River flow prediction method and device based on acoustic wave inversion, equipment and medium

Abstract

This invention provides a method, apparatus, equipment, and medium for river flow forecasting based on acoustic wave inversion, comprising: acquiring acoustic signals non-contactly collected by a hydrophone array; preprocessing and extracting features from the acoustic signals to obtain acoustic features; constructing a physically constrained deep learning model; the physically constrained deep learning model takes acoustic features and water level as input and cross-sectional average flow velocity as output, and during model training, incorporates calculation deviations determined according to a preset calculation method as physical constraints into the loss function to obtain a trained physically constrained deep learning model; inputting the real-time acquired acoustic signals, after preprocessing and feature extraction, into the trained physically constrained deep learning model for inversion to obtain real-time flow; and generating flow forecast results for multiple future times by rolling updates based on the real-time flow. This method achieves non-contact, all-weather monitoring, exhibits small inversion errors in complex environments, strong extrapolation capability under extreme conditions, and ensures physical consistency.

Description

Technical Field

[0001] This invention relates to the field of river flow forecasting technology, and in particular to a method, apparatus, equipment and medium for river flow forecasting based on acoustic inversion. Background Technology

[0002] River flow is a core parameter for hydrological forecasting and flood control scheduling. Traditional contact flow measurement equipment (such as propeller current meters) is easily damaged by harsh conditions such as high sediment content and ice periods; non-contact radar flow measurement is affected by water surface fluctuations and floating objects, resulting in decreased accuracy.

[0003] In related technologies, physical models (such as the Saint-Venant equations) rely on detailed parameters such as cross-sections and roughness, making them difficult to apply to areas without data; purely data-driven models (such as LSTM) lack physical constraints and have poor extrapolation capabilities for extreme floods; and physical-data fusion methods are mostly loosely coupled, with physical laws only used for post-processing verification and unable to be deeply embedded in the training process. Therefore, there is an urgent need for a non-contact, strongly physically constrained, and adaptable online monitoring and forecasting method for river flow in complex environments. Summary of the Invention

[0004] In view of this, the purpose of the present invention is to provide a method, apparatus, equipment and medium for river flow forecasting based on acoustic wave inversion, which reduces the loss generated by copper busbars by determining magnetic shielding.

[0005] In a first aspect, embodiments of the present invention provide a method for river flow forecasting based on acoustic wave inversion. The method includes: acquiring acoustic signals non-contactly collected by a hydrophone array; preprocessing and extracting features from the acoustic signals to obtain acoustic features; constructing a physically constrained deep learning model; the physically constrained deep learning model takes the acoustic features and water level as input and the cross-sectional average flow velocity as output, and during model training, incorporates the calculation deviation determined according to a preset calculation method as a physical constraint term into the loss function to obtain a trained physically constrained deep learning model; inputting the real-time collected acoustic signals, after preprocessing and feature extraction, into the trained physically constrained deep learning model for inversion to obtain real-time flow; and connecting the real-time flow to a hydrological forecasting model, generating flow forecast results for multiple future times through rolling forecast updates.

[0006] In a preferred embodiment of the present invention, the method for deploying the hydrophone array includes: deploying the hydrophone array on stable bank slopes on both sides of the river channel, with the array direction parallel to the river flow direction; dynamically adjusting the water depth of the hydrophones according to the current water level so that they are located in the mainstream area with a smaller velocity gradient; and determining the spacing between adjacent hydrophones based on the spatial sampling theorem so that the array length covers the main area of ​​velocity change in the river cross section.

[0007] In a preferred embodiment of the present invention, the above-mentioned preprocessing of the acoustic signal includes: performing wavelet multi-scale decomposition on the acquired acoustic signal to obtain low-frequency approximation coefficients and multiple high-frequency detail coefficients; estimating the noise standard deviation of the high-frequency detail coefficients at each decomposition scale using the median method, and setting an adaptive threshold based on the characteristic frequency band; performing thresholding on the high-frequency detail coefficients at each scale using a soft threshold function based on the adaptive threshold, while keeping the low-frequency approximation coefficients unchanged; and performing inverse wavelet transform on the processed high-frequency detail coefficients at each scale and the original low-frequency approximation coefficients to reconstruct the denoised acoustic signal.

[0008] In a preferred embodiment of the present invention, the above-mentioned feature extraction of the acoustic signal includes: performing a fast Fourier transform on the preprocessed acoustic signal to extract frequency domain features; performing a short-time Fourier transform on the preprocessed acoustic signal to generate a time-frequency graph, and extracting time-frequency domain depth features from the time-frequency graph through a pre-trained convolutional neural network; and calculating the time-domain statistics and entropy features of the preprocessed acoustic signal to obtain statistical features.

[0009] In a preferred embodiment of the present invention, the aforementioned physical constraint deep learning model includes: an input layer for receiving fused features composed of acoustic features and water level values; a multi-layer LSTM temporal coding layer for capturing the temporal dynamic relationship of the fused features and outputting high-dimensional temporal features; a fully connected output layer for mapping the high-dimensional temporal features to cross-sectional average flow velocity; and a physical constraint layer for calculating the Manning formula theoretical flow velocity based on the input water level value and pre-stored cross-sectional morphology data, and calculating the deviation between the theoretical flow velocity and the cross-sectional average flow velocity output by the physical constraint deep learning model as a physical loss term.

[0010] In a preferred embodiment of the present invention, the above-mentioned method of preprocessing and extracting features from the real-time acquired acoustic signal and then inputting it into a trained physical constraint deep learning model for inversion to obtain the real-time flow rate includes: preprocessing and extracting features from the real-time acquired acoustic signal and inputting it together with the synchronously acquired water level value into the trained physical constraint deep learning model to obtain the current cross-sectional average flow velocity; determining the cross-sectional area based on the synchronously acquired water level value and pre-stored river cross-sectional morphology data; and obtaining the real-time flow rate based on the current cross-sectional average flow velocity and the cross-sectional area.

[0011] In a preferred embodiment of the present invention, real-time flow is input into a hydrological forecasting model, and flow forecast results for multiple future times are generated through rolling forecast updates. This includes: inputting real-time flow into the hydrological forecasting model to update historical flow sequences; inputting the updated historical flow sequences, meteorological forcing data, and acoustic features into a trained hydrological forecasting model; and performing iterative multi-step prediction using the hydrological forecasting model. Specifically, after predicting the flow at time t+1, the flow at time t+1 is used as one of the inputs for the next time step to predict the flow at time t+2, and the flow forecast sequence for multiple future times is recursively generated.

[0012] Secondly, embodiments of the present invention also provide a river flow forecasting device based on acoustic wave inversion. The device includes: an acoustic signal acquisition module for acquiring acoustic signals non-contactly collected by a hydrophone array; a processing module for preprocessing and feature extraction of the acoustic signals to obtain acoustic features; a model building module for constructing a physically constrained deep learning model; the physically constrained deep learning model takes acoustic features and water level as input and cross-sectional average flow velocity as output, and during model training, incorporates the calculation deviation determined according to a preset calculation method as a physical constraint term into the loss function to obtain a trained physically constrained deep learning model; a real-time flow acquisition module for inputting the real-time collected acoustic signals, after preprocessing and feature extraction, into the trained physically constrained deep learning model for inversion to obtain real-time flow; and a flow forecast result generation module for connecting the real-time flow to the hydrological forecasting model and generating flow forecast results for multiple future times through rolling forecast updates.

[0013] Thirdly, embodiments of the present invention also provide an electronic device, including a processor and a memory, wherein the memory stores computer-executable instructions that can be executed by the processor, and the processor executes the computer-executable instructions to implement the river flow forecasting method based on acoustic inversion described in the first aspect above.

