Millimeter wave radar-based tiny meteorological change real-time monitoring method

Abstract

The invention discloses a micro meteorological change real-time monitoring method based on a millimeter wave radar, and relates to the technical field of meteorological monitoring. The method comprises the following steps: carrying out dynamic acquisition on millimeter wave radar signals and carrying out coordinate system conversion; converting the converted information into a programmable chirp signal through a DDS (Direct Digital Synthesizer) chip; performing space-time adaptive filtering processing on the received signal to eliminate interference; abnormal phase points are detected, and a dynamic noise base model is established; inputting the preprocessed phase data into a pre-trained DPGN network, and outputting a weather result map; a lightweight engine is deployed at a radar end to process the weather result map; when it is detected that the weather gradient exceeds a threshold value, weather early warning is triggered. The adaptive chirp waveform modulation is adopted, the frequency slope and the pulse width of the radar transmitting signal are dynamically adjusted according to the real-time environmental noise level, the meteorological monitoring sensitivity and the small change detection efficiency are improved, and the environmental noise interference is optimized.

Description

Technical Field

[0001] The present invention belongs to the technical field of meteorological monitoring, and particularly relates to a real-time monitoring method for minute meteorological changes based on a millimeter-wave radar. Background Art

[0002] With the frequent occurrence of extreme weather events, the demand for real-time monitoring of minute meteorological disturbances (such as microburst, local wind shear, urban heat island effect, etc.) is becoming increasingly urgent. The traditional meteorological monitoring technology mainly has the following technical bottlenecks: (1) Insufficient detection sensitivity Existing meteorological radars (such as X-band Doppler radars) are limited by wavelength characteristics (3 - 5 cm), and have limited detection capabilities for micro-meteorological disturbances at sub-kilometer scales (diameter < 500 m). According to the WMO meteorological observation standard, the minimum recognizable threshold of the existing system in wind speed mutation monitoring is about 2 m / s, and it is unable to effectively capture sudden micro-scale dangerous weather (such as a microburst with a diameter of 50 - 100 m) in key areas such as airports and ports.

[0003] (2) Limited accuracy of multi-parameter inversion Traditional methods estimate meteorological parameters based on the linear regression model (Z - V algorithm) of radar reflectivity factor (Z) and Doppler velocity (V), and it is difficult to separate the coupling effects of parameters such as temperature, humidity, and air pressure. Research shows that under complex terrain conditions, the inversion error of the traditional algorithm for relative humidity can reach ±15%, resulting in inaccurate prediction of disturbance propagation.

[0004] (3) Weak environmental anti-interference ability Existing millimeter-wave radar systems use fixed waveform parameters (such as standard chirp signals) and are prone to generating false echoes in complex environments such as heavy rainfall and electromagnetic interference. Experimental data shows that when the rainfall intensity exceeds 30 mm / h, the false alarm rate of traditional CFAR detection will increase from 5% to more than 40%, seriously affecting the monitoring reliability.

[0005] (4) Low real-time processing efficiency Existing solutions mostly adopt the "radar acquisition + cloud processing" architecture, and the data transmission and processing delay generally exceeds 1 minute. The evolution speed of micro-meteorological disturbances can reach 10 - 20 m / s, and the existing system is difficult to generate warning signals in a timely manner, resulting in a warning lag of up to 6 minutes in a wind shear accident at an airport in 2018.

[0006] In recent years, although there have been research attempts to introduce deep learning into meteorological monitoring, there are still two major technical defects: ① The network model is not deeply coupled with the radar physical observation equation, resulting in the lack of physical interpretability of the prediction results; ② The hardware cooperation problem of millimeter-wave signal processing and neural network calculation has not been solved, and the real-time performance at the edge end is difficult to break through the second-level threshold. Summary of the Invention

[0007] The object of the present invention is to provide a real-time monitoring method for minute meteorological changes based on a millimeter-wave radar. By adopting adaptive chirp waveform modulation and using a deep phase gradient network, the phase perturbation components at different spatial scales are separated through a cascaded U-Net structure, and the frequency slope and pulse width of the radar transmission signal are dynamically adjusted according to the real-time environmental noise level, solving the problems of insufficient meteorological monitoring sensitivity, lag in detecting minute changes, and environmental noise interference of existing millimeter-wave radars.

[0008] To solve the above technical problems, the present invention is realized through the following technical solutions: The present invention is a real-time monitoring method for minute meteorological changes based on a millimeter-wave radar, including the following steps: Step S1: Dynamically collect millimeter-wave radar signals and convert the radar signal scan data from the polar coordinate system to the X-Y plane rectangular coordinate system; Step S2: Generate a programmable chirp signal from the converted information through a DDS chip; Step S3: Perform space-time adaptive filtering processing on the received signal to eliminate interference; Step S4: Detect abnormal phase points and establish a dynamic noise floor model; Step S5: Input the preprocessed phase data into a pre-trained DPGN network, and the network outputs a weather result map; Step S6: Deploy a lightweight engine at the radar end to process the weather result map; Step S7: When it is detected that the weather gradient exceeds the threshold, trigger a weather warning.

