A method and system for predicting low-altitude invasion behavior of unmanned aerial vehicles based on deep learning

Abstract

The present invention discloses a method and system for predicting the low-altitude intrusion behavior of unmanned aerial vehicles based on deep learning, and belongs to the field of radar detection technology. The present invention first uses a radar to transmit and receive radar reflection signals in the monitoring area, then extracts the pulse morphology distortion value and the pulse width distortion value of the signal in each time period and constructs a corresponding sequence, and simultaneously performs time-frequency transformation to calculate the reflection frequency difference to construct a reflection frequency difference sequence; then calculates the development value of each sequence respectively; finally, inputs each sequence into a prediction network, and obtains the low-altitude intrusion risk value of the unmanned aerial vehicle based on the development value compensation. The present invention solves the problem of low accuracy in detecting low-altitude intrusion behavior of unmanned aerial vehicles in the prior art.

Description

Technical Field

[0001] The present invention relates to the field of radar detection technology, and in particular to a method and system for predicting low-altitude intrusion behavior of unmanned aerial vehicles based on deep learning. Background Art

[0002] With the rapid development of drone technology, low-flying drones are increasingly used in civil and commercial fields. However, this also brings potential safety risks, especially low-altitude intrusions that may occur in cities and sensitive areas. Existing monitoring technologies mainly rely on traditional radar systems to detect the presence of drones by transmitting signals and receiving reflected waves. The radar system transmits short pulse signals and receives their reflected waves, and then calculates the round-trip time of the signal to determine the target location.

[0003] However, low-altitude drone intrusions are characterized by rapid movement, and the low-altitude environment is extremely complex with a large number of other objects. These factors cause radar reflection waves to face serious interference. Buildings, trees, and the ground will all produce complex signal echoes, making it difficult for traditional radar systems to accurately detect low-altitude drone intrusions. Summary of the invention

[0004] In view of the above-mentioned deficiencies in the prior art, the present invention provides a method and system for predicting low-altitude intrusion behavior of unmanned aerial vehicles based on deep learning, which solves the problem of low accuracy in detecting low-altitude intrusion behavior of unmanned aerial vehicles in the prior art.

[0005] In order to achieve the above-mentioned invention object, the technical solution adopted by the present invention is: a method for predicting low-altitude invasion behavior of unmanned aerial vehicles based on deep learning, comprising the following steps:

[0006] S1. Use radar to transmit radar signals to the monitoring area and receive radar reflected signals;

[0007] S2. For the radar reflection signal received in each time period, extract the pulse shape distortion value and the pulse width distortion value, and construct a pulse shape distortion sequence and a pulse width distortion sequence respectively;

[0008] S3, performing time-frequency transformation on the radar reflection signal received in each time period, calculating the reflection frequency difference, and constructing a reflection frequency difference sequence;

[0009] S4, calculating the morphological distortion development value for the pulse morphological distortion sequence, calculating the width distortion development value for the pulse width distortion sequence, and calculating the reflection frequency difference development value for the reflection frequency difference sequence;

[0010] S5. Input the pulse morphology distortion sequence, pulse width distortion sequence and reflection frequency difference sequence into the UAV low-altitude intrusion behavior prediction network, and obtain the UAV low-altitude intrusion risk value based on the compensation of the morphology distortion development value, width distortion development value and reflection frequency difference development value.

[0011] Further, S2 includes the following sub-steps:

[0012] S21, in the radar reflection signal received in each time period, marking a signal segment higher than the amplitude threshold as a pulse;

[0013] S22, extracting a local maximum value and a local minimum value in a pulse, and calculating a pulse morphology distortion value;

[0014] S23, calculating a pulse width distortion value according to the width of each pulse;

[0015] S24, arranging the pulse morphology distortion values ​​of each time period in chronological order to construct a pulse morphology distortion sequence;

[0016] S25. Arrange the pulse width distortion values ​​of each time period in chronological order to construct a pulse width distortion sequence.

[0017] Furthermore, the formula for calculating the pulse morphology distortion value in S22 is: , where r shape is the pulse shape distortion value, A max,k,j is the jth local maximum value in the kth pulse of the radar reflection signal received in a time period, A min,k,j is the jth local minimum value of the kth pulse of the radar reflection signal received in a time period, K max,k is the number of local maxima in the kth pulse of the radar reflection signal received in a time period, K min,k is the number of local minima in the kth pulse of the radar reflection signal received in a time period, k and j are positive integers, and K is the number of pulses in the radar reflection signal received in a time period.

