Space-time-frequency integrated detection method and system based on artificial intelligence

Abstract

The present invention discloses a space-time-frequency integrated detection method and system based on artificial intelligence, the system comprising: a time domain processing module for digitally down-converting an input N-channel intermediate frequency signal to obtain a baseband I / Q signal; a space domain processing module for utilizing a space domain N-channel subarray interference suppression algorithm to realize space domain adaptive interference suppression, forming a space domain detection channel and two compensation channels; a frequency domain processing module for outputting three-way two-dimensional range Doppler complex (RD) images after pulse compression and moving target detection of three-channel signals; an AI detection and recognition integration module for converting the two-dimensional range Doppler complex image of the detection channel into a two-dimensional range Doppler RGB color image, performing detection and recognition based on an intelligent YORO IDeR algorithm and outputting the spatial three-dimensional information, Doppler dimension information and target type information of the target. The present invention retains all the information of the echo, and adopts an interference suppression algorithm to effectively suppress noise and improve signal quality; the present invention realizes the integration of target detection and recognition, and improves the performance of target detection and recognition in a complex electromagnetic environment.

Description

Technical Field

[0001] The present invention relates to the field of radar signal processing technology, and in particular to a space-time-frequency integrated detection method and system based on artificial intelligence. Background Art

[0002] Phased array radar, also known as electronic scanning radar, realizes antenna beam scanning by electronic methods and has the advantages of flexible beam scanning and multi-beam forming. Traditional mechanical scanning radar changes the direction of antenna beam by rotating the array, while phased array radar uses amplitude and phase controllable antenna units. It only needs to control the feed phase of each antenna unit in the array to change the direction of antenna beam, thereby realizing rapid scanning of antenna beam.

[0003] In an electronic warfare environment, radar systems are inevitably subject to various passive and active electromagnetic interferences. These interference signals enter the receiving system through the side lobes or main lobes of the receiving antenna pattern, greatly reducing the signal detection capability of the receiver. The anti-interference performance of radar has become an important indicator to measure its performance, and array antenna technology is an important means to improve the anti-interference capability of radar. The adaptive digital beamforming technology of array antenna enables the radar to automatically adopt the corresponding anti-interference scheme according to the interference characteristics, and can dynamically adjust the working parameters as the interference changes to achieve a certain optimal performance.

[0004] The traditional phased array radar signal processing uses an amplitude and phase calibration module to calibrate the amplitude and phase of the four-channel signal, and convert it into a sum channel, an azimuth difference channel, and an elevation difference channel. The data of the three channels are pulse compressed and then used for moving target detection. After that, the DC operation is removed and the data is passed to the constant false alarm module. The sum channel performs speed constant false alarm and distance constant false alarm. At the same time, the speed constant false alarm of the two difference channels takes out the target point, and the distance constant false alarm performs a secondary constant false alarm operation to confirm whether the previously taken out point can be identified as a target, and finally calculates the target information. In the beamforming process, it is necessary to suppress interference signals as much as possible to improve signal quality. At the same time, when facing a complex signal environment, amplitude and phase calibration is relatively lacking in suppressing interference signals. The constant false alarm detection used in the target information calculation process requires the use of complex algorithms and technologies, including signal processing, feature extraction, classifier design, etc., which increases the complexity and development difficulty of the system. In addition, constant false alarm detection requires processing and analysis of a large number of radar signals, which consumes a lot of computing resources and time. Moreover, the CFAR algorithm is sensitive to environmental changes, and its performance may be affected by environmental changes. For example, factors such as weather and terrain may cause the detection performance to degrade. Summary of the invention

[0005] The present invention proposes a space-time-frequency integrated detection method based on artificial intelligence.

[0006] The technical solution to achieve the purpose of the present invention is: a space-time-frequency integrated detection method based on artificial intelligence, comprising:

[0007] Perform digital down-conversion and digital orthogonal demodulation on the intermediate frequency echo signal of the N-channel of the received two-dimensional phased array antenna to obtain the echo baseband I / Q signal; perform amplitude and phase calibration on the echo baseband I / Q signal;

[0008] Subarray beamforming is performed on the signal after amplitude and phase calibration, and adaptive interference suppression processing is performed using the spatial N-channel subarray interference suppression algorithm to obtain the spatial interference suppression weighting coefficient; the N-channel baseband signal is weighted and summed using the spatial interference suppression weighting coefficient to complete the subarray-level adaptive digital beamforming and convert it into a spatial information detection signal; the N-channel baseband signal is weighted and subtracted using the spatial interference suppression weighting coefficient to obtain two-way spatial information compensation signals;

[0009] Pulse compression is performed on the spatial information detection signal and the two-way spatial information compensation signal, and the moving target detection algorithm is used to process the two-dimensional range Doppler complex image of the detection channel and the two-way compensation channel pointed by the beam where the target is located;