[0009] As a preferred technical solution, in step S1, the formula for dynamically collecting millimeter-wave radar signals and converting them from the polar coordinate system to the rectangular plane coordinate system is as follows: ; In the formula, respectively represent the projected longitude and latitude under the conversion from polar coordinates to rectangular coordinates, represents the longitude and latitude of the radar base station, represents the radar body storage yard, represents the radar elevation angle, represents the direction angle, A represents the latitude distance, and B represents the longitude distance.

[0010] As a preferred technical solution, in step S2, a 77 GHz millimeter-wave radar (bandwidth 4 GHz) is used, 8 groups of MIMO arrays are deployed in the vertical direction, a programmable chirp signal is generated through a DDS chip, and the waveform parameters (pulse width adjustable from 50 to 200 ns) are adjusted according to the real-time signal-to-noise ratio.

[0011] As a preferred technical solution, the DDS chip adopts ADI AD9914; when the DDS chip dynamically controls the signal-to-noise ratio to adjust the waveform parameters, the radar receiving end calculates the background noise power in real time and calculates the signal-to-noise ratio; when the calculated signal-to-noise ratio is less than 15 dB, the pulse width expansion mode is triggered, and T gradually increases from 100 ns to 200 ns; when the calculated signal-to-noise ratio is greater than 25 dB, the high-resolution mode is started, T is reduced to 50 ns, and at the same time, the ADC sampling frequency is increased; the specific chirp waveform modulation formula is as follows: ; In the formula, represents the complex form of the time-domain radar signal, represents the time-varying amplitude modulation function, the complex exponential form, which characterizes the phase modulation component, represents the carrier frequency phase term, is the noise feedback modulation coefficient, represents the dynamic phase perturbation term.

[0012] As a preferred technical solution, in step S3, the space-time adaptive filtering processing flow is as follows: Step S31: Construct a three-dimensional space-time snapshot matrix through space diversity reception , in the formula, is the number of array elements, is the number of pulses, is the number of range cells; Step S32: Based on the building feature shaking characteristics, construct a dynamic interference space-time steering vector, and the formula is as follows: , in the formula, is the space steering vector, is the time steering vector, represents the Kronecker product; Step S33: Screen the interference sample area without target signals in the Kepler region to obtain the interference covariance matrix : ; In the formula, L is the training sample, H is the conjugate transpose, represents the i-th space-time snapshot vector, represents the conjugate transpose of ; Step S34: Solve the STAP optimal weight vector W for dimensionality reduction processing, and the specific calculation formula is as follows: ; in the formula, is the target signal steering vector, denotes the joint covariance matrix of interference and noise, H denotes the conjugate transpose element of the matrix, and -1 denotes the matrix inversion operation; the dimensionality reduction process uses the multistage Wiener filter (MSWF) to reduce the computational complexity from to ; Step S35: Obtain the signal after interference suppression through space-time weighting processing , ; Step S36: Based on the phase gradient network, perform secondary suppression on the micro-motion interference remaining after filtering to improve the signal-to-interference-plus-noise ratio by more than 15 dB.

[0013] As a preferred technical solution, in step S4, the specific process of detecting abnormal phase points and establishing a dynamic noise floor model is as follows: Step S41: Construct a multi-dimensional feature space, input the original phase matrix , and calculate the five-dimensional feature vector of each phase point. The specific formula is as follows: ; In the formula, denotes the spatial coordinate index, denotes the time series index, denotes the spectral entropy feature parameter, denotes the original phase data in three-dimensional space, respectively denote the spatial gradient of the phase on the x-axis, the spatial gradient of the phase on the y-axis, and the gradient of the phase change over time; This feature vector significantly improves the detection sensitivity of the millimeter-wave radar to micro-meteorological events (such as local wind shear and sudden changes in temperature and humidity) by fusing spatio-temporal gradients and spectral characteristics, and can achieve an accuracy of 0.1 mm in building structure deformation monitoring; Step S42: Construct a spatio-temporal probability map, and calculate the abnormal prior probability of each network point based on the historical data within the sliding window; the prior probability calculation formula is as follows: ; In the formula, is the attenuation coefficient, is the number of times determined to be abnormal within the past frames; converting historical abnormal records into probability guidance can improve the effective sample extraction efficiency by more than 40%; Step S43: Parallelly fit three types of candidate models, and calculate the weighted residuals for each model; the three types of candidate models include a linear phase drift model, a quadratic fluctuation model, and a vortex propagation model; calculate the weighted residuals for each model, and break through the limitations of traditional single-model assumptions through physically driven multi-model competition; Step S44: Automatically adjust the internal and external thresholds according to the welding noise base, and determine whether it is an abnormal point; the calculation formula is as follows: ; In the formula, Q3 is the third quartile, and IQR is the interquartile range; the specific formula for determining whether it is an abnormal point is as follows: If the residual , it is determined as an inlier; if , start the spectral entropy verification; if , it is directly determined as an abnormal point; Step S45: Construct a noise tensor, and construct the inlier data of continuous K frames into a four-dimensional tensor , and generate a dynamic base; the calculation formula of the dynamic base is as follows: ; In the formula, represents the noise base value of the position in the three-dimensional space at time t, represents the sum of r from 1 to R, respectively represent the distribution pattern of the rth spatial basis vector in the i direction, the distribution pattern of the rth spatial basis vector in the j direction, and the response value of the rth temporal basis function at time t. The noise base model can complete the response to sudden disturbances (such as bird flocks) within 3 scanning periods (about 1.5 seconds); Step S46: For the detected abnormal electric set, establish a spatio-temporal correlation graph for abnormal propagation backtracking, use the Louvain algorithm to detect closely related abnormal clusters, remove isolated noise points, and verify whether each abnormal cluster conforms to the fluid continuity equation, and the clusters that do not meet the physical constraints are removed.