[0018] Furthermore, the formula for calculating the pulse width distortion value in S23 is: , where r width is the pulse width distortion value, T k is the width of the kth pulse in the radar reflection signal received in a time period, τ is the transmitted pulse width, | | is the absolute value operation, K is the number of pulses in the radar reflection signal received in a time period, and k is a positive integer.

[0019] Furthermore, S3 includes the following sub-steps:

[0020] S31, performing time-frequency transformation on the radar reflection signal received in each time period to obtain multiple amplitudes and frequencies;

[0021] S32, selecting the frequency corresponding to the maximum amplitude as the reflection frequency;

[0022] S33, subtracting the transmission frequency from the reflection frequency to obtain a reflection frequency difference;

[0023] S34. Arrange the reflection frequency differences in each time period in chronological order to construct a reflection frequency difference sequence.

[0024] Furthermore, the formula for calculating the morphological distortion development value in S4 is: , where γ shape is the morphological distortion development value, r shape,t is the pulse shape distortion value at the tth moment in the pulse shape distortion sequence, r shape,t-1 is the pulse shape distortion value at the t-1th moment in the pulse shape distortion sequence, N shape is the number of pulse shape distortion values ​​in the pulse shape distortion sequence, and t is the time number;

[0025] The formula for calculating the width distortion development value in S4 is: , where γ width is the width distortion development value, r width,t is the pulse width distortion value at the tth moment in the pulse width distortion sequence, r width,t-1 is the pulse width distortion value at the t-1th moment in the pulse width distortion sequence, N width is the number of pulse width distortion values ​​in the pulse width distortion sequence;

[0026] The formula for calculating the reflection frequency difference development value in S4 is: , where γ fs is the reflection frequency difference development value, r fs,t is the reflection frequency difference at the tth moment in the reflection frequency difference sequence, r fs,t-1 is the reflection frequency difference at the t-1th moment in the reflection frequency difference sequence, N fs is the number of reflection frequency differences in the reflection frequency difference sequence.

[0027] Furthermore, the UAV low-altitude intrusion behavior prediction network in S5 includes: a first LSTM unit, a second LSTM unit, a third LSTM unit, a first feature enhancement unit, a second feature enhancement unit, a third feature enhancement unit, a first fully connected layer, a second fully connected layer, a third fully connected layer, a first compensation unit Z1, a second compensation unit Z2, a third compensation unit Z3 and a weighted layer;

[0028] The input end of the first LSTM unit is used to input a pulse shape distortion sequence; the input end of the second LSTM unit is used to input a pulse width distortion sequence; the input end of the third LSTM unit is used to input a reflection frequency difference sequence;

[0029] The input end of the first feature enhancement unit is connected to the output end of the first LSTM unit, and its output end is connected to the input end of the first fully connected layer; the input end of the second feature enhancement unit is connected to the output end of the second LSTM unit, and its output end is connected to the input end of the second fully connected layer; the input end of the third feature enhancement unit is connected to the output end of the third LSTM unit, and its output end is connected to the input end of the third fully connected layer; the first input end of the first compensation unit Z1 is connected to the output end of the first fully connected layer, and its second input end is used to input the morphological distortion development value; the first input end of the second compensation unit Z2 is connected to the output end of the second fully connected layer, and its second input end is used to input the width distortion development value; the first input end of the third compensation unit Z3 is connected to the output end of the third fully connected layer, and its second input end is used to input the reflection frequency difference development value; the input end of the weighted layer is respectively connected to the output end of the first compensation unit Z1, the output end of the second compensation unit Z2 and the output end of the third compensation unit Z3, and its output end serves as the output end of the UAV low-altitude invasion behavior prediction network.

[0030] Furthermore, the expressions of the first feature enhancement unit, the second feature enhancement unit and the third feature enhancement unit are all: , where H is the output feature vector of the first feature enhancement unit, the second feature enhancement unit or the third feature enhancement unit, X is the input feature vector of the first feature enhancement unit, the second feature enhancement unit or the third feature enhancement unit, and MP one is a one-dimensional maximum pooling layer, is element-wise multiplication.

[0031] Furthermore, the expression of the first compensation unit Z1 is: ,in, is the first risk component after compensation, ζ shape is the first risk component before compensation, e is a natural constant, γ shape is the morphological distortion development value;

[0032] The expression of the second compensation unit Z2 is: ,in, is the second risk component after compensation, ζ width is the second risk component before compensation, γ width is the width distortion development value;

[0033] The expression of the third compensation unit Z3 is: ,in, is the third risk component after compensation, ζ fs is the third risk component before compensation, γ fs is the reflection frequency difference development value.

