River network monitoring and early warning system, method and device

Abstract

The application discloses a river network monitoring and early warning system and method, and mainly solves the technical problems of slow response and high artificial dependence of existing river network management. Through the operation of the river network detection and early warning system, the dependence of river network management on artificial inspection and manual operation is greatly reduced, especially the operation cost of the area with wide geographical range and poor environment is reduced. Compared with the traditional flood treatment, the river network detection and early warning system can discover potential risks earlier and more accurately, so that valuable time is saved for the start and disposal of emergency plans, and the disaster loss can be effectively reduced. In addition, through digital twin simulation, the river network system can be displayed in a panoramic and dynamic manner, and the decision can be simulated, so that the scientificity and efficiency of water resource scheduling and management can be improved.

Description

Technical Field

[0001] This invention relates to the field of water conservancy monitoring technology, specifically to a river network monitoring and early warning system, method, and equipment. Background Technology

[0002] In the fields of smart water management and urban flood control and drainage, the management of river networks, pumping stations, gates, and underground pipe networks is crucial. Currently, the mainstream operation and maintenance model still heavily relies on periodic manual on-site inspections and static hydraulic model analysis based on historical hydrological data.

[0003] The inventors discovered through long-term practical research that existing static models, unable to incorporate real-time dynamic data, often result in analytical results that lag and deviate from actual conditions. They frequently require manual on-site inspections to detect problems (such as water level exceeding warning levels or sudden changes in pipeline pressure) and report them before countermeasures can be initiated, leading to delayed decision-making and missed optimal control windows. Furthermore, manual inspections require significant manpower and resources, are extremely inefficient, and are limited by inspection cycles, making them even more challenging in geographically complex or sparsely populated areas. Failure to promptly detect and quickly locate and address sudden water network issues can easily lead to serious public safety accidents, environmental incidents, and substantial economic losses.

[0004] The information disclosed in this background section is intended only to enhance the understanding of the background technology of this disclosure and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0005] In view of at least one of the above technical problems, this disclosure provides a river network monitoring and early warning system and method, which aims to solve the technical problems of slow response and high dependence on manual labor in existing river network management.

[0006] According to the first aspect of this disclosure, a river network monitoring and early warning system is provided, comprising a data processing module for processing river network data collected by a sensor array, a water body floating object identification module for identifying the type and location of floating objects in the water body, a physical conversion module for converting the identification information output by the water body floating object identification module into a corresponding river network accessibility coefficient matrix, a water level and flow prediction module that corresponds to the fusion of a U-net encoder-decoder architecture and a bidirectional gated loop unit and corrects the pipeline flow capacity parameters based on the river network accessibility coefficient matrix, an abnormal early warning module that generates risk-level early warning information based on the real-time data of the sensor array, the prediction results of the water level and flow prediction module, and the identification results of the water body floating object identification module, and a three-dimensional digital module for constructing a digital twin scene and accessing real-time monitoring data and prediction and early warning information; the pipeline accessibility coefficient is used as one of the input feature channels of the water level and flow prediction module to quantitatively characterize the attenuation state of drainage flow capacity at each spatial grid location in the gridded river network geographic space.

[0007] In some embodiments of this disclosure, the data processing module includes a data missing layer, an anomaly correction layer, a specific filtering layer, and a normalization layer. The data missing layer is used to check the river network data according to the time series and fill in the missing data using linear interpolation or pre- and post-filling. The anomaly correction layer constructs a corresponding data statistics window and identifies outliers using 3-Sigma or interquartile range methods according to the data distribution within the data statistics window, correcting outliers exceeding the upper and lower bounds to the corresponding boundary values. The specific filtering layer performs exponential moving average filtering or Gaussian filtering on water temperature and water level data, and median filtering or Savitzky-Golay filtering on turbidity data after identifying the data type. The normalization layer is used for normalizing the filtered data.

[0008] In some embodiments of this disclosure, the data processing module further includes a data calibration layer, which calculates the baseline drift trend and signal-to-noise ratio of the sensor array based on normalized data, and calibrates or compensates the corresponding sensors when the baseline drift trend and signal-to-noise ratio exceed a set threshold.