[0014] As a preferred technical solution, in the step S5, the DPGN network separates the phase perturbation components at different spatial scales through a cascaded U-Net structure; the DPGN consists of three-level networks, specifically including: The first-level network of the DPGN network extracts the original phase matrix of the millimeter-wave reflection signal ; The second-level network of the DPGN network reversely calculates the meteorological parameter perturbation through the Doppler-refractive index coupling equation ; The third-level network of the DPGN network establishes a perturbation propagation model by combining LSTM time series prediction.

[0015] As a preferred technical solution, in the step S5, the DPGN network separates the phase perturbation components at different spatial scales through a cascaded U-Net structure; the DPGN consists of three-level networks, specifically including: The first - stage network of the DPGN network extracts the original phase matrix of the millimeter - wave reflection signal ; The second - stage network of the DPGN network reversely calculates the meteorological parameter perturbations through the Doppler - refractive index coupling equation ; Based on the propagation characteristics of electromagnetic waves in inhomogeneous media, the Doppler - refraction establishes the coupling equation between the refractive index perturbation and the Doppler velocity v: ; In the formula, n0 is the average atmospheric refractive index, represents the air flow velocity, c represents the speed of light; Decompose into: temperature humidity atmospheric pressure function: ; In the formula, is the atmospheric medium sensitivity coefficient; Input feature: Receive the phase gradient tensor output by the first - stage network , combined with the original radar Doppler spectrum to construct a four - dimensional input feature (space×time×spectrum);‌ The third - stage network of the DPGN network establishes a perturbation propagation model by combining LSTM time - series prediction; The three - stage network receives the four - dimensional meteorological perturbation tensor output by the second - stage network, , where the spatial resolution is 10m*10m and the time step is 30 seconds / frame; The LSTM adopts a bidirectional peephole LSTM, inserts a spatio - temporal attention module in the encoder - decoder structure, and the encoder extracts the hidden state of the historical sequence and calculates the attention weight: ; In the formula, Q is the decoder state and d is the dimension parameter.

[0016] As a preferred technical solution, when training the DPGN network, the training data set adopts typical scenarios including typhoon eyes, urban heat islands, etc.; The DPGN network outputs a weather result map; The weather result map includes a micro - air - pressure fluctuation distribution map, a water - vapor condensation nucleus diffusion prediction vector field, and a perturbation propagation heat map.

[0017] As a preferred technical solution, in step S6, the lightweight engine processing flow at the radar end is as follows: Step S61: Perform orthogonal mixing mirror interference suppression on the millimeter - wave original signal, use Hilbert transform to restore the phase consistency of the I / Q channels, and eliminate more than 90% of the signal distortion; Step S62: Establish a dynamic baseline compensation model, and update the environmental noise floor in real time through a sliding window (500 ms) to compensate for the phase shift caused by temperature drift; Step S63: Convert the multi-channel radar echo into a four-dimensional tensor structure (range × azimuth × elevation × Doppler), and achieve a standard input size of 32×32×16×64 through zero padding; Step S64: Construct a circular buffer using a double-buffer mechanism to ensure that the data acquisition and processing pipeline latency < 5 ms; Step S65: Deploy the density of the cascaded filter bank to extract data features; Step S66: Construct a cascaded network architecture, and achieve parallelism in the three stages of preprocessing - inference - postprocessing through an asynchronous pipeline design; The cascaded network structure includes: Pre-network: Improve U-Net to extract the barometric pressure fluctuation characteristics of a 10m×10m grid; Core network: Spatiotemporal convolutional GRU predicts the disturbance propagation path in the next 5 minutes; Post-processing network: Conditional generative adversarial network (CGAN) optimizes the visualization effect of the heat map; Step S67: Establish a dynamic conversion model between the polar coordinate system and the Cartesian coordinate system to achieve sub-meter alignment of the radar point cloud and the GIS map, and use the Hungarian algorithm for multi-frame data association with an azimuth angle resolution accuracy of 0.1°; Step S68: Automatically adjust the detection sensitivity according to the environmental visibility, and update the model weights in real time through an online learning module to adapt to sudden meteorological condition changes.

