Method for realizing target online detection by using lightweight gabp network
By employing dual-mode laser and millimeter-wave detection and a lightweight GABP network, the limitations of fuze hardware were overcome, enabling rapid and accurate target identification and processing in complex battlefield environments, thus improving identification accuracy and adaptability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NANJING UNIV OF SCI & TECH
- Filing Date
- 2023-08-29
- Publication Date
- 2026-05-12
AI Technical Summary
Existing technologies have limited hardware capabilities for fuses in complex battlefield environments, making it difficult to achieve rapid and accurate target identification and processing, and they are not adaptable enough to complex environmental interference.
A lightweight GABP network is designed by combining laser and millimeter-wave dual-mode detection with a GA genetic algorithm to optimize the BP network, sparse the weight matrix, and achieve online target detection.
To improve the accuracy and reliability of target identification in complex battlefield environments, reduce computational load and storage space requirements, and adapt to the accurate identification of different targets.
Smart Images
Figure CN117111077B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of signal detection, specifically relating to a method for online target detection using a lightweight GABP network, which is particularly suitable for laser millimeter-wave dual-mode detection radar. Background Technology
[0002] Since the continuous development of networked and intelligent technologies, researchers have been proposing various mathematical algorithms combined with artificial intelligence to solve the problems of massive amounts of data and slow information processing in different fields. Neural network models, which omit complex manual feature extraction steps, possess good generalization ability, superior flexibility, universality, and high accuracy, thus showing great potential in the field of target recognition. Especially for dual-mode laser and millimeter-wave detection, neural network models can compensate for the shortcomings of traditional single detection methods. They can change detection modes or recombine composite modes according to different detection targets or different external environments, adapting to the constantly changing battlefield environment and target characteristics, thereby improving the accuracy of weapon target recognition.
[0003] In special military battlefield environments, the fuze, as the core of damage control in the terminal countermeasures of weapon systems, needs to accurately detect target-related information and rapidly process it to obtain target characteristics and target-missile intersection parameters, thereby achieving optimal fuze-warfare coordination control and efficient damage. Applying neural networks to fuze target recognition in land battlefields significantly improves accuracy compared to traditional methods, addressing the need for precise target identification. However, this requires a large amount of collected data, and the limited capabilities of fuze detectors and processors to support such large datasets are a key issue. Furthermore, in actual battlefield situations, complex environments exist, including electromagnetic fields, clouds, smoke, lightning, rain, snow, sea surface moisture, and clutter, presenting numerous unknown interfering targets. Improving the accuracy and speed of battlefield target classification and identification is also a significant challenge.
[0004] To address the aforementioned two issues and achieve accurate and rapid identification of military targets across a wide-area battlefield encompassing land, sea, air, and space, researchers are currently focusing on designing different neural network model structures, increasing network depth, and expanding network width. To address the problem of convolutional neural networks requiring a large number of training samples, Wu et al. proposed an end-to-end deep learning framework called Multi-View Prototype Network (MVPN). MVPN can learn features jointly from multiple views of a 3D shape and the prototype representation of each class. It can then perform identification by querying the closest class prototype in the embedding space. Furthermore, it improves the loss function to reduce environmental interference (Wu, Zizhao, PingYang, and Yigang Wang. "MVPN: Multi-View Prototype Network for 3D ShapeRecognition." IEEE Access 7 (2019): 130363-130372.). Zhao et al. proposed an airborne lidar point cloud classification method based on transfer learning. By using a feature image generation strategy based on the spatial distribution of point clouds, they extracted multi-scale and multi-view depth features through transfer learning. A two-layer CNN deep learning network was designed to reduce dimensionality, fuse and learn high-level features, ultimately achieving good classification accuracy with shorter training time and fewer training samples (Zhao, Chuan, et al. "ALS point cloud classification with small training data set based on transfer learning." IEEE Geoscience and Remote Sensing Letters 17.8(2019): 1406-1410.). Liu et al. studied the impact of cloud and fog interference at different visibility levels on the laser fuze of a certain type of missile. Low visibility due to clouds and fog led to false alarms in the laser fuze due to the influence of suspended particles in the air. They obtained the false alarm conditions for the laser fuze, confirming that cloud and fog environmental conditions need to be considered during the use of laser fuze weapons. (Liu Yun, Peng Xingge, Zhang Jun, Jiang Fei. Research on anti-cloud and fog interference of laser fuze of a certain type of missile [J]. Science and Technology Innovation and Application, 2022, 12(03):18-20.)