[0014] Fourthly, embodiments of the present invention also provide a computer-readable storage medium storing computer-executable instructions. When the computer-executable instructions are invoked and executed by a processor, the computer-executable instructions cause the processor to implement the river flow forecasting method based on acoustic wave inversion described in the first aspect.

[0015] The embodiments of the present invention bring the following beneficial effects: This invention provides a method, apparatus, equipment, and medium for river flow forecasting based on acoustic wave inversion. It acquires acoustic signals non-contactly via a hydrophone array, preprocesses and extracts features from the acoustic signals to obtain acoustic features, and constructs a physically constrained deep learning model. The physical constrained deep learning model takes acoustic features and water level as input and average cross-sectional flow velocity as output. During model training, calculation deviations determined according to a preset calculation method are included as physical constraints in the loss function, resulting in a trained physical constrained deep learning model. Real-time acquired acoustic signals, after preprocessing and feature extraction, are input into the trained physical constrained deep learning model for inversion to obtain real-time flow. The real-time flow is then integrated into a hydrological forecasting model, and flow forecasts for multiple future time periods are generated through rolling updates. This method achieves non-contact, all-weather monitoring, exhibits small inversion errors in complex environments, strong extrapolation capabilities under extreme conditions, and ensures physical consistency.

[0016] Other features and advantages of this disclosure will be set forth in the following description, or some features and advantages may be inferred from the description or determined without doubt, or may be learned by practicing the techniques described above.

[0017] To make the above-mentioned objects, features and advantages of this disclosure more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0018] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 A flowchart illustrating a river flow forecasting method based on acoustic inversion, provided for an embodiment of the present invention; Figure 2 This is a schematic diagram of the hydrophone array layout provided in an embodiment of this application; Figure 3 An overall architecture diagram for the physical constraint deep learning model provided in this embodiment of the invention; Figure 4 A flowchart of another river flow forecasting method based on acoustic inversion provided in an embodiment of the present invention; Figure 5 A flowchart illustrating another method for river flow forecasting based on acoustic inversion provided in this embodiment of the invention; Figure 6 A flowchart of the rolling forecast provided in this embodiment of the invention; Figure 7 A schematic diagram of a river flow forecasting device based on acoustic inversion provided in an embodiment of the present invention; Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Dynamic calibration of hydrological model parameters and runoff forecasting for data-free areas are current frontier challenges and urgent engineering needs in hydrological research. The International Association of Hydrological Sciences (IAHS) has listed "Data-Free Watershed Forecasting" (PUB) as a core scientific problem in 21st-century hydrological research. Globally, a large number of small and medium-sized watersheds, transboundary rivers, and remote mountainous areas lack measured hydrological data, making traditional hydrological models difficult to apply due to the inability to effectively calibrate parameters. In my country, approximately 60% of small and medium-sized rivers have insufficient hydrological station density, severely restricting flood warnings, water resource allocation, and flood and drought disaster prevention capabilities during the flood season. With the intensification of global climate change and the increasing frequency of extreme hydrological events, higher demands are placed on runoff forecasting capabilities in data-free areas.

[0022] The existing technology has the following main shortcomings: (1) Traditional physical models rely on high-precision inputs, making them difficult to apply to areas without data. Although physical models, such as the Saint-Venant equations, have clear mechanisms, they require detailed parameters such as river cross-sections, roughness, gradient, and soil properties, which are difficult to obtain in areas without data. Although related technologies have attempted to solve the data gap problem through physical information neural networks (PINN), their training data still relies on simulations generated by physical models such as HEC-RAS. Essentially, they are still limited by simulation accuracy and have a systematic bias between simulation and reality, making it impossible to truly get rid of the dependence on physical model parameters.

[0023] (2) Pure data-driven models lack physical constraints, resulting in poor extrapolation and transferability. Deep learning models, such as LSTM, demonstrate strong nonlinear fitting capabilities in runoff prediction. However, pure data-driven methods rely entirely on historical observation data, exhibiting poor extrapolation capabilities under extreme conditions (such as floods exceeding standard levels) and lacking physical interpretability. When applied to areas without data, the lack of local training data severely hinders the model's generalization ability, making it difficult to achieve cross-basin transfer applications.

[0024] (3) The existing physical-data fusion methods are not deep enough and fail to achieve true mechanism embedding. Some studies have attempted to combine physical models with deep learning models, but most of them adopt the approach of "outcome selection" or "serial prediction". Physical laws are only used as the basis for post-processing verification or model selection, and physical constraints are not deeply embedded in the model training process. This loose coupling method is difficult to guarantee the physical rationality of the model output. Especially in areas without data and lacking measured data constraints, the model is prone to producing physically distorted prediction results.

[0025] In summary, it is necessary to develop a new runoff forecasting method that does not rely on measured hydrological data, can deeply integrate physical mechanisms and data-driven approaches, and has good extrapolation capabilities and cross-basin migration performance, in order to solve the "bottleneck" problem of hydrological forecasting in areas without data.

[0026] Based on this, the present invention provides a method, apparatus, equipment, and medium for river flow forecasting based on acoustic inversion. This method acquires acoustic signals non-contactly via a hydrophone array, preprocesses and extracts features from the acoustic signals to obtain acoustic features, and constructs a physically constrained deep learning model. The physical constrained deep learning model takes acoustic features and water level as input and average cross-sectional flow velocity as output. During model training, calculation deviations determined according to a preset calculation method are included as physical constraints in the loss function, resulting in a trained physical constrained deep learning model. The real-time acquired acoustic signals, after preprocessing and feature extraction, are input into the trained physical constrained deep learning model for inversion to obtain real-time flow. The real-time flow is then integrated into a hydrological forecasting model, and by continuously updating the forecast, flow forecasts for multiple future time periods are generated. This method achieves non-contact, all-weather monitoring, exhibits small inversion errors in complex environments, strong extrapolation capability under extreme conditions, and ensures physical consistency.

[0027] To facilitate understanding of this embodiment, a method for predicting river flow based on acoustic inversion disclosed in this embodiment of the invention will first be described in detail.

[0028] Example 1 This invention provides a method for river flow forecasting based on acoustic inversion. Figure 1 A flowchart illustrating a river flow forecasting method based on acoustic inversion, provided as an embodiment of the present invention. Figure 1 As shown, the river flow forecasting method based on acoustic inversion may include the following steps: Step S101: Acquire the acoustic signal obtained by non-contact acquisition of the hydrophone array.

[0029] Among them, the acoustic signals acquired by the hydrophone array without contact are passive acoustic waves naturally generated by the water flow.

[0030] The generation of sound waves in flowing water mainly stems from the following three physical mechanisms: (1) Turbulent pressure fluctuations. Turbulent vortex motion in water flow generates pressure fluctuations. The relationship between sound pressure level p and flow velocity U can be expressed as: ,in, For water density, Mach number. Dominant frequency of sound wave and integral scale of turbulence. Related: This part of the sound wave carries information about the mainstream flow velocity.

[0031] (2) Riverbed friction and particle collision. The interaction between water flow and the riverbed and bedload particles generates broadband sound waves, with a sound pressure level of [missing information]. It is positively correlated with flow velocity U and riverbed roughness n: Where R is the hydraulic radius. This part of the acoustic wave reflects the resistance characteristics of the riverbed.

