An Automatic Calibration Method for Water Accumulation Monitors Based on Adaptive Joint Feature Filtering
By using adaptive joint feature filtering and variational Bayesian inference, the problems of detection failure and high false alarm rate of millimeter-wave radar water accumulation monitoring technology in static and slow water level change scenarios are solved, realizing full-scenario detection and automatic calibration, and improving detection accuracy and system stability.
Patent Information
- Application Number
- CN202611158313.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-31
- Publication Date
- 2026-11-10
AI Technical Summary
Existing millimeter-wave radar water accumulation monitoring technology fails to detect water accumulation in static or slow water level change scenarios, suffers from high false alarm rates due to false echo interference, and lacks an automatic calibration mechanism, resulting in low detection accuracy and efficiency.
An adaptive joint feature filtering method is adopted, which combines the Doppler frequency domain and echo waveform time domain features. The adaptive switching of water accumulation state is achieved through variational Bayesian inference, and calibration is automatically triggered when the system drifts. The process includes Doppler feature extraction, waveform feature extraction, joint feature fusion and automatic calibration.
It enables full-scene detection of static, slowly changing, and dynamic water accumulation, reduces false alarm rate, ensures measurement accuracy and system stability, eliminates the influence of ambient temperature changes on measurement, and achieves automatic calibration without human intervention.
Smart Images

Figure CN123066698A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of waterlogging monitoring technology, and specifically to an automatic calibration method for waterlogging monitoring instruments based on adaptive joint feature filtering. Background Technology
[0002] Urban flooding is a significant disaster affecting the safe operation of cities, and road water accumulation monitoring is a core component of flood early warning. Water accumulation monitoring technology based on millimeter-wave radar is widely used due to its all-weather operation and high ranging accuracy. It measures water depth by transmitting frequency-modulated continuous wave signals and receiving reflected echoes from the water surface.
[0003] Current millimeter-wave radar water accumulation monitoring technology mainly employs a pure Doppler filtering method for signal processing. Its core principle is to utilize the Doppler frequency shift generated by the relative motion of the target object to detect water accumulation. However, this method has the following technical shortcomings when dealing with real-world scenarios of urban road water accumulation monitoring: Firstly, in scenarios with static water accumulation or a nearly still liquid surface, the reflected signal produces little or no Doppler frequency shift. The gain of traditional adaptive filters is severely attenuated in the low Doppler region, leading to a deterioration in the signal-to-noise ratio of the detection output. Secondly, in scenarios with slow water level changes, the Doppler signal is submerged by environmental noise, resulting in a low detection probability. The fundamental reason for this is that existing technologies process Doppler frequency domain information and echo waveform time domain information separately. While static water accumulation lacks Doppler characteristics, it has a stable waveform envelope; while dynamic water accumulation has significant Doppler characteristics, its waveform changes drastically. Single-domain processing methods cannot cover all water level change scenarios.
[0004] Secondly, when there are floating objects around the monitoring point, the radar echo exhibits multi-peak characteristics. Traditional fixed threshold detectors cannot effectively distinguish between real echoes from the water surface and false echoes from floating objects, resulting in a high false alarm rate and insufficient effective monitoring rate.
[0005] Third, existing monitoring instruments generally lack effective automatic calibration mechanisms. Insufficient temperature drift compensation accuracy leads to large measurement errors under extreme temperature conditions; the accumulation of contaminants on the sensor surface causes a decrease in sensitivity during long-term operation; traditional equipment relies on manual periodic calibration, resulting in severe distortion of measurement data within the calibration period.
[0006] In summary, existing technologies suffer from technical problems such as failure to detect static water accumulation, missed detection of slow water level changes, high false alarm rate due to false echo interference, and lack of automatic calibration capabilities. Summary of the Invention
[0007] The purpose of this invention is to overcome the shortcomings of the prior art. This application provides an automatic calibration method for water accumulation monitoring instruments based on adaptive joint feature filtering, which aims to solve the technical problems mentioned in the background art, such as low calibration efficiency, poor accuracy, failure of static water accumulation detection, missed detection of slow water level changes, and high false alarm rate caused by false echo interference.
[0008] To achieve the above objectives, this invention provides an automatic calibration method for a water accumulation monitoring instrument based on adaptive joint feature filtering, applied to a control system consisting of a millimeter-wave radar sensor, a signal conditioning module, an analog-to-digital conversion module, a digital signal processor, a storage module, and a temperature and humidity sensor. The method is characterized by the following steps: Step S1, Radar Echo Signal Acquisition and Preprocessing: A millimeter-wave radar sensor transmits a frequency-modulated continuous wave radar signal to the monitoring area and receives radar echo signals reflected from the water surface and floating objects. The millimeter-wave radar uses the 77GHz frequency band, with a radio frequency bandwidth of 4GHz, a range resolution better than 5cm, a maximum detection distance of 10m, and a ranging accuracy of ±2mm. Preferably, the starting frequency of the frequency-modulated continuous wave radar signal is 76GHz, the ending frequency is 81GHz, and the modulation period is 1ms.
[0009] The receiving link mixes the radio frequency echo signal with the local oscillator signal to generate an intermediate frequency (IF) signal. The IF signal is then sequentially amplified by low noise and filtered by a bandpass filter. The amplified and filtered IF signal enters the analog-to-digital converter (ADC) module for digital sampling. The ADC has a sampling rate of 20 MSPS and a digital sampling sequence length of 2048 points.
[0010] Simultaneously, the temperature and humidity sensors collect ambient temperature and humidity signals and transmit them to the digital signal processor via a data bus. The processor uses a table to look up the corresponding sound velocity correction coefficient based on the ambient temperature, which is then used for subsequent distance calculation compensation. Preferably, the temperature sensor is a platinum resistance temperature sensor with a measurement range of -40℃ to +85℃ and a measurement accuracy of ±0.3℃; the humidity sensor is a capacitive humidity sensor with a measurement range of 0% to 100% relative humidity and a measurement accuracy of ±2% relative humidity. Step S2, Doppler feature extraction: The digital sampling sequence obtained in step S1 is subjected to frequency domain transformation processing to extract Doppler velocity spectrum information and form a Doppler feature vector.
[0011] Specifically, a Hamming window is applied to the digital sampling sequence to reduce spectral leakage. The formula for calculating the coefficients of the Hamming window function is as follows:
[0012] Where N is the number of points in the Fast Fourier Transform (FFT), taking the value of 1024 points, and n is the sampling point index. The length of the digital sampling sequence is 2048 points, from which the first 1024 points are extracted for windowing and subsequent FFT processing. The windowed signal is represented as follows: , where x(n) is the original sampling sequence.
[0013] Subsequently, a 1024-point Fast Fourier Transform (FFT) was performed on the windowed signal. The FFT operation employed a radix-2 decimation-time algorithm, implemented in a digital signal processor using a butterfly arithmetic unit. The transform result X(k) is a complex sequence containing amplitude and phase information of 512 independent frequency components. Continuous acquisition... Distance-dimensional FFT results for each frequency modulation cycle ( (Value is 256), forming a slow time-dimensional sequence for the same distance cell along the frame number direction; a second FFT is performed on the slow time-dimensional data sequence, and the frequency domain result after the second FFT is assumed to be... .
