Method and System for Identifying Slow Moving Targets on the Sea Surface by an Unmanned Aerial Vehicle Based on Passive Radar Signals

Abstract

The present invention discloses a method and system for identifying slow-moving targets on the sea surface carried by an unmanned aerial vehicle based on passive radar signals. The shipborne search radar, which is the main electromagnetic radiation source of typical slow-moving targets on the sea surface, is used as the source for the passive radar to obtain electromagnetic radiation signals. Based on the radar parameters and operation modes of ship targets, the passive radar signals of typical sea surface targets are simulated through a simulation method. On this basis, a passive radar signal sample library of typical slow-moving targets on the sea surface is established to obtain the sample basis for data-driven target recognition. Furthermore, the sample signals are input into the target recognition system at the backend of the passive radar system to achieve the recognition of typical targets, that is, the identification of target categories. The present invention can provide advanced technical guidance for the construction of intelligent target recognition methods and systems driven by large samples, help expand the perception channels of slow-moving targets on the sea surface of unmanned aerial vehicle platforms, and enrich the means of battlefield reconnaissance and target surveillance.

Description

Technical Field

[0001] The present invention belongs to the technical field of pattern recognition, and particularly relates to a method and system for identifying slow-moving targets on the sea surface carried by an unmanned aerial vehicle based on passive radar signals. Background Art

[0002] The slow-moving targets on the sea surface described in the present invention can be defined from two aspects: ① Macroscopically, the slow-moving targets on the sea surface are mainly relative to the fast movement of airborne or missile-borne radars. That is, general sea surface moving targets, such as various moving ships, all belong to the category of slow-moving targets, but some other moving targets, such as cruise missiles flying over the sea surface, are not included; ② Spectrally, the slow-moving targets on the sea surface are relative to stationary targets. Ordinary airborne active radar systems only consider the focusing of stationary targets. For slow-moving targets, their echo signals fall into the main lobe clutter area. Since their motion parameters are different from those of stationary targets, they usually appear as defocused or shifted information in the stationary radar image. The slow-moving targets are mixed with the main clutter and are not easily detected and located, but these targets are often sea surface ships, boats and other targets that are of military interest.

[0003] A passive radar, also known as a passive radar, is a radar that detects targets by receiving the microwave energy radiated by the target itself or reflected by other sources without using a transmitter to emit energy. A passive radar system is a system that can detect and analyze the signals of target radiation sources and obtain radiation source information from them. The passive radar system itself does not emit energy, but passively receives the electromagnetic signals radiated or reflected by the target itself or non-cooperative radiation sources to track and locate the target. This makes it impossible for detection equipment and anti-radiation missiles to capture, track and attack the passive radar using electromagnetic signals. The passive radar system has a simple structure, a small size, can be installed on a mobile platform, is easy to deploy and has a low maintenance cost. Since the passive radar does not emit signals to irradiate the target, it is not easily perceived by the other party and generally has no problem of being interfered, and can work all day and all weather; it does not require frequency allocation, so it can be deployed in areas where conventional radars cannot be deployed.

[0004] The passive radar system has the following characteristics: First, it has a wide bandwidth, and the passive radar system can receive the radiation signals of various active radars; second, it has a long operating range and good concealment performance. Under the same sensitivity conditions, it can receive signals much farther than ordinary radars. Since it does not emit signals itself, it is not easily detected by the enemy, which can improve its own survival ability; third, it has a large dynamic range. Since the radiation sources are far and near, the incoming signals vary greatly in strength. To complete the simultaneous detection of strong and weak signals, a logarithmic amplifier must be used to achieve a large dynamic range.

[0005] Currently, the application research of passive radar technology has been launched in many countries. Universities such as the Defense Advanced Research Projects Agency of the US Department of Defense, the University of Washington, and the Georgia Institute of Technology, as well as companies such as Raytheon, have also carried out relevant research. In Europe, the UK has studied passive coherent radar and "Celldar", and France and Italy have also developed prototype systems. Russia and the Czech Republic are also engaged in similar research.

[0006] Passive radar systems can be classified according to the detection object or configuration method. According to the configuration method, passive radars are divided into two categories: fixed (ground-based) and mobile (installed on platforms such as submarines, ships, aircraft, and ground vehicles). The detection targets of passive radars can be radars, communication radio stations, or other radio radiation sources, or they can be targets that only reflect radio signals. Passive radars can be divided into two categories according to different detection objects: detecting and tracking using the self-radiation of the detected target, and detecting and tracking using the electromagnetic waves emitted by an external illumination source. An unmanned aerial vehicle (UAV)-borne sea surface slow moving target recognition system based on passive radar signals involved in the present invention belongs to the mobile configuration and is an intelligent recognition system for classifying the radar self-radiation of slow moving targets such as sea surface ships.

[0007] Typical passive radar systems include the following:

[0008] The US "Silent Sentry" radar. This passive radar uses the 50 - 80 MHz continuous wave signals emitted by commercial FM radio and television stations to detect, track, and monitor moving targets in the area. The system consists of a large dynamic range digital receiver, a phased array receiving antenna, a high-performance parallel processor with a gigaflop per second operation speed and its software. Tests have shown that its tracking range for a target with a radar cross section of 10 m2 can reach 180 km, and can reach 220 km after improvement. It can track more than 200 targets simultaneously, and the resolution interval is 15 m. The "Silent Sentry" can be installed on buildings and fixed structures, or on aircraft, trucks, and shelters for rapid deployment. They generally use the coastal area broadcast system as the illumination source. The system on a submarine is installed on the periscope and uses an omnidirectional antenna to provide warnings for helicopters or coastal reconnaissance aircraft. The fixed "Silent Sentry" system can achieve full airspace coverage, conduct real-time three-dimensional tracking and monitoring of targets, update data 8 times per second, is not affected by weather conditions, and has low system cost and maintenance cost. The phased array antenna uses commercial components and is installed on a building, with an antenna size of 2.3×2.5 m. The system has an observation range of 120°, uses digital beamforming technology to achieve coverage of the entire sector, and can track fixed-wing aircraft, rotary-wing aircraft, cruise missiles, and ballistic missiles.

[0009] British "Honeycomb" radar. This radar system can detect, track and identify moving targets on land, at sea and in the air, including vehicles moving in the woods. In theory, it can detect ground targets within 10 - 15 kilometers and large aircraft within 100 kilometers in the field environment. When a target enters the detection area, it causes the reflection of the radiation waves of cellular phones, and these reflections are detected by one or more cellular phone radars. The detection data is transmitted to the central control system in real time through the communication network, where the data is processed to determine the position and speed of the target. In addition to reflecting the radiation signals of cellular phone base stations, this radar system can also use acoustic sensors to detect the noise radiated by the target, which helps to determine the target's position. For air defense, the phased array receiver of this system can adopt an unfolded structure or be integrated into a camouflage net. When used for passive early warning, it can be integrated into a tank or an early warning aircraft. This system can also be used for coastal surveillance, battlefield reconnaissance, special intelligence collection (such as secretly monitoring airport activities without personnel), border security and offshore operations, etc. The "Honeycomb" radars deployed along the coastline can be used for tasks such as ship detection, assisting search and rescue, navigation, and warning. It can also be installed on an early warning aircraft to form a passive early warning capability, secretly monitoring the airspace and detecting aircraft more than 100 kilometers away.

[0010] Czech "Vera - E" radar. "Vera - E" is the latest model of the "Vera" radar series, which can detect, locate, identify and track air, ground and sea targets. The maximum detection range for air targets is 450 kilometers, and it can identify targets and generate air target images. The "Vera - E" system consists of four parts: an analysis and processing center is in the middle, and three signal receiving stations are distributed in an arc - like shape around it, with a distance of more than 50 kilometers between stations. The analysis and processing center is deployed in a sheltered vehicle, with a complete computer system as well as communication, command and control systems. The signal receiving stations are carried by heavy - duty trucks and can be flexibly deployed. When the receiving antenna support is erected, it is 17 meters high and covers an area of 9×12 meters. Three people can erect the antenna and enter the monitoring state within 1 hour. The antenna is in a cylindrical structure, with low power consumption and extremely high reliability. The mean time between failures reaches 2000 hours, and it can withstand strong winds of 30 m / s.

