Evaluation Method for Anti-Interference Effect of Convolutional Neural Network Optimized by Sparrow Algorithm

Abstract

The present invention discloses a method for evaluating the anti-interference effect of optimizing a convolutional neural network using a sparrow algorithm, aiming to solve the problem that the weights of evaluation indicators and the hyperparameters of the convolutional neural network need to be set manually. The implementation steps of the present invention include: using the entropy weight method to assign weight values to the anti-interference evaluation indicators; generating a training set; constructing a convolutional neural network; using the sparrow algorithm to optimize the hyperparameters of the convolutional neural network; training the convolutional neural network; and evaluating the anti-interference effect. The present invention has the advantages that the accuracy of the anti-interference effect evaluation can be improved without manually setting the weights of the evaluation indicators and the hyperparameters of the convolutional neural network.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar communication, and further relates to a method for real-time evaluation of anti-jamming effect based on Sparrow Search Algorithm-Optimized Convolutional Neural Network (SSA-CNN) in the field of electronic countermeasures. The present invention can be used to evaluate the effectiveness of radar against jamming signals. Background Art

[0002] Radar countermeasure is an important part of electronic countermeasure, and the anti-jamming effect of radar is one of the important indicators of radar systems. However, the existing evaluations of anti-jamming effects are all for the anti-jamming analysis of a specific interference environment and a specific radar. Most radar engineers use traditional methods to evaluate the anti-deception jamming and anti-suppression jamming effects of radar from a qualitative perspective, rather than from a quantitative perspective to comprehensively describe. However, in actual scenarios, the interference environment is constantly changing, and radar adopts different anti-jamming means according to different interference environments. Both are dynamic change processes. With the complex and changeable electromagnetic interference environment, radar needs to have a certain anti-jamming ability to better play its effectiveness.

[0003] University of Electronic Science and Technology proposed a method for evaluating the anti-jamming effect of radar in its patent document "A Method for Real-Time Evaluation of Radar Anti-Jamming Efficiency" (Application No.: 201410453974; Publication No.: CN 104239712 A). The implementation scheme of this method is as follows: First, according to the changes in the technical and tactical indicators of the radar in the interference environment, select the initial indicators to establish an anti-jamming evaluation index system. Second, reduce the initial index system through the rough set attribute reduction method to establish the final evaluation index. Third, through the Delphi algorithm, obtain the group decision matrix, and then use the chi-square least squares method to rank the decision matrix to construct the final comprehensive weight vector. Use the interference pattern recognition technology based on decision tree to identify the interference pattern in the real-time environment and evaluate the anti-jamming effect of the radar. This method can obtain the performance loss of the radar in the real-time interference environment. However, the still existing deficiency of this method is that the weight is assigned by human factors, and the evaluation result of the radar anti-jamming effect is easily affected subjectively.

[0004] He Xiaofeng et al. disclosed a method for evaluating the anti-jamming effect of radar using a neural network model in their published paper "Research on Indexes of Radar Comprehensive Anti-jamming Effectiveness Evaluation Method" (Radar & ECM, 2020, 40(03): 11-15). The implementation method of this method is as follows: First step, select evaluation indexes and establish an anti-jamming evaluation index system including radar power change, accuracy improvement, false target rejection rate, true target recognition rate (TTR), as well as track quality and anti-jamming measures of anti-jamming measures. Second step, according to different radar performances and different degrees of interference, establish an evaluation set and distinguish anti-jamming evaluation values according to "excellent, good, general, poor, bad". Third step, input the index change detection data under different environments into the neural network. Fourth step, train the neural network and correct the weights and thresholds of neurons through iterative training. Fifth step, output the anti-jamming effect evaluation value of the BP neural network. This method only seeks the help of experts when obtaining samples. This method can evaluate the anti-jamming effect in real time. Once the neural network is trained, no other external conditions are required, reducing the influence of subjective factors and random factors. However, this method requires artificial setting of network parameters. Summary of the Invention

[0005] The object of the present invention is to provide an online evaluation method for anti-jamming effect of optimizing a convolutional neural network using a sparrow algorithm in view of the deficiencies of the above-mentioned existing technologies, aiming to solve the problems of artificial weighting of anti-jamming evaluation indexes in the existing technologies, which is greatly affected by subjectivity, and the need for artificial setting of neural network parameters in the existing technologies.