[0014] The Doppler velocity spectrum is extracted using a peak search method. For each distance cell in the slow time dimension, the power spectrum of the corresponding Doppler frequency component is calculated. The power spectrum calculation formula is:
[0015] Peak search was performed on the power spectrum of all range cells to obtain the Doppler velocity spectrum. The relationship between Doppler frequency and target radial velocity is as follows:
[0016] in The wavelength is the radar wavelength (the wavelength corresponding to the 77GHz band is approximately 3.9mm). The frequency is the Doppler frequency. The Doppler velocity spectrum has 512 dimensions, forming the Doppler eigenvector. .
[0017] Step S3, Waveform Feature Extraction: The digital sampling sequence obtained in step S1 is synchronously subjected to envelope detection processing, and time-domain feature parameters are extracted from the envelope sequence to form a waveform feature vector.
[0018] Envelope detection was achieved using the Hilbert transform, which was implemented through a finite impulse response (FIR) filter. The FIR filter employed the Parks-McClellan optimal equal-ripple design method, with a passband range of 0.1 to 0.9 times the Nyquist frequency, a stopband attenuation greater than 60 dB, and a filter order of 128. The sequence after envelope detection underwent multi-stage decimation filtering, reducing the resolution to 1 kHz.
[0019] Four types of temporal feature parameters are extracted from the envelope sequence. Rise time Defined as the time it takes for the envelope to rise from 10% peak value to 90% peak value, reflecting the establishment speed of the echo signal. Peak amplitude Defined as the maximum value of the envelope sequence, reflecting the intensity of the target echo. The echo energy E is defined as the discrete integral approximation of the square of the envelope amplitude over the time window, i.e. Peak width Defined as the duration during which the envelope exceeds 50% of the peak value, it reflects the time spread characteristics of the echo pulse.
[0020] The four types of time-domain feature parameters mentioned above constitute the waveform feature vector W(t), which has a dimension of 512. Each feature parameter is normalized to eliminate the influence of dimensional differences on the subsequent fusion algorithm.
[0021] For echo signals exhibiting multi-peak characteristics after envelope detection, dynamic time warping (DTW) is used for matching and clustering to calculate the DTW distance between each echo pulse. Echoes with a DTW distance less than a preset threshold are grouped into the same category. Echoes whose waveform characteristics deviate from the typical water surface echo pattern are identified as false echoes of floating objects and suppressed. Echoes that conform to the typical water surface echo characteristics are retained for subsequent processing.
[0022] Step S4, Joint Feature Fusion: The Doppler feature vector obtained in step S2 and the waveform feature vector obtained in step S3 are weighted and concatenated according to dynamic weighting coefficients. The fusion result is then subjected to dimensionality reduction and normalization to obtain the joint feature vector.
[0023] The mathematical expression for weighted splicing and fusion is:
[0024] in and These are dynamic weighting coefficients. This represents a vector concatenation operation. The dynamic weighting coefficients are determined based on the feature variance.
[0025] in The variance of the Doppler eigenvectors. denoted as the variance of the waveform eigenvector. It satisfies the normalization condition.
[0026] The weighted concatenated joint feature vector F has a dimension of 1024. To reduce the computational complexity and storage overhead of subsequent algorithms, dimensionality reduction is performed. Principal Component Analysis (PCA) is used for this dimensionality reduction. First, a data matrix containing M historical samples is constructed. Where M is 1000. Calculate the mean vector of the data matrix. Covariance Matrix :
[0027] The eigenvalues and eigenvectors of the covariance matrix are obtained through eigenvalue decomposition. The eigenvectors corresponding to the first 64 largest eigenvalues are then used to form the projection matrix. Dimensionality reduction transforms a 1024-dimensional feature vector to a 64-dimensional space:
[0028] in, It is the projection matrix composed of the eigenvectors corresponding to the first 64 largest eigenvalues.
[0029] The eigenvectors after dimensionality reduction are then normalized using the L2 norm.
[0030] Normalization ensures that the energy of the joint feature vector remains constant, eliminating the impact of feature amplitude fluctuations on the classifier.
[0031] Step S5, Variational Bayesian Inference: The joint feature vector obtained in step S4 is probabilistically classified using the variational Bayesian inference method, and the output includes at least the posterior probabilities corresponding to three detection states: static water accumulation, slowly changing water accumulation, and dynamic water accumulation.
[0032] Preferably, the observation model adopts a Gaussian mixture model (GMM), which includes three Gaussian components, corresponding to three detection states: static water accumulation, slowly changing water accumulation, and dynamic water accumulation. Let the latent class variable be C, with values ranging from {static, slowly changing, dynamic}. The parameters of each Gaussian component include the mean vector. Covariance Matrix ,in The conditional probability of the observation model is:
[0033] in, These correspond to three states: static, slow, and dynamic. The goal of variational Bayesian inference is to approximate the posterior distribution. , where Z is the set of latent class variables. Introducing variational distribution. The approximate posterior is solved by minimizing the KL divergence between the variational distribution and the true posterior distribution.
[0034] E-step: Fix model parameters and calculate the posterior distribution of latent variables:
[0035] M-step to fix the distribution of latent variables and update model parameters. ,in Here are the prior mixing coefficients:
[0036] The iterative process repeats the E-step and M-step until convergence. The convergence criterion is that the change in the lower bound of evidence (ELBO) between two consecutive iterations is less than... Or reach the maximum number of iterations of 50.
[0037] After the iteration converges, the posterior probabilities of each category are calculated using Bayes' theorem:
[0038] The posterior probability ranges from 0 to 1, reflecting the confidence level that the joint feature vector belongs to each category.
[0039] Step S6, Adaptive mode switching: Determine the current water accumulation status based on the posterior probability output in step S5, adaptively switch to the corresponding detection mode, and set the corresponding integration time and detection threshold.
[0040] Pattern determination uses a threshold comparison strategy: When the static posterior probability When the water level is high, it is determined to be in a static water accumulation mode. In this case, the integration time is set to 30 seconds, and the detection threshold is set to 0.5 mm. A longer integration time can effectively smooth out random noise and improve the signal-to-noise ratio for detecting static targets.
[0041] When slow posterior probability and When the water level changes slowly, the system is identified as a slowly changing water accumulation mode. This mode is suitable for applications where the water level changes slowly at a rate of 0.5 mm / s to 5 mm / s. The integration time is set to 10 seconds, and the detection threshold is set to 0.2 mm.
[0042] When the dynamic posterior probability and and When the water level changes rapidly or is flowing, the system is identified as being in dynamic water accumulation mode. This mode is suitable for scenarios where the water level changes rapidly or where there is flow. The integration time is set to 2 seconds, and the detection threshold is set to 1.0 mm.
