Cross section suspended load sediment concentration cooperative monitoring method and device

Abstract

The application discloses a cross-section suspended load sediment concentration cooperative monitoring method, relates to the technical field of hydrological sediment monitoring, and collects sample multi-source heterogeneous data and actual sediment data corresponding to a full cross-section of a to-be-monitored river channel, obtains sample input data and true value labels after pretreatment, and is used for training a multi-source data space-time fusion network of attention mechanism dynamic fusion multi-source features; real-time multi-source heterogeneous data of the full cross-section of the to-be-monitored river channel is collected, the trained multi-source data space-time fusion network is used, each vertical line predicted sediment concentration is acquired, and then real-time cross-section average sediment concentration is obtained, the problems that the prior art is non-real-time, labor intensity is large, spatial representativeness is insufficient and cross-section suspended load sediment concentration cannot be directly acquired are solved, and spatial representativeness, time continuity, measurement accuracy and automatic multi-dimensional requirements can be considered.

Description

Technical Field

[0001] This application relates to the field of hydrological sediment monitoring technology, specifically to a method and device for coordinated monitoring of suspended sediment content in cross sections. Background Technology

[0002] The suspended sediment concentration at cross-sections is a core hydrological parameter characterizing the suspended sediment content in rivers, estuaries, and other water bodies. Its accurate measurement and analysis play a crucial foundational role in fields such as river channel erosion and deposition evolution analysis, water project operation and scheduling, water resource management and utilization, sediment transport laws and mechanisms, and the impact of global change and human activities. Existing measurement techniques are mainly divided into three categories, all of which have significant limitations:

[0003] (1) Traditional instrumental methods (such as horizontal samplers and time-integrated samplers): Discrete water samples are collected manually on a cross-section using finite vertical lines and finite measurement points on the vertical lines. Laboratory analysis yields accurate but isolated data. This method is labor-intensive, dangerous, and inefficient. It cannot capture continuous changes and has an inherent error of insufficient spatial representativeness when extrapolating the cross-sectional average from sparse data.

[0004] (2) Single-point continuous monitoring method (such as fixed optical / acoustic sand measuring instrument): continuous and automatic monitoring is carried out at a fixed point on the cross section. Although the time continuity is achieved, its measurement value only represents the mud and sand condition of a small local area where the sensor is located. It cannot reflect the spatial distribution (lateral and vertical) of sand content on the cross section. It is extremely risky to use it directly to represent the average value of the cross section, especially when the cross section is uneven, the error is huge.

[0005] (3) Spectral surface inversion method (e.g., based on satellite, UAV or ground multispectral instrument): The suspended sediment content of surface water is inverted by acquiring the spectral characteristics of water bodies. It has the advantages of wide spatial coverage, non-contact and speed. However, its inversion accuracy is greatly affected by atmospheric conditions, water surface glare and the complexity of water composition (e.g. chlorophyll, yellow substances interference), and usually only reflects the sediment information of surface water, making it difficult to directly obtain the cross-sectional spatial vertical average or even the cross-sectional average sediment content.

[0006] It can be determined that the existing methods for measuring the average sediment content of cross sections have the following significant drawbacks:

[0007] (1) Non-real-time: The traditional manual sand measurement method involves a long process of sampling, testing and analysis, which cannot obtain the real-time change process of sand content and makes it difficult to capture key hydrological events such as the rapid movement of sand peaks.

[0008] (2) High labor intensity and high risk: Especially under conditions of high flood, high sediment content or severe weather, manual sand measurement is labor-intensive and poses safety hazards.

[0009] (3) Insufficient spatial representativeness: Traditional manual sand measurement methods, single-point photoelectric / spectral / acoustic online sand measurement instruments, and multispectral surface sand measurement instruments all have discrete point sampling, which makes it difficult to accurately reflect the spatial distribution of suspended sediment content in the entire cross section (especially when there is stratification or transverse gradient). The average value of the cross section calculated by finite vertical lines has a large error.

[0010] (4) The real-time suspended sediment concentration of the cross section cannot be directly obtained: It is difficult to integrate into the modern hydrological automatic monitoring and reporting system, and cannot meet the requirements of smart water conservancy for data real-time performance, continuity and cross section spatial distribution.

[0011] Therefore, there is currently a lack of a real-time monitoring method for suspended sediment concentration at cross-sections that can balance spatial representativeness, temporal continuity, measurement accuracy, and automation. There is an urgent need for a novel integrated solution that can leverage the advantages of multi-source data and overcome the limitations of single-technology approaches. Summary of the Invention

[0012] The purpose of this application is to provide a method and device for collaborative monitoring of suspended sediment content in cross sections, which solves the problems of non-real-time nature, high labor intensity, high risk, insufficient spatial representativeness and / or inability to directly obtain the suspended sediment content in cross sections in the existing technology.

[0013] This application is achieved through the following technical solution:

[0014] The first aspect of this application provides a method for coordinated monitoring of suspended sediment concentration in cross sections, including:

[0015] Collect multi-source heterogeneous data of the entire cross-section of the river to be monitored, as well as the actual sediment concentration data corresponding to the multi-source heterogeneous data of the samples.

[0016] Preprocess the sample multi-source heterogeneous data and the actual sand-containing data corresponding to the sample multi-source heterogeneous data to obtain the sample input data and the truth label corresponding to the sample input data.

[0017] Based on the sample input data and the ground truth labels corresponding to the sample input data, a multi-source data spatiotemporal fusion network that dynamically integrates multi-source features through an attention mechanism is trained to learn the mapping relationship from multi-source data to the sand content of the cross-section vertical line, and the trained multi-source data spatiotemporal fusion network is obtained.

[0018] Collect real-time multi-source heterogeneous data corresponding to the entire cross-section of the river to be monitored, and preprocess the real-time multi-source heterogeneous data to obtain real-time input data;

[0019] The trained multi-source data spatiotemporal fusion network is scheduled to analyze the real-time input data and obtain the predicted sediment concentration of each vertical line on the entire cross section of the river to be monitored.