[0018] As a preferred technical solution, in step S7, detect the feature-level abnormal information, calculate the microbarometric pressure gradient, extract the curl feature of the water vapor diffusion vector field, detect the local vortex generation points to establish an adaptive threshold model based on historical data, and perform the division of the warning level; and deploy an online learning module to correct the model parameters using the warning results (update period ≤ 10 min), and synchronize the multi-node knowledge bases through the federated learning framework to improve the edge collaborative warning ability.

[0019] The present invention has the following beneficial effects: (1) By adopting adaptive chirp waveform modulation, the present invention dynamically adjusts the frequency slope and pulse width of the radar transmission signal according to the real-time environmental noise level, uses a deep phase gradient network to separate the phase disturbance components of different spatial scales, dynamically adjusts the frequency slope and pulse width of the radar transmission signal according to the real-time environmental noise level, improves the meteorological monitoring sensitivity and the detection efficiency of minute changes, and optimizes the environmental noise interference.

[0020] (2) The present invention breaks through the limitations of traditional software noise reduction through hardware-level real-time waveform optimization, achieves the optimal matching of signal transmission parameters and electromagnetic environment at the hardware level, improves the system sensitivity to -110 dBm, and combines 8 groups of vertical MIMO arrays with dynamic chirp technology to establish the temperature gradient monitoring ability of 0.1 °C / m within a range of 300 meters. The fast reconfiguration feature of the DDS chip (<10 ns switching time) supports the millisecond-level response to sudden meteorological events.

[0021] (3) Through dynamic interference suppression, the present invention can eliminate the Doppler spread caused by building sway, achieve full-band coverage from overall building sway to local vibration using MSWF, and implement 1 ms-level delay processing on the FPGA to meet the second-level refresh rate requirement for meteorological monitoring. (4) Through the combination of spatio-temporal feature reconstruction, hybrid optimization algorithm and online learning mechanism, the present invention realizes high-precision dynamic modeling of the propagation process of micro-meteorological disturbances, embeds physical field features into the LSTM gating mechanism, and breaks through the interpretability bottleneck of traditional pure data-driven models. Of course, it is not necessary for any product implementing the present invention to achieve all the above-mentioned advantages simultaneously. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for describing the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention, and those of ordinary skill in the art can obtain other drawings based on these drawings without creative efforts.

[0023] Figure 1 It is a flowchart of a real-time monitoring method for micro-meteorological changes based on millimeter-wave radar of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0024] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0025] In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0026] In order to make the purpose, technical solutions and advantages of the present application more clear, the following is combined with the attached Figure 1Examples are provided to further elaborate on this application. It should be understood that the specific examples described herein are for purposes of explaining this application only and are not intended to limit this application.

[0027] Please refer to Figure 1 As shown, the present invention is a real-time monitoring method for minute meteorological changes based on a millimeter-wave radar, including the following steps: Step S1: Dynamically collect millimeter-wave radar signals and convert the radar signal scan data from the polar coordinate system to the X-Y plane rectangular coordinate system; Step S2: Generate a programmable chirp signal from the converted information through a DDS chip; Step S3: Perform space-time adaptive filtering on the received signal to eliminate interference; Step S4: Detect abnormal phase points and establish a dynamic noise floor model; Step S5: Input the preprocessed phase data into a pre-trained DPGN network, and the network outputs a weather result map; Step S6: Deploy a lightweight engine at the radar end to process the weather result map; Step S7: When the detected weather gradient exceeds the threshold, trigger a weather warning.

[0028] In step S1, the formula for dynamically collecting millimeter-wave radar signals and converting them from the polar coordinate system to the rectangular plane coordinate system is as follows: ; In the formula, respectively represent the projected longitude and latitude under the conversion from polar coordinates to rectangular coordinates, represents the longitude and latitude of the radar base station, represents the radar body storage yard, represents the radar elevation angle, represents the direction angle, A represents the latitude distance, and B represents the longitude distance.

[0029] In step S2, a 77GHz millimeter-wave radar (bandwidth 4GHz) is used, 8 groups of MIMO arrays are deployed in the vertical direction, a programmable chirp signal is generated through a DDS chip, and the waveform parameters (pulse width adjustable from 50 - 200ns) are adjusted according to the real-time signal-to-noise ratio.