[0034] A UAV low-altitude intrusion behavior prediction system based on deep learning, including: a transmitting and receiving subsystem, a distortion sequence construction subsystem, a frequency difference sequence construction subsystem, a development value calculation subsystem and a prediction subsystem;

[0035] The transmitting and receiving subsystem is used to transmit radar signals to the monitoring area and receive radar reflected signals;

[0036] The distortion sequence construction subsystem is used to extract the pulse shape distortion value and the pulse width distortion value of the radar reflection signal received in each time period, and respectively construct the pulse shape distortion sequence and the pulse width distortion sequence;

[0037] The frequency difference sequence construction subsystem is used to perform time-frequency transformation on the radar reflection signal received in each time period, calculate the reflection frequency difference, and construct the reflection frequency difference sequence;

[0038] The development value calculation subsystem is used to calculate the morphology distortion development value for the pulse morphology distortion sequence, calculate the width distortion development value for the pulse width distortion sequence, and calculate the reflection frequency difference development value for the reflection frequency difference sequence;

[0039] The prediction subsystem is used to input the pulse morphology distortion sequence, pulse width distortion sequence and reflection frequency difference sequence into the UAV low-altitude intrusion behavior prediction network, and obtain the UAV low-altitude intrusion risk value based on the compensation of the morphology distortion development value, width distortion development value and reflection frequency difference development value.

[0040] The beneficial effects of the present invention are:

[0041] 1. The present invention extracts the pulse morphology distortion value and pulse width distortion value in the radar reflection signal and constructs the corresponding sequence, which can deeply explore the reflection situation of the UAV to the radar signal. Compared with the simple signal analysis of traditional radar, the capture of the target characteristics of the UAV is more accurate, and the interference of the reflected signals of other objects in the complex low-altitude environment is effectively reduced, thereby improving the accuracy of detection.

[0042] 2. The present invention performs time-frequency transformation on the radar reflection signal, calculates the reflection frequency difference, reflects the frequency disturbance caused by the moving UAV, constructs a reflection frequency difference sequence, and highlights the frequency difference information generated when the UAV moves rapidly at low altitude, thereby making up for the problem of insufficient feature extraction of traditional methods in complex dynamic environments.

[0043] 3. The present invention calculates the development value of morphological distortion, width distortion and reflection frequency difference, and compensates for the features in the UAV low-altitude intrusion behavior prediction network based on these development values. It can dynamically reflect the changing trend of the UAV flight status, making the prediction network's prediction of the UAV low-altitude intrusion behavior more forward-looking and accurate, and effectively overcoming the problem of low detection accuracy caused by the inability to accurately grasp the changing trend of UAV flight in the prior art.

[0044] 4. The present invention calculates the morphological distortion development value, the width distortion development value and the reflection frequency difference development value, which reflects the change of the reflection signal over time, further eliminates the influence of signal echoes generated by buildings, trees, ground, etc., and improves the accuracy of detection of low-altitude intrusion behavior of drones. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is a flowchart of a method for predicting low-altitude intrusion behavior of drones based on deep learning;

[0046] Figure 2 Schematic diagram of the structure of the drone low-altitude intrusion behavior prediction network. DETAILED DESCRIPTION

[0047] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0048] Embodiment 1, as Figure 1 As shown, a method for predicting low-altitude intrusion behavior of unmanned aerial vehicles based on deep learning includes the following steps:

[0049] S1. Use radar to transmit radar signals to the monitoring area and receive radar reflected signals;

[0050] S2. For the radar reflection signal received in each time period, extract the pulse shape distortion value and the pulse width distortion value, and construct a pulse shape distortion sequence and a pulse width distortion sequence respectively;

[0051] S3, performing time-frequency transformation on the radar reflection signal received in each time period, calculating the reflection frequency difference, and constructing a reflection frequency difference sequence;

[0052] S4, calculating the morphological distortion development value for the pulse morphological distortion sequence, calculating the width distortion development value for the pulse width distortion sequence, and calculating the reflection frequency difference development value for the reflection frequency difference sequence;

[0053] S5. Input the pulse morphology distortion sequence, pulse width distortion sequence and reflection frequency difference sequence into the UAV low-altitude intrusion behavior prediction network, and obtain the UAV low-altitude intrusion risk value based on the compensation of the morphology distortion development value, width distortion development value and reflection frequency difference development value.

