CN-122020308-A - Radar working mode identification method and system based on quantum free search mechanism evolution BP neural network

CN122020308ACN 122020308 ACN122020308 ACN 122020308ACN-122020308-A

Abstract

The invention provides a radar working mode identification method and system based on a quantum free search mechanism evolution BP neural network, and belongs to the field of radar signal processing. The method aims to solve the problems that the existing evolutionary algorithm-based neural network-based radar working mode identification work is prone to being in local optimum, insufficient in convergence precision and weak in expandability. According to the invention, a small part of training set is used for training in the quantum free search stage, a high-potential area is rapidly positioned in a parameter space by introducing direction perception quantum rotation gate updating and self-adaptive quantum variation operation, a high-quality initial weight is generated, low-efficiency convergence caused by random initialization is avoided, and a large part of training set is used for training in the BP neural network optimization stage, and four steps of forward propagation, error calculation, reverse propagation and parameter updating are used for continuous iteration, so that network parameters are gradually optimized, a loss function tends to be minimum, and model performance is continuously improved.

Inventors

WANG JIAQI
GAO HONGYUAN
LIU QINGLING
GU XIAOYUAN
ZHANG WEI
ZHANG YAN
WEI JIALIN
WU HUAN
GU Yu

Assignees

哈尔滨工程大学

Dates

Publication Date: 20260512
Application Date: 20260130

Claims (10)