[0009] In some embodiments of this disclosure, the water body floating object recognition module includes a backbone network for extracting feature maps from images, a neck network for feature fusion and outputting a fused feature map, and a prediction head for interpreting the fused feature map and outputting detection boxes and categories.

[0010] In some embodiments of this disclosure, the backbone network is a cross-stage local connectivity structure and redundant feature channels are removed by a channel pruning algorithm; the neck network is a path aggregation network architecture and integrates channel and / or spatial attention blocks at its feature fusion nodes for adaptive weighted fusion of features from different layers or channels; the convolutional layers of the backbone network and / or the neck network employ depth-separable convolution operators; the prediction head separates the target location regression and category classification tasks and synchronously outputs target information within different pixel ranges through multi-scale detection layers.

[0011] In some embodiments of this disclosure, the water level and flow prediction module includes an encoder for extracting multi-scale spatial features from the input multi-channel spatiotemporal tensor and storing output feature maps at each level; a bottleneck layer for serializing the features output by the encoder and learning their cross-temporal dependencies; a decoder for fusing the feature maps stored at the corresponding level of the encoder through a skip connection mechanism and reconstructing spatiotemporal features; and an output layer for generating future time-series prediction values ​​of flow and water level at key cross-sections of the river network based on the reconstructed features output by the decoder.

[0012] In some embodiments of this disclosure, the encoder includes an encoding processing block consisting of four cascaded layers, including a three-dimensional convolutional layer and a spatial downsampling layer; the bottleneck layer includes a bidirectional gated loop block for dependency modeling of sensor features via a bidirectional temporal loop mechanism; and the decoder includes a decoding processing block consisting of four cascaded layers, including a spatial upsampling layer and a three-dimensional convolutional layer.

[0013] In some embodiments of this disclosure, the anomaly warning module includes an instant diagnosis layer that performs real-time anomaly judgment based on preset physical thresholds and Boolean logic rules, a trend warning layer that identifies potential risks based on a historical disaster precursor pattern database and the matching similarity between the change rate and combination characteristics of real-time data and historical patterns, and a prediction diagnosis layer that performs risk warning based on the time-series prediction values ​​output by the water level and flow prediction module.

[0014] According to a second aspect of this disclosure, a river network monitoring and early warning method is provided, which is implemented based on the aforementioned river network monitoring and early warning system, and specifically includes the following steps: (1) The monitoring data output by the sensor array is processed and calibrated based on the data processing module; (2) The floating objects in the water body are identified in real time by the water body floating object identification module, and the water body floating objects are identified, and the river network accessibility coefficient matrix is ​​generated at the same time. (3) Construct a multi-channel spatiotemporal tensor containing a static spatial feature field, a dynamic monitoring field and the river network accessibility coefficient matrix, input the multi-channel spatiotemporal tensor into the water level and flow prediction module, and output the flow and water level time series prediction values ​​of key sections of the river network in a specific future time period. (4) Risk prediction is made based on real-time monitoring data, the time-series predicted values ​​of flow and water level and the corresponding database of historical disaster precursors, and the risk prediction results are input into the three-dimensional digital twin scene constructed by the three-dimensional digital module for display. (5) Control the river network regulation facilities according to the risk prediction.

[0015] According to a third aspect of this disclosure, a river network monitoring and early warning device is provided, comprising a data acquisition unit for collecting river network hydrological data, a data processing unit communicatively connected to the data acquisition unit and embedded with the river network monitoring and early warning system, an execution unit for controlling river network facilities, and an edge computing device deployed at corresponding control nodes of the river network and communicatively connected to the data processing unit; the data acquisition unit includes a sensor array and an image acquisition device; the edge computing device embeds a water body floating object identification module, a water level and flow prediction module, and an emergency decision rule base.