[0010] The two-dimensional range Doppler complex image of the detection channel is converted into a two-dimensional range Doppler RGB color image, and the two-dimensional range Doppler RGB color image is subjected to target detection and recognition to obtain the target type. At the same time, the coordinate information of the detected target is obtained, and the corresponding target distance and Doppler data value are calculated to complete the distance and speed information fitting; the target information is indexed into the two compensation channels to complete the target airspace information fitting; the target distance, Doppler data value and airspace directivity information constitute the target's spatial three-dimensional information and Doppler dimension information, and are merged with the target type information to finally obtain the target's spatial three-dimensional coordinates, Doppler dimension information and target type information.

[0011] Preferably, the specific method for the spatial domain processing module to perform subarray beamforming on the signal output from the time domain processing module after the amplitude and phase calibration is:

[0012] The N channels are evenly divided into L sub-arrays without overlap, and the sub-array beamforming signals of the L sub-arrays are:

[0013]

[0014] In the formula, ω i is an external input fixed weighting coefficient, k=1...L.

[0015] Preferably, the spatial domain processing module uses a spatial domain N-channel subarray interference suppression algorithm to perform adaptive interference suppression processing to obtain a spatial domain interference suppression weighting coefficient, and the specific formula is:

[0016]

[0017] Among them, R sub is the sub-array level covariance matrix, α T (θ0) is the subarray-level steering vector, ω subk is the spatial interference suppression weighting coefficient.

[0018] Preferably, the sub-array level covariance matrix is ​​specifically:

[0019]

[0020] Among them, y k (t) is the sample received by the kth sub-array level at time t, t is a positive integer, and M is the number of snapshots, i.e., the number of samples.

[0021] Preferably, the N-channel baseband signals are weighted and summed using the spatial interference suppression weighting coefficient to output, so as to complete the subarray-level adaptive digital beamforming, and the converted spatial information detection channel Z is specifically:

[0022] Z=y1ω sub1 +y2ω sub2 +…+y L ω subL

[0023] The spatial interference suppression weighting coefficient is used to weight the difference of the N-channel baseband signal to obtain two-way spatial information compensation channel Z α and Z β , specifically:

[0024]

[0025] In the formula, ω subk is the spatial interference suppression weighting coefficient of the kth sub-array.

[0026] Preferably, the AI ​​detection and recognition integrated module uses the YORO IDeR algorithm to perform target detection and recognition on color images, and the specific method is as follows:

[0027] Use convolutional neural networks to extract features from images and capture target information in images;

[0028] The target location information and its category are obtained through an independent coordinate feature projection head and classification feature projection head respectively. The information output by the coordinate feature projection head is passed through a multi-layer perceptron network to regress and estimate the location information of the target in the image, and obtain the target center point offset and relative length and width offset;

[0029] The detection information of the target is obtained by calculating the relative center point offset and the relative length offset, and the exact position of the target is located to obtain the coordinate information of the target in the two-dimensional range-Doppler RGB color image;

[0030] According to the range cells and Doppler cells corresponding to the two-dimensional range-Doppler RGB color image, the corresponding range and Doppler data values are obtained;

[0031] The feature vector output by the classification feature projection head passes through the RA-CNN network to output the target type and complete target recognition.

[0032] Preferably, the specific method for the AI detection and recognition integrated module to calculate the corresponding target range and Doppler data values and complete the velocity information fitting is as follows:

[0033] Select the target point and two adjacent points in the velocity dimension for velocity fitting. The target point is denoted as a0, the amplitude is denoted as V0, the two adjacent points before and after the target point are denoted as a1 and a2 respectively, and the amplitudes are V1 and V2 respectively; the position of the velocity gate after fitting is denoted as V;

[0034] Judge the magnitudes of V1 and V2. If V1≥V2, then select a0 and a1 for velocity fitting, and the fitting formula is

[0035]

[0036] If V1<V2, then select a0 and a2 for velocity fitting, and the fitting formula is

[0037]

[0038] Preferably, the specific method for the AI detection and recognition integrated module to calculate the corresponding target range and Doppler data values and complete the range information fitting is as follows:

[0039] Select the target point and two adjacent points on the coordinate axis before and after in the range dimension for range fitting. The target point is denoted as b0, the amplitude is denoted as R0, the two adjacent points before and after it are denoted as b1 and b2 respectively, and the amplitudes are R1 and R2 respectively. The position of the range gate after fitting is denoted as R;

[0040] Judge the magnitudes of R1 and R2. If R1≥R2, then select b0 and b1 for range fitting, and the fitting formula is:

[0041]

[0042] If R1<R2, select b0 and b2 for range fitting, and the fitting formula is:

[0043]