[0032] (3) Water surface ripples and bubble oscillations. Water surface ripples and bubble bursts generate high-frequency sound waves (>1 kHz), which are related to the degree of air entrainment in the water flow. The acquisition system uses synchronous sampling, with a sampling frequency of... The sampling accuracy is no less than 24 bits to ensure the complete recording of weak sound wave signals.

[0033] Regarding the selection of hydrophones, wide-bandwidth, high-sensitivity piezoelectric hydrophones can be chosen, with a frequency response range of 0.1Hz to 10kHz and a sensitivity of not less than -170dBre1V / μPa. This frequency band covers the main sound wave energy generated by the friction between water flow and the riverbed and riverbank, while effectively avoiding low-frequency interference from wind and wave noise (mainly concentrated in <10Hz) and mechanical vibration noise (usually <100Hz).

[0034] Specifically, the deployment method of the hydrophone array may include: deploying the hydrophone array on stable bank slopes on both sides of the river channel, with the array direction parallel to the river flow direction; dynamically adjusting the water depth of the hydrophones according to the current water level so that they are located in the mainstream area with a small velocity gradient; and determining the spacing between adjacent hydrophones based on the spatial sampling theorem so that the array length covers the main area of ​​velocity change in the river cross section.

[0035] For ease of understanding, Figure 2 This is a schematic diagram of the hydrophone array layout provided in the embodiments of this application, as shown below. Figure 2 As shown, a linear hydrophone array is deployed on each of the stable bank slopes on both sides of the river channel, with the array direction parallel to the river flow direction. The deployment scheme meets the following principles: (1) Non-contact deployment on the shore. The hydrophone is fixed to the shore extension bracket, with an immersion depth of [missing information]. satisfy ,in, This is the current water level. This is the relative immersion depth coefficient (taken as 0.3~0.6). To avoid a safety margin (0.2~0.5 m) in the water surface boundary layer, the design ensures that the hydrophone is located in the mainstream region with a small velocity gradient, while avoiding the disturbance layer of wind and waves on the water surface.

[0036] (2) Array spacing optimization. Spacing between adjacent hydrophones. Based on the spatial sampling theorem and the determination of acoustic wave coherence length: ,in, The speed of sound in water, The highest frequency of the target signal. The coherence coefficient (taken as 0.5~0.8). When hour, However, considering the ease of project implementation, the actual cost was... Insufficient spatial sampling can be compensated for by array signal processing algorithms (such as beamforming).

[0037] (3) Array length and number. The array length L covers the main areas of flow velocity variation in the river cross-section: Where B is the width of the river channel. The coverage factor is 0.3 to 0.6.

[0038] Step S102: Preprocess and extract features from the acoustic signal to obtain acoustic features.

[0039] Specifically, preprocessing the acoustic signal may include: performing wavelet multi-scale decomposition on the acquired acoustic signal to obtain low-frequency approximation coefficients and multiple high-frequency detail coefficients; estimating the noise standard deviation of the high-frequency detail coefficients at each decomposition scale using the median method, and setting an adaptive threshold based on the characteristic frequency band; using a soft thresholding function to perform thresholding on the high-frequency detail coefficients at each scale based on the adaptive threshold, while keeping the low-frequency approximation coefficients unchanged; and performing inverse wavelet transform on the processed high-frequency detail coefficients at each scale and the original low-frequency approximation coefficients to reconstruct the denoised acoustic signal.

[0040] To extract effective underwater acoustic signals from complex environmental noise, the following multi-level suppression strategy can be adopted: (1) Spatial filtering. Using array beamforming technology, the target direction signal is enhanced through time delay compensation and weighted summation: ,in, For the i-th hydrophone to receive the signal, The time delay is relative to the reference point. These are the weighting coefficients. By designing directional beams, interference noise from non-target directions such as riverbanks and water surfaces can be effectively suppressed.

[0041] (2) Adaptive noise cancellation. A reference hydrophone (buried underground on the riverbank) is deployed to collect environmental vibration noise, and the relevant noise components are subtracted from the underwater acoustic signal through adaptive filtering: ,in For underwater acoustic signals, For reference noise, The adaptive filter coefficients are iteratively updated using the LMS algorithm.

[0042] (3) Time-frequency domain threshold denoising. Adaptive threshold denoising based on wavelet transform preserves the characteristic frequency bands related to flow velocity.

[0043] Step A1: Wavelet multi-scale decomposition. This involves processing the acquired raw hydrophone signal. Perform J-level wavelet decomposition to decompose it into a low-frequency approximation coefficient. and multiple high-frequency detail coefficients ( Different decomposition scales correspond to different frequency bands, with high-frequency detail coefficients mainly carrying the transient components of the signal, while low-frequency approximation coefficients reflect the macroscopic trend of the signal. Daubechies wavelet bases (such as db4) are used as decomposition basis functions because they have good time-frequency localization characteristics and are suitable for processing non-stationary underwater acoustic signals.

[0044] Step A2: Noise level estimation. At each decomposition scale j, using the detail coefficients... Estimate the noise standard deviation at this scale A robust median estimation method is used: This estimation method is not sensitive to a small number of outliers and can accurately reflect the background noise level at various scales.

[0045] Step A3: Adaptive Threshold Calculation. Calculate the corresponding denoising threshold based on the noise level at each scale. A soft threshold function is used, and the threshold calculation formula is as follows: ,in, The length of the detail coefficient at this scale. This is a frequency band adaptive adjustment factor. The key innovation lies in... Setting: For the decomposition scale within the characteristic frequency band (0.1Hz~5kHz), The values ​​are set to small (e.g., 0.6~0.8) to preserve the acoustic components related to flow velocity; for scales outside the characteristic frequency band, A larger value (e.g., 1.2~1.5) is used to enhance the noise suppression effect.

[0046] Step A4: Thresholding and Wavelet Coefficient Reconstruction. For the detail coefficients at each decomposition scale... Thresholding is performed. Specifically, for the detail coefficients of the j-th layer... When the absolute value of the coefficient is greater than the threshold corresponding to that scale When the absolute value of a coefficient is less than or equal to the threshold, it is shrunk by a threshold unit to zero, preserving its sign information and subtracting the threshold from its absolute value. When the absolute value of a coefficient is less than or equal to the threshold, it is directly set to zero. This soft-thresholding method ensures the continuity of the processed coefficients, avoiding the pseudo-Gibbs oscillations that may be introduced by abrupt changes at the threshold point in hard-thresholding, thus ensuring the smoothness and physical authenticity of the reconstructed signal waveform. Meanwhile, low-frequency approximation coefficients... Keep the original values ​​unchanged to fully preserve the macroscopic trend information of the signal—this part of the coefficients carries the overall energy characteristics of the water flow sound wave and has a high correlation with the average flow velocity of the cross section. Therefore, it is not advisable to perform threshold processing to avoid the loss of key information.

[0047] Step A5: Wavelet Inverse Transform Reconstructs the Signal. The detail coefficients at each scale, after soft thresholding, are input together with the original low-frequency approximation coefficients into the wavelet inverse transform to reconstruct the denoised underwater acoustic signal. The wavelet inverse transform is the reverse process of the forward decomposition, enabling precise synthesis of the coefficients at each decomposition scale into the original time domain. Due to the completeness and orthogonality of wavelet basis functions, the reconstruction process maintains the complete temporal synchronization of the signal, ensuring that the denoised signal is strictly aligned with the original signal in the time domain, providing a reliable data foundation for subsequent time-series analysis of acoustic features. Through the above processing, the denoised signal achieves a significant improvement in signal-to-noise ratio, effectively filtering out environmental interference such as wind, waves, and mechanical vibrations; simultaneously, the acoustic components within the flow velocity-related characteristic frequency band are preserved to the greatest extent, and the waveform distortion rate after denoising is controlled within 5%, providing high-quality, low-distortion source data with complete physical information preservation for subsequent acoustic feature extraction and flow inversion models.