[0016] One or more technical solutions provided in this application embodiment have at least one of the following technical effects or advantages: The operation of the river network detection and early warning system greatly reduces the reliance on manual inspection and operation in river network management, especially reducing operating costs in geographically extensive and environmentally harsh areas; compared to traditional flood control, the river network detection and early warning system can detect potential risks earlier and more accurately, buying valuable time for the activation and handling of emergency plans, and effectively mitigating disaster losses; in addition, digital twin simulation can achieve a panoramic and dynamic display of the river network system and simulate decision-making, thereby helping to improve the scientific nature and efficiency of water resource scheduling and management. Furthermore, by establishing a link between video monitoring and hydrological models through the river network accessibility coefficient matrix, the blockage information identified by the floating debris identification module is integrated into the water level and flow rate change prediction of the water level and flow rate prediction module, improving prediction reliability. Attached Figure Description

[0017] Figure 1 This is a three-dimensional visualization diagram of the operational status of a river network in one embodiment of this application.

[0018] Figure 2 This is a terminal control interface for a river network monitoring and early warning system running in one embodiment of this application.

[0019] Figure 3 This is a schematic diagram of floating object identification in one embodiment of this application. Detailed Implementation

[0020] The programs involved or relied upon in the following embodiments are all conventional or simple programs in this technical field, and those skilled in the art can make conventional selections or adaptive adjustments according to specific application scenarios. Unless otherwise specified, the devices involved in the following embodiments are all conventional commercially available products.

[0021] To better understand the technical solution of this application, the above technical solution will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] To address the technical problems of slow response and high reliance on manual labor in existing river network management, this example discloses a river network monitoring and early warning system. Based on river network sensor data and static spatial characteristics, it predicts the risk of urban flooding caused by heavy rainfall at the source and its spatial distribution. In addition, based on real-time video streams, it automatically identifies key floating objects that may block drainage pipe network inlets and bridge culverts, achieving early warning of physical obstacles.

[0023] To ensure the reliability of the model input data, the monitoring data is processed and calibrated in this embodiment. Specifically, after the sensor data enters the river network monitoring and early warning system, the data processing module of the system first parses the data, extracting specific monitoring values, sensor IDs, data type tags (such as water temperature, water level, turbidity, etc.), and timestamps. The data is then organized into raw data points with structured labels, such as using the representation of {type, value, time}. In this example, the data processing module also includes a data missing layer, an anomaly correction layer, a specific filtering layer, and a normalization processing layer. The data missing layer of the data processing module checks the data stream of each sensor according to the time sequence, detecting and filling missing values. In this example, when the gap between missing data is less than a preset threshold (such as less than 3 consecutive sampling periods), a linear interpolation operation is performed, using the valid data before and after the missing point to calculate the filling value; for longer gaps, forward filling is used, that is, the last valid value is used to fill the gap. After filling in missing values, the anomaly correction layer performs anomaly processing on the data stream to identify and correct extreme outliers caused by transient interference or brief sensor malfunctions. Specifically, in this embodiment, when identifying anomalies in the data stream, a dynamic statistical window (e.g., the 100 most recent data points) is first constructed for each sensor. By calculating the statistics of the data within the window, outliers are identified using the 3-Sigma criterion for data that conforms to a normal distribution (e.g., water temperature under steady-state conditions). For data with unknown distribution or skewed distribution (e.g., turbidity), the interquartile range method is used to calculate upper and lower bounds, and data points exceeding the selected statistical boundaries are marked as outliers. In addition, to avoid time series breaks in this example, marked outliers are not directly deleted. Instead, outliers exceeding the upper bound are corrected to the upper bound value, and similarly, outliers exceeding the lower bound are corrected to the lower bound value. This smooths out data peaks and troughs while maintaining data continuity, creating favorable conditions for subsequent filtering operations.

