Wireless charging living foreign body detection and removal system based on millimeter wave radar

Abstract

The present invention relates to the field of wireless charging technology, and specifically discloses a wireless charging living foreign body detection system based on millimeter wave radar, which applies the millimeter wave radar's ability to detect vital signs and the method of 3D scattered point modeling of radar reflection waves to the detection of living foreign bodies at the wireless charging transmitting end of new energy vehicles; the millimeter wave radar sensor and the AI ​​deep learning model are combined to propose a high-precision detection system that can identify the type of living bodies. The advantages are: the high stability, non-contact detection, high environmental adaptability of the millimeter wave radar, and the improvement of the deep learning model training algorithm with increasing capabilities, which can efficiently collect data, obtain extremely low probability of wrong judgment, reduce maintenance costs, etc., forming a set of high-precision, high-accuracy, intelligent three-in-one new energy vehicle wireless charging station foreign body detection system.

Description

Technical Field

[0001] The present invention relates to the technical field of wireless charging, and in particular to a wireless charging living foreign body detection system based on millimeter wave radar. Background Art

[0002] The wireless charging system transmits electrical energy to the receiving device through the principle of electromagnetic induction. If there are foreign objects on the charging device, such as metal objects or magnetic objects, they may interfere with the transmission, resulting in poor power transmission or even failure to charge. If these foreign objects cannot be detected in time, it will not only waste energy, but also bring safety hazards to the device or user, and may even cause dangers such as fire or explosion.

[0003] While preventing metal objects from heating up is sufficient for other applications of wireless power transfer, inductive charging systems for electric vehicles require a special protection system against living things.

[0004] There is a large air gap between the primary and secondary coils of commercial high-power electric vehicle wireless charging systems, as well as a high-intensity magnetic field. Especially for higher transmission powers, the magnetic field strength exceeds the regulatory limits for general public exposure.

[0005] In real-world situations, living objects must be prevented from being exposed to the magnetic field under the vehicle. For example, children catch a ball that rolls under the vehicle, or the driver reaches out to pick up the car keys that have fallen to the ground. Although it is almost unthinkable that an unattended toddler could crawl under a car, this scenario still represents a relevant application for Living Object Protection (LOP) systems. In addition, there are widely discussed applications such as preventing long-term exposure of pets that sleep during charging.

[0006] Since the objects entering the air gap may be non-living objects that are not directly or indirectly threatened by the alternating magnetic field, the optimized LOP system needs to be able to distinguish between biological and non-biological objects. This can be achieved based on some specific attributes, such as temperature, shape, and the motion profile of the object. In the traditional machine vision combined with deep learning model, the performance of the deep learning model is greatly affected by the network structure and hyperparameters, especially the hyperparameters, which often require manual selection, and the image acquisition frequency and resolution of the equipment will also seriously affect the effect of machine vision. In addition, factors such as weather and light will also seriously affect the acquisition of images, thereby greatly reducing the recognition accuracy of existing technologies.

[0007] Based on the above problems, the traditional LOP system solution combining machine vision with big data artificial intelligence will not be able to meet the requirements of high precision, high stability, and high environmental adaptability in the actual commercial model. Therefore, it is particularly important to develop a LOP system based on millimeter wave radar for the Wireless Power Translation (WPT) system of new energy vehicles. Summary of the invention

[0008] The present invention provides a wireless charging living foreign body detection and expulsion system based on millimeter wave radar, and solves the technical problem of how to detect and expel living bodies in a wireless charging area with high precision and high stability based on millimeter wave radar.

[0009] In order to solve the above technical problems, the present invention provides a wireless charging living foreign body detection system based on millimeter wave radar, including a millimeter wave radar module, a digital signal processing module, and a big data and artificial intelligence comprehensive analysis module;

[0010] The millimeter wave radar module is used to transmit a detection signal for the wireless charging area at a first frequency, and receive reflected radar waves at the first frequency to obtain a first intermediate frequency signal and send it to the digital signal processing module;

[0011] The digital signal processing module is used to receive the first intermediate frequency signal and perform fast Fourier transform on it to obtain a three-dimensional FFT heat map and send it to the big data and artificial intelligence comprehensive analysis module;

[0012] The big data and artificial intelligence comprehensive analysis module is used to analyze the three-dimensional FFT heat map to see whether there is a living body invading the wireless charging area, and if so, send a frequency change signal to the millimeter wave radar module, otherwise continue to receive data for analysis;

[0013] The millimeter wave radar module is also used to, when receiving the frequency change signal, transmit a detection signal for the wireless charging area at a second frequency, and receive the reflected radar wave at the second frequency to obtain a second intermediate frequency signal and send it to the digital signal processing module;

[0014] The digital signal processing module is also used to receive the second intermediate frequency signal and convert it into a radar 3D scatter plot and send it to the big data and artificial intelligence comprehensive analysis module;

[0015] The big data and artificial intelligence comprehensive analysis module is also used to classify living organisms according to the radar 3D scatter plot to obtain the type of invading living organisms.

[0016] Preferably, the big data and artificial intelligence comprehensive analysis module analyzes the three-dimensional FFT heat map to see whether there is a living body invading the wireless charging area, specifically comprising the steps of:

[0017] A1, reading the FFT graph in the three-dimensional FFT heat map;

[0018] A2, using a constant false alarm algorithm to detect a target on the FFT graph, if the target is detected, proceed to the next step, otherwise continue detection;

[0019] A3, obtaining the distance between each target and the millimeter wave radar module, and determining whether the minimum distance therein is within the detection range, if so, proceeding to the next step, if not, returning to step A1 to read the FFT graph of the next cycle;

[0020] A4, calculate the emission energy of the target corresponding to the minimum distance, and determine whether its reflected energy is between the preset minimum threshold and maximum threshold. If yes, proceed to the next step, otherwise return to step A1 to read the FFT graph of the next cycle;

[0021] A5. Calculate the phase change of the target corresponding to the minimum distance when the motion displacement is ΔR. If the phase change Δφ b If it is a preset value, it indicates that a living body has invaded the wireless charging area, otherwise it indicates that the person invading the wireless charging area is not a living body.

[0022] Preferably, in step A5, the phase shift amount Δφ b The default values ​​are: λ represents the wavelength corresponding to the radar wave.

