, sim(z i ,z k ) is the cosine similarity, and τ represents the temperature hyperparameter.
[0022] The step S2-2 comprises the following steps:
[0023] 1) The preprocessed laser echo signal is input into the backbone network, passing through the input layer, multiple Bottleneck layers, and the output layer in sequence;
[0024] Input layer, including: 3×3 convolution kernel, batch normalization and h-swish activation function;
[0025] Bottleneck layer, used to extract multi-scale features using depth-wise separable convolution and SE attention module;
[0026] Output layer, used for global average pooling to generate a one-dimensional feature vector;
[0027] 2) The one-dimensional feature vector is input into the projection head and passes through the first fully connected layer and the second fully connected layer in sequence.
[0028] The first fully connected layer is used to increase the dimension to 512 and pass the ReLU activation function;
[0029] The second fully connected layer is used to reduce the dimension to 128 and perform L2 normalization to generate a unit embedding vector.
[0030] In step S2, the training data set is constructed by fusing multi-source heterogeneous data, including the following steps:
[0031] a) Using civil aircraft, car models, cones, plates, and drones as target equivalents to collect laser echo data from multiple angles and materials;
[0032] b) Use SolidWorks and Zemax simulation to generate laser echo images with different incident angles and reflectivities;
[0033] c) Mapping infrared / radar / visible light data into laser-like reflection features through a depth map generation algorithm, namely:
[0034]
[0035] in, is a mapping function based on the physical reflection model, and θ is the material reflectivity and incident angle parameters.
[0036] The step S4 comprises the following steps:
[0037] S4-1: Select a 3D size template of the corresponding category in the local model library based on the candidate box shape and target classification results. The local model library includes: differentiated 3D size templates, material reflection characteristics, and motion trajectory models for aircraft and drones;
[0038] S4-2: Use deep networks to estimate the spatial scale factor for geometric inversion, namely:
[0039]
[0040] in, is the 3D template corresponding to the target category, Φ() is the scale factor estimation network, and BayesUncertainty is the Bayesian posterior confidence;
[0041] S4-3: Fusion of Kalman predicted motion state to correct positioning drift.
[0042] In step S5, the adaptive Kalman filter performs the following steps:
[0043] S5-1: Encode the historical observation sequence through the lightweight Transformer and dynamically predict the state transfer matrix A and the observation matrix H, that is:
[0044]
[0045] Among them, y t-k is a historical observation sequence, and Transformer is a lightweight self-attention encoder;
[0046] S5-2: Embed the predicted shift matrix A and observation matrix H into the Kalman filter process and output the target velocity, acceleration and trajectory change.
[0047] The step S5-2 is specifically as follows:
[0048] (1) Perform state prediction based on the predicted shift matrix A:
[0049]
[0050] Among them, x t-1 is the state vector at the previous moment, including: position vector and velocity six-dimensional vector, A t is the state transition matrix dynamically predicted by the lightweight Transformer;
[0051] (4) Covariance prediction:
[0052]
[0053] Where Q is the preset process noise covariance matrix;
[0054] (3) According to the observation matrix H, the Kalman gain is calculated:
[0055]
[0056] Among them, H t is the observation matrix dynamically predicted by the lightweight Transformer, and R is the preset observation noise covariance matrix;
[0057] (5) Perform state update and covariance update, and output dynamic parameters, that is, output target velocity, acceleration and trajectory change.
[0058] The method further includes: deploying the algorithms of steps S1 to S5 to the fuze microprocessor through model pruning and quantization, specifically:
[0059] S6-1: Using depthwise separable convolution and attention-guided channel pruning to compress the number of parameters;
[0060] S6-2: Perform INT8 quantization on the entire network to fuse cross convolutional layers to reduce hardware load;
[0061] S6-3: Perform BatchNorm layer fusion to optimize computational efficiency.
[0062] The present invention has the following beneficial effects and advantages:
[0063] 1. The present invention improves detection accuracy: Through deep feature extraction and target classification, the present invention improves the target detection accuracy of the laser fuze system in complex environments, especially in multi-target, high-speed motion and strong interference environments, the system can still effectively identify and track targets.
