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

Abstract

The invention provides a radar working mode recognition method and system based on a quantum free search mechanism evolution BP neural network, and belongs to the field of radar signal processing. The objective of the invention is to solve the problems of easy local optimum, insufficient convergence precision and weak expandability in radar working mode recognition based on an evolutionary algorithm and a neural network. According to the method, training is performed by using a small part of training sets in a quantum free search stage, a high-potential region is quickly positioned in a parameter space by introducing direction perception quantum revolving door updating and adaptive quantum mutation operation, a high-quality initial weight is generated, and low-efficiency convergence caused by random initialization is avoided; in the BP neural network optimization stage, most training sets are used for training, and continuous iteration is performed through four steps of forward propagation, error calculation, back propagation and parameter updating, so that network parameters are gradually optimized, a loss function tends to be minimum, and the model performance is continuously improved.

Description

Technical Field

[0001] This invention relates to the field of radar signal processing technology, and more specifically, to a method and system for radar operating mode recognition based on a quantum free search mechanism-evolved BP neural network. Background Technology

[0002] Radar pattern recognition is a technique that analyzes radar signal parameters and behavioral characteristics to determine its current functional status. As a crucial component of electronic reconnaissance, its significance extends beyond simply monitoring the radar system's own status. Furthermore, by analyzing the operational status of enemy radar, it indirectly reveals the level of electronic threat posed by the enemy. Therefore, accurate pattern recognition plays a vital role in military confrontations.

[0003] Existing radar operating pattern recognition methods can be mainly categorized as follows: statistical learning-based methods, behavioral reasoning-based methods, machine learning-based methods, and deep learning-based methods. From a mathematical perspective, radar operating pattern recognition can be viewed as constructing disjoint subsets within a multidimensional pulse parameter space containing parameters such as PRI (Pulse Repetition Interval) and carrier frequency, and labeling each subset with a category, assigning the same label to all pulse sequences within the same pattern. With the increasing complexity of modern battlefield environments, the diversification of radar signal types, and the continuous improvement of jamming countermeasures, traditional radar operating pattern recognition methods such as template matching, threshold decision-making, and statistical histogram methods have significant shortcomings in terms of real-time performance, adaptability, and robustness. Radar operating pattern recognition methods based on deep learning and intelligent algorithms are gaining increasing popularity. Common deep learning-based radar operating pattern recognition methods include feedforward neural networks, convolutional neural networks, recurrent neural networks, and Transformer and self-attention models. With the rise of quantum intelligent optimization algorithms, radar operating mode recognition methods combining quantum swarm intelligence and neural networks have gained increasing attention due to their high efficiency and robustness. Therefore, designing a radar operating mode recognition method with high convergence accuracy, strong robustness, and easy scalability has significant theoretical and practical value.

[0004] By reviewing relevant literature, Ren Cheng et al. proposed an improved free search algorithm in their paper "An Improved Free Search Algorithm" published in *Microcomputer Information* (2012, Vol. 28, No. 10, pp. 454-455, 460). This improved algorithm introduces a dynamic stretching technique into the original free search algorithm (FS). It modifies the objective function through two function transformations, eliminating local extrema and forcing the algorithm to explore unexplored potential optimal regions, effectively avoiding premature convergence and enhancing global search capabilities while preventing getting trapped in local optima. However, the stretching parameter needs to be manually set. Inappropriate parameter selection may lead to overstretching, resulting in the loss of potential optimal solutions, or understretching, preventing effective escape from local extrema. Furthermore, the algorithm relies on the initial detection results. If the initial search does not detect an extremum close to the global optimum, subsequent stretching may guide the algorithm away from the correct direction, requiring multiple iterations for correction. Jia Bangling, Shi Yanling, Jiang Lei, et al., published "Radar Operating Pattern Recognition Based on Convolutional Neural Networks" in *Science and Technology Innovation and Application* (2023, Vol. 13, No. 22, pp. 15-18), proposing a radar operating pattern recognition method based on convolutional neural networks. This method automatically extracts high-dimensional semantic information directly from the time-frequency feature map of the original pulse descriptor by constructing multi-level convolutional-pooling modules, avoiding the complexity of manual feature engineering. In the network design, residual connections and batch normalization techniques are used to enhance the model's convergence stability, and a channel attention mechanism is introduced to dynamically weight key frequency band features. However, experimental results show that its recognition accuracy and convergence speed still need improvement. In their paper "Adaptive Weighted Hybrid Integration Learning-Based Radar Operating Mode Recognition Method" published in the *Journal of Army Engineering University* (2024, Vol.3, No.6, pp.10-16), Wang Zhiteng, Ji Cunxiao, Liu Chang, and others proposed a WWCD intelligent optimization algorithm. This algorithm performs global optimization by simulating the physical behavior of water wave diffusion, refraction, and reflection, gradually converging to the optimal solution through multiple iterative diffusion iterations. However, the algorithm requires preset hyperparameters such as diffusion ratio and jump ratio, and the optimization process relies on experience, increasing the difficulty of engineering implementation.

[0005] Existing literature indicates that common radar operating pattern recognition methods primarily include those based on statistical learning, behavioral reasoning, traditional machine learning, and deep learning. These methods all suffer from low convergence accuracy, weak robustness, slow convergence speed, and limited scalability. The most common solution to these problems is to combine recently proposed swarm intelligence methods with neural network frameworks in the field of radar operating pattern recognition. The combination of novel swarm intelligence methods and neural network frameworks typically achieves higher convergence accuracy than traditional methods, offering some improvement over the aforementioned issues. However, since some newly proposed methods still suffer from long processing times and susceptibility to local optima, specific optimizations and improvements are needed to address these problems. Summary of the Invention

[0006] The technical problem to be solved by this invention is:

[0007] To address the problems of existing radar operating mode recognition methods based on evolutionary algorithms combined with neural networks, such as easy getting trapped in local optima, insufficient convergence accuracy, and weak scalability.