[0044] Preferably, the AI ​​detection and recognition integrated module indexes the target information into two compensation channels, and the specific formula for completing the target spatial domain information fitting is:

[0045]

[0046] In the formula, Δ represents the difference channel, ∑ represents the sum channel, and θ t is the relative offset angle, θ0 is the axial direction angle of the antenna, and F(θ) is the directivity function of the antenna voltage

[0047] The present invention also proposes an artificial intelligence-based space-time-frequency integrated detection system, comprising a time domain processing module, a space domain processing module, a frequency domain processing module and an AI detection and recognition integrated module, wherein the time domain processing module is used to digitally down-convert and digitally orthogonally demodulate the intermediate frequency echo signal of the N-channel received two-dimensional phased array antenna to obtain an echo baseband I / Q signal; perform amplitude and phase calibration on the echo baseband I / Q signal;

[0048] The spatial domain processing module is used to perform subarray beamforming on the signal after amplitude and phase calibration output from the time domain processing module, and use the spatial domain N-channel subarray interference suppression algorithm to perform adaptive interference suppression processing to obtain the spatial domain interference suppression weighting coefficient; use the spatial domain interference suppression weighting coefficient to weight and sum the N-channel baseband signal to output, complete the subarray-level adaptive digital beamforming, and convert it into a spatial domain information detection signal; use the spatial domain interference suppression weighting coefficient to weight the difference of the N-channel baseband signal to obtain two-way spatial domain information compensation signals;

[0049] The frequency domain processing module is used to perform pulse compression on the spatial information detection signal and the two-way spatial information compensation signal output by the spatial domain processing module, and use the moving target detection algorithm to process and obtain the detection channel and the two-way compensation channel two-dimensional range Doppler complex image of the beam pointing to the target;

[0050] The AI ​​detection and recognition integrated module is used to convert the two-dimensional range Doppler complex image of the detection channel output by the frequency domain processing module into a two-dimensional range Doppler RGB color image, and perform target detection and recognition on the two-dimensional range Doppler RGB color image to obtain the target type. At the same time, the coordinate information of the detected target is obtained, and the corresponding target distance and Doppler data value are calculated to complete the distance and speed information fitting; the target information is indexed into the two compensation channels to complete the target airspace information fitting; the target distance, Doppler data value and airspace directivity information constitute the target's spatial three-dimensional information and Doppler dimension information, and are merged with the target type information to finally obtain the target's spatial three-dimensional coordinates, Doppler dimension information and target type information.

[0051] Compared with the prior art, the present invention has the following significant advantages: the present invention retains all the information of the radar echo and effectively suppresses noise through the N-channel interference suppression algorithm, thereby improving the signal quality. When facing a complex environment, the present invention can handle more interference than the conventional subarray-based beamforming method; compared with the conventional radar signal processing algorithm, the consumption of hardware resources can be greatly reduced, and the calculation convergence speed is improved; the intelligent YORO IDeR algorithm is used to complete the detection and identification of the target in the signal domain, and the detection and identification integration based on the AI ​​algorithm is realized; the AI ​​algorithm is used to replace the traditional constant false alarm algorithm, thereby reducing the complexity and development difficulty of the system, improving the performance of target detection and identification in a complex electromagnetic environment, and avoiding the system errors caused by environmental changes.

[0052] The present invention is described in further detail below in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 This is a schematic diagram of a space-time-frequency integrated detection system based on artificial intelligence in the present invention.

[0054] Figure 2 Schematic diagram of a linear frequency modulation signal time domain waveform and signal spectrum in an embodiment of the present invention.

[0055] Figure 3 Schematic diagram of the time domain waveform and signal spectrum of the linear frequency modulation signal after orthogonal demodulation in an embodiment of the present invention.

[0056] Figure 4 Schematic diagram of the output waveform of the digitized antenna in an embodiment of the present invention.

[0057] Figure 5 Schematic diagram of subarray synthesis in an embodiment of the present invention. DETAILED DESCRIPTION

[0058] Example 1

[0059] A space-time-frequency integrated detection method based on artificial intelligence comprises the following steps:

[0060] The intermediate frequency echo signal of the N-channel of the two-dimensional phased array antenna is digitally down-converted, and then the digital down-conversion is completed through a low-pass filter and the echo baseband I / Q signal is output; the obtained echo baseband I / Q signal is amplitude-phase calibrated;

[0061] The signal after amplitude and phase calibration is first subjected to subarray beamforming, and then adaptive interference suppression processing is performed using the spatial N-channel subarray interference suppression algorithm to output the spatial interference suppression weighting coefficient; the baseband signal is weighted and summed using the spatial interference suppression weighting coefficient to complete the subarray-level adaptive digital beamforming and converted into the spatial information detection signal Z of the spatial information detection channel; the baseband signal is weighted and subtracted using the spatial interference suppression weighting coefficient and converted into the spatial information compensation signal (Z) of the two-way spatial information compensation channel. α and Z β ), the specific process is:

