River surface flow velocity and flow measurement device based on active shock wave and space-time doppler-deep learning fusion analysis

Abstract

The application discloses a river surface flow velocity and flow measurement device based on active shock wave and space-time Doppler-deep learning fusion analysis, which comprises a vibration body emission device, a data acquisition device, a data processing and control module, a Doppler-deep learning fusion feature construction module, a physical inversion module, a deep learning flow velocity / flow regression module and a flow calculation and output unit. The application improves the signal-to-noise ratio of the tracer through the active shock wave, combines the space-time Doppler-deep learning fusion feature engineering and the wave physical inversion, realizes the stable non-contact measurement of the river flow under the condition of weak light or even night, and can realize the high-reliability non-contact flow measurement of the river under the condition of insufficient natural texture, weak light or even night in the river channel (open channel or water body such as lake and reservoir).

Description

Technical Field

[0001] This invention belongs to the field of hydrological measurement and hydrodynamic monitoring technology, specifically relating to a device and apparatus for measuring river surface velocity and flow rate based on active shock wave and spatiotemporal Doppler-deep learning fusion analysis. Background Technology

[0002] Currently, river flow measurement mainly includes two categories of methods: contact and non-contact. Contact methods, such as current metering and acoustic Doppler current profiler (ADCP), require the deployment of personnel and equipment during floods, which poses problems such as high safety risks, low automation, and inability to achieve long-term continuous monitoring.

[0003] Non-contact flow measurement technologies have developed rapidly in recent years, mainly including microwave radar flow measurement, LSPIV (Large Scale Particle Image Velometry) based on video images, and STIV (Space-Time Image Velometry). Video flow measurement methods acquire images of the river surface using cameras on the bank or bridge, and use textures such as water surface foam, floating objects, and natural ripples as tracers to calculate the surface velocity, and then combine it with cross-sectional geometric parameters to estimate the flow rate.

[0004] Existing video flow measurement / STIV technologies have the following typical problems:

[0005] (1) Heavy dependence on natural tracers and lighting conditions: Most methods assume that there is sufficient density of foam, floating objects or natural ripple texture on the water surface and require good visible light conditions. Once the water surface is flat or the lighting is insufficient, the texture signal-to-noise ratio in the spatiotemporal image decreases significantly, resulting in unstable flow velocity inversion.

[0006] (2) The correspondence between texture features and physical flow velocity is not clear enough: Some schemes only extract the brightness slope in the spatiotemporal image or calculate the optical flow on the image sequence, without making full use of the physical laws of water surface ripples and lacking the physical constraints of decomposing the flow velocity from the wave velocity, which makes the algorithm more sensitive to noise and environmental changes.

[0007] (3) Limited application scope of deep learning: Some existing video flow measurement systems have tried to introduce deep learning for water body recognition, floating object recognition or water level recognition, but few systems have considered the fusion of high-dimensional spatiotemporal features of ripples in spatiotemporal profiles and Doppler phase velocity features, and the degree of integration between physical modeling and data-driven models is not high.

[0008] In summary, there is an urgent need for a new video flow measurement method and device that can work stably even under conditions of insufficient natural texture, low light, or even nighttime, integrate wave physics and deep learning at the algorithm level, and can achieve real-time on-site processing through dedicated hardware. Summary of the Invention

[0009] To address the aforementioned issues, embodiments of the present invention propose a river surface velocity and flow measurement device and apparatus based on active shock wave and spatiotemporal Doppler-deep learning fusion analysis.

[0010] The present invention relates to a river surface velocity and flow measurement device based on active shock wave and spatiotemporal Doppler-deep learning fusion analysis, comprising: a shock wave transmitter installed on the riverbank; a data acquisition device for acquiring video sequences of water surface ripples and water level data generated by the shock wave transmitter; a data processing and control module for preprocessing the acquired water surface ripple video sequences to construct spatiotemporal profile images; a Doppler-deep learning fusion feature construction module for parameter splicing of the processed data to construct event-level fusion features for velocity inversion; a physical inversion module for calculating the theoretical phase velocity under still water conditions based on water level data and obtaining a physical initial estimate of the river surface velocity; a deep learning velocity / flow regression module for calculating the corrected river surface velocity; and a flow calculation and output unit for calculating, storing, displaying, or remotely transmitting the river flow.