[0024] To improve the effectiveness of filtering, in this embodiment, the special filtering layer performs differentiated filtering based on the physical characteristics of different data. Specifically, in this example, the data type of each data label is first checked during filtering. For data types labeled as water temperature or water level, considering that such physical quantities usually change slowly and the noise is mostly high-frequency random fluctuations, exponential moving average or Gaussian filtering is used to effectively smooth out small fluctuations and retain the trend information of its slow changes. For data types labeled as turbidity, since there are transient spikes (impulse noise) in the data in this example, median filtering is used to replace the impulse noise with the neighborhood median, so as to eliminate abnormal spikes while better protecting the signal edges (such as the start and end points of turbidity changes). In other embodiments, Savitzky-Golay filtering is used to smooth the data through local polynomial fitting, which preserves the higher-order moment characteristics of the signal to the greatest extent while reducing noise. The filtered data is normalized by a normalization layer before output, which maps the values ​​from different sensors with different dimensions to a specified scaling range. This allows us to obtain standardized, high-quality datasets.

[0025] In some other embodiments, the data processing module also includes a data calibration layer, which establishes a corresponding sensor data calibration mechanism. Specifically, the data calibration layer analyzes normalized long-term data, calculates long-term baseline drift trends and signal-to-noise ratio indicators, and when it is found that the drift trend or noise level of a certain sensor exceeds the reliability threshold set for its type, it is determined that the sensor needs calibration or maintenance, and then hardware calibration or software numerical compensation is performed to ensure the reliability and validity of the data.

[0026] To enable early warning of physical obstacles by automatically identifying critical floating objects that may block drainage pipe network inlets and bridge culverts under the strict resource constraints of edge computing devices, the river network monitoring and early warning system in this embodiment also includes a water body floating object identification module. This module acquires video streams or image frames by deploying high-definition image acquisition devices (such as webcams) at key locations in the river (such as under bridges, pipe inlets, and narrow river sections). Then, it automatically identifies, classifies, locates, and counts various floating objects (such as domestic waste, dead branches and leaves, and large obstacles) in the water body based on a lightweight target detection algorithm using a convolutional neural network.

[0027] In this embodiment, the water body floating object recognition module first adjusts the river image captured by the image acquisition device to a fixed size and performs pixel normalization. Then, a lightweight neural network identifies the image, finds the location and category of the floating objects, and outputs the coordinate frame, category and confidence level of each floating object.

[0028] Specifically, in this example, the water-based floating object recognition module consists of three tightly coupled parts: a backbone network, a neck network, and a prediction head. The backbone network acts as a feature extractor, rapidly scanning the entire image and converting it into a series of multi-layered feature maps rich in semantic information. In this example, the backbone network employs an improved CSP structure, which, based on the core segmentation-transformation-merging structure of CSP, splits and processes the feature map channels before merging them. This significantly reduces computational redundancy and memory consumption while deepening the network and mitigating the gradient vanishing problem. Furthermore, to quickly extract discriminative features from the image while avoiding excessive computation, depthwise separable convolutions are used in the backbone network layers instead of standard convolutions. Specifically, in the CSP transformation stage, a small number of standard convolutions are retained in the shallow layers to more fully capture initial details and textures (such as the subtle differences between the edges of floating objects and water ripples), while depthwise separable convolutions are applied comprehensively in the middle and deep layers. This reduces computational complexity and the number of parameters by an order of magnitude without changing the receptive field. Depthwise separable convolution decomposes standard convolution into two steps: channel-wise convolution and pointwise convolution. First, each input channel is convolved individually to extract spatial features. Then, a 1×1 convolution is used to mix the channel information, thereby reducing computational cost. The theoretical computational cost and number of parameters can be reduced to approximately 1 / 9 of standard convolution. In this embodiment, after feature extraction by the backbone network, three different resolution feature maps are output: a high-resolution feature map rich in detail and suitable for finding small objects; a medium-resolution feature map balancing detail and semantics; and a low-resolution feature map with a large receptive field suitable for finding large objects.