[0043] The mode switching mechanism employs a combination of hard switching and soft hold. Hard switching occurs only after three consecutive frames that meet the switching condition, preventing frequent mode jumps caused by transient noise or sudden interference. The soft hold mechanism provides a 15-second buffer period after mode switching; if the posterior probability reverts to the original mode during this buffer period, the system immediately switches back to the original mode, ensuring smoothness and stability during mode switching.
[0044] Step S7, Automatic Calibration Triggering and Execution: Continuously monitor the system operating status parameters. When a drift in system parameters is detected, the calibration program is automatically triggered to perform automatic calibration of the system parameters.
[0045] The drift monitoring module continuously calculates the deviation between the statistical mean of the joint eigenvectors and the initial calibration value. The deviation is calculated using the L2 norm as a measure.
[0046] in This represents the statistical mean of the joint feature vectors at the current time. This is the baseline mean value during the initial system calibration.
[0047] The calibration trigger conditions include at least one of the following: Cumulative drift Exceeding the preset threshold by 2mm (primary triggering method); The continuous running time reaches the preset calibration cycle of 48 hours; When the ambient temperature changes by more than 15°C (auxiliary triggering method, which complements the cumulative drift monitoring); The first measurement after the system is powered on and reset.
[0048] The automatic calibration procedure is triggered when any condition is met. Among them, monitoring the cumulative drift between the statistical mean of the joint feature vector and the initial calibration value is the main calibration triggering method, which can comprehensively reflect the system performance degradation caused by various factors such as sensor aging, contaminant accumulation, and circuit parameter drift. Ambient temperature change triggering is an auxiliary calibration triggering method, mainly used to provide rapid supplementary calibration when the temperature changes drastically. The two complement each other and jointly ensure the measurement accuracy of the system under all operating conditions.
[0049] The automatic calibration procedure includes the following stages: Baseline Acquisition Phase: Lasts 5 seconds, radar transmission is disabled, and only ambient noise data is acquired. The digital signal processor calculates the noise mean and noise variance.
[0050] Where M is the number of noise sampling points, with a value of 5000.
[0051] Model update phase: lasts 3 seconds, updating the background noise model based on the collected noise baseline.
[0052] in The inherent noise variance of the sensor is determined through testing during the system design phase.
[0053] Filter reinitialization phase: lasts 2 seconds, reinitializing the prior distribution parameters of the variational Bayesian filter, including the mean vector, covariance matrix, and mixing coefficients for each category.
[0054] in These are the prior mixing coefficients, with initial values set according to typical application scenarios.
[0055] Recovery monitoring phase: After completing the calibration parameter update, normal water accumulation monitoring mode is restored.
[0056] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention effectively extracts the Doppler frequency components of water surface reflected echoes by performing a second FFT on the distance dimension FFT results of multiple consecutive frequency modulation cycles in a slow time dimension. Combined with a peak search method, the Doppler velocity spectrum is obtained, enabling the effective extraction of weak Doppler signals in both static and slowly changing water accumulation scenarios. This provides reliable Doppler feature input for subsequent joint feature fusion. In dynamic water accumulation scenarios, Doppler features dominate the joint features to enhance motion information, and dynamic weighting adjustments are made based on feature variance anomalies to highlight feature dimensions with stronger discriminative power. Therefore, it achieves effective coverage of water level change scenarios from static to dynamic, overcoming the technical bottleneck of single-domain processing methods being unable to cover the entire scenario. This invention constructs waveform feature vectors by extracting multi-dimensional time-domain features such as rise time, peak amplitude, echo energy, and peak width from the envelope sequence, enabling real water surface echoes and floating object echoes to form a separable distribution in the feature space. Furthermore, it combines dynamic time warping matching clustering to perform similarity clustering on multi-peak echoes, identifying and suppressing echoes whose waveform features deviate from typical water surface echo feature patterns as false echoes, thereby effectively reducing the system's false alarm rate. This invention continuously tracks and monitors the statistical mean of the joint feature vector, using statistical distance to measure its deviation from the initial calibration benchmark. It comprehensively senses system performance degradation caused by the coupling of multiple factors such as sensor aging, temperature drift, and contaminant accumulation. When the deviation exceeds a preset range, it automatically triggers a complete calibration process including benchmark noise acquisition, background noise model updating, and filter prior distribution reinitialization. The entire calibration process requires no manual intervention. Simultaneously, by synchronously acquiring ambient temperature and continuously compensating for the sound velocity of the ranging parameters, the influence of ambient temperature changes on measurement accuracy is eliminated, ensuring measurement consistency across the entire temperature range. Attached Figure Description
[0057] Figure 1 This is a schematic diagram of the principle framework of the adaptive joint feature filtering algorithm in this invention; Figure 2This is a flowchart of the automatic calibration method for the water accumulation monitor in this invention; Figure 3 This is a schematic diagram of the parallel processing flow of Doppler feature extraction and waveform feature extraction in this invention; Figure 4 This is a schematic diagram of the variational Bayesian adaptive filtering and mode switching process in this invention. Detailed Implementation
[0058] To make the objectives, technical solutions, and advantages of this invention clearer and more complete, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention in any way. Example
[0059] This embodiment discloses an automatic calibration method for a water accumulation monitoring instrument based on adaptive joint feature filtering, covering the entire signal processing and mode switching mechanism from radar echo signal acquisition to automatic calibration execution. This method achieves accurate detection of all scenarios—static water accumulation, slowly changing water accumulation, and dynamic water accumulation—by jointly utilizing Doppler frequency domain features and echo waveform time domain features. Furthermore, it achieves adaptive switching of detection modes and automatic calibration of system parameters through variational Bayesian inference.
[0060] The automatic calibration system for the water accumulation monitor in this embodiment includes core hardware components such as a millimeter-wave radar sensor, a signal conditioning module, an analog-to-digital converter, a digital signal processor, a storage module, and a temperature and humidity sensor. These components are connected via a high-speed data bus, forming a complete signal acquisition, processing, and storage chain.
[0061] The millimeter-wave radar sensor employs a 77GHz frequency-modulated continuous wave system with a 4GHz RF bandwidth, a range resolution of 4cm, a maximum detection range of 10m, and a ranging accuracy of ±2mm. The radar module comprises three sub-modules: a transmitting antenna, a receiving antenna, and an RF front-end circuit. The transmitting antenna uses a microstrip array antenna structure with a gain of 12dBi, and its radiation pattern covers the main area of the monitoring region. The receiving antenna uses a dual-channel MIMO structure, achieving angle measurement through spatial diversity, effectively improving the spatial resolution of the target. The RF front-end circuit integrates core components such as a voltage-controlled oscillator, power amplifier, low-noise amplifier, and mixer to perform RF signal generation, amplification, and down-conversion processing.
[0062] The signal conditioning module comprises two functional units: a low-noise amplifier and a bandpass filter. The low-noise amplifier, with a noise figure of less than 2dB and a gain of 30dB, amplifies the radar echo signal to a power level suitable for subsequent processing while suppressing introduced additional noise as much as possible. The bandpass filter, with a passband range of 200MHz to 8GHz and a stopband rejection greater than 40dB, filters out out-of-band interference signals and image frequency components, ensuring the signal-to-noise ratio of the intermediate frequency signal.