[0020] Based on the predicted sediment concentration of each vertical line on the entire cross section of the river to be monitored, the real-time average sediment concentration of the corresponding cross section of the entire cross section of the river to be monitored is obtained.

[0021] In one possible implementation, the method further includes:

[0022] Acquire manually tested data from human-computer interaction input; the manually tested data includes multi-source heterogeneous data and manually tested sand content;

[0023] Based on the artificial test data, the trained multi-source data spatiotemporal fusion network is adjusted, and the adjusted multi-source data spatiotemporal fusion network is used to analyze the sediment content in the subsequent cross-section suspended sediment concentration collaborative monitoring process.

[0024] In one possible implementation, the process involves collecting multi-source heterogeneous data of the entire cross-section of the river to be monitored, as well as the actual sediment load data corresponding to the multi-source heterogeneous data, including:

[0025] For the entire cross-section of the river to be monitored, spatial continuous spectral image data of the water body are collected at a preset sampling frequency.

[0026] For the entire cross-section of the river to be monitored, real-time continuous sediment concentration time series data are collected along each vertical line of the entire cross-section of the river to be monitored at a preset sampling frequency.

[0027] The actual sediment concentration data of each vertical line in the full cross section of the river to be monitored were obtained by manual analysis and laboratory precision analysis, and the actual sediment concentration data corresponding to the multi-source heterogeneous data of the sample were obtained.

[0028] In one possible implementation, the sample multi-source heterogeneous data and the corresponding actual sand-containing data are preprocessed to obtain sample input data and the corresponding truth labels, including:

[0029] The sample multi-source heterogeneous data and the actual sand-containing data corresponding to the sample multi-source heterogeneous data are synchronized in time and / or registered in space to obtain the registered sample multi-source heterogeneous data and the actual sand-containing data corresponding to the sample multi-source heterogeneous data.

[0030] Based on the spectral images in the registered multi-source heterogeneous data of the samples, the spectral reflectance curves of the water surface pixels and spatial context features corresponding to each preset calculation vertical line are obtained to obtain image features;

[0031] Based on the real-time continuous sediment concentration time series data in the multi-source heterogeneous data of the registered samples, the real-time sediment concentration time series corresponding to the measuring points on each preset calculation vertical line is determined to obtain the sensor characteristics;

[0032] Obtain water level, flow rate, flow velocity, flow direction, water temperature and / or cross-sectional water depth distribution to obtain auxiliary data features;

[0033] The image features, sensor features, and auxiliary data features are used together as sample input data;

[0034] Based on the actual sediment-containing data corresponding to the multi-source heterogeneous data of the sample, the actual sediment-containing data of each measuring point on the preset calculation vertical line are subjected to velocity-weighted integral or arithmetic average to obtain the true value label corresponding to the sample input data.

[0035] In one possible implementation, the attention mechanism dynamically fuses multi-source data spatiotemporal fusion network, comprising:

[0036] The system consists of a spectral feature extraction subnetwork composed of a one-dimensional convolutional neural network, a temporal feature extraction subnetwork composed of a long short-term memory network, an auxiliary feature encoding subnetwork composed of a fully connected network, a multi-source attention fusion layer, and an output layer. These structures are connected sequentially to establish a physical relationship between water and sediment physical quantities and neural network parameters.

[0037] In one possible implementation, based on the sample input data and the ground truth labels corresponding to the sample input data, a multi-source data spatiotemporal fusion network that dynamically integrates multi-source features using an attention mechanism is trained to learn the mapping relationship from multi-source data to the sediment content of the cross-section vertical line, thereby obtaining the trained multi-source data spatiotemporal fusion network, including:

[0038] The loss function is constructed using mean squared error and absolute error;

[0039] Using the sample input data as input, the ground truth label corresponding to the sample input data as the expected output, and minimizing the loss function as the objective, the backpropagation algorithm is used to train the multi-source data spatiotemporal fusion network that dynamically integrates multi-source features through the attention mechanism, thereby obtaining the trained multi-source data spatiotemporal fusion network.

[0040] In one possible implementation, the loss function is constructed using mean squared error and absolute error as follows:

[0041]

[0042] in, Represents the loss function. Indicates the truth label, This represents the model's predicted data. This represents the threshold parameter.

[0043] In one possible implementation, the trained multi-source data spatiotemporal fusion network is scheduled to analyze the real-time input data to obtain the predicted sediment concentration of each preset calculation vertical line on the entire cross-section of the river channel to be monitored, including:

[0044] The real-time input data is used as the input to the trained multi-source data spatiotemporal fusion network, and the predicted value output by the trained multi-source data spatiotemporal fusion network is obtained to obtain the predicted sediment concentration of each preset calculation vertical line on the entire cross section of the river to be monitored.

[0045] In one possible implementation, based on the predicted sediment concentration of each preset calculation vertical line on the entire cross-section of the river channel to be monitored, the real-time average sediment concentration of the corresponding cross-section is obtained as follows:

[0046]

[0047] in, This represents the real-time average sediment concentration of the entire cross-section of the river channel to be monitored. This represents the predicted sediment concentration of the i-th pre-defined calculation vertical line across the entire cross-section of the river to be monitored, where N represents the total number of vertical lines. This represents the local cross-sectional area represented by the i-th preset calculated vertical line.

[0048] The second aspect of this application provides a cross-sectional suspended sediment concentration co-monitoring device, comprising:

[0049] The first data acquisition module is used to acquire sample multi-source heterogeneous data corresponding to the full cross-section of the river to be monitored, as well as the actual sediment content data corresponding to the sample multi-source heterogeneous data.

[0050] The data preprocessing module is used to preprocess the sample multi-source heterogeneous data and the actual sand-containing data corresponding to the sample multi-source heterogeneous data to obtain the sample input data and the truth labels corresponding to the sample input data.

[0051] The model training module is used to train the multi-source data spatiotemporal fusion network that dynamically integrates multi-source features based on the sample input data and the ground truth labels corresponding to the sample input data, so as to learn the mapping relationship from multi-source data to the sand content of the cross-section vertical line and obtain the trained multi-source data spatiotemporal fusion network.

