Radar threat level prediction method based on attention mechanism and LSTM

Abstract

The present invention discloses a radar threat level prediction method based on an attention mechanism and LSTM. The method aims to address the problems of incomplete threat level prediction datasets constructed by existing radar threat level prediction methods, insufficient objectivity and rationality in the prediction of radar threat levels, inability to highlight threat attributes that are more critical to the threat level, and low prediction accuracy. The implementation steps of the present invention include: generating a training set containing radar threat attributes and radar threat levels; constructing an Attention-LSTM neural network that combines an attention mechanism and LSTM; optimizing the hyperparameters of the Attention-LSTM neural network using a genetic algorithm; training the Attention-LSTM network using the training set and the hyperparameters optimized by the genetic algorithm; and predicting the radar threat level using the trained Attention-LSTM network. The present invention has the advantages of a more complete constructed dataset, more objective and reasonable prediction results, and higher prediction accuracy.

Description

Technical Field

[0001] The present invention relates to the field of communications technology, and more specifically, to a radar threat level prediction method based on an attention mechanism and a long short-term memory (LSTM) network in the field of radar communications technology. The present invention can be used to predict the threat level of an opposing radar emitter. Background Art

[0002] In radar countermeasures, radar threat level prediction involves scientifically estimating the likely future trends and levels of threat levels based on the objective evolution of past radar threat developments and changes, using various analytical methods. Accurate radar threat level prediction facilitates the proactive implementation of various threat countermeasures and is crucial for maintaining electromagnetic situational initiative. However, existing radar threat level prediction methods utilize incomplete datasets, resulting in inadequate and unbiased predictions. Furthermore, they fail to highlight the more critical threat attributes, compromising prediction accuracy and potentially leading to a loss of electromagnetic situational initiative.

[0003] Henan University has proposed a radar-based method for predicting aerial target threats in its patent application, "Aerial Target Threat Prediction Method Based on an Improved GM(1,1) Model" (Application Number: CN 202210087937.1; Publication Number: CN 114519262A). The method involves the following steps: First, the threat value of the incoming aerial target is obtained, the inherent residuals in the traditional GM(1,1) model are corrected, and the weights are determined based on the magnitude of each component in the first-order cumulative sequence. Second, the sum of squared residuals is weighted according to the time weight sequence, and the optimal problem is solved to determine the time parameter. Third, the residual sequence, after Fourier series fitting, is fed back to the GM(1,1) model to obtain the final improved GM(1,1) model. Fourth, the threat value of the incoming aerial target is calculated based on the improved prediction model, achieving accurate prediction of the threat value of the aerial target. The shortcomings of this method are that it does not consider the impact of the radar's threat attributes on the threat level in the future. The data set used for target threat prediction is only a one-dimensional threat level vector, which does not include the radar's threat attributes. The constructed data set is incomplete and cannot objectively and reasonably predict the radar threat level.

[0004] Xu Gongguo et al. proposed a target threat degree prediction and evaluation method based on the Elman neural network in their published paper "Target Threat Degree Prediction and Evaluation Based on Improved Elman Neural Network" (Journal of Detection and Control, 2017, 39(03):101-106). The specific steps of this method are: first, select the target threat degree influencing attributes. Second, use the quantum particle swarm optimization (QPSO) algorithm to improve the Elman neural network. Third, construct a dynamic prediction and evaluation model of the target threat degree based on the optimized QPSO-Elman neural network. Fourth, use the constructed prediction and evaluation model to evaluate the target threat degree. The disadvantage of this method is that it gives the same weight to each threat attribute of the target threat degree, and fails to highlight the more critical threat attributes that affect the threat degree prediction effect, resulting in a low prediction accuracy. Summary of the Invention

[0005] The purpose of the present invention is to address the deficiencies in the above-mentioned prior art and propose a radar threat level prediction method based on an attention mechanism and LSTM, so as to solve the problems that the threat level prediction data set constructed by the existing radar threat level prediction method is incomplete, the prediction of the radar threat level is not objective and reasonable, and the more critical threat attributes that affect the threat level prediction effect are not highlighted, resulting in a low accuracy rate in predicting the radar threat level.