[0006] The idea of realizing the object of the present invention is as follows: When generating a training set, the present invention constructs interference patterns, relative positions of radars and jammers, frequency domain overlap degree, time domain overlap degree, signal-to-interference ratio, time-bandwidth product, carrier frequency volatility, pulse repetition frequency volatility, interference suppression degree, direction of suppression interference, number of anti-suppression interferences, remaining false alarm rate, direction of deception interference, number of anti-deception interferences, discovery time of deception interference, and interference environment indexes. These index values all change dynamically with the interference environment. The entropy weight method is used to weight the anti-jamming evaluation indexes. The entropy weight method makes a horizontal comparison of the multiple measurement results of each index, and the degree of dispersion of the results is used as the entropy value. With different measurement results each time, the index weights are not fixed, but change dynamically with the reconnaissance parameters, solving the characteristic of the need for artificial weighting in the existing technologies. The present invention uses the sparrow algorithm to optimize the model of the existing convolutional neural network, can automatically set the selection of network structure and number of layers, internal parameters, and select the optimal hyperparameters that minimize the error rate of the validation set. When evaluating the anti-jamming effect, inputting the test data and the optimal hyperparameters into the convolutional neural network can quickly extract feature parameters and obtain the corresponding anti-jamming evaluation value, solving the characteristic of the need for artificial setting of neural network parameters in the existing technologies.

[0007] The specific steps to achieve the object of the present invention are as follows:

[0008] Step 1: Use the entropy weight method to assign weight values to each anti-interference evaluation index in the sample set:

[0009] Step 1.1: Select 16 anti-interference evaluation indexes to form a sample, and select m samples to form a sample set

[0010] Step 1.2: Perform normalization operation on the samples;

[0011] Step 1.3: Calculate the entropy value of each anti-interference evaluation index after normalization:

[0012]

[0013] Among them, E j is the entropy value of the j-th index after normalizing the sample, ln represents the logarithmic operation with base e, n represents the number of anti-interference evaluation indexes, ∑ represents the summation operation, m represents the number of samples, and x ij ' represents the j-th index of the i-th sample after normalization;

[0014] Step 1.4: Calculate the weight of each anti-interference evaluation index after normalization:

[0015]

[0016] Among them, β j is the weight of the j-th index after normalizing the sample, E j is the entropy value of the j-th index after normalizing the sample, ∑ represents the summation operation, and n represents the number of anti-interference evaluation indexes;

[0017] Step 2: Generate a training set:

[0018] Step 2.1: Select m samples to form a sample set, where m≥6000. Use the Technique for Order Preference by Similarity to Ideal Solution (TOPSIS) algorithm and combine the weights obtained by the entropy weight method to calculate the anti-interference effect evaluation value of each sample;

[0019] Step 2.2: Combine all the normalized samples and their corresponding anti-interference effect evaluation values to form a training set;

[0020] Step 3: Construct a convolutional neural network:

[0021] Construct a 7-layer convolutional neural network, and its structure is connected in series in turn: the first convolutional layer, the first pooling layer, the second convolutional layer, the second pooling layer, the third convolutional layer, the third pooling layer, and the fully connected layer;

[0022] Set the number of convolution kernels of the first to third convolutional layers to 4, 8, and 16 in sequence, and set the size of the convolution kernels to 1×3; the first to third pooling layers all adopt the maximum pooling method, the size of the pooling kernels is set to 1×2, and the pooling stride is set to 1×2;

[0023] Step 4, use the sparrow algorithm to optimize the hyperparameters of the convolutional neural network:

[0024] Step 4.1, set the training parameters and hyperparameters;

[0025] Step 4.2, calculate the fitness function value of each sparrow at the current iteration and sort them to find the best and worst fitness, where the fitness function is the reciprocal of the mean squared error (MSE) of the training set;

[0026] Step 4.3, use the following formula to update the position of the discoverers:

[0027]