[0062] Perform subarray beamforming and divide the N channels into L subarrays evenly and without overlap. Then the subarray beamforming signals of the L subarrays are:

[0063]

[0064] Where ω i is the external input fixed weighting coefficient, n is the uniform non-overlapping division of N channels, that is k = 1...L;

[0065] The spatial interference suppression weighting coefficient is calculated using the spatial interference suppression algorithm. The following formula determines the spatial interference suppression weighting coefficient:

[0066]

[0067]

[0068] Where R sub is the sub-array level covariance matrix, α T (θ0) is the external input subarray-level steering vector, and the sample received by the subarray level at time t is y k (t), t is a positive integer, M is the number of snapshots, i.e. the number of samples;

[0069] The subarray-level adaptive beamforming is completed by using the spatial interference suppression weighting coefficient and converted into a detection channel and two compensation channels. Let the detection channel be Z and the two compensation channels be Z respectively. α and Z β , three-way signal is Figure 5 And the following formula determines:

[0070] Z=y1ω sub1 +y2ω sub2 +…+y L ω subL

[0071]

[0072] where ω subis the spatial interference suppression weighting coefficient.

[0073] The spatial information detection signal Z and the two-way spatial information compensation signal (Z α and Z β ) A total of three channels of data are pulse compressed, and after the pulse compression is completed, a moving target detection algorithm is used to detect a moving target, and a two-dimensional range Doppler (RD) complex image of the detection channel and two compensation channels pointed by the beam where the target is located is obtained;

[0074] The RD complex image of the airspace information detection channel is preprocessed and converted into a two-dimensional range Doppler RGB color image. The YORO IDeR algorithm is used to detect and identify the color image, output the target type, and obtain the coordinate information of the detected target. The coordinates are output to the distance and speed fitting module and indexed to the two compensation channels at the same time; the corresponding target distance and Doppler data value are calculated to complete the distance and speed information fitting; the target airspace information fitting is completed through the two-dimensional range Doppler complex image of the two compensation channels; after the fitting is completed, the target distance, Doppler data value and airspace directivity information are output, merged with the target type information, and the target's spatial three-dimensional information, Doppler dimension information and target type information are output. The specific process is as follows:

[0075] 1) First, the detection channel RD complex image is preprocessed and converted into a two-dimensional range Doppler RGB color image;

[0076] 2) Use convolutional neural network to extract features from the preprocessed image to capture local and global information in the image; then obtain the target location information and category through independent coordinate feature projection head and classification feature projection head respectively. The information output by the coordinate feature projection head will be regressed through a multi-layer perceptron network to estimate the location information of the target in the image, and obtain the normalized center point offset and relative length and width of the target. By solving the normalized center point offset and relative length, the detection information of the target can be obtained and the precise position of the target can be located, and the coordinates of the target in the two-dimensional range Doppler RGB color image can be obtained. The coordinates are output to the distance and speed fitting module and indexed to the two compensation channels; then the corresponding Doppler data value is calculated based on the coordinate information; the feature vector output by the classification feature projection head passes through the RA-CNN network to output the target type and complete the target recognition.

[0077] 3) After the YORO IDeR algorithm detects the target in the image, it outputs the coordinate information on the RD map of the spatial information detection channel, selects the target point and two surrounding points in the speed dimension for speed fitting, the target point is recorded as a0, the amplitude is recorded as V0, and the two points on the left and right are recorded as a1 and a2, with amplitudes of V1 and V2 respectively. The speed gate position after fitting is recorded as V.

[0078] First, compare the magnitudes of V1 and V2. If V1 ≥ V2, then select a0 and a1 for velocity fitting; if V1 < V2, then select a0 and a2 for velocity fitting.

[0079] If V1 ≥ V2, the fitting formula is

[0080]

[0081] If V1 < V2, the fitting formula is

[0082]

[0083] For distance fitting, use the same method:

[0084] Select the target point and the two points before and after the coordinate axis in the distance dimension for distance fitting. The target point is denoted as b0, the amplitude is denoted as R0, and the two points before and after it are denoted as b1 and b2, with amplitudes R1 and R2 respectively. The position of the distance gate after fitting is denoted as R.

[0085] First, compare the magnitudes of R1 and R2. If R1 ≥ R2, then select b0 and b1 for distance fitting; if R1 < R2, then select b0 and b2 for distance fitting.

[0086] If R1 ≥ R2, the fitting formula is

[0087]

[0088] If R1 < R2, the fitting formula is

[0089]

[0090] After distance and velocity fitting are completed, output the target distance and velocity information.