[0005] Current research has made some progress in algorithm research in terms of training time, computational load and accuracy, but all of them are carried out on powerful GPU platforms, which cannot meet the detection requirements of fuse hardware with low integration, low module functional complexity and low single detection accuracy in actual battlefields. At the same time, the impact of various complex environments on the battlefield is considered in a relatively simplistic way. Summary of the Invention
[0006] The purpose of this invention is to provide a method for online target detection using a lightweight GABP network, which is suitable for dual-mode laser and millimeter-wave detection. It can efficiently and accurately identify targets and can also be applied to weak hardware fuses.
[0007] The technical solution to achieve the purpose of this invention is as follows: a method for online target detection using a lightweight GABP network, comprising the following steps:
[0008] Step 1: Use a dual-mode laser and millimeter-wave sampling circuit to acquire laser echo signals and millimeter-wave echo signals from different targets. The laser echo signals and millimeter-wave echo signals from different targets are processed by the receiving channel to obtain laser beat signals and millimeter-wave beat signals, respectively. After extraction by windowed Fourier transform and sampling by ADC, digital laser beat signals and digital millimeter-wave beat signals of different targets are obtained. The digital beat signals are then processed in the FPGA. The digital laser beat signals and digital millimeter-wave beat signals are fused using a weighted average method to obtain the spatial distribution point sequence and the distribution sequence of each frequency band of the corresponding targets. Proceed to Step 2.
[0009] Step 2: Build a lightweight GABP neural network:
[0010] The number of neurons in the BP neural network is designed, and the threshold, weights and biases of the BP network are optimized and updated through the GA genetic algorithm. The weight and bias matrix in the BP neural network is sparsed to achieve the lightweighting of the BP network, and thus a lightweight GABP neural network is established.
[0011] Proceed to step 3.
[0012] Step 3: Train the lightweight GABP neural network using the spatial distribution point sequence and the distribution sequence of each frequency band of the target to obtain the lightweight GABP neural network model:
[0013] The spatially distributed point sequence and the distribution sequence of each frequency band are input into a lightweight GABP neural network, and the threshold and weights of the network are continuously updated to obtain a lightweight GABP neural network model for recognizing and detecting different targets.
[0014] Proceed to step 4.
[0015] Step 4: A lightweight GABP neural network model is used to detect the laser echo signal and millimeter-wave echo signal acquired by the laser millimeter-wave dual-mode sampling circuit in real time to identify the target type. Simultaneously, to more objectively evaluate the prediction performance of the GA-BP neural network model, the mean square error function is used to assess the overall prediction performance of the model.
[0016] Compared with the prior art, the significant advantages of this invention are:
[0017] (1) This paper adopts laser and millimeter wave dual-mode detection sampling, which can make full use of the advantages of the two detection systems, complement each other, realize the detection of multiple physical characteristics of the target in complex battlefield environment, effectively improve the detection reliability and accuracy of single-system fuse, and realize the precision strike of weapon system.
[0018] (2) This invention uses a GABP network to achieve target detection. The GA genetic algorithm is used to optimize and update the weights, biases and thresholds of the BP network. The GABP network is designed as a target classifier, which has strong nonlinear mapping ability and adaptive self-learning ability, and can achieve accurate identification of different targets.
[0019] (3) The present invention adopts lightweight operation of GABP network, and realizes the partial parameter merging and deletion operation of GABP network weight bias matrix by compressing sparse rows, which can reduce the amount of computation and the storage space of target feature library, and realize online recognition of GABP network. Attached Figure Description
[0020] Figure 1 This is a flowchart of the method for online target detection using a lightweight GABP network as described in this invention.
[0021] Figure 2 This is a block diagram illustrating the design of the method for online target detection using a lightweight GABP network as described in this invention.
[0022] Figure 3 This is a system flowchart of the GABP neural network of the present invention. Detailed Implementation
[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.
[0024] The following section will further introduce the specific implementation method, as well as the technical difficulties and inventive points of this invention, using this design example as an example.