[0011] The exciter launching device includes: a base fixed to the riverbank; a support pole vertically fixed above the base; a cantilever crossbar vertically installed on the upper side of the support pole and extending horizontally above the river measurement section; an exciter release mechanism, which is electromagnetically controlled and installed below the end of the cantilever crossbar; and an exciter, which is a spherical counterweight placed inside the exciter release mechanism.

[0012] The data acquisition device includes: a high-speed camera mounted on a cantilever crossbar, with a camera range covering the area where the exciter enters the water; a water level measuring device installed at the end of the cantilever crossbar, located between the high-speed camera and the exciter release mechanism, used to acquire real-time water depth data of the measurement section; and a synchronous triggering unit used to transmit a signal from a microcontroller to the high-speed camera, triggering the high-speed camera to start recording, at the same time the release mechanism releases the exciter.

[0013] The exciter release mechanism includes a hollow cylindrical shell with an open bottom. The top of the cylindrical shell is fixedly connected to the cantilever crossbar via a connecting rod. An exciter delivery assembly is fixedly installed inside the cylindrical shell.

[0014] The exciter delivery assembly is fixed to the top of the inside of the cylindrical shell. The bottom of the exciter delivery assembly is open, and an electromagnetic switch is installed at the bottom opening. The exciter is placed inside the exciter delivery assembly.

[0015] The spatiotemporal profile image is generated by the following steps:

[0016] S101. Collect a large number of ripple video frames excited by the exciter under different hydrological and lighting conditions;

[0017] S102. Generate corresponding wavefront position binary mask labels for each frame of image through manual annotation or semi-automatic algorithm, and generate a standard sample library; perform wavefront pixel-level segmentation on each frame of image through convolutional neural network to obtain the binary mask of the ripple region.

[0018] S103. Extract brightness / texture values ​​only along the main direction and several transverse or oblique profiles in the ripple region with high segmentation confidence, and stack them over time to form a spatiotemporal profile image;

[0019] S104. Using the ripple video frame as input and the corresponding binary mask of the front position of the water wave as the supervision label, train the CNN-UNet model using a deep learning framework until its loss function converges.

[0020] The construction of event-level fusion features for flow velocity inversion includes Doppler spectral analysis and deep learning spatiotemporal feature encoding and fusion, specifically including the following steps:

[0021] S201. Perform a two-dimensional fast Fourier transform or short-time Fourier transform on the spatiotemporal profile to obtain the frequency-wavenumber spectrum; or use the Radon transform to search for the principal energy direction on the spatiotemporal plane;

[0022] S202. Extract the frequencies of the principal energy directions in the spectral domain. With wavenumber Or extract the observed phase velocity C corresponding to the spatiotemporal slope. observed ;

[0023] S203. Construct Doppler characteristics such as the mean, variance, and robust statistics of observed phase velocities at the event level;

[0024] S204. Using 3D CNN or CNN+ and other temporal networks, the spatiotemporal profiles of multiple survey lines are jointly encoded to obtain high-dimensional spatiotemporal feature vectors;

[0025] S205. Doppler features, depth features, and environmental variables such as water depth, wind speed, and wind direction are concatenated to form event-level fusion features, which are then used for subsequent flow velocity / flow regression learning.

[0026] The Doppler spectrum analysis includes the following steps:

[0027] First, the spatial resolution Δs between pixel coordinates and actual spatial coordinates is determined based on camera calibration parameters, and the inter-frame time interval Δt is determined according to the video frame rate. Gray values, texture intensity, or wavefront response intensity are extracted from the water surface ripple video sequence along the main channel direction or a preset survey line direction to construct a spatiotemporal profile image. The spatiotemporal profile tensor formula for a single shock wave event is as follows:

[0028]

[0029] in, .

[0030] In the formula, Δs is the scaling factor between the image pixel coordinate system and the world coordinate system, in meters per pixel; Δt is the inter-frame time interval of the video, in seconds per frame; x i = i·Δs represents the spatial position along the survey line, i = 0,1,…,N x-1 ;t n = n·Δt represents time, n = 0,1,…,N t-1 .

[0031] Then, a two-dimensional Fourier transform is performed on the spatiotemporal profile image to obtain the frequency-wavenumber spectrum:

[0032]

[0033] in, ; .