[0054] In this embodiment, the transmitted radar signal is: , where y(t) is the transmitted radar signal, A is the amplitude, sin is the sinusoidal signal, f is the transmission frequency, t is the time, τ is the transmitted pulse width, rect(t / τ) is the pulse shaping function, and φ is the phase.

[0055] In this embodiment, the transmission frequency is 24 GHz.

[0056] In this embodiment, S2 includes the following sub-steps:

[0057] S21, in the radar reflection signal received in each time period, marking a signal segment higher than the amplitude threshold as a pulse;

[0058] S22, extracting a local maximum value and a local minimum value in a pulse, and calculating a pulse morphology distortion value;

[0059] S23, calculating a pulse width distortion value according to the width of each pulse;

[0060] S24, arranging the pulse morphology distortion values ​​of each time period in chronological order to construct a pulse morphology distortion sequence;

[0061] S25. Arrange the pulse width distortion values ​​of each time period in chronological order to construct a pulse width distortion sequence.

[0062] In a pulse signal segment, if the amplitude of a signal point is greater than the amplitudes of its adjacent left and right signal points (that is, the amplitude is the largest in its local neighborhood), the amplitude of the signal point is called a local maximum.

[0063] In a pulse signal segment, if the amplitude of a signal point is smaller than the amplitudes of its adjacent left and right signal points (that is, the amplitude is the smallest in its local neighborhood), the amplitude of the signal point is called a local minimum.

[0064] In this embodiment, the amplitude threshold is set to the amplitude mean of the radar reflection signal.

[0065] The size and structure of the drone will cause the echo pulse width to change. The time it takes for radar waves reflected from different parts of the drone to reach the radar is different, which will cause the echo pulse width to be widened; and the drone has multiple strong reflection points (such as the junction of the wing and the fuselage), which will cause the echo pulse to be distorted with multiple peaks. Therefore, the present invention marks the signal segment above the amplitude threshold as a pulse, extracts the local maximum and local minimum in each pulse to reflect the drone's reflection of the radar signal, and then calculates the pulse width distortion value to reflect the drone's widening of the pulse.

[0066] In this embodiment, the formula for calculating the pulse morphology distortion value in S22 is: , where rshape is the pulse shape distortion value, A max,k,j is the jth local maximum value in the kth pulse of the radar reflection signal received in a time period, A min,k,j is the jth local minimum value of the kth pulse of the radar reflection signal received in a time period, K max,k is the number of local maxima in the kth pulse of the radar reflection signal received in a time period, K min,k is the number of local minima in the kth pulse of the radar reflection signal received in a time period, k and j are positive integers, and K is the number of pulses in the radar reflection signal received in a time period.

[0067] The present invention calculates the ratio of the local maximum value to the local minimum value, which reflects the difference between the peak and the valley in the pulse, and then combines the number of the local maximum value and the local minimum value to reflect the degree of distortion of the pulse.

[0068] In this embodiment, the formula for calculating the pulse width distortion value in S23 is: , where r width is the pulse width distortion value, T k is the width of the kth pulse in the radar reflection signal received in a time period, τ is the transmitted pulse width, | | is the absolute value operation, K is the number of pulses in the radar reflection signal received in a time period, and k is a positive integer.

[0069] The present invention calculates the difference between the width of the pulse in the reflected signal and the width of the emitted pulse, reflecting the widening of the pulse.

[0070] In this embodiment, S3 includes the following sub-steps:

[0071] S31, performing time-frequency transformation on the radar reflection signal received in each time period to obtain multiple amplitudes and frequencies;

[0072] S32, selecting the frequency corresponding to the maximum amplitude as the reflection frequency;

[0073] S33, subtracting the transmission frequency from the reflection frequency to obtain a reflection frequency difference;

[0074] S34. Arrange the reflection frequency differences in each time period in chronological order to construct a reflection frequency difference sequence.

[0075] In this embodiment, a radar signal of a single frequency is transmitted. Therefore, after the present invention performs time-frequency transformation on the radar reflection signal, only the frequency corresponding to the maximum amplitude is selected, and the reflection frequency difference is obtained by subtracting the transmission frequency from the reflection frequency. This can effectively highlight the signal characteristics of the target drone among many signals, distinguish it from the background interference signal, and greatly improve the recognition of the target signal.

[0076] The present invention calculates the reflection frequency difference and can reflect the change of the motion state of the UAV in real time and accurately. Whether it is acceleration, deceleration or turning, it can be reflected through the change of frequency difference.