[0048] (4) This application proposes a flow velocity-sound pressure spatial correlation function to guide array optimization, defining the cross-correlation function of the signals received by the $i$ and $j$-th hydrophones: Theoretical analysis shows that there is an analytical relationship between this correlation function and the cross-sectional velocity distribution U(y): ,in For the sound source intensity distribution, Let be the time delay difference between the sound waves propagating from position y to the two hydrophones. Based on this relationship, the flow velocity distribution can be inverted using the array cross-correlation matrix, providing physical prior knowledge for subsequent deep learning models.

[0049] Specifically, feature extraction of acoustic signals may include: performing a fast Fourier transform on the preprocessed acoustic signal to extract frequency domain features; performing a short-time Fourier transform on the preprocessed acoustic signal to generate a time-frequency graph, and extracting time-frequency domain depth features from the time-frequency graph using a pre-trained convolutional neural network; and calculating the time-domain statistics and entropy features of the preprocessed acoustic signal to obtain statistical features.

[0050] The goal of acoustic feature extraction is to comprehensively characterize the hydraulic information contained in the acoustic signal of water flow from multiple dimensions, including the time domain, frequency domain, and time-frequency domain. This application extracts the following three types of core features: (1) Frequency domain feature extraction Fast Fourier Transform is performed on the denoised underwater acoustic signal to extract the following frequency domain feature parameters.

[0051] ① Main frequency The frequency corresponding to the maximum amplitude in the spectrum represents the dominant frequency component of the water flow sound wave. Theoretical analysis shows that there is a negative correlation between the dominant frequency and the average flow velocity U at the cross section. This relationship stems from the interaction between the turbulent integral scale and the flow velocity.

[0052] ②-3dB bandwidth The bandwidth corresponds to the frequency range where the spectral amplitude drops to half of its maximum value, reflecting the concentration of sound wave energy. The bandwidth is related to the intensity of flow velocity pulsation; the more intense the pulsation, the wider the bandwidth.

[0053] ③ Sub-band energy ratio The 0–5 kHz frequency band is divided into K sub-bands (e.g., 0–500 Hz, 500–1000 Hz, …, 4500–5000 Hz), and the proportion of energy in each sub-band to the total energy is calculated. The energy distribution of different sub-bands has different sensitivities to parameters such as flow velocity, riverbed roughness, and sediment content, and can provide multi-dimensional hydraulic state information.

[0054] ④ Spectral moment characteristics: Calculate the zeroth moment of the spectrum (Total Energy), Second Moment (Spectral variance), fourth moment (Spectral kurtosis) characterizes the overall shape characteristics of the spectrum.

[0055] (2) Time-frequency domain feature extraction The time-frequency plot of the signal is constructed using short-time Fourier transform or wavelet transform, and high-dimensional deep features are extracted using deep learning techniques. ① Time-frequency graph generation: Perform a short-time Fourier transform on the signal, set the window length to 1024 points, the overlap rate to 50%, and generate a graph of size [size missing]. The time-frequency diagram matrix retains information in both the time and frequency dimensions.

[0056] ②CNN Deep Feature Extraction: The time-frequency map is input into a pre-trained convolutional neural network (such as VGG16 or ResNet18), and the feature map before the fully connected layers is extracted as a deep feature vector. This deep feature can automatically learn high-order nonlinear patterns in the time-frequency map and capture complex sound wave-flow mapping relationships that are difficult to represent using traditional hand-crafted features.

[0057] ③ Temporal feature fusion: The deep features of multiple consecutive time windows are spliced ​​together to form a temporal feature sequence, which provides input for the subsequent LSTM network to capture temporal correlations.

[0058] (3) Statistical characteristics and higher-order characteristics Extract the time-domain and higher-order statistical features of the signal as a supplement to the frequency-domain features: Time-domain statistics, such as mean, variance, root mean square, peak factor, skewness, and kurtosis, reflect the amplitude distribution characteristics of a signal.

[0059] Entropy characteristics: such as sample entropy, approximate entropy, and permutation entropy, characterize the complexity and irregularity of a signal and are closely related to turbulence intensity.

[0060] Fractal dimension: The fractal dimension of a signal is calculated using the box counting method, reflecting the self-similarity characteristics of the signal.

[0061] The core principle of constructing the source data feature library is to ensure the authenticity and reliability of the feature-traffic mapping relationship by basing it on real-world measured data. The specific construction process is as follows: (1) Synchronous Data Acquisition and Labeling. During the model training phase, the following three types of data are collected synchronously: ① Acoustic signal: acquired through a hydrophone array and obtained after noise reduction processing in step one; ② Water level data: collected synchronously through a shoreline water level gauge, with the sampling frequency consistent with the acoustic signal; ③True value of flow: The cross-sectional flow rate is obtained through precision measurement equipment such as ADCP and used as the true value label for model training.

[0062] All data is strictly synchronized with timestamps to ensure that the temporal correspondence between acoustic characteristics and true flow values ​​is accurate.

[0063] (2) Feature normalization and standardization. The extracted features are normalized to eliminate dimensional differences: ,in, The mean of the features, The standard deviation of the feature is denoted as . The normalized feature values ​​all follow a standard normal distribution, which is beneficial for the training convergence of deep learning models.

[0064] (3) Hierarchical Dataset Construction. Considering that the mapping relationship between acoustic features and flow rate may differ at different flow rates, the dataset is hierarchically divided according to flow rate: ① Low flow layer: The flow rate is less than 30% of the multi-year average flow rate; ②Middle flow layer: The flow rate is between 30% and 70% of the multi-year average flow rate; ③ High flow layer: The flow rate is greater than 70% of the multi-year average flow rate.

[0065] Within each traffic layer, the dataset is divided into training, validation, and test sets in an 8:1:1 ratio. This hierarchical partitioning strategy ensures that the model receives sufficient training samples at each traffic level, avoiding underfitting issues under extreme conditions.

[0066] (4) Data Augmentation and Enlargement. To address the issue of scarce samples in the low-flow layer, data augmentation techniques are employed to expand the sample size: ① Noise enhancement: Add artificial noise with different signal-to-noise ratios (10~30dB) to the original signal to enhance the model's robustness to noise; ②Time stretching: Stretch or compress the signal along the time axis (scaling factor 0.8~1.2) to simulate signal changes under different water flow velocities; ③ Spectrum shift: Apply small perturbations (perturbation amplitude less than 5%) to the wavelet coefficients in the frequency domain to enhance the generalization ability of the model.

[0067] Step S103: Construct a physical constraint deep learning model.

[0068] The physical constraint deep learning model takes acoustic wave features and water level as input and cross-sectional average flow velocity as output. During the model training process, the calculation deviation determined according to the preset calculation method is included as a physical constraint term in the loss function to obtain the trained physical constraint deep learning model.

[0069] The physical constraint deep learning model can include: an input layer for receiving fused features composed of acoustic features and water level values; a multi-layer LSTM temporal coding layer for capturing the temporal dynamic relationship of the fused features and outputting high-dimensional temporal features; a fully connected output layer for mapping the high-dimensional temporal features to the cross-sectional average flow velocity; and a physical constraint layer for calculating the Manning formula theoretical flow velocity based on the input water level value and pre-stored cross-sectional morphology data, and calculating the deviation between the theoretical flow velocity and the cross-sectional average flow velocity output by the physical constraint deep learning model as a physical loss term.