[0034] In the formula, For frequency-wavenumber spectrum, Let p be the p-th discrete spatial wavenumber, where p = 0, 1, ..., N. x-1 , ; For spatial frequency, ; f is the q-th discrete angular frequency, in rad / s; q q represents the time frequency, in Hz. p is the spectral number in the spatial direction; q is the spectral number in the temporal direction.

[0035] Furthermore, the frequency-wavenumber energy spectrum is defined as follows:

[0036]

[0037] Extract the main energy direction from the energy spectrum:

[0038]

[0039]

[0040] in, and These are the spatial and temporal frequency indices corresponding to the principal energy direction, i.e. ( , ) is to make the spectrum energy The largest set of index numbers; and These are the characteristic spatial wavenumber and characteristic angular frequency corresponding to the principal direction of this energy, respectively.

[0041] The calculation of the river surface velocity includes the following steps:

[0042] S1. Based on the wavenumber of the characteristic space and characteristic angular frequency Calculate the phase velocity of the observed wave:

[0043]

[0044] In the formula, C observed Indicates the observed wave phase velocity;

[0045] S2. Based on the current water depth h obtained from the water level measuring device, calculate the theoretical still water wave velocity C0(h) under shallow water long wave conditions. The formula is:

[0046]

[0047] In the formula, g is the acceleration due to gravity, and h is the current water depth;

[0048] S3. The initial physical estimate U of the river surface velocity is obtained from the difference between the observed wave phase velocity and the theoretical wave velocity in still water. phys The formula is:

[0049]

[0050] In the formula, U phys C0 represents the surface velocity of the river and C0 represents the theoretical still water wave velocity.

[0051] The deep learning flow rate / flow regression module will U phy C observed , , Doppler spectral eigenvectors Deep learning spatiotemporal feature vectors and environmental feature vectors The pre-trained deep learning regression model is input together.

[0052] Among them, the Doppler spectral eigenvectors This feature vector is used to characterize the dominant propagation characteristics and stability of ripples in a spatiotemporal profile during an active shock event. It is composed of statistics on the observed wave phase velocity, spectral parameters of the dominant ripple, and spectral energy concentration.

[0053]

[0054] In the formula, M represents the number of survey lines or the number of time windows. The average flow velocity from M measurements. Let C be the variance of the flow velocity measured in M ​​measurements. max It is the maximum flow rate, C min It is the minimum flow rate.

[0055] The aforementioned feature vectors can simultaneously reflect the magnitude of the ripple propagation speed, the consistency between different measurement lines, the significance of the main ridge of the spectrum, and the reliability of the measurement results, providing input features for subsequent deep learning flow velocity correction models.

[0056] Deep learning spatiotemporal feature vectors Encoded and obtained by 3D CNN, CNN-LSTM, or temporal convolutional networks:

[0057]

[0058] In the formula, X is a deep learning spatiotemporal coding network; X is the multichannel spatiotemporal profile tensor input to the deep learning network.

[0059] The deep learning spatiotemporal feature vector Spatiotemporal coding network The input tensor X, composed of multiple spatiotemporal profile images, is encoded to obtain the result.

[0060] Environmental feature vector Composed of environmental variables at the time of shooting, the vector consists of:

[0061]

[0062] In the formula, h is the current water depth, in meters; W is the current wind speed, in meters per second; θ w This represents the angle between the current wind direction and the river channel direction. Optionally, environmental factors such as air temperature, water temperature, light intensity, and image signal-to-noise ratio may also be included.

[0063] Will U phy C observed , , Doppler spectral eigenvectors Deep learning spatiotemporal feature vectors and environmental feature vectors The input is a pre-trained deep learning regression model, which outputs the final surface flow velocity. Deep learning regression model formula:

[0064]

[0065] In the formula, U final f represents the final surface flow rate corrected by deep learning. θ ( ) represents a deep learning regression model with parameter θ; The Doppler spectral eigenvector can include C observed The mean, variance, phase velocity distribution on multiple measurement lines, and spectral energy concentration, etc. These are deep feature vectors extracted by 3D CNN, CNN-LSTM, or other spatiotemporal coding networks; This is an environmental feature vector, which may include water depth h, wind speed, wind direction, light intensity, etc.

[0066] Furthermore, the average flow velocity at the cross-section needs to be corrected based on the shape of the cross-section at the site. The specific correction formula is as follows:

[0067]

[0068] In the formula, α is the cross-sectional average velocity, in m / s; α is the correction factor for converting surface velocity to cross-sectional average velocity, which is dimensionless and can be determined through field calibration, historical flow measurement data, or hydraulic models.