[0008] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0009] This invention provides a radar operating mode recognition method based on a quantum free search mechanism-evolved BP neural network, comprising the following steps:

[0010] S100. Construct and divide radar signal feature datasets for radar operating mode recognition;

[0011] S200. Initialize the BP neural network and determine its structure and activation function.

[0012] S300, Initialize the quantum free search population and define quantum positions;

[0013] S400: Map the quantum position of each quantum individual to obtain the position of the quantum individual and convert it into BP neural network model parameters;

[0014] S500: Use the small portion of the training set obtained in step S100 to evaluate the fitness value of the quantum individuals and perform quantum elite selection. Then initialize the global optimal fitness and the global optimal position vector.

[0015] S600 uses a direction-aware quantum rotation gate to update the quantum positions of non-elite quantum individuals in the quantum free search population, and uses an adaptive quantum mutation strategy to update the probability amplitude of some quantum elite individuals to enhance the diversity of the quantum free search population.

[0016] S700. Use the small portion of the training set obtained in step S100 to evaluate the fitness of all quantum individuals, perform quantum elite selection and global optimal update, and determine the global optimal fitness and global optimal position vector.

[0017] S800: Determine if the maximum number of iterations has been reached. If the maximum number of iterations has been reached, terminate the iteration and continue execution; otherwise, return to step S600 and continue iterating until the maximum number of iterations is reached.

[0018] S900 outputs the globally optimal position vector obtained through quantum free search using a small portion of the training set, which is then used as the initial weights and threshold of the BP neural network. The majority of the training set is used to train the model using the traditional backpropagation method to improve the model's prediction accuracy.

[0019] S1000: Check again whether the maximum number of iterations has been reached. If the maximum number of iterations has been reached, terminate the iteration and continue to the next step; otherwise, return to step S900 and continue iterating until the maximum number of iterations is reached.

[0020] S1100: The trained BP neural network with optimal weights and thresholds is used as a classifier for radar operating mode recognition to perform radar operating mode recognition on the radar signal feature data collected in real time.

[0021] Further, in step S100, the radar signal feature dataset used for radar working mode recognition is divided into a training set and a test set; wherein the training set, i.e. the total training set, is used to train the model and optimize the model parameters, and the test set is used to evaluate the final performance of the model after training and to test its generalization ability.

[0022] Total training set Divide into a small portion of the training set and most of the training set A small portion of the training set The proportion of the total training set is Most of the training set The proportion of the total training set is ,and ; a small portion of the training set Used as a phase of quantum free search, most of the training set reserve.

[0023] Further, in step S200, the radar signal feature dataset constructed in step S100 is used to construct a system input vector. The output category probability distribution vector is constructed as the system output vector. Based on the system input / output sequence Determine the number of nodes in the input layer of the BP neural network. Number of hidden layer nodes and the number of output layer nodes ; Increase the number of nodes in the input layer of the BP neural network Equal to system input vector The dimension, the number of nodes in the network output layer Equal to the system output vector The dimension;

[0024] Use symbols To represent the hidden layer The nth neuron and the input layer The connection weights of each neuron, using symbols To represent the output layer The first neuron and the hidden layer The connection weights between neurons, using symbols To represent the hidden layer The threshold of each neuron, using the symbol To represent the output layer The threshold of a neuron;

[0025] Given the hidden layer Neuron activation function ,in Indicates the hidden layer number 1 The linearly weighted input of the n neurons before activation function processing; given the nth output layer... The activation function of each neuron is: ,in Indicates the output layer number The linearly weighted input of each neuron before the activation function processing.

[0026] Furthermore, in step S300, let the number of quantum individuals in the quantum free search population be... The quantum position of a quantum individual is determined by... It is represented by qubits, where the number of qubits The same number of parameters as the BP neural network; the first The generation The first quantum individual A 1-dimensional quantum bit is represented as And it satisfies the normalization condition. ;

[0027] No. The generation The quantum position of a quantum individual is defined as follows: ,in , qubit and They are defined as , , Let the initial number of iterations be... Randomly generate the quantum positions of the quantum free search population within the domain of the qubit, and let the maximum number of iterations be... .

[0028] Furthermore, in step S400, the maximum boundary is set. and minimum boundary They are respectively and The corresponding position is obtained by mapping the quantum positions of all quantum individuals. The generation The quantum individual of the first The mapping equation for dimension is ,in , , , ;

[0029] The mapping equation is used to obtain the first... The generation The position vector of each quantum individual The obtained first The generation The position vector of each quantum individual This is mapped to a set of parameter vectors of a BP neural network.

[0030] Further, in step S500, the first The generation The position vector obtained by mapping individual quantum entities Load the BP neural network; set the initial weights between the input layer and the hidden layer to be... ,in The threshold is ,in Let the initial weights between the hidden layer and the output layer be... ,in The threshold is ,in , That is, the number of qubits;

[0031] For a small portion of the training set Conduct an evaluation; obtain the first The generation The fitness function value of each quantum individual is ,in For a small portion of the training set The number of samples within, of which This indicates that the first [item / number] has been obtained. A BP neural network with alternating network parameters for the first The identification results of each sample Indicates the first The actual category of each sample The indicator function is defined as follows: when the recognition result of the BP neural network is the same as the actual category of the sample, the output value of the indicator function is 1; when the recognition result of the neural network is different from the actual category of the sample, the output value of the indicator function is 0.