[0011] In the research of related technologies of passive radar systems, certain progress has been made in core technology fields at home and abroad, such as time-frequency synchronization, direct wave parameter estimation, direct wave interference suppression, target detection and positioning. However, current target detection based on passive radar mainly focuses on air targets, and the key point is to determine whether there is a specific target in the acquired radar signals; for sea targets, due to their slow moving speed, small Doppler frequency shift, and the fact that the ocean environment is more complex than the sky, the target echo is strongly interfered by direct wave side lobes, ground clutter and sea clutter. Therefore, the detection of slow-moving sea targets faces many difficulties and challenges. There are no relevant literatures on the design and implementation of a slow-moving sea target recognition system for unmanned aerial vehicle (UAV) - borne applications based on a large amount of passive radar signal data and intelligent target classification algorithms. Summary of the Invention

[0012] To overcome the deficiencies of the prior art, the present invention provides a method and system for UAV - borne slow-moving sea target recognition based on passive radar signals. The main electromagnetic radiation source of typical slow-moving sea targets, namely shipborne search radar, is used as the source for the passive radar to obtain electromagnetic radiation signals. Based on the radar parameters and operation modes of ship targets, the passive radar signals of typical sea targets are simulated by a simulation method. On this basis, a passive radar signal sample library of typical slow-moving sea targets is established to obtain the sample basis for data-driven target recognition. Furthermore, the sample signals are input into the target recognition system at the back end of the passive radar system to achieve the recognition of typical targets, that is, the identification of target categories. Relying on a large number of passive radar signal samples and a high-precision deep neural network model, the present invention proposes an effective method for slow-moving sea target recognition, which can provide advanced technical guidance for the construction of data-driven intelligent target recognition methods and systems, help expand the perception channels of slow-moving sea targets on UAV - borne platforms, and enrich battlefield reconnaissance and target surveillance means.

[0013] The technical solutions adopted by the present invention to solve its technical problems are as follows:

[0014] Step 1: Construction of a passive radar simulation sample library for slow-moving sea targets;

[0015] Step 1-1: Simulation and generation of passive radar signals of slow-moving sea targets;

[0016] Step 1-1-1: Define a large sea ship, namely Target 1, whose shipborne search radar electromagnetic radiation spectrum consists of radar signals in the frequency bands of 0.85 - 0.94 GHz, 0.95 - 1.2 GHz, 2.9 - 3.1 GHz, and 5.4 - 5.8 GHz;

[0017] Define a medium - large sea ship, namely Target 2, whose shipborne search radar electromagnetic radiation spectrum consists of radar signals in the frequency bands of 0.85 - 0.94 GHz and 0.95 - 1.2 GHz;

[0018] Different frequency electronic devices equipped according to Target 1 and Target 2 simulate corresponding electromagnetic signals by performing time-frequency domain transformation and pulse compression on the linear frequency modulation (LFM) signal, and use the spectrum of the simulated electromagnetic signal as a feature.

[0019] Step 1-1-2: Assume that the passive radar received signal of Target 1 contains four types of radar electromagnetic radiation signals with frequencies of 0.85 - 0.94 GHz, 0.95 - 1.2 GHz, 2.9 - 3.1 GHz, and 5.4 - 5.8 GHz respectively. Assume it is a sine wave with an amplitude of 1, plus Gaussian noise with a mean of 0 and a variance of 0.01. The radar startup is random.

[0020] The passive radar received signal of Target 2 contains radar electromagnetic radiation signals with frequencies of 0.85 - 0.94 GHz and 0.95 - 1.2 GHz respectively. Assume it is a sine wave with an amplitude of 1, plus Gaussian noise with a mean of 0 and a variance of 0.01. The radar startup is random.

[0021] Simulate Target 3, i.e., a stationary sea target, using Gaussian noise with a mean of 0 and a variance of 0.01.

[0022] Step 1-1-3: Set a random variable r, whose value is determined by a pseudo-random number uniformly distributed between 0 and 1:

[0023] r = rand(1)

[0024] Generate three types of samples of Target 1, Target 2, and Target 3 respectively according to probabilities P of 0.6, 0.3, and 0.1, that is:

[0025] P = [0.6 0.3 0.1]

[0026] Set the target frequency matrix FR as:

[0027] FR = [0.85 0.94; 0.95 1.2; 2.9 3.1; 5.4 5.8]

[0028] For Target 1, the radar radiation signal received by the passive radar is the superposition of 4 sine waves with different frequencies and amplitudes of 1 from 4 different types of radars. Assume the generated mixed frequency is Fs, then for the 4 different types of radars, Fs(j) is:

[0029] Fs(j) = FR(j,1) + r * (FR(j,2) - FR(j,1))

[0030] Among them, j = 1, 2, 3, 4, representing the serial number of the radar model;

[0031] For target 2, the generated mixed frequency Fs only contains two types of radars, A and B. The radar radiation signals received by the passive radar are the superposition of two sine waves with different frequencies and an amplitude of 1 from radars A and B. Its target frequency matrix FR is as follows:

[0032] FR' = [0.85 0.94; 0.95 1.2]

[0033] Fs'(j) for two different types of radars is:

[0034] Fs'(j) = FR'(j,1) + r * (FR'(j,2) - FR'(j,1))

[0035] where j = 1, 2, representing the serial numbers of radar models;

[0036] Based on the mixed frequency, the calculation method for the one-dimensional time-domain waveform signal s generated by target 1 at the receiving end through simulation is:

[0037] s = sin(2π * Fs(j) * t) + randn(size(s)) * 0.1, j = 1, 2, 3, 4

[0038] where randn() is the standard normal distribution random number generation function, and size(s) is the length of the signal s;

[0039] The calculation method for the one-dimensional time-domain waveform signal s' generated by target 2 at the receiving end through simulation is:

[0040] s' = sin(2π * Fs'(j) * t) + randn(size(s')) * 0.1, j = 1, 2

[0041] where randn() is the standard normal distribution random number generation function, and size(s') is the length of the signal s';

[0042] The frequency-domain signal F is the Fourier transform of the corresponding time-domain signal. For target 1, the calculation method is:

[0043] F = FFT(s)

[0044] Step 1-2: Adding noise and augmenting the passive radar signal samples;

[0045] Adding random noise with different intensity levels to the passive radar time-domain signal. For target 1, the time-domain waveform signal s after adding noise Addnoise is:

[0046] s Addnoise = s + randn(size(s)) * 0.1

[0047] Introduce the signal-to-noise ratio SNR, and its calculation formula is:

[0048]

[0049] Wherein, P signal is the signal power, and P noise is the noise power;

[0050] For a sinusoidal signal with a period of T, its power is simplified to:

[0051]

[0052] That is, the average power of the sinusoidal signal over one period is equal to the square of the amplitude divided by 2. Considering that the passive radar signal is composed of the superposition of 4 sinusoidal waves with different frequencies and an amplitude of 1, the signal power is

[0053] P signal = 4 * P 正弦 = 2A 2

[0054] For Gaussian white noise with a mean of 0, the noise power is the variance.