[0030] The 77GHz millimeter-wave radar includes an array topology structure and a radio frequency front-end architecture; Among them, the array topology structure is that 8 groups of MIMO arrays are deployed in the vertical direction, and each group of arrays contains 24 transmitting units (Tx) and 32 receiving units (Rx), and they are arranged in a vertical staggered manner: the transmitting array spacing: ; the receiving array spacing: ; Spatial diversity is achieved through orthogonal coding to form 8 independent virtual channels, enhancing the vertical angle resolution to 0.5°; The RF front-end architecture includes a transmit link and a receive link; The transmit link includes: Baseband chip: TI AWR2243 (supporting 4GHz instantaneous bandwidth); Power amplifier: Qorvo QPF7255 (output power 20dBm, efficiency 38%); Phase calibration: Integrated temperature compensation circuit, phase error <0.5°.

[0031] The receive link includes: Low-noise amplifier: NXP BGA725L6 (noise figure 2.1dB); Mixer: ADIHMC773A (conversion loss 6dB, IP1dB = 15dBm); ADC sampling: 14 bits @ 1.2GSPS, quantization noise -75dBFS.

[0032] The chirp parameter data model is: ; In the formula, is the frequency modulation slope, , and the pulse width (50 - 200ns) is adjustable.

[0033] Set the waveform reconstruction mechanism: Initialize the pulse width , and the frequency modulation slope .

[0034] The real-time adjustment rule is: ; where = 0.7 is the empirical coefficient, SNR is calculated through a sliding window, represents the preset signal-to-noise ratio threshold, represents the real-time measured environmental signal-to-noise ratio.

[0035] The DDS chip uses ADI AD9914; When the DDS chip dynamically controls the signal-to-noise ratio to adjust the waveform parameters, the radar receiving end calculates the background noise power in real time, and the calculation formula is: ; Calculate the signal-to-noise ratio, and the calculation formula is: , the update period is set to 1ms; When the calculated signal-to-noise ratio is less than 15dB, the pulse width expansion mode is triggered, and T gradually increases from 100ns to 200ns; When the calculated signal-to-noise ratio is greater than 25dB, the high-resolution mode is started, T is reduced to 50ns, and at the same time, the ADC sampling frequency is increased to 2.4GSPS; The specific chirp waveform modulation formula is as follows: ; In the formula, represents the complex form of the time-domain radar signal, represents the time-varying amplitude modulation function, The complex exponential form represents the phase modulation component. It represents the carrier frequency phase term. is the noise feedback modulation coefficient. It represents the dynamic phase perturbation term.

[0036] This design converts environmental noise into waveform modulation resources, breaks through the limitations of the traditional fixed modulation mode of radar signals, realizes the collaborative optimization of signal-to-noise ratio and resolution, and is especially suitable for the micro-meteorological disturbance detection scenario.

[0037] In step S3, the space-time adaptive filtering processing flow is as follows: Step S31: Adopt a dual-polarized MIMO array (such as a 24×24 element layout), and construct a three-dimensional space-time snapshot matrix through spatial diversity reception. , where is the number of array elements, is the number of pulses, is the number of range cells; Step S32: Based on the shaking characteristics of building features, construct a dynamic interference space-time steering vector, and the formula is as follows: , where is the spatial steering vector, is the temporal steering vector, represents the Kronecker product; Step S33: Screen the interference sample area without target signals in the Kepler region to obtain the interference covariance matrix : ; where L is the training sample, H is the conjugate transpose, represents the i-th space-time snapshot vector, represents the conjugate transpose of; Add a regularization term to improve the matrix condition number and suppress small eigenvalue perturbations; Step S34: Solve the STAP optimal weight vector W for dimensionality reduction processing. The specific calculation formula is as follows: ; where is the target signal steering vector, represents the joint covariance matrix of interference and noise, H represents the conjugate transpose element of the matrix, and -1 represents the matrix inverse operation; The dimensionality reduction processing uses the multistage Wiener filter (MSWF) to reduce the computational complexity from to ; Step S35: Obtain the signal after interference suppression through space-time weighted processing , ; Step S36: Based on the phase gradient network, perform secondary suppression on the micro-motion interference remaining after filtering to increase the signal-to-interference-plus-noise ratio by more than 15 dB.