[0025] Combination Figures 1-3 A method for online target detection using a lightweight GABP network, comprising the following steps:
[0026] Step 1: A dual-mode laser and millimeter-wave sampling circuit is used to acquire laser echo signals and millimeter-wave echo signals from different targets. These signals are processed by a receiving channel to obtain laser beat signals and millimeter-wave beat signals, respectively. After extraction via windowed Fourier transform and ADC sampling, digitized laser beat signals and digitized millimeter-wave beat signals for different targets are obtained. These digitized beat signals are then processed within the FPGA. A weighted average method is used to fuse the digitized laser beat signals and digitized millimeter-wave beat signals to obtain the spatial distribution point sequence and frequency band distribution sequence of the corresponding targets, as detailed below:
[0027] A weighted average method is used to fuse digital laser beat signals and digital millimeter-wave beat signals to obtain the point sequence of spatial distribution of the corresponding target and the distribution sequence of each frequency band.
[0028] The key lies in using a dual-mode laser and millimeter-wave sampling circuit to acquire laser echo signals and millimeter-wave echo signals from different targets. These signals are then further processed by a multi-channel ADC acquisition unit to obtain digital laser beat signals and digital millimeter-wave beat signals for each target. A weighted average is then applied, multiplying the laser beat signals and digital millimeter-wave beat signals by their respective weights and summing the results. Finally, the sum is divided by the total weight of 1 to obtain the fused laser and millimeter-wave signal. ,Right now:
[0029] (1)
[0030] in Indicates laser beat signal, Indicates millimeter-wave beat signal, and All represent the fusion weights. and The values are 0.5 respectively.
[0031] Proceed to step 2.
[0032] Step 2: Establish a lightweight GABP neural network: Design the number of nodes for each neuron in the BP neural network. Optimize and update the threshold, weights, and biases of the BP network using the GA genetic algorithm. Perform sparse matrix operations on the weight and bias matrices in the BP neural network to achieve lightweighting, thus establishing a lightweight GABP neural network. Figure 3 The details are as follows:
[0033] S21: Design the BP network topology.
[0034] Designing a backpropagation (BP) network requires determining the number of neurons in the input, output, and hidden layers. The function expression is as follows:
[0035] (2)
[0036] In the formula m and N represent the number of neurons in the input layer, the output layer, and the hidden layer, respectively. For constant terms, The sigmoid function is used as the activation function. The initial weight threshold is a random constant between [-1, 1], and a complete BP network topology is established.
[0037] S22: Optimize and update the weights, thresholds, and biases of the BP network using the selection, crossover, and mutation operations of the GA genetic algorithm, and establish a weight-bias matrix using the weights, thresholds, and biases:
[0038] The expression for the mutation function is as follows:
[0039] (3)
[0040] (4)
[0041] In the formula, These represent the row and column numbers of the current individual, respectively. For individuals The upper bound maximum value, For individuals The lower bound minimum value, For each individual's fitness, It is a random number. This represents the current evolution count. To the maximum number of evolutions that can be achieved, It is the range of values. The random constant.
[0042] S23: Perform sparsification on the obtained weight and bias matrices to achieve lightweighting of the BP network. This is achieved by compressing sparse rows to perform partial parameter merging and deletion operations on the BP network weight and bias matrices.
[0043] First, the weights and bias matrices are determined as the current index values. Then, the weights and bias matrices are iterated to determine if any target feature data with the same index value exists in the original target feature library. If not, the weights and bias matrices are de-zeroed and sparsified by introducing a 0-norm constraint and a row index slicing function to obtain a sparsified weight matrix. sparse bias matrix At the same time, it is added to the target feature library; if it exists, then S3 is executed.
[0044] The expression that should be satisfied for de-zeroing sparsity is as follows:
[0045] (5)
[0046] (6)
[0047] In the formula, , , These are the weights, biases, and bias matrices. , , The number of non-zero elements in the neutron. To limit the maximum number of non-zero elements in a matrix; For sparse weight matrix, For sparse bias matrix, For the sparsified threshold matrix, This is a function for slicing the weight matrix by row index. This is a function for slicing rows by index of a bias matrix. This is a function for slicing rows in the threshold matrix.
[0048] Step 3: Train the lightweight GABP neural network using the spatial distribution point sequence and the distribution sequence of each frequency band of the target to obtain the lightweight GABP neural network model, as follows:
[0049] S31: The lightweight GABP neural network is trained using the spatial distribution point sequence of the target and the distribution sequence of each frequency band to obtain the lightweight GABP neural network model.