[0069] The average flow rate of the cross section can be derived from the average flow velocity corrected for the cross section.

[0070]

[0071] In the formula, Q is the cross-sectional flow rate of the river, with the unit being m³ / s, and A(h) represents the cross-sectional area of ​​the river channel below the water surface at a water depth of h, that is, the area is a function of the water depth.

[0072] The beneficial effects of this invention are that it improves the tracer signal-to-noise ratio by using active shock waves, and combines spatiotemporal Doppler-deep learning fusion feature engineering and wave physics inversion to achieve stable non-contact measurement of river flow under low light or even nighttime conditions. It can optionally achieve real-time on-site processing through FPGA, and has strong engineering application value. It can achieve highly reliable non-contact flow measurement in rivers (open channels or lakes and reservoirs) with insufficient natural texture, low light or even nighttime conditions. Attached Figure Description

[0073] Figure 1 This is a schematic diagram of the structure of the exciter launching device of the present invention.

[0074] Figure 2 This is a schematic diagram of the exciter release mechanism of the present invention.

[0075] Figure 3 This is a schematic diagram of the radar level gauge and radar level gauge camera of the present invention.

[0076] Figure 4 This is a schematic diagram of the binary mask labels of the first ring of ripples generated after the exciter of the present invention falls into the water surface at three different times.

[0077] Figure label:

[0078] 1. Base; 2. Support pole; 3. Distribution box; 4. Solar panel; 5. High-speed camera; 6. Radar water level gauge camera; 7. Radar water level gauge; 8. Exciter release mechanism; 9. Diagonal brace; 10. Cantilever beam; 11. Exciter; 12. Exciter release assembly; 13. Connecting rod one; 14. Solenoid valve; 15. Connecting rod two. Detailed Implementation

[0079] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0080] The present invention relates to a river surface velocity and flow measurement device based on active shock wave and spatiotemporal Doppler-deep learning fusion analysis, comprising: a shock body emitting device, a data acquisition device, a data processing and control module, a Doppler-deep learning fusion feature construction module, a physical inversion module, a deep learning velocity / flow regression module, and a flow calculation and output unit.

[0081] The exciter launching device is installed on the riverbank to generate highly repeatable and consistent water surface ripples at the measurement cross-section. Exciters of the same mass are launched into the river at the same angle. A high-speed camera captures spatiotemporal dynamic images of the spreading ripples caused by the exciter's entry into the water under the action of the flowing water. Video of the ripple diffusion process is also acquired simultaneously, and a dataset is constructed.

[0082] The data acquisition device is used to simultaneously record video sequences of water surface ripples and water level data generated by the excitation device. Water depth data of the river cross-section is obtained through water gauges and video acquisition; if meteorological station data is available, wind speed data can also be acquired simultaneously.

[0083] like Figures 1-4As shown, the exciter launching device includes: a base 1, a supporting pole 2, a cantilever crossbar 10, an exciter release mechanism 8, and an exciter 11. The base 1 is fixed to the riverbank; the supporting pole 2 is vertically fixed above the base 1; a solar panel 4 is installed on the top of the supporting pole 2, and a distribution box 3 is installed on one side of the lower part. The cantilever crossbar 10 is vertically installed on one side of the upper part of the supporting pole 2 and extends horizontally above the river measurement section; the exciter release mechanism 8 is electromagnetically controlled and is installed below the end of the cantilever crossbar 10; the exciter 11 is a spherical counterweight, and the exciter 11 is placed inside the exciter release mechanism 8. A diagonal brace 9 is also provided between the cantilever crossbar 10 and the supporting pole 2.

[0084] A ball-throwing port is provided at the top of the connection between the cantilever crossbar 10 and the diagonal brace 9 for throwing the exciter 11. The section of the cantilever crossbar 10 from the ball-throwing port to the exciter release mechanism 8 is hollow, and the inner diameter of the cantilever crossbar 10 is larger than the outer diameter of the exciter 11. The cantilever crossbar 10 is inclined from the root to the end (slope of 1%) to ensure that the exciter can slide along the cantilever crossbar 10 into the exciter release mechanism 8.