[0077] In this embodiment, the formula for calculating the morphological distortion development value in S4 is: , where γ shape is the morphological distortion development value, r shape,t is the pulse shape distortion value at the tth moment in the pulse shape distortion sequence, r shape,t-1 is the pulse shape distortion value at the t-1th moment in the pulse shape distortion sequence, N shape is the number of pulse shape distortion values ​​in the pulse shape distortion sequence, and t is the time number;

[0078] The formula for calculating the width distortion development value in S4 is: , where γ width is the width distortion development value, r width,t is the pulse width distortion value at the tth moment in the pulse width distortion sequence, r width,t-1 is the pulse width distortion value at the t-1th moment in the pulse width distortion sequence, N width is the number of pulse width distortion values ​​in the pulse width distortion sequence;

[0079] The formula for calculating the reflection frequency difference development value in S4 is: , where γ fs is the reflection frequency difference development value, r fs,t is the reflection frequency difference at the tth moment in the reflection frequency difference sequence, r fs,t-1 is the reflection frequency difference at the t-1th moment in the reflection frequency difference sequence, N fs is the number of reflection frequency differences in the reflection frequency difference sequence.

[0080] The present invention calculates the morphological distortion development value, the width distortion development value and the reflection frequency difference development value to reflect the deterioration of the pulse morphological distortion value, the pulse width distortion value and the reflection frequency difference, thereby characterizing whether the UAV is approaching or moving away.

[0081] like Figure 2 As shown, the UAV low-altitude invasion behavior prediction network in S5 includes: a first LSTM unit, a second LSTM unit, a third LSTM unit, a first feature enhancement unit, a second feature enhancement unit, a third feature enhancement unit, a first fully connected layer, a second fully connected layer, a third fully connected layer, a first compensation unit Z1, a second compensation unit Z2, a third compensation unit Z3 and a weighted layer;

[0082] The input end of the first LSTM unit is used to input a pulse shape distortion sequence; the input end of the second LSTM unit is used to input a pulse width distortion sequence; the input end of the third LSTM unit is used to input a reflection frequency difference sequence;

[0083] The input end of the first feature enhancement unit is connected to the output end of the first LSTM unit, and its output end is connected to the input end of the first fully connected layer; the input end of the second feature enhancement unit is connected to the output end of the second LSTM unit, and its output end is connected to the input end of the second fully connected layer; the input end of the third feature enhancement unit is connected to the output end of the third LSTM unit, and its output end is connected to the input end of the third fully connected layer; the first input end of the first compensation unit Z1 is connected to the output end of the first fully connected layer, and its second input end is used to input the morphological distortion development value; the first input end of the second compensation unit Z2 is connected to the output end of the second fully connected layer, and its second input end is used to input the width distortion development value; the first input end of the third compensation unit Z3 is connected to the output end of the third fully connected layer, and its second input end is used to input the reflection frequency difference development value; the input end of the weighted layer is respectively connected to the output end of the first compensation unit Z1, the output end of the second compensation unit Z2 and the output end of the third compensation unit Z3, and its output end serves as the output end of the UAV low-altitude invasion behavior prediction network.

[0084] In this embodiment, the expressions of the first feature enhancement unit, the second feature enhancement unit and the third feature enhancement unit are all: , where H is the output feature vector of the first feature enhancement unit, the second feature enhancement unit or the third feature enhancement unit, X is the input feature vector of the first feature enhancement unit, the second feature enhancement unit or the third feature enhancement unit, and MP one is a one-dimensional maximum pooling layer, is element-wise multiplication.

[0085] The window size of the max pooling layer is set to 3.

[0086] The present invention uses LSTM units (the first, second, and third LSTM units) to process pulse morphology distortion sequences, pulse width distortion sequences, and reflection frequency difference sequences respectively, which can effectively mine the time series characteristics and long-term dependencies in these sequences, adapt to the dynamic changes of UAV low-altitude intrusion behavior in the time dimension, and provide rich time series information for subsequent analysis.

[0087] The present invention can further strengthen the features of each sequence through the feature enhancement unit (the first, second, and third feature enhancement units), highlight key information, and improve the expressiveness and discrimination of the features, which helps to more accurately identify the feature patterns related to the low-altitude intrusion of drones and improve the accuracy of prediction.

[0088] The present invention calculates the first risk component through the first fully connected layer, calculates the second risk component through the second fully connected layer, and calculates the third risk component through the third fully connected layer. Based on the compensation of morphological distortion development values, width distortion development values ​​and reflection frequency difference development values, the comprehensive utilization of multi-source feature information is realized, and multiple aspects of the UAV flight status are fully considered, so that the model can capture signs of low-altitude intrusion of UAVs from multiple angles, enhance the robustness and generalization ability of the model, and improve the reliability of the prediction of the risk value of low-altitude intrusion of UAVs.