[0032] Define the proportion of elites as , will the In the quantum free search population, the top-ranked individuals are arranged in descending order of fitness value. A quantum individual of a given degree is defined as a quantum elite, among which... This represents the floor function;

[0033] The individual with the highest fitness in the quantum free search population is called the leader of the population. Let the fitness of the leader in the first generation of the quantum free search population be . Let the position vector of the leader in the first generation of quantum free search population be... Initialize the global optimal fitness and the global optimal position vector ; make , .

[0034] Furthermore, in step S600, during the evolutionary process, the behavior of the leader with the highest fitness will induce learning in non-elite quantum individuals, thus... The generation The first non-elite quantum individual The qubits are updated to , of which The generation The first non-elite quantum individual The formula for updating the dynamic rotation angle for each qubit is: , , Indicates the fundamental rotation coefficient. Indicates the decay index, Is Random numbers between This represents the function that takes the absolute value.

[0035] Define the probability of mutation as ,in The initial mutation probability is used; using a quantum NOT gate, the probability magnitudes of the mutated qubits are swapped, and the selected qubit is... The generation The first quantum elite individual The probability amplitude of each qubit is updated to .

[0036] Further, in step S700, after updating the quantum positions of all quantum individuals in the quantum free search population using a direction-aware quantum rotation gate and an adaptive quantum mutation strategy, the first... The generation The quantum position expression for each quantum individual is: ;

[0037] Mapping the quantum positions of all quantum individuals in the quantum free-search population to their positions, the mapping equation is: ,in , ; Using the mapping equation, we obtain the first The generation The position vector of each quantum individual , will the The generation The position vector of each quantum individual Load the BP neural network to obtain the first... The initial weights between the input layer and the hidden layer are: ;No. The threshold is Let the first The initial weights between the hidden layer and the output layer are: ;No. The threshold is ;

[0038] According to the fitness function, we have ,in Indicates obtaining the first Neural network parameters for the first generation The identification results of each sample;

[0039] Define the proportion of elites as , will the In the quantum free search population, the top-ranked individuals are arranged in descending order of fitness value. A quantum individual of the 1st degree is defined as a quantum elite; let the 1st degree be defined as a quantum elite. The fitness of the leader in a quantum free search population is Let the first The position vector of the leader in the quantum free search population is ,Will With until the Global optimal fitness up to date Compare; when the fitness of the leader in this quantum free-search population Higher than until the Global optimal fitness up to date Update the global optimal fitness at that time. and the global optimal position vector ; make , Otherwise, the global optimal fitness and the global optimal position vector remain unchanged, that is, .

[0040] Further, in step S900, the training period is set to... Output the globally optimal position vector The initial weights and thresholds of the network are assigned to these values; the initial weights between the input layer and the hidden layer are assigned to these values. The initial threshold value is assigned as follows: The initial weights between the hidden layer and the output layer are assigned as follows: The initial threshold value is assigned as follows: The optimal weights and thresholds obtained by evolving a BP neural network through a quantum free search mechanism are used as the majority of the training set. Initial weights and thresholds for training a BP neural network;

[0041] Using most of the training set When training neural network parameters, the first The training process is shown below:

[0042] Hidden layer output calculation: Let the input vector be... Assuming the hidden layer output is Then there is ,in Let be the activation function of the hidden layer, and take... , , ;

[0043] Output layer output calculation: Let the weighted sum of the output layer inputs be... Then there is The BP neural network predicts the output as follows: Then there is ;

[0044] Error calculation: Let the expected output be... Calculate the network prediction error ,in ;

[0045] Weight update: based on network prediction error Update network connection weights and , making and ;in For learning rate;

[0046] Threshold update: based on network prediction error Update network node thresholds and ;in , ; , .

[0047] The radar operating mode recognition system based on the quantum free search mechanism evolutionary BP neural network has a program module corresponding to the above steps, and executes the steps in the radar operating mode recognition method based on the quantum free search mechanism evolutionary BP neural network during runtime.

[0048] Compared with the prior art, the beneficial effects of the present invention are:

[0049] (1) In radar operating mode recognition tasks, traditional BP neural networks are prone to getting trapped in local optima due to their random initialization method, which limits the model performance. However, the quantum free search mechanism introduced in this invention uses a quantum free search population to explore the parameter space in parallel, realizing a global search of the network parameter space, which greatly reduces the risk of getting trapped in local minima and enhances the stability and robustness of the radar operating mode recognition model.

[0050] (2) In radar operating mode recognition tasks, the quantum free search mechanism of this invention can quickly find a better solution region in the early stage, which significantly reduces the number of iterations required in the subsequent neural network optimization stage. Experimental results show that, under the same hardware conditions, the method of this invention reduces the number of iterations required by traditional neural network optimization by 20% to 30%, and greatly improves the convergence speed.

[0051] (3) Compared with traditional BP neural networks, the BP neural network based on quantum free search mechanism evolution of the present invention exhibits superior recognition performance in radar operating mode recognition tasks, with an accuracy improvement of approximately 2% compared to traditional methods. Furthermore, the radar operating mode recognition model based on quantum free search mechanism evolution of the present invention only requires setting a few hyperparameters such as population size and maximum number of iterations during actual deployment. In addition, since the present invention does not rely on a large number of sensitive hyperparameter tuning processes during actual deployment, the algorithm configuration process is greatly simplified, thereby lowering the threshold for practical engineering applications.

