CN-117111077-B - Method for realizing on-line detection of target by using lightweight GABP network

CN117111077BCN 117111077 BCN117111077 BCN 117111077BCN-117111077-B

Abstract

The invention discloses a method for realizing on-line detection of a target by a lightweight GABP network suitable for a laser millimeter wave dual-mode detection radar, which comprises the steps of converting an input laser echo analog signal and a millimeter wave echo signal into a point sequence with spatial distribution by an AD high-speed sampling technology, carrying out spectrum analysis, generating a distribution sequence of each frequency band, and jointly forming an algorithm input characteristic sequence by the two sequences; and obtaining a lightweight GABP neural network model through offline training and updating, and recording weights and thresholds of different targets in a full-standby target feature library to realize detection and identification of different target types. The invention realizes the real-time detection of the on-line targets, enhances the capability of resisting the interferences of smoke dust, sweep frequency interference, foil strips and the like by adopting the laser/millimeter wave dual-mode detection, establishes a full-standby target feature library of various environmental interferences, and can be used for quickly and accurately identifying the targets under the high dynamic state of the fuze.

Inventors

LI HAOJIE
Hou Zhaoqi
ZHA BINGTING
ZHENG ZHEN
WANG CHENGJUN

Assignees

南京理工大学

Dates

Publication Date: 20260512
Application Date: 20230829

Claims (7)