[0063] The analog-to-digital converter (ADC) module employs a 12-bit resolution ADC with a sampling rate of 20 MSPS and an analog input bandwidth of 8 MHz. The ADC front-end integrates an anti-aliasing filter and a sample-and-hold circuit. The anti-aliasing filter uses an elliptic filter design, with passband ripple less than 0.1 dB and stopband attenuation greater than 60 dB, effectively preventing high-frequency signal aliasing. The aperture jitter of the sample-and-hold circuit is less than 1 ps, ensuring accuracy during high-speed sampling.
[0064] The digital signal processor (DSP) uses an ARM Cortex-M7 core microcontroller chip with a maximum clock speed of 480MHz, and features a floating-point arithmetic unit and a digital signal processing instruction set. The DSP module is responsible for the real-time processing of core algorithms such as FFT operations, Hilbert transform, PCA dimensionality reduction, and variational Bayesian filtering. The processor integrates 512KB of flash memory and 256KB of SRAM to meet the requirements for program storage and data caching.
[0065] The storage module consists of ferroelectric RAM (FRAM) and electrically erasable programmable read-only memory (EEPROM). The FRAM has a capacity of 64KB and is used to store calibration parameters, filter status, and system configuration information. It is resistant to 10¹³ erase / write cycles, ensuring long-term reliable operation. The EEPROM has a capacity of 512KB and is used to store historical monitoring data and event logs, supporting data retention even when power is off.
[0066] The temperature and humidity sensors include platinum resistance temperature sensors and capacitive humidity sensors. The temperature sensor uses a PT1000 platinum resistance element, with a measurement range of -40℃ to +85℃ and an accuracy of ±0.3℃. The humidity sensor uses an HS3001 capacitive humidity chip, with a measurement range of 0% to 100% relative humidity and an accuracy of ±2% relative humidity. Both sensors communicate with the main control chip via an I²C digital interface to achieve real-time acquisition of environmental parameters.
[0067] Step S1: The millimeter-wave radar sensor transmits a frequency-modulated continuous wave signal to the monitoring area according to a preset modulation period. The modulation waveform uses linear frequency modulation, with a starting frequency of 76 GHz, an ending frequency of 81 GHz, and a modulation period of 1 ms. The transmitted signal is captured by the receiving antenna after reflection from the water surface or floating objects. The transmit power is set to +10 dBm, and the total gain of the receiving link is 60 dB. After considering antenna loss and feeder loss, the overall equivalent noise figure is approximately 5 dB. The receiver's intermediate frequency bandwidth is 4 MHz, corresponding to a time resolution of 0.25 μs, and the theoretical limit of the ranging resolution is 3.75 cm.
[0068] The receiving link first mixes the RF echo signal with the local oscillator signal to generate an intermediate frequency (IF) signal. The mixed IF signal is then amplified by a low-noise amplifier and filtered by a bandpass filter to remove image interference and out-of-band noise. Subsequently, the IF signal enters the analog-to-digital converter (ADC) module for digital sampling. The sampling clock is generated by a crystal oscillator with a frequency accuracy of ±25 ppm to ensure accurate sampling timing.
[0069] The sampled digital sequence has a length of 2048 points, a sampling rate of 20 MSPS, and a corresponding time window of 102.4 μs. After decimation filtering, the digital sequence is reduced to an effective sampling rate of 1 MHz for subsequent signal processing.
[0070] Simultaneously, temperature and humidity sensors collect environmental parameters and transmit the data to a digital signal processor via an I²C bus. The processor uses a table to look up the corresponding sound speed correction coefficient based on the ambient temperature, which is then used for subsequent distance calculation and compensation.
[0071] Step S2: Perform frequency domain analysis on the acquired radar echo signal to extract Doppler velocity spectrum information reflecting the target's motion state.
[0072] First, a Hamming window function is applied to the sampled sequence to reduce spectral leakage. The coefficients of the window function are calculated using the following formula:
[0073] Where N is the number of points in the Fast Fourier Transform (FFT), taking the value of 1024 points, and n is the sampling point index. The length of the digital sampling sequence is 2048 points. The first 1024 points are truncated for windowing and subsequent FFT processing. The windowed signal is represented as follows: , This is the original sampling sequence.
[0074] Subsequently, a 1024-point Fast Fourier Transform (FFT) was performed on the windowed signal. The FFT operation employed a radix-2 decimation-time algorithm, implemented in a DSP using a butterfly arithmetic unit. The transform result X(k) is a complex sequence containing amplitude and phase information for 512 independent frequency components. Continuous acquisition was performed. Distance-dimensional FFT results for each frequency modulation cycle ( (Value is 256), forming a slow time-dimensional sequence for the same distance cell along the frame number direction; a second FFT is performed on the slow time-dimensional data sequence, and the frequency domain result after the second FFT is assumed to be... .
[0075] The Doppler velocity spectrum is extracted using a peak search method. For each distance cell in the slow time dimension, the power spectrum of the corresponding Doppler frequency component is calculated. The formula for calculating the power spectrum is:
[0076] Peak search was performed on the power spectrum of all range cells to obtain the Doppler velocity spectrum. The relationship between Doppler frequency and target radial velocity is as follows:
[0077] Where λ is the radar wavelength, and the wavelength corresponding to the 77GHz band is approximately 3.9mm. The frequency is the Doppler frequency.
[0078] The Doppler velocity spectrum has 512 dimensions, corresponding to the amplitude information of 512 discrete frequency components. These amplitude values constitute the Doppler eigenvectors. This is used for subsequent joint feature fusion processing.
[0079] Step S3: Perform time-domain analysis on the radar echo signal and extract waveform feature parameters that reflect the shape of the echo envelope.
[0080] First, envelope detection is performed on the original sampled sequence. Envelope detection is achieved using the Hilbert transform, and the calculation formula is as follows:
[0081] The Hilbert transform is implemented using a finite impulse response (FIR) filter. The filter design employs the Parks-McClellan optimal equal-ripple design method, with a passband range of 0.1 to 0.9 times the Nyquist frequency, a stopband attenuation greater than 60 dB, and a filter order of 128. This 128th-order FIR filter achieves sufficient transition bandwidth and stopband rejection performance while maintaining phase linearity.
[0082] Sequence obtained from envelope detection After multiple stages of decimation filtering, the resolution is reduced to 1kHz. The length of the extracted envelope sequence is 512 points, which matches the dimension of the Doppler feature vector.
[0083] Four types of temporal feature parameters are extracted from the envelope sequence. Rise time Defined as the time it takes for the envelope to rise from 10% peak value to 90% peak value, reflecting the establishment speed of the echo signal. Peak amplitude Defined as the maximum value of the envelope sequence, reflecting the intensity of the target echo. The echo energy E is defined as the discrete integral approximation of the square of the envelope amplitude over the time window, i.e. Peak width Defined as the duration during which the envelope exceeds 50% of the peak value, it reflects the time spread characteristics of the echo pulse.