[0028] where, represents the position of the h-th discoverer in the g-th dimension after the (t + 1)-th update, n = 1, 2,..., d, d represents the dimension of the parameters to be optimized, represents the position of the h-th discoverer in the g-th dimension at the t-th iteration; exp represents the exponential operation with the natural constant e as the base, α represents a decimal randomly selected in the interval (0, 1], K represents the maximum number of iterations, R2 represents the early warning value, R2 ∈ [0, 1], ST represents the safety value, ST ∈ [0.5, 1]; Q represents a random number obeying the standard normal distribution, and L represents a matrix of size 1×d with all elements being 1;

[0029] Step 4.4, use the following formula to update the position of the followers:

[0030]

[0031] where, represents the position of the x-th follower in the y-th dimension after the (t + 1)-th update, y = 1, 2,..., d, d represents the dimension of the parameters to be optimized, Q represents a random number obeying the standard normal distribution, exp represents the exponential operation with the natural constant e as the base, represents the globally worst position at the t-th iteration, represents the position of the x-th follower in the y-th dimension at the t-th iteration, and i represents the total number of sparrows; represents the best position of the discoverers at the (t + 1)-th iteration, ∑ represents the summation operation, rand{-1, 1} represents generating a random number of -1 or 1, and | | represents the modulo operation;

[0032] Step 4.5, randomly select 10% of the individuals in the sparrow population as vigilant ones, and use the following formula to update the positions of the selected vigilant ones:

[0033]

[0034] where, represents the position of the v-th dimension of the u-th vigilant one after the (t + 1)-th update, n = 1, 2,..., d, and d represents the dimension of the parameters to be optimized, represents the global optimal position after the t-th iteration; β represents a random number obeying the standard normal distribution; | | represents the modulo operation; represents the position of the v-th dimension of the u-th vigilant one after the t-th iteration, represents the global worst position after the t-th iteration, M represents the moving direction of the vigilant one, M ∈ [-1, 1], f u represents the fitness of the current vigilant one, f w represents the current global worst fitness, ε represents a decimal close to 0 to avoid meaninglessness, f g represents the current global optimal fitness;

[0035] Step 4.6, judge whether the fitness value reaches the minimum. If so, terminate the iteration to obtain the optimal hyperparameters. Otherwise, execute Step 4.2;

[0036] Step 5, train the convolutional neural network:

[0037] Input the training set and the optimal hyperparameters into the convolutional neural network, and use the backpropagation gradient descent method to iteratively update the parameters of each layer of the convolutional neural network until the loss function of the convolutional neural network converges, so as to obtain the trained convolutional neural network;

[0038] Step 6, evaluate the anti-interference effect:

[0039] Adopt the same method as in Step 1.2 to normalize each sample and evaluation value for evaluating the anti-interference effect; input the normalized samples and evaluation values into the optimized convolutional neural network, and output the anti-interference evaluation result of the sample.

[0040] Compared with the prior art, the present invention has the following advantages:

[0041] First, the present invention adopts the entropy weight method to assign weights to each anti-interference evaluation index of each sample in the training set, overcomes the deficiency of the prior art that requires artificial setting of index weights, and enables the present invention to have the advantage of objectively evaluating the radar anti-interference effect in real time.

[0042] Second, the present invention optimizes the hyperparameters of the convolutional neural network using the sparrow algorithm, and inputs the training set and hyperparameters into the convolutional neural network, so that when training the CNN model, the error rate of the validation set reaches the lowest. It overcomes the deficiency of the prior art that requires artificial setting of neural network parameters, and improves the accuracy of real-time evaluation of the anti-interference effect of the present invention. Description of the Drawings

[0043] Figure 1 It is the flowchart of the present invention. Detailed Embodiment

[0044] Next, in combination with Figure 1 ..., the specific implementation steps of the embodiments of the present invention will be further described in detail.

[0045] Step 1, using the entropy weight method, assign weight values to all anti-interference evaluation indicators in the sample set.