[0052] (4) In practical applications, the sample data for certain special radar operating modes is often limited (small sample size). The radar operating mode recognition method based on the quantum free search mechanism and evolutionary BP neural network proposed in this invention still has robustness advantages under small sample conditions. The quantum free search mechanism of this invention uses a quantum free search population to explore the parameter space in parallel. It does not rely on a large number of samples to estimate gradient information, but guides the quantum free search population to converge to the optimal solution region after fully exploring the parameter space through direction-aware quantum rotation gate updates and adaptive mutation operations. This global search mechanism can more reliably cover the parameter space when the sample size is insufficient, avoid getting trapped in local minima, and maintain high performance. Attached Figure Description

[0053] Figure 1 This is a flowchart of the radar operating mode recognition method based on the quantum free search mechanism evolving BP neural network in an embodiment of the present invention;

[0054] Figure 2 This is a comparison of the optimization performance of the quantum free search mechanism of the present invention and the traditional free search mechanism in the optimization stage of a small portion of the training set in an embodiment of the present invention;

[0055] Figure 3 This is a comparison of the test accuracy of the quantum free search mechanism evolution BP neural network of the present invention and the traditional BP neural network in an embodiment of the present invention;

[0056] Figure 4 This is a confusion matrix diagram of the quantum free search mechanism evolution BP neural network of the present invention on the test set in an embodiment of the present invention;

[0057] Figure 5 The ROC curve of the quantum free search mechanism evolution BP neural network of the present invention on the test set is shown in the embodiment of the present invention. Detailed Implementation

[0058] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0059] Specific Implementation Plan 1: Combining Figure 1 As shown, this invention provides a radar operating mode recognition method based on a quantum free search mechanism-evolved BP neural network, comprising the following steps:

[0060] S100. Construct and divide radar signal feature datasets for radar operating mode recognition;

[0061] The radar signal feature dataset includes pulse repetition interval, pulse width, carrier frequency, and bandwidth. The radar signal feature dataset used for radar operating mode recognition is divided into a training set and a test set. The training set is used to train the model and optimize the model parameters so that it can learn effective features and patterns from the data. The test set is used to evaluate the final performance of the model after training and to test its generalization ability.

[0062] All training sets (i.e., the total training set) is divided into a smaller training set. and most of the training set A small portion of the training set The proportion of the total training set is Most of the training set The proportion of the total training set is ,and ; a small portion of the training set Used as a phase of quantum free search, most of the training set reserve;

[0063] S200. Initialize the BP neural network and determine its structure and activation function.

[0064] The radar signal feature dataset constructed in step S100 is used to construct the system input vector. The output category probability distribution vector is constructed as the system output vector. Based on the system input / output sequence Determine the number of nodes in the input layer of the BP neural network. Number of hidden layer nodes and the number of output layer nodes ; Increase the number of nodes in the input layer of the BP neural network Equal to system input vector The dimension, the number of nodes in the network output layer Equal to the system output vector The dimension;

[0065] Use symbols To represent the hidden layer The nth neuron and the input layer The connection weights of each neuron, using symbols To represent the output layer The first neuron and the hidden layer The connection weights between neurons, using symbols To represent the hidden layer The threshold of each neuron, using the symbol To represent the output layer The threshold of a neuron;

[0066] Given the hidden layer Neuron activation function ,in Indicates the hidden layer number 1 The linearly weighted input of the n neurons before activation function processing; given the nth output layer... The activation function of each neuron is: ,in Indicates the output layer number The linearly weighted input of each neuron before the activation function processing. The activation function serves to convert the linear weighted result of each neuron in the output layer into a probability distribution;

[0067] S300, Initialize the quantum free search population and define quantum positions;

[0068] Let the number of quantum individuals in the quantum free search population be . The quantum position of a quantum individual is determined by... It is represented by qubits, where the number of qubits The number of parameters is the same as that of a BP neural network; a qubit, as the smallest unit of information, is a complex pair, the th... The generation The first quantum individual A dimensional qubit can be represented as And it satisfies the normalization condition. ;

[0069] No. The generation The quantum position of a quantum individual is defined as follows: ,in , qubit They are defined as , , Let the initial number of iterations be... Randomly generate the quantum positions of the quantum free search population within the domain of the qubit, and let the maximum number of iterations be... ;

[0070] S400: Map the quantum position of each quantum individual to obtain the position of the quantum individual and convert it into BP neural network model parameters;

[0071] Since the initial weights and thresholds of a BP neural network are generally... Random numbers between these ranges have a maximum bound. and minimum boundary Restricted to and The corresponding position is obtained by mapping the quantum positions of all quantum individuals. The generation The quantum individual of the first The mapping equation for dimension is ,in , , , ;

[0072] The mapping equation is used to obtain the first... The generation The position vector of each quantum individual The obtained first The generation The position vector of each quantum individual This is mapped to a set of parameter vectors in a BP neural network; obviously, the position obtained after mapping the quantum position... Since it meets the boundary constraints of the BP neural network, the mapped position vector can be directly regarded as the parameter vector of the BP neural network.

[0073] S500, Using the small portion of the training set obtained in step S100 The fitness values ​​of quantum individuals are evaluated, and quantum elite selection is performed. Then, the global optimal fitness and global optimal position vector are initialized.