[0084] These four types of feature parameters constitute the waveform feature vector W(t), which has a dimension of 512. Each element in the feature vector is normalized to eliminate the influence of dimensional differences on the subsequent fusion algorithm.
[0085] Step S4: Weight the Doppler feature vector and the waveform feature vector, and then perform dimensionality reduction and normalization.
[0086] The mathematical expression for weighted splicing and fusion is:
[0087] in and These are the weighting coefficients. The weighting coefficients are determined using a dynamic allocation strategy based on the characteristic variance. First, the variances of the Doppler eigenvector and the waveform eigenvector are calculated separately. and The formula for calculating the weighting coefficient is:
[0088] This allocation strategy ensures that feature vectors with larger variance receive higher weights, thus retaining more discriminative information in the joint features. It satisfies the normalization condition.
[0089] The weighted concatenated joint feature vector F has a dimension of 1024. To reduce the computational complexity and storage overhead of subsequent algorithms, dimensionality reduction is required. Principal component analysis is used as the dimensionality reduction algorithm. First, a data matrix containing M historical samples is constructed. Where M is 1000. Calculate the mean vector of the data matrix. Covariance Matrix :
[0090] The eigenvalues and eigenvectors of the covariance matrix are obtained through eigenvalue decomposition. The eigenvectors corresponding to the first 64 largest eigenvalues are used to form the projection matrix. The projection matrix The feature vectors are calculated and determined during system initialization and updated periodically in subsequent operation using a sliding window method. The window length is 1000 frames, and the update period is 10 minutes. Dimensionality reduction transforms the 1024-dimensional feature vectors to a 64-dimensional space, resulting in... .
[0091] The eigenvectors after dimensionality reduction are then normalized using the L2 norm.
[0092] Normalization ensures that the energy of the joint feature vector remains constant, eliminating the impact of feature amplitude fluctuations on the classifier.
[0093] Step S5: Use the variational Bayesian inference method to perform probabilistic classification on the joint feature vector and output the posterior probabilities of three detection states: static water accumulation, slow change, and dynamic water accumulation.
[0094] This step embeds variational Bayesian inference into the physical observation model of water accumulation monitoring, enabling mathematical inference methods to serve the classification of physical states. The observation model employs a Gaussian mixture model, containing three Gaussian components, each corresponding to one of the three detection states. Let the latent class variable be C, with values ranging from {static, slow, dynamic}. The parameters of each Gaussian component include the mean vector. Covariance Matrix ,in The conditional probability of the observation model is:
[0095] The goal of variational Bayesian inference is to approximate the posterior distribution. , where Z is the set of latent class variables. Introducing variational distribution. The approximate posterior is solved by minimizing the KL divergence between the variational distribution and the true posterior distribution.
[0096] E-step: Fix model parameters and calculate the posterior distribution of latent variables:
[0097] M-step to fix the distribution of latent variables and update model parameters. Where π_k is the prior mixing coefficient:
[0098] The iterative process repeats the E-step and M-step until convergence. The convergence criterion is that the change in the lower bound of evidence between two consecutive iterations is less than... Or reach the maximum number of iterations of 50.
[0099] After the iteration converges, the posterior probabilities of each category are calculated using Bayes' theorem:
[0100] The posterior probability ranges from 0 to 1, reflecting the confidence level that the joint feature vector belongs to each category.
[0101] Step S6: Determine the current water accumulation status based on the posterior probability result, switch to the corresponding detection mode, and set the corresponding integration time and detection threshold.
[0102] The pattern determination uses a threshold comparison strategy. The determination criteria for the static water accumulation pattern are as follows: When this condition is met, the system determines that there is still or nearly still water within the monitoring area. In this case, the integration time is set to 30 seconds, and the detection threshold is set to 0.5 mm. A longer integration time can effectively smooth out random noise and improve the signal-to-noise ratio for detecting static targets.
[0103] The criteria for determining a slow change pattern are: and This mode is suitable for applications where the water level changes slowly at a rate of 0.5 mm / s to 5 mm / s. The integration time is set to 10 seconds, and the detection threshold is set to 0.2 mm. The moderate integration time balances response speed and detection stability.
[0104] The criteria for determining the dynamic water accumulation mode are: and and This mode is suitable for scenarios with rapid water level changes or where there is flow. The integration time is set to 2 seconds, and the detection threshold is set to 1.0 mm. The short integration time ensures that the system can respond quickly to changes in water level.
[0105] The mode switching mechanism employs a combination of hard switching and soft hold. Hard switching occurs only after three consecutive frames that meet the switching condition, preventing frequent mode jumps caused by transient noise or sudden interference. The soft hold mechanism provides a 15-second buffer period after mode switching; if the posterior probability reverts to the original mode during this buffer period, the system immediately switches back to the original mode, ensuring smoothness and stability during mode switching.
[0106] Step S7: Continuously monitor the system operating status parameters. When a drift in system parameters is detected, automatically trigger the calibration procedure and perform the calibration operation.
[0107] The core concept of this step is to achieve fully automated calibration without additional hardware references through "signal feature drift monitoring." The drift monitoring module continuously calculates the deviation between the statistical mean of the joint eigenvectors and the initial calibration value. The deviation is calculated using the L2 norm as a measure.
[0108] in This represents the statistical mean of the joint feature vectors at the current time. This is the baseline mean value during the initial system calibration.
[0109] The calibration trigger conditions include the following four scenarios: cumulative drift exceeding a preset threshold of 2 mm (primary trigger method); no calibration performed for 48 consecutive hours; ambient temperature change exceeding 15℃ (auxiliary trigger method); and the first measurement after system power-on reset. Meeting any one of these conditions triggers the automatic calibration procedure. Cumulative drift monitoring, as the primary trigger method, comprehensively reflects multiple factors such as sensor aging, contaminant accumulation, and circuit parameter drift. Temperature triggering, as an auxiliary method, is mainly used to provide rapid supplementary calibration during drastic temperature changes; the two methods complement each other.
[0110] The automatic calibration procedure comprises four execution phases. The baseline acquisition phase lasts 5 seconds, during which the radar sensor shuts down its transmission signal and acquires only ambient noise data. The digital signal processor calculates the noise mean and noise variance.
[0111] Where M is the number of noise sampling points, with a value of 5000.
[0112] The model update phase lasts for 3 seconds. The background noise model is updated based on the collected noise baseline.
[0113] in The inherent noise variance of the sensor is determined through testing during the system design phase.
[0114] The filter re-initialization phase lasts 2 seconds. It re-initializes the prior distribution parameters of the variational Bayesian filter, including the mean vectors, covariance matrices, and mixing coefficients for each class.
[0115] in These are the prior mixing coefficients, with initial values set according to typical application scenarios.
[0116] After the calibration parameters are updated during the recovery monitoring phase, the system resumes normal water accumulation monitoring mode. Intermediate data and status information during the calibration process are written to the calibration log area of the storage module for subsequent analysis and traceability.