[0046] Step 1.1, each sample in the embodiments of the present invention selects 16 anti-interference evaluation indicators. The 16 anti-interference effect indicators include interference pattern, relative position of radar and jammer, frequency domain overlap, time domain overlap, signal-to-interference ratio, time-bandwidth product, carrier frequency fluctuation, pulse repetition frequency fluctuation, interference suppression ratio, direction of suppression interference, number of anti-suppression interference, remaining false alarm rate, direction of deception interference, number of anti-deception interference, deception interference discovery time, and interference environment. A total of 6000 samples are selected in the embodiments of the present invention to form a sample set.

[0047] Step 1.2, perform the following normalization operation on each sample in the sample set:

[0048]

[0049] where x ij ' represents the j-th index of the i-th sample in the normalized sample set, j represents the serial number of the anti-interference index, corresponding to 16 indexes respectively. j = 1 represents the interference pattern, j = 2 represents the relative position of the radar and the jammer, j = 3 represents the frequency domain overlap, j = 4 represents the time domain overlap, j = 5 represents the signal-to-interference ratio, j = 6 represents the time-bandwidth product, j = 7 represents the carrier frequency fluctuation, j = 8 represents the pulse repetition frequency fluctuation, j = 9 represents the interference suppression ratio, j = 10 represents the direction of suppression interference, j = 11 represents the number of anti-suppression interference, j = 12 represents the remaining false alarm rate, j = 13 represents the direction of deception interference, j = 14 represents the number of anti-deception interference, j = 15 represents the deception interference discovery time, j = 16 represents the interference environment. x ij represents the j-th index of the i-th sample point in the sample set, x jmin represents the minimum value of the j-th index in the sample set, x jmax represents the maximum value of the j-th index in the sample set.

[0050] Step 1.3, calculate the entropy value of each anti-interference evaluation index after normalization:

[0051]

[0052] Among them, E j represents the entropy value of the j-th index after normalizing the samples, ln represents the logarithmic operation with base e, n represents the number of anti-interference evaluation indexes, ∑ represents the summation operation, m represents the number of samples, and x ij ' represents the j-th index of the i-th sample after normalization.

[0053] Step 1.4, calculate the weight of each anti-interference evaluation index after normalization:

[0054]

[0055] Among them, β j represents the weight of the j-th index after normalizing the samples, E j represents the entropy value of the j-th index after normalizing the samples, ∑ represents the summation operation, and n represents the number of anti-interference evaluation indexes.

[0056] Step 2, generate the training set.

[0057] Step 2.1, according to the weights of each anti-interference evaluation index obtained by the entropy weight method, use the Technique for Order Preference by Similarity to an Ideal Solution (TOPSIS) to evaluate the anti-interference effect of each sample, and obtain the evaluation value corresponding to the sample. Among them, each sample contains the weight values of 16 anti-interference evaluation indexes:

[0058]

[0059] Among them, y i represents the anti-interference effect evaluation value of each sample, represents the square root operation, ∑ represents the summation operation, min represents the minimum value, β j represents the weight of the j-th index after normalizing the samples, x ij ' represents the j-th index of the i-th sample after normalization, x mj represents the j-th index of the m-th sample after normalization, n represents the number of anti-interference evaluation indexes, and min represents the minimum value.

[0060] Step 2.2, form the training set by combining all the normalized samples and their corresponding anti-interference effect evaluation values.

[0061] Step 3, construct a convolutional neural network:

[0062] Construct a 7-layer convolutional neural network, and its structure is connected in series in turn as follows: the first convolutional layer, the first pooling layer, the second convolutional layer, the second pooling layer, the third convolutional layer, the third pooling layer, and the fully connected layer;

[0063] Set the number of convolutional kernels of the first to third convolutional layers to 4, 8, and 16 in turn, and set the size of the convolutional kernels to 1×3; the first to third pooling layers all adopt the maximum pooling method, the size of the pooling kernels is set to 1×2, and the pooling stride is set to 1×2;

[0064] Step 4, use the sparrow algorithm to optimize the hyperparameters of the convolutional neural network:

[0065] Step 4.1, set the training parameters including: the number of sparrow populations is 10, the maximum number of iterations is 20, the ratio of discoverers to followers is 7:3, and the hyperparameters include: the range of the learning rate is 0.0005 - 0.01, the range of the batch size is 16 - 100, the range of the kernel size is 1 - 20, and the range of the number of kernels is 1 - 20.