[0050] The spatial distribution sequence of different targets and the distribution sequence of each frequency band are input into the lightweight GABP neural network for multiple iterations of training. The number of training times for each target is 50 to 300. The threshold and weights of the lightweight GABP neural network are continuously updated until the GABP neural network converges, and the lightweight GABP neural network model is obtained.
[0051] Proceed to step 4.
[0052] Step 4: A lightweight GABP neural network model is used to detect the laser echo signal and millimeter-wave echo signal acquired by the laser millimeter-wave dual-mode sampling circuit in real time to identify the target type. Simultaneously, to more objectively evaluate the prediction performance of the GA-BP neural network model, the mean square error function is used to evaluate the overall prediction performance of the model, as detailed below:
[0053] S41: Use a trained lightweight GABP network model to detect dual-mode signals and identify target types.
[0054] When training a lightweight GABP neural network, the learning rate is... The training value is 0.01, the number of training steps is 50, and the training precision is 0.00001.
[0055] The error function is the mean squared error function, which descends along the negative gradient direction. After applying the GA genetic algorithm, the minimum value of the mean squared error is defined as the fitness function value. A smaller fitness function value indicates more accurate training. The BP network is updated and optimized using the GA genetic algorithm until the network output error accuracy reaches the target accuracy requirement, at which point the fitness function value converges to its minimum, and the learning process ends.
[0056] (7)
[0057] In the formula, , These are samples from the training set and the test set, respectively. Solve for the mean square error function.
[0058] Example
[0059] The present invention describes a method for online target detection using a lightweight GABP network, based on... Figure 1 Flowchart of a method for online target detection using lightweight GABP networks Figure 2 The implementation scheme of the lightweight GABP network for online target detection is illustrated in the diagram. The laser echo analog signal and millimeter-wave echo signal are subjected to spectral analysis and fusion processing. The processed echo signal sequence is used as model input. Based on the model of this invention, accurate identification of different point cloud targets can be achieved. Subsequently, a complete target feature library is established and installed on the designed embedded platform, which is then ported to a weak hardware fuze, providing methodological support for rapid and efficient target identification in battlefield environments.
Claims
1. A method for online target detection using a lightweight GABP neural network, characterized in that, The steps are as follows: Step 1: Use a dual-mode laser and millimeter-wave sampling circuit to acquire laser echo signals and millimeter-wave echo signals from different targets. The laser echo signals and millimeter-wave echo signals from different targets are processed by the receiving channel to obtain laser beat signals and millimeter-wave beat signals, respectively. After extraction by windowed Fourier transform and sampling by ADC, digital laser beat signals and digital millimeter-wave beat signals of different targets are obtained. The above digital laser beat signals and digital millimeter-wave beat signals are processed in the FPGA. The weighted average method is used to fuse the digital laser beat signals and digital millimeter-wave beat signals to obtain the spatial distribution point sequence and the distribution sequence of each frequency band of the corresponding target. Proceed to Step 2. Step 2: Build a lightweight GABP neural network: The number of nodes in each neuron of the BP neural network is designed. The threshold, weights and biases of the BP neural network are optimized and updated through the GA genetic algorithm. The weight and bias matrix in the BP neural network is sparsed to achieve the lightweighting of the BP neural network, and then a lightweight GABP neural network is established. Proceed to step 3; Step 3: Train the lightweight GABP neural network using the spatial distribution point sequence and the distribution sequence of each frequency band of the target to obtain the lightweight GABP neural network model: The spatially distributed point sequence and the distribution sequence of each frequency band are input into a lightweight GABP neural network, and the threshold and weights of the network are continuously updated to obtain a lightweight GABP neural network model for recognizing and detecting different targets. Proceed to step 4; Step 4: Use a lightweight GABP neural network model to detect the laser echo signal and millimeter-wave echo signal acquired by the laser millimeter-wave dual-mode sampling circuit in real time to identify the type of target; at the same time, in order to more objectively measure the prediction effect of the GABP neural network model, the mean square error function index is used to evaluate the overall prediction effect of the model.