[0055] By introducing the signal-to-noise ratio SNR for calculation, the signal-to-noise ratio of the passive radar time-domain signal after adding Gaussian noise of different intensity levels is obtained;

[0056] Steps 1 - 3: Generate a passive radar signal data sample library for slow-moving targets on the sea surface;

[0057] Based on Step 1 - 1, the one-dimensional time-domain waveform signal of a single target sample is simulated. The length of the signal sequence is 10 s, and through sampling at a time interval of 0.01 s, a 1001-dimensional discrete time series is obtained, and the amplitude value of each sampling point is the value of the one-dimensional time-domain waveform signal s;

[0058] The discrete time series is stored in the.txt document format, using the variable-length UTF-8 encoding method, which supports Chinese characters;

[0059] After simulating a single target sample, multiple target types and samples of each type are generated to construct a passive radar signal data sample library for slow-moving targets on the sea surface;

[0060] An index for each sample is established, and this index is stored in the.txt document format;

[0061] Step 2: Target recognition based on passive radar signals and multi-scale one-dimensional convolutional neural networks;

[0062] The front-end signal acquisition device of the unmanned aerial vehicle (UAV)-borne passive radar system collects the electromagnetic radiation signals of typical slow-moving targets on the sea surface, that is, the time-domain waveform signals of three types of targets. Based on the constructed multi-scale one-dimensional CNN classifier for passive radar signals, supervised learning is carried out through three types of target samples generated by simulation, and different features of passive radar signals of three typical sea surface targets are automatically learned to realize the recognition of three typical sea surface targets, namely, target 1, target 2, and target 3, that is, the judgment of three target types; specifically as follows:

[0063] Step 2-1: Construct a multi-scale one-dimensional CNN classifier for passive radar signals;

[0064] The input end of the multi-scale one-dimensional CNN classifier is a one-dimensional passive radar signal, which is a one-dimensional time-domain signal with 1001 sampling points. After the signal is input, it first passes through a one-dimensional convolutional layer conv1 and a max-pooling layer pool1, and then enters the convolutional layer paths of three scales with convolutional kernels of 3, 5, and 7 respectively for convolution and pooling operations. Each convolutional layer path also undergoes 3 layers of convolution + max-pooling processing. Then, through the splicing of the three path tensors, it is sent to the max-pooling layer pool2. Finally, through two fully connected layers (fc1, fc2), it is sent to the softmax activation function, that is, the output layer in the recognition problem of three types of targets is obtained;

[0065] Step 2-2: Based on the input-output mode of the recognition of slow-moving targets on the sea surface by passive radar signals;

[0066] The input signal is designed as a general one-dimensional time-domain signal. According to the sampling number parameter settings generated by the simulation of the sample signal, the signal length is 1001, and this length can also be adjusted according to the sample generation requirements. For the actual radar time-domain signal, an embedded data acquisition card is used to collect one-dimensional discrete signals of different lengths as needed. On this basis, a sample signal matrix is constructed to facilitate subsequent CNN operation.

[0067] The output end is designed with a total of three category labels for target 1, target 2, and target 3. After Softmax operation, the probability values of the current sample belonging to the three types of targets are directly output, and the category with the largest probability value is the recognition result of the current sample.

[0068] Preferably, the pooling layers of the multi-scale one-dimensional CNN classifier all adopt max pooling. The kernel size of pooling layer pool2 is 8×8, the stride is 1, and the padding is 0. The kernel sizes of all other pooling layers are 2×2, the strides are 2, and the padding is 0. The fully connected layer fc1 realizes the fusion of the feature matrices, fusing the three output feature matrices of the three channels into one feature matrix. Therefore, the number of input features in_features is 3, and the number of output features out_features is 1. The feature matrix is input into the second fully connected layer fc2 after being transformed in matrix shape. The number of input features in_features of fc2 is 3, and the number of output features out_features is 1, finally realizing the classification of three types of targets.

[0069] A UAV-borne sea surface slow moving target recognition system based on passive radar signals, comprising a functional framework and a hardware subsystem;

[0070] The functional framework includes a passive radar sensor module, a passive radar signal sampling module, a passive radar signal processing module, a sea surface slow moving target recognition module, an information transmission module, an intelligent control module, and a target recognition result output identification module;

[0071] The passive radar sensor module is a front-end receiving device for acquiring passive radar signals of sea surface slow moving targets, capable of acquiring the radar electromagnetic radiation signals of targets, that is, one-dimensional time-domain signals mixed by multiple sine waves. This signal provides test samples for the target recognition system. At the same time, this module can also acquire the azimuth information of the target relative to itself, and through coordinate transformation, acquire the absolute position information of the target, that is, the longitude and latitude coordinate information of the target;

[0072] The passive radar signal sampling module discretely samples the one-dimensional time-domain signals acquired by the passive radar sensor module, collects one-dimensional discrete signals of different lengths as needed, and constructs a sample signal matrix to facilitate subsequent operation and processing of target recognition algorithms;

[0073] The passive radar signal processing module preprocesses the acquired passive radar raw signals, including filtering, denoising, and data arrangement;

[0074] The sea surface slow moving target recognition module includes a high-performance embedded arithmetic unit, a memory, a high-speed interface device, and a target recognition algorithm, and completes the storage of the target recognition algorithm model, the storage of input samples, and the recognition of input target samples, realizing the classification of sea surface slow moving targets, that is, the determination of target categories;

[0075] The information transmission module completes the wired and wireless transmission functions of target recognition data and target coordinate data. Wired transmission is responsible for the internal data information transmission of the unmanned aerial vehicle (UAV)-borne slow-moving target recognition system on the sea surface. Wireless transmission is responsible for the data information transmission between the system and the ground control station. For short-range wireless communication, wireless digital data link sending and receiving devices are directly used to realize the information transmission between the system and the ground control station. For long-range wireless communication, data links and communication relay devices are used to realize the information transmission between the system and the ground control station.

[0076] The intelligent control module controls the normal and orderly operation of the passive radar sensor module, the slow-moving target recognition module on the sea surface, and the information transmission module through the transmission of control signals in the UAV-borne slow-moving target recognition system on the sea surface. It can also independently plan and make decisions on the slow-moving target detection task on the sea surface according to the target recognition result.

[0077] The ground display control module is deployed on the host of the ground control station. Through the wirelessly obtained target coordinates and target recognition information, it realizes the visual output and positioning identification of the target in the digital map of the remote ground station, and completes the real-time tracking and character overlay function of the target.

[0078] The hardware subsystem includes an FPGA data reception and processing subsystem, an Ascend AI acceleration operation and control subsystem, an information transmission subsystem, and corresponding memories, which jointly complete functions such as radar signal sampling, signal processing, slow-moving target recognition on the sea surface, intelligent control, and transmission of recognition results.

[0079] The FPGA data reception and processing subsystem is implemented by a general-purpose image processing FPGA, and completes functions such as data decoding, sampling, data distribution format conversion, and data storage after the input of passive radar signals. Among them, the passive radar signal sampling module samples the input one-dimensional time-domain signal of the passive radar as needed through a sampling controller using an acquisition card with a general-purpose analog-to-digital conversion module to obtain discrete time-domain signals.

[0080] The Ascend AI acceleration operation and control subsystem consists of Ascend operation and control devices. The board is designed modularly, and completes functions such as storage of the target recognition algorithm model, recognition of input target samples, storage of recognition results, and interactive control with other subsystems.

[0081] The FPGA data receiving and processing subsystem sends the processed passive radar sample data to the Ascend AI acceleration operation and control subsystem through the PCIE interface to complete the determination of the target category, and synchronously records the target azimuth, coordinates, and characteristic parameters; the image data information is interacted through USB, SDI, network, and PAL interfaces, and the communication protocol is CAN; the instruction interaction information between the Ascend AI acceleration operation and control subsystem and the FPGA data receiving and processing subsystem is transmitted through the UART interface, and the image information is interacted through the PCIE interface.

[0082] The information transmission subsystem completes the wired transmission within the system and the wireless transmission function with the remote ground control station; for wireless communication within a distance of 10KM, it is implemented using an Ethernet & serial port wireless digital data link module. Its on-board transmitter can provide both network port and serial port data input and output methods simultaneously, with a working frequency of 300MHz - 6GHz, a throughput of 12Mbps, supporting the TCP / IP network protocol, and supporting point-to-point and point-to-multipoint working modes; for wireless communication beyond a distance of 10km, relying on data link and communication relay system equipment, it realizes the information transmission between the system and the ground control station.