[0066] Step 4.2, calculate the fitness function value of each sparrow at the current iteration and sort it to find the best and worst fitness. Among them, the fitness function is the reciprocal of the mean squared error (MSE) of the training set.

[0067] Step 4.3, use the following formula to update the position of the discoverer:

[0068]

[0069] Among them, represents the position of the h-th discoverer in the g-th dimension after the (t + 1)-th update, n = 1, 2,..., d, where d represents the dimension of the parameter to be optimized, represents the position of the h-th discoverer in the n-th dimension at the t-th iteration; exp represents the exponential operation with the natural constant e as the base, α represents a decimal randomly selected in the interval (0, 1], K represents the maximum number of iterations, R2 represents the warning value, R2 ∈ [0, 1], ST represents the safety value, ST ∈ [0.5, 1]; Q represents a random number obeying the standard normal distribution, and L represents a matrix of size 1×d with all elements being 1.

[0070] Step 4.4, use the following formula to update the position of the follower:

[0071]

[0072] Among them, denotes the position of the $y$-th dimension of the $x$-th follower after the $(t + 1)$-th update, where $y = 1, 2, \ldots, d$, $d$ represents the dimension of the parameters to be optimized, $Q$ represents a random number following the standard normal distribution, and exp represents the exponential operation with the natural constant $e$ as the base. denotes the globally worst position at the $t$-th iteration. denotes the position of the $y$-th dimension of the $x$-th follower at the $t$-th iteration, and $i$ represents the total number of sparrows. denotes the best position of the discoverer at the $(t + 1)$-th iteration, $\sum$ represents the summation operation, rand{-1,1} represents generating a random number of -1 or 1, and $| |$ represents the modulo operation.

[0073] Step 4.5, randomly select 10% of the individuals in the sparrow population as the guardians, and use the following formula to update the positions of the selected guardians:

[0074]

[0075] where denotes the position of the $v$-th dimension of the $u$-th guardian after the $(t + 1)$-th update, $n = 1, 2, \ldots, d$, and $d$ represents the dimension of the parameters to be optimized. denotes the globally optimal position after the $t$-th iteration; $\beta$ represents a random number following the standard normal distribution; $| |$ represents the modulo operation; denotes the position of the $v$-th dimension of the $u$-th guardian after the $t$-th iteration. denotes the globally worst position after the $t$-th iteration, $M$ represents the moving direction of the guardian, $M \in [-1, 1]$, and $f$ u represents the fitness of the current guardian, and $f$ w represents the current globally worst fitness, $\varepsilon$ represents a small number close to 0 to avoid meaninglessness, and $f$ g represents the current globally optimal fitness.

[0076] Step 4.6, determine whether the fitness value reaches the minimum. If so, terminate the iteration to obtain the optimal hyperparameters; otherwise, execute Step 4.2.

[0077] Step 5, train the convolutional neural network:

[0078] Input the training set and the optimal hyperparameters into the convolutional neural network, and use the backpropagation gradient descent method to iteratively update the parameters of each layer of the convolutional neural network until the loss function of the convolutional neural network converges, obtaining the trained convolutional neural network.

[0079] The loss function is as follows:

[0080]

[0081] Among them, Loss represents the loss function of the convolutional neural network, ∑ represents the summation operation, N represents the total number of anti-interference evaluation samples in the training set, k represents the k-th sample, and p k represents the true value of the anti-interference effect of the k-th sample in the training set, lg represents the logarithmic operation with base 10, and q k represents the predicted value of the anti-interference effect of the k-th sample in the training set.

[0082] Step 6, evaluate the anti-interference effect:

[0083] Adopt the same method as in Step 1.2 to normalize each sample and evaluation value of the evaluated anti-interference effect; input the normalized samples and evaluation values into the optimized convolutional neural network, and output the anti-interference evaluation result of the sample.

[0084] The following further illustrates the effect of the present invention in combination with simulation experiments:

[0085] 1. Simulation experiment conditions:

[0086] The software platform for the simulation experiment of the present invention is: Windows 10 operating system and matlab2021b.