[0070] The physical constraint deep learning model in this application adopts a multi-input single-output architecture design. For ease of understanding, 4, as shown... Figure 3As shown, the physical constraint deep learning model includes: an input layer, where acoustic feature vectors and water level values ​​are input separately; a feature fusion layer, which merges the two types of features through a concatenation operation; an LSTM temporal coding layer, consisting of three stacked LSTM layers, which progressively extracts features at different time scales and ultimately outputs high-dimensional temporal features; an output layer, where a fully connected layer maps the LSTM output to the cross-sectional average flow velocity; and a physical constraint layer, which calculates the theoretical flow velocity using Manning's formula and compares it with the model's output flow velocity to obtain the physical loss; simultaneously, it combines the water level to calculate the actual flow rate and compares it with the ground truth to obtain the data loss; the total loss is fed back to update the model parameters, thus embedding the physical constraints.

[0071] Regarding input feature design and fusion: (1) Sound wave feature vector The input acoustic wave features include the three types of features extracted in step two: frequency domain features, dominant frequency, and so on. -3dB bandwidth Sub-band energy ratio ( ) Spectral moment characteristics Time-frequency domain deep features, which are D-dimensional deep feature vectors extracted from the time-frequency plot through a pre-trained CNN; statistical features, such as root mean square, skewness, kurtosis, sample entropy, and fractal dimension.

[0072] All features, after normalization, are concatenated into a sound wave feature vector. ,in, This represents the total dimension of the features.

[0073] (2) Water level input The synchronously acquired water level value H serves as an auxiliary input, forming the model's input space together with the acoustic features. The introduction of water level information serves the following purposes: it provides cross-sectional morphology information; combined with known cross-sectional geometry, the water level can be directly converted into the flow area. ; To assist in physical constraint calculations, the hydraulic radius R in the Manning formula needs to be determined based on the water level and cross-sectional shape; to enhance the physical consistency of the model, since there is a theoretical monotonic relationship between flow rate and water level, the introduction of water level can constrain the physical rationality of the model output.

[0074] (3) Feature fusion The acoustic characteristics and water level values ​​are fused using a splicing method: .

[0075] Regarding the LSTM time-series coding layer: Considering the complex time dynamic relationship between acoustic signals and flow velocity (such as the continuity of water flow and the lag in flood peak propagation), this application uses a multilayer long short-term memory network (LSTM) to capture temporal features.

[0076] (1) LSTM unit structure The LSTM unit effectively solves the gradient vanishing problem of traditional RNNs through a gating mechanism. Its core calculation is as follows: Input Gate: .

[0077] Candidate memories: .

[0078] Memory update: .

[0079] Output gate: .

[0080] Hidden state: .

[0081] (2) Multi-layer stacking design This application embodiment employs a 3-layer LSTM stacked structure, with parameters for each layer shown in Table 1 below. The multi-layer stacked structure enables the model to learn the sound wave-flow mapping relationship at different time scales. The first layer focuses on second-level pulsation features, while the second and third layers focus on minute-level process evolution features.

[0082] Table 1:

[0083] (3) Timing window settings The input time window length T is set to 10 minutes, and the sampling frequency is... =20Hz, meaning each input sample contains =Feature sequences corresponding to 12,000 sampling points. The window length was optimized to cover the main time scales of the water flow process while avoiding the introduction of too much redundant information.

[0084] Regarding the design of the physical constraint layer: The physical constraint layer is the core innovation of this invention. Its goal is to force the output results to conform to the basic physical laws of fluid mechanics during the model training process.

[0085] ① Selection of physical equation: Manning formula. This invention selects the Manning formula as the physical constraint, and its expression is: .

[0086] Where V is the average flow velocity across the cross section; n is the Manning roughness coefficient, reflecting the riverbed resistance characteristics; R is the hydraulic radius (m), R=A / P, where A is the cross-sectional area and P is the wetted perimeter; and S is the water surface gradient, reflecting the driving force of the flow. The core reason for choosing the Manning formula in this invention is that the natural sound waves of flowing water contain information about the riverbed roughness. The broadband components (>1kHz) in the sound wave signal are closely related to riverbed friction and particle collisions, and the roughness coefficient $n$ in the Manning formula is precisely the parameter characterizing this physical process. By learning the equivalent representation of n from the sound wave characteristics through a deep learning model, a deep fusion of physical equations and data features can be achieved.

[0087] ①Design of physical loss function The core of the physical constraint layer is to use the deviation between the theoretical flow velocity calculated by the Manning formula and the flow velocity predicted by the model as the loss term.

[0088] First, based on the input water level H and the pre-stored river channel cross-sectional morphology data, the cross-sectional area of ​​the water passage is calculated. and hydraulic radius The cross-sectional shape is determined by N pre-surveyed cross-sectional points. Determine the width of the water surface corresponding to water level $H$. ,area Wet Week It can be obtained through linear interpolation or numerical integration.

[0089] Then, calculate the theoretical flow rate according to Manning's formula: ,in The equivalent roughness coefficient, in this invention, is not a fixed constant but is implicitly learned by the model from acoustic wave characteristics. This design enables the model to dynamically adjust based on real-time acoustic wave signals. This reflects changes in riverbed conditions (such as increased roughness during the flood season and decreased roughness during the ice season).

[0090] Finally, define the physical loss function: .

[0091] ③ Total Loss Function. The model's total loss function is a weighted average of the data fitting term and the physical constraint term: .

[0092] in, , ; This is the physical constraint weighting coefficient, used to balance the importance of data-driven approaches and physical laws. Its value range is usually 0.01 to 0.1.

[0093] Regarding model training strategies: (1) Training data preparation Using the source data feature library constructed in step two, the dataset is divided into training, validation, and test sets in an 8:1:1 ratio. The training set is used for model parameter updates, the validation set is used for hyperparameter tuning and early stopping detection, and the test set is used for final performance evaluation.

[0094] (2) Optimizer and learning rate The Adam optimizer is used for parameter updates, with an initial learning rate set to 0.001, and a cosine annealing learning rate decay strategy is employed. .

[0095] in, =0.001 is the initial learning rate. =1×10-5 is the minimum learning rate, and T is the total number of training rounds.

[0096] (3) Physical constraint weight adjustment strategy Physical constraint weights The settings directly affect the model's behavior. This application's embodiments employ a dynamic adjustment strategy: ① Initial training phase (first 20% of training): Set it to a small value (0.01) to allow the model to learn the data distribution first; ②Mid-training phase (20%~60% T): Gradually increase to 0.05, and introduce physical constraints to guide model learning; ③ Late training phase (60%~100% T): Keep it at 0.05 to balance the model between data fit and physical consistency.

[0097] Step S104: After preprocessing and feature extraction, the real-time acquired acoustic signal is input into the trained physical constraint deep learning model for inversion to obtain the real-time traffic flow.

[0098] Real-time flow inversion and online monitoring are the core application components of the technical solution in this application. This step involves deploying the offline-trained physical constraint deep learning model at the actual river monitoring site, continuously processing the acoustic signals collected in real time by the hydrophone array, achieving minute-level online flow inversion through model inference, and outputting the results to a data platform or connecting to a forecasting system.

[0099] Step S105: Integrate real-time flow data into the hydrological forecasting model and generate flow forecast results for multiple future time periods by continuously updating the forecast.

[0100] The high-precision flow data obtained through real-time inversion serves as the input or correction target for the hydrological forecasting model, enabling rolling predictions of future flow rates. Unlike traditional physical models, this step employs a Long Short-Term Memory (LSTM) network as the hydrological forecasting model, fully leveraging its advantages in time series forecasting and forming a closed-loop system of "data-driven + real-time correction" with acoustic inversion technology.