[0091] 4) After the YORO IDeR algorithm performs target detection on the color image, output the coordinate information on the detection channel RD map and complete the indexing to the two compensation channels. Also find the target points on the two compensation channels, cooperate with the detection channel to complete the sum-difference ratio amplitude angle measurement, and obtain the target airspace information, which is determined by the following formula:

[0092]

[0093] Where Δ represents the difference channel, ∑ represents the sum channel, θ t is the relative offset angle, θ0 is the antenna axial direction angle, F(θ) is the directivity function of the antenna voltage, θ t , θ0 is the target airspace information.

[0094] Combine the distance, Doppler data values, and airspace directivity information of the target with the target type information, and output the three-dimensional spatial information, Doppler dimension information, and target type information of the target.

[0095] Example 2

[0096] See also Figure 1 to Figure 4 This embodiment provides a space-time-frequency integrated detection system based on artificial intelligence. The present invention not only retains all the information of the radar echo signal, but also uses artificial intelligence algorithms to replace the traditional constant false alarm algorithms in subsequent radar processing such as detection and identification, thereby reducing the complexity and development difficulty of the system, improving the performance of target detection and identification in complex electromagnetic environments, and avoiding system errors caused by environmental changes.

[0097] In this embodiment, a drone is released as a target and the radar system is activated.

[0098] Time domain processing module: receives the N-channel intermediate frequency input of the two-dimensional phased array antenna, completes digital down-conversion and amplitude and phase calibration, and then sends it to the spatial domain processing module.

[0099] A linear frequency modulation signal with a signal time width of 2us, a center frequency of 125MHz, and a bandwidth of 32MHz is selected as the transmitting signal to detect and search for UAVs flying in the airspace. The antenna points to the receiving signal in the airspace through beam forming. The received signal is sampled by a bandpass to ensure that the original signal x(t) is restored from the discrete signal x(n) without distortion. The analog signal x(t) is sampled by AD to obtain the digital sequence x(n), and then multiplied with the two orthogonal digital signals cos(ω0n) and -sin(ω0n) respectively to obtain the orthogonal demodulated signal x(n). Each signal is By performing complex multiplication operations, the calculated digital signals can ensure that the amplitude and phase of each signal are aligned with the first signal. It needs to be measured in a dark room. After completing the amplitude and phase calibration, the signal is output to the spatial domain processing module.

[0100] Spatial processing module: The signal after amplitude and phase calibration output from the time domain processing module is first subjected to subarray beamforming, and then adaptive interference suppression processing is performed using the spatial domain N-channel subarray interference suppression algorithm to output the spatial domain interference suppression weighting coefficient; the baseband signal is weighted and summed using the spatial domain interference suppression weighting coefficient to complete the subarray-level adaptive digital beamforming and convert it into the spatial domain information detection signal Z of the spatial domain information detection channel; the baseband signal is weighted and subtracted using the spatial domain interference suppression weighting coefficient and converted into the spatial domain information compensation signal (Z) of the two-way spatial domain information compensation channel. α and Z β ); output spatial information detection signal Z and two-way spatial information compensation signal (Z α and Z β ) to the frequency domain processing module;

[0101] In a further embodiment, the N channels are evenly divided into L sub-arrays without overlap, and the sub-array beamforming signals of the L sub-arrays are:

[0102]

[0103] Where ω i is the external input fixed weighting coefficient, n is the uniform non-overlapping division of N channels, that is k = 1...L;

[0104] The interference suppression weighting coefficient is calculated using the following interference suppression algorithm expression:

[0105]

[0106] Where R sub is the sub-array level covariance matrix, α T (θ0) is the external input subarray-level steering vector, and the sample received by the subarray level at time t is y k (t), t is a positive integer, M is the number of snapshots, i.e. the number of samples;

[0107] After obtaining the weighted coefficients, the subarray-level adaptive beamforming is completed and converted into a detection channel and two compensation channels. The detection channel is Z and the two compensation channels are Z. α and Z β :

[0108] Z=y1ω sub1 +y2ω sub2 +…+y L ω subL

[0109]

[0110] where ω sub is the spatial interference suppression weighting coefficient.

[0111] The detection channel and two compensation channel signals are output to the frequency domain processing module.

[0112] Frequency domain processing module: completes pulse compression and moving target detection for the detection signal and two compensation signals obtained by the spatial domain processing module, and outputs a two-dimensional range Doppler complex image of the beam pointing to the target;

[0113] Pulse compression is accomplished through time domain matched filtering, where an FIR filter is used to convolve the matched filter coefficient h(n) with the three-way signal. The matched filter coefficient is conjugate with the received signal mirror image, and as the filter order increases, the number of multiplication and addition operations required by the signal also increases dramatically.