[0087] The parameter settings of the training set and test set used in the simulation experiment of the present invention are as follows: The training set of the present invention is the training set generated in Steps 1 and 2 of the specific embodiments of the present invention, which contains 6000 samples and corresponding anti-interference evaluation values. Then, a test set is generated in the same way, and this test set contains 2000 samples and corresponding anti-interference effect evaluation values, and the test set is used to evaluate the anti-interference effect evaluation result of the present invention.

[0088] In the simulation experiment of the present invention, the training parameters of the sparrow algorithm are set as follows: the population number is 10, the maximum number of iterations is 20, the ratio of discoverers to followers is 7:3, the range of the learning rate is 0.0005 - 0.01, the range of the batch size is 16 - 100, the range of the kernel size is 1 - 20, and the range of the number of kernels is 1 - 20.

[0089] In the simulation experiment of the present invention, the parameters of the convolutional neural network are set as follows: the learning rate is 0.001, the training batch size is 30, and the total number of training epochs is 300.

[0090] 2. Simulation content and its result analysis:

[0091] The simulation experiment of the present invention uses the method of the present invention and the convolutional neural network method of the prior art to evaluate the anti-interference effect on the test set respectively. After normalizing the 2,000 samples in the test set, the method of the present invention uses the sparrow algorithm for optimization, and the optimal hyperparameters obtained are: the batch size is 22, the learning rate is 0.01, the number of convolutional kernels in the first to third convolutional layers are 10, 8, and 19 respectively, and the sizes of the convolutional kernels are 1×1, 1×2, and 1×3 respectively. The normalized samples in the test set and the optimal hyperparameters are input into the convolutional neural network to obtain the anti-interference effect evaluation value of the present invention. The convolutional neural network method directly trains the normalized samples in the test set to obtain the anti-interference effect evaluation value of this method. The anti-interference evaluation results of the convolutional neural network optimized by the sparrow algorithm and the anti-interference effect evaluation results of the convolutional neural network not optimized by the sparrow algorithm are recorded respectively.

[0092] In the simulation experiment, the convolutional neural network CNN method of the prior art adopted refers to the convolutional neural network regression prediction method proposed by Ma Qiaoyu in his published paper "Prediction of shear wave velocity based on one-dimensional convolutional neural network" (Lithologic Reservoirs, 2021, 33(04): 111-120).

[0093] To illustrate the effect of the present invention, the determination coefficient, mean square error, mean absolute error, and mean absolute percentage error of the method of the present invention and the convolutional neural network method of the prior art are calculated respectively:

[0094]

[0095]

[0096]

[0097]

[0098] Among them, R 2 represents the absolute coefficient, ∑ represents the summation operation, i represents the i-th sample, l represents the number of samples, y i is the true value of the sample, is the predicted value of the sample, represents the average value of the samples, MSE represents the mean square error, MAE represents the mean absolute error, MAPE represents the mean absolute percentage error, and finally the anti-interference effect prediction results are shown in Table 1.

[0099] Table 1 Comparison table of anti-interference effect evaluation between CNN and the method of the present invention

[0100]

[0101] As can be seen from Table 1, the determination coefficient R of the method of the present invention2 The determination coefficient is 0.9835, the mean square error MSE is 1.2907, the mean absolute error MAE is 0.8485, the mean absolute percentage error MAPE is 23.8242%. The determination coefficient of the CNN without SSA optimization is 0.9474, the mean square error is 9.6186, the mean absolute error MAE is 2.4255, and the mean absolute percentage error MAPE is 69.5308%. The absolute coefficient of the CNN optimized by SSA is higher than that of the non-optimized CNN, and the mean square error, mean absolute error, and mean absolute percentage error are all reduced, which proves that the present invention improves the accuracy of real-time evaluation of anti-interference effect.