[0052] The second data acquisition module is used to acquire real-time multi-source heterogeneous data corresponding to the entire cross-section of the river to be monitored, and to preprocess the real-time multi-source heterogeneous data to obtain real-time input data.

[0053] The data analysis module is used to schedule the trained multi-source data spatiotemporal fusion network to analyze the real-time input data and obtain the predicted sediment concentration of each preset calculation vertical line on the full cross section of the river to be monitored.

[0054] The result output module is used to obtain the real-time average sediment concentration of the entire cross-section of the river under monitoring based on the predicted sediment concentration of each preset calculation vertical line on the entire cross-section of the river under monitoring.

[0055] Compared with the prior art, this application has the following advantages and beneficial effects:

[0056] This application provides a method for collaborative monitoring of suspended sediment concentration in a cross-section. It involves collecting multi-source heterogeneous data and actual sediment concentration data corresponding to the entire cross-section of the river to be monitored. After preprocessing, sample input data and ground truth labels are obtained, which are used to train an attention mechanism to dynamically fuse multi-source data spatiotemporal fusion network. Then, real-time multi-source heterogeneous data of the entire cross-section of the river to be monitored is collected. Using the trained multi-source data spatiotemporal fusion network, the predicted sediment concentration of each vertical line is obtained, thus yielding the real-time average sediment concentration of the cross-section. This method solves the problems of non-real-time performance, high labor intensity, insufficient spatial representativeness, and inability to directly obtain the suspended sediment concentration of the cross-section in existing technologies. It can simultaneously meet the multi-dimensional requirements of spatial representativeness, temporal continuity, measurement accuracy, and automation. Attached Figure Description

[0057] To more clearly illustrate the technical solutions of the exemplary embodiments of this application, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings:

[0058] Figure 1 A flowchart of a method for coordinated monitoring of suspended sediment content in cross sections provided in this application embodiment;

[0059] Figure 2 A schematic diagram of the structure of a cross-sectional suspended sediment content co-monitoring device provided in this application embodiment;

[0060] Figure 3 A cross-sectional view of the data acquisition device provided in the embodiments of this application;

[0061] Figure 4 A top view of the data acquisition device layout on a cross section, as provided in the embodiments of this application;

[0062] The attached diagram shows the markings and corresponding component names:

[0063] 201-First data acquisition module, 202-Data preprocessing module, 203-Model training module, 204-Second data acquisition module, 205-Data analysis module, 206-Result output module. Detailed Implementation

[0064] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the embodiments and accompanying drawings. The illustrative embodiments and descriptions of this application are only for explaining this application and are not intended to limit this application.

[0065] like Figure 1 As shown in the embodiment of this application, a method for coordinated monitoring of suspended sediment content in a cross section is provided, including:

[0066] S101. Collect multi-source heterogeneous data of samples corresponding to the full cross-section of the river to be monitored, as well as the actual sediment content data corresponding to the multi-source heterogeneous data of the samples.

[0067] S102. Preprocess the sample multi-source heterogeneous data and the actual sand-containing data corresponding to the sample multi-source heterogeneous data to obtain the sample input data and the truth label corresponding to the sample input data.

[0068] S103. Based on the sample input data and the ground truth labels corresponding to the sample input data, train the multi-source data spatiotemporal fusion network that dynamically integrates multi-source features through the attention mechanism to learn the mapping relationship from multi-source data to the sand content of the cross-section vertical line, and obtain the trained multi-source data spatiotemporal fusion network.

[0069] S104. Collect real-time multi-source heterogeneous data corresponding to the entire cross-section of the river to be monitored, and preprocess the real-time multi-source heterogeneous data to obtain real-time input data.

[0070] S105. The trained multi-source data spatiotemporal fusion network is scheduled to analyze the real-time input data to obtain the predicted sediment concentration of each vertical line on the full cross section of the river channel to be monitored.

[0071] S106. Based on the predicted sediment concentration of each vertical line on the full cross section of the river to be monitored, obtain the real-time average sediment concentration of the full cross section of the river to be monitored.

[0072] The core objective of the collaborative monitoring method for suspended sediment concentration in river sections provided in this application embodiment is to achieve high-frequency, high-precision, and automated monitoring of suspended sediment concentration distribution in river sections by using three collaborative monitoring methods: non-contact multispectral scanning, underwater fixed-point continuous monitoring, and manual comparative verification.

[0073] In one possible implementation, the method further includes:

[0074] Acquire manually tested data from human-computer interaction input; the manually tested data includes multi-source heterogeneous data and manually tested sand content;

[0075] Based on the artificial test data, the trained multi-source data spatiotemporal fusion network is adjusted, and the adjusted multi-source data spatiotemporal fusion network is used to analyze the sediment content in the subsequent cross-section suspended sediment concentration collaborative monitoring process.

[0076] For example, when a new batch of manually tested data... Upon arrival, the system does not undergo a complete retraining; instead, it employs an incremental learning strategy. Using the current model weights as initial values, it iterates through a small batch of new data, fine-tuning the model parameters to quickly adapt to new river conditions while retaining its memory of historical patterns. Indicates the time of manual testing. Indicates in At time i, the "true value" of the i-th vertical line is the average sand content of the vertical line.

[0077] In one possible implementation, the process involves collecting multi-source heterogeneous data of the entire cross-section of the river to be monitored, as well as the actual sediment load data corresponding to the multi-source heterogeneous data, including:

[0078] For the entire cross-section of the river to be monitored, spatial continuous spectral image data of the water body are collected at a preset sampling frequency.

[0079] For example, one or more multispectral instruments can be fixedly mounted on supports, bridges, or drone platforms along the riverbank, with their field of view covering the entire cross-section of the river to be measured. This sub-unit automatically scans the cross-section at a set frequency (e.g., per minute) to acquire spatial continuous spectral image data of the water body, where each pixel or cell contains spectral information reflecting the optical characteristics of the surface water at that location.

