UAV reconnaissance and countermeasure method based on deep learning

Abstract

The present invention relates to the technical field of data processing, and specifically discloses a method for detecting and countering unmanned aerial vehicles based on deep learning, comprising the following steps: obtaining a training sample set, the training sample set comprising a measured echo signal or a simulated echo signal; using the training sample set to train a neural network model to obtain a trained neural network model; the loss function of the neural network model is an improved binary cross entropy loss function; the binary cross entropy loss is obtained by weighting the loss values ​​of all samples using each loss weight; when the sample is a measured echo signal, the loss weight is 1; when the sample is a simulated echo signal, the loss weight is; using the trained neural network model to detect the echo signal to be tested, and judging the probability of containing a drone signal; when the probability of containing a drone signal is greater than a threshold, taking countermeasures. The present invention improves the performance of the model, thereby improving the accuracy of the model in identifying drones.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a drone reconnaissance and countermeasure method based on deep learning. Background Art

[0002] Unmanned aerial vehicles (UAVs) are aircraft controlled by remote control or autonomous programs and are used in many fields, such as aerial photography, surveying and mapping, agricultural monitoring, and emergency rescue. However, the development of drone technology has also brought new challenges and threats, including privacy violations and security threats, which has prompted the development of technologies to detect and counter illegal drones with unauthorized or malicious behavior. Drone reconnaissance and countermeasures are aimed at detecting, identifying and preventing threats from illegal drones, protecting important facilities and public safety, and maintaining national security and social stability. They are important directions for the development of modern technology and tactics.

[0003] The identification of drones usually adopts the radar detection method. Radar detection is that the radar system emits electromagnetic waves and uses the principle of electromagnetic wave reflection to analyze the received echo signals to determine whether there are drones. One of the echo signal analysis techniques in the prior art is to train a neural network model by collecting measured echo signals, and then input the echo signals to be detected into the trained neural network model to determine whether there are drones. However, a large number of measured echo signals are required when training the neural network model. Therefore, a large amount of manpower and material resources will be consumed in the process of collecting echo signals. Not only is the cost high, but the collection is also difficult. Especially for some expensive drones, the price may exceed the scientific research budget of researchers. At the same time, the increase in drone models will also lead to higher collection costs. Therefore, in order to reduce costs, the neural network model is often trained by combining simulated echo signals and measured echo signals, which can reduce costs and collection difficulties.

[0004] For example, in the Chinese patent application document with authorization announcement number CN118566873B, a method for rotorcraft target recognition based on micro-Doppler features and deep learning is disclosed, including the following steps: simulating the echo signals of a single-rotor four-blade UAV, a single-rotor three-blade UAV, and a four-rotor UAV respectively; obtaining the echo signals of a single-rotor four-blade UAV, a single-rotor three-blade UAV, and a four-rotor UAV respectively through theoretical calculations; performing time-frequency transformation to obtain a time-frequency graph data set; dividing it into a training set and a test set; constructing and training a ResNet18-SVM neural network based on an attention mechanism; and classifying the time-frequency graphs of the three types of rotorcraft to achieve target recognition of the three types of rotorcraft.

[0005] Each time the neural network model is iterated, the cross entropy loss function is used to calculate the cross entropy loss. The smaller the cross entropy loss, the better the performance of the neural network model and the closer the prediction is to the truth. The existing cross entropy loss is the average of the formula that calculates the difference between the predicted value of all echo signals and their label values. However, since the simulation software cannot completely restore the measured echo signal, there is a difference between the simulated echo signal and the measured echo signal. Some simulated echo signals of poor quality will cause a large cross entropy loss, resulting in poor performance of the trained neural network model. The accuracy of identifying drones using this neural network model is low, and it is impossible to detect illegal drones in time and implement countermeasures. Summary of the invention

[0006] The present invention provides a UAV detection and countermeasure method based on deep learning, aiming to solve the technical problem that the accuracy of the neural network model in identifying UAVs in the prior art is low and illegal UAVs cannot be detected in time to implement countermeasures.