[0085] The exciter release mechanism 8 includes a hollow cylindrical shell with an open bottom. The top of the cylindrical shell is fixedly connected to the cantilever crossbar 1 via a connecting rod 13. The connecting rod 13 is a hollow structure with open ends. The two ends of the connecting rod 13 are respectively connected to the cantilever crossbar 10 and the exciter release assembly 12. The exciter release assembly 12 is fixedly installed inside the cylindrical shell.

[0086] The exciter delivery assembly 12 is fixed inside the top of the cylindrical shell. The top of the exciter delivery assembly 12 is connected to the bottom of the connecting rod 13. The bottom of the exciter delivery assembly 12 is open, and an electromagnetic switch is set at the bottom opening. Multiple exciters 11 are placed inside the exciter delivery assembly 12.

[0087] Reply: To avoid the increased energy consumption and higher failure rate caused by complex structures, this application adopts a passive current-limiting release structure of "single-row ball storage cylinder + bottom throttling ball outlet + single electromagnetic baffle". Its single-ball release principle is as follows: The exciter release assembly 12 includes a vertical ball storage cylinder and a bottom ball outlet channel, with an electromagnetic control latch or solenoid valve 14 installed between the vertical ball storage cylinder and the bottom ball outlet channel. A bottom ball outlet is provided at the bottom of the vertical ball storage cylinder.

[0088] Let the diameter of the exciter be D, and the inner diameter of the vertical ball storage cylinder be D. c The diameter of the bottom ball outlet, i.e., the opening when the solenoid valve is open, is D. o The bottom ball outlet channel is L o In structural design, the following conditions must be met:

[0089]

[0090] In the formula, D c Slightly larger than the exciter diameter D, this ensures that multiple exciters are arranged in a single vertical row within the vertical ball storage cylinder, preventing two or more exciters from entering the ball storage cylinder side-by-side or simultaneously being squeezed into the ball outlet; D o Slightly larger than the exciter diameter D, it ensures the exciter can pass smoothly through the ball outlet while limiting the lateral space of the ball outlet, allowing only one exciter to enter the outlet channel at any given time; L o Used to form a channel of finite length, so that a time interval is generated between two adjacent exciters during the ball ejection process.

[0091] One exciter can be placed into the exciter delivery assembly each time, or the number of exciters delivered each time the solenoid valve is opened can be controlled by adjusting the opening time of the solenoid valve 14.

[0092] The data acquisition device includes a high-speed camera 5 and a water level measuring device. The high-speed camera 5 is mounted on the cantilever crossbar 10, and its imaging range covers the water entry area of ​​the exciter 11. The water level measuring device is installed at the end of the cantilever crossbar 10, located between the high-speed camera 5 and the exciter release mechanism 8, and is used to acquire real-time water depth data of the measurement section. A synchronous triggering unit is used to transmit a signal from the microcontroller to the high-speed camera at the same time as the release mechanism releases the exciter, triggering the high-speed camera to start recording.

[0093] The water level measuring device includes a radar water level gauge 7 and a radar water level gauge camera 6. The function of the radar water level gauge camera 6 is to infer the water depth based on the water level, and then calculate the still water wave velocity at a specific water depth to infer the actual flow velocity. The top of the radar water level gauge camera 6 is mounted on the cantilever crossbar 10 through a connecting rod 2 15. The radar water level gauge camera 6 is a spherical camera.

[0094] The data processing and control module is used to preprocess the video sequence and, based on the wavefront segmentation results, construct a spatiotemporal profile image of water surface brightness or texture changing over time along the main channel direction and at least one preset profile line.

[0095] Spatiotemporal profile images are generated by the following steps:

[0096] S101. Collect a large number of ripple video frames excited by the exciter under different hydrological and lighting conditions;

[0097] S102. Generate corresponding wavefront position binary mask labels for each frame of image through manual annotation or semi-automatic algorithm, and generate a standard sample library; perform wavefront pixel-level segmentation on each frame of image through convolutional neural network to obtain the binary mask of the ripple region.

[0098] S103. Extract brightness / texture values ​​only along the main direction and several horizontal or oblique profiles in regions with high segmentation confidence, and stack them over time to form a spatiotemporal profile image;

[0099] S104. Using the ripple video frame as input and the corresponding wavefront position mask as the supervision label, train the CNN-UNet model using a deep learning framework until its loss function converges.