[0029] After feature extraction, feature fusion is performed through the neck network, exchanging information among the three sets of feature maps. This allows each resolution to see details and understand semantics simultaneously, ensuring that features at each scale contain rich contextual information. Specifically, in this embodiment, the neck network adopts a path aggregation network architecture: deep high-level semantic features, such as low-resolution feature maps, are upsampled (i.e., enlarged) and concatenated with shallow detailed features, such as medium-resolution feature maps; shallow features, such as medium-resolution feature maps, are downsampled (shrunken) and concatenated with deep features, such as low-resolution feature maps. However, considering that simple concatenation or addition fusion assumes all input features are equally important, which may degrade model performance in complex aquatic environments, an adaptive feature weighting mechanism is introduced in this example. This mechanism assigns appropriate importance weights to feature maps from different levels and resolutions during the fusion process, thereby optimizing feature representations before fusion. Specifically, in this example, this weighting mechanism is implemented by integrating channels and spatial attention blocks, or by using a bidirectional weighted feature pyramid fusion strategy. For example, when detecting small floating objects, the network can adaptively assign higher weights to high-resolution feature maps containing more details. Thus, through feature fusion, the feature map at each scale simultaneously contains low-level detailed information that helps with localization and high-level semantic information that helps with classification, thereby achieving the ability and robustness to detect floating objects of various sizes.

[0030] The prediction head interprets the fused high-quality feature maps, outputting detection boxes and categories. Specifically, traditional YOLO heads use a single convolution to predict both categories and boxes simultaneously, potentially causing task conflicts. Therefore, this example employs a decoupling head, separating object classification (e.g., determining whether a branch is a plastic bag) and bounding box regression tasks. The decoupling head uses two parallel branch networks: a classification branch and a regression branch. The classification branch consists of a convolutional layer, outputting the probability of each anchor box belonging to each category; the regression branch also consists of a convolutional layer, outputting the coordinate offset and size scaling of each anchor box. The detection head performs predictions on feature maps at three scales: large-scale feature maps detect small objects, and small-scale feature maps detect large objects. This avoids the conflict between the two tasks in traditional coupled designs, improving accuracy and convergence. Furthermore, in this example, an identical decoupling head is connected in parallel to all three scale feature maps output by the neck network, allowing the model to detect small objects in large-scale feature maps, medium-scale objects in medium-scale feature maps, and large objects in small-scale feature maps.

[0031] After the water-based floating object recognition module is established, its model is trained. In this example, floating object images from real-world water scenes are collected and labeled to construct a dataset. In this embodiment, the collected images cover monitoring images from different river channels, pipeline inlets, weather conditions, lighting conditions, and seasonal scenarios. Floating objects in the images are finely labeled, and the bounding boxes must be closely aligned with the targets. At the data input end, the front-end video stream is uniformly cropped to a 640×640 pixel three-channel RGB image, which serves as the standard input tensor for the model. The output end is defined as a structured vector containing the bounding box center coordinates, confidence scores, and class probabilities. In some embodiments, data augmentation is performed on the constructed dataset to improve the model's generalization ability under extreme weather and low-light conditions. In this example, Mosaic augmentation is used, stitching four random images together into one for training, forcing the model to learn to recognize targets at a smaller scale. In other embodiments, MixUp augmentation is used, randomly linearly mixing two images to improve the ability to distinguish blurred targets or target edges. In this embodiment, the AdamW optimizer is used during training, with a weight decay coefficient of 0.025. The learning strategy employs linear warm-up and cosine annealing, and the learning rate is set to 10 at the beginning of training. -3 Cosine annealing was used to guide a smooth decline in the later stages of training, and the batch size was set to 32 to balance memory load. In the loss function, Focal Loss was used for classification, which addresses the extreme class imbalance between foreground floating objects and complex water backgrounds by reducing the weight of a large number of easily classifiable background samples, making the model more focused on difficult-to-classify or sparse targets. For floating objects with irregular boundaries and varied shapes, CIoULoss was used for regression loss in this example, which not only optimizes the overlap of bounding boxes but also considers the matching degree of center point distance and aspect ratio.