[0007] A deep learning-based drone detection and countermeasure method of the present invention comprises the following steps:

[0008] Acquire a training sample set, wherein the training sample set includes a plurality of samples, and the samples are measured echo signals or simulated echo signals;

[0009] The neural network model is trained using the training sample set to obtain a trained neural network model; the loss function of the neural network model is an improved binary cross entropy loss function; the binary cross entropy loss is obtained by weighting the loss values ​​of all samples using each loss weight; when the sample is a measured echo signal, the loss weight is 1; when the sample is a simulated echo signal, the loss weight is for: ; In the formula, For the The simulated echo signal and the corresponding The similarity of the measured echo signals, is the total number of the corresponding measured echo signals; the similarity With 1 and The simulated echo signal, The difference in the degree of difference of the time-frequency data corresponding to the measured echo signals is positively correlated;

[0010] The trained neural network model is used to detect the echo signal to be tested and determine the probability of containing a drone signal; when the probability of containing a drone signal is greater than the threshold, countermeasures are taken.

[0011] In the above scheme, since different loss weights are assigned to each sample according to the similarity of the time-frequency data of the simulated echo signal and the measured echo signal when calculating the binary cross entropy loss, the focus on samples with poor quality during model training can be reduced, and more attention can be paid to samples with good quality, thereby improving the performance of the model and the accuracy of the model in identifying drones, so that illegal drones can be discovered in time and countermeasures can be implemented.

[0012] Preferably, the degree of difference for:

[0013] ;

[0014] In the formula, For the The time-frequency data of the simulated echo signal The moment The power corresponding to the frequency, For the The time-frequency data of the measured echo signal The moment The power corresponding to the frequency, is the total number of frequencies, is the total number of moments, is the standard normalization function.

[0015] In the above scheme, by comparing the average values ​​of the power corresponding to the simulated echo signal and the measured echo signal at all times and all frequencies, the difference between the two samples can be reflected, and then the similarity of the two samples can be calculated.

[0016] Preferably, the similarity for:

[0017]

[0018] In the formula, For the The simulated echo signal and the The similarity of the signal-to-noise ratios corresponding to the measured echo signals, is the label value of the simulated echo signal, is the label value of the measured echo signal, and a label value of 1 indicates that the simulated echo signal or the measured echo signal contains a drone signal, and a label value of 0 indicates that the simulated echo signal or the measured echo signal does not contain a drone signal.

[0019] In the above scheme, by obtaining the similarity of the signal-to-noise ratio of the simulated echo signal and the measured echo signal, when comparing the similarity of the two samples, the influence of the noise environment on the calculation result is taken into account, so that the calculation result is more accurate.

[0020] Preferably, the similarity for:

[0021] ;

[0022] In the formula, For the The signal-to-noise ratio of a simulated echo signal, For the The signal-to-noise ratio of the measured echo signal, The natural constant The exponential function of base .

[0023] Preferably, the binary cross entropy loss for:

[0024] ;

[0025] In the formula, is the loss weight, is 1 or , For the The loss value of samples is is the total number of samples;

[0026] No. The loss value of samples for:

[0027] ;

[0028] In the formula, For the The label value of a sample, when the label value is 1, it indicates that a drone signal is present, and when the label value is 0, it indicates that no drone signal is present; For the The probability of predicting the sample contains a drone signal.

[0029] Preferably, the time-frequency data corresponding to the simulated echo signal and the measured echo signal are obtained by short-time Fourier transform or continuous wavelet transform.

[0030] Preferably, the neural network model is a ResNet-18 neural network model.

[0031] In the above scheme, the ResNet-18 neural network model overcomes the degradation problem by introducing residual connections. Compared with traditional convolutional neural networks, it can greatly reduce the number of parameters and improve training efficiency, making it suitable for a variety of tasks.

[0032] Preferably, the countermeasures include jamming electronic signals of the UAV and attacking it with laser weapons.

[0033] Preferably, the measured echo signal is collected by a radar system, and the simulated echo signal is generated by electromagnetic simulation software.

[0034] In the above scheme, the simulated echo signal is generated by electromagnetic simulation software, which can significantly reduce the cost, improve the sample collection efficiency and enrich the sample types.

[0035] Preferably, the threshold value ranges from 0.6 to 0.8.