[0100] Doppler-Deep Learning Fusion Feature Construction Module:

[0101] ① Used to perform two-dimensional spectral transformation or Radon transform on the spatiotemporal profile image to extract the frequency-wavenumber ridge or spatiotemporal slope characterizing the ripple propagation phase velocity, and obtain the observed phase velocity C. observed and its statistical characteristics;

[0102] ②The deep learning spatiotemporal coding submodule is a convolutional neural network, a three-dimensional convolutional network, or a convolutional-temporal hybrid network. It performs end-to-end coding of spatiotemporal profile images and outputs high-dimensional feature vectors representing the spatiotemporal morphology of ripples.

[0103] ③ Feature fusion submodule, used to integrate C observed The event-level fusion features for flow velocity inversion are constructed by splicing the Doppler spectral features with deep learning spatiotemporal features and environmental parameters such as current water depth, wind speed and wind direction.

[0104] Doppler spectral analysis and deep learning spatiotemporal feature encoding and fusion specifically include the following steps:

[0105] S201. Perform a two-dimensional fast Fourier transform or short-time Fourier transform on the spatiotemporal profile to obtain the frequency-wavenumber spectrum; or use the Radon transform to search for the principal energy direction on the spatiotemporal plane;

[0106] S202. Extract the characteristic spatial wavenumbers of the principal energy directions in the spectral domain. and characteristic angular frequency Or extract the observed phase velocity C corresponding to the spatiotemporal slope. observed ;

[0107] S203. Construct Doppler characteristics such as the mean, variance, and robust statistics of observed phase velocities at the event level;

[0108] S204. Using 3D CNN or CNN+ and other temporal networks, the spatiotemporal profiles of multiple survey lines are jointly encoded to obtain high-dimensional spatiotemporal feature vectors;

[0109] S205. Doppler features, depth features, and environmental variables such as water depth, wind speed, and wind direction are concatenated to form event-level fusion features, which are then used for subsequent flow velocity / flow regression learning.

[0110] The Doppler spectrum analysis includes the following steps:

[0111] First, the spatial resolution Δs between pixel coordinates and actual spatial coordinates is determined based on camera calibration parameters, and the inter-frame time interval Δt is determined according to the video frame rate. Gray values, texture intensity, or wavefront response intensity are extracted from the water surface ripple video sequence along the main channel direction or a preset survey line direction to construct a spatiotemporal profile image, as shown in the following formula:

[0112] I (x i ,t n )

[0113] in, .

[0114] In the formula, Δs is the scaling factor between the image pixel coordinate system and the world coordinate system, in meters per pixel; Δt is the inter-frame time interval of the video, in seconds per frame; x i = i·Δs represents the spatial position along the survey line, i = 0,1,…,N x-1 ;t n = n·Δt represents time, n = 0,1,…,N t-1 .

[0115] Then, a two-dimensional Fourier transform is performed on the spatiotemporal profile image to obtain the frequency-wavenumber spectrum:

[0116]

[0117] in, ; .

[0118] In the formula, For frequency-wavenumber spectrum, Let p be the p-th discrete spatial wavenumber, where p = 0, 1, ..., N. x-1 , ; For spatial frequency, ; f is the q-th discrete angular frequency, in rad / s; q This refers to the time frequency, measured in Hz.

[0119] Furthermore, the frequency-wavenumber energy spectrum is defined as follows:

[0120]

[0121] Extract the main energy direction from the energy spectrum:

[0122]

[0123]

[0124] in, and These are the spatial and temporal frequency indices corresponding to the main energy direction, respectively. and These are the characteristic spatial wavenumber and characteristic angular frequency corresponding to the principal direction of this energy, respectively.

[0125] Based on the wavenumber of the feature space and characteristic angular frequency Calculate the phase velocity of the observed wave:

[0126]

[0127] Based on the current water depth h obtained from the water level measuring device, calculate the theoretical wave velocity in still water under shallow water long-wave conditions:

[0128]

[0129] In the formula, g is the acceleration due to gravity, and h is the current water depth.

[0130] The physical inversion module obtains the initial physical estimate of the river surface velocity from the difference between the observed wave phase velocity and the theoretical wave velocity in still water.