[0023] Preferably, the big data and artificial intelligence comprehensive analysis module is also used to classify living bodies according to the radar 3D scatter plot, specifically comprising the steps of:

[0024] B1. Using a point cloud registration algorithm to stitch multiple local point clouds in the radar 3D scattergram into a global three-dimensional point cloud;

[0025] B2. Fitting the global three-dimensional point cloud using a RANSAC algorithm to obtain a RANSAC image; fitting the global three-dimensional point cloud using a three-dimensional SAR algorithm to obtain a SAR image;

[0026] B3. Input the RANSAC image into a first deep learning network to obtain a first classification result; input the SAR image into a second deep learning network with the same structure to obtain a second classification result;

[0027] B4. Comprehensively consider the first classification result and the second classification result to obtain a final classification result.

[0028] Preferably, in step B2, the three-dimensional SAR algorithm used is a range Doppler algorithm, and its algorithm flow includes the steps of:

[0029] B21, performing distance dimension FFT and orientation dimension FFT on the global three-dimensional point cloud respectively to convert the data into a two-dimensional frequency domain;

[0030] B22, perform range-dimensional pulse compression and secondary range compression in the two-dimensional frequency domain;

[0031] B23, perform IFFT on the distance dimension;

[0032] B24, performing motion correction and azimuth matching filtering on the distance dimension IFFT;

[0033] B25, perform azimuth dimension IFFT;

[0034] B26. Obtain imaging results.

[0035] Preferably, after step B4, the method further comprises the following steps:

[0036] B5. Re-send the final classification result to the first deep learning network and the second deep learning network to train the first deep learning network and the second deep learning network in a mutually optimized manner.

[0037] Preferably, the millimeter wave radar module is installed on the side of the wireless charging station, where it can directly illuminate the wireless charging area, and the digital signal processing module and the big data and artificial intelligence comprehensive analysis module are deployed in the cloud;

[0038] The first frequency selects a high-frequency electromagnetic wave with a transmission frequency between 77 GHz and 81 GHz;

[0039] The second frequency is a low-frequency electromagnetic wave with a transmission frequency of 24 GHz.

[0040] The present invention also provides a millimeter wave radar-based wireless charging living foreign body expulsion system, comprising the big data and artificial intelligence comprehensive analysis module as described above, and also comprising a wireless charging system controller and an expulsion module;

[0041] The big data and artificial intelligence comprehensive analysis module is also used to send a power reduction instruction to the wireless charging system controller when analyzing that a living body has invaded the wireless charging area; and is also used to send the type of the invading living body to the expulsion module when the type of the invading living body is classified;

[0042] The wireless charging system controller is used to control the wireless charging system to reduce power after receiving the instruction to reduce power;

[0043] The expulsion module is used to send a corresponding expulsion signal according to the type of the received invading living body.

[0044] Preferably, the big data and artificial intelligence comprehensive analysis module is also used to send a command to stop charging to the wireless charging system controller when it is analyzed that there are still living beings in the wireless charging area within a preset time period and when it is classified that the invading living being is a human, and send a command to start charging to the wireless charging system controller after the living being leaves the wireless charging area.

[0045] Preferably, when the invading living body is a human, the expulsion module sends out a voice warning signal; when the invading living body is an animal, the expulsion module sends out a corresponding ultrasonic signal for the animal species and simultaneously sends out a voice expulsion signal.

[0046] The wireless charging living foreign body detection and removal system based on millimeter wave radar provided by the present invention has the following beneficial effects:

[0047] 1. Use millimeter-wave radar, take advantage of the micro-Doppler effect of living things, and detect the tiny movements of living things to detect human beings or animals, so as to realize the identification of living and non-living things in the wireless charging transmitter. Compared with the traditional monitoring form combining visual cameras and artificial intelligence, millimeter-wave radar has the advantages of high environmental adaptability, reliability and real-time performance;

[0048] 2. The millimeter wave radar sends millimeter wave signals and receives reflected signals to obtain information such as the distance, speed and direction of the target. When processing this information, it is converted into a scatter plot, and the algorithm is used to filter the three-dimensional scatter plot and build a three-dimensional model to build a deep learning model. The model is continuously trained to improve the judgment accuracy of the model through a large amount of data processing, so that it has the ability of unmanned detection;

[0049] 3. As a new type of monitoring equipment, millimeter wave radar is added to the wireless charging foreign object detection system, which improves the system's environmental adaptability, system stability and flexibility. Combined with the training of deep learning models, it has a good fit with the comprehensive monitoring needs mentioned above. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a schematic diagram of the basic components of millimeter wave radar and the basic operation flow of radar;

[0051] Figure 2 A simple FFT transformation is performed on the preset 27-30 GHz millimeter wave radar module and then the FFT heat map is fed back to the PC host computer;

[0052] Figure 3 This is the basic flow chart for millimeter wave radar vital signs detection;

[0053] Figure 4 Schematic diagram of energy detection for millimeter wave radar;

[0054] Figure 5 The two-dimensional straight line fitting diagram: (a) two-dimensional plane point set; (b) the best straight line model obtained by fitting;

[0055] Figure 6 It is the flow chart of SnapNet;

[0056] Figure 7 It is the SegCloud flow chart;

[0057] Figure 8 This is the PointNet network structure diagram;

[0058] Fig. 9 Simplified flowchart for liveness detection through CNN;

[0059] Fig.10 Simplified flowchart for animal dataset identification;

[0060] Fig.11 It is the flow chart of RD algorithm;

[0061] Fig.12 It is the flow chart of CS algorithm;

[0062] Fig.13 It is the flow chart of BP algorithm;

[0063] Fig.14 This is the topology diagram of the CLC-S type WPT system. DETAILED DESCRIPTION

[0064] The following specifically illustrates the implementation mode of the present invention in conjunction with the accompanying drawings. The embodiments are provided for illustrative purposes only and cannot be understood as limiting the present invention. The accompanying drawings are provided for reference and illustration only and do not constitute a limitation on the scope of patent protection of the present invention, because many changes can be made to the present invention without departing from the spirit and scope of the present invention.

[0065] The embodiment of the present invention first provides a wireless charging living foreign body detection system based on millimeter wave radar, including a millimeter wave radar module, a digital signal processing module, and a big data and artificial intelligence comprehensive analysis module;

[0066] The millimeter wave radar module is used to transmit a detection signal for the wireless charging area at a first frequency, and receive reflected radar waves at the first frequency to obtain a first intermediate frequency signal and send it to the digital signal processing module;

[0067] The digital signal processing module is used to receive the first intermediate frequency signal and perform a fast Fourier transform on it to obtain a three-dimensional FFT heat map and send it to the big data and artificial intelligence comprehensive analysis module;

[0068] The big data and artificial intelligence comprehensive analysis module is used to analyze the three-dimensional FFT heat map to see if there is a living body invading the wireless charging area. If so, a frequency change signal is sent to the millimeter wave radar module. If not, data continues to be received for analysis.