Claims

1. An anti-interference effect evaluation method for optimizing a convolutional neural network using a sparrow algorithm, characterized in that, The entropy weight method is used to assign weights to the anti-interference evaluation indicators, generate a training set containing anti-interference evaluation indicators, and optimize a convolutional neural network with an evaluation function using the sparrow algorithm; the specific steps of this evaluation method are as follows: Step 1, use the entropy weight method to assign weight values to each anti-interference evaluation indicator in the sample set: Step 1.1, select 16 anti-interference evaluation indicators to form a sample, and select m samples to form a sample set Step 1.2, perform a normalization operation on the sample; Step 1.3, calculate the entropy value of each anti-interference evaluation indicator after normalization: Among them, E j is the entropy value of the j-th index after normalizing the samples. ln represents the logarithmic operation with base e, n represents the number of anti-interference evaluation indicators, ∑ represents the summation operation, m represents the number of samples, and x ij ' represents the j-th index of the i-th sample after normalization; Step 1.4, calculate the weight of each anti-interference evaluation indicator after normalization: Among them, β j is the weight of the j-th index after normalizing the sample, and E j is the entropy value of the j-th index after normalizing the sample. ∑ represents the summation operation, and n represents the number of anti-interference evaluation indicators; Step 2, generate a training set: Step 2.1, select m samples to form a sample set, where m≥6000, use the Technique for Order Preference by Similarity to Ideal Solution (TOPSIS) algorithm, combined with the weights obtained by the entropy weight method, to calculate the anti-interference effect evaluation value of each sample; Step 2.2, form a training set with all the normalized samples and their corresponding anti-interference effect evaluation values; Step 3, construct a convolutional neural network: Construct a 7-layer convolutional neural network, and its structure is connected in series in turn: the first convolutional layer, the first pooling layer, the second convolutional layer, the second pooling layer, the third convolutional layer, the third pooling layer, and the fully connected layer; Set the number of convolutional kernels in the first to third convolutional layers to 4, 8, and 16 in turn, and set the size of the convolutional kernels to 1×3; the first to third pooling layers all use the maximum pooling method, the size of the pooling kernels is set to 1×2, and the pooling stride is set to 1×2; Step 4, use the sparrow algorithm to optimize the hyperparameters of the convolutional neural network: Step 5, train the convolutional neural network: Input the training set and the optimal hyperparameters into the convolutional neural network, and use the backpropagation gradient descent method to iteratively update the parameters of each layer of the convolutional neural network until the loss function of the convolutional neural network converges, and obtain the trained convolutional neural network; Step 6, evaluate the anti-interference effect: Use the same method as in Step 1.2 to normalize each sample and evaluation value for evaluating the anti-interference effect; input the normalized samples and evaluation values into the optimized convolutional neural network, and output the anti-interference evaluation result of the sample.

2. The anti-interference effect evaluation method for optimizing a convolutional neural network using a sparrow algorithm according to claim 1, characterized in that The anti-interference effect indicators described in Step 1.1 include interference patterns, relative positions of the radar and the jammer, frequency domain overlap, time domain overlap, signal-to-interference ratio, time-bandwidth product, carrier frequency volatility, pulse repetition frequency volatility, interference suppression ratio, direction of barrage jamming, number of anti-barrage jamming, residual false alarm rate, direction of deception jamming, number of anti-deception jamming, discovery time of deception jamming, and interference environment.

3. The anti-interference effect evaluation method for optimizing a convolutional neural network using a sparrow algorithm according to claim 1, wherein The normalization operation described in Step 1.2 is realized by the following formula: Among them, x ij ' represents the jth index of the i-th sample in the normalized sample set, j represents the serial number of the anti-interference index, corresponding to 16 indicators, j = 1 represents the interference pattern, j = 2 represents the relative position of the radar and the jammer, j = 3 represents the frequency domain overlap, j = 4 represents the time domain overlap, j = 5 represents the signal-to-interference ratio, j = 6 represents the time-bandwidth product, j = 7 represents the carrier frequency fluctuation, j = 8 represents the repetition frequency fluctuation, j = 9 represents the interference suppression, j = 10 represents the suppression interference direction, j = 11 represents the number of anti-suppression interference, j = 12 represents the residual false alarm rate, j = 13 represents the deceptive interference direction, j = 14 represents the number of anti-deceptive interference, j = 15 represents the deceptive interference detection time, j = 16 represents the interference environment, x ij represents the jth index in the i-th sample point in the sample set, x jmin represents the minimum value of the jth indicator in the sample set, x jmax Indicates the maximum value of the j-th indicator in the sample set.