[0006] The present invention achieves its objectives by constructing a set of radar threat attributes, including distance, altitude, azimuth, carrier frequency, pulse width, repetition rate, and phase, when generating a dataset. These attributes are then evaluated to determine the corresponding threat level at each time point. This dataset, comprised of both radar threat attributes and threat level, addresses the incomplete threat level prediction datasets typically found in existing threat level prediction methods. The present invention combines the attention mechanism (Attention) with an LSTM neural network to construct an Attention-LSTM neural network based on the Attention mechanism and LSTM. The hyperparameters of the Attention-LSTM neural network are then optimized using a genetic algorithm. The Attention mechanism assigns different weights to factors influencing the radar threat level, highlighting more critical factors. Furthermore, the genetic algorithm can be used to quickly find the optimal parameter combination within the model's parameter space. This allows the present invention to highlight the more important factors influencing radar threat level prediction results, helping the model make more accurate judgments and thus improving model prediction accuracy. The Attention-LSTM neural network is trained using a training set consisting of radar threat attributes such as distance, altitude, azimuth, carrier frequency, pulse width, repetition rate, and phase, as well as the radar threat level. This gives the trained Attention-LSTM neural network good predictive capabilities, and a genetic algorithm is used to optimize the model's hyperparameters. When predicting the threat level, passing the test data through the trained Attention-LSTM neural network can highlight the key factors that affect the threat level, thereby more accurately predicting the radar threat level.

[0007] To achieve the above objectives, the implementation scheme of the present invention includes the following:

[0008] Step 1: Generate training set:

[0009] Step 1.1: Generate a time series of each radar threat attribute changing over time when the enemy airborne radar emitter switches between three different working modes within 10 seconds.

[0010] Step 1.2: Sample the time-varying sequence of each threat attribute at a sampling interval of 1 ms to obtain the historical time series corresponding to each threat attribute.

[0011] Step 1.3: Use the approximate ideal solution sorting algorithm to comprehensively calculate the threat attribute values ​​of each identical sampling point in all threat attribute time series to obtain the threat level value of the sampling point;

[0012] In step 1.4, each threat attribute and threat level of each sampling point are combined into a sample, and a threat level time series is obtained according to the time sequence of the sampling points; all radar threat attribute time series and threat level time series are combined into a sample set;

[0013] Step 1.5: Normalize the samples in the sample set to form a training set;

[0014] Step 2: Build the Attention-LSTM neural network:

[0015] Step 2.1: Build a 5-layer Attention-LSTM neural network with the following structure: input layer, Attention unit, one LSTM hidden layer, fully connected layer, and output layer. The Attention unit consists of an Attention input layer, an Attention hidden layer, and an Attention output layer connected in series.

[0016] In step 2.2, set the number of neurons in the input layer to 10×8 and the activation function to Sigmoid. Set the number of neurons in the attention input layer of the attention unit to 10×8 and the number of neurons in the attention hidden layer to 10×1. Set the number of neurons in the attention output layer to 10×8. Set the range of the number of neurons in the LSTM hidden layer to [1,200]. Set the number of neurons in the fully connected layer to 10×1 and the activation function to ReLU. Set the number of neurons in the output layer to 1 and the activation function to Sigmoid.

[0017] Step 3: Use genetic algorithm to optimize the hyperparameters of the Attention-LSTM neural network:

[0018] Step 3.1: Encode the network hyperparameters and initialize the population, and set the relevant initial parameters, including crossover probability and mutation probability;

[0019] Step 3.2, calculate the individual fitness function value, where the fitness function is the inverse of the mean square error (MSE) of the training set;

[0020] Step 3.4, obtain offspring through selection, crossover, and mutation processes;

[0021] Step 3.5: Determine whether the currently searched optimal individual fitness value has reached the optimal value, or whether the number of iterations has reached the maximum value. If so, execute step 4 after obtaining the optimized hyperparameters. Otherwise, execute step 3.2.

[0022] Step 4: Train the Attention-LSTM neural network:

[0023] The training set and optimized hyperparameters are input into the Attention-LSTM neural network. The Adam optimization algorithm is used to iteratively update the weights of each layer of the Attention-LSTM neural network until the loss function of the Attention-LSTM neural network converges, thus obtaining a trained Attention-LSTM neural network.

[0024] Step 5: Predict radar threat level:

[0025] Using the same method as step 1, the radar parameters of the radar threat level to be predicted are sampled and normalized, the processed samples are input into the trained Attention-LSTM neural network, and the threat level prediction result is output.

[0026] Compared with the prior art, the present invention has the following advantages:

[0027] First, the training set constructed by the present invention includes radar threat attributes and corresponding threat levels. Threat attributes include range, altitude, azimuth, carrier frequency, pulse width, repetition rate, and phase. This overcomes the shortcomings of existing threat level prediction methods, which construct training sets that only include threat levels or some threat attributes. This method provides a more complete dataset and more objective and reasonable prediction results.