[0069] The millimeter wave radar module is also used to, when receiving the frequency change signal, transmit a detection signal for the wireless charging area at a second frequency, and receive the reflected radar wave at the second frequency to obtain a second intermediate frequency signal and send it to the digital signal processing module;

[0070] The digital signal processing module is also used to receive the second intermediate frequency signal and convert it into a radar 3D scatter plot and send it to the big data and artificial intelligence comprehensive analysis module;

[0071] The big data and artificial intelligence comprehensive analysis module is also used to classify living organisms based on the radar 3D scatter plot to obtain the type of invading living organisms.

[0072] This embodiment applies the millimeter wave radar's ability to detect vital signs and the method of 3D scattered point modeling of radar reflection waves to the detection of living foreign objects at the wireless charging transmitter of new energy vehicles; the millimeter wave radar sensor and the AI ​​deep learning model are combined to propose a high-precision detection system that can identify the type of living things. The advantages are: the high stability, non-contact detection, high environmental adaptability of the millimeter wave radar, and the improvement of the deep learning model training algorithm with increasing capabilities, which can efficiently collect data, obtain extremely low probability of misjudgment, reduce maintenance costs, etc., forming a high-precision, high-accuracy, intelligent three-in-one new energy vehicle wireless charging station foreign object detection system.

[0073] Among them, the millimeter-wave radar module is installed on the side of the wireless charging station, where it can directly illuminate the wireless charging area. The digital signal processing module and the big data and artificial intelligence comprehensive analysis module are deployed in the cloud.

[0074] like Figure 1 As shown in the figure, the millimeter wave radar module is mainly composed of a TX (transmit) antenna, an RX (receive) antenna, a synthesizer, a mixer (Mixer) and an ADC sampler. When the millimeter wave radar module starts working, the synthesizer first generates a linear frequency modulation pulse and transmits it by the transmitting antenna (TX antenna). The transmitted wave is reflected by the object to form a reflected linear modulation pulse and is captured by the receiving antenna (RX antenna). The mixer combines the TX signal x1 represented by equation (1) and the RX signal x2 represented by equation (2) to generate an intermediate frequency (IF) signal x represented by equation (3). out .

[0075]

[0076]

[0077] Among them, ω1, Respectively represent the instantaneous frequency and phase of the TX signal, ω2, Indicates the instantaneous frequency and phase of the RX signal. Intermediate frequency (IF) signal x out There is an instantaneous frequency that is equal to the difference between the instantaneous frequencies of the two input sinusoidal functions. The intermediate frequency (IF) signal x out The phase of the mixer is equal to the phase difference of the two input signals. The digital signal processing module (DSP) performs FFT processing on the intermediate frequency IF signal output by the mixer to obtain the following Figure 2 3D FFT heat map shown).

[0078] x out =sin[(ω1-ω2)t+(φ1-φ2)] (3)

[0079] The big data and artificial intelligence comprehensive analysis module has a complete set of algorithm processes for vital signs detection. Figure 3 As shown, the specific steps include:

[0080] A1. Read the FFT graph in the 3D FFT heat map;

[0081] A2. Use the constant false alarm algorithm to detect the target on the FFT graph. If the target is detected, proceed to the next step. If not, continue the detection.

[0082] A3, obtaining the distance between each target and the millimeter wave radar module, and determining whether the minimum distance is within the detection range, if so, proceeding to the next step, otherwise returning to step A1 to read the FFT graph of the next cycle;

[0083] A4, calculate the emission energy of the target corresponding to the minimum distance, and determine whether its reflected energy is between the preset minimum threshold and maximum threshold. If yes, proceed to the next step, otherwise return to step A1 to read the FFT graph of the next cycle;

[0084] A5. Calculate the phase change of the target corresponding to the minimum distance when the motion displacement is ΔR. If the phase change Δφ b If it is the preset value, it indicates that a living body has invaded the wireless charging area, otherwise it indicates that the intruder is not a living body.

[0085] The energy detection in step A4 is to filter out situations that are obviously inconsistent. For example, at the same distance, if a non-metallic object similar to the one in the photo is placed in front of the radar, its reflected energy will be very weak; on the contrary, if a metal object such as a monitor or a mobile phone is placed in front of the radar, its reflected energy will be very strong. Energy detection is to avoid such special situations.

[0086] The principle of energy detection is as follows Figure 4 As shown, Figure 4The horizontal axis is the frequency domain, which represents the distance from the radar, and the vertical axis is the relative energy amplitude. The broken line where point A is located is the result of FFT modulus of the original IF signal of the radar. The target point detected after target detection is A. The sum of the amplitudes of point A and the three points on its left and right is the energy of object A. Figure 4 It is represented as point B above. Figure 4 The other two curves represent the two thresholds (maximum energy and minimum energy) of the energy reflected by the living body in front of the radar. These two thresholds are derived from the data collected under different distance conditions. If point B is between the two thresholds, it means that the target has passed the energy detection and can enter the next step, i.e., the stationary detection in step A5.

[0087] Regarding stationary detection, the principle is described as follows.

[0088] After the millimeter wave radar module receives the target reflected signal, it obtains the original signal through a series of signal preprocessing. Formula (4) is the radar transmitting FMCW wave signal equation, where A T is the transmitted signal amplitude, φ(t) is the phase noise of the transmitter, f c is the starting frequency of the linear frequency modulation signal, Β is the bandwidth of the linear frequency modulation signal, and t represents time. Assuming that the target is at a distance R0 and has a time-varying displacement x(t), the distance change between the radar and the target can be expressed as R(t)=R0+x(t), where x(t) is the chest motion function. The radar echo signal is the reflected signal after encountering the target, and there is a t d The time delay, t d =2R(t) / c. d Substituting into equation (4) we get the echo signal equation (5), where α is the attenuation of the amplitude during signal propagation.

[0089]

[0090]

[0091] In the radar receiver, the radar receiving signal is coherently mixed with the transmitting signal, and then the signal is obtained after I / Q demodulation. Equation (6) is the signal obtained after coherent mixing and I / Q demodulation, A R is the received signal amplitude, x is x(t). The frequency of the intermediate frequency signal obtained by mixing is f b , where the phase change due to the time delay caused by the distance is recorded as φ b (t), expressed as formula (8), where π(B / T c )t d 2The part is very small in practical application, so b (t) can be ignored. Δφ(t)=φ(tt d ) is the phase noise.