[0038] In step S4, the specific process of detecting abnormal phase points and establishing a dynamic noise floor model is as follows: Step S41: Construct a multi-dimensional feature space and input the original phase matrix , and calculate the five-dimensional feature vector of each phase point. The specific formula is as follows: ; In the formula, represents the spatial coordinate index, represents the time series index, represents the spectral entropy feature parameter, represents the original phase data in three-dimensional space, respectively represent the spatial gradient of the phase on the x-axis, the spatial gradient of the phase on the y-axis, and the gradient of the phase change over time; This feature vector significantly improves the detection sensitivity of the millimeter-wave radar to micro-meteorological events (such as local wind shear and sudden changes in temperature and humidity) by fusing spatio-temporal gradients and spectral characteristics, and can achieve an accuracy of 0.1 mm in building structure deformation monitoring; Step S42: Construct a spatio-temporal probability map, and calculate the abnormal prior probability of each network point based on the historical data within the sliding window; the prior probability calculation formula is as follows: ; In the formula, is the attenuation coefficient, is the number of times determined to be abnormal within the past frames; converting historical abnormal records into probability guidance improves the effective sample extraction efficiency by more than 40%; Step S43: Parallelly fit three types of candidate models, and calculate the weighted residuals for each model; the three types of candidate models include a linear phase drift model, a quadratic fluctuation model, and a vortex propagation model; calculate the weighted residuals for each model, and break through the limitations of traditional single-model assumptions through physically driven multi-model competition; Step S44: Automatically adjust the internal and external thresholds according to the welding noise floor, and determine whether it is an abnormal point; the calculation formula is as follows: ; In the formula, Q3 is the third quartile and IQR is the interquartile range; the specific formula for determining whether it is an abnormal point is as follows: If the residual , it is determined to be an inlier; if , start spectral entropy verification; if , it is directly determined to be an abnormal point; Step S45: Construct a noise tensor and construct the inlier data of consecutive K frames into a four-dimensional tensor , and generate a dynamic basis; the calculation formula of the dynamic basis is as follows: ; In the formula, represents the noise basis value of the position in three-dimensional space at time t, represents the sum of r from 1 to R, respectively represent the distribution pattern of the r-th spatial basis vector in the i direction, the distribution pattern of the r-th spatial basis vector in the j direction, and the response value of the r-th temporal basis function at time t. The noise basis model can complete the response to sudden interference (such as a flock of birds) within 3 scanning cycles (about 1.5 seconds); Step S46: For the detected abnormal electric set, establish a spatio-temporal correlation graph for abnormal propagation backtracking, use the Louvain algorithm to detect closely related abnormal clusters, eliminate isolated noise points, and verify whether each abnormal cluster conforms to the fluid continuity equation. Clusters that do not meet the physical constraints are eliminated.

[0039] In step S5, the DPGN network separates the phase perturbation components at different spatial scales through a cascaded U-Net structure; the DPGN consists of three-level networks, specifically including: The first-level network of the DPGN network extracts the original phase matrix of the millimeter-wave reflection signal ; The second-level network of the DPGN network inversely calculates the meteorological parameter perturbation through the Doppler-refractive index coupling equation ; The third-level network of the DPGN network establishes a perturbation propagation model by combining LSTM time series prediction.

[0040] In step S5, the DPGN network separates the phase perturbation components at different spatial scales through a cascaded U-Net structure; the DPGN consists of three-level networks, specifically including: The first-level network of the DPGN network extracts the original phase matrix of the millimeter-wave reflection signal ; The second-level network of the DPGN network inversely calculates the meteorological parameter perturbation through the Doppler-refractive index coupling equation ; The Doppler-refraction is based on the propagation characteristics of electromagnetic waves in a non-uniform medium to establish a coupling equation between the refractive index perturbation and the Doppler velocity v: ; In the formula, n0 is the average atmospheric refractive index, represents the air flow velocity, c represents the speed of light; substitute Decomposed into: temperature humidity atmospheric pressure functions: ; In the formula, is the sensitivity coefficient of the atmospheric medium; Input feature: Receive the phase gradient tensor output by the first-level network and combine it with the original radar Doppler spectrum to construct a four-dimensional input feature (space × time × spectrum);‌ The third-level network of the DPGN network establishes a perturbation propagation model by combining LSTM time series prediction; the third-level network receives the four-dimensional meteorological perturbation tensor output by the second-level network, where the spatial resolution is 10m * 10m and the time step is 30 seconds / frame; the LSTM uses a bidirectional peephole LSTM, and a spatio-temporal attention module is inserted into the encoder-decoder structure. The encoder extracts the hidden state of the historical sequence and calculates the attention weight: ; In the formula, Q is the decoder state and d is the dimension parameter.

[0041] When training the DPGN network, the training data set adopts typical scenarios including typhoon eyes, urban heat islands, etc.; the DPGN network outputs a weather result map; the weather result map includes a microbarometric fluctuation distribution map, a water vapor condensation nucleus diffusion prediction vector field, and a perturbation propagation heat map.