[0028] Second, the present invention uses a genetic algorithm to optimize the Attention-LSTM model to predict the radar threat level. The Attention mechanism can highlight the threat attributes that have a greater impact on the prediction results, and the genetic algorithm can be used to quickly find the optimal parameter combination in the parameter space of the model, overcoming the shortcomings of existing threat level prediction methods that have inaccurate prediction results. This enables the present invention to make more accurate judgments on the radar threat level at future moments and has the advantage of higher prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0030] In order to more clearly illustrate the technical solutions and technical effects of the present invention, the embodiments or technical solutions in the prior art will be further introduced below in combination with the embodiments and the accompanying drawings. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0031] Refer to the attached Figure 1 , the implementation steps of the embodiment of the present invention are further described in detail.

[0032] Step 1: Generate a training set.

[0033] In an embodiment of the present invention, an airborne radar emitter, whose radar threat level is to be predicted, switches between three different operating modes: close-range search mode, long-range search mode, and tracking mode, performing uniformly variable linear motion over a 10-second period. The airborne radar emitter first operates in close-range search mode, maintains this mode for a period of time, then switches to long-range search mode. After maintaining this mode for a period of time, it switches to tracking mode. After maintaining this mode for a period of time, it switches back to long-range mode again. This three-mode switching continues until the total duration reaches 10 seconds.

[0034] The duration of each mode is shown in Table 1:

[0035] Table 1 Summary of duration of three modes

[0036] Working Mode Duration (ms) Proximity Search [3,4,5,6] Long-range search [2,3,4] track [0,1,2]

[0037] The duration of the close search mode is randomly selected from the durations [3, 4, 5, 6] listed in Table 1. The duration of the long search mode is randomly selected from the durations [2, 3, 4] listed in Table 1. The duration of the tracking mode is randomly selected from the durations [0, 1, 2] listed in Table 1. Radar threat attributes in the three operating modes of close search, long search, and tracking all have seven types: range, altitude, azimuth, phase, carrier frequency, pulse width, and repetition rate. The time series of the radar emitter's altitude and range is obtained by performing uniformly variable linear motion on the radar emitter and is independent of the operating mode. Threat attribute values ​​such as azimuth, phase, carrier frequency, pulse width, and repetition rate change with each operating mode. In short-range search mode, the radar's azimuth changes slowly, its phase is incoherent, and the carrier frequency randomly jumps between four fixed values: 5.34 GHz, 5.37 GHz, 5.39 GHz, and 5.41 GHz. The pulse width randomly jumps between 42 μs and 50 μs, and the repetition rate randomly jumps between 0.7 kHz, 0.75 kHz, 0.8 kHz, and 0.88 kHz. In long-range search mode, the radar's azimuth changes instantaneously, its phase is incoherent, the carrier frequency randomly jumps between 5.4 GHz and 5.41 GHz, the pulse width randomly jumps between 20 μs and 27 μs, and the repetition rate randomly jumps between 2.46 kHz and 2.5 kHz, all within a certain range. In tracking mode, the radar's bearing remains constant, its phase is coherent, the carrier frequency randomly jumps between 5.6 GHz and 5.61 GHz, the pulse width randomly jumps between 5 μs and 12 μs, and the repetition rate randomly jumps between 6.41 kHz and 6.67 kHz. As the radar emitter switches between its three operating modes multiple times within 10 seconds, a time series of the radar emitter's operating modes is generated. Based on this operating mode time series, a time series of each threat attribute value can be derived.

[0038] In the embodiment of the present invention, the time-varying sequence of each threat attribute value within the 10-second duration is sampled at a sampling interval of 1 ms to obtain the historical time series corresponding to each threat attribute value, including distance, altitude, azimuth, phase, carrier frequency, pulse width, and repetition rate. The total number of sampling points and the total data length of each threat attribute value time series are both 10,000.

[0039] Using the approximate ideal solution sorting algorithm, the threat attribute values ​​of each identical sampling point in the seven threat attribute time series are comprehensively calculated. The threat degree value of the sampling point is obtained as follows, and a total of 10,000 threat degree values ​​are obtained:

[0040] threat i=T(D i ,H i ,A i ,P i ,F i ,B i ,R i )

[0041] Among them, threat i Indicates the threat level value of the i-th sampling point, i=1,2,....N, N represents the total number of sampling points, D i represents the radar distance of the i-th sampling point, H i represents the radar height of the i-th sampling point, A i represents the radar azimuth of the i-th sampling point, P i represents the radar phase of the i-th sampling point, F i represents the radar carrier frequency of the i-th sampling point, B i represents the radar pulse width of the i-th sampling point, R i represents the radar repetition frequency of the i-th sampling point, and T() represents the operation of approximating the ideal sorting algorithm;

[0042] Each threat attribute and threat level of each sampling point is taken as a sample and expressed as (D i ,H i ,A i ,P i ,F i ,B i ,R i ,threat i 10,000 threat level values ​​are organized into a threat level time series according to the order of sampling point time. The seven radar threat attribute time series and the threat level time series form a sample set.