[0036] The beneficial effects are:

[0037] In the scheme of the present invention, the loss weight of each sample is obtained according to the similarity of the time-frequency data of the simulated echo signal and the measured echo signal, and the loss values ​​of all samples are weighted by using the loss weights to obtain the improved binary cross entropy loss, so as to complete the training of the neural network model. Since different loss weights are assigned according to the quality of each sample when calculating the binary cross entropy loss, the focus on samples with poor quality during model training can be reduced, and more attention can be paid to samples with good quality, thereby improving the performance of the model, improving the accuracy of the model in identifying drones, and timely discovering illegal drones and implementing countermeasures. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] By reading the following detailed description with reference to the accompanying drawings, the above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily understood. In the accompanying drawings, several embodiments of the present invention are shown in an exemplary and non-limiting manner, and the same or corresponding reference numerals represent the same or corresponding parts, wherein:

[0039] Figure 1 This is a flowchart of the steps of a deep learning-based drone reconnaissance and countermeasure method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0040] Embodiments of the present invention are described in detail below, and examples of the embodiments are shown in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to be used to explain the present invention, but should not be understood as limiting the present invention.

[0041] like Figure 1 As shown, the present invention provides a UAV reconnaissance and countermeasure method based on deep learning, comprising the following steps:

[0042] S1. Obtain a training sample set, where the training sample set includes a plurality of samples, and the samples are measured echo signals or simulated echo signals.

[0043] The purpose of this step is to obtain a training sample set to train the neural network model based on the training sample set. The measured echo signal is a signal actually collected by a radar system, such as a millimeter wave radar system. The simulated echo signal is a signal generated by electromagnetic simulation software, which can be simulated by modeling according to the measured scene in the electromagnetic simulation software to obtain a simulated echo signal. In this step, a mixed sample of the measured echo signal and the simulated echo signal is used as a training sample set, which not only avoids the problem of high cost caused by only using the measured echo signal, but also enriches the types of samples, so that the training effect of the neural network model is better.

[0044] Specifically, the measured echo signal includes a measured echo signal containing a drone signal and a measured echo signal without a drone signal. The simulated echo signal includes a simulated echo signal containing a drone signal and a simulated echo signal without a drone signal. Since the acquisition cost of the measured echo signal is high, the number of measured echo signals can be set to be less than the number of simulated echo signals. For example, the number of measured echo signals containing drone signals and the number of measured echo signals without drone signals are set to 10. The number of simulated echo signals containing drone signals and the number of simulated echo signals without drone signals are set to 400. Of course, the number of samples can also be adjusted according to actual conditions.

[0045] S2. Use the training sample set to train the neural network model to obtain a trained neural network model.

[0046] In this step, the neural network model uses the ResNet-18 neural network model. The ResNet-18 neural network model overcomes the degradation problem by introducing residual connections. Compared with traditional convolutional neural networks, it can greatly reduce the number of parameters and improve training efficiency, making it suitable for a variety of tasks. Of course, the VGG-16 neural network model can also be used.

[0047] Training a neural network model using a training sample set belongs to the prior art, which generally includes the following steps:

[0048] S21. Initialization parameters: First, the parameters in the neural network model need to be initialized. Usually, the random initialization method can be used to initialize the parameters.

[0049] S22, forward propagation: by inputting each sample of the training sample set into the neural network model, calculating the output of each layer until the final prediction result is obtained. The prediction result of the present invention is the probability that the sample contains a drone.

[0050] S23, calculating the loss: comparing the prediction result of the neural network model with the true label of the sample, calculating the gap between the prediction result and the true label, which is usually represented by a loss function. The loss function used in the present invention is a binary cross entropy loss function.

[0051] S24, Back Propagation: The gradient of the loss function for each parameter is calculated through the back propagation algorithm. The back propagation process propagates the gradient from the output layer to the input layer in order to adjust the parameters of each layer.

[0052] S25, parameter update: Based on the calculated gradient information, use an optimization algorithm, such as the gradient descent algorithm, to update the parameters in the neural network model to minimize the loss function.

[0053] S26, repeated iteration: Repeat the process of forward propagation, loss calculation, back propagation and parameter update until the set stopping condition is reached, such as reaching the maximum number of iterations or the loss function converges.

[0054] S27. Model evaluation: Use a validation set or a test set to evaluate the performance of the trained neural network model in order to adjust parameters or improve the model structure.