[0117] The software implementation process of the automatic calibration module is as follows: The drift monitoring submodule continuously calculates at a period of 1 second. ,when When a threshold is exceeded, a trigger signal is sent to the trigger condition judgment submodule. This submodule simultaneously receives temperature change signals and runtime timing signals; when either trigger condition is met, the calibration execution submodule is activated. The calibration execution submodule sequentially performs benchmark acquisition, model update, and filter reinitialization, and upon completion, feeds back the status to the main control loop. All of these submodules are implemented as software tasks in the main controller, exchanging data via a message queue.
[0118] This embodiment also includes a temperature compensation step to correct the impact of ambient temperature changes on radar ranging accuracy. A temperature sensor collects ambient temperature data in real time and transmits it to a digital signal processor via an I²C bus.
[0119] The processor's temperature compensation module obtains the sound speed correction coefficient from a table based on the measured temperature. The relationship between sound speed and temperature is achieved through piecewise linear interpolation, with the interpolation table covering a temperature range of -40℃ to +85℃, in 1℃ increments. The corrected ranging formula is as follows:
[0120] in This is the original radar ranging result. The speed of sound at the reference temperature, The speed of sound at the current temperature, This is the corrected distance value.
[0121] This temperature compensation mechanism, combined with variational Bayesian adaptive filtering, ensures that the measurement error is controlled within ±1.5% over a wide temperature range of -30℃ to +60℃.
[0122] It should be noted that the temperature compensation described above is a continuous, real-time correction, which is a different level of operation from the "automatic calibration triggered when the ambient temperature changes by more than 15°C" step S7: temperature compensation is a real-time correction at the signal level, used to compensate for the impact of sound speed changes with temperature on ranging accuracy; temperature-triggered calibration is a re-initialization at the system parameter level, used to address model parameter mismatch caused by drastic temperature changes. The two are independent and complementary.
[0123] In this embodiment, the waveform feature extraction module also integrates a false echo suppression function. When floating objects are present in the monitoring area, the radar echo exhibits multi-peak characteristics. The false echo suppression algorithm uses the Dynamic Time Warping (DTW) method to match and cluster the multi-peak echoes.
[0124] The DTW algorithm calculates the similarity metric between different echo pulses. Let two echo pulse sequences be A and B, with lengths m and n respectively. The DTW distance is defined as:
[0125] in , representing the i-th point of sequence A With the j-th point of sequence B The squared Euclidean distance between them. The optimal matching path is solved using dynamic programming, grouping echoes with high similarity into the same category.
[0126] The waveform characteristics of false echoes differ significantly from those of genuine water surface echoes. Water surface echoes exhibit stable statistical characteristics in rise time and peak width, while echoes generated by floating objects show shorter rise times and narrower peak widths. By comparing the DTW distance between the echo characteristics and typical water surface echo characteristics, it is possible to determine whether an echo is false and suppress it.
[0127] Application example: In the application scenario of monitoring water accumulation in low-lying areas of urban roads, the water accumulation monitoring instrument in this application example is installed on street lamp poles or traffic sign poles in sections of roads prone to water accumulation. The radar sensor is installed at a downward tilt, with a vertical coverage distance of 0.5m to 6m.
[0128] One night, rainfall began in the monitored area. As the rainfall intensified, water gradually formed on the roads. The millimeter-wave radar continuously transmitted FMCW signals, with a sampling rate set to 1 MHz and a sampling sequence length of 2048 points. The sampled data was first processed by Hamming windowing, followed by a 1024-point FFT transform to extract the Doppler velocity spectrum.
[0129] Initially, the surface velocity of the accumulated water is below 0.5 mm / s, and the Doppler frequency shift is close to zero. However, at this time, the echo waveform envelope is stable, with a rise time of approximately 2 ms and a peak amplitude that remains constant within a preset range. The variational Bayesian filter calculates the posterior probability based on the joint eigenvector. The integration value reaches 0.85, meeting the criteria for static water accumulation mode. The system switches to static water accumulation mode with an integration time of 30 seconds and a detection threshold of 0.5mm. When the water depth reaches 10mm, the system generates an alarm message and uploads it to the monitoring center.
[0130] As rainfall continues and the catchment area increases, a weak flow begins to appear on the surface of the accumulated water, with the flow velocity increasing to 1 mm / s to 2 mm / s. At this time, a weak velocity component appears in the Doppler velocity spectrum, and the waveform envelope begins to change slowly. The posterior probability of the variational Bayesian filter changes. It rose to 0.7. The value drops to 0.3. After three consecutive frames meet the criteria for slow change mode, the system switches to slow change mode, the integration time is adjusted to 10 seconds, and the detection threshold is adjusted to 0.2 mm.
[0131] During the peak of the rainstorm, the flow of water accelerates, with velocities exceeding 5 mm / s. The dynamic components of the Doppler velocity spectrum are significantly enhanced, and the waveform envelope undergoes drastic changes. It rose to 0.75. and All values were below the threshold. The system determined it to be in dynamic water accumulation mode, shortened the integration time to 2 seconds, and adjusted the detection threshold to 1.0 mm. The rapidly changing integration parameters ensured that the system could respond promptly to sudden changes in water level.
[0132] After the rain stops, the water flow gradually slows down, and the system gradually switches back to the slow-changing mode based on the posterior probability, finally reverting to the static water accumulation mode. The 15-second soft hold buffer effectively avoids frequent mode jumps under critical conditions.
[0133] During operation, the temperature sensor continuously monitors changes in ambient temperature. When the temperature change exceeds 15°C, the automatic calibration procedure is triggered. The system enters the calibration process, sequentially completing reference acquisition, model update, and filter reinitialization. After calibration, the prior distribution parameters of the variational Bayesian filter are restored to their initial state, ensuring that measurement accuracy is not affected by temperature drift.
[0134] This application example fully verifies the applicability and reliability of the method in this embodiment for monitoring water accumulation in all scenarios.
[0135] To further illustrate the technical effects of the present invention, the following comparative description will be provided through embodiments and two comparative examples.
[0136] Comparative Example 1: Pure Doppler Filtering Method The same hardware environment as the above embodiments is used, but the signal processing method is replaced with a common pure Doppler filtering method in the prior art. This method only uses Doppler frequency domain information for water accumulation detection, and uses a fixed threshold to detect peaks in the Doppler velocity spectrum. When the peak amplitude exceeds a preset threshold, water accumulation is determined to be present. This method does not extract waveform time-domain features, does not perform feature fusion, and does not have an automatic calibration function.
[0137] Comparative Example 2: Fixed Fusion Method The hardware environment is exactly the same as in the above embodiments, but the feature fusion method is replaced with fixed-weight concatenation fusion, that is, α and β are fixed at 0.5 and are not dynamically adjusted according to the statistical characteristics of the feature variance. Other processing steps (including Doppler feature extraction, waveform feature extraction, variational Bayes classification, mode switching and automatic calibration) are consistent with the embodiments.