1. A radar working mode identification method based on a quantum free search mechanism evolution BP neural network is characterized by comprising the following steps: s100, constructing and dividing a radar signal characteristic data set for identifying a radar working mode; s200, initializing a BP neural network, and determining a neural network structure and an excitation function; s300, initializing a quantum free searching population and defining a quantum position; S400, mapping the quantum position of each quantum unit to obtain the position of the quantum unit and converting the position of each quantum unit into BP neural network model parameters; s500, evaluating the fitness value of the quantum individuals by using the small training set obtained by dividing in the step S100, selecting quantum elite, and then initializing global optimal fitness and global optimal position vectors; s600, updating quantum positions of non-elite quantum individuals in the quantum free searching population by using a direction-sensing quantum revolving door, and updating probability breadth of partial quantum elite individuals by using a self-adaptive quantum variation strategy so as to enhance diversity of the quantum free searching population; S700, evaluating the fitness of all quantum individuals by using the small training set obtained by dividing in the step S100, selecting quantum elite and updating global optimum, and determining global optimum fitness and global optimum position vectors; S800, judging whether the maximum iteration number is reached, if so, stopping iteration, and continuing to execute downwards, otherwise, returning to the step S600, and continuing iteration until the maximum iteration number is reached; S900, outputting a global optimal position vector obtained by quantum free search by using a small part of training set, correspondingly using the global optimal position vector as an initial weight and a threshold value of the BP neural network, and training by using a traditional back propagation method by using a large part of training set so as to improve the model prediction accuracy; S1000, judging whether the maximum iteration number is reached again, if so, stopping iteration, and continuing to execute downwards, otherwise, returning to the step S900, and continuing iteration until the maximum iteration number is reached; s1100, taking the trained BP neural network with the optimal weight and the threshold value as a classifier for radar working mode identification, and carrying out radar working mode identification on the radar signal characteristic data acquired in real time.
2. The method for identifying the radar working mode based on the quantum free search mechanism evolution BP neural network according to claim 1, wherein in step S100, a radar signal characteristic data set for identifying the radar working mode is divided into a training set and a testing set, wherein the training set is a total training set and is used for training a model and optimizing model parameters, and the testing set is used for evaluating the final performance of the model after training is completed and checking the generalization capability of the model; Total training set Dividing into small training sets And most of training set A small portion of the training set The proportion of the total training set is Most training sets The proportion of the total training set is And (2) and A small part of training set Used as a quantum free searching stage, most training sets And (5) reserving.
3. The method for identifying radar operation mode based on quantum free search mechanism evolution BP neural network according to claim 2, wherein in step S200, the radar signal characteristic dataset constructed in step S100 is constructed as a system input vector The output class probability distribution vector is constructed as a system output vector According to the input-output sequence of the system Determining the node number of the BP neural network input layer Number of hidden layer nodes And output layer node number The BP neural network is enabled to input the node number of the layer Equal to the system input vector Number of nodes in network output layer Equal to the system output vector Is a dimension of (2); Using symbols To represent hidden layer number Individual neurons and input layer Connection weights of individual neurons using symbols To represent the output layer Individual neurons and hidden layer Connection weights between individual neurons using symbols To represent hidden layer number Threshold of individual neurons, using symbols To represent the output layer A threshold for individual neurons; Given hidden layer number Individual neuron excitation function Wherein Indicating hidden layer number Linear weighted input of each neuron before excitation function processing, given output layer first The individual neuron excitation function is Wherein Representing output layer number The neurons are linearly weighted before the excitation function process.
4. The method for identifying radar working mode based on quantum free search mechanism evolution BP neural network of claim 3, wherein in step S300, the number of quantum individuals in the quantum free search population is set as The quantum position of the quantum unit is formed by Represented by a number of qubits, where the number of qubits The number of parameters is the same as that of BP neural network Substitute for the first Individual quantum of the first The dimensional qubits are expressed as And meets the normalization condition ; First, the Substitute for the first The quantum positions of individual quantum units are defined as Wherein , Quantum bit And Respectively defined as , , Let the initial iteration number be Randomly generating quantum positions of quantum free searching population in a quantum bit definition domain, and enabling the maximum iteration number to be 。
5. The method for identifying radar operating mode based on quantum free search mechanism evolution BP neural network according to claim 4, wherein in step S400, a maximum boundary is set And a minimum boundary Respectively is And Mapping the quantum positions of all quantum units to obtain corresponding positions, the first Substitute for the first Individual quantum number The mapping equation of the dimension is Wherein , , , ; Mapping by a mapping equation to obtain the first Substitute for the first Position vector of individual quanta The obtained first Substitute for the first Position vector of individual quanta Mapped to a set of parameter vectors of the BP neural network.
6. The method for identifying radar operation mode based on quantum free search mechanism evolution BP neural network according to claim 5, wherein in step S500, the method comprises the steps of Substitute for the first Position vector mapped by individual quantum units Loading BP neural network, setting the initial weight between input layer and hidden layer as Wherein Threshold value is Wherein Setting the initial weight between the hidden layer and the output layer as Wherein Threshold value is Wherein , Namely the number of qubits; For small part training set Evaluating to obtain the first Substitute for the first The fitness function value of each quantum individual is Wherein For a small part of training set Number of samples in, wherein Indicating that obtain the first BP neural network pair of network parameters The result of the identification of the individual samples, Represent the first The actual class of the individual samples is that, For identifying the function, when the identification result of the BP neural network is the same as the actual category of the sample, the output value of the function is indicated to be 1, and when the identification result of the BP neural network is different from the actual category of the sample, the output value of the function is indicated to be 0; Definition of elite ratio as Will be at the first The generation amount is arranged in descending order according to the fitness value and then is positioned in front in the free searching population The quantum units of bits are defined as quantum elites, in which Representing a downward rounding function; the individual with the highest fitness in the quantum free searching population is called the leader of the quantum free searching population, and the fitness of the leader in the first-generation quantum free searching population is set as Setting the position vector of the leader in the first-generation quantum free search population as Initializing global optimum fitness And a global optimal position vector And making it , 。
7. The method for identifying radar working mode of BP neural network based on quantum free search mechanism evolution according to claim 6, wherein in step S600, in the evolution process, the captain behavior with highest fitness causes learning of non-elite quantum individuals, and the captain behavior is the first to Substitute for the first Individual non-elite quantum units The individual qubits are updated to Wherein the first Substitute for the first Individual non-elite quantum units The dynamic rotation angle updating formula corresponding to each quantum bit is as follows , , The basic rotation coefficient is represented by a reference number, Represents the decay index of the sample, Is at A random number between the two random numbers, Representing an absolute value function; Defining the probability of variation occurrence as Wherein For initial mutation probability, using quantum NOT gate to exchange probability amplitude of mutated qubit, to be selected Substitute for the first Individual quantum elite individual first The probability amplitude of the individual qubits is updated to 。
8. The method for identifying radar working mode based on quantum free search mechanism evolution BP neural network of claim 7, wherein in step S700, after the quantum positions of all quantum individuals in the quantum free search population are updated by using a direction-aware quantum rotation gate and an adaptive quantum variation strategy, a first result is obtained Substitute for the first The quantum position expression of each quantum unit is ; Mapping the quantum positions of all quantum individuals in the quantum free search population into positions, wherein the mapping equation is that Wherein , Mapping to obtain the first using the mapping equation Substitute for the first Position vector of individual quanta Will be at the first Substitute for the first Position vector of individual quanta Loading BP neural network to obtain the first The initial weight between the input layer and the hidden layer is First, a third step The threshold is Setting the first The initial weight between the hidden layer and the output layer is First, a third step The threshold is ; According to the fitness function Wherein Representation acquisition of the first Neural network pair of network parameters Recognition results of the individual samples; Definition of elite ratio as Will be at the first The generation amount is arranged in descending order according to the fitness value and then is positioned in front in the free searching population The quantum units of the bit are defined as quantum elite, set up The adaptability of the leader in the generation quantum free search population is that Let the first The position vector of the leader in the generation sub-free search population is Will be And up to the first Global optimum adaptation to generations Comparing, when the quantum is free to search the adaptability of the leader in the population Above until the first Global optimum adaptation to generations When the global optimum fitness is updated And a global optimal position vector And making it , Otherwise, the global optimum fitness and the global optimum position vector are kept unchanged, that is, 。
9. The method for identifying radar operation mode based on quantum free search mechanism evolution BP neural network according to claim 8, wherein in step S900, training period is set as Outputting the global optimal position vector Assigning the initial weight value to be the initial weight value and the threshold value of the network, and assigning the initial weight value between the input layer and the hidden layer to be Initial threshold value assignment of The initial weight value between the hidden layer and the output layer is assigned as Initial threshold value assignment of The optimal weight and threshold obtained by evolving BP neural network through quantum free search mechanism are used as most training sets Training an initial weight and a threshold value of the BP neural network; Using most training sets When training the neural network parameters, the first The secondary training process is as follows: hidden layer output calculation, setting input vector Suppose the hidden layer output is Then there is Wherein To conceal the layer excitation function, take , , ; Output layer output calculation, namely setting the input weighted sum of the output layer as Then there is BP neural network prediction output is Then there is ; Error calculation, setting desired output Calculating network prediction errors Wherein ; Updating weight according to network prediction error Updating network connection weights And So that And is also provided with Wherein Is the learning rate; threshold updating based on network prediction error Updating network node thresholds And Wherein , ; , 。
10. The radar working mode identification system based on the BP neural network evolved by the quantum free search mechanism is characterized by comprising a program module corresponding to the steps of any one of the claims 1-9, and the steps in the radar working mode identification method based on the BP neural network evolved by the quantum free search mechanism are executed in running.