[0074] The first The generation The position vector obtained by mapping individual quantum entities Load the BP neural network; set the initial weights between the input layer and the hidden layer to be... ,in The threshold is ,in Let the initial weights between the hidden layer and the output layer be... ,in The threshold is ,in , That is, the number of qubits;

[0075] For a small portion of the training set Conduct an evaluation; obtain the first The generation The fitness function value of each quantum individual is ,in For a small portion of the training set The number of samples within, of which This indicates that the first [item / number] has been obtained. A BP neural network with alternating network parameters for the first The identification results of each sample Indicates the first The actual category of each sample The indicator function is used for identification. When the identification result of the BP neural network is the same as the actual category of the sample, the output value of the indicator function is 1; when the identification result of the neural network is different from the actual category of the sample, the output value of the indicator function is 0. The fitness value reflects the fitness of the quantum position of different quantum individuals and their corresponding positions. The higher the fitness value, the stronger the fitness of the quantum position of the quantum individual and its corresponding positions. The quantum free search population is arranged in descending order according to the fitness value.

[0076] Define the proportion of elites as , will the In the quantum free search population, the top-ranked individuals are arranged in descending order of fitness value. A quantum individual of a given degree is defined as a quantum elite, among which... This represents the floor function; to prevent the excellent experience gained by quantum elites from being lost during the iteration process, the quantum positions of all quantum elites are retained and do not participate in subsequent direction-aware quantum rotation gate update operations;

[0077] The individual with the highest fitness in a quantum free-search population is called the leader of that population. The behavior of the leader has a significant impact on the entire quantum free-search population. Let the fitness of the leader in the first generation of the quantum free-search population be denoted as . Let the position vector of the leader in the first generation of quantum free search population be... Initialize the global optimal fitness and the global optimal position vector ; make ;

[0078] S600 uses a direction-aware quantum rotation gate to update the quantum positions of non-elite quantum individuals in the quantum free search population, and uses an adaptive quantum mutation strategy to update the probability amplitude of some quantum elite individuals to enhance the diversity of the quantum free search population.

[0079] In the process of evolution, the behavior of the leader, which exhibits the best performance and highest fitness, will be learned by non-elite quantum individuals, thus influencing the behavior of the first... The generation The first non-elite quantum individual The qubits are updated to , of which The generation The first non-elite quantum individual The formula for updating the dynamic rotation angle for each qubit is: , ,in It serves to control the rotation intensity, and its value gradually decreases with the number of iterations. It can be used for large-angle rotations in the early stages to achieve global exploration, and small-angle adjustments in the later stages to achieve local optimization. Indicates the fundamental rotation coefficient. The decay exponent represents the nonlinear decay rate that controls the rotational intensity. It acts as a directional indicator, guiding non-elite quantum individuals to learn from the leader with the highest fitness. Is The random number between these values ​​helps prevent the algorithm from getting stuck in a fixed pattern of local optima. This represents the function that takes the absolute value.

[0080] Some elite quantum individuals undergo mutations with a certain probability during evolution. These mutations help to break through directional constraints and maintain population diversity. The probability of mutation is defined as follows: ,in This represents the initial mutation probability; the mutation probability is higher in the early stages, which helps enhance the ability to explore unknown regions; the mutation probability decreases in the later stages, allowing for greater focus on local development; this applies to the qubits of all quantum elite individuals. The probability of the mutated qubit changes; using a quantum NOT gate, the probability magnitudes of the mutated qubits are swapped, and the selected qubit... The generation The first quantum elite individual The probability amplitude of each qubit is updated to ;

[0081] S700, Using the small portion of the training set obtained in step S100 The fitness of all quantum individuals is evaluated, quantum elite selection and global optimal update are performed, and the global optimal fitness and global optimal position vector are determined.

[0082] After updating the quantum positions of all quantum individuals in the quantum free search population using a direction-aware quantum rotation gate and an adaptive quantum mutation strategy, the first... The generation The quantum position expression for each quantum individual is: ,in , ;

[0083] Mapping the quantum positions of all quantum individuals in the quantum free-search population to their positions, the mapping equation is: ,in ; Using the mapping equation, we obtain the first The generation The position vector of each quantum individual , will the The generation The position vector of each quantum individual Load the BP neural network to obtain the first... The initial weights between the input layer and the hidden layer are: ,in ;No. The threshold is ,in Let the first The initial weights between the hidden layer and the output layer are: ,in ;No. The threshold is ,in ;

[0084] According to the fitness function, we have ,in Indicates obtaining the first Neural network parameters for the first generation The indicator function outputs 1 when the BP neural network's recognition result matches the sample's actual class; and 0 when the neural network's recognition result differs from the sample's actual class. The quantum free search population is then sorted in descending order of fitness values.

[0085] Define the proportion of elites as , will the In the quantum free search population, the top-ranked individuals are arranged in descending order of fitness value. A quantum individual of a given degree is defined as a quantum elite, among which... Let represent the floor function; to prevent the loss of valuable experience gained by quantum elites during iteration, the quantum positions of all quantum elites are retained and do not participate in subsequent direction-aware quantum rotation gate update operations; the individual with the highest fitness in the quantum free search population is called the leader of the quantum free search population; let the ... The fitness of the leader in a quantum free search population is Let the first The position vector of the leader in the quantum free search population is ,Will With until the Global optimal fitness up to date Compare; when the fitness of the leader in this quantum free-search population Higher than until the Global optimal fitness up to date Update the global optimal fitness at that time. and the global optimal position vector ; make Otherwise, the global optimal fitness and the global optimal position vector remain unchanged, even if... ;

[0086] S800, Judgment Has the maximum number of iterations been reached? If the maximum number of iterations is reached If the iteration terminates, continue execution; otherwise, let... Return to step S600 to continue;

[0087] S900, output using a small portion of the training set The globally optimal position vector obtained through quantum free search is used as the initial weights and threshold of a backpropagation (BP) neural network, and most of the training set is used for this purpose. The model's prediction accuracy can be further improved by training it using the traditional backpropagation method.