[0080] For the entire cross-section of the river to be monitored, real-time continuous sediment concentration time series data are collected along each vertical line of the entire cross-section of the river to be monitored at a preset sampling frequency.

[0081] Several (no fewer than two) representative key vertical lines (such as the main flow zone and the shore transition zone) can be selected on the cross section, and underwater spectrophotometers based on photoelectric sensors can be fixedly installed. These sensors sample at a set frequency, providing real-time continuous time series data of sediment concentration at their location (specific vertical line, specific water depth), serving as "anchored" continuous observation points on the cross section.

[0082] The actual sediment concentration data of each vertical line in the full cross section of the river to be monitored were obtained by manual analysis and laboratory precision analysis, and the actual sediment concentration data corresponding to the multi-source heterogeneous data of the sample were obtained.

[0083] Actual sediment concentration data serves as the basis for system calibration and verification. Water samples can be collected periodically (e.g., daily, weekly, monthly, or intensively during floods) using traditional and recognized reliable instrumental methods (e.g., horizontal samplers, time-integrated samplers) at different depths along multiple vertical lines, including the spectral scanning area and fixed sensor positions. After precise laboratory analysis, a series of high-precision, high-spatial-resolution (vertical) baseline sediment concentration data can be obtained across the cross-section.

[0084] In one possible implementation, the sample multi-source heterogeneous data and the corresponding actual sand-containing data are preprocessed to obtain sample input data and the corresponding truth labels, including:

[0085] The sample multi-source heterogeneous data and the actual sand-containing data corresponding to the sample multi-source heterogeneous data are synchronized in time and / or registered in space to obtain the registered sample multi-source heterogeneous data and the actual sand-containing data corresponding to the sample multi-source heterogeneous data.

[0086] The heterogeneous data collected by the above three sub-units are precisely synchronized in time (based on GPS / BeiDou time synchronization) and spatially registered (mapping the pixel coordinates of the spectral image with the cross-sectional geographic coordinates and the fixed sensor position coordinates), and transmitted in real time to the central server or edge computing device via wired or wireless network (such as fiber optic, 4G / 5G, BeiDou), and the central server or edge computing device executes the data analysis method described in the embodiments of this application.

[0087] Based on the spectral images in the registered multi-source heterogeneous data of the samples, the spectral reflectance curves of the water surface pixels and spatial context features corresponding to each preset calculation vertical line are obtained to obtain image features;

[0088] Based on the real-time continuous sediment concentration time series data in the multi-source heterogeneous data of the registered samples, the real-time sediment concentration time series corresponding to the measuring points on each preset calculation vertical line is determined to obtain the sensor characteristics;

[0089] For example, taking the current time t as a reference, the sediment-containing time series uses historical window data with t as the endpoint and a length of L, i.e., {Sk(t−L+1),Sk(t−L+2),…,Sk(t)}. Here, L is a hyperparameter (typically 5 to 20 sampling periods), the purpose of which is to utilize the time inertia of sediment concentration (settlement and suspension processes require time) to capture changing trends and improve the model's ability to predict dynamic processes.

[0090] Regarding the processing logic for vertical lines and measuring points: Vertical lines can be used as the unit. Typically, only one fixed-point sensor is installed on each key vertical line (e.g., at a depth of 0.6 times the water surface). Therefore, each vertical line directly corresponds to a real-time sediment concentration time series Sk(t), eliminating the need to combine multiple measuring points. Setting multiple measuring points on each vertical line is for a baseline manual testing subunit (manual sampling), rather than fixed continuous sensors.

[0091] Extension from sparse points to the entire cross section: Since the number of fixed sensors K is relatively small (2-3 vertical lines), while the number of computational vertical lines N is relatively large (dozens), the LSTM temporal features hkt of K discrete points can be interpolated onto all N vertical lines through a spatial interpolation layer to generate a spatially continuous temporal feature field fitemp, thereby solving the problem of insufficient spatial coverage.

[0092] Installing multiple depth sensors along each vertical line would lead to problems such as input dimension explosion, cost and maintenance difficulties, and mismatch between ground truth labels. Manual sampling typically only yields the average sediment concentration along the vertical line, while sediment concentration varies significantly at different depths. Assigning the same average ground truth value to all measuring points would cause the model to learn incorrect physical relationships. Therefore, this application chooses the vertical line as the basic unit, which is reasonable and fulfills the invention's purpose.

[0093] Obtain water level, flow rate, flow velocity, flow direction, water temperature and / or cross-sectional water depth distribution to obtain auxiliary data features;

[0094] The image features, sensor features, and auxiliary data features are used together as sample input data;

[0095] The sample input data is based on a specific time t (usually the time tm of the manual test) and includes the following three types of data:

[0096] The spectral image features at time t are obtained by extracting the spectral reflectance curve Ri(λ) of the water surface pixel corresponding to each calculated vertical line i from the spectral image covering the entire cross section, as spatial coverage information. This feature represents the optical properties of the surface water at each location on the cross section at time t.

[0097] The auxiliary hydrological data at time t, including water level H(t), cross-sectional average flow velocity V(t), and water temperature T(t), are used as global background features.

[0098] The real-time sediment concentration time series is obtained for L sampling times prior to time t. For K fixed underwater sensors deployed along the cross-section (each sensor located at a fixed depth point on a key vertical line, i.e., one measuring point per vertical line), sediment concentration data are collected within a historical window of length L, ending at time t, forming a time series: {Sk(t−L+1),Sk(t−L+2),…,Sk(t)}, k=1,2,…,K; where Sk(τ) is the sediment concentration value of the k-th sensor at time τ (obtained through a preliminary transformation model), τ=t−L+1,t−L+2,...,t. This series reflects the sediment concentration change trend of each fixed point within the recent time window.

[0099] Output label (for training only): The high-precision average sediment concentration CiGT(t) of each vertical line (corresponding to the calculated vertical line of the spectral image) obtained by manual testing at time t.