[0092] y(x)=αA R e j(x) (6)

[0093]

[0094]

[0095] In the process of non-contact vital sign signal detection, the distance function R(t) = R0 + x(t) is used to further simplify the radar echo signal to obtain x(t), which is equation (7), where t d =2R0 / c (c represents the propagation speed of radar signal in the air), so The radar signal uses FMCW wave, and the duration of the entire signal is 0<t<T c , so the frequency of the entire signal time period (B / T c )t<<f c , so in the theoretical analysis of radar echo signals, (2B / cT c ) part can be ignored in formula (8), and formulas (6), (7), and (8) are transformed into formulas (9), (10), and (11) as a whole.

[0096] y(x)=αA R e j(x) (9)

[0097]

[0098]

[0099] The radar echo function is simplified to obtain the chest motion function x(t) represented by equation (10), and the final radar echo original signal is the intermediate frequency signal obtained by mixing. Equation (10) only contains the phase change and phase noise obtained by the forward and backward movement of the chest wall. The subsequent radar echo signal modeling by software will be based on equation (10) as the main theoretical basis. The corresponding y(x) shown in equation (9) is transformed into the following equation (12):

[0100]

[0101] Among them, λ is the wavelength corresponding to the millimeter wave radar wave.

[0102] The echo signal obtained by formula (10) can be transformed into the phase value of the signal by Fourier transform. The radar measures the phase change of the FMCW echo signal within the target distance unit to measure the tiny vibration. Formula (13) is the relationship between the slight change in distance and the phase. When the target motion displacement is ΔR, the phase difference before and after is Δφ b This is also the core of detecting vital signs signals through millimeter wave radar. In the process of vital signs signal detection, the phase obtained by radar echo is used to represent the chest motion function x(t). The wavelength λ of the millimeter wave radar wave in this experiment is 3.8mm. When ΔR has a displacement of 1mm, the corresponding phase change Δφ b =π.

[0103]

[0104] The above detection method can be used to determine whether the intruding object is a living body. If the feedback result is yes, the radar switches to the 24GHz low-frequency millimeter wave band to perform radar wave scanning on the intruding target. The millimeter wave radar scatter plot is a graph of the intensity of the scattered echo signal received by the millimeter wave radar and time. First, the millimeter wave radar uses the TX low-frequency transmitting antenna to transmit 24GHz low-frequency millimeter waves, and then receives the electromagnetic wave signal scattered back by the target object. These reflected signals are digitized, and by distinguishing the strength of each different signal, different coordinate points can be displayed three-dimensionally at different positions in the radar 3D scatter plot. This information is presented in the form of a scattered point cloud to form a millimeter wave radar scatter plot.

[0105] The big data and artificial intelligence comprehensive analysis module is also used to classify living organisms based on the radar 3D scatter plot, which specifically includes the following steps:

[0106] B1. Use the point cloud registration algorithm to stitch multiple local point clouds in the radar 3D scatter plot into a global 3D point cloud;

[0107] B2. Use the RANSAC algorithm to fit the global three-dimensional point cloud to obtain a RANSAC image; use the three-dimensional SAR algorithm to fit the global three-dimensional point cloud to obtain a SAR image;

[0108] B3. Input the RANSAC image into the first deep learning network to obtain a first classification result; input the SAR image into the second deep learning network with the same structure to obtain a second classification result;

[0109] B4. Comprehensively consider the first classification result and the second classification result to obtain the final classification result.

[0110] For step B1, the point cloud in the millimeter-wave radar scatter plot is generally a small-scale local point cloud, and a point cloud registration algorithm is needed to stitch multiple local point clouds into a global three-dimensional point cloud of the scene. The most commonly used point cloud registration algorithms are the Normal Distribution Transform (NDF) algorithm and the Iterative Closest Point (ICP) algorithm. The core idea is to perform feature matching through the overlapping parts between different local point clouds. NDF assumes that the point cloud obeys a normal distribution and solves a spatial transformation involving rotation and translation so that a point cloud is most likely to be located on the scanning plane of another point cloud after transformation. ICP solves a spatial transformation involving rotation and translation so that the sum of the distances between each point and the nearest point of another reference point is minimized after the transformation.

[0111] For step B2, the points in the data set are divided into "Inliers" and "Outliers". "Inliers" refer to data that fit the model and are useful for model fitting; "Outliers" refer to data that do not conform to the consistency of the model description and are useless for model fitting, that is, noise points in the data set. The RANSAC algorithm has good robustness and can calculate parameters that conform to a certain geometric model in a given set of noisy data.

[0112] The following is an example of a straight line fitting to illustrate the application of RANSAC. The example is as follows: fit the straight line model parameters with the highest support for the sample points from a two-dimensional plane point set. The point set contains both in-points and out-points. The general process of fitting the parameters is as follows:

[0113] Step 1: Assume that the model to be fitted is a straight line, and the equation is y=ax+b. Only two points are needed to determine a straight line. First, two points are randomly selected from the data to solve the equation, and the values ​​of parameters a and b are obtained, thus determining a preliminary hypothesized model.

[0114] Step 2: Set the distance threshold to d max , traverse all the points in the data set, calculate the distance d from each point to the line y = ax + b, if the distance d < d max , then the point is considered to be on the straight line y=ax+b, that is, the point is an internal point. Otherwise, the point is determined to be an external point. For each straight line equation determined in the previous step, the number of internal points N is counted;

[0115] Step 3: According to the size of the data set, set a parameter S to determine the support of the data set for the current model. If N>S, save the current model parameters, otherwise continue to the next round of iteration;

[0116] Step 4: Repeat the above steps and iterate continuously. The model corresponding to the maximum value N is the best fitting model of the straight line.

[0117] like Figure 5 As shown, Figure 5 (a) is a point set on a two-dimensional plane containing both internal and external points. Figure 5 (b) is the best straight line model obtained by fitting using the RANSAC algorithm. Figure 5 It can be seen from the results that the existence of outliers does not affect the final model fitting results.

[0118] Based on a process similar to these steps, the global 3D point cloud is fitted using the RANSAC algorithm to obtain a RANSAC image.

[0119] The basic idea of ​​the RANSAC algorithm can be described as: given a data set P containing inliers and outliers, and a model M that requires at least n samples to solve the position parameters, first randomly select n points from the data set to form a subset X1, which is used to solve the parameters of the model M to obtain the initial model M1; then fit the points in the data set P to the model M1, and count the number of inliers N. The principle of adaptation is that the distance from the point to the model M1 is less than a given threshold Δ. The larger the value of N, the greater the support of the subset composed of inliers for the model. After continuous iteration, the model M corresponding to the largest value of N is the best fitting model.