4. The anti-interference effect evaluation method for optimizing a convolutional neural network using the sparrow algorithm according to claim 1, characterized in that: The steps of using the sparrow algorithm to optimize the hyperparameters of the convolutional neural network described in Step 4 are as follows: The first step is to set the training parameters and hyperparameters; The second step is to calculate the fitness function value of each sparrow at the current iteration and sort them to find the best and worst fitness, where the fitness function is the reciprocal of the Mean Squared Error (MSE) of the training set; The third step is to update the position of the discoverer using the following formula: Among them, represents the position of the g-th dimension of the h-th discoverer after the (t + 1)-th update, where n = 1, 2, ..., d, and d represents the dimension of the parameter to be optimized. represents the position of the g-th dimension of the h-th discoverer at the t-th iteration; exp represents the exponential operation with the natural constant e as the base, α represents a decimal randomly selected within the interval (0, 1], K represents the maximum number of iterations, R2 represents the warning value, where R2 ∈ [0, 1], ST represents the safety value, where ST ∈ [0.5, 1]; Q represents a random number following the standard normal distribution, and L represents a matrix of size 1 × d with all elements being 1. Step 4: Update the follower's position using the following formula: Among them, represents the position of the y-th dimension of the x-th follower after the (t + 1)-th update, where y = 1, 2,..., d, d represents the dimension of the parameter to be optimized, Q represents a random number following the standard normal distribution, and exp represents the exponential operation with the natural constant e as the base. represents the global worst position at the t-th iteration. represents the position of the y-th dimension of the x-th follower at the t-th iteration, and i represents the total number of sparrows. represents the best position of the discoverer at the (t + 1)-th iteration, ∑ represents the summation operation, rand{-1, 1} represents generating a random number of -1 or 1, and || represents the modulo operation. Step 5: Randomly select 10% of the individuals in the sparrow population as sentinels, and update the positions of the selected sentinels using the following formula: Among them, represents the position of the $u$-th sentinel at the $(t + 1)$-th update in the $v$-th dimension, where $n = 1, 2, \cdots, d$, and $d$ represents the dimension of the parameters to be optimized; represents the global optimal position after the $t$-th iteration; $\beta$ represents a random number following the standard normal distribution; represents the position of the $u$-th sentinel in the $v$-th dimension after the $t$-th iteration, represents the global worst position after the $t$-th iteration, $M$ represents the moving direction of the sentinel, $M\in[-1, 1]$, and $f$ u represents the fitness of the current sentinel, and $f$ w represents the current global worst fitness, $\varepsilon$ represents a small number close to 0 to avoid meaninglessness, and $f$ g represents the current global optimal fitness; Step 6: Determine whether the fitness value has reached the minimum. If so, terminate the iteration to obtain the optimal hyperparameters; otherwise, execute Step 2.

5. The anti-interference effect evaluation method for optimizing a convolutional neural network using the sparrow algorithm according to claim 4 is characterized in that: The training parameters and hyperparameters refer to: the training parameters include: the number of sparrow populations is 10, the maximum number of iterations is 20, and the ratio of discoverers to followers is 7:3; the hyperparameters include: the range of the learning rate is 0.0005 - 0.01, the range of the batch size is 16 - 100, the range of the kernel size is 1 - 20, the range of the number of kernels is 1 - 20, and the range of the number of iterations is determined according to the convergence degree of the convolutional neural network loss function.

6. The anti-interference effect evaluation method for optimizing a convolutional neural network using a sparrow algorithm according to claim 1, characterized in that: The loss function described in Step 5 is as follows: Among them, Loss represents the loss function of the convolutional neural network, ∑ represents the summation operation, N represents the total number of anti-interference evaluation samples in the training set, k represents the k-th sample, and p k represents the true value of the anti-interference effect of the k-th sample in the training set, lg represents the logarithmic operation with base 10, and q k represents the predicted value of the anti-interference effect of the k-th sample in the training set.

Images