[0100] Time window length L: Typical value is 10 to 20 (if the sampling frequency is once per minute, it corresponds to the past 10 to 20 minutes).

[0101] Number of fixed sensors K: no less than 2 (such as three key vertical lines on the left, middle and right).

[0102] The number of vertical lines N is calculated to be much greater than K, typically 20 to 50, and is used for the final output of suspended sediment concentration distribution in the cross section.

[0103] Based on the above definition, the fusion model of the present invention accepts the joint input of "current moment wide-area spatial information (spectrum) + recent time dynamics (historical time series sensor data) + current global environmental conditions (auxiliary hydrological data)" during training, and has the ability to model the spatiotemporal continuous sediment concentration field.

[0104] Based on the actual sediment-containing data corresponding to the multi-source heterogeneous data of the sample, the actual sediment-containing data of each measuring point on the preset calculation vertical line are subjected to velocity-weighted integral or arithmetic average to obtain the true value label corresponding to the sample input data.

[0105] For example, the acquired spectral image cube can be radiometrically calibrated, atmospherically corrected (if necessary), and geometrically corrected. Then, based on the cross-sectional spatial coordinates, the spectral reflectance curve of the water surface pixel corresponding to each pre-defined calculated vertical line is extracted from the image. , where i is the vertical index and λ is the band. Simultaneously, spatial context features of these pixels are extracted, such as the coefficient of variation of the spectra of adjacent pixels, to characterize the local homogeneity of the water surface.

[0106] The raw signal (such as backscattering intensity) of the fixed-point optical / acoustic sediment analyzer can be filtered and denoised. A preliminary conversion model, calibrated manually, can then be used to convert the sensor's raw output signal into an engineering estimate of sediment concentration. The calibration method involves simultaneously collecting water samples at no fewer than eight different sediment concentration gradients at the sensor installation location, and obtaining the true sediment concentration S through laboratory analysis. lab The power function model S=a⋅I is adopted. b (I represents the sensor signal, and a and b are undetermined coefficients) For data pairs (I, S) lab Least squares fitting was performed to determine the coefficients a and b. This transformation model provides a continuous, dimensional initial sediment concentration time series for subsequent multi-source data fusion. And its system bias will be dynamically corrected in the iterative update of the fusion model.

[0107] It can collect auxiliary hydrological data closely related to sediment transport, including water level. ,flow Cross-sectional average velocity Water temperature and the cross-sectional water depth distribution derived from historical data. These data serve as physical constraints on the model, enhancing its generalization ability.

[0108] Water samples can be collected manually at regular intervals. The sediment concentration at each vertical line and measuring point can be obtained through laboratory analysis. The high-precision average sediment concentration of each manually measured vertical line can then be calculated by weighted integration of flow velocity or by arithmetic mean (depending on the definition of vertical mean). This serves as the ground truth for model training. This represents the time of the m-th manual test.

[0109] In one possible implementation, the attention mechanism dynamically fuses multi-source data spatiotemporal fusion network, comprising:

[0110] The system consists of a spectral feature extraction subnetwork composed of a one-dimensional convolutional neural network, a temporal feature extraction subnetwork composed of a long short-term memory network, an auxiliary feature encoding subnetwork composed of a fully connected network, a multi-source attention fusion layer, and an output layer. These structures are connected sequentially to establish a physical relationship between water and sediment physical quantities and neural network parameters.

[0111] For example, this application embodiment constructs a deep learning model called Multi-Source DataFusion Network (MSDF-Net). The core idea of ​​this model is to use artificial benchmark data as a "teacher" to learn how to dynamically correct and fuse real-time acquired wide-area spectral information and continuous point source information.

[0112] The MSDF-Net network architecture consists of three parallel sub-networks and a core fusion layer:

[0113] Spectral Feature Extraction Subnetwork (Spectral-Net): Employs a one-dimensional convolutional neural network (1D-CNN) to process the spectral reflectance curves of each computational vertical line. Its mathematical expression is:

[0114]

[0115] in, It is a feature vector representing deep features related to sediment content extracted from the water surface spectrum of this vertical line.

[0116] Temporal-Net: Employs a Long Short-Term Memory (LSTM) network to process historical sediment concentration time series {S1(t), S2(t), ..., S...} from K fixed-point sensors. K The time window is L, and its mathematical expression is:

[0117]

[0118] in, This represents the hidden state of the LSTM network at the current time t, encoding the dynamic change pattern of sediment concentration at the k-th sensor location. Then, a spatial interpolation layer is used to extract the temporal characteristics of these discrete locations. Combined with cross-sectional water depth distribution As covariates, they are diffused to all computational perpendiculars i using the differential form of Gaussian process regression or inverse distance weighted interpolation (IDW) to obtain a spatially continuous temporal characteristic field. .

[0119]

[0120] in, Represents the spatial interpolation function. Represents the set of hidden states of all K sensors. This represents the water depth along the i-th vertical line.

[0121] Aux-Net (Aux Feature Encoding Subnetwork): Encodes the global auxiliary data [H(t),V(t),T(t)] into a global feature vector through a fully connected network (FCN). It is worth noting the cross-sectional water depth distribution It has been fully utilized as a covariate in the spatial interpolation operation of the temporal feature extraction subnetwork to generate a spatially temporal feature field with physical consistency. Therefore, the input is not repeated in the auxiliary feature encoding sub-network to avoid information redundancy and model structure complexity. This design ensures that the functions of each sub-network are clear and the information flow is reasonable, reflecting the node division of labor and collaborative integration ideas of this invention.

[0122] Multi-source attention fusion layer: This is one of the core innovations of this application. For each computational vertical line i, the feature vector obtained from the spectral and temporal fields is... and and the global features after replication The data is then stitched together. Then, an attention mechanism is introduced to allow the model to automatically learn which data source to trust more under different hydrological conditions.