[0064] 2. The present invention enhances robustness: The present invention adopts adaptive Kalman filtering and deep reinforcement learning algorithm, which can maintain the stability and reliability of the system in complex dynamic environments and interference conditions, and significantly reduce the probability of false triggering or missed triggering.
[0065] 3. Real-time response of the present invention: Based on deep learning and adaptive filtering algorithms, the present invention can quickly process laser reflection signals, ensure real-time requirements, and adapt to the detection needs of high-speed targets.
[0066] 4. The present invention has high deployment feasibility: the present invention undergoes model compression and quantization processing to ensure that the algorithm can run efficiently in resource-constrained hardware. BRIEF DESCRIPTION OF THE DRAWINGS
[0067] Figure 1 Schematic diagram of the convolutional neural network of the device of the present invention;
[0068] Figure 2 Flowchart of the method for improving the target detection accuracy of laser fuze according to the present invention. DETAILED DESCRIPTION
[0069] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.
[0070] like Figure 2 As shown in FIG, a flow chart of a method for improving the target detection accuracy of a laser fuze is provided. The method of the present invention for improving the target detection accuracy of a laser fuze includes the following steps:
[0071] Step 1: Laser signal preprocessing
[0072] First, the received laser reflection signal is preprocessed, mainly including denoising, normalization, and signal enhancement. A filter is used to remove noise components from the signal, and an adaptive algorithm is used to adjust the signal gain according to the real-time environment to ensure signal stability and quality in low signal-to-noise ratio environments.
[0073] Denoising: Use Wavelet Transform to denoise the signal. Wavelet Transform can effectively decompose the signal and remove noise in the frequency domain, improving the clarity and reliability of the signal.
[0074] Adaptive gain adjustment: According to the intensity of environmental noise, the gain of the laser signal is adjusted in real time to avoid the signal being too strong or too weak, ensuring the optimization of signal quality.
[0075] Step 2: Deep feature extraction and target detection
[0076] The present invention adopts convolutional neural network (such as Figure 1 As shown in the figure, it is used to extract deep feature information from the laser reflection signal and perform target detection. By training the convolutional neural network, the system can learn the characteristics of different targets and identify key features such as the target's shape and motion trajectory. In the laser fuze system, due to its closed application environment, difficulty in obtaining annotations, and high testing costs, the small amount of data and scarcity of labels have become a major bottleneck for training deep models. To this end, the present invention introduces a self-supervised learning mechanism in the laser fuze for the first time. Since the laser echo data of real combat targets obtained in live-fire tests are costly, time-consuming, and highly restricted, the present invention uses multi-source heterogeneous data fusion to generate training samples and construct a target echo signal data set that can be used for fuze training, including target shape, posture, material, reflection characteristics, etc.
[0077] Constructing a training dataset by fusing multi-source heterogeneous data includes the following steps:
[0078] a) Using civil aircraft, car models, cones, plates, and drones as target equivalents to collect laser echo data from multiple angles and materials;
[0079] b) Use SolidWorks and Zemax simulation to generate laser echo images with different incident angles and reflectivities;
[0080] c) Mapping infrared / radar / visible light data into laser-like reflection features through a depth map generation algorithm, namely:
[0081]
[0082] in, is a mapping function based on the physical reflection model, and θ is the material reflectivity and incident angle parameters.
[0083] The convolutional neural network is trained using the SimCLR self-supervised contrastive learning architecture to complete unsupervised representation learning of laser echo signals. Signal enhancement, rotation, occlusion, and other transformations are used as "positive and negative sample pairs" to generate multiple enhanced views of the same target, driving the network to learn highly discriminative feature representations. After self-supervised training, small-sample fine-tuning is performed, achieving good classification and detection performance with only a minimal amount of field-measured fuze data. This strategy significantly reduces reliance on labeled data, saves training costs, accelerates deployment, and meets the practical limitations of weapon systems in actual combat deployment environments.