[0032] In other embodiments, considering that each filter in a convolutional layer is responsible for extracting a specific feature, but not all filters are equally important, some filters may output near-zero activation or repeatedly extract features similar to other filters. To identify and remove redundant channels with low contribution, and to achieve module lightweighting, it is necessary to facilitate subsequent deployment on edge computing devices. In this example, channel pruning is used to perform sparsity compression on redundant feature channels in the backbone network. The L1 norm is used to evaluate the importance of each convolutional layer channel, and weight parameters with low contribution to the final decision are removed. Specifically, for a convolutional layer, the L1 norm of each filter is calculated as the sum of the absolute values ​​of all weights of that filter. The smaller the norm of the filter, the smaller its activation value and the lower its importance. Then, according to the pruning rate preset by the layer, the channel to be removed is selected, and the filter corresponding to the selected channel is removed. The input channel dimension corresponding to the pruned channel in the next layer's convolutional weights is also removed. In addition, after pruning the trained model, there is a risk of performance degradation. In this example, the original training data is used to train the pruned model for a small number of epochs to adapt it to the new structure. Referring to Figure N, the average detection accuracy of the water body floating object recognition module in the actual deployment environment reaches over 92%, fully meeting the business needs of real-time obstacle removal management of urban pipe networks.

[0033] The output of the floating object recognition module is a structured list of detections, each item including bounding box coordinates, object category, and confidence level. However, the output is a discrete list of objects based on image pixel coordinates. To enable its use in river network monitoring and control, this embodiment of the river network monitoring and early warning system also includes a physical conversion module. This module converts the visual information output by the floating object recognition module into physical parameters usable by the subsequent water level and flow prediction module. Specifically, in this example, the input to the physical conversion module is the detection list output by the floating object recognition module. Since this list cannot be directly used by the subsequent water level and flow prediction module, the physical conversion module first performs coordinate mapping, converting the image pixel coordinates of the detection boxes into actual geographic coordinates based on the mapping relationship between camera calibration parameters and the geographic information system. Then, it calculates the impact domain, specifically the physical range of the blockage impact based on the object category and size. For example, a large obstacle might affect a range of 10 meters downstream. The affected area is then mapped onto a 64×64 spatial grid required by the subsequent water level and flow prediction module. For each affected grid, a local accessibility coefficient between 0 and 1 is calculated based on the object type and coverage level. For example, the module assigns different Manning roughness correction coefficients to floating ice and domestic waste. A local accessibility coefficient of 1 indicates that the pipeline is completely unobstructed, while a local accessibility coefficient of 0 indicates complete blockage. Finally, these are merged into a 64×64 two-dimensional array as the river network accessibility coefficient matrix.

[0034] To enable the prediction of flow and water level at key sections of the river network over a future period, thereby providing a basis for decision-making in flood control, drought relief, and water resource allocation, this embodiment of the river network monitoring and early warning system also includes a water level and flow prediction module. Considering that traditional time series models (such as ARIMA-LSTM) can effectively process single-point time series, but struggle to effectively integrate key spatial information such as topography and pipeline layout, this example employs a lightweight U-RNN model that integrates U-net and Bi-GRU. This spatiotemporal fusion architecture not only predicts "when" flooding will occur but also accurately predicts "where" flooding will occur. Specifically, the U-shaped encoder-decoder of U-net is used to extract and compress multi-scale spatial features such as topography and pipeline layout; the bidirectional GRU is introduced in the decoding segment to capture the long- and short-term temporal dependencies of sensor data such as water temperature, water level, and turbidity.

[0035] Specifically, in this embodiment, the water level and flow prediction module includes an input layer, an encoder, a bottleneck layer, a decoder, and an output layer. For the input layer of the water level and flow prediction module, the input data is a five-dimensional tensor including batch size, time step, spatial grid height, spatial grid width, and the number of feature channels. In this example, the water level and flow prediction module specifically receives data from the past 12 hours, spaced at 10-minute intervals, for a total of 72 time steps. Spatially, the target river network area is divided into a 64×64 regular grid. Furthermore, the input data in this example contains six feature channels, which are, in order: dynamic water level depth, real-time cumulative rainfall, digital elevation model height, river network topology, water turbidity, and river network accessibility matrix. The river network accessibility matrix is ​​generated and continuously updated by the floating object recognition module and the physical conversion module based on the location, type, and size of the floating objects identified by the camera.