[0101] The river flow forecasting method based on acoustic inversion provided in this invention can acquire acoustic signals non-contactly using a hydrophone array, preprocess and extract features from the acoustic signals to obtain acoustic features, and construct a physically constrained deep learning model. The physical constrained deep learning model takes acoustic features and water level as input and average cross-sectional flow velocity as output. During model training, calculation deviations determined according to a preset calculation method are included as physical constraints in the loss function to obtain a trained physical constrained deep learning model. The real-time acquired acoustic signals, after preprocessing and feature extraction, are input into the trained physical constrained deep learning model for inversion to obtain real-time flow. The real-time flow is then integrated into a hydrological forecasting model, and flow forecasts for multiple future time periods are generated through rolling updates. This method achieves non-contact, all-weather monitoring, has small inversion errors in complex environments, strong extrapolation capabilities under extreme conditions, and ensures physical consistency.

[0102] Example 2 This invention also provides another method for river flow forecasting based on acoustic wave inversion. This method is implemented based on the method in the above embodiments. The method focuses on describing the specific implementation of the real-time acquisition of acoustic wave signals, which are preprocessed and feature extracted, and then input into a trained physical constraint deep learning model for inversion to obtain the real-time flow.

[0103] Figure 4 A flowchart of another river flow forecasting method based on acoustic inversion provided in an embodiment of the present invention is shown below. Figure 4 As shown, the real-time acquired acoustic signal, after preprocessing and feature extraction, is input into a trained physical constraint deep learning model for inversion to obtain the real-time traffic flow. This process may include the following steps: Step S201: After preprocessing and feature extraction, the real-time acquired acoustic signal is input into the trained physical constraint deep learning model along with the synchronously acquired water level value to obtain the current cross-sectional average flow velocity.

[0104] Regarding real-time acoustic signal acquisition and preprocessing: The hydrophone array continuously acquires data at a fixed sampling frequency of fs=20kHz, and the data stream from each channel is transmitted to the industrial control computer in real time through a synchronous acquisition module. The acquisition system adopts a ring buffer design to cache the raw signal from the most recent T minutes (e.g., 10 minutes) to support sliding window processing. For each newly arriving data block (e.g., a data segment every 1 second), time-frequency domain threshold denoising processing is immediately performed.

[0105] Regarding sound wave feature extraction: On the denoised signal, a feature extraction algorithm is used to calculate the features in Table 2 below in real time.

[0106] Table 2:

[0107] Finally, the feature vector Xacoustic(t) at the current moment is concatenated with the synchronously collected water level H(t) to form the model input Xinput(t), which is a technological advantage.

[0108] Regarding model inference: The input vector Xinput(t) is fed into the physically constrained deep learning model that has been loaded into memory, and forward computation is performed: The model inference is based on the network structure and parameters determined during the training phase, and only includes forward propagation, resulting in low computational cost.

[0109] Step S202: Determine the cross-sectional area of ​​the water passage based on the synchronously collected water level values ​​and the pre-stored river cross-sectional morphology data.

[0110] Step S203: Based on the current average flow velocity and cross-sectional area, obtain the real-time flow rate.

[0111] Regarding flow calculation and output: (1) Flow calculation Combined with the cross-sectional average flow velocity output by the model Given the real-time water level H(t), the cross-sectional area A(H(t)) is calculated using pre-stored river channel cross-sectional morphology data, and then the real-time flow rate is obtained: The cross-sectional morphology data was obtained through preliminary surveying.

[0112] (2) Output frequency and smoothing The system outputs a flow rate value every minute, which is the average of all valid predictions within that minute. When water level changes drastically or the quality of the acoustic signal deteriorates, the system can combine quality indicators to determine the validity of the data and remove or interpolate outliers.

[0113] (3) Data storage and uploading The flow rate, velocity, water level, and corresponding acoustic characteristics output every minute are stored in a local database in time series format and uploaded to a data center or hydrological forecasting platform in real time via 4G / 5G network.

[0114] Example 3 This invention also provides another method for river flow forecasting based on acoustic inversion; this method is implemented on the basis of the method in the above embodiments; the method focuses on describing the specific implementation of integrating real-time flow into the hydrological forecasting model and generating flow forecast results for multiple future times through rolling forecast updates.

[0115] Figure 5 A flowchart of another river flow forecasting method based on acoustic inversion provided in this embodiment of the invention is shown below. Figure 5 As shown, this method of integrating real-time flow data into a hydrological forecasting model and generating flow forecasts for multiple future time periods through rolling forecast updates can include the following steps: Step S301: Integrate real-time flow data into the hydrological forecasting model to update the historical flow sequence.

[0116] Step S302: Input the updated historical flow sequence, meteorological forcing data, and acoustic characteristics into the trained hydrological forecasting model.

[0117] Step S303: Use the hydrological forecasting model to perform iterative multi-step prediction.

[0118] In this process, after predicting the flow rate at time t+1, the flow rate at time t+1 is used as one of the inputs for the next time step to predict the flow rate at time t+2, and the flow rate forecast sequence for multiple future time steps is generated recursively.

[0119] Regarding the architecture of the LSTM hydrological forecasting model: The LSTM model is specifically designed for time series forecasting of river flow, and it can capture the nonlinear characteristics and long-term dependencies of the runoff process.

[0120] (1) Model input design The input to the LSTM forecast model can include the three types of data shown in Table 3 below.

[0121] Table 3:

[0122] The input vector can be represented as: Th is the length of the historical window, which is usually 24 to 72 hours.

[0123] (2) LSTM prediction process The LSTM forecasting model uses a deep neural network structure. Specifically, the model consists of an input layer, three LSTM layers, two Dropout layers, a fully connected layer, and an output layer.

[0124] The dimensionality of the input layer is determined by the number of input features, specifically including the sum of the dimensions of meteorological forcing variables (rainfall, evaporation, and temperature), the window length of the historical flow sequence, and the acoustic features. After the input layer, a first LSTM layer with 128 hidden units is connected, returning a complete output sequence to pass temporal information to subsequent networks. Following the first LSTM layer is a first Dropout layer with a dropout rate of 0.3 to prevent overfitting. Next is a second LSTM layer with 64 hidden units, also returning a complete output sequence, followed by another Dropout layer with a dropout rate of 0.3. The third LSTM layer has 32 hidden units, returning only the output of the last time step, compressing the entire time series information into a fixed-dimensional feature vector. This vector is input to a fully connected layer, undergoing a linear transformation and mapping to a single output value, i.e., the predicted future flow rate.

[0125] Based on this, the model adopts an iterative multi-step prediction strategy, that is, using the input data to predict the flow rate at time t+1; and then using the predicted values... Using this as input, and combining it with other known variables, predict the flow rate at time t+2; repeat the above process to generate the flow rate sequence for the entire forecast period: ,in, The forecast period length can be set according to business needs (e.g., 24 hours, 72 hours).

[0126] (3) Real-time rolling forecast process For ease of understanding, Figure 6 A flowchart of the rolling forecast provided in the embodiments of the present invention, such as Figure 6 As shown, the rolling forecast is executed cyclically with a fixed time step (e.g., 1 hour), forming a system of "acoustic wave inversion → model update → rolling prediction → re-inversion".

[0127] Main loop (solid arrow): A complete rolling forecast is executed every hour, starting from obtaining the latest acoustic inversion flow and ending with the output of the new forecast results, forming a continuously updated closed loop.