[0088] Set the training period to Output the globally optimal position vector The initial weights and thresholds of the network are assigned to these values; the initial weights between the input layer and the hidden layer are assigned to these values. The initial threshold value is assigned as follows: The initial weights between the hidden layer and the output layer are assigned as follows: The initial threshold value is assigned as follows: The optimal weights and thresholds obtained by evolving a BP neural network through a quantum free search mechanism are used as the majority of the training set. Initial weights and thresholds for training a BP neural network;

[0089] Using most of the training set When training neural network parameters, the first The training process is shown below:

[0090] Hidden layer output calculation: Let the input vector be... Assuming the hidden layer output is Then there is ,in Let be the activation function of the hidden layer, and take... , The number of nodes in the input layer. , This represents the number of hidden layer nodes.

[0091] Output layer output calculation: Let the weighted sum of the output layer inputs be... Then there is The BP neural network predicts the output as follows: Then there is ;

[0092] Error calculation: Let the expected output be... Calculate the network prediction error ,in ;

[0093] Weight update: based on network prediction error Update network connection weights and , making and ;in For learning rate;

[0094] Threshold update: based on network prediction error Update network node thresholds and ;in , ; , ;

[0095] S1000, Judgment Has the maximum number of iterations been reached? If the maximum number of iterations is reached If the iteration terminates, continue execution; otherwise, let... Return to step S900 to continue;

[0096] S1100: The trained BP neural network with optimal weights and thresholds is used as a classifier for radar operating mode recognition to perform radar operating mode recognition on the radar signal feature data collected in real time.

[0097] Specific Implementation Scheme 2: The present invention provides a radar operating mode recognition system based on a quantum free search mechanism evolving BP neural network. The system has a program module corresponding to the above steps, and executes the steps in the radar operating mode recognition method based on a quantum free search mechanism evolving BP neural network when running.

[0098] The other combinations and connections in this implementation scheme are the same as in Specific Implementation Scheme 1.

[0099] Simulation Experiment

[0100] The BP neural network trained with optimal weights and thresholds according to this invention is used as a classifier for radar operating mode recognition. Radar operating mode recognition is performed on test set feature data that is completely independent of the training process. The performance of the radar operating mode recognition results on the test set is evaluated by constructing a confusion matrix and plotting ROC curves.

[0101] Parameter settings

[0102] The radar signal feature dataset includes radar signal feature datasets for six typical operating modes: velocity search, track-plus-search, range-while-track, multi-target tracking, search-while-track, and single-target tracking. This dataset contains 10,000 samples, each consisting of five time steps representing five consecutive pulse transmit-receive cycles. Each time step records the pulse characteristic parameters (pulse repetition interval, pulse width, carrier frequency, and bandwidth) within that cycle. The specific ranges of these parameters are shown in Table 1.

[0103] Table 1

[0104]

[0105] The radar signal feature dataset was divided into two parts: 85% was used as the training set, and then the training set was further divided into two parts, with 15% used as a smaller training set. 70% was allocated to the majority of the training set. The remaining 15% is allocated to the test set, but this method is not limited to this allocation method.

[0106] Set the number of nodes in the input layer of the BP neural network. The value is 20, the number of hidden layer nodes The value is 128, the number of output layer nodes. The value is 6.

[0107] The performance comparison between the quantum free search mechanism and the traditional free search mechanism in the optimization stage of a small portion of the training set is shown in the figure below. Figure 2 As shown, the parameter settings for the traditional free search mechanism are found in the paper "An Improved Free Search Algorithm" published by Ren Cheng et al. in *Microcomputer Information* (2012, Vol. 28, No. 10, pp. 454-455+460). To ensure fairness, the population size and maximum number of iterations are set to be the same for both mechanisms. The number of quantum individuals within the quantum free search population is... The value is set to 100, which represents the maximum number of iterations. The number of individuals in the traditional free search population was set to 100, and the maximum number of iterations was set to 200. Experimental results show that the quantum free search mechanism can rapidly increase the global optimal fitness value in the early stages, indicating that it has strong global search capabilities and fast convergence characteristics. After 150-180 iterations, the fitness value tends to stabilize, eventually converging at a fitness value of approximately 0.73. This convergence process exhibits a typical "rapid ascent-stable convergence" pattern, demonstrating the good performance of the quantum free search mechanism. Comparing the quantum free search mechanism with the traditional free search mechanism, the results show that the quantum free search mechanism has a faster convergence speed and an optimal fitness value that is about 3% higher than that of the traditional free search mechanism.

[0108] To comprehensively compare the performance gap between BP neural networks evolving from quantum free search mechanisms and traditional BP neural networks, a comparison of the test accuracy of BP neural networks evolving from quantum free search mechanisms and traditional BP neural networks is presented as follows: Figure 3 As shown, the training cycle Set it to 200. For example... Figure 3 The BP neural network representing the evolution of the quantum free search mechanism achieved basic convergence after 150 training epochs, while the traditional BP neural network converged after about 200 epochs. Furthermore, the BP neural network with the quantum free search mechanism achieved approximately 2% higher recognition accuracy than the traditional BP neural network. This demonstrates that the BP neural network with the quantum free search mechanism exhibits significant advantages over the traditional BP neural network in terms of convergence speed and accuracy.