[0084] The self-supervised learning mechanism uses the SimCLR contrastive learning framework, which includes:
[0085] S2-1: Apply random enhancement transformations to the laser echo image to generate positive sample pairs; the enhancement operations include: adding noise, brightness adjustment, local occlusion, affine transformation, geometric rotation, and frequency domain perturbation;
[0086] S2-2: Features are extracted through a lightweight CNN backbone network and converted into embedding vectors through the projection head;
[0087] 1) The preprocessed laser echo signal is input into the backbone network, passing through the input layer, multiple Bottleneck layers, and the output layer in sequence;
[0088] Input layer, including: 3×3 convolution kernel, batch normalization and h-swish activation function;
[0089] Bottleneck layer, used to extract multi-scale features using depth-wise separable convolution and SE attention module;
[0090] Output layer, used for global average pooling to generate a one-dimensional feature vector;
[0091] 2) The one-dimensional feature vector is input into the projection head and passes through the first fully connected layer and the second fully connected layer in sequence.
[0092] The first fully connected layer is used to increase the dimension to 512 and pass the ReLU activation function;
[0093] The second fully connected layer is used to reduce the dimension to 128 and perform L2 normalization to generate a unit embedding vector.
[0094] S2-3: The contrastive loss function is used to optimize the network to minimize the embedding distance of enhanced views of the same target, that is:
[0095]
[0096] Among them, z i ,z j The enhanced view embedding vector representing the same object, sim(z i ,z k ) is the cosine similarity, and τ represents the temperature hyperparameter.
[0097] Multi-scale feature extraction: Use the trained lightweight CNN network to extract multi-scale features from the preprocessed echo signal; use the Region Proposal Network (RPN) to generate candidate target boxes;
[0098] In the object detection module, traditional methods (such as sliding windows) are computationally intensive and inefficient when searching for potential target locations. To improve the speed and accuracy of generating candidate target boxes, this paper introduces a region proposal network (RPN) as an additional branch of the convolutional neural network. This network automatically and efficiently generates regions that may contain targets (i.e., candidate boxes). A softmax head is used to perform preliminary target category classification, and the output includes category confidence, location information, shape, and structural features for subsequent localization modules. This multi-scale processing can adapt to targets of varying sizes and speeds, improving target recognition accuracy.
[0099] Step 3: Target classification and precise positioning
[0100] Based on target detection and preliminary classification results, a corresponding model is selected from the local target model library. This model library is one of the biggest innovations of this invention compared to traditional laser ranging and positioning. Its core functions are: establishing feature priors for different target types to assist in resolving the ambiguity problem of laser signals; using the selected model to fit the current target reflection characteristics and optimize the error, combined with a depth fine-tuning mechanism to improve positioning accuracy; introducing Bayesian uncertainty estimation in the process of matching the echo signal with the trajectory model, and outputting the positioning confidence interval for reference by the trigger judgment module;
[0101] Precise positioning: This system combines the shape (height, width, and texture) of the candidate box with the 3D size template and reflection characteristics of the corresponding category in the local target model library. A deep network is used to estimate the spatial scale factor, perform geometric inference, and incorporate the current motion state predicted by the Kalman ensemble to correct for drift. This system no longer relies on single reflection intensity or time delay, but instead integrates modeling across multiple dimensions including category, shape, size, material, and trajectory to build a more robust positioning system.
[0102] In the above-mentioned precise positioning step, the local model library includes: differentiated 3D size templates, material reflection characteristics and motion trajectory models of aircraft and drones;
[0103] Use the deep network to estimate the spatial scale scaling factor for geometric inversion, namely:
[0104]
[0105] in, is the 3D template corresponding to the target category, Φ() is the scale factor estimation network, and BayesUncertainty is the Bayesian posterior confidence;
[0106] Step 4: Target tracking and motion state estimation
[0107] In actual combat, laser echoes often experience "short-term loss". The Neural-Kalman state estimation Kalman filter algorithm is used to estimate and predict the target's motion state. Based on target detection and classification, combined with the target's historical trajectory information, the target's speed, acceleration and other dynamic information are accurately estimated, which stabilizes target judgment, increases the probability of hitting, and reduces the risk of accidental bombing or accidental contact.