[0123]

[0124] in, and These are trainable weights and biases. α is the attention weight, and The fused composite feature vector is:

[0125]

[0126] The output layer of MSDF-Net is a fully connected layer that fuses features. Mapped to the predicted value of the average sediment concentration along the i-th calculated vertical line. for:

[0127]

[0128] in, Indicates a fully connected layer;

[0129] In one possible implementation, based on the sample input data and the ground truth labels corresponding to the sample input data, a multi-source data spatiotemporal fusion network that dynamically integrates multi-source features using an attention mechanism is trained to learn the mapping relationship from multi-source data to the sediment content of the cross-section vertical line, thereby obtaining the trained multi-source data spatiotemporal fusion network, including:

[0130] The loss function is constructed using mean squared error and absolute error;

[0131] Using the sample input data as input, the ground truth label corresponding to the sample input data as the expected output, and minimizing the loss function as the objective, the backpropagation algorithm is used to train the multi-source data spatiotemporal fusion network that dynamically integrates multi-source features through the attention mechanism, thereby obtaining the trained multi-source data spatiotemporal fusion network.

[0132] The training objective of this model is to ensure that the predicted values ​​on all manually measured vertical lines are accurate. with truth value The error between them is minimized or within the acceptable range. The loss function used is Huber Loss, which combines the advantages of mean squared error and absolute error, and is more robust to outliers; the loss function is constructed using mean squared error and absolute error as follows:

[0133]

[0134] in, Represents the loss function. Indicates the truth label, This represents the model's predicted data. This represents the threshold parameter.

[0135] Based on the aforementioned loss function, the backpropagation algorithm is used to optimize all weights of the network, enabling the model to learn how to accurately invert the vertical sand content from multi-source data.

[0136] In one possible implementation, the trained multi-source data spatiotemporal fusion network is scheduled to analyze the real-time input data to obtain the predicted sediment concentration of each preset calculation vertical line on the entire cross-section of the river channel to be monitored, including:

[0137] The real-time input data is used as the input to the trained multi-source data spatiotemporal fusion network, and the predicted value output by the trained multi-source data spatiotemporal fusion network is obtained to obtain the predicted sediment concentration of each preset calculation vertical line on the entire cross section of the river to be monitored.

[0138] In one possible implementation, based on the predicted sediment concentration of each preset calculation vertical line on the entire cross-section of the river channel to be monitored, the real-time average sediment concentration of the corresponding cross-section is obtained as follows:

[0139]

[0140] in, This represents the real-time average sediment concentration of the entire cross-section of the river channel to be monitored. This represents the predicted sediment concentration of the i-th pre-defined calculation vertical line across the entire cross-section of the river to be monitored, where N represents the total number of vertical lines. This represents the local cross-sectional area represented by the i-th pre-defined calculated perpendicular line (usually divided using the Thiessen polygon method or the midpoint method of adjacent perpendicular lines).

[0141] For example, after training is complete, the system enters real-time monitoring mode. At each calculation time t, the real-time collected data stream propagates forward through MSDF-Net, instantly yielding the sand content of all N calculated vertical lines. , ,..., Finally, the final result was calculated using the cross-sectional segmentation and area weighting method.

[0142] Compared with the prior art, the present invention has the following outstanding advantages:

[0143] Real-time monitoring of suspended sediment content in cross sections with high spatiotemporal resolution was achieved: high spatial resolution was obtained through spectral scanning, high temporal resolution was obtained through underwater sensors, and correction was performed through manual measurement data. After fusion, the limitations of each data point were overcome, and for the first time, real-time reconstruction and monitoring of the "spatiotemporal continuous field" of suspended sediment content in cross sections was realized.

[0144] Measurement accuracy and reliability are significantly improved: High-precision but discrete manual measurement data is innovatively used as a benchmark and "anchor point" and deeply integrated into the real-time monitoring system, which effectively corrects the bias of the data inversion system and sensor drift, ensuring the absolute accuracy and physical reliability of the output results.

[0145] The system boasts a high level of automation and intelligence: With the support of regular manual calibration, the system can operate fully automatically most of the time, outputting the average sand content of the cross section in real time, greatly freeing up manpower, and enabling timely capture of the sediment transport process of the cross section.

[0146] It balances comprehensiveness and economy: by using relatively low-cost spectral scanning and a small number of fixed-point sensors, combined with periodic (non-continuous) manual measurements, it achieves monitoring results that are close to those that can only be achieved by "densely deploying sensor arrays", and the system construction and maintenance costs are more advantageous.

[0147] It possesses strong adaptive and evolutionary capabilities: the fusion algorithm has machine learning capabilities, can continuously optimize itself as new data accumulates, and can adapt to long-term processes such as river channel changes and changes in hydrological conditions, ensuring the long-term effectiveness of the technology.

[0148] like Figure 2 As shown in the figure, this application provides a cross-sectional suspended sediment concentration co-monitoring device, comprising:

[0149] The first data acquisition module 201 is used to acquire sample multi-source heterogeneous data corresponding to the full cross section of the river to be monitored, as well as the actual sediment content data corresponding to the sample multi-source heterogeneous data.

[0150] The data preprocessing module 202 is used to preprocess the sample multi-source heterogeneous data and the actual sand-containing data corresponding to the sample multi-source heterogeneous data to obtain the sample input data and the truth label corresponding to the sample input data.

[0151] The model training module 203 is used to train the multi-source data spatiotemporal fusion network that dynamically integrates multi-source features based on the sample input data and the ground truth labels corresponding to the sample input data, so as to learn the mapping relationship from multi-source data to the sand content of the cross-section vertical line and obtain the trained multi-source data spatiotemporal fusion network.

[0152] The second data acquisition module 204 is used to acquire real-time multi-source heterogeneous data corresponding to the entire cross-section of the river to be monitored, and to preprocess the real-time multi-source heterogeneous data to obtain real-time input data.

[0153] Data analysis module 205 is used to schedule the trained multi-source data spatiotemporal fusion network to analyze the real-time input data and obtain the predicted sediment concentration of each preset calculation vertical line on the full cross section of the river to be monitored.

[0154] The result output module 206 is used to obtain the real-time average sediment concentration of the entire cross-section of the river to be monitored based on the predicted sediment concentration of each preset calculation vertical line on the entire cross-section of the river to be monitored.