[0138] Test conditions and evaluation indicators A typical 24-hour workday during the rainy season, characterized by variable weather conditions, was selected for testing. During the test, the ambient temperature ranged from 22℃ to 34℃, the cumulative rainfall was 45.6 mm, and the maximum hourly rainfall intensity was 28.2 mm / h. The monitoring scenario was a low-lying section of an urban road with varying degrees of water accumulation and flow.
[0139] During the test, three comparative verification methods were set up around the monitoring point: (1) manual gauging method - the actual water depth was measured on-site by maintenance personnel every hour; (2) video monitoring - the actual state of the waterlogged area was recorded by a high-definition camera; (3) rain gauge - the changes in rainfall intensity were recorded. The detection accuracy of each method was calculated based on the measurement value of the manual gauging method.
[0140] The key performance indicators of the three systems under the same conditions were recorded and compared in detail. The comparison results are shown in Table 1.
[0141] The significant advantages of the technical solution of the present invention can be clearly seen through the analysis of the data in Table 1.
[0142] Table 1: Performance Comparison Table of Examples, Comparative Example 1, and Comparative Example 2
[0143] Analysis of the data in Table 1 shows that the detection probability of this invention reaches 96.2% in static water accumulation scenarios, while that of Comparative Example 1 is only 38.5%. This is mainly because the pure Doppler filtering method in Comparative Example 1 cannot generate an effective Doppler frequency shift in static scenes, resulting in near-failure in detection. Although the fixed fusion method in Comparative Example 2 extracts waveform features, the contribution of waveform features is suppressed in some scenarios due to the lack of a dynamic weighting strategy based on feature variance, resulting in a detection probability of only 82.3%. This invention, through a dynamic weighting mechanism, automatically increases the weight of waveform features in static water accumulation scenarios, significantly improving the detection probability.
[0144] Regarding the detection probability of slowly changing water accumulation, the embodiment of the present invention achieves 94.8%, compared to 57.1% in Comparative Example 1 and 86.5% in Comparative Example 2. The embodiment of the present invention improves upon Comparative Example 1 by 37.7 percentage points, fully demonstrating the superiority of dual-domain joint feature extraction in weak signal scenarios.
[0145] Regarding the false alarm rate of false echoes, the embodiment of the present invention has a false alarm rate of only 3.2%, which is significantly lower than 18.6% in Comparative Example 1 and 8.5% in Comparative Example 2. This is mainly due to the ability of the waveform feature extraction module in the present invention to distinguish between real water surface echoes and false echoes of floating objects, as well as the effective application of the dynamic time warping matching clustering algorithm.
[0146] Regarding long-term operational stability, Comparative Example 1 showed a 12.8% decrease in detection probability after 30 days of continuous operation, primarily due to the lack of an automatic calibration mechanism, sensor aging, and model parameter mismatch caused by environmental changes. Comparative Example 2, while possessing an automatic calibration function, still experienced a 4.3% decrease due to the inability of fixed fusion weights to adapt to scene changes. This invention, through a fully automatic calibration mechanism based on signal feature drift monitoring (using the drift of the joint feature vector statistical mean as the primary trigger and temperature change as an auxiliary trigger), controlled the decrease to 1.2%, verifying the long-term effectiveness of the automatic calibration method.
[0147] To meet the needs of different application scenarios, this invention also provides an alternative implementation scheme for an automatic calibration method for water accumulation monitoring instruments based on adaptive joint feature filtering. This scheme is optimized for water accumulation monitoring in enclosed spaces such as underground passages and tunnels.
[0148] Water accumulation monitoring in enclosed spaces has the following characteristics: short monitoring distance, typically ranging from 0.2m to 3m; relatively enclosed environment with a narrow temperature variation range but drastic humidity fluctuations; and a complex electromagnetic propagation environment with multiple reflections from walls and ceilings. To address these characteristics, the alternative solution adjusts the hardware configuration and algorithm parameters accordingly.
[0149] The millimeter-wave radar sensor uses the 24GHz band to reduce hardware costs, while maintaining a 4GHz RF bandwidth to ensure sufficient range resolution. The transmit power is reduced to +5dBm to meet electromagnetic compatibility requirements. The sampling rate is adjusted to 500ksps to match the shorter monitoring distance.
[0150] The bandpass filter in the signal conditioning module has a passband range adjusted to 100MHz to 4GHz, with passband ripple less than 0.2dB and stopband rejection greater than 50dB. The lower passband upper limit helps filter out high-frequency wall reflection interference.
[0151] In the Doppler feature extraction stage, the number of FFT transform points remains unchanged at 1024, but the windowing function is changed to a Blackman window to achieve better sidelobe suppression performance. The coefficients of the Blackman window are:
[0152] In the waveform feature extraction stage, the order of the Hilbert transform filter is increased to 256 to improve the accuracy of envelope detection. In the time-domain feature extraction stage, two parameters, kurtosis coefficient and skewness coefficient, are added to enhance the ability to identify multi-peak echoes.
[0153] In the joint feature fusion stage, considering the more severe interference from multi-peak echoes within a closed space, the weighting coefficient β of the waveform feature vector is appropriately increased. The adjustment strategy for the dynamic weighting coefficient is changed to:
[0154] The value of k is 1.5, which is used to increase the weight of waveform features in the joint features.
[0155] In the false echo suppression mechanism, the matching window width of the DTW algorithm is expanded from ±50 points to ±100 points to accommodate long-delay echoes generated by reflections from walls within enclosed spaces. The threshold for determining echo validity is correspondingly increased; only echoes with a DTW distance less than the threshold are considered valid water surface echoes.
[0156] In the automatic calibration trigger conditions, the temperature change threshold has been reduced from 15℃ to 10℃ to accommodate the relatively small temperature variations within enclosed spaces. The calibration cycle has been shortened from 48 hours to 24 hours to ensure long-term operational stability.
[0157] This alternative solution has achieved good results in the actual deployment of underground passage water accumulation monitoring. The monitoring coverage rate reached over 98%, and the false alarm rate was controlled within 3%, effectively meeting the application requirements for water accumulation monitoring in enclosed spaces.
[0158] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.
Claims
1. An automatic calibration method for a water accumulation monitoring instrument based on adaptive joint feature filtering, applied to a control system consisting of a millimeter-wave radar sensor, a signal conditioning module, an analog-to-digital conversion module, a digital signal processor, a storage module, and a temperature and humidity sensor, characterized in that, Includes the following steps: Step S1: Transmit frequency-modulated continuous wave radar signals to the monitoring area and receive radar echo signals formed by water surface reflection; perform frequency mixing processing on the radar echo signals to generate intermediate frequency signals; amplify, filter, and perform analog-to-digital conversion processing on the intermediate frequency signals to obtain digital sampling sequences; Step S2: Perform frequency domain transformation on the digital sampling sequence to extract Doppler velocity spectrum information and construct Doppler feature vectors; Step S3: Perform envelope detection processing on the digital sampling sequence, extract time-domain feature parameters from the envelope sequence, and construct a waveform feature vector; Step S4: The Doppler feature vector and the waveform feature vector are weighted and concatenated according to dynamic weighting coefficients, and the fusion result is subjected to dimensionality reduction and normalization to obtain a joint feature vector; Step S5: Use the variational Bayesian inference method to perform probability classification on the joint feature vector. The observation model of the variational Bayesian inference adopts a Gaussian mixture model, whose three Gaussian components correspond to three physical states: static water accumulation, slowly changing water accumulation, and dynamic water accumulation, respectively. The output includes at least the posterior probabilities corresponding to the three detection states: static water accumulation, slowly changing water accumulation, and dynamic water accumulation. Step S6: Determine the current water accumulation status based on the posterior probability, adaptively switch to the corresponding detection mode, and set the corresponding detection parameters; Step S7: Continuously monitor the system operating status parameters. When a drift in system parameters is detected, automatically trigger the calibration program to perform automatic calibration of the system parameters.