[0108] In particular, the present invention designs a lightweight Transformer structure to encode historical observation data sequences and predict the state transition matrix A and observation matrix H of the current frame; the predicted A and H are embedded in the traditional Kalman filter process to complete the adaptive tracking of target speed, acceleration, and trajectory changes. Compared with the traditional AKF, it has stronger ability to cope with nonlinear dynamics and is particularly suitable for aircraft changing trajectory or avoiding flying targets.
[0109] The adaptive Kalman filter performs the following steps:
[0110] S5-1: Encode the historical observation sequence through the lightweight Transformer and dynamically predict the state transfer matrix A and the observation matrix H, that is:
[0111]
[0112] Among them, y t-k is a historical observation sequence, and Transformer is a lightweight self-attention encoder;
[0113] S5-2: Embed the predicted shift matrix A and observation matrix H into the Kalman filter process and output the target velocity, acceleration and trajectory change.
[0114] (1) Perform state prediction based on the predicted shift matrix A:
[0115]
[0116] Among them, x t-1 is the state vector at the previous moment, including: position vector and velocity six-dimensional vector, A tis the state transition matrix dynamically predicted by the lightweight Transformer;
[0117] (2) Covariance prediction:
[0118]
[0119] Where Q is the preset process noise covariance matrix;
[0120] (3) According to the observation matrix H, the Kalman gain is calculated:
[0121]
[0122] Among them, H t is the observation matrix dynamically predicted by the lightweight Transformer, and R is the preset observation noise covariance matrix;
[0123] Perform state update and covariance update, and output dynamic parameters, that is, output target velocity, acceleration and trajectory change.
[0124] Step 5: Lightweight model deployment
[0125] The network parameters are greatly compressed by using techniques such as structural sparsification, depthwise separable convolution, and attention-guided channel pruning. All network structures are quantized to INT8, and cross-fusion convolution layers are used to reduce the hardware memory read and write load.
[0126] Example: For model training and feature learning:
[0127] like Figure 1 This example illustrates how the convolutional neural network model used for fuze target recognition in the present invention is trained through self-supervised learning, and how a deep feature extraction network with both generalization and discrimination capabilities is obtained in data-scarce scenarios. It should be noted that the following training process occurs before the actual deployment of the fuze system and falls within the model-building phase of the present invention. After training, the resulting model is solidified and deployed in a quantized form within the fuze system, and does not involve on-board real-time learning or self-update mechanisms.
[0128] 1. Training data construction method
[0129] Given the high cost of collecting laser reflection data from real targets in weapon systems, traditional training methods using large numbers of labeled samples are difficult to implement. Therefore, this paper uses the following multi-source heterogeneous data fusion method to construct a self-supervised training dataset:
[0130] Non-military object replacement modeling
[0131] Representative civilian aircraft, car models, aircraft models, cones, plates, drones, and other objects are used as structural equivalents of military targets. Using a track motion system, rotating hanger, and attitude change platform, laser reflection echo data is collected at different angles, materials, and distances, including time delay, intensity distribution, and signal-to-noise ratio changes.
[0132] Simulation data synthesis
[0133] Using the 3D modeling tool SolidWorks and optical simulation software Zemax, we constructed a target geometry model and a laser beam path simulation environment. We set different wavelengths, incident angles, and material reflectivity to generate large quantities of laser echo image data.
[0134] Equivalent Visual Data Mapping Method
[0135] The existing infrared / radar / visible light image data is converted into laser-like reflection feature maps through depth map generation algorithms, reflection model mapping and other methods, and input into the self-supervised model as enhanced data.
[0136] The final dataset size is controlled at 5,000 sets of target scene samples, all of which are unlabeled raw echo data.
[0137] 2. Self-supervised feature learning method
[0138] The present invention adopts the SimCLR structure as the self-supervised training framework, and the training process includes the following steps:
[0139] Positive and negative sample construction
[0140] Each laser echo image is subjected to two random enhancement transformations to generate a "positive sample pair". The enhancement operations include: adding noise, brightness adjustment, local occlusion, affine transformation, geometric rotation, frequency domain perturbation, etc.