[0083] The beneficial effects of the present invention are as follows:

[0084] 1. The method of the present invention relies on the simulation of unmanned aerial vehicle (UAV)-borne passive radar signals. Based on a large number of passive radar signal samples and a high-precision deep neural network model, an effective method for identifying slow-moving targets on the sea surface is proposed, which can provide advanced technical guidance for the construction of intelligent target recognition methods and systems driven by large samples, contribute to expanding the perception channels of slow-moving targets on the sea surface for UAV-borne platforms, and enrich the means of battlefield reconnaissance and target surveillance. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] Figure 1 It is a schematic diagram of the simulation of the passive radar electromagnetic signal sample for Target 1;

[0086] Figure 2 It is a schematic diagram of the simulation of the passive radar electromagnetic signal sample for Target 2;

[0087] Figure 3 It is a schematic diagram of the simulation of the passive radar electromagnetic signal sample for Target 3;

[0088] Figure 4 It is a schematic diagram of the passive radar electromagnetic signal sample;

[0089] Figure 5 It is a schematic diagram of the simulation index of the passive radar electromagnetic signal sample;

[0090] Figure 6 It is a schematic diagram of a sample library containing 30,000 passive radar signal samples;

[0091] Figure 7 Schematic diagram of the recognition framework for slow-moving sea targets carried by unmanned aerial vehicles (UAVs) based on passive radar signals;

[0092] Figure 8 Schematic diagram of a multi-scale one-dimensional CNN classifier for passive radar signals;

[0093] Figure 9 Training and validation loss curve of passive radar signals;

[0094] Figure 10 Training and validation accuracy curve of passive radar signals;

[0095] Figure 11 Schematic diagram of the confusion matrix of the recognition results of three types of targets;

[0096] Figure 12 Schematic diagram of the output statistics of the 10-round test results of slow-moving sea target recognition;

[0097] Figure 13 Logic function diagram of the UAV-borne slow-moving sea target recognition system based on passive radar signals;

[0098] Figure 14 Configuration function diagram of the UAV-borne slow-moving sea target recognition system based on passive radar signals. Detailed implementation method

[0099] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0100] The object of the present invention is to provide a design and implementation method for a UAV-borne slow-moving sea target recognition system based on passive radar signals, which is based on a large amount of passive radar signal data and intelligent target classification algorithms and is oriented to UAV-borne application scenarios. This method can realize the automatic recognition of typical UAV-borne slow-moving sea targets. To achieve the above object, the method of the present invention includes the construction of a passive radar simulation sample library for typical slow-moving sea targets, the design of a classifier for the target recognition system, and the design and implementation of a UAV-borne target recognition system.

[0101] The present invention does not develop the front-end signal acquisition device of the passive radar system, but uses the shipborne search radar, which is the main electromagnetic radiation source of typical slow-moving sea targets, as the source for the passive radar to obtain electromagnetic radiation signals. Based on the radar parameters and operation modes of ship targets, the passive radar signals of typical sea targets are simulated by a simulation method. On this basis, a passive radar signal sample library for typical slow-moving sea targets is established to obtain the sample basis for data-driven target recognition. Then, the sample signals are input into the target recognition system at the back end of the passive radar system to realize the recognition of typical targets, that is, the identification of target categories.

[0102] 1. Construction of a passive radar simulation sample library for typical slow-moving sea targets;

[0103] 1.1. Simulation and generation of passive radar signals for slow-moving sea targets;

[0104] To detect a target with a passive radar, it is necessary to start from two aspects: the airspace and the frequency domain. Detection in the airspace: The passive radar searches for signals within the working frequency band in the search sector. As long as the received signal reaches the sensitivity of the passive radar receiver, the passive radar can detect the target. Detection in the frequency domain: The electromagnetic radiation frequency of the radiation source must fall within the effective working frequency range of the passive radar. Only then is it possible for the passive radar to receive the target signal and detect the target. This invention belongs to frequency domain detection and mainly realizes target recognition based on the classification of electromagnetic radiation frequency signals of slow-moving sea targets.

[0105] Taking typical slow-moving sea targets as an example, their shipborne search radar equipment mainly includes:

[0106] ① A certain type A air search radar;

[0107] Operating in the E / F band, with a working frequency of 2900 - 3100 MHz, a peak power of 2400 KW, an average power of 35 KW, and a range of 407 KM (σ = 5 m 2 ) and 231 KM (σ = 1 m 2 ). The measurement accuracy is 0.6° (azimuth) and 411 M (altitude). Its antenna is a 5.48 * 5.18 M planar array, and the antenna rotation speed is 15 r / min.

[0108] ② A certain type B air search radar;

[0109] Operating in the C / D band, with a working frequency of 850 - 942 MHz, a scanning rate of 6 and 12 r / min, horizontal and vertical beam widths of 3.3° and 3.4° in azimuth respectively, a polarization mode of horizontal polarization, a transmit power of 360 KW (peak) and 13 KW (average), and a range greater than 463 KM.

[0110] ③ A certain type C sea search radar;

[0111] Operating in the G band, with a working frequency of 5400 - 5800 MHz, repetition frequencies of 2400, 1200, and 700 Hz, pulse widths of 1, 0.25, and 0.1 μs, a peak transmit power of 280 KW, a scanning period of 4 S, and an instrument-measured range of 104 KM.

[0112] ④ A certain type D target tracking radar;

[0113] Operating in the D band, with a working frequency of 950 - 1200 MHz, a peak power of 200 KW, a range of 37 KM (σ = 1 m²) and 185 KM (secondary surveillance), an antenna size of 3.28 * 8.15 * 1.93 M, and an antenna rotation speed of 15 or 30 r / min.

[0114] For a specific type of sea - surface ship target, such as:

[0115] For a large sea - surface ship target 1, the electromagnetic radiation spectrum of its ship - borne search radar mainly consists of radar signals at 0.85 - 0.94 GHz (type B radar), 0.95 - 1.2 GHz (type D radar), 2.9 - 3.1 GHz (type A radar), and 5.4 - 5.8 GHz (type C radar);

[0116] For a medium - large sea - surface ship target 2, the electromagnetic radiation spectrum of its ship - borne search radar mainly consists of radar signals at 0.85 - 0.94 GHz (type B radar) and 0.95 - 1.2 GH (type D radar).

[0117] According to the different - frequency electronic devices equipped on target 1 and target 2, the present invention simulates corresponding electromagnetic signals through signal - processing methods such as time - frequency domain transformation and pulse compression of the linear - frequency - modulated LFM signal, and uses the spectrum of the simulated electromagnetic signal as a feature for subsequent identification.

[0118] The specific method is as follows:

[0119] Suppose the passive radar received signal of target 1 mainly contains four types of radar electromagnetic radiation signals with frequencies of 0.85 - 0.94 GHz, 0.95 - 1.2 GHz, 2.9 - 3.1 GHz, and 5.4 - 5.8 GHz, which are set as sine waves with an amplitude of 1, plus Gaussian noise with a mean of 0 and a variance of 0.01, and the radar is randomly turned on.

[0120] The passive radar received signal of target 2 mainly contains radar electromagnetic radiation signals with frequencies of 0.85 - 0.94 GHz and 0.95 - 1.2 GH, which are set as sine waves with an amplitude of 1, plus Gaussian noise with a mean of 0 and a variance of 0.01, and the radar is randomly turned on.

[0121] To increase target diversity and enhance the classifier's recognition ability for negative - class samples, the present invention also simulates target 3, that is, sea - surface stationary targets such as islands and lighthouses, which are simulated using Gaussian noise with a mean of 0 and a variance of 0.01.

[0122] The specific method is as follows:

[0123] Set a random variable r, whose value is determined by a pseudo - random number uniformly distributed between 0 and 1:

[0124] r = rand(1)

[0125] Generate three types of samples, namely target 1, target 2, and target 3, according to probabilities P of 0.6, 0.3, and 0.1 respectively, that is:

[0126] P = [0.6 0.3 0.1]

[0127] Set the target frequency matrix FR as:

[0128] FR = [0.85 0.94; 0.95 1.2; 2.9 3.1; 5.4 5.8]

[0129] For target 1, the radar radiation signal received by the passive radar is the superposition of 4 sine waves with different frequencies and an amplitude of 1 from four different types of radars. Let the generated mixed frequency be Fs. Then, for the four different types of radars, Fs(j) is:

[0130] Fs(j) = FR(j,1) + r * (FR(j,2) - FR(j,1))

[0131] where j = 1, 2, 3, 4, representing the serial numbers of the radar types;

[0132] For target 2, the generated mixed frequency Fs only contains two types of radars, A and B. The radar radiation signal received by the passive radar is the superposition of 2 sine waves with different frequencies and an amplitude of 1 from radars A and B. Its target frequency matrix FR is:

[0133] FR' = [0.85 0.94; 0.95 1.2]

[0134] For the two different types of radars, Fs'(j) is:

[0135] Fs'(j) = FR'(j,1) + r * (FR'(j,2) - FR'(j,1))

[0136] where j = 1, 2, representing the serial numbers of the radar types.