2. The method for online target detection using a lightweight GABP neural network according to claim 1, characterized in that, In step 1, a weighted average method is used to fuse the digital laser beat signal and the digital millimeter-wave beat signal, as follows: Laser echo and millimeter-wave echo data are processed by a multi-channel ADC acquisition unit to obtain digital laser beat signals and digital millimeter-wave beat signals for different targets. A weighted average is then applied, multiplying the laser beat signal and the digital millimeter-wave beat signal by their respective weights and summing the results. Finally, the sum is divided by the total weight of 1 to obtain the fused laser-millimeter-wave signal. ,Right now: (1)， in Indicates laser beat signal, Indicates millimeter-wave beat signal, and All represent the fusion weights. and The values are 0.5 respectively.
3. The method for online target detection using a lightweight GABP neural network according to claim 1, characterized in that, Step 2 involves building a lightweight GABP neural network, as detailed below: S21: Design the topology of the BP neural network: The number of neurons in the input layer, output layer, and hidden layer of the BP neural network are determined separately, and the function expressions are as follows: (2)， In the formula m and N represent the number of neurons in the input layer, the output layer, and the hidden layer, respectively. For constant terms, The sigmoid function is used as the activation function. The initial weight threshold is a random constant between [-1, 1], and a complete BP neural network topology is established. S22: Optimize and update the weights, thresholds, and biases of the BP neural network using the selection, crossover, and mutation operations of the GA genetic algorithm, and establish a weight bias matrix using the weights, thresholds, and biases; The expression for the mutation function is as follows: (3)， (4)， In the formula, These represent the row and column numbers of the current individual, respectively. For individuals The upper bound maximum value, For each individual's fitness, It is a random number. This represents the current evolution count. To the maximum number of evolutions that can be achieved, It is the range of values. The random constant; S23: Based on the weight bias matrix obtained in step S22, perform a sparsification matrix operation to achieve lightweighting of the BP neural network; The sparsification matrix operation includes merging and deleting some parameters of the BP neural network weight bias matrix by compressing sparse rows: First, the weight bias matrix is determined as the current index value. Then, the weight bias matrix is traversed to determine whether there is target feature data in the original target feature library that is the same as the current index value. If it does not exist, the weight bias matrix is reduced to zero and sparsified by introducing a 0-norm constraint and a row index slicing function to obtain the sparsified weight matrix. sparse bias matrix At the same time, add it to the target feature library; if it exists, proceed to step 3.
4. The method for online target detection using a lightweight GABP neural network according to claim 3, characterized in that, The expression that should be satisfied for de-zeroing sparsity is as follows: (5)， (6)， In the formula, , , These are the weight matrices Bias matrix Threshold matrix The number of non-zero elements in the neutron. To limit the maximum number of non-zero elements in a matrix; For sparse weight matrix, For sparse bias matrix, For the sparsified threshold matrix, This is a function for slicing the weight matrix by row index. This is a function for slicing rows by index of a bias matrix. This is a function for slicing rows in the threshold matrix.
5. The method for online target detection using a lightweight GABP neural network according to claim 1, characterized in that, In step 3, the lightweight GABP neural network is trained using the spatial distribution point sequence and the distribution sequence of each frequency band of the target, resulting in a lightweight GABP neural network model, as follows: The spatial distribution sequence of different targets and the distribution sequence of each frequency band are input into the lightweight GABP neural network for multiple iterations of training. The number of training times for each target is 50 to 300. The threshold and weights of the lightweight GABP neural network are continuously updated until the GABP neural network converges, and the lightweight GABP neural network model is obtained.
6. The method for online target detection using a lightweight GABP neural network according to claim 5, characterized in that, When training a lightweight GABP neural network, the learning rate is... The training value is 0.01, the number of training steps is 50, and the training precision is 0.00001.
7. The method for online target detection using a lightweight GABP neural network according to claim 1, characterized in that, In step 4, the error function is the mean squared error function, which descends along the negative gradient direction. After applying the GA genetic algorithm, the minimum value of the mean squared error is defined as the fitness function value. A smaller fitness function value indicates more accurate training. The BP neural network is updated and optimized using a GA genetic algorithm until the network output error accuracy reaches the target accuracy requirement, at which point the fitness function value converges to its minimum, and the learning process ends. (7)， In the formula, , These are samples from the training set and the test set, respectively. Let be the mean square error function.