[0042] In step S6, the lightweight engine processing flow deployed at the radar end is as follows: Step S61: Perform quadrature mixing mirror interference suppression on the millimeter-wave original signal, use Hilbert transform to restore the phase consistency of the I / Q channels, and eliminate more than 90% of the signal distortion; Step S62: Establish a dynamic baseline compensation model, and update the environmental noise floor in real time through a sliding window (500ms) to compensate for the phase shift caused by temperature drift‌; Step S63: Convert the multi-channel radar echo into a four-dimensional tensor structure (range × azimuth × elevation × Doppler), and achieve a standard input size of 32×32×16×64 through zero padding‌; Step S64: Use a double-buffer mechanism to construct a circular buffer to ensure that the data acquisition and processing pipeline latency < 5ms‌; Step S65: Deploy the density of the cascaded filter bank to extract data features; Step S66: Construct a cascaded network architecture, and achieve parallelism of the three stages of preprocessing - inference - postprocessing through an asynchronous pipeline design; the cascaded network structure includes: Pre - processing network: Improve the U - Net to extract the barometric pressure fluctuation characteristics of a 10m×10m grid; Core network: Use spatio - temporal convolutional GRU to predict the disturbance propagation path in the next 5 minutes; Post - processing network: Use conditional generative adversarial network (CGAN) to optimize the visualization effect of the heat map; Step S67: Establish a dynamic conversion model between the polar coordinate system and the Cartesian coordinate system to achieve sub - meter alignment of radar point cloud and GIS map. Use the Hungarian algorithm for multi - frame data association with an azimuth resolution accuracy of 0.1°; Step S68: Automatically adjust the detection sensitivity according to the environmental visibility, and update the model weights in real - time through the online learning module to adapt to sudden meteorological condition changes.

[0043] In step S7, detect the feature - level abnormal information, calculate the micro - barometric pressure gradient, extract the curl feature of the water vapor diffusion vector field, detect the local vortex generation points, establish an adaptive threshold model based on historical data, and divide the warning levels; and deploy an online learning module to correct the model parameters using the warning results (update period ≤ 10 min), and synchronize the multi - node knowledge bases through the federated learning framework to enhance the edge collaborative warning ability.

[0044] The warning level division is as follows in the table:

[0045] It should be noted that in the above - mentioned system embodiments, the various units included are only divided according to functional logic, but are not limited to the above - mentioned division, as long as the corresponding functions can be realized; in addition, the specific names of the functional units are only for the convenience of mutual distinction and do not limit the protection scope of the present invention.

[0046] In addition, those of ordinary skill in the art can understand that all or part of the steps in the methods of the above - mentioned embodiments can be completed by instructing relevant hardware through a program, and the corresponding program can be stored in a computer - readable storage medium.

[0047] The preferred embodiments of the present invention disclosed above are only used to help illustrate the present invention. The preferred embodiments do not describe all the details in detail, nor limit the invention to the specific embodiments described. Obviously, many modifications and variations can be made according to the content of this specification. These embodiments are selected and specifically described in this specification to better explain the principle and practical application of the present invention, so that those skilled in the art can understand and utilize the present invention well. The present invention is only limited by the claims and their full scope and equivalents.

Claims

1. A real-time monitoring method for minute meteorological changes based on millimeter-wave radar, characterized in that, It includes the following steps: Step S1: Dynamically collect millimeter-wave radar signals, and convert the radar signal scan data from the polar coordinate system to the X-Y plane rectangular coordinate system; Step S2: Generate a programmable chirp signal from the converted information through a DDS chip; Step S3: Perform space-time adaptive filtering processing on the received signal to eliminate interference; Step S4: Detect abnormal phase points and establish a dynamic noise floor model; Step S5: Input the preprocessed phase data into a pre-trained DPGN network, and the network outputs a weather result map; Step S6: Deploy a lightweight engine at the radar end to process the weather result map; Step S7: When it is detected that the weather gradient exceeds the threshold, trigger a weather warning.

2. The real-time monitoring method for minute meteorological changes based on a millimeter-wave radar according to claim 1, wherein In the said step S1, the formula for dynamically collecting millimeter-wave radar signals and converting them from the polar coordinate system to the rectangular plane coordinate system is as follows: ； In the formula, respectively represent the projected longitude and latitude after the polar coordinates are converted to the rectangular coordinate system, represents the longitude and latitude of the radar base station, represents the radar body storage yard, represents the radar elevation angle, represents the direction angle, A represents the latitude distance, and B represents the longitude distance.

3. A real-time monitoring method for minute meteorological changes based on millimeter-wave radar according to claim 1, characterized in that In the said step S2, a 77GHz millimeter-wave radar is used, 8 groups of MIMO arrays are deployed in the vertical direction, a programmable chirp signal is generated through a DDS chip, and the waveform parameters are adjusted according to the real-time signal-to-noise ratio.