[0128] Incremental learning (dashed arrow): Triggered daily at set times, it uses newly accumulated data to incrementally train the LSTM model, update the model parameters, and enable the model to adapt to changes in the watershed.

[0129] Prediction-driven: The updated historical flow sequence, together with meteorological forcing and acoustic features, constitutes the input of the LSTM model, driving the model to iteratively generate flow forecasts for multiple future time periods.

[0130] Specific operations: Forecast frequency: Forecast results are updated every hour. Forecast period: Set according to business needs; flood warnings are generally 24 to 72 hours. Output content: Hourly flow rate process line for the foreseeable future period.

[0131] Regarding the evaluation of forecast results: The system evaluates the forecast results in real time for each round and calculates the indicators in Table 4 below.

[0132] Table 4:

[0133] Through the rolling forecasting mechanism, this application embodiment deeply integrates acoustic wave inversion technology with LSTM deep learning forecasting model, forming a complete technology chain of "acoustic wave acquisition → physical constraint inversion → deep learning rolling forecasting", providing a brand-new solution for the construction of smart water conservancy and digital twin watersheds.

[0134] The above embodiments achieve the following effects: (1) Non-contact and highly safe. The hydrophone is installed on the shore and has no contact with the water, which completely avoids the risk of equipment damage such as impact from high-velocity floating objects and abrasion from high sand content during the flood season, and realizes true all-weather unattended monitoring.

[0135] (2) Strong adaptability to complex environments. Utilizing the stable propagation characteristics of sound waves in water, it is less affected by water surface fluctuations, floating objects, and severe weather. Under conditions where traditional radar current measurement fails, such as high sediment content, low flow velocity, and ice periods, the inversion error is small, significantly better than existing non-contact technologies.

[0136] (3) Physical constraints ensure high accuracy. The Manning formula is embedded into the deep learning loss function, and the physical mechanism and data-driven approach are deeply integrated by dynamically correcting the equivalent roughness. The extrapolation capability under extreme conditions is significantly enhanced.

[0137] (4) Integrated monitoring-forecasting closed loop. Real-time inverted flow is integrated into the LSTM rolling forecasting system to realize a complete technology chain of "acoustic wave acquisition → physical constraint inversion → rolling forecasting", which improves the accuracy of flood forecasting and buys valuable time for flood control scheduling.

[0138] Example 4 Corresponding to the above method embodiments, this invention provides a river flow forecasting device based on acoustic wave inversion. Figure 7 A schematic diagram of a river flow forecasting device based on acoustic inversion provided in an embodiment of the present invention is shown below. Figure 7 As shown, the river flow forecasting device based on acoustic inversion may include: The acoustic signal acquisition module 401 is used to acquire acoustic signals non-contactly collected by the hydrophone array.

[0139] The processing module 402 is used to preprocess and extract features from the acoustic signal to obtain acoustic features.

[0140] The model building module 403 is used to build a physical constraint deep learning model. The physical constraint deep learning model takes acoustic features and water level as input and cross-sectional average flow velocity as output. During the model training process, the calculation deviation determined according to the preset calculation method is included as a physical constraint term in the loss function to obtain the trained physical constraint deep learning model.

[0141] The real-time flow acquisition module 404 is used to input the real-time acquired acoustic wave signal into the trained physical constraint deep learning model after preprocessing and feature extraction to obtain the real-time flow.

[0142] The flow forecast result generation module 405 is used to input real-time flow into the hydrological forecast model and generate flow forecast results for multiple future times by rolling forecast updates.

[0143] The river flow forecasting device based on acoustic inversion provided in this invention can acquire acoustic signals non-contactly from a hydrophone array, preprocess and extract features from the acoustic signals to obtain acoustic features, and construct a physically constrained deep learning model. The physical constrained deep learning model takes acoustic features and water level as input and average cross-sectional flow velocity as output. During model training, calculation deviations determined according to a preset calculation method are included as physical constraints in the loss function to obtain a trained physical constrained deep learning model. The real-time acquired acoustic signals, after preprocessing and feature extraction, are input into the trained physical constrained deep learning model for inversion to obtain real-time flow. The real-time flow is then integrated into a hydrological forecasting model, and flow forecasts for multiple future time periods are generated through rolling updates. This method achieves non-contact, all-weather monitoring, has small inversion errors in complex environments, strong extrapolation capabilities under extreme conditions, and ensures physical consistency.

[0144] In some embodiments, the acoustic signal acquisition module is further configured to deploy the hydrophone array on the stable bank slopes on both sides of the river channel, with the array direction parallel to the river flow direction; dynamically adjust the water depth of the hydrophones according to the current water level so that they are located in the mainstream area with a small velocity gradient; and determine the spacing between adjacent hydrophones based on the spatial sampling theorem so that the array length covers the main area of ​​the river cross-section velocity change.

[0145] In some embodiments, the processing module is further configured to perform wavelet multi-scale decomposition on the acquired acoustic signal to obtain low-frequency approximation coefficients and multiple high-frequency detail coefficients; estimate the noise standard deviation of the high-frequency detail coefficients at each decomposition scale using the median method, and set an adaptive threshold based on the characteristic frequency band; perform threshold processing on the high-frequency detail coefficients at each scale using a soft threshold function based on the adaptive threshold, while keeping the low-frequency approximation coefficients unchanged; and perform inverse wavelet transform on the processed high-frequency detail coefficients at each scale and the original low-frequency approximation coefficients to reconstruct the denoised acoustic signal.

[0146] In some embodiments, the processing module is further configured to perform a fast Fourier transform on the preprocessed acoustic signal to extract frequency domain features; perform a short-time Fourier transform on the preprocessed acoustic signal to generate a time-frequency graph, and extract time-frequency domain depth features from the time-frequency graph through a pre-trained convolutional neural network; and calculate the time-domain statistics and entropy features of the preprocessed acoustic signal to obtain statistical features.

[0147] In some embodiments, a physically constrained deep learning model includes: an input layer for receiving fused features composed of acoustic features and water level values; a multi-layer LSTM temporal coding layer for capturing the temporal dynamic relationship of the fused features and outputting high-dimensional temporal features; a fully connected output layer for mapping the high-dimensional temporal features to cross-sectional average flow velocity; and a physical constraint layer for calculating the Manning formula theoretical flow velocity based on the input water level value and pre-stored cross-sectional morphology data, and calculating the deviation between the theoretical flow velocity and the cross-sectional average flow velocity output by the physically constrained deep learning model as a physical loss term.

[0148] In some embodiments, the real-time flow acquisition module is further configured to preprocess and extract features from the real-time acquired acoustic signal, and input it together with the synchronously acquired water level value into a trained physical constraint deep learning model to obtain the current cross-sectional average flow velocity; determine the cross-sectional area based on the synchronously acquired water level value and the pre-stored river cross-sectional morphology data; and obtain the real-time flow rate based on the current cross-sectional average flow velocity and the cross-sectional area.

[0149] In some embodiments, the flow forecast result generation module is further configured to input real-time flow into the hydrological forecast model to update the historical flow sequence; input the updated historical flow sequence, meteorological forcing data, and acoustic features into the trained hydrological forecast model; and perform iterative multi-step prediction using the hydrological forecast model; wherein, after predicting the flow at time t+1, the flow at time t+1 is used as one of the inputs for the next time step to predict the flow at time t+2, and the flow forecast sequence for multiple future time steps is recursively generated.

[0150] The device provided in this embodiment of the invention has the same implementation principle and technical effect as the aforementioned method embodiment. For the sake of brevity, any parts not mentioned in the device embodiment can be referred to the corresponding content in the aforementioned method embodiment.