[0089] In this embodiment, the expression of the first compensation unit Z1 is: ,in, is the first risk component after compensation, ζ shape is the first risk component before compensation, e is a natural constant, γ shape is the morphological distortion development value;

[0090] The expression of the second compensation unit Z2 is: ,in, is the second risk component after compensation, ζ width is the second risk component before compensation, γ width is the width distortion development value;

[0091] The expression of the third compensation unit Z3 is: ,in, is the third risk component after compensation, ζ fs is the third risk component before compensation, γ fs is the reflection frequency difference development value.

[0092] The present invention can dynamically adjust the corresponding risk components according to the morphological distortion development value, the width distortion development value and the reflection frequency difference development value through the expression of the compensation unit. This enables the risk assessment to reflect the changing trend of the UAV flight status in real time. When these development values ​​change, the compensated risk components change accordingly, which improves the adaptability and accuracy of the risk assessment to the dynamic behavior of the UAV. The larger the development value, the more obvious the amplification effect on the risk component, which helps to more significantly reflect the possible changes in the intrusion risk of the UAV in the risk assessment, making the system more sensitive to situations with a higher intrusion risk trend and improving the early warning capability of potential intrusion behaviors.

[0093] The weighted layer is used to weight the compensated first risk component, the compensated second risk component and the compensated third risk component to obtain the UAV low-altitude intrusion risk value.

[0094] Embodiment 2, a UAV low-altitude intrusion behavior prediction system based on deep learning, comprising: a transmitting and receiving subsystem, a distortion sequence construction subsystem, a frequency difference sequence construction subsystem, a development value calculation subsystem and a prediction subsystem;

[0095] The transmitting and receiving subsystem is used to transmit radar signals to the monitoring area and receive radar reflected signals;

[0096] The distortion sequence construction subsystem is used to extract the pulse shape distortion value and the pulse width distortion value of the radar reflection signal received in each time period, and respectively construct the pulse shape distortion sequence and the pulse width distortion sequence;

[0097] The frequency difference sequence construction subsystem is used to perform time-frequency transformation on the radar reflection signal received in each time period, calculate the reflection frequency difference, and construct the reflection frequency difference sequence;

[0098] The development value calculation subsystem is used to calculate the morphology distortion development value for the pulse morphology distortion sequence, calculate the width distortion development value for the pulse width distortion sequence, and calculate the reflection frequency difference development value for the reflection frequency difference sequence;

[0099] The prediction subsystem is used to input the pulse morphology distortion sequence, pulse width distortion sequence and reflection frequency difference sequence into the UAV low-altitude intrusion behavior prediction network, and obtain the UAV low-altitude intrusion risk value based on the compensation of the morphology distortion development value, width distortion development value and reflection frequency difference development value.

[0100] The specific implementation process of Example 2 is the same as that of Example 1.

[0101] The present invention extracts the pulse morphology distortion value and pulse width distortion value in the radar reflection signal and constructs the corresponding sequence, so as to deeply explore the reflection situation of the UAV to the radar signal. Compared with the simple signal analysis of traditional radar, the present invention can capture the target characteristics of the UAV more accurately, effectively reduce the interference of the reflected signals of other objects in the complex low-altitude environment, and improve the accuracy of detection.

[0102] The present invention performs time-frequency transformation on the radar reflection signal, calculates the reflection frequency difference, reflects the frequency disturbance caused by the moving UAV, constructs a reflection frequency difference sequence, and highlights the frequency difference information generated when the UAV moves rapidly at low altitude, thereby making up for the problem of insufficient feature extraction of traditional methods in complex dynamic environments.

[0103] The present invention calculates the development value of morphological distortion, width distortion and reflection frequency difference, and compensates for the characteristics in the UAV low-altitude intrusion behavior prediction network based on these development values. It can dynamically reflect the changing trend of the UAV flight status, making the prediction network's prediction of the UAV low-altitude intrusion behavior more forward-looking and accurate, and effectively overcoming the problem of low detection accuracy in the prior art due to the inability to accurately grasp the changing trend of the UAV flight.

[0104] The present invention calculates the morphological distortion development value, the width distortion development value and the reflection frequency difference development value, reflects the change of the reflection signal over time, further eliminates the influence of signal echoes generated by buildings, trees, ground, etc., and improves the accuracy of detection of low-altitude intrusion behavior of unmanned aerial vehicles.