4. A real-time monitoring method for minute meteorological changes based on a millimeter-wave radar according to claim 3, characterized in that, The said DDS chip adopts ADI AD9914; when the DDS chip dynamically controls the signal-to-noise ratio to adjust the waveform parameters, the radar receiving end calculates the background noise power in real time and performs signal-to-noise ratio calculation; when the calculated signal-to-noise ratio is less than 15dB, the pulse width expansion mode is triggered, and T gradually increases from 100ns to 200ns; when the calculated signal-to-noise ratio is greater than 25dB, the high-resolution mode is started, T is reduced to 50ns, and at the same time the ADC sampling frequency is increased. The specific chirp waveform modulation formula is as follows: ； wherein, represents the complex form of the time-domain radar signal, represents the time-varying amplitude modulation function, the complex exponential form, characterizing the phase modulation component, represents the carrier frequency phase term, is the noise feedback modulation coefficient, represents the dynamic phase perturbation term.

5. A real-time monitoring method for minute meteorological changes based on a millimeter-wave radar according to claim 1, characterized in that, In the said step S3, the space-time adaptive filtering processing flow is as follows: Step S31: Construct a three-dimensional space-time snapshot matrix through space diversity reception; Step S32: Based on the building feature shaking feature, Construct a dynamic interference space-time steering vector; Step S33: Screen the interference sample area without target signals in the Kepler region to obtain the interference covariance matrix; Step S34: Solve the STAP optimal weight vector W for dimensionality reduction processing; Step S35: Obtain the signal Y after interference suppression through space-time weighted processing; Step S36: Perform secondary suppression on the micro-motion interference remaining after filtering based on the phase gradient network.

6. The real-time monitoring method for minute meteorological changes based on millimeter-wave radar according to claim 1, wherein, In the said step S4, the specific process of detecting abnormal phase points and establishing a dynamic noise floor model is as follows: Step S41: Construct a multi-dimensional feature space and calculate the five-dimensional feature vector of each phase point; Step S42: Construct a spatio-temporal probability map, and calculate the abnormal prior probability of each network point based on the historical data within the sliding window; Step S43: Parallelly fit three types of candidate models and calculate the weighted residuals for each model; Step S44: Automatically adjust the internal and external thresholds according to the welding noise floor and determine whether it is an abnormal point; Step S45: Construct a noise tensor and generate a dynamic base; Step S46: Establish a spatio-temporal correlation graph for the detected abnormal electric set to perform abnormal propagation backtracking.

7. A real-time monitoring method for minute meteorological changes based on a millimeter-wave radar according to claim 1, characterized in that In the said step S5, the DPGN network separates the phase perturbation components of different spatial scales through a cascaded U-Net structure; The said DPGN consists of three-level networks, specifically including: The first - stage network of the DPGN network extracts the original phase matrix of the millimeter - wave reflection signal ; The second-level network of the DPGN network reversely calculates meteorological parameter perturbations through the Doppler-refractive index coupling equation ; The third - level network of the DPGN network establishes a perturbation propagation model by combining LSTM time - series prediction.

8. The real-time monitoring method for minute meteorological changes based on millimeter-wave radar according to claim 7, characterized in that When training the DPGN network, the training data set adopts typical scenarios including the typhoon eye and urban heat island; the DPGN network outputs a weather result map; the weather result map includes a micro - air - pressure fluctuation distribution map, a water - vapor condensation nucleus diffusion prediction vector field, and a perturbation propagation heat map.

9. A real-time monitoring method for minute meteorological changes based on a millimeter-wave radar according to claim 1, characterized in that, In step S6, the lightweight engine processing flow is deployed at the radar end as follows: Step S61: Perform quadrature mixing mirror interference suppression on the millimeter - wave raw signal, and use Hilbert transform to restore the phase consistency of the I / Q channels; Step S62: Establish a dynamic baseline compensation model, and update the environmental noise floor in real - time through a sliding window to compensate for the phase shift caused by temperature drift; Step S63: Convert the multi - channel radar echo into a four - dimensional tensor structure (range × azimuth × elevation × Doppler), and achieve a standard input size of 32×32×16×64 through zero padding; Step S64: Construct a circular buffer using a double - buffer mechanism; Step S65: Deploy a cascaded filter - bank density to extract data features; Step S66: Construct a cascaded network architecture, and realize the parallelism of the pre - processing, inference, and post - processing stages through an asynchronous pipeline design; Step S67: Establish a dynamic conversion model between the polar coordinate system and the Cartesian coordinate system, and use the Hungarian algorithm for multi - frame data association; Step S68: Automatically adjust the detection sensitivity according to the environmental visibility, and update the model weights in real - time through an online learning module to adapt to sudden meteorological condition changes.

10. A real-time monitoring method for minute meteorological changes based on a millimeter-wave radar according to claim 1, characterized in that, In step S7, detect the feature - level abnormal information, calculate the micro - air - pressure gradient, extract the curl feature of the water - vapor diffusion vector field, detect the local vortex generation points to establish an adaptive threshold model based on historical data, and perform the division of the warning level; and deploy an online learning module, use the warning results to correct the model parameters, and synchronize the multi - node knowledge bases through the federated learning framework.

Images