2. The automatic calibration method for a water accumulation monitoring instrument based on adaptive joint feature filtering according to claim 1, characterized in that, In step S1: the frequency-modulated continuous wave radar signal has a preset start frequency, a preset end frequency, and a preset modulation period; the analog-to-digital conversion has a preset sampling rate, and the digital sampling sequence has a preset length; In step S2, before performing frequency domain transformation on the digital sampling sequence, a window function is applied to the digital sampling sequence to reduce spectral leakage. The frequency domain transformation is a fast Fourier transform with a preset number of transformation points. The Doppler velocity spectrum is extracted using a peak search method, which obtains the Doppler velocity spectrum by performing a peak search on the power spectrum of all distance cells in the slow time dimension. In step S3, the envelope detection is achieved through Hilbert transform, which is implemented through a finite impulse response (FIR) filter, which has a preset order. The time-domain characteristic parameters include rise time, peak amplitude, echo energy, and peak width. The rise time is the time it takes for the envelope to rise from a first preset percentage peak value to a second preset percentage peak value. The peak amplitude is the maximum value of the envelope sequence. The echo energy is the integral discrete approximation of the square of the envelope amplitude within a time window. The peak width is the duration during which the envelope is higher than a third preset percentage peak value.
3. The automatic calibration method for a water accumulation monitoring instrument based on adaptive joint feature filtering according to claim 1, characterized in that, In step S4: the dynamic weighting coefficients are determined based on the feature variances. The Doppler feature weighting coefficient is equal to the variance of the Doppler feature vector divided by the sum of the variances of the Doppler feature vector and the waveform feature vector. The waveform feature weighting coefficient is equal to the variance of the waveform feature vector divided by the sum of the variances of the Doppler feature vector and the waveform feature vector. The dimensionality reduction process uses principal component analysis to map the weighted and concatenated 1024-dimensional feature vector to a 64-dimensional space. The normalization process uses L2 norm normalization to scale the dimensionality-reduced feature vector to a unit length.
4. The automatic calibration method for a water accumulation monitoring instrument based on adaptive joint feature filtering according to claim 1, characterized in that, In step S5: the observation model of the variational Bayesian inference adopts a Gaussian mixture model, which includes three Gaussian components corresponding to static water accumulation, slowly changing water accumulation, and dynamic water accumulation, respectively; the variational inference is performed iteratively through E-step and M-step until convergence, the E-step fixes the model parameters and calculates the posterior distribution of the latent variables, and the M-step fixes the latent variable distribution and updates the model parameters; the convergence criterion is that the change in the lower bound of evidence between two adjacent iterations is less than a preset threshold or the preset maximum number of iterations is reached.
5. The automatic calibration method for a water accumulation monitoring instrument based on adaptive joint feature filtering according to claim 1, characterized in that, In step S6: When the static posterior probability is greater than the first preset threshold, it is determined to be a static water accumulation mode, the integration time is set to the first time value, and the detection threshold is set to the first detection threshold. When the slow posterior probability is greater than the second preset threshold and the static posterior probability is not greater than the first preset threshold, it is determined to be a slow-changing water accumulation mode, the integration time is set to the second time value, and the detection threshold is set to the second detection threshold. When the dynamic posterior probability is greater than the third preset threshold and the static posterior probability is not greater than the first preset threshold and the slow posterior probability is not greater than the second preset threshold, it is determined to be a dynamic water accumulation mode, the integration time is set to the third time value, and the detection threshold is set to the third detection threshold. The adaptive handover adopts a mechanism that combines hard handover and soft hold: the hard handover condition is that the handover condition is met for N consecutive frames before the mode switch is performed; the soft hold mechanism sets a buffer period after the mode switch, and if the posterior probability reverts to the original mode during the buffer period, the mode is switched back to the original mode.
6. The automatic calibration method for a water accumulation monitoring instrument based on adaptive joint feature filtering according to claim 1, characterized in that, In step S7, the calibration triggering conditions include at least one of the following: the cumulative drift between the statistical mean of the joint feature vector and the initial calibration value exceeds a preset threshold; the continuous running time reaches a preset calibration cycle; the ambient temperature change exceeds a preset temperature threshold; and the first measurement is taken after the system is powered on and reset. Among these, monitoring the cumulative drift between the statistical mean of the joint feature vector and the initial calibration value is the primary calibration triggering method, while the ambient temperature change triggering method is an auxiliary calibration triggering method. The cumulative drift is measured using the L2 norm to measure the deviation between the statistical mean of the joint feature vector and the initial calibration value. The automatic calibration procedure includes: During the baseline acquisition phase, the radar transmission signal is turned off and environmental noise data is collected to calculate the noise mean and noise variance. During the model update phase, the background noise model is updated based on the collected noise benchmark. During the filter reinitialization phase, the prior distribution parameters of the variational Bayesian filter are restored to their initial state. During the monitoring recovery phase, the water accumulation monitoring mode will be restored after the calibration parameters are updated.
7. The automatic calibration method for a water accumulation monitoring instrument based on adaptive joint feature filtering according to claim 1, characterized in that, Also includes: The temperature compensation step involves synchronously acquiring ambient temperature signals and correcting the ranging parameters in real time based on the mathematical relationship between temperature and sound speed. The false echo suppression step involves dynamic time-warping matching and clustering of multi-peak echoes to distinguish between real echoes from the water surface and false echoes from floating objects.
8. The automatic calibration method for a water accumulation monitoring instrument based on adaptive joint feature filtering according to claim 2, characterized in that, The frequency modulated continuous wave radar signal operates at a frequency of 77 GHz; the analog-to-digital conversion sampling rate is 20 MSPS.
9. The automatic calibration method for a water accumulation monitoring instrument based on adaptive joint feature filtering according to claim 1, characterized in that, In monitoring water accumulation in enclosed spaces, the dynamic weighting coefficient is determined as follows: The variance of the Doppler eigenvectors is denoted as The variance of the waveform eigenvector is denoted as The first weighted adjustment coefficient is denoted as k; the Doppler characteristic weighting coefficient... Weighting coefficients based on waveform characteristics Calculate according to the following formulas respectively: ; in, The first weighted adjustment coefficient Greater than 1.