[0155] The cross-sectional suspended sediment concentration monitoring device provided in this application can execute the technical solution described in any of the above embodiments, and its principle and beneficial effects are similar.

[0156] Based on the above technical solution, a preferred embodiment of this application will be described in detail below with reference to the accompanying drawings.

[0157] like Figure 3 and Figure 4 As shown in the diagram, this device is deployed at the basic cross-section of a hydrological station. At the three key vertical lines on the left, middle, and right sides of the cross-section, an underwater online sediment analyzer (underwater fixed-point continuous sensing subunit, with sliding rail mounting on the left and right banks for adjustable vertical position, and a float + anchor mounting in the middle of the cross-section to accommodate the need for a fixed mounting base in the middle of the cross-section; in adverse conditions, this can be omitted in the middle of the river channel) measures the sediment concentration at different heights above the bed surface. On a pillar on the right bank of the cross-section, an imaging spectrometer (non-contact spectral scanning subunit) is installed, with a top-down view covering the entire water surface of the cross-section. The station building is equipped with sampling lead weights and sample processing equipment (part of the baseline manual testing subunit).

[0158] The data integration and transmission unit transmits the images from the spectrometer every minute, the data from the three sand measuring instruments every minute, and the data from the automatic water level gauge and flow meter (auxiliary features) to the data processing and fusion unit (server) in the station building in real time via fiber optic network or GPRS.

[0159] The implementation process of the fusion algorithm is as follows (corresponding to...) Figure 3 ):

[0160] 1. Data Acquisition and Preprocessing: The system acquires spectral images, readings from 3 or 2 online sediment analyzers, water level, flow rate, and other data per minute. The data processing unit automatically performs the preprocessing described in step one, extracting spectral features from 5 or more preset calculation vertical lines. Processing sensor data to obtain And integrate auxiliary data H(t), V(t), T(t).

[0161] 2. Initial Training Phase: In the initial period after system installation, a rigorous manual test will be conducted once daily (the arrangement of vertical lines and measuring points will be determined based on river width and water depth). During each manual test, all the aforementioned multi-source data will be recorded simultaneously. A total of 150 sets of "multi-source data-true value" sample pairs covering high, medium, and low sediment concentration levels will be accumulated.

[0162] 3. Model Training: Divide the 150 datasets into training and validation sets in an 8:2 ratio. Construct and initialize the MSDF-Net model (Spectral-Net uses a 3-layer 1D-CNN, Temporal-Net uses a 2-layer LSTM, and the attention layer is a 3D softmax). Use Huber Loss as the loss function and the Adam optimizer for training. Monitor model performance on the validation set to prevent overfitting. After training, save the model.

[0163] 4. Real-time Application Phase: After training, the system enters real-time monitoring mode. Every minute, the system automatically acquires the latest data and inputs it into the MSDF-Net model. The model propagates forward and instantly outputs the average sand content of 5 or more calculated vertical lines at the current moment. For example, during floods, the spectral signal may become saturated due to high sediment content, and the attention mechanism will automatically reduce the weight of spectral features. Increase the temporal characteristics of underwater sensors This trust ensures the accuracy of predictions.

[0164] 5. Calculation of cross-sectional average area: Based on the area of ​​6 pre-divided sections (dynamically adjusted according to the number of vertical lines). According to the formula

[0165]

[0166] Calculate the average sediment content of the cross section and store and publish it in real time.

[0167] 6. Model Updates: Monthly manual calibration is performed. New data is added to the training set, and incremental training with small batches and low learning rates is conducted starting from the current model to iteratively update the model and ensure its long-term adaptability.

[0168] This application creatively combines the precision advantages of traditional methods, the temporal advantages of continuous monitoring, and the spatial advantages of spectroscopic technology through deep fusion of multi-source data, providing a revolutionary solution for measuring suspended sediment concentration in cross sections.

[0169] This application creatively combines the precision advantages of traditional methods, the temporal advantages of continuous monitoring, and the spatial advantages of spectroscopic technology through deep fusion of multi-source data, providing a revolutionary solution for measuring suspended sediment concentration in cross sections.

[0170] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented in various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0171] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0172] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0173] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0174] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0175] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for monitoring the concentration of suspended sediment in a cross-section, characterized in that, include: Collect multi-source heterogeneous data of the entire cross-section of the river to be monitored, as well as the actual sediment concentration data corresponding to the multi-source heterogeneous data of the samples. Preprocess the sample multi-source heterogeneous data and the actual sand-containing data corresponding to the sample multi-source heterogeneous data to obtain the sample input data and the truth label corresponding to the sample input data. Based on the sample input data and the ground truth labels corresponding to the sample input data, a multi-source data spatiotemporal fusion network that dynamically integrates multi-source features through an attention mechanism is trained to learn the mapping relationship from multi-source data to the sand content of the cross-section vertical line, and to obtain the trained multi-source data spatiotemporal fusion network. Collect real-time multi-source heterogeneous data corresponding to the entire cross-section of the river to be monitored, and preprocess the real-time multi-source heterogeneous data to obtain real-time input data; The trained multi-source data spatiotemporal fusion network is scheduled to analyze the real-time input data and obtain the predicted sediment concentration of each vertical line on the entire cross section of the river to be monitored. Based on the predicted sediment concentration of each vertical line on the entire cross section of the river to be monitored, the real-time average sediment concentration of the corresponding cross section of the entire cross section of the river to be monitored is obtained.

2. The method according to claim 1, wherein, The method also includes: Acquire manually tested data from human-computer interaction input; the manually tested data includes multi-source heterogeneous data and manually tested sand content; Based on the artificial test data, the trained multi-source data spatiotemporal fusion network is adjusted, and the adjusted multi-source data spatiotemporal fusion network is used to analyze the sediment content in the subsequent cross-section suspended sediment concentration collaborative monitoring process.