[0043] According to the following formula, each sample in the sample set is normalized:

[0044]

[0045] Among them, x uv ' represents the vth radar threat attribute of the uth sample in the normalized sample set, v represents the sequence number of the radar threat attribute, v = 1 represents the distance between the radar in the sample set and the friendly target, v = 2 represents the height between the radar in the sample set and the friendly target, v = 3 represents the azimuth of the radar in the sample set, v = 4 represents the carrier frequency of the radar, v = 5 represents the pulse width of the radar in the sample set, v = 6 represents the pulse repetition frequency of the radar in the sample set, v = 7 represents the phase of the radar in the sample set, v = 8 represents the threat level of the radar in the sample set, x uvrepresents the attribute of the vth radar characteristic parameter in the uth sample in the sample set, x vmax represents the maximum value of the vth radar threat attribute in the sample set, x vmin Represents the minimum value of the vth radar threat attribute in the sample set.

[0046] All normalized samples form the training set.

[0047] Step 2: Build the Attention-LSTM neural network.

[0048] Build a 5-layer Attention-LSTM neural network with the following structure: input layer, Attention unit, an LSTM hidden layer, fully connected layer, and output layer. The Attention unit consists of an Attention input layer, an Attention hidden layer, and an Attention output layer connected in series.

[0049] The number of neurons in the input layer of the Attention-LSTM neural network is set to 10×8, and the activation function is Sigmoid. The number of neurons in the Attention input layer of the Attention unit in the Attention-LSTM neural network is set to 10×8, and the number of neurons in the Attention hidden layer is 10×1. The number of neurons in the Attention output layer is set to 10×8. The range of the number of neurons in the LSTM hidden layer of the Attention-LSTM neural network is set to [1,200]. The number of neurons in the fully connected layer of the Attention-LSTM neural network is set to 10×1, and the activation function is ReLU. The number of neurons in the output layer of the Attention-LSTM neural network is set to 1, and the activation function is Sigmoid.

[0050] Step 3: Use genetic algorithm to optimize the hyperparameters of the Attention-LSTM neural network.

[0051] A genetic algorithm is used to optimize the Attention-LSTM neural network's learning rate, number of hidden layer neurons, and dropout rate. Hyperparameter optimization aims to find the hyperparameters that maximize the performance of the Attention-LSTM neural network on the training set. Hyperparameters are different from the model parameters in Step 2 and need to be set before training.

[0052] The genetic algorithm optimizes the Attention-LSTM neural network by searching for optimal parameter combinations within a parameter search space, with the objective function being to minimize prediction error. The steps are as follows: First, determine the parameter search space for the hyperparameters to be optimized. The parameter search space for the learning rate is [0, 1], the parameter search space for hidden layer neurons is [1, 200], and the parameter search space for dropout is [0.1, 0.6]. Second, encode the parameters to be optimized and initialize the population, setting the crossover probability to 0.3 and the mutation probability to 0.6. Third, calculate the individual fitness value using the individual fitness function, which is the inverse of the mean squared error (MSE) of the training set. Fourth, obtain offspring through selection, crossover, and mutation. Fifth, determine whether the currently found optimal individual fitness value is optimal or whether the number of iterations has reached the maximum. If so, execute step 4 after obtaining the optimized hyperparameters. Otherwise, recalculate the fitness value.

[0053] Step 4: Train the Attention-LSTM neural network.

[0054] The training set and the hyperparameters optimized by the genetic algorithm are input into the Attention-LSTM neural network. The Adam optimization algorithm is used to iteratively update the weights of each layer model of the Attention-LSTM neural network until the loss function of the Attention-LSTM neural network converges, thus obtaining a trained Attention-LSTM neural network.

[0055] The loss function of the Attention-LSTM neural network is as follows:

[0056]

[0057] Where Loss represents the loss function of the Attention-LSTM neural network, ∑· represents the summation operation, N represents the total number of radar threat level samples in the training set, and p k represents the true value of the radar threat level of the kth sample in the training data set, lg represents the logarithmic operation with base 10, and q k represents the radar threat level prediction value of the kth sample.

[0058] Step 5: Predict the radar threat level.

[0059] Using the same method as step 1, the radar parameters of the radar threat level to be predicted are sampled and normalized, the processed samples are input into the trained Attention-LSTM neural network, and the threat level prediction result is output.