[0109] To further analyze the confusion between different categories, the training set was used with a large number of quantum individuals. Value set to 100, maximum number of iterations Set to 200, training cycles The model is trained with a setting of 200, and then evaluated using a test set that is completely independent of the training set. A schematic diagram of the test set confusion matrix is ​​shown below. Figure 4 The ROC curves for the test set are shown in the figure below. Figure 5 As shown in the diagram, analysis of the diagonal elements of the confusion matrix reveals that the number of correctly classified test set samples dominates for each category. Simultaneously, the AUC values ​​for all categories are above 0.95, far exceeding the 0.5 benchmark for random classification, indicating that the model maintains excellent discriminative performance across different classification thresholds. The diagonal elements and AUC values ​​of the confusion matrix show that the vast majority of samples are classified well. The classification accuracy for velocity search and single-target tracking is the best, consistently achieving correct classification for these two categories. The recognition accuracy for range-based tracking and multi-target tracking also reaches high levels. However, there is some confusion between the tracking-plus-search and search-based tracking modes, resulting in lower recognition accuracy. Overall, the BP neural network evolved from the quantum free search mechanism demonstrates good classification performance and strong class discrimination ability in multi-class recognition tasks.

[0110] While the present invention has been disclosed above, its scope of protection is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, and all such changes and modifications will fall within the scope of protection of the present invention.

Claims

1. A radar operating mode recognition method based on a quantum free search mechanism evolving BP neural network, characterized in that, Includes the following steps: S100. Construct and divide radar signal feature datasets for radar operating mode recognition; S200. Initialize the BP neural network and determine its structure and activation function. S300, Initialize the quantum free search population and define quantum positions; S400: Map the quantum position of each quantum individual to obtain the position of the quantum individual and convert it into BP neural network model parameters; S500: Use the small portion of the training set obtained in step S100 to evaluate the fitness value of the quantum individuals and perform quantum elite selection. Then initialize the global optimal fitness and the global optimal position vector. S600 uses a direction-aware quantum rotation gate to update the quantum positions of non-elite quantum individuals in the quantum free search population, and uses an adaptive quantum mutation strategy to update the probability amplitude of some quantum elite individuals to enhance the diversity of the quantum free search population. S700. Use the small portion of the training set obtained in step S100 to evaluate the fitness of all quantum individuals, perform quantum elite selection and global optimal update, and determine the global optimal fitness and global optimal position vector. S800: Determine if the maximum number of iterations has been reached. If the maximum number of iterations has been reached, terminate the iteration and continue execution; otherwise, return to step S600 and continue iterating until the maximum number of iterations is reached. S900 outputs the globally optimal position vector obtained through quantum free search using a small portion of the training set, which is then used as the initial weights and threshold of the BP neural network. The majority of the training set is used to train the model using the traditional backpropagation method to improve the model's prediction accuracy. S1000: Check again whether the maximum number of iterations has been reached. If the maximum number of iterations has been reached, terminate the iteration and continue to the next step; otherwise, return to step S900 and continue iterating until the maximum number of iterations is reached. S1100: The trained BP neural network with optimal weights and thresholds is used as a classifier for radar operating mode recognition to perform radar operating mode recognition on the radar signal feature data collected in real time.

2. The radar operating mode recognition method based on a quantum free search mechanism evolving BP neural network according to claim 1, characterized in that: In step S100, the radar signal feature dataset used for radar working mode recognition is divided into a training set and a test set; the training set, i.e. the total training set, is used to train the model and optimize the model parameters, while the test set is used to evaluate the final performance of the model after training and to test its generalization ability. The total training set Divide into a small portion of the training set and most of the training set A small portion of the training set The proportion of the total training set is Most of the training set The proportion of the total training set is ,and ; a small portion of the training set Used as a phase in the quantum free search, most of the training set reserve.

3. The radar operating mode recognition method based on a quantum free search mechanism evolving BP neural network according to claim 2, characterized in that: In step S200, the radar signal feature dataset constructed in step S100 is used to construct a system input vector. The output category probability distribution vector is constructed as the system output vector. Based on the system input / output sequence Determine the number of nodes in the input layer of the BP neural network. Number of hidden layer nodes and the number of output layer nodes ; Increase the number of nodes in the input layer of the BP neural network Equal to system input vector The dimension, the number of nodes in the network output layer Equal to the system output vector The dimension; Use symbols To represent the hidden layer The nth neuron and the input layer The connection weights of each neuron, using symbols To represent the output layer The first neuron and the hidden layer The connection weights between neurons, using symbols To represent the hidden layer The threshold of each neuron, using the symbol To represent the output layer The threshold of each neuron; Given the hidden layer activation function of each neuron ,in Indicates the hidden layer number 1 The linearly weighted input of the n neurons before activation function processing; given the nth output layer... The activation function of each neuron is: ,in Indicates the output layer number The linearly weighted input of each neuron before processing by the activation function.

4. The radar operating mode recognition method based on a quantum free search mechanism evolving BP neural network according to claim 3, characterized in that: In step S300, let the number of quantum individuals in the quantum free search population be... The quantum position of a quantum individual is determined by... It is represented by qubits, where the number of qubits The same number of parameters as the BP neural network; the first The generation The first quantum individual A 1-dimensional quantum bit is represented as And it satisfies the normalization condition. ; No. The generation The quantum position of a quantum individual is defined as follows: ,in , qubit and They are defined as , , ; Let the initial number of iterations be Randomly generate the quantum positions of the quantum free search population within the domain of the qubit, and let the maximum number of iterations be... .