1. A method for realizing target online detection by using a lightweight GABP neural network is characterized by comprising the following steps: Step 1, acquiring laser echo signals and millimeter wave echo signals of different targets by utilizing a laser millimeter wave dual-mode sampling circuit, respectively processing the laser echo signals and the millimeter wave echo signals of the different targets through a receiving channel to obtain laser beating signals and millimeter wave beating signals, extracting and ADC (analog to digital) sampling the laser beating signals and the millimeter wave beating signals of the different targets through windowed Fourier transform, processing the digital laser beating signals and the digital millimeter wave beating signals in an FPGA (field programmable gate array), and performing data fusion on the digital laser beating signals and the digital millimeter wave beating signals by adopting a weighted average method to obtain a point sequence of spatial distribution of the corresponding targets and a distribution sequence of each frequency band, and transferring to step 2; Step2, establishing a lightweight GABP neural network: designing the number of each neuron node of the BP neural network, optimizing and updating the threshold value, the weight and the bias of the BP neural network through a GA genetic algorithm, and performing sparse matrix operation on a weight bias matrix in the BP neural network to realize the light weight of the BP neural network so as to establish a lightweight GABP neural network; Turning to step 3; Training the lightweight GABP neural network by utilizing a point sequence of spatial distribution of a target and a distribution sequence of each frequency band to obtain a lightweight GABP neural network model: Inputting the spatially distributed point sequence and the distributed sequence of each frequency band into a lightweight GABP neural network, and continuously updating the threshold value and the weight value of the network to obtain a lightweight GABP neural network model for identifying and detecting different targets; Turning to step 4; And 4, detecting the laser echo signals and the millimeter wave echo signals acquired by the laser millimeter wave dual-mode sampling circuit by utilizing the lightweight GABP neural network model in real time to identify the type of the target, and evaluating the overall prediction effect of the model by adopting a mean square error function index for objectively measuring the prediction effect of the GABP neural network model.
2. The method for realizing the on-line detection of the target by using the lightweight GABP neural network according to claim 1, wherein in the step 1, a weighted average method is adopted to perform data fusion on the digitized laser beat signal and the digitized millimeter wave beat signal, and the method is specifically as follows: the laser echo signal and millimeter wave echo signal data are processed through a multichannel ADC acquisition unit to obtain digital laser beating signals and digital millimeter wave beating signals of different targets, the laser beating signals and the digital millimeter wave beating signals are multiplied by weights to be added by weighted average, and finally the sum of the weights is divided by the total weight 1 to obtain a laser millimeter wave fusion signal The method comprises the following steps: (1), Wherein the method comprises the steps of Representing the laser beat signal, Representing the millimeter wave beat signal, And All of which represent the fusion weights and, And The values were 0.5 respectively.
3. The method for realizing on-line target detection by using a lightweight GABP neural network according to claim 1, wherein the lightweight GABP neural network is established in step 2, specifically comprising the following steps: s21, designing a BP neural network topological structure: the neuron numbers of an input layer, an output layer and a hidden layer of the BP neural network are respectively determined, and the function expression is as follows: (2), In the middle of M and N are respectively corresponding to the number of neurons of an input layer, the number of neurons of an output layer and the number of neurons of a hidden layer, Is a constant term which is used to determine the degree of freedom, Using sigmoid function as excitation function The initial weight threshold is a random constant between [ -1,1], and a complete BP neural network topological structure is established; S22, optimizing and updating weight, threshold and bias of the BP neural network by using selection, crossing and mutation operations of a GA genetic algorithm, and constructing write value bias matrixes by using the weight, the threshold and the bias; The variant function expression is as follows: (3), (4), In the formula, Representing the number of rows and columns of the current individual, For individuals Is set to be at the upper bound maximum of (c), For the fitness of each individual, In the form of a random number, For the current number of evolutions, In order to achieve the maximum number of evolutions that can be achieved, Is a value range Random constants of (a); s23, on the basis of the weight bias matrix obtained in the step S22, performing a sparse matrix operation to realize the light weight of the BP neural network, wherein the sparse matrix operation comprises partial parameter merging and deleting operation of the weight bias matrix of the BP neural network by compressing sparse rows: firstly, determining a weight bias matrix as a current index value, secondly, traversing the weight bias matrix, judging whether target feature data which is the same as the current index value exists in an original target feature library, if not, carrying out zero removal sparseness on the weight bias matrix, introducing 0 norm constraint and a line index slicing function to obtain a sparsified weight matrix Sparse bias matrix And (3) if the target feature library exists, executing the step (3).
4. The method for realizing on-line target detection by using a lightweight GABP neural network according to claim 3, wherein the expression satisfied by the nulling sparseness is as follows: (5), (6), In the formula, 、、 Respectively, weight matrix Bias matrix Threshold matrix The number of non-zero elements in the matrix, To limit the maximum number of non-zero elements in the matrix; to sparse weight matrix, To sparse bias matrix, In order to sparsify the matrix of threshold values, The slicing functions are indexed for the weight matrix rows, To index the slicing functions for the offset matrix rows, The slicing function is indexed for the threshold matrix row.
5. The method for realizing the on-line detection of the target by using the lightweight GABP neural network according to claim 1, wherein in the step 3, the lightweight GABP neural network is trained by using a point sequence of spatial distribution of the target and a distribution sequence of each frequency band, so as to obtain a lightweight GABP neural network model, which is specifically as follows: And (3) inputting the spatially distributed point sequences of different targets and the distributed sequences of each frequency band into the lightweight GABP neural network for multiple iterative training, wherein the training times of each target are 50-300, and continuously updating the threshold value and the weight value of the lightweight GABP neural network until the GABP neural network converges to obtain a lightweight GABP neural network model.
6. The method for achieving online detection of a target using a lightweight GABP neural network of claim 5, wherein the learning rate is a rate of training the lightweight GABP neural network The training step number is 50, and the training precision is 0.00001.
7. The method for realizing on-line target detection by using a lightweight GABP neural network as set forth in claim 1, wherein in step 4, the error function is a mean square error function, which decreases in the negative gradient direction, and after the GA genetic algorithm is used, the minimum value for solving the mean square error is defined as the fitness function value The smaller the fitness function value is, the more accurate the training is, the more the GA genetic algorithm is used for updating and optimizing the BP neural network until the accuracy of the network output error reaches the target accuracy requirement, the fitness function value is converged to the minimum value, and the learning is finished: (7), In the formula, 、 The samples are respectively a training set and a testing set; Is a mean square error function.