[0137] Based on the mixed frequency, the calculation method for the one-dimensional time-domain waveform signal s generated by target 1 at the receiving end through simulation is:

[0138] s = sin(2π * Fs(j) * t) + randn(size(s)) * 0.1, j = 1, 2, 3, 4

[0139] where randn() is the standard normal distribution random number generation function, and size(s) is the length of the signal s.

[0140] The calculation method for the one-dimensional time-domain waveform signal s' generated by target 2 at the receiving end through simulation is:

[0141] s' = sin(2π * Fs'(j) * t) + randn(size(s')) * 0.1, j = 1, 2

[0142] where randn() is the standard normal distribution random number generation function, and size(s') is the length of the signal s'.

[0143] The frequency-domain signal F is the Fourier transform of the corresponding time-domain signal. Taking Target 1 as an example, the calculation method is as follows:

[0144] F = FFT(s)

[0145] Through corresponding programming implementation in MATLAB, the time-domain waveform and amplitude-frequency characteristics of the passive radar received signal are as Figure 1 , Figure 2 and Figure 3 .

[0146] 1.2. Passive radar signal sample noise addition and augmentation;

[0147] To more realistically simulate the electromagnetic radiation signals received by passive radar in actual application scenarios, different Gaussian noises are considered to be added to the generated time-domain waveform signals. On the one hand, this method can increase the scale of samples, achieve the augmentation of sample signals, and better support the training of deep neural network models; on the other hand, it can improve the robustness of target recognition algorithms and make the algorithms more in line with actual application requirements.

[0148] Add different levels of random noise to the passive radar time-domain signal, such as adding Gaussian noises with a mean of 0 and variances of 0.01, 0.25, 1, 1.26, 1.59, 2, 4, 25, 100, 200 respectively. The specific noise addition method is as follows: For example, directly add a standard normal distribution Gaussian noise with a mean of 0 and a variance of 0.01 (standard deviation of 0.1) to the time-domain signal s. The time-domain waveform signal s Addnoise is as follows:

[0149] s Addnoise = s + randn(size(s)) * 0.1

[0150] To quantitatively describe the intensity of the noise relative to the un-noised signal, the concept of signal-to-noise ratio (SNR) is introduced, and its calculation formula is:

[0151]

[0152] That is, SNR represents the ratio of the useful signal power to the noise power, where the signal power is also called the energy or intensity of the signal. The unit of SNR is generally in decibels (dB), and its calculation formula is:

[0153]

[0154] For a sinusoidal signal with a period of T, its power can be simplified as:

[0155]

[0156] That is, the average power of the sinusoidal signal over one period is equal to the square of the amplitude divided by 2. Considering that the passive radar signal is composed of the superposition of 4 sinusoidal waves with different frequencies and an amplitude of 1, therefore, in the present invention, the power of the signal is:

[0157] P signal = 4 * P 正弦 = 2A 2

[0158] For Gaussian white noise with a mean of 0, the noise power is the variance.

[0159] Using the passive radar signal simulation generation method in the previous section and combining with the noise addition operation, noise-free pure signal samples and corresponding noisy signal samples can be obtained. Through the SNR calculation method, the signal-to-noise ratios of the passive radar time-domain signal with different intensity levels of Gaussian noise added can be obtained as shown in Table 1.

[0160] Table 1 Signal-to-noise ratios of the passive radar time-domain signal with different intensity levels of Gaussian noise added

[0161]

[0162]

[0163] 1.3. Generate a passive radar signal data sample library for slow-moving targets on the sea surface;

[0164] Based on the passive radar signal simulation generation method, the one-dimensional time-domain waveform signal of a single target sample can be simulated. The length of the signal sequence is 10 s. By sampling at a time interval of 0.01 s, a 1001-dimensional discrete time series can be obtained, and the amplitude of each sampling point is the value of the one-dimensional time-domain waveform signal s.

[0165] To enhance the generality and convenience of sample data storage and reading, in the present invention, the discrete time series is stored in the.txt document format, using the variable-length UTF-8 encoding method, which supports Chinese characters. As Figure 4 shown is the sequence storage style of the first target sample "signal_1.txt".

[0166] By simulating a single target sample, multiple target types and a large number of samples of each type can be generated as needed to construct a passive radar signal data sample library of slow-moving targets on the sea surface. Due to the large number of target samples, it is necessary to establish an index for each sample. The index is also stored in the .txt file format in the present invention. Figure 5 The following is the sample library index "passive_list.txt" of 30,000 samples generated by target 1, target 2, and target 3 in a ratio of 6:3:1. In this index, each row of data represents the relevant information of a target sample. The data has 6 columns in total. The first column represents the target type, and the label values are target 1, target 2, and target 3; the second to fifth columns correspond to the frequency values of the four types of radars A, B, C, and D in the mixed signal of the sample; the sixth column is the storage location and sample name of each sample.

[0167] By collecting the generated passive radar signal samples into folders, you can get a sample library stored in numerical order. Figure 6 The following is a sample library "Passive 1" containing 30,000 passive radar signal samples, each of which is added with Gaussian noise with a mean of 0 and a variance of 0.01. The sample library is stored in the form of folders, and each .txt file in it is a target sample.

[0168] 2. Target recognition based on passive radar signals and multi-scale one-dimensional convolutional neural networks;

[0169] 2.1. Framework for identifying slow-moving targets on the sea surface carried by unmanned aerial vehicles based on passive radar signals;

[0170] Based on the front-end signal acquisition device of the passive radar system on UAV, a UAV-borne sea surface slow-moving target recognition framework based on passive radar signals is designed. Figure 7 In this framework, the front-end signal acquisition device of the passive radar system on the unmanned aerial vehicle collects the electromagnetic radiation signals of typical slow-moving targets on the sea surface, that is, the time domain waveform signals of the three types of targets, and based on the constructed multi-scale one-dimensional CNN classifier for passive radar signals, supervised learning is performed through the three types of target samples generated by simulation, and the different characteristics of the passive radar signals of the three types of typical sea surface targets can be automatically learned, so as to realize the recognition of the three types of typical sea surface targets, namely, target 1, target 2, and target 3, that is, the discrimination of the three types of targets.

[0171] 2.2. Construct a multi-scale one-dimensional CNN classifier for passive radar signals;

[0172] Aiming at one-dimensional passive radar time domain signals, the present invention designs and constructs a multi-scale one-dimensional convolutional neural network classifier, such as Figure 8As shown in the figure. The input end of the classifier is a one-dimensional passive radar signal, which is a one-dimensional time-domain signal with 1001 sampling points. After the signal is input, it first passes through a one-dimensional convolutional layer (conv1) and a max pooling layer (pool1), and then enters the convolutional layer paths of three scales with convolutional kernels of 3, 5, and 7 respectively for convolution and pooling operations. For example, in the 3×1 convolutional kernel layer path, the signal undergoes 3 layers of convolution + max pooling processing (conv3_1, pool3_1, conv3_2, pool3_2, conv3_3, pool3_3) respectively. The 5×1 convolutional kernel layer path (conv5_1, pool5_1, conv5_2, pool5_2, conv5_3, pool5_3) and the 7×1 convolutional kernel layer path (conv7_1, pool7_1, conv7_2, pool7_2, conv7_3, pool7_3) are similar. Finally, through the concatenation of the tensors of the three paths, it is sent to the max pooling layer (pool2). Finally, through two fully connected layers (fc1, fc2), it is sent to the softmax activation function, and the output layer in the three-class target recognition problem can be obtained.

[0173] All pooling layers of this multi-scale one-dimensional CNN classifier use max pooling. The kernel size of pooling layer pool2 is 8×8, the stride is 1, and the padding is 0. The kernel size of all other pooling layers is 2×2, the stride is 2, and the padding is 0. The fully connected layer fc1 realizes the fusion of feature matrices, which can fuse the three output feature matrices of the three path scales into one feature matrix. Therefore, the number of input features in_features is 3, and the number of output features out_features is 1. The feature matrix is input to the second fully connected layer fc2 after transforming the matrix shape. The number of input features in_features of fc2 is 3, and the number of output features out_features is 1, finally realizing the classification of three classes of targets. The detailed parameters of the model are shown in Table 2.