[0120] Traditional point cloud classification methods are based on artificial features designed by experts according to geometric constraints and statistical constraints. Deep learning-based methods can use the network to automatically extract features from data, which can achieve better classification results than traditional methods. Deep learning was originally used for image processing, and three-dimensional point clouds are disordered, irregular, and unstructured compared to images. Therefore, point clouds need to be transformed accordingly before entering the network. According to the different formats of input network point cloud data, deep learning-based point cloud classification methods can be divided into three categories: multi-view-based, voxel-based, and point-based methods.

[0121] The core idea of ​​the multi-view based method is to use Figure 2 D images are used to represent 3D point cloud data. SnapNet is one of the few point cloud classification networks based on multiple views. The principle process is as follows Figure 6 As shown, the classification steps for point clouds are as follows:

[0122] Step 1: Point cloud preprocessing, extracting the point cloud, calculating the features of the points (such as normals or local noise) and generating a mesh;

[0123] Step 2: Use the SegNet-based deep segmentation network and fusion residual correction to provide a semantic label for each pixel on the input pair of RGB images and depth maps;

[0124] Step 3: Project the semanticized image into three-dimensional space to obtain the classification result of the three-dimensional point cloud.

[0125] The voxel-based method uses voxels as input, which solves the problem of point cloud disorder and unstructuredness. Similar to pixels in a two-dimensional image, the voxelized point cloud can be directly processed by a 3D convolutional network. SegCloud is the most representative voxel-based point cloud classification network. The principle process is as follows: Figure 7 As shown, the steps for processing point clouds are as follows:

[0126] Step 1: voxelization of point cloud;

[0127] Step 2: Use a 3D fully convolutional network to generate coarse downsampled voxel labels;

[0128] Step 3: Use trilinear interpolation to convert the rough output of voxels to the original point cloud;

[0129] Step 4: Infer the results obtained in the previous step through the three-dimensional fully connected conditional random field to obtain the classification results of all points.

[0130] Although SegCloud uses voxels as processing objects, which greatly improves efficiency compared to directly processing point clouds, its computational complexity and memory consumption are still large. Some scholars have used octree structures to replace voxels, which can better improve the efficiency of point cloud processing.

[0131] Point-based methods directly use point clouds as input, and the network directly processes the input 3D points. The most representative of this type of method is PointNet, which is the first deep network that directly performs point cloud classification based on point clouds. Its network structure is as follows: Figure 8 As shown in the figure, the scene point cloud is used as input, the input point cloud is transformed and the feature transformation is performed, and then the maximum pooling layer is used to integrate the features together. The final segmentation network is an extension of the previous classification network, which connects the global features with the local features, and finally outputs k categories and the scores corresponding to each category.

[0132] In subsequent research, the network optimized and improved based on PointNet, the point cloud classification network with convolution, and the point cloud classification network with graph structure achieved better classification results. Compared with the traditional point cloud classification method, the point cloud classification method based on deep learning has greatly improved the accuracy and efficiency.

[0133] In order to improve the recognition accuracy of the system, while running the above RANSAC algorithm, the system will also put the data into the three-dimensional SAR algorithm for modeling. That is, the three-dimensional SAR algorithm is used to fit the global three-dimensional point cloud to obtain a SAR image. The two sets of algorithms simultaneously calculate the output results, and compare and output the results, which can not only improve the accuracy of the output results; the output results are re-sent into the deep learning model, and the model can be continuously optimized during the system's use, so that the recognition accuracy rate is continuously improved. Therefore, after step B4, the following steps are also included:

[0134] B5. Send the final classification result back to the first deep learning network and the second deep learning network to train the first deep learning network and the second deep learning network in a mutually optimized manner.

[0135] SAR (Synthetic Aperture Radar) is a technology used to produce radar images. It uses platforms such as aircraft or satellites to transmit radar pulses and receive reflected signals to obtain ground information. SAR algorithms are a series of mathematical methods for processing SAR data to obtain images. SAR images are characterized by high resolution, all-weather, and all-day, but there are also some difficulties, such as multiple reflections and speckle noise. In order to overcome these problems, complex algorithms are needed to process data and obtain high-quality images. Common SAR algorithms include:

[0136] 1. Range Doppler algorithm: This is one of the most commonly used SAR processing algorithms. It uses the relationship between range and Doppler shift to synthesize images. This algorithm can provide high-resolution images, but requires a lot of computing resources.

[0137] 2. SAR algorithm based on time domain analysis: This type of algorithm converts SAR data into the time domain and obtains SAR images by processing time series data. The most commonly used algorithms are the Omega-K algorithm and the Chirp Scaling algorithm.

[0138] 3. SAR algorithms based on polarization analysis: This type of algorithm uses the polarization information of SAR data for processing to obtain more accurate classification and analysis of objects. Common algorithms include polarization interferometry and polarization scattering decomposition.

[0139] 4. SAR algorithm based on compressed sensing: This type of algorithm uses compressed sensing theory to achieve compressed sampling and reconstruction of SAR signals, thereby reducing the cost of data acquisition and processing. Common algorithms include compressed sensing based on sparse representation and compressed sensing based on low-rank representation.

[0140] In this system, we first analyze the target motion model, establish the echo model, analyze the echo spectrum, then introduce the classic RD, CS, and BP imaging algorithms, and finally model the target scattering center to obtain the target simulated SAR image.

[0141] The echo signal form of targets at the same distance and different azimuths is the same, but the time delay in the azimuth is different. Because the time delay does not affect the spectrum range of the signal, the echoes of targets at the same distance and different azimuths overlap in the range-Doppler domain. The Range-Doopler (RD) algorithm uses this feature to perform range migration correction in the range-Doppler domain, and each range unit only needs to be corrected once.

[0142] The RD algorithm process is as follows Fig.11 As shown, the specific steps include:

[0143] B21, perform distance dimension FFT and orientation dimension FFT on the global three-dimensional point cloud respectively to convert the data into two-dimensional frequency domain;

[0144] B22, perform range-dimensional pulse compression and secondary range compression in the two-dimensional frequency domain;

[0145] B23, perform IFFT on the distance dimension;

[0146] B24, performing motion correction and azimuth matching filtering on the distance dimension IFFT;

[0147] B25, perform azimuth dimension IFFT;

[0148] B26. Obtain imaging results.

[0149] First, perform FFT on the range dimension and azimuth dimension respectively to convert the raw data into the two-dimensional frequency domain. Then, perform range dimension pulse compression in the two-dimensional frequency domain. Since the range dimension defocus is more serious under large oblique angles, secondary range compression is required. Next, perform motion correction and azimuth matched filtering on the range dimension IFFT in the range-Doppler domain. Finally, perform azimuth IFFT to obtain the imaging result.