[0131]

[0132] The deep learning flow rate / flow regression module will U phy C observed , , Doppler spectral eigenvectors Deep learning spatiotemporal feature vectors and environmental feature vectors The common input is the trained deep learning regression model:

[0133]

[0134] In the formula, U final f represents the final surface flow rate corrected by deep learning. θ ( ) represents a deep learning regression model with parameter θ; The Doppler spectral eigenvector can include C observed The mean, variance, phase velocity distribution on multiple measurement lines, and spectral energy concentration, etc. These are deep feature vectors extracted by 3D CNN, CNN-LSTM, or other spatiotemporal coding networks; This is an environmental feature vector, which may include water depth h, wind speed, wind direction, light intensity, etc.

[0135] The average flow velocity of the cross-section needs to be corrected according to the shape of the cross-section at the site. The specific correction formula is as follows:

[0136]

[0137] In the formula, α is the cross-sectional average velocity, in m / s; α is the correction factor for converting surface velocity to cross-sectional average velocity, which is dimensionless and can be determined through field calibration, historical flow measurement data, or hydraulic models.

[0138] The average flow rate of the cross section can be derived from the average flow velocity corrected for the cross section.

[0139]

[0140] In the formula, Q is the cross-sectional flow rate of the river, with the unit being m³ / s, and A(h) represents the cross-sectional area of ​​the river channel below the water surface at a water depth of h.

[0141] Flow calculation and output unit: combining current water depth h, cross-sectional geometric parameters, and It calculates the flow rate of a cross-section and stores, displays, or transmits the data remotely.

[0142] Optionally, the river surface velocity and flow measurement device based on active shock wave and spatiotemporal Doppler-deep learning fusion analysis of the present invention further includes an FPGA real-time processing unit: real-time processing of video data streams is realized based on field-programmable gate array (FPGA), including spatiotemporal profile generation, Doppler spectrum operation and deep learning inference, reducing power consumption and improving system real-time performance and integration.

[0143] Although the above embodiments have been shown and described, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Any changes, modifications, substitutions and variations made to the above embodiments by those skilled in the art are within the protection scope of the present invention.

Claims

1. A device for measuring river surface velocity and flow rate based on active shock wave and spatiotemporal Doppler-deep learning fusion analysis, characterized in that, include: A vibratory body launching device, wherein the vibratory body launching device is installed on the bank of a river; The data acquisition device is used to acquire video sequences of water surface ripples and water level data generated by the exciter transmitter. The data processing and control module is used to preprocess the acquired water surface ripple video sequence to construct a spatiotemporal profile image. A Doppler-deep learning fusion feature construction module is used to concatenate parameters of the processed data to construct event-level fusion features for flow velocity inversion. The physical inversion module calculates the theoretical phase velocity under still water conditions based on water level data and obtains the initial physical estimate of the river surface velocity. A deep learning flow velocity / flow regression module is used to input the physical initial estimate and event-level fusion features into a trained regression model and output the corrected river surface flow velocity. A flow calculation and output unit is used to calculate river flow and store, display or remotely transmit it.

2. The river surface velocity and flow measurement device based on active shock wave and spatiotemporal Doppler-deep learning fusion analysis according to claim 1, characterized in that, The exciter launching device includes: A base, which is fixed to the riverbank; A support pole is vertically fixed above the base; A cantilever crossbar is vertically installed on the upper side of the support column and extends horizontally above the river channel measurement section. The exciter release mechanism is electromagnetically controlled and is installed below the end of the cantilever crossbar. The exciter is a spherical counterweight, which is placed inside the exciter release mechanism.

3. The river surface velocity and flow measurement device based on active shock wave and spatiotemporal Doppler-deep learning fusion analysis according to claim 2, characterized in that, The data acquisition device includes: A high-speed camera is mounted on a cantilever crossbar, and its imaging range covers the area where the exciter enters the water. A water level measuring device is installed at the end of a cantilever crossbar, located between a high-speed camera and an exciter release mechanism, and is used to acquire real-time water depth data of the measuring section. The synchronous triggering unit is used to transmit a signal to the high-speed camera via a microcontroller at the same time as the release mechanism releases the exciter, thereby triggering the high-speed camera to start recording.

4. The river surface velocity and flow measurement device based on active shock wave and spatiotemporal Doppler-deep learning fusion analysis according to claim 3, characterized in that, The exciter release mechanism includes a hollow cylindrical shell with an open bottom. The top of the cylindrical shell is fixedly connected to the cantilever crossbar via a connecting rod. An exciter delivery assembly is fixedly installed inside the cylindrical shell.