[0105] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for predicting low-altitude intrusion behavior of unmanned aerial vehicles based on deep learning, characterized in that: The following steps are involved: S1. Use radar to transmit radar signals to the monitoring area and receive radar reflected signals; S2. For the radar reflection signal received in each time period, extract the pulse shape distortion value and the pulse width distortion value, and construct a pulse shape distortion sequence and a pulse width distortion sequence respectively; S3, performing time-frequency transformation on the radar reflection signal received in each time period, calculating the reflection frequency difference, and constructing a reflection frequency difference sequence; S4, calculating the morphological distortion development value for the pulse morphological distortion sequence, calculating the width distortion development value for the pulse width distortion sequence, and calculating the reflection frequency difference development value for the reflection frequency difference sequence; S5. Input the pulse morphology distortion sequence, pulse width distortion sequence and reflection frequency difference sequence into the UAV low-altitude intrusion behavior prediction network, and obtain the UAV low-altitude intrusion risk value based on the compensation of the morphology distortion development value, width distortion development value and reflection frequency difference development value.

2. The method for predicting low-altitude intrusion behavior of unmanned aerial vehicles based on deep learning according to claim 1 is characterized in that: The S2 comprises the following sub-steps: S21, in the radar reflection signal received in each time period, marking a signal segment higher than the amplitude threshold as a pulse; S22, extracting a local maximum value and a local minimum value in a pulse, and calculating a pulse morphology distortion value; S23, calculating a pulse width distortion value according to the width of each pulse; S24, arranging the pulse morphology distortion values ​​of each time period in chronological order to construct a pulse morphology distortion sequence; S25. Arrange the pulse width distortion values ​​of each time period in chronological order to construct a pulse width distortion sequence.

3. The method for predicting low-altitude intrusion behavior of unmanned aerial vehicles based on deep learning according to claim 2 is characterized in that: The formula for calculating the pulse morphology distortion value in S22 is: , where r shape is the pulse shape distortion value, A max,k,j is the jth local maximum value in the kth pulse of the radar reflection signal received in a time period, A min,k,j is the jth local minimum value of the kth pulse of the radar reflection signal received in a time period, K max,k is the number of local maxima in the kth pulse of the radar reflection signal received in a time period, K min,k is the number of local minima in the kth pulse of the radar reflection signal received in a time period, k and j are positive integers, and K is the number of pulses in the radar reflection signal received in a time period.

4. The method for predicting low-altitude intrusion behavior of unmanned aerial vehicles based on deep learning according to claim 2 is characterized in that: The formula for calculating the pulse width distortion value in S23 is: , where r width is the pulse width distortion value, T k is the width of the kth pulse in the radar reflection signal received in a time period, τ is the transmitted pulse width, | | is the absolute value operation, K is the number of pulses in the radar reflection signal received in a time period, and k is a positive integer.

5. The method for predicting low-altitude intrusion behavior of unmanned aerial vehicles based on deep learning according to claim 1 is characterized in that: The S3 comprises the following sub-steps: S31, performing time-frequency transformation on the radar reflection signal received in each time period to obtain multiple amplitudes and frequencies; S32, selecting the frequency corresponding to the maximum amplitude as the reflection frequency; S33, subtracting the transmission frequency from the reflection frequency to obtain a reflection frequency difference; S34. Arrange the reflection frequency differences in each time period in chronological order to construct a reflection frequency difference sequence.

6. The method for predicting low-altitude intrusion behavior of unmanned aerial vehicles based on deep learning according to claim 1 is characterized in that: The formula for calculating the morphological distortion development value in S4 is: , where γ shape is the morphological distortion development value, r shape,t is the pulse shape distortion value at the tth moment in the pulse shape distortion sequence, r shape,t-1 is the pulse shape distortion value at the t-1th moment in the pulse shape distortion sequence, N shape is the number of pulse shape distortion values ​​in the pulse shape distortion sequence, and t is the time number; The formula for calculating the width distortion development value in S4 is: , where γ width is the width distortion development value, r width,t is the pulse width distortion value at the tth moment in the pulse width distortion sequence, r width,t-1 is the pulse width distortion value at the t-1th moment in the pulse width distortion sequence, N width is the number of pulse width distortion values ​​in the pulse width distortion sequence; The formula for calculating the reflection frequency difference development value in S4 is: , where γ fs is the reflection frequency difference development value, r fs,t is the reflection frequency difference at the tth moment in the reflection frequency difference sequence, r fs,t-1 is the reflection frequency difference at the t-1th moment in the reflection frequency difference sequence, N fs is the number of reflection frequency differences in the reflection frequency difference sequence.