Description

Method for realizing on-line detection of target by using lightweight GABP network Technical Field The invention belongs to the field of signal detection, and particularly relates to a method for realizing on-line detection of a target by using a lightweight GABP network, which is particularly suitable for a laser millimeter wave dual-mode detection radar. Background Since the continuous development of networking intelligent technology, researchers continuously propose different mathematical algorithms and artificial intelligence to solve the problems of large data volume and slow information processing in different fields. The method omits the complicated manual design feature extraction step based on the neural network model, has better generalization capability, superior flexibility, universality and high accuracy, and has great potential in the field of target identification. The neural network model can make up for the defects of the traditional single detection mode, can change the detection mode or recombine the composite mode according to different detection targets or different external environments, is suitable for continuous changes of battlefield environments and target characteristics, and improves the accuracy rate of weapon to target identification. Aiming at the situation of special military battlefield, the fuze is used as a damage control core in the terminal countermeasure of a weapon system, the related information of a target needs to be accurately detected and rapidly processed to obtain the target characteristic and the bullet-target intersection parameter, and then optimal fight cooperation control is carried out to realize efficient damage. The neural network is used for target recognition of the fuze in the land battlefield, compared with the traditional method, the accuracy is greatly improved, the requirement of the fuze on accurate recognition of a specific target can be met, but the requirement can be met only by acquiring enough data, and the key problem is that the fuze detector, the processor and other hardware capabilities are limited and cannot support large data. Meanwhile, in the actual battlefield situation, complex environments such as electromagnetism, cloud fog, smoke dust, thunder, rain and snow, sea water vapor, clutter and the like exist, a plurality of unknown interference targets exist, and the problem of improving the classification recognition precision and speed of the battlefield targets is also a great problem. In order to solve the two problems and realize accurate and rapid identification of military targets in a wide-area battlefield of the sea, the land, the air and the space, currently, scholars mainly conduct researches on designing different neural network model structures, deepening network depth, increasing network width and the like. Wu and the like aim at the problem that a large number of training samples are needed by a convolutional neural network, an end-to-end deep learning framework of a Multi-view prototype network (Multi-View Prototype Network, MVPN) is provided, the MVPN can represent common learning characteristics from multiple views of a 3D shape and prototypes of each class, then the closest class prototypes can be found in an embedded space to inquire for identification, meanwhile, an airborne laser radar point cloud classification method based on transfer learning is provided for reducing interference improvement loss functions (Wu, Zizhao, Ping Yang, and Yigang Wang. "MVPN: Multi-View Prototype Network for 3D Shape Recognition." IEEE Access 7 (2019): 130363-130372.).Zhao and the like of environments and the like, characteristic image generation strategies of point cloud space distribution are utilized, depth characteristics of multiple scales and multiple views are extracted by utilizing transfer learning, two layers of CNN deep learning networks are designed for reducing dimensionality, fusion and learning of advanced characteristics are achieved, finally, the influence of different visibility cloud mist interference on laser fuses of a certain model is studied under the condition of less training samples, the cloud mist effect can cause the laser fuse virtual fuse when the cloud effect is low, and the virtual fuse condition is confirmed by taking the condition of the virtual fuse into consideration in the process of the environment condition of the virtual fuse. (Liu Yun, peng Xinge, zhang Jun, jiang Fei. Research on anti-cloud and mist interference of certain missile laser fuze [ J ]. Technological innovation and application, 2022,12 (03): 18-20.) The current research has achieved a certain result in the aspects of training time, calculated amount and accuracy, but the algorithm research is operated on a powerful GPU platform, so that the detection requirements of fuze hardware with low actual battlefield integration level, low module function complexity and low single detection precision can not be met, and meanwhile, the infl