5. The river surface velocity and flow measurement device based on active shock wave and spatiotemporal Doppler-deep learning fusion analysis according to claim 4, characterized in that, The exciter delivery assembly is fixed to the top of the inside of the cylindrical shell. The bottom of the exciter delivery assembly is open, and an electromagnetic switch is installed at the bottom opening. The exciter is placed inside the exciter delivery assembly.

6. The river surface velocity and flow measurement device based on active shock wave and spatiotemporal Doppler-deep learning fusion analysis according to claim 1, characterized in that, The spatiotemporal profile image is generated by the following steps: S101. Collect a large number of ripple video frames excited by the exciter under different hydrological and lighting conditions; S102. Generate corresponding wavefront position binary mask labels for each frame of image through manual annotation or semi-automatic algorithm, and generate a standard sample library; perform wavefront pixel-level segmentation on each frame of image through convolutional neural network to obtain the binary mask of the ripple region. S103. Extract brightness / texture values ​​only along the main direction and several transverse or oblique profiles in the ripple region with high segmentation confidence, and stack them over time to form a spatiotemporal profile image; S104. Using the ripple video frame as input and the corresponding binary mask of the front position of the water wave as the supervision label, train the CNN-UNet model using a deep learning framework until its loss function converges.

7. The river surface velocity and flow measurement device based on active shock wave and spatiotemporal Doppler-deep learning fusion analysis according to claim 6, characterized in that, The construction of event-level fusion features for flow velocity inversion includes Doppler spectral analysis and deep learning spatiotemporal feature encoding and fusion, specifically including the following steps: S201. Perform a two-dimensional fast Fourier transform or short-time Fourier transform on the spatiotemporal profile to obtain the frequency-wavenumber spectrum; or use the Radon transform to search for the principal energy direction on the spatiotemporal plane; S202. Extract the characteristic angular frequencies of the principal energy directions in the spectral domain. With characteristic space wavenumber Or extract the observed phase velocity C corresponding to the spatiotemporal slope. observed ; S203. Construct Doppler characteristics such as the mean, variance, and robust statistics of observed phase velocities at the event level; S204. Using 3D CNN or CNN+ and other temporal networks, the spatiotemporal profiles of multiple survey lines are jointly encoded to obtain high-dimensional spatiotemporal feature vectors; S205. Doppler features, deep features, and environmental variables are concatenated to form event-level fusion features, which are then used for subsequent velocity / flow regression learning.

8. The river surface velocity and flow measurement device based on active shock wave and spatiotemporal Doppler-deep learning fusion analysis according to claim 7, characterized in that, The formula for the frequency-wavenumber spectrum is: In the formula, For frequency-wavenumber spectrum, Let p be the p-th discrete spatial wavenumber, where p = 0, 1, ..., N. x-1 ; Let q be the qth discrete angular frequency; ; Δs is the scaling factor between the image pixel coordinate system and the world coordinate system, in meters per pixel; Δt is the inter-frame time interval of the video, in seconds per frame; x i = i·Δs represents the spatial position along the survey line, i = 0,1,…,N x-1 ;t n = n·Δt represents time, n = 0,1,…,N t-1 p is the spectral number in the spatial direction; q is the spectral number in the temporal direction.

9. The river surface velocity and flow measurement device based on active shock wave and spatiotemporal Doppler-deep learning fusion analysis according to claim 8, characterized in that, The calculation of the river surface velocity includes the following steps: S1. Based on the wavenumber of the characteristic space and characteristic angular frequency Calculate the phase velocity of the observed wave: In the formula, C observed Indicates the observed wave phase velocity; S2. Calculate the theoretical wave velocity in still water under shallow water long-wave conditions based on the water depth h obtained from the water level measuring device: In the formula, g is the acceleration due to gravity, and h is the current water depth; S3. Calculate the physical initial estimate U of the river surface velocity. phys The formula is: In the formula, U phys C0 is the initial physical estimate of the surface velocity of the river, and C0 is the theoretical still water wave velocity.

10. The river surface velocity and flow measurement device based on active shock wave and spatiotemporal Doppler-deep learning fusion analysis according to claim 9, characterized in that, The average flow rate of the river cross-section is: In the formula, Q represents the cross-sectional flow of the river. Let A(h) be the cross-sectional average flow velocity, and let A(h) represent the cross-sectional area of ​​the river channel below the water surface at a water depth of h.

Images