3. The method according to claim 1, wherein, Collect multi-source heterogeneous data of the entire cross-section of the river to be monitored, as well as the actual sediment content data corresponding to the multi-source heterogeneous data, including: For the entire cross-section of the river to be monitored, spatial continuous spectral image data of the water body are collected at a preset sampling frequency. For the entire cross-section of the river to be monitored, real-time continuous sediment concentration time series data are collected along each vertical line of the entire cross-section of the river to be monitored at a preset sampling frequency. The actual sediment concentration data of each vertical line in the full cross section of the river to be monitored were obtained by manual analysis and laboratory precision analysis, and the actual sediment concentration data corresponding to the multi-source heterogeneous data of the sample were obtained.

4. The method for coordinated monitoring of suspended sediment content in cross-sections according to claim 3, characterized in that, Preprocessing is performed on the multi-source heterogeneous data of the samples and the actual sand-containing data corresponding to the multi-source heterogeneous data of the samples to obtain the sample input data and the ground truth labels corresponding to the sample input data, including: The sample multi-source heterogeneous data and the actual sand-containing data corresponding to the sample multi-source heterogeneous data are synchronized in time and / or registered in space to obtain the registered sample multi-source heterogeneous data and the actual sand-containing data corresponding to the sample multi-source heterogeneous data. Based on the spectral images in the registered multi-source heterogeneous data of the samples, the spectral reflectance curves of the water surface pixels and spatial context features corresponding to each preset calculation vertical line are obtained to obtain image features; Based on the real-time continuous sediment concentration time series data in the multi-source heterogeneous data of the registered samples, the real-time sediment concentration time series corresponding to the measuring points on each preset calculation vertical line is determined to obtain the sensor characteristics; Obtain water level, flow rate, flow velocity, flow direction, water temperature and / or cross-sectional water depth distribution to obtain auxiliary data features; The image features, sensor features, and auxiliary data features are used together as sample input data; Based on the actual sediment-containing data corresponding to the multi-source heterogeneous data of the sample, the actual sediment-containing data of each measuring point on the preset calculation vertical line are subjected to velocity-weighted integral or arithmetic average to obtain the true value label corresponding to the sample input data.

5. The method according to claim 1, wherein, The attention mechanism dynamically fuses multi-source data spatiotemporal fusion networks, including: The system consists of a spectral feature extraction subnetwork composed of a one-dimensional convolutional neural network, a temporal feature extraction subnetwork composed of a long short-term memory network, an auxiliary feature encoding subnetwork composed of a fully connected network, a multi-source attention fusion layer, and an output layer. These structures are connected sequentially to establish a physical relationship between water and sediment physical quantities and neural network parameters.

6. The method according to claim 1, wherein Based on the sample input data and the corresponding ground truth labels, a multi-source data spatiotemporal fusion network that dynamically integrates multi-source features using an attention mechanism is trained to learn the mapping relationship from multi-source data to the sand content of the cross-section vertical line, and to obtain the trained multi-source data spatiotemporal fusion network, including: The loss function is constructed using mean squared error and absolute error; Using the sample input data as input, the ground truth label corresponding to the sample input data as the expected output, and minimizing the loss function as the objective, the backpropagation algorithm is used to train the multi-source data spatiotemporal fusion network that dynamically integrates multi-source features through the attention mechanism, thereby obtaining the trained multi-source data spatiotemporal fusion network.

7. The method according to claim 6, wherein, The loss function is constructed using mean squared error and absolute error as follows: ； in, Represents the loss function. Indicates the truth label, This represents the model's predicted data. This represents the threshold parameter.

8. The method according to claim 1, wherein, The trained multi-source data spatiotemporal fusion network is used to analyze the real-time input data to obtain the predicted sediment concentration of each preset calculation vertical line on the entire cross-section of the river channel to be monitored, including: The real-time input data is used as the input to the trained multi-source data spatiotemporal fusion network, and the predicted value output by the trained multi-source data spatiotemporal fusion network is obtained to obtain the predicted sediment concentration of each preset calculation vertical line on the entire cross section of the river to be monitored.

9. The method according to claim 8, wherein, Based on the predicted sediment concentration of each preset calculation vertical line on the entire cross-section of the river to be monitored, the real-time average sediment concentration of the corresponding cross-section of the entire cross-section of the river to be monitored is obtained as follows: ； wherein, represents the real-time average sediment concentration of the whole cross section of the river to be monitored, represents the predicted sediment concentration of the ith pre-designed calculation vertical line on the whole cross section of the river to be monitored, and N represents the total number of vertical lines, represents the local cross-sectional area represented by the ith pre-designed calculation vertical line.

10. A cross-section suspended sediment concentration synergic monitoring device for performing the cross-section suspended sediment concentration synergic monitoring method according to any one of claims 1 to 9, characterized in that, include: The first data acquisition module is used to acquire sample multi-source heterogeneous data corresponding to the full cross-section of the river to be monitored, as well as the actual sediment content data corresponding to the sample multi-source heterogeneous data. The data preprocessing module is used to preprocess the sample multi-source heterogeneous data and the actual sand-containing data corresponding to the sample multi-source heterogeneous data to obtain the sample input data and the truth labels corresponding to the sample input data. The model training module is used to train the multi-source data spatiotemporal fusion network that dynamically integrates multi-source features based on the sample input data and the ground truth labels corresponding to the sample input data, so as to learn the mapping relationship from multi-source data to the sand content of the cross-section vertical line and obtain the trained multi-source data spatiotemporal fusion network. The second data acquisition module is used to acquire real-time multi-source heterogeneous data corresponding to the entire cross-section of the river to be monitored, and to preprocess the real-time multi-source heterogeneous data to obtain real-time input data. The data analysis module is used to schedule the trained multi-source data spatiotemporal fusion network to analyze the real-time input data and obtain the predicted sediment concentration of each preset calculation vertical line on the full cross section of the river to be monitored. The result output module is used to obtain the real-time average sediment concentration of the entire cross-section of the river under monitoring based on the predicted sediment concentration of each preset calculation vertical line on the entire cross-section of the river under monitoring.

Images