[0060] In order to evaluate the prediction effect of the present invention, the mean square error (MSE) and mean absolute error (MAE) are selected as evaluation indicators to calculate the error between the true value and the predicted value. The smaller the values ​​of MSE and MAE, the higher the prediction accuracy.

[0061] The calculation formulas of the MSE and MAE are as follows:

[0062]

[0063]

[0064] Where N represents the total number of radar threat level samples in the test set, y k Indicates the true threat level value of the kth sample in the test set, y k =1,2,...,N, represents the predicted threat level value of the kth sample in the test set,

Claims

1. A radar threat level prediction method based on attention mechanism and LSTM, characterized in that: Generate a training set containing radar threat attributes and radar threat levels. Use the training set and genetic algorithm-optimized hyperparameters to train the constructed Attention-LSTM neural network. The steps of this prediction method include the following: Step 1: Generate training set: Step 1.1: Generate a time-varying sequence of each radar threat attribute of the enemy airborne radar emitter in three different operating modes within 10 seconds. Step 1.2: Sample the time-varying sequence of each radar threat attribute at a sampling interval of 1 ms to obtain the time series corresponding to each sampling point of each threat attribute. Step 1.3: Use the approximate ideal solution sorting algorithm to comprehensively calculate the threat attribute values ​​of each identical sampling point in the sequence of all threat attributes changing over time to obtain the threat level value of the sampling point; In step 1.4, each threat attribute at each sampling point and the threat level corresponding to that sampling point are combined into a sample, and a threat level time series is obtained according to the chronological order of the sampling points. The time series of all radar threat attributes and the threat level time series are combined into a sample set. Step 1.5: Normalize each sample in the sample set to form a training set; Step 2: Build the Attention-LSTM neural network: Step 2.1: Build a 5-layer Attention-LSTM neural network. Its structure is connected in series: input layer, Attention unit, an LSTM hidden layer, fully connected layer, and output layer. The Attention unit is composed of the Attention input layer, Attention hidden layer, and Attention output layer connected in series. In step 2.2, set the number of neurons in the input layer of the Attention-LSTM neural network to 10×8, and the activation function to Sigmoid. Set the number of neurons in the Attention input layer of the Attention unit to 10×8, and the number of neurons in the Attention hidden layer to 10×1. Set the number of neurons in the Attention output layer to 10×8. Set the range of the number of neurons in the LSTM hidden layer to [1,200]. Set the number of neurons in the fully connected layer to 10×1, and the activation function to ReLU. Set the number of neurons in the output layer to 1, and the activation function to Sigmoid. Step 3: Use genetic algorithm to optimize the hyperparameters of the Attention-LSTM neural network: Step 3.1: Encode the network hyperparameters and initialize the population, and set the relevant initial parameters, including crossover probability and mutation probability; Step 3.2, calculate the individual fitness function value, where the fitness function is the inverse of the mean square error (MSE) of the training set; Step 3.4, obtain offspring through selection, crossover, and mutation processes; Step 3.5: Determine whether the currently searched optimal individual fitness value has reached the optimal value, or whether the number of iterations has reached the maximum value. If so, execute step 4 after obtaining the optimized hyperparameters. Otherwise, execute step 3.

2. Step 4: Train the Attention-LSTM neural network: The training set and optimized hyperparameters are input into the Attention-LSTM neural network. The Adam optimization algorithm is used to iteratively update the weights of each layer of the Attention-LSTM neural network until the loss function of the Attention-LSTM neural network converges, thus obtaining a trained Attention-LSTM neural network. Step 5: Predict radar threat level: Using the same method as step 1, the radar parameters of the radar threat level to be predicted are sampled and normalized in sequence, and the processed samples are input into the trained Attention-LSTM neural network to output the threat level prediction result.

2. The radar threat level prediction method based on attention mechanism and LSTM according to claim 1 is characterized in that: The radar threat attributes described in step 1.1 include distance, altitude, azimuth, carrier frequency, pulse width, repetition rate, and phase.

3. The radar threat level prediction method based on attention mechanism and LSTM according to claim 1 is characterized in that: The network hyperparameters described in step 3.1 include: learning rate, hidden layer neurons, and forgetting rate dropout.

4. The radar threat level prediction method based on the attention mechanism and LSTM according to claim 1 is characterized in that: The loss function of the Attention-LSTM neural network described in step 4 is as follows: Among them, Loss represents the loss function of the Attention-LSTM neural network, ∑ represents the summation operation, N represents the total number of radar threat level samples in the training set, and p k represents the true value of the radar threat level of the kth sample in the training set, lg represents the logarithmic operation with base 10, and q k Represents the predicted value of the radar threat level of the kth sample in the training set.

Images