5. The radar operating mode recognition method based on a quantum free search mechanism evolving BP neural network according to claim 4, characterized in that: In step S400, the maximum boundary is set. and minimum boundary They are respectively and The corresponding position is obtained by mapping the quantum positions of all quantum individuals. The generation The quantum individual of the first The mapping equation for dimension is: ,in , , , ; The mapping equation is used to obtain the first... The generation The position vector of each quantum individual The obtained first The generation The position vector of each quantum individual This is mapped to a set of parameter vectors of a BP neural network.

6. The radar operating mode recognition method based on a quantum free search mechanism evolving BP neural network according to claim 5, characterized in that: In step S500, the first The generation The position vector obtained by mapping individual quantum entities Load the BP neural network; set the initial weights between the input layer and the hidden layer to be... ,in The threshold is ,in Let the initial weights between the hidden layer and the output layer be... ,in The threshold is ,in , That is, the number of qubits; For a small portion of the training set Conduct an evaluation; obtain the first The generation The fitness function value of each quantum individual is ,in For a small portion of the training set The number of samples within, of which This indicates that the first [item / number] has been obtained. A BP neural network with alternating network parameters for the first The identification results of each sample Indicates the first The actual category of each sample The indicator function is defined as follows: when the recognition result of the BP neural network is the same as the actual category of the sample, the output value of the indicator function is 1; when the recognition result of the neural network is different from the actual category of the sample, the output value of the indicator function is 0. Define the proportion of elites as , will the In the quantum free search population, the top position is determined by sorting the populations in descending order of fitness value. A quantum individual with a specific quantum position is defined as a quantum elite, among which... This represents the floor function; The individual with the highest fitness in the quantum free search population is called the leader of the population. Let the fitness of the leader in the first generation of the quantum free search population be . Let the position vector of the leader in the first generation of quantum free search population be... Initialize the global optimal fitness and the global optimal position vector ; make , .

7. The radar operating mode recognition method based on a quantum free search mechanism evolving BP neural network according to claim 6, characterized in that: In step S600, during the evolutionary process, the behavior of the leader with the highest fitness will induce learning in non-elite quantum individuals, thus... The generation The first non-elite quantum individual The qubits are updated to , of which The generation The first non-elite quantum individual The formula for updating the dynamic rotation angle for each qubit is: , , Indicates the fundamental rotation coefficient. Indicates the decay index, Is Random numbers between This represents the function that takes the absolute value. Define the probability of mutation as ,in The initial mutation probability is used; using a quantum NOT gate, the probability magnitudes of the mutated qubits are swapped, and the selected qubit is... The generation The first quantum elite individual The probability amplitude of each qubit is updated to .

8. The radar operating mode recognition method based on a quantum free search mechanism evolving BP neural network according to claim 7, characterized in that: In step S700, after updating the quantum positions of all quantum individuals in the quantum free search population using a direction-aware quantum rotation gate and an adaptive quantum mutation strategy, the first... The generation The quantum position expression for each quantum individual is: ; Mapping the quantum positions of all quantum individuals in the quantum free-search population to their positions, the mapping equation is: ,in , ; Using the mapping equation, we obtain the first The generation The position vector of each quantum individual , will the The generation The position vector of each quantum individual Load the BP neural network to obtain the first... The initial weights between the input layer and the hidden layer are: ;No. The threshold is Let the first The initial weights between the hidden layer and the output layer are: ;No. The threshold is ; According to the fitness function, we have ,in Indicates obtaining the first Neural network parameters for the first generation The identification results of each sample; Define the proportion of elites as , will the In the quantum free search population, the top position is determined by sorting the populations in descending order of fitness value. A quantum individual of the 1st degree is defined as a quantum elite; let the 1st degree be defined as a quantum elite. The fitness of the leader in a quantum free search population is Let the first The position vector of the leader in the quantum free search population is ,Will With until the Global optimal fitness up to date Compare; the fitness of the leader in this quantum free-search population. Higher than until the Global optimal fitness up to date Update the global optimal fitness at that time. and the global optimal position vector ; make , Otherwise, the global optimal fitness and the global optimal position vector remain unchanged, that is, .

9. The radar operating mode recognition method based on a quantum free search mechanism evolving BP neural network according to claim 8, characterized in that: In step S900, the training period is set to... Output the globally optimal position vector The initial weights and thresholds of the network are assigned to these values; the initial weights between the input layer and the hidden layer are assigned to these values. The initial threshold value is assigned as follows: The initial weights between the hidden layer and the output layer are assigned as follows: The initial threshold value is assigned as follows: ; The optimal weights and thresholds obtained by evolving a BP neural network through a quantum free search mechanism are used as the majority of the training set. Initial weights and thresholds for training a BP neural network; Using most of the training set When training neural network parameters, the first The training process is shown below: Hidden layer output calculation: Let the input vector be... Assuming the hidden layer output is Then there is ,in Let be the activation function of the hidden layer, and take... , , ; Output layer output calculation: Let the weighted sum of the output layer inputs be... Then there is The BP neural network predicts the output as follows: Then there is ; Error calculation: Let the expected output be... Calculate the network prediction error ,in ; Weight update: based on network prediction error Update network connection weights and , making and ;in For learning rate; Threshold update: based on network prediction error Update network node thresholds and ;in , ; , .

10. A radar operating mode recognition system based on a quantum free search mechanism evolving BP neural network, characterized in that: The system has a program module corresponding to the steps of any one of the claims 1-9 above, and executes the steps in the radar operating mode recognition method based on quantum free search mechanism evolution BP neural network described above when running.

Images