[0141] The remaining samples are considered negative samples;
[0142] Neural network structure
[0143] The backbone network uses a lightweight convolutional neural network (MobileNetV3-Small), and the output is converted into an embedding vector through the projection head for comparative training;
[0144] Contrastive loss optimization
[0145] The contrastive loss function is used to optimize the network parameters so that the embedding vectors of enhanced views of the same object are closer and the embedding distances of different objects are farther, thereby achieving spontaneous learning of discriminative feature representations.
[0146] This training process can complete effective target feature learning without manual labeling of information.
[0147] 3. Small Sample Fine-tuning and Model Deployment
[0148] After the self-supervised pre-training is completed, the present invention further performs fine-tuning training on a small number of measured samples (e.g., 20-50 images of each type of target) collected by the fuze test platform. The method is as follows:
[0149] Freeze the feature extraction layer and only train the Softmax classification head to identify the target category;
[0150] If the computing power of the deployment platform allows, a layer-by-layer unfreezing strategy can be used to fine-tune the entire network;
[0151] The obtained model is further optimized through network structure pruning, channel sparsification, INT8 quantization, BatchNorm fusion, etc., and finally exported into a deployment format and deployed into the fuze system.
[0152] In summary, in combination with the embodiments, the present invention addresses the problem of decreased detection accuracy of laser fuzes in complex environments and seeks a method to improve the target detection accuracy of laser fuzes, which is of great significance in the field of laser fuzes.
[0153] Those skilled in the art will understand that the above description is only a preferred embodiment of the present invention, and the features described in the various embodiments and / or claims of the present disclosure may be combined or coupled in various ways, even if such a combination or coupling is not explicitly described in the present disclosure. It is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art may still modify the technical solutions described in the aforementioned embodiments or make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention shall be included in the scope of protection of the present invention.
[0154] Although preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they are aware of the basic inventive concepts. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention. Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, the present invention is intended to include such changes and modifications as fall within the scope of the claims and their equivalents.
Claims
1. A method for improving the target detection accuracy of a laser fuze, characterized in that: The following steps are involved: Step S1: obtaining a laser reflection signal through a laser transmitter and receiver module, and preprocessing the laser reflection signal to obtain a preprocessed laser reflection signal; Step S2: Train the convolutional neural network through a self-supervised learning mechanism, and construct a training data set by fusing multi-source heterogeneous data to complete pre-training; Step S3: Extract multi-scale features from the preprocessed signal through the trained convolutional neural network, generate candidate target boxes based on the region proposal network, and obtain the target classification result; Step S4: Based on the target classification result, the 3D size template and reflection characteristics of the corresponding category in the local target model library are called, and the precise position of the target is located by combining the Bayesian uncertainty estimation; Step S5: Adopting the adaptive Kalman filter algorithm regulated by the neural network to further optimize the positioning accuracy of the target and realize the target motion state tracking.
2. The method for improving target detection accuracy of a laser fuze according to claim 1, characterized in that: The pre-processing in step S1 includes: denoising, normalization and adaptive gain adjustment; Wherein, the denoising process uses wavelet transform to achieve frequency domain noise removal; The adaptive gain adjustment optimizes the signal gain in real time according to the ambient noise intensity.
3. The method for improving target detection accuracy of a laser fuze according to claim 1, characterized in that: The convolutional neural network uses MobileNetV3-Small as the backbone network. After self-supervised pre-training, it uses the measured data of the fuse to perform small sample fine-tuning. The small sample fine-tuning adopts freezing the feature layer to train the Softmax classification head, or unfreezing the layer by layer to perform fine-tuning of the entire network.
4. The method for improving target detection accuracy of a laser fuze according to claim 1, characterized in that: In step S2, the self-supervised learning mechanism adopts the SimCLR contrastive learning framework, which specifically includes: S2-1: Apply random enhancement transformations to the laser echo image to generate positive sample pairs; the enhancement operations include: adding noise, brightness adjustment, local occlusion, affine transformation, geometric rotation, and frequency domain perturbation; S2-2: Features are extracted through a lightweight CNN backbone network and converted into embedding vectors through the projection head; S2-3: The contrastive loss function is used to optimize the network to minimize the embedding distance of enhanced views of the same target, that is: Among them, z i ,z j The enhanced view embedding vector representing the same object, sim(z i ,z k ) is the cosine similarity, and τ represents the temperature hyperparameter.