Description

Radar working mode identification method and system based on quantum free search mechanism evolution BP neural network Technical Field The invention relates to the technical field of radar signal processing, in particular to a radar working mode identification method and system based on a quantum free search mechanism evolution BP neural network. Background The radar working mode identification is a technology for judging the current functional state of the radar by analyzing the radar signal parameters and the behavior characteristics. The radar working mode identification is used as an important link in electronic reconnaissance, the significance of the radar working mode identification is not only limited to the state monitoring of a radar system, but also indirectly reveals the electronic threat degree formed by enemies to the my through the analysis of the working state of the enemy radar. Thus, accurate operation pattern recognition has a crucial role in military countermeasure. The existing radar working mode identification method mainly comprises an identification method based on statistical learning, an identification method based on behavior reasoning, an identification method based on machine learning and an identification method based on deep learning. From a mathematical perspective, radar operation pattern recognition can be seen as constructing mutually disjoint subsets in a multi-dimensional pulse parameter space containing parameters such as PRI (Pulse Repetition Interval ) and carrier frequency, and classifying each subset such that all pulse sequences in the same pattern are given the same label. Along with the complexity of modern battlefield environments, the diversification of radar signal types and the continuous promotion of interference countermeasure means, the traditional radar working mode identification methods such as a template matching method, a threshold decision method, a statistical histogram method and the like have obvious defects in real-time performance, self-adaptability and robustness. Radar working mode identification methods based on deep learning and intelligent algorithms are becoming more popular. Common radar working mode identification methods based on deep learning include a feedforward neural network method, a convolutional neural network method, a cyclic neural network method, a transducer, a self-attention model method and the like. With the rise of quantum intelligent optimization algorithm, the radar working mode identification method combining throughput sub-group intelligence and neural network is more and more important for the high efficiency and the robust performance. Therefore, the design of the radar working mode identification method with high convergence accuracy, strong robustness and easy expansion has important theoretical value and practical significance. By searching related documents Ren Cheng and the like, an improved free search algorithm is proposed in an improved free search algorithm published in microcomputer information (2012, vol.28, no.10, pp.454-455, 460), a dynamic stretching technology is introduced on the basis of an original free search algorithm (FS), an objective function is modified through two function transformations, a local extremum region is removed, and the forced algorithm searches a potential optimal region which is not developed, so that early-maturing convergence is effectively avoided, global searching capacity is enhanced and local optimality is avoided. However, the stretching amplitude parameter needs to be set manually, and incorrect parameter selection may cause excessive stretching, thereby losing the potential optimal stretching or insufficient stretching, and thus the local extremum cannot be effectively escaped. And the algorithm depends on an initial detection result, if the initial search does not detect an extreme point close to global optimum, the subsequent stretching can lead the algorithm to deviate from the correct direction, and repeated iterative correction is needed. Gu Bangling, shi Yanling, jiang Lei et al, published in "technical innovation and application" (2023, vol.13, no.22, pp.15-18), "convolutional neural network-based radar operation pattern recognition" propose a convolutional neural network-based radar operation pattern recognition method. The method directly and automatically extracts high-dimensional semantic information from the time-frequency characteristic diagram of the original pulse description word by constructing a multi-level convolution-pooling module, thereby avoiding the complexity of artificial characteristic engineering. In the network design, the convergence stability of the model is enhanced by adopting residual connection and batch normalization technology, and the channel attention mechanism dynamic weighting key frequency band characteristics are introduced. But the test result shows that the recognition accuracy and the convergence rate are still to be impro