[0114] After the signal passes through the matched filter, it passes through the -35dB Taylor window to suppress the signal sidelobes, and finally outputs a peak signal with a narrow time width and relatively high main and side lobes.

[0115] The signal of 256 pulse repetition cycles is rearranged. Since a total of 256 points of Fourier transform are performed, the number of Doppler units is 0-255, and 0, 1 and 254, 255 are Doppler units near the Doppler frequency of 0. Deleting them can suppress the ground static and slow-moving clutter to obtain the data of the same distance unit, and suppress the static or low-speed interference signal by removing DC and deleting the Doppler frequency point, so that the average value of the data of the same distance unit is zero, and the interference of static clutter such as ground objects is suppressed; the signal spectrum leakage is prevented by adding windows, using Taylor window, the main sidelobe ratio is set to 35dB, and then the speed information of the moving target is obtained by FFT operation. Output the two-dimensional range Doppler broadband complex "image" to the artificial intelligence detection and recognition integrated module.

[0116] Artificial intelligence detection and recognition integrated module: The two-dimensional range Doppler complex image of the detection channel obtained by the frequency domain processing module is converted into a two-dimensional range Doppler RGB color image, and the YORO IDeR algorithm is used to detect and identify the color image, output the target type, and obtain the coordinate information of the detected target at the same time, and use it to calculate the corresponding target distance and Doppler data value to complete the distance and speed information fitting; the target information is indexed into the two compensation channels to complete the target spatial information fitting, and the target distance, Doppler data value and spatial directionality information constitute the target's spatial three-dimensional information and Doppler dimension information; finally, the target's spatial three-dimensional coordinates, Doppler dimension information and target type information are output.

[0117] First, the input complex image data is converted into amplitude image data. Then all data are standardized to the range of [0,1] to obtain normalized amplitude data. After that, the single-band amplitude image is loaded into the ENVI software that supports false color synthesis, and the preset color table is loaded. The color table divides the real numbers in the range of [0,1] into different data intervals and assigns different colors to each data interval. According to the preset color table, the amplitude data in different numerical ranges can be assigned different RGB values, thereby generating the two-dimensional range Doppler RGB color image required for the subsequent neural network model.

[0118] The target detection of the input two-dimensional range-Doppler RGB color image is realized by relying on the YORO IDeR algorithm. First, a convolutional neural network is used to extract features from the image to capture local and global information in the image. Then, the target position information and the category to which it belongs are obtained through an independent coordinate feature projection head and a classification feature projection head respectively. The information output by the coordinate feature projection head will pass through a multi-layer perceptron network to perform regression estimation on the position information of the target in the image, and obtain the normalized center point offset and relative length and width of the target. By resolving the normalized center point offset and relative length, the detection information of the target can be obtained and the exact position of the target can be located, and the coordinates of the target in the two-dimensional range-Doppler RGB color image can be obtained. Then, according to the range cells and Doppler cells corresponding to the two-dimensional range-Doppler RGB color image, the corresponding range and Doppler data values are obtained. The feature vector output by the classification feature projection head passes through the RA-CNN network to output the target type and complete target recognition. The output detection channel coordinate information is indexed to two compensation channels to complete the fitting of the target spatial domain information. Since the image coordinates output by the model are floating-point numbers, the range values and Doppler values obtained by this module are more accurate than the range cells and Doppler gates obtained by conventional radar systems.

[0119] After obtaining the velocity and range coordinate information, velocity and range information fitting is performed. In the velocity dimension, we select the target point and two surrounding points for velocity fitting. The target point is denoted as a0, and the amplitude is denoted as V0. The two points on the left and right are denoted as a1 and a2 respectively, and the amplitudes are V1 and V2 respectively. The position of the velocity gate after fitting is denoted as V.

[0120] First, judge the magnitudes of V1 and V2. If V1≥V2, then select a0 and a1 for velocity fitting; if V1<V2, then select a0 and a2 for velocity fitting.

[0121] If V1≥V2, the fitting formula is

[0122]

[0123] If V1<V2, the fitting formula is

[0124]

[0125] For range fitting, use the same method:

[0126] In the range dimension, select the target point and two points before and after the coordinate axis for range fitting. The target point is denoted as b0, and the amplitude is denoted as R0. The two points before and after it are denoted as b1 and b2 respectively, and the amplitudes are R1 and R2 respectively. The position of the range gate after fitting is denoted as R.

[0127] First, determine the magnitudes of R1 and R2. If R1 ≥ R2, then select b0 and b1 for distance fitting; if R1 < R2, then select b0 and b2 for distance fitting.

[0128] If R1 ≥ R2, the fitting formula is

[0129]

[0130] If R1 < R2, the fitting formula is

[0131]

[0132] After the distance and speed fitting are completed, output the target distance and speed information.