[0174] Table 2 Parameters of the multi-scale one-dimensional CNN classifier for passive radar signals

[0175]

[0176]

[0177] 2.3 Input and output mode for the recognition of slow-moving targets on the sea surface based on passive radar signals;

[0178] The unmanned aerial vehicle (UAV)-borne sea surface slow-moving target recognition system based on passive radar signals involved in the present invention designs a general input and output mode for the recognition of slow-moving targets on the sea surface.

[0179] At the input end of the sea surface slow moving target recognition system, based on the characteristics of passive radar signals, the present invention designs the input signal into a general one-dimensional time domain signal. According to the sampling number parameter settings generated by the simulation of the sample signal, the signal length is 1001, and this length can also be adjusted according to the sample generation requirements; for the actual radar time domain signal, in order to improve the data acquisition efficiency and reduce the data acquisition workload of the core arithmetic unit in the embedded system, the present invention can adopt an embedded data acquisition card to collect one-dimensional discrete signals of different lengths as needed. On this basis, a sample signal matrix is constructed to facilitate subsequent CNN operation.

[0180] At the output end of the sea surface slow moving target recognition system, based on the number of categories of typical sea surface slow moving targets, the present invention designs the output end into three category labels: target 1, target 2, and target 3. After Softmax operation, the probability values that the current sample belongs to the three types of targets are directly output, and the type with the largest probability value is the recognition result of the current sample; based on the number of sample categories and the training of the model with a sufficient number of samples, the output end of the present invention can also support the recognition of sea surface slow moving targets with more than three categories.

[0181] In the presentation of the recognition results of typical sea surface slow moving targets, the present invention uses a variety of methods to describe the performance of the algorithm in the training and testing processes. For example Figure 9 is the training and validation loss curve of the passive radar signal, which shows the change trends of the loss functions of the training sample set and the validation sample set during the learning process; Figure 10 is the training and validation accuracy curve of the passive radar signal, which shows the change trends of the overall correct recognition rates of the three types of targets of the training sample set and the validation sample set during the learning process; Figure 11 is the confusion matrix of the recognition results of the three types of targets, which shows the recognition accuracy of each type of target among the three types of targets and the misclassification situations of the three types of targets; to test the stability of the algorithm in multiple trainings and tests, the present invention conducts 10 rounds of sea surface slow moving target recognition training and testing. Figure 12 As shown in the figure is the output statistics of the 10 rounds of training and testing results, which are stored in the general Excel file format, including the correct rate of each round of testing, as well as the average training time and testing time of each round.

[0182] During the simulation experiment, Gaussian noise with different intensity levels (mean value is 0, variances are 0.01, 0.25, 1, 1.26, 1.5876, 2, 4, 25, 100, 200 respectively) was added to the passive radar signal. The generated sample size was 30,000, and the ratio of the three types of samples was 6:3:1. In the experiment, the data was divided into a training set, a validation set, and a test set according to the ratio of 7:2:1. Through comprehensive experiments, the recognition rate of MSCNN under different signal-to-noise ratios was investigated, and the results are shown in Table 3. It can be seen from the experimental results that the model has a 100% recognition accuracy when the signal-to-noise ratio is greater than or equal to 1; when the variance of Gaussian noise is increased to 25, the accuracy of the model on the test set drops to 0.92777, indicating that the model has strong robustness to Gaussian white noise; further increasing the noise intensity, the recognition rate of the model begins to drop significantly.

[0183] Table 3 Recognition results of slow-moving sea targets in passive radar signals with different levels of Gaussian noise introduced

[0184]

[0185] 3. Implementation method of the unmanned aerial vehicle (UAV)-borne slow-moving sea target recognition system;

[0186] 3.1 Functional framework of the UAV-borne slow-moving sea target recognition system;

[0187] For UAV-borne applications, the present invention designs a UAV-borne slow-moving sea target recognition system framework based on passive radar signals. This system includes: ① passive radar sensor module, ② passive radar signal sampling module, ③ passive radar signal processing module, ④ slow-moving sea target recognition module, ⑤ information transmission module, ⑥ intelligent control module, ⑦ target recognition result output identification module. Among them, the five modules ②③④⑤⑥ constitute the main body of the system of the present invention, and the composition and connection relationship of each module are as Figure 13 shown.

[0188] The functions of each module of the UAV-borne slow-moving sea target recognition system based on passive radar signals are as follows:

[0189] ① Passive radar sensor module;

[0190] This module is the front-end receiving device for obtaining the passive radar signals of slow-moving sea targets. It can obtain the radar electromagnetic radiation signals of the targets, that is, one-dimensional time-domain signals composed of a mixture of multiple sine waves. This signal provides test samples for the target recognition system; at the same time, this module can also obtain the azimuth information of the target relative to itself. Through coordinate transformation, the absolute position information of the target, that is, the longitude and latitude coordinate information of the target, can be obtained.

[0191] ② Passive radar signal sampling module;

[0192] This module can perform discrete sampling on the one-dimensional time-domain signals obtained by the passive radar system, collect one-dimensional discrete signals of different lengths as needed, and construct a sample signal matrix to facilitate subsequent operations and processing of target recognition algorithms. It is mainly implemented by an embedded high-speed data acquisition card, which can improve the data acquisition efficiency and reduce the data acquisition workload of the core arithmetic unit in the embedded system.

[0193] ③ Passive radar signal processing module;

[0194] This module mainly preprocesses the collected raw passive radar signals, including filtering, denoising, data arrangement, etc.

[0195] ④ Sea surface slow moving target recognition module;

[0196] This module mainly consists of a high-performance embedded arithmetic unit, a memory, high-speed interface devices, and target recognition algorithms, etc. It mainly completes the storage of the target recognition algorithm model, the storage of input samples, the recognition of input target samples, and realizes the classification of sea surface slow moving targets, that is, the determination of target categories.

[0197] ⑤ Information transmission module;

[0198] This module mainly completes the wired and wireless transmission functions of information such as target recognition data and target coordinate data; wired transmission is mainly responsible for the transmission of necessary data information inside the unmanned aerial vehicle (UAV)-borne sea surface slow moving target recognition system; wireless transmission is mainly responsible for the data information transmission between the system and the ground control station; for short-distance wireless communication, wireless digital data link sending and receiving devices can be directly used to realize the information transmission between the system and the ground control station; for long-distance wireless communication, devices such as data links and communication relays can be used to realize the information transmission between the system and the ground control station.

[0199] ⑥ Intelligent control module;

[0200] This module mainly controls the normal and orderly operation of other modules such as the passive radar sensor module, the sea surface slow moving target recognition module, and the information transmission module through the transmission of control signals in the UAV-borne sea surface slow moving target recognition system; and can perform autonomous planning, decision-making, etc. on the sea surface slow moving target detection task according to the target recognition results.

[0201] ⑦ Ground display control module;

[0202] This module is deployed on the host of the ground control station. It mainly realizes the visual output and positioning identification of the target in the digital map of the remote ground station through the wirelessly obtained target coordinates and target recognition information, and completes functions such as real-time tracking and character overlay of the target.

[0203] 3.2, Main device configuration and implementation method;

[0204] In the UAV-borne sea surface slow moving target recognition system based on passive radar signals in the present invention, its hardware implementation mainly consists of an FPGA data receiving and processing subsystem, an Ascend AI acceleration operation and control subsystem, an information transmission subsystem, and corresponding memories, jointly completing functions such as radar signal sampling, signal processing, sea surface slow moving target recognition, intelligent control, and transmission of recognition results. Its system configuration and functional connections are as Figure 14 shown.