[0150] The following is an explanation of the formulas related to range matching filtering, secondary range compression, range migration correction, and azimuth matching filtering.

[0151] (1) Distance Dimension Matched Filtering

[0152] Based on equations (2)-(6), the distance dimension of the echo signal is expressed as:

[0153]

[0154] Among them, k r is the leveling rate, S r() is the distance dimension of the echo signal.

[0155] Therefore, the reference signal of the matched filter is:

[0156] h r (τ) = w τ (τ)exp(jπk r τ 2 ) (15)

[0157] After matched filtering, the echo is a narrow pulse in the form of Sinc, which achieves range-dimensional focusing.

[0158] (2) Secondary distance compression

[0159] Observe the two-dimensional frequency domain expression (16), the phase From the leveling rate k r The phase component of the transmitted signal is offset by distance matching filtering, and the other phase component θ(f t ,f τ ) The third term of the expanded formula (17) is generated by the range-azimuth coupling and is defined as θ SRC (f t ,f τ ):

[0160]

[0161] Secondary distance compression on θ SRC (f t ,f τ ) compensation, frequency response H SRC (f t ,f τ )for:

[0162]

[0163] In the case of direct side view, the Doppler frequency f t Mainly concentrated around 0Hz, phase θ SRC (f t ,f τ ) has little effect. However, when the oblique angle is large, the frequency f t Improve, phase θ SRC (f t ,f τ ) becomes larger, so k m and the original modulation frequency k r If the secondary distance compression is not performed, it will lead to phase mismatch and cause distance dimension defocusing.

[0164] (3) Range Migration Correction

[0165] In the case of direct side view, the migration in the range-Doppler domain can be simplified to equation (18):

[0166]

[0167] When the oblique angle is large, simplification cannot be performed and the motion correction is:

[0168]

[0169] (4) Azimuth Matched Filtering

[0170] The Doppler signal expression is as follows:

[0171]

[0172] The frequency response of the azimuth matched filter is given by equation (21):

[0173]

[0174] The RD algorithm has a high resolution, and each scattering point has a good focusing effect in range and azimuth, with low side lobes. However, the side lobes are slightly raised in the oblique scene, which is due to the derivation of the range Doppler spectrum and the phase θ SRC (f t ,f τ ) is expanded, the high-order phase is ignored. As the oblique angle increases, the impact of this approximation on imaging becomes more obvious, so the RD algorithm needs to be modified for large front oblique imaging.

[0175] Chirp Scaling (CS) multiplies the linear frequency modulation signal with the scaling function to change the zero frequency position. After matched filtering, the pulse is time-shifted to achieve range migration correction. The CS algorithm avoids interpolation operations and reduces time complexity. The following first introduces the basic principle of CS, and then combines the SAR echo expression to derive the scaling equation.

[0176] The single pulse linear frequency modulation receiving signal is as follows:

[0177] s(τ)=exp[jπk r (τ-τ0) 2 ] (twenty two)

[0178] Assume the variable scale equation is (23):

[0179]

[0180] s(τ) multiplied by the scaling equation s cs (τ) Formula (24):

[0181]

[0182] It can be seen from equation (24) that the zero frequency appears at τ0 / 1+α after scaling. Because the peak value of the linear frequency modulation signal after pulse compression is located at the zero frequency position, the scaling equation can cause the compressed pulse to produce a time shift, and the offset Δτ is equation (25):

[0183]

[0184] Δτ is proportional to τ0. After scaling, the modulation frequency becomes (1+α)k r , a change in modulation frequency will result in a change in bandwidth.

[0185] The process of CS algorithm is as follows Fig.12 As shown. First, perform an azimuth FFT and use the variable scaling equation (26) to complement the RCMC. Then, perform a range FFT, use equation (27) to complete the range matched filtering, and use equation (28) to perform phase compensation to complete the consistent RCMC. Next, perform a range IFFT, transform back to the range-Doppler domain, use equation (29) to complete the azimuth matched filtering, and use equation (30) for phase compensation. Finally, perform an azimuth IFFT to complete the imaging.

[0186]

[0187]

[0188]

[0189]

[0190]

[0191] The back projection (BP) algorithm images in the time domain and maps the echo after range compression to the user-defined imaging surface. The BP algorithm is relatively simple and flexible, and the imaging surface is not restricted. It can image the target under any geometric relationship. However, the algorithm has a high time complexity and requires traversal and interpolation of all pixel points on the imaging surface.

[0192] The expression after echo distance dimension matched filtering (31) is:

[0193]

[0194] In the formula, q r [τ-2R(t) / c] represents the range-dimensional compressed pulse. Azimuth focusing can be achieved by coherently accumulating the echo units corresponding to the distance R(t).

[0195] Define the imaging surface I and divide the imaging surface into grids. First, perform distance-dimensional pulse compression, and then image all cells in I cyclically. Take cell (i, j) as an example, calculate the distance change history R of the celli,j (n), n is the pulse number. i,j (n) Get the echo data position k = (2R i,j (n) / c)f s k may not be an integer, and it is necessary to interpolate the echo at position k. In formula (31), the phase -4π(f c / c)R(t) changes with time, and the phase needs to be compensated during coherent accumulation. In summary, the imaging value of cell (i, j) is formula (32):

[0196]

[0197] The BP algorithm process is as follows Fig.13 As shown. First, determine the imaging space and divide the imaging surface into cells. Then perform distance-dimensional pulse compression on the original echo. Next, image all cells of the imaging surface cyclically:

[0198] (1) Calculate the distance R between the imaging point and the SAR platform at all slow time sampling moments i,j (n);

[0199] (2) By R i,j (n) Get the distance sampling point subscript k and interpolate in the distance dimension to get the echo value s r (n,k);

[0200] (3) For s r (n, k) Doppler phase compensation is performed, and each slow time echo is coherently superimposed;

[0201] (4) After the cycle is completed for all cells, the image of the imaging area is obtained.

[0202] In the results of the front and side views, pulse broadening or increased side lobe power in the range dimension occurs. This is because the imaging points are relatively concentrated, and the echoes are aliased in the time domain. When the BP algorithm performs echo projection on a certain imaging unit, it may introduce echoes from other imaging units at the same time. In the oblique view scene, the difference in the slant distance between imaging points is larger than that in the front and side views, so the degree of echo time domain aliasing is less, and the imaging effect is better.

[0203] The BP algorithm is more flexible, can freely define the imaging surface, and has a higher resolution. However, it requires a lot of interpolation operations, and the time complexity reaches o(N 3 ), which brings difficulties to real-time imaging.