[0133] Output the detection channel coordinate information and index it to the two compensation channels. Then, find the target points on the two compensation channels as well, and cooperate with the sum channel to complete the sum-difference amplitude comparison for angle measurement to obtain the azimuth and elevation information.

[0134]

[0135] Where Δ represents the difference channel, ∑ represents the sum channel, θ t is the relative offset angle, θ0 is the antenna axial direction angle, and F(θ) is the antenna voltage directivity function.

[0136] After the above steps, it can be proven that this radar has achieved detection and recognition of the target. The target information is as shown in the figure: distance: 1032.52 meters, azimuth: 39.672°, elevation: 56.315°, speed: 10.697 m / s, and the radar can distinguish that the target type is a drone, achieving the integration of detection and recognition based on artificial intelligence, with higher detection accuracy and stronger anti-interference ability.

[0137] In the above embodiments, all functions can be realized, or some functions can be realized according to needs.

[0138] In addition, it should be noted that, in the above specific embodiments, the various specific technical features described can be combined in any appropriate manner without contradiction. To avoid unnecessary repetition, the present invention will not separately describe various possible combination methods.

Claims

1. A space-time-frequency integrated detection method based on artificial intelligence, characterized in that: include: Perform digital down-conversion and digital orthogonal demodulation on the intermediate frequency echo signal of the N-channel of the received two-dimensional phased array antenna to obtain the echo baseband I / Q signal; Perform amplitude and phase calibration on the echo baseband I / Q signal; Subarray beamforming is performed on the signal after amplitude and phase calibration, and adaptive interference suppression processing is performed using the spatial N-channel subarray interference suppression algorithm to obtain the spatial interference suppression weighting coefficient; The spatial interference suppression weighting coefficient is used to perform weighted summing and output of the N-channel baseband signals, completing the subarray-level adaptive digital beamforming and converting it into a spatial information detection signal; the spatial interference suppression weighting coefficient is used to perform weighted difference calculation on the N-channel baseband signals to obtain two-way spatial information compensation signals; Pulse compression is performed on the spatial information detection signal and the two-way spatial information compensation signal, and the moving target detection algorithm is used to process the two-dimensional range Doppler complex image of the detection channel and the two-way compensation channel pointed by the beam where the target is located; The AI ​​detection and recognition integrated module converts the two-dimensional range Doppler complex image of the detection channel into a two-dimensional range Doppler RGB color image, and performs target detection and recognition on the two-dimensional range Doppler RGB color image to obtain the target type. At the same time, it obtains the coordinate information of the detected target, calculates the corresponding target distance and Doppler data value, and completes the distance and speed information fitting; indexes the target information into the two compensation channels to complete the target airspace information fitting; the target distance, Doppler data value and airspace directivity information constitute the target's spatial three-dimensional information and Doppler dimension information, and merges them with the target type information to finally obtain the target's spatial three-dimensional coordinates, Doppler dimension information and target type information; Among them, the AI ​​detection and recognition integrated module uses the YORO IDeR algorithm to detect and recognize targets in color images. The specific method is as follows: Use convolutional neural networks to extract features from images and capture target information in images; The target location information and its category are obtained through an independent coordinate feature projection head and classification feature projection head respectively. The information output by the coordinate feature projection head is passed through a multi-layer perceptron network to regress and estimate the location information of the target in the image, and obtain the target center point offset and relative length and width offset; By solving the relative center point offset and relative length offset, the detection information of the target is obtained and the precise position of the target is located, and the coordinate information of the target in the two-dimensional range Doppler RGB color image is obtained; According to the range unit and Doppler unit corresponding to the two-dimensional range Doppler RGB color image, corresponding range and Doppler data values ​​are obtained; The feature vector output by the classification feature projection head passes through the RA-CNN network to output the target type and complete target recognition.

2. The space-time-frequency integrated detection method based on artificial intelligence according to claim 1 is characterized in that: The specific method of performing subarray beamforming by the spatial domain processing module on the signal after amplitude and phase calibration output from the time domain processing module is as follows: The N channels are evenly divided into L sub-arrays without overlap, and the sub-array beamforming signals of the L sub-arrays are: In the formula, ω i is an external input fixed weighting coefficient, 3. The space-time-frequency integrated detection method based on artificial intelligence according to claim 2 is characterized in that: In the spatial processing module, the spatial N-channel subarray interference suppression algorithm is used to perform adaptive interference suppression processing to obtain the spatial interference suppression weighting coefficient. The specific formula is: Among them, R sub is the sub-array level covariance matrix, α T (θ0) is the subarray-level steering vector, ω subk is the spatial interference suppression weighting coefficient.

4. The space-time-frequency integrated detection method based on artificial intelligence according to claim 3 is characterized in that: The sub-array level covariance matrix is ​​specifically: Among them, y k (t) is the sample received by the kth sub-array level at time t, t is a positive integer, and M is the number of snapshots, i.e., the number of samples.