[0036] In this example, the encoder employs a four-level downsampling structure of U-Net, consisting of multiple convolutional and pooling layers. It progressively downsamples to extract multi-scale spatial features. Each level contains two 3×3 convolutional layers and one 2×2 max-pooling layer for downsampling. Each convolutional layer is followed by BatchNorm and ReLU activation. Through the encoder's four downsampling steps, the number of channels increases progressively from 6 to 256, while the spatial size is compressed from 64×64 to 4×4. Furthermore, this example uses skip connections to buffer the feature maps output from each encoder level, preserving fine-grained spatial topological information at the corresponding scale (such as the precise locations of drainage outlets and gates), thus providing anchor points for subsequent decoder reconstruction. Specifically, in the first level of the encoder, the dynamic river network accessibility coefficient matrix is ​​element-wise multiplied with the static river network topological feature map to achieve real-time correction and modulation of the river network's basic flow capacity. Therefore, when the blockage coefficient of a certain node decreases, the water level and flow prediction module can simulate the abnormal surge in local water level caused by poor drainage. By converting the discrete obstacle information of visual recognition into a continuous physical resistance field, the water level and flow prediction module can predict the risk of urban flooding caused by sudden blockage in a forward-looking manner, which significantly improves the scientific nature and timeliness of the early warning.

[0037] The bottleneck layer is used to serialize the spatial features output by the encoder and capture long-term dependencies in the time dimension using a bidirectional RGU. Simultaneously, it integrates real-time pipeline accessibility information to achieve spatiotemporal information fusion. Specifically, in this embodiment, the 4×4×256 feature map of each time step of the encoder's final output is expanded in the spatial dimension to convert the three-dimensional spatial features into a one-dimensional vector and achieve serialization. Further, this sequence is input into a two-layer bidirectional gated recurrent unit network for processing. The Bi-GRU captures complex spatiotemporal dependencies such as the top-down propagation law of water flow and the bottom-up backing effect through forward and backward scanning, including the time lag effect between upstream water flow and downstream response. Then, feature recombination is performed, taking the output of the last time step of the bidirectional RGU to obtain a 256-dimensional vector as the context representation of the entire sequence. This compresses the time series information into a fixed-length vector, which serves as the input to the subsequent decoder. In this process, since the local congestion information contained in the real-time river network accessibility coefficient matrix from the visual channel has been deeply integrated with the static pipeline topology features during the encoding stage, the learning of bidirectional RGU can dynamically assess how such local resistance changes will affect the future temporal evolution of the entire river network system.

[0038] The bottleneck layer processes the multi-scale spatial features of the input data extracted by the encoder and then passes these features to the decoder for sampling and reconstruction. In this example, the decoder employs a four-level structure symmetrical to the encoder, progressively upsampling the spatial dimensions of the feature map through transposed convolutional layers. During feature fusion, each decoder level uses skip connections to directly concatenate high-resolution spatial detail features cached in the corresponding encoder layer into the current feature. This ensures that the internal representation of the water level and flow prediction module, after deep temporal abstraction, remains precisely aligned with the physical spatial location in the real world, effectively avoiding the spatial information loss problem in deep networks.

[0039] Furthermore, on the full-basin feature field (64×64 grid) output by the decoder, the system locates the corresponding grid region based on the predefined geographical location of key cross-sections. Then, through region pooling or a dedicated small convolutional module, it extracts a feature vector representing the hydrological state of the cross-section from the features of that region. This feature vector is then fed to three parallel fully connected prediction heads to generate predicted flow and water level sequences for the next 6 hours, 24 hours, and 72 hours, respectively. In some embodiments, the decoder output is mapped to a regional water accumulation heatmap via a 1×1 convolution to visually display the spatial distribution of water accumulation.

[0040] Furthermore, when training the model for the water level and flow prediction module, considering that directly processing long-sequence spatiotemporal data would lead to problems such as gradient instability, huge memory consumption, and mismatch between the initial hidden state and the data context during the initial training phase, this example employs a two-stage training method based on sliding