[0204] The steps of SAR image slice simulation are as follows:

[0205] (1) Set the SAR platform operating parameters;

[0206] (2) Calculate the RCS of the sampling point on the target surface according to the incident angle;

[0207] (3) Calculate the echo power based on the transmitted signal parameters, scattering point RCS, distance and other information to generate the echo;

[0208] (4) Add the echoes from all scattering points and superimpose the noise;

[0209] (5) Use imaging algorithms to process the echoes to obtain SAR images.

[0210] Machine learning and deep learning have become the trend of SAR image classification and recognition. This system uses classic machine learning algorithms and convolutional neural networks to classify aircraft and vehicle SAR images. Live target images are generated through simulation, and animal images are derived from the ImageNet public dataset. Based on the deep learning method, two convolutional neural networks are designed to classify SAR images. Classification experiments are conducted on ImageNet animal targets to analyze the performance of the algorithm.

[0211] As a deep learning model with high learning efficiency and easy training, CNN is the most commonly used in animal recognition and is constantly improving. Based on CNN, people have enhanced the feature extraction of images by alternating and optimizing structures such as convolutional layers, pooling layers, and fully connected layers, and have enhanced learning capabilities by adjusting the number of network layers, and further trained to improve recognition performance. In 2014, Chen et al. found that the learning ability of CNN is significantly better than the traditional Bag of visual words (BOW) model, but the accuracy of animal species classification through CNN is only 38%. In 2016, Okafor et al. improved the CNN architecture by reducing the number of neurons in the fully connected layer based on the Wild-Anim dataset, simplified the recognition process, and improved the recognition efficiency. Based on the Snapshot Serengeti dataset (the largest existing animal labeling dataset), Norouzzadeh et al. used a two-stage multi-task learning detection method to identify 48 species and analyze other characteristics such as their behavior, with an accuracy rate of more than 93.8%. Shi et al. identified 40 Siberian tigers by constructing a DCNN (Deep Convoluted Neural Network) framework, with an accuracy rate of 93.5%, and significantly shortened the running time compared to other neural networks. Based on the Darknet-53 framework, Lamaojie et al. realized the recognition of some livestock animal images through data enhancement, with an accuracy rate of 87.9%. Wang Wencheng et al. used the ResNet50 network framework to classify and identify 10 species of fish, with an accuracy rate of 93.3%. Ma Mengyuan used the DCNN network to classify 70 species of Lepidoptera insects, with a recognition rate of 99.8%. Guo et al. identified 41 species of primates based on Tri-AI technology, with an accuracy rate of up to 94.1%. This technology is not only suitable for multi-species recognition, but also for night data analysis. In addition, CNN can be combined with other neural network architectures, such as RNN's LSTM, GAN, etc., to enhance feature extraction capabilities, further optimize the network structure, and improve recognition accuracy. The CNN-based animal recognition process can be seen in Fig. 9 .

[0212] There are three main methods for processing streaming data streams using neural networks: one is to use multiple sequence fusion, that is, to use CNN to extract modeling features of each frame and input them into the LSTM network, and finally output multiple labels; the second is to use a dual-stream architecture, that is, spatial stream and temporal motion stream, which are classified through multiple convolutional layers and finally the results are fused, which is the current mainstream route for streaming data stream target recognition; the third is to use three-dimensional convolution (3D Convolutional Neural Network, 3DCNN) method, which requires preprocessing of the video, and then use 3D convolution kernels to extract features and output results. The basic target recognition simplified process is as follows: Fig.10 .

[0213] Compare the results obtained by the above RANSAC algorithm and SAR algorithm, select the results with greater probability, and return the results to the deep learning model of the two algorithms for identification and modeling; continuously train the deep learning model in a mutually optimized manner to achieve the effect of training the model in use. This method can not only expand the training set for model training, but also continuously improve the reliability of the model output results. Finally, high precision and high accuracy can be achieved.

[0214] Based on the above wireless charging living foreign body detection system, an embodiment of the present invention further provides a wireless charging living foreign body removal system based on millimeter wave radar, including the above big data and artificial intelligence comprehensive analysis module, and also including a wireless charging system controller and a removal module;

[0215] The big data and artificial intelligence comprehensive analysis module is also used to send a power reduction instruction to the wireless charging system controller when analyzing that a living organism has invaded the wireless charging area; it is also used to send the type of the invading living organism to the expulsion module when the type of the invading living organism is classified;

[0216] The wireless charging system controller is used to control the wireless charging system to reduce power after receiving an instruction to reduce power;

[0217] The expulsion module is used to send out a corresponding expulsion signal according to the type of the received invading living body.

[0218] The big data and artificial intelligence comprehensive analysis module is also used to send a stop charging instruction to the wireless charging system controller when it is analyzed that there are still living beings in the wireless charging area within a preset time period and when it is classified that the invading living being is a human. After the living being leaves the wireless charging area, a start charging instruction is sent to the wireless charging system controller.

[0219] When the intruding living being is a human, the expulsion module sends out a voice warning signal (for example: Danger! Danger! Please leave the current wireless charging area as soon as possible; or various sounds such as beeps and whistles that can drive people away as soon as possible). When the intruding living being is an animal, the expulsion module sends out a corresponding ultrasonic signal for the animal species and a voice expulsion signal (human voice to drive away animals). Taking advantage of the sensitivity of animals to high-frequency waves and their different reactions to sound waves of different frequencies, the ultrasonic waves of specific frequencies are emitted to influence the behavior of animals and make them escape from the wireless charging area.

[0220] Fig.14 The topology diagram of the CLC-S type WPT system is shown in Figure 2. If you need to reduce the power or turn off the wireless charging, you can adjust the primary DC-DC power supply voltage or turn off the DC-DC to achieve the function.

[0221] In summary, the wireless charging living foreign body detection and removal system based on millimeter wave radar provided by the embodiment of the present invention has the following beneficial effects:

[0222] 1. Use millimeter-wave radar, take advantage of the micro-Doppler effect of living things, and detect the tiny movements of living things to detect human beings or animals, so as to realize the identification of living and non-living things in the wireless charging transmitter. Compared with the traditional monitoring form combining visual cameras and artificial intelligence, millimeter-wave radar has the advantages of high environmental adaptability, reliability and real-time performance;

[0223] 2. The millimeter wave radar sends millimeter wave signals and receives reflected signals to obtain information such as the distance, speed and direction of the target. When processing this information, it is converted into a scatter plot, and the algorithm is used to filter the three-dimensional scatter plot and build a three-dimensional model to build a deep learning model. The model is continuously trained to improve the judgment accuracy of the model through a large amount of data processing, so that it has the ability of unmanned detection;

[0224] 3. As a new type of monitoring equipment, millimeter wave radar is added to the wireless charging foreign object detection system, which improves the system's environmental adaptability, system stability and flexibility. Combined with the training of deep learning models, it has a good fit with the comprehensive monitoring needs mentioned above.