7. The method for predicting low-altitude intrusion behavior of unmanned aerial vehicles based on deep learning according to claim 1 is characterized in that: The UAV low-altitude invasion behavior prediction network in S5 includes: a first LSTM unit, a second LSTM unit, a third LSTM unit, a first feature enhancement unit, a second feature enhancement unit, a third feature enhancement unit, a first fully connected layer, a second fully connected layer, a third fully connected layer, a first compensation unit Z1, a second compensation unit Z2, a third compensation unit Z3 and a weighted layer; The input end of the first LSTM unit is used to input a pulse shape distortion sequence; the input end of the second LSTM unit is used to input a pulse width distortion sequence; the input end of the third LSTM unit is used to input a reflection frequency difference sequence; The input end of the first feature enhancement unit is connected to the output end of the first LSTM unit, and its output end is connected to the input end of the first fully connected layer; the input end of the second feature enhancement unit is connected to the output end of the second LSTM unit, and its output end is connected to the input end of the second fully connected layer; the input end of the third feature enhancement unit is connected to the output end of the third LSTM unit, and its output end is connected to the input end of the third fully connected layer; the first input end of the first compensation unit Z1 is connected to the output end of the first fully connected layer, and its second input end is used to input the morphological distortion development value; the first input end of the second compensation unit Z2 is connected to the output end of the second fully connected layer, and its second input end is used to input the width distortion development value; the first input end of the third compensation unit Z3 is connected to the output end of the third fully connected layer, and its second input end is used to input the reflection frequency difference development value; the input end of the weighted layer is respectively connected to the output end of the first compensation unit Z1, the output end of the second compensation unit Z2 and the output end of the third compensation unit Z3, and its output end serves as the output end of the UAV low-altitude invasion behavior prediction network.

8. The method for predicting low-altitude intrusion behavior of unmanned aerial vehicles based on deep learning according to claim 7 is characterized in that: The expressions of the first feature enhancement unit, the second feature enhancement unit and the third feature enhancement unit are all: , where H is the output feature vector of the first feature enhancement unit, the second feature enhancement unit or the third feature enhancement unit, X is the input feature vector of the first feature enhancement unit, the second feature enhancement unit or the third feature enhancement unit, and MP one is a one-dimensional maximum pooling layer, is element-wise multiplication.

9. The method for predicting low-altitude intrusion behavior of unmanned aerial vehicles based on deep learning according to claim 7 is characterized in that: The expression of the first compensation unit Z1 is: ,in, is the first risk component after compensation, ζ shape is the first risk component before compensation, e is a natural constant, γ shape is the morphological distortion development value; The expression of the second compensation unit Z2 is: ,in, is the second risk component after compensation, ζ width is the second risk component before compensation, γ width is the width distortion development value; The expression of the third compensation unit Z3 is: ,in, is the third risk component after compensation, ζ fs is the third risk component before compensation, γ fs is the reflection frequency difference development value.

10. A UAV low-altitude intrusion behavior prediction system based on deep learning, implemented based on the UAV low-altitude intrusion behavior prediction method based on deep learning according to any one of claims 1 to 9, characterized in that: include: Transmitting and receiving subsystem, distortion sequence construction subsystem, frequency difference sequence construction subsystem, development value calculation subsystem and prediction subsystem; The transmitting and receiving subsystem is used to transmit radar signals to the monitoring area and receive radar reflected signals; The distortion sequence construction subsystem is used to extract the pulse morphology distortion value and the pulse width distortion value of the radar reflection signal received in each time period, and respectively construct a pulse morphology distortion sequence and a pulse width distortion sequence; The frequency difference sequence construction subsystem is used to perform time-frequency transformation on the radar reflection signal received in each time period, calculate the reflection frequency difference, and construct the reflection frequency difference sequence; The development value calculation subsystem is used to calculate the morphology distortion development value for the pulse morphology distortion sequence, calculate the width distortion development value for the pulse width distortion sequence, and calculate the reflection frequency difference development value for the reflection frequency difference sequence; The prediction subsystem is used to input the pulse morphology distortion sequence, pulse width distortion sequence and reflection frequency difference sequence into the UAV low-altitude intrusion behavior prediction network, and obtain the UAV low-altitude intrusion risk value based on the compensation of the morphology distortion development value, the width distortion development value and the reflection frequency difference development value.

Images