[0055] Among them, in step S23 of the prior art, the binary cross entropy loss is the arithmetic mean of the loss values ​​of all samples. The loss value of each sample is the difference between the predicted value of the sample and its label value. The smaller the binary cross entropy loss, the better the performance of the neural network model and the closer the prediction is to the truth. However, since the simulation software cannot completely restore the measured echo signal, there is a difference between the simulated echo signal and the measured echo signal, and some simulated echo signals with poor quality have the same weight as the echo signals with good quality, which will lead to a large binary cross entropy loss, making the performance of the trained neural network model poor, and the accuracy of identifying drones using the neural network model is low, and it is impossible to detect illegal drones in time and implement countermeasures.

[0056] Therefore, in order to prevent the samples of simulated echo signals with poor quality from adversely affecting the performance of the neural network model, it is necessary to reduce the weight of the samples of simulated echo signals with poor quality in the training of the neural network model. In order to distinguish from the weight of samples in the prior art (the weight of samples in the prior art can be considered as 1), the weights assigned to samples in the present invention are called loss weights.

[0057] The core of the present invention is to assign different loss weights according to the quality of the simulated echo signal. Specifically, a simulated echo signal with better quality is assigned a larger loss weight so that it is paid more attention in the model training, and a simulated echo signal with poorer quality is assigned a smaller loss weight so that it is paid less attention in the model training.

[0058] Since training a neural network model belongs to the prior art, the present invention only describes in detail the improved process of calculating the loss in step S23, which includes the following steps:

[0059] S231. Obtain the loss weight of each sample.

[0060] When the sample is a measured echo signal, the sample is considered to be a real signal and is assigned the maximum loss weight. When the sample is a simulated echo signal, the corresponding loss weight is obtained by comparing the difference between the simulated echo signal and the measured echo signal. Specifically, the loss weight of the simulated echo signal is obtained by the following steps:

[0061] S2311. Obtain time-frequency data of the measured echo signal and the simulated echo signal.

[0062] Drones are usually equipped with multiple rotors, which produce unique micro-motions when rotating. This micro-motion causes the echo signal to contain additional time-varying frequency signals. This phenomenon is called micro-Doppler features. By extracting these features, drones can be identified and the recognition accuracy and reliability of drones can be improved. Time-frequency data has a strong ability to characterize micro-Doppler features, so it is necessary to obtain time-frequency data of samples.

[0063] The time-frequency data of the sample can be obtained by short-time Fourier transform. Short-time Fourier transform is a time-frequency analysis method used to analyze non-stationary signals. Its basic principle is to divide the signal into several small segments, each of which is assumed to be stationary, and then perform Fourier transform on each segment to obtain the frequency information of the signal. By moving these small segments and repeating this process, the frequency information of the signal at different time points can be obtained, thereby obtaining time-frequency data. When the time-frequency data is represented on a time-frequency graph, the horizontal axis represents the time, the vertical axis represents the frequency, and the color represents the power. The power is usually expressed as the square of the amplitude. Therefore, the time-frequency graph can easily obtain the power at any time and any frequency.

[0064] In some alternative embodiments, the time-frequency data of the samples may be obtained by continuous wavelet transform.

[0065] S2312. Obtain the similarity of the signal-to-noise ratio between the simulated echo signal containing the UAV signal and the measured echo signal.

[0066] The measured echo signal containing the drone signal also contains noise signals, including wind noise generated by the rotor, electromagnetic interference, motor aging and jamming, and other environmental noises, such as temperature, density, and air pressure. In order to identify the simulated echo signals with poor quality, it is necessary to compare each simulated echo signal with all the measured echo signals. Only when the noise environment is similar, can the interference error caused by noise be avoided by comparing the simulated echo signal containing the drone with the measured echo signal. The simulated echo signal and the measured echo signal without the drone signal are both noise signals and can be directly compared.

[0067] Therefore, it is necessary to first calculate the similarity of the signal-to-noise ratio between the simulated echo signal containing the drone and the measured echo signal. The signal-to-noise ratio is the ratio between the signal and the noise, which directly affects the signal quality and transmission reliability.