5. The method for improving target detection accuracy of a laser fuze according to claim 4, characterized in that: The step S2-2 comprises the following steps: 1) The preprocessed laser echo signal is input into the backbone network and passes through the input layer, multiple Bottleneck layers, and output layer in sequence; Input layer, including: 3×3 convolution kernel, batch normalization and h-swish activation function; Bottleneck layer, used to extract multi-scale features using depth-wise separable convolution and SE attention module; Output layer, used for global average pooling to generate a one-dimensional feature vector; 2) The one-dimensional feature vector is input into the projection head and passes through the first fully connected layer and the second fully connected layer in sequence. The first fully connected layer is used to increase the dimension to 512 and pass the ReLU activation function; The second fully connected layer is used to reduce the dimension to 128 and perform L2 normalization to generate a unit embedding vector.
6. The method for improving target detection accuracy of a laser fuze according to claim 1, characterized in that: In step S2, the training data set is constructed by fusing multi-source heterogeneous data, including the following steps: a) Using civil aircraft, car models, cones, plates, and drones as target equivalents to collect laser echo data from multiple angles and materials; b) Use SolidWorks and Zemax simulation to generate laser echo images with different incident angles and reflectivities; c) Mapping infrared / radar / visible light data into laser-like reflection features through a depth map generation algorithm, namely: in, is a mapping function based on the physical reflection model, and θ is the material reflectivity and incident angle parameters.
7. The method for improving target detection accuracy of a laser fuze according to claim 1, characterized in that: The step S4 comprises the following steps: S4-1: Based on the candidate box shape and target classification results, a 3D size template of the corresponding category in the local model library is selected. The local model library includes: differentiated 3D size templates, material reflection characteristics, and motion trajectory models for aircraft and drones; S4-2: Use deep networks to estimate the spatial scale factor for geometric inversion, namely: in, is the 3D template corresponding to the target category, Φ() is the scale factor estimation network, and BayesUncertainty is the Bayesian posterior confidence; S4-3: Fusion of Kalman predicted motion state to correct positioning drift.
8. The method for improving target detection accuracy of a laser fuze according to claim 1, characterized in that: In step S5, the adaptive Kalman filter performs the following steps: S5-1: Encode the historical observation sequence through the lightweight Transformer and dynamically predict the state transfer matrix A and the observation matrix H, that is: Among them, y t-k is a historical observation sequence, and Transformer is a lightweight self-attention encoder; S5-2: Embed the predicted shift matrix A and observation matrix H into the Kalman filter process and output the target velocity, acceleration and trajectory change.
9. The method for improving target detection accuracy of a laser fuze according to claim 8, characterized in that: The step S5-2 is specifically as follows: (1) Perform state prediction based on the predicted shift matrix A: Among them, X t-1 is the state vector at the previous moment, including: position vector and velocity six-dimensional vector, A t is the state transition matrix dynamically predicted by the lightweight Transformer; (2) Covariance prediction: Where Q is the preset process noise covariance matrix; (3) According to the observation matrix H, the Kalman gain is calculated: Among them, H t is the observation matrix dynamically predicted by the lightweight Transformer, and R is the preset observation noise covariance matrix; (3) Perform state update and covariance update, and output dynamic parameters, that is, output target velocity, acceleration and trajectory change.
10. The method for improving target detection accuracy of a laser fuze according to claim 1, characterized in that: Also includes: The algorithm of steps S1 to S5 is deployed to the fuze microprocessor through model pruning and quantization, specifically: S6-1: Using depthwise separable convolution and attention-guided channel pruning to compress the number of parameters; S6-2: Perform INT8 quantization on the entire network to fuse cross convolutional layers to reduce hardware load; S6-3: Perform BatchNorm layer fusion to optimize computational efficiency.