[0225] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.

Claims

1. A wireless charging living foreign body detection system based on millimeter wave radar, characterized in that: Including millimeter wave radar module, digital signal processing module, big data and artificial intelligence comprehensive analysis module; The millimeter wave radar module is used to transmit a detection signal for the wireless charging area at a first frequency, and receive reflected radar waves at the first frequency to obtain a first intermediate frequency signal and send it to the digital signal processing module; The digital signal processing module is used to receive the first intermediate frequency signal and perform fast Fourier transform on it to obtain a three-dimensional FFT heat map and send it to the big data and artificial intelligence comprehensive analysis module; The big data and artificial intelligence comprehensive analysis module is used to analyze the three-dimensional FFT heat map to see whether there is a living body invading the wireless charging area, and if so, send a frequency change signal to the millimeter wave radar module, otherwise continue to receive data for analysis; The big data and artificial intelligence comprehensive analysis module analyzes the three-dimensional FFT heat map to see whether there is a living body invading the wireless charging area, specifically including the steps of: A1, reading the FFT graph in the three-dimensional FFT heat map; A2, using a constant false alarm algorithm to detect a target on the FFT graph, if the target is detected, proceed to the next step, otherwise continue detection; A3, obtaining the distance between each target and the millimeter wave radar module, and determining whether the minimum distance therein is within the detection range, if so, proceeding to the next step, if not, returning to step A1 to read the FFT graph of the next cycle; A4, calculate the emission energy of the target corresponding to the minimum distance, and determine whether its reflected energy is between the preset minimum threshold and maximum threshold. If yes, proceed to the next step, otherwise return to step A1 to read the FFT graph of the next cycle; A5. Calculate the phase change of the target corresponding to the minimum distance when the motion displacement is ΔR. If the phase change Δφ b If it is a preset value, it indicates that a living body has invaded the wireless charging area, otherwise it indicates that the intruder is not a living body; The millimeter wave radar module is also used to, when receiving the frequency change signal, transmit a detection signal for the wireless charging area at a second frequency, and receive the reflected radar wave at the second frequency to obtain a second intermediate frequency signal and send it to the digital signal processing module; The digital signal processing module is also used to receive the second intermediate frequency signal and convert it into a radar 3D scatter plot and send it to the big data and artificial intelligence comprehensive analysis module; The big data and artificial intelligence comprehensive analysis module is also used to classify living organisms according to the radar 3D scatter plot to obtain the type of invading living organisms.

2. The millimeter wave radar-based wireless charging living foreign body detection system according to claim 1, characterized in that: In step A5, the phase change Δφ b The default values ​​are: λ represents the wavelength corresponding to the radar wave.

3. The millimeter wave radar-based wireless charging living foreign body detection system according to claim 1, characterized in that: The big data and artificial intelligence comprehensive analysis module is also used to classify living bodies according to the radar 3D scatter plot, specifically including the steps of: B1. Using a point cloud registration algorithm to stitch multiple local point clouds in the radar 3D scattergram into a global three-dimensional point cloud; B2. Fitting the global three-dimensional point cloud using a RANSAC algorithm to obtain a RANSAC image; fitting the global three-dimensional point cloud using a three-dimensional SAR algorithm to obtain a SAR image; B3. Inputting the RANSAC image into a first deep learning network to obtain a first classification result; Inputting the SAR image into a second deep learning network with the same structure to obtain a second classification result; B4. Comprehensively consider the first classification result and the second classification result to obtain a final classification result.

4. The millimeter wave radar-based wireless charging living foreign body detection system according to claim 3 is characterized in that: In step B2, the three-dimensional SAR algorithm used is a range Doppler algorithm, and its algorithm flow includes the following steps: B21, performing distance dimension FFT and orientation dimension FFT on the global three-dimensional point cloud respectively to convert the data into a two-dimensional frequency domain; B22, perform range-dimensional pulse compression and secondary range compression in the two-dimensional frequency domain; B23, perform IFFT on the distance dimension; B24, performing motion correction and azimuth matching filtering on the distance dimension IFFT; B25, perform azimuth dimension IFFT; B26. Obtain imaging results.

5. The millimeter wave radar-based wireless charging living foreign body detection system according to claim 3 is characterized in that: After step B4, the method further comprises the following steps: B5. Re-send the final classification result to the first deep learning network and the second deep learning network to train the first deep learning network and the second deep learning network in a mutually optimized manner.

6. The millimeter wave radar-based wireless charging living foreign body detection system according to claim 1, characterized in that: The millimeter wave radar module is installed on the side of the wireless charging station, where it can directly illuminate the wireless charging area, and the digital signal processing module and the big data and artificial intelligence comprehensive analysis module are deployed in the cloud; The first frequency selects a high-frequency electromagnetic wave with a transmission frequency between 77 GHz and 81 GHz; The second frequency is a low-frequency electromagnetic wave with a transmission frequency of 24 GHz.

7. A wireless charging living foreign body removal system based on millimeter wave radar, characterized by: It comprises the wireless charging living foreign body detection system based on millimeter wave radar as claimed in any one of claims 1 to 6, and also comprises a wireless charging system controller and a drive-away module; The big data and artificial intelligence comprehensive analysis module is also used to send a power reduction instruction to the wireless charging system controller when analyzing that a living body has invaded the wireless charging area; and is also used to send the type of the invading living body to the expulsion module when the type of the invading living body is classified; The wireless charging system controller is used to control the wireless charging system to reduce power after receiving the instruction to reduce power; The expulsion module is used to send a corresponding expulsion signal according to the type of the received invading living body.

8. The wireless charging living foreign body removal system based on millimeter wave radar according to claim 7 is characterized in that: The big data and artificial intelligence comprehensive analysis module is also used to send a stop charging instruction to the wireless charging system controller when it is analyzed that there are still living beings in the wireless charging area within a preset time period and when it is classified that the invading living being is a human, and send a start charging instruction to the wireless charging system controller after the living being leaves the wireless charging area.

9. The wireless charging living foreign body removal system based on millimeter wave radar according to claim 8 is characterized in that: When the invading living body is a human, the expulsion module sends out a voice warning signal; when the invading living body is an animal, the expulsion module sends out a corresponding ultrasonic signal for the animal species and a voice expulsion signal at the same time.

Images