[0068] The signal and noise of the simulated echo signal containing the drone signal are generated by electromagnetic simulation software, so its signal-to-noise ratio can be directly obtained by electromagnetic simulation software. The signal of the measured echo signal containing the drone signal is obtained by modeling according to the measured scene in the electromagnetic simulation software, and its noise is the echo signal without the drone in the measurement environment corresponding to the measured echo signal, and then the signal-to-noise ratio of the measured echo signal can be obtained.

[0069] The first The simulated echo signal and the The similarity of the signal-to-noise ratios corresponding to the measured echo signals for:

[0070]

[0071] In the formula, For the The signal-to-noise ratio of a simulated echo signal, For the The signal-to-noise ratio of the measured echo signal, The natural constant The exponential function of base .

[0072] In this step, by obtaining the similarity of the signal-to-noise ratio of the simulated echo signal and the measured echo signal, when comparing the two samples, the influence of the noise environment on the calculation result is taken into account, so that the calculation result is more accurate.

[0073] In some alternative embodiments, the similarity may be a normalized value of the ratio of the corresponding signal-to-noise ratios of the two samples.

[0074] S2313. Obtain the difference between the simulated echo signal and the measured echo signal.

[0075] The degree of difference for:

[0076] ;

[0077] In the formula, For the The time-frequency data of the simulated echo signal The moment The power corresponding to the frequency, For the The time-frequency data of the measured echo signal The moment The power corresponding to the frequency, is the total number of frequencies, is the total number of moments, is the standard normalization function.

[0078] In this step, by comparing the average values ​​of the power corresponding to the simulated echo signal and the measured echo signal at all times and all frequencies, the difference between the two samples can be reflected, and then the similarity of the two samples can be calculated.

[0079] S2314. Obtain the similarity between the simulated echo signal and the measured echo signal.

[0080] The similarity between the simulated echo signal containing the UAV signal and the measured echo signal is positively correlated with the similarity of the corresponding signal-to-noise ratio and negatively correlated with the degree of difference.

[0081] Specifically, similarity for:

[0082]

[0083] In the formula, For the The simulated echo signal and the The similarity of the signal-to-noise ratios corresponding to the measured echo signals, is the label value of the simulated echo signal, is the label value of the measured echo signal, and a label value of 1 indicates that the simulated echo signal or the measured echo signal contains a drone signal, and a label value of 0 indicates that the simulated echo signal or the measured echo signal does not contain a drone signal.

[0084] S2315. Obtain the loss weight of each sample.

[0085] When the sample is a measured echo signal, since it is actually collected by the radar system, its loss weight can be set to the maximum, and the maximum value is 1.

[0086] When the sample is a simulated echo signal, the loss weight for:

[0087] ;

[0088] In the formula, For the The simulated echo signal and the corresponding The similarity of the measured echo signals, is the total number of corresponding measured echo signals.

[0089] Through the above steps, taking the measured echo signal as a benchmark, by calculating the similarity between the simulated echo signal and the measured echo signal, the loss weight corresponding to the simulated echo signal can be obtained, and then different attention can be given to different samples during model training, which can improve the training performance of the model.

[0090] S232. Obtain the binary cross entropy loss of the improved binary cross entropy loss function.

[0091] The binary cross entropy loss is obtained by weighting the loss values ​​of all samples using each loss weight. Among them, the binary cross entropy loss is the weighted average of the loss values ​​of all samples calculated by the improved binary cross entropy loss function.

[0092] Specifically, the binary cross entropy loss for:

[0093] ;

[0094] In the formula, is the loss weight, is 1 or , For the The loss value of samples is is the total number of samples;

[0095] No. The loss value of samples for:

[0096] ;

[0097] In the formula, For the The label value of each sample. When the label value is 1, it means that there is a drone signal. When the label value is 0, it means that there is no drone signal. For the The probability of predicting the sample contains a drone signal.

[0098] S3. Use the trained neural network model to detect the echo signal to be tested and determine the probability of containing a drone signal; when the probability of containing a drone signal is greater than a threshold, take countermeasures.

[0099] In this step, the threshold value ranges from 0.6 to 0.8, and preferably, the threshold value is 0.7. When the probability of containing a drone signal is greater than the threshold, the drone can be countered by electronic signal interference and laser weapon attack.