[0205] (1) FPGA data receiving and processing subsystem;

[0206] This subsystem is implemented by a general-purpose image processing FPGA, mainly completing functions such as data decoding, sampling, data distribution format conversion, and data storage after the input of passive radar signals. The passive radar signal sampling module in it can, through a sampling controller, use an acquisition card with a general-purpose analog-to-digital converter (ADC) module to sample the input one-dimensional time-domain passive radar signal as needed to obtain discrete time-domain signals. During the simulation generation of the sea surface slow moving target passive radar signal, for a 10-second passive radar signal, the system collected 1001 data points, and the acquisition frequency of the semi-physical simulation system was approximately 100 Hz. The passive radar signal processing module, under the control of the intelligent control module and with the support of the operation and storage unit of the target recognition module, completes the preprocessing function of the collected original passive radar signal, mainly including signal filtering, noise removal, data normalization, data transformation and arrangement, etc., and finally converts the discrete sampling points into a 1×1001 data matrix (actually a 1001-dimensional vector) and stores it, similar to two-dimensional image data, for the processing of the multi-scale convolutional neural network in the target recognition module. During the sampling and processing process, the FPGA processor performs fast data interaction and read / write through the on-board 256 Mb or 512 Mb flash memory, and the processed data is stored in the 4 GB or 8 GB on-board memory.

[0207] (2) Ascend AI acceleration operation and control subsystem;

[0208] The core sea surface slow moving target recognition module in this subsystem consists of Ascend operation and control devices. The board is designed modularly, mainly completing functions such as storage of the target recognition algorithm model, recognition of input target samples, storage of recognition results, and interaction control with other subsystems. The Ascend operation and control devices are equipped with an Atlas A2 intelligent acceleration operation core, integrated with a Huawei Ascend processor, having a secure and reliable NPU architecture, with a high-performance computing power level of more than 20 TOPS, and having strong industrial stability, capable of realizing UAV-borne image recognition, image classification, etc. at the edge side. Based on the Ascend processor, the platform also integrates a MindStudio development environment and rich peripheral interfaces, facilitating quick access to development.

[0209] Under the support of the Ascend on-board control system, the FPGA data reception and processing subsystem sends the processed passive radar sample data to the Ascend Atlas A2 intelligent acceleration computing module through the PCIE interface to complete the determination of the target category, and synchronously records parameters such as the target azimuth, coordinates, and characteristics. The image data information is interacted through USB, SDI, network, and PAL interfaces, etc., and the communication protocol is CAN. The command interaction information between the Ascend AI acceleration computing and control subsystem and the FPGA data reception and processing subsystem is transmitted through the UART interface, and the image information is interacted through the PCIE interface. This subsystem can also be equipped with an embedded multimedia memory card with a relatively large capacity of 64G or 128G, which is mainly used for storing the target recognition model and recognition results, and can also be used for storing a relatively large number of training samples, facilitating the algorithm debugging of the system at the edge side.

[0210] (3) Information transmission subsystem;

[0211] The information transmission subsystem mainly completes the wired transmission within the system and the wireless transmission function with the remote ground control station. For wireless communication within a distance of 10KM, an Ethernet & serial port wireless digital data link module (NanoDDL) can be used to achieve it. Its on-board transmitter can provide both network port and serial port data input and output methods at the same time, with a working frequency of 300MHz to 6GHz, a throughput of 12Mbps, supporting all network protocols such as TCP / IP, supporting point-to-point and point-to-multipoint working modes, and having good long-distance and high-bandwidth communication performance. For wireless communication over a longer distance, the information transmission between the system and the ground control station can be achieved relying on data link and communication relay and other standard equipment.

Claims

1. A method for identifying slow-moving targets on the sea surface carried by an unmanned aerial vehicle based on passive radar signals, characterized in that, It includes the following steps: Step 1: Construction of a passive radar simulation sample library for slow-moving targets on the sea surface; Step 1-1: Simulation generation of passive radar signals for slow-moving targets on the sea surface; Step 1-1-1: Define a large sea-going ship, i.e., Target 1, whose electromagnetic radiation spectrum of the shipborne search radar consists of radar signals with frequencies of 0.85 - 0.94 GHz, 0.95 - 1.2 GHz, 2.9 - 3.1 GHz, and 5.4 - 5.8 GHz; Define a medium to large sea-going ship, i.e., Target 2, whose electromagnetic radiation spectrum of the shipborne search radar consists of radar signals with frequencies of 0.85 - 0.94 GHz and 0.95 - 1.2 GHz; According to the different frequency electronic devices equipped on Target 1 and Target 2, through time-frequency domain transformation and pulse compression of the linear frequency modulation (LFM) signal, simulate and generate the corresponding electromagnetic signals, and use the spectrum of the simulated electromagnetic signals as features; Step 1-1-2: Assume that the passive radar received signal of Target 1 contains four types of radar electromagnetic radiation signals with frequencies of 0.85 - 0.94 GHz, 0.95 - 1.2 GHz, 2.9 - 3.1 GHz, and 5.4 - 5.8 GHz, set as sine waves with an amplitude of 1, plus Gaussian noise with a mean of 0 and a variance of 0.01, and the radar is randomly turned on; The passive radar received signal of Target 2 contains radar electromagnetic radiation signals with frequencies of 0.85 - 0.94 GHz and 0.95 - 1.2 GHz, set as sine waves with an amplitude of 1, plus Gaussian noise with a mean of 0 and a variance of 0.01, and the radar is randomly turned on; Simulate Target 3, i.e., a stationary target on the sea surface, and Target 3 is simulated using Gaussian noise with a mean of 0 and a variance of 0.01; Step 1-1-3: Set a random variable r, whose value is determined by a pseudo-random number uniformly distributed between 0 and 1: r = rand(1) Generate three types of samples of Target 1, Target 2, and Target 3 respectively according to probabilities P of 0.6, 0.3, and 0.1, that is: P＝[0.60.30.1] Set the target frequency matrix FR as: FR = [0.85 0.94; 0.95 1.2; 2.9 3.1; 5.4 5.8] For Target 1, the radar radiation signal received by the passive radar is the superposition of 4 sine waves with different frequencies and amplitudes of 1 for four different types of radars. Let the generated mixed frequency be Fs, then for Fs(j) of the 4 different types of radars: Fs(j) = FR(j,1) + r*(FR(j,2) - FR(j,1)) where j = 1, 2, 3, 4, representing the serial number of the radar type; For Target 2, the generated mixed frequency Fs only contains two types of radars, A and B. The radar radiation signal received by the passive radar is the superposition of 2 sine waves with different frequencies and amplitudes of 1 for radars A and B, and its target frequency matrix FR is: FR' = [0.85 0.94; 0.95 1.2] For Fs'(j) of the 2 different types of radars: Fs'(j) = FR'(j,1) + r*(FR'(j,2) - FR'(j,1)) where j = 1, 2, representing the serial number of the radar type; Based on the mixed frequency, the calculation method for the target 1 to simulate and generate the one-dimensional time-domain waveform signal s at the receiving end is as follows: s = sin(2π*Fs(j)*t) + randn(size(s))*0.1, j = 1, 2, 3, 4 where randn() is the standard normal distribution random number generation function, and size(s) is the length of the signal s; The calculation method for the target 2 to simulate and generate the one-dimensional time-domain waveform signal s' at the receiving end is as follows: s' = sin(2π*Fs'(j)*t) + randn(size(s'))*0.1, j = 1, 2 where randn() is the standard normal distribution random number generation function, and size(s') is the length of the signal s'; The frequency-domain signal F is the Fourier transform of the corresponding time-domain signal. For target 1, the calculation method is: F = FFT(s) Step 1-2: Adding noise and augmenting the passive radar signal samples; Add random noise with different intensity levels to the passive radar time-domain signal. For target 1, the time-domain waveform signal s after adding noise is Addnoise as follows: s Addnoise = s + randn(size(s)) * 0.1 Introduce the signal-to-noise ratio SNR, and its calculation formula is: Wherein, P signal is the signal power, and P noise is the noise power; For a sinusoidal signal with a period of T, its power is simplified to: That is, the average power of the sinusoidal signal over one period is equal to the square of the amplitude divided by 2. Considering that the passive radar signal is composed of the superposition of 4 sine waves with different frequencies and an amplitude of 1, the signal power is P signal = 4 * P 正弦 = 2A 2 For Gaussian white noise with a mean of 0, the noise power is the variance; By introducing the signal-to-noise ratio SNR for calculation, the signal-to-noise ratio after adding Gaussian noise of different intensity levels to the passive radar time-domain signal is obtained; Step 1-3: Generating a database of passive radar signal data samples for slow-moving targets on the sea surface; Based on Step 1-1, simulate and obtain the one-dimensional time-domain waveform signal of a single target sample. The length of the signal sequence is 10 s, and through sampling at a time interval of 0.01 s, a 1001-dimensional discrete time series is obtained. The amplitude of each sampling point is the value of the one-dimensional time-domain waveform signal s; Store the discrete time series in the.txt document format, using the variable-length UTF-8 encoding method, which supports Chinese characters; After simulating a single target sample, generate multiple target types and samples of each type to construct a database of passive radar signal data samples for slow-moving targets on the sea surface; Establish an index for each sample, and store the index in the.txt document format; Step 2: Target recognition based on passive radar signals and a multi-scale one-dimensional convolutional neural network; The front-end signal acquisition device of the UAV-borne passive radar system collects the electromagnetic radiation signals of typical slow-moving targets on the sea surface, that is, the time-domain waveform signals of three types of targets. Based on the constructed multi-scale one-dimensional CNN classifier for passive radar signals, through supervised learning with the three types of target samples generated by simulation, automatically learn the different characteristics of the passive radar signals of three types of typical sea surface targets, and realize the recognition of three types of typical sea surface targets, namely target 1, target 2, and target 3, that is, the judgment of three types of target types; specifically as follows: Step 2-1: Construct a multi-scale one-dimensional CNN classifier for passive radar signals; The input end of the multi-scale one-dimensional CNN classifier is a one-dimensional passive radar signal, which is a one-dimensional time-domain signal with 1001 sampling points. After the signal is input, it first passes through a one-dimensional convolutional layer conv1 and a max pooling layer pool1, and then is sent to the convolutional layer paths of three scales with convolutional kernels of 3, 5, and 7 respectively for convolution and pooling operations. Each convolutional layer path also undergoes 3 layers of convolution + max pooling processing. Then, through the concatenation of the three-path tensors, it is sent to the max pooling layer pool2. Finally, through two fully connected layers (fc1, fc2), it is sent to the softmax activation function, and thus the output layer in the three-class target recognition problem is obtained. Step 2-2: Input-output mode for the recognition of slow-moving targets on the sea surface based on passive radar signals. The input signal is designed as a general one-dimensional time-domain signal. According to the sampling number parameter settings generated by simulating the sample signal, the signal length is 1001, and this length can also be adjusted according to the sample generation requirements. For the actual radar time-domain signal, an embedded data acquisition card is used to collect one-dimensional discrete signals of different lengths as needed. On this basis, a sample signal matrix is constructed to facilitate subsequent CNN operation. The output end is designed with a total of three category labels: target 1, target 2, and target 3. After Softmax operation, the probability values of the current sample belonging to the three types of targets are directly output, and the type with the largest probability value is the recognition result of the current sample.