5. The space-time-frequency integrated detection method based on artificial intelligence according to claim 2 is characterized in that: The weighted summation output of the N-channel baseband signals is performed using the spatial domain interference suppression weighting coefficients to complete the subarray-level adaptive digital beamforming. The converted spatial domain information detection signal Z is specifically as follows: Z=y1ω sub1 +y2ω sub2 +…+y L oh subL The spatial interference suppression weighted coefficient is used to weight the difference of the N-channel baseband signal to obtain the two-way spatial information compensation signal Z α and Z β , specifically: In the formula, ω subk is the spatial interference suppression weighting coefficient of the kth sub-array.

6. The space-time-frequency integrated detection method based on artificial intelligence according to claim 1 is characterized in that: In the AI detection and recognition integration module, the specific method for calculating the corresponding target distance and Doppler data values and completing the velocity information fitting is as follows: Select the target point and two adjacent points in the velocity dimension for velocity fitting. The target point is denoted as a0, and its amplitude is denoted as V0. The two adjacent points before and after the target point are denoted as a1 and a2, and their amplitudes are V1 and V2 respectively. The position of the velocity gate after fitting is denoted as V; Judge the magnitudes of V1 and V2. If V1≥V2, then select a0 and a1 for velocity fitting, and the fitting formula is If V1<V2, then select a0 and a2 for velocity fitting, and the fitting formula is 7. The space-time-frequency integrated detection method based on artificial intelligence according to claim 1 is characterized in that: In the AI detection and recognition integration module, the specific method for calculating the corresponding target distance and Doppler data values and completing the distance information fitting is as follows: Select the target point and two adjacent points on the coordinate axis in the distance dimension for distance fitting. The target point is denoted as b0, and its amplitude is denoted as R0. The two adjacent points before and after it are denoted as b1 and b2, and their amplitudes are R1 and R2 respectively. The position of the distance gate after fitting is denoted as R; Judge the magnitudes of R1 and R2. If R1≥R2, then select b0 and b1 for distance fitting, and the fitting formula is: If R1<R2, select b0 and b2 for distance fitting, and the fitting formula is:

8. The space-time-frequency integrated detection method based on artificial intelligence according to claim 1 is characterized in that: The AI detection and recognition integration module indexes the target information into the two compensation channels. The specific formula for completing the target spatial domain information fitting is: In the formula, Δ represents the difference channel, ∑ represents the sum channel, θ t is the relative offset angle, θ0 is the axial direction angle of the antenna, and F(θ) is the directivity function of the antenna voltage.

9. An artificial intelligence-based space-time-frequency integrated detection system based on any one of the methods of claims 1 to 8, characterized in that: It includes a time domain processing module, a spatial domain processing module, a frequency domain processing module, and an AI detection and recognition integration module. The time domain processing module is used to perform digital down-conversion and digital quadrature demodulation on the intermediate frequency echo signals of the N channels of the received two-dimensional phased array antenna to obtain the echo baseband I / Q signals; Perform amplitude and phase calibration on the echo baseband I / Q signals; The spatial domain processing module is used to perform subarray beamforming on the signals after amplitude and phase calibration output by the time domain processing module, and perform adaptive interference suppression processing using the spatial domain N-channel subarray interference suppression algorithm to obtain the spatial domain interference suppression weighting coefficients; The weighted summation output of the N-channel baseband signals is performed using the spatial domain interference suppression weighting coefficients to complete the subarray-level adaptive digital beamforming and convert it into a spatial domain information detection signal; the weighted difference of the N-channel baseband signals is obtained using the spatial domain interference suppression weighting coefficients to obtain two spatial domain information compensation signals; The frequency domain processing module is used to perform pulse compression on the spatial domain information detection signal and the two spatial domain information compensation signals output by the spatial domain processing module, and perform processing using the moving target detection algorithm to obtain the two-dimensional range-Doppler complex images of the detection channel and the two compensation channels where the target is located; The AI ​​detection and recognition integrated module is used to convert the two-dimensional range Doppler complex image of the detection channel output by the frequency domain processing module into a two-dimensional range Doppler RGB color image, and perform target detection and recognition on the two-dimensional range Doppler RGB color image to obtain the target type. At the same time, the coordinate information of the detected target is obtained, and the corresponding target distance and Doppler data value are calculated to complete the distance and speed information fitting; the target information is indexed into the two compensation channels to complete the target airspace information fitting; the target distance, Doppler data value and airspace directivity information constitute the target's spatial three-dimensional information and Doppler dimension information, and are merged with the target type information to finally obtain the target's spatial three-dimensional coordinates, Doppler dimension information and target type information.

Images