[0100] In the deep learning-based drone detection and countermeasure method of the present invention, the loss weight of each sample is obtained according to the similarity of the time-frequency data of the simulated echo signal and the measured echo signal, and the loss values ​​of all samples are weighted by each loss weight to obtain an improved binary cross entropy loss to complete the training of the neural network model. Since different loss weights are assigned according to the quality of each sample when calculating the binary cross entropy loss, the focus on samples with poor quality during model training can be reduced, and more attention can be paid to samples with good quality, thereby improving the performance of the model, improving the accuracy of the model in identifying drones, and timely discovering illegal drones and implementing countermeasures.

[0101] In the description of this specification, “plurality” means at least two, such as two, three or more, etc., unless otherwise clearly and specifically defined.

[0102] Although this specification has shown and described a number of embodiments of the present invention, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Those skilled in the art will conceive of many modifications, changes and alternatives without departing from the ideas and spirit of the present invention. It should be understood that in the practice of the present invention, various alternatives to the embodiments of the present invention described herein may be employed.

Claims

1. A UAV reconnaissance and countermeasure method based on deep learning, characterized in that: The steps include: Acquire a training sample set, wherein the training sample set includes a plurality of samples, and the samples are measured echo signals or simulated echo signals; The neural network model is trained using the training sample set to obtain a trained neural network model; the loss function of the neural network model is an improved binary cross entropy loss function; the binary cross entropy loss is obtained by weighting the loss values ​​of all samples using each loss weight; when the sample is a measured echo signal, the loss weight is 1; when the sample is a simulated echo signal, the loss weight is for: ; In the formula, For the The simulated echo signal and the corresponding The similarity of the measured echo signals, is the total number of the corresponding measured echo signals; the similarity With 1 and The simulated echo signal, The difference in the degree of difference of the time-frequency data corresponding to the measured echo signals is positively correlated; Use the trained neural network model to detect the echo signal to be tested and determine the probability of containing drone signals; when the probability of containing drone signals is greater than the threshold, take countermeasures; The degree of difference for: ; In the formula, For the The time-frequency data of the simulated echo signal The moment The power corresponding to the frequency, For the The time-frequency data of the measured echo signal The moment The power corresponding to the frequency, is the total number of frequencies, is the total number of moments, is the standard normalization function; The similarity for: In the formula, For the The simulated echo signal and the The similarity of the signal-to-noise ratios corresponding to the measured echo signals, is the label value of the simulated echo signal, is the label value of the measured echo signal, and a label value of 1 indicates that the simulated echo signal or the measured echo signal contains a drone signal, and a label value of 0 indicates that the simulated echo signal or the measured echo signal does not contain a drone signal.

2. The deep learning-based drone detection and countermeasure method according to claim 1, characterized in that: The similarity for: ； In the formula, For the The signal-to-noise ratio of a simulated echo signal, For the The signal-to-noise ratio of the measured echo signal, The natural constant The exponential function of base .

3. The deep learning-based drone detection and countermeasure method according to claim 1, characterized in that: The binary cross entropy loss for: ； In the formula, is the loss weight, is 1 or , For the The loss value of samples is is the total number of samples; No. The loss value of samples for: ； In the formula, For the The label value of a sample, when the label value is 1, it indicates that a drone signal is present, and when the label value is 0, it indicates that no drone signal is present; For the The probability of predicting the sample contains a drone signal.

4. The method for detecting and countering unmanned aerial vehicles based on deep learning according to claim 1, characterized in that: The time-frequency data corresponding to the simulated echo signal and the measured echo signal are obtained through short-time Fourier transform or continuous wavelet transform.

5. The deep learning-based drone detection and countermeasure method according to claim 1, characterized in that: The neural network model is a ResNet-18 neural network model.

6. The deep learning-based drone detection and countermeasure method according to claim 1, characterized in that: The countermeasures include jamming electronic signals against drones and using laser weapons to attack them.

7. The deep learning-based drone detection and countermeasure method according to claim 1, characterized in that: The measured echo signal is collected by a radar system, and the simulated echo signal is generated by electromagnetic simulation software.

8. The deep learning-based drone detection and countermeasure method according to claim 1, characterized in that: The threshold value ranges from 0.6 to 0.8.

Images