2. The method for identifying slow-moving targets on the sea surface carried by an unmanned aerial vehicle based on passive radar signals according to claim 1, wherein, All pooling layers of the multi-scale one-dimensional CNN classifier adopt max pooling. The kernel size of the pooling layer pool2 is 8×8, the stride is 1, and the padding is 0. The kernel sizes of all other pooling layers are 2×2, the stride is 2, and the padding is 0. The fully connected layer fc1 realizes the fusion of the feature matrices, fusing the three output feature matrices of the three paths into one feature matrix. Therefore, the input feature number in_features is 3, and the output feature number out_features is 1. The feature matrix is input to the second fully connected layer fc2 after transforming the matrix shape. The input feature number in_features of fc2 is 3, and the output feature number out_features is 1, finally realizing the classification of the three types of targets.

3. A target recognition system adopting the target recognition method as described in claim 1, characterized in that, It includes a functional framework and a hardware subsystem. The functional framework includes a passive radar sensor module, a passive radar signal sampling module, a passive radar signal processing module, a slow-moving target recognition module on the sea surface, an information transmission module, an intelligent control module, and a target recognition result output identification module. The passive radar sensor module is the front-end receiving device for obtaining the passive radar signals of slow-moving targets on the sea surface. It can obtain the radar electromagnetic radiation signals of the targets, that is, one-dimensional time-domain signals composed of a mixture of multiple sine waves, which provide test samples for the target recognition system. At the same time, this module can also obtain the azimuth information of the target relative to itself, and through coordinate transformation, obtain the absolute position information of the target, that is, the longitude and latitude coordinate information of the target. The passive radar signal sampling module discretely samples the one-dimensional time-domain signal obtained by the passive radar sensor module, collects one-dimensional discrete signals of different lengths as required, and constructs a sample signal matrix to facilitate subsequent operation and processing of the target recognition algorithm; The passive radar signal processing module preprocesses the collected passive radar raw signal, including filtering, denoising, and data arrangement; The sea surface slow-moving target recognition module consists of a high-performance embedded arithmetic unit, a memory, a high-speed interface device, and a target recognition algorithm, and completes the storage of the target recognition algorithm model, the storage of input samples, and the recognition of input target samples, realizing the classification of sea surface slow-moving targets, that is, the determination of target categories; The information transmission module completes the wired and wireless transmission functions of target recognition data and target coordinate data information; wired transmission is responsible for the data information transmission within the airborne sea surface slow-moving target recognition system; wireless transmission is responsible for the data information transmission between the system and the ground control station; for short-distance wireless communication, directly use wireless digital data link sending and receiving devices to realize the information transmission between the system and the ground control station; for long-distance wireless communication, use a data link and a communication relay device to realize the information transmission between the system and the ground control station; The intelligent control module controls the normal and orderly operation of the passive radar sensor module, the sea surface slow-moving target recognition module, and the information transmission module through the transmission of control signals in the airborne sea surface slow-moving target recognition system; and can autonomously plan and make decisions on the sea surface slow-moving target detection task according to the target recognition result; The ground display control module is deployed on the host of the ground control station, and through the wirelessly obtained target coordinates and target recognition information, realizes the visual output and positioning identification of the target in the digital map of the remote ground station, and completes the real-time tracking and character overlay function of the target; The hardware subsystem includes an FPGA data reception and processing subsystem, an Ascend AI acceleration operation and control subsystem, an information transmission subsystem, and corresponding memories, which jointly complete functions such as radar signal sampling, signal processing, sea surface slow-moving target recognition, intelligent control, and transmission of recognition results; The FPGA data reception and processing subsystem is implemented by a general-purpose image processing FPGA, and completes functions such as data decoding, sampling, data distribution format conversion, and data storage after the passive radar signal is input; among them, the passive radar signal sampling module uses a sampling controller and an acquisition card with a general-purpose analog-to-digital conversion module to sample the input passive radar one-dimensional time-domain signal as required to obtain discrete time-domain signals; The Ascend AI acceleration operation and control subsystem is composed of Ascend operation and control devices, and the board adopts a modular design, and completes functions such as storage of the target recognition algorithm model, recognition of input target samples, storage of recognition results, and interactive control with other subsystems; The FPGA data receiving and processing subsystem sends the processed passive radar sample data to the Ascend AI acceleration operation and control subsystem through the PCIE interface to complete the determination of the target category, and synchronously records the target azimuth, coordinates, and characteristic parameters; the image data information is interacted through the USB, SDI, network, and PAL interfaces, and the communication protocol is CAN; the instruction interaction information between the Ascend AI acceleration operation and control subsystem and the FPGA data receiving and processing subsystem is transmitted through the UART interface, and the image information is interacted through the PCIE interface; The information transmission subsystem completes the wired transmission within the system and the wireless transmission function with the remote ground control station; for wireless communication within a distance of 10KM, an Ethernet & serial port wireless digital data link module is used to implement it. Its on-board transmitter can provide both network port and serial port data input and output methods at the same time, with a working frequency of 300MHz to 6GHz, a throughput of 12Mbps, supporting the TCP / IP network protocol, and supporting point-to-point and point-to-multipoint working modes; for wireless communication over a distance of more than 10km, relying on the data link and communication relay system equipment to realize the information transmission between the system and the ground control station.

Images