[0151] Example 5 This invention also provides an electronic device for running the above-described river flow forecasting method based on acoustic inversion; see also Figure 8 The diagram shows the structure of an electronic device, which includes a memory 500 and a processor 501. The memory 500 stores one or more computer instructions, which are executed by the processor 501 to implement the aforementioned river flow forecasting method based on acoustic wave inversion.

[0152] Furthermore, Figure 8 The electronic device shown also includes a bus 502 and a communication interface 503. The processor 501, the communication interface 503, and the memory 500 are connected via the bus 502.

[0153] The memory 500 may include high-speed random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 503 (which can be wired or wireless), such as the Internet, wide area network, local area network, metropolitan area network, etc. The bus 502 can be an ISA bus, PCI bus, or EISA bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 8 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0154] Processor 501 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of processor 501 or by instructions in software form. Processor 501 can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this invention can be directly manifested as execution by a hardware decoding processor, or execution by a combination of hardware and software modules in the decoding processor. The software module can reside in a readily available storage medium in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory 500, and processor 501 reads information from memory 500 and, in conjunction with its hardware, completes the steps of the method described in the foregoing embodiments.

[0155] This invention also provides a computer-readable storage medium storing computer-executable instructions. When these computer-executable instructions are called and executed by a processor, they cause the processor to implement the aforementioned river flow forecasting method based on acoustic wave inversion. For specific implementation details, please refer to the method embodiments, which will not be repeated here.

[0156] The computer program product for a river flow forecasting method based on acoustic inversion provided in this embodiment of the invention includes a computer-readable storage medium storing non-volatile program code executable by a processor. The instructions included in the program code can be used to execute the methods described in the preceding method embodiments. For specific implementation details, please refer to the method embodiments, which will not be repeated here.

[0157] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0158] In the several embodiments provided by this invention, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the coupling or direct coupling or communication connection shown or discussed may be through some communication interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0159] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0160] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0161] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0162] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for predicting river flow based on acoustic wave inversion, characterized in that, The method includes: Acquire acoustic signals from a hydrophone array using non-contact acquisition; The acoustic signal is preprocessed and its features are extracted to obtain acoustic features; A physical constraint deep learning model is constructed. The physical constraint deep learning model takes the acoustic wave features and water level values ​​as inputs and the cross-sectional average flow velocity as output. During the model training process, the calculation deviation determined according to the preset calculation method is included as a physical constraint term in the loss function to obtain the trained physical constraint deep learning model. After the real-time collected acoustic signals are preprocessed and feature extracted, they are input into the trained physical constraint deep learning model for inversion to obtain the real-time traffic flow. The real-time flow rate is integrated into the hydrological forecasting model, and the flow rate forecast results for multiple future time periods are generated through rolling forecast updates.

2. The method according to claim 1, characterized in that, The deployment method of the hydrophone array includes: The hydrophone array is deployed on the stable bank slopes on both sides of the river channel, with the array direction parallel to the river flow direction. The depth of the hydrophone is dynamically adjusted according to the current water level to ensure that it is located in the mainstream area with a smaller velocity gradient. The spacing between adjacent hydrophones is determined based on the spatial sampling theorem, so that the array length covers the main region of the river cross-section flow velocity variation.

3. The method according to claim 1, characterized in that, Preprocessing the acoustic signal includes: The acquired acoustic signal is decomposed using wavelet multi-scale decomposition to obtain low-frequency approximation coefficients and multiple high-frequency detail coefficients. The noise standard deviation of high-frequency detail coefficients at each decomposition scale is estimated using the median method, and an adaptive threshold is set based on the characteristic frequency band. A soft thresholding function is used to threshold the high-frequency detail coefficients at each scale based on the adaptive threshold, while keeping the low-frequency approximation coefficients unchanged; The processed high-frequency detail coefficients at each scale are subjected to inverse wavelet transform with the original low-frequency approximation coefficients to reconstruct the denoised acoustic signal.

4. The method according to claim 2, characterized in that, Feature extraction of the acoustic signal includes: The preprocessed acoustic signal is subjected to a fast Fourier transform to extract frequency domain features; The preprocessed acoustic signal is subjected to short-time Fourier transform to generate a time-frequency map, and time-frequency domain depth features are extracted from the time-frequency map through a pre-trained convolutional neural network. The time-domain statistics and entropy characteristics of the preprocessed acoustic signal are calculated to obtain the statistical characteristics.

5. The method according to claim 1, characterized in that, The physical constraint deep learning model includes: The input layer is used to receive the fused features composed of acoustic features and water level values; A multi-layer LSTM temporal coding layer is used to capture the temporal dynamic relationship of the fused features and output high-dimensional temporal features; A fully connected output layer is used to map the high-dimensional temporal features into cross-sectional average flow velocity; The physical constraint layer is used to calculate the Manning formula theoretical flow velocity based on the input water level value and pre-stored cross-sectional morphology data, and to calculate the deviation between the theoretical flow velocity and the cross-sectional average flow velocity output by the physical constraint deep learning model as a physical loss term.

6. The method according to claim 1, characterized in that, The real-time acquired acoustic signal, after preprocessing and feature extraction, is input into the trained physical constraint deep learning model for inversion to obtain the real-time traffic flow, including: After the real-time collected acoustic signals are preprocessed and feature extracted, they are input together with the synchronously collected water level values ​​into the trained physical constraint deep learning model to obtain the current cross-sectional average flow velocity. Based on the synchronously collected water level values ​​and the pre-stored river cross-sectional morphology data, the cross-sectional area of ​​the water passage is determined. The real-time flow rate is obtained based on the current average flow velocity at the cross-section and the cross-sectional area of ​​the water passage.

7. The method according to claim 1, characterized in that, The step of integrating the real-time flow rate into the hydrological forecasting model and generating flow forecast results for multiple future time periods through rolling forecast updates includes: Real-time flow data is incorporated into hydrological forecasting models to update historical flow sequences; The updated historical flow sequence, meteorological forcing data, and acoustic features are input together into the trained hydrological forecasting model. The hydrological forecasting model is used for iterative multi-step prediction. After predicting the flow rate at time t+1, the flow rate at time t+1 is used as one of the inputs for the next time step to predict the flow rate at time t+2. This process recursively generates a flow rate forecast sequence for multiple future time steps.

8. A river flow forecasting device based on acoustic wave inversion, characterized in that, The device includes: The acoustic signal acquisition module is used to acquire acoustic signals non-contactly collected by the hydrophone array; The processing module is used to preprocess and extract features from the acoustic signal to obtain acoustic features; The model building module is used to build a physical constraint deep learning model. The physical constraint deep learning model takes the acoustic features and water level values ​​as inputs and the cross-sectional average flow velocity as output. During the model training process, the calculation deviation determined according to the preset calculation method is included as a physical constraint term in the loss function to obtain the trained physical constraint deep learning model. The real-time traffic acquisition module is used to input the real-time collected acoustic wave signal into the trained physical constraint deep learning model for inversion after preprocessing and feature extraction, so as to obtain the real-time traffic. The flow forecast result generation module is used to input the real-time flow into the hydrological forecast model and generate flow forecast results for multiple future times by rolling forecast updates.

9. An electronic device, characterized in that, The method includes a processor and a memory, the memory storing computer-executable instructions that can be executed by the processor, the processor executing the computer-executable instructions to implement the river flow forecasting method based on acoustic inversion as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when invoked and executed by a processor, cause the processor to implement the river flow forecasting method based on acoustic inversion as described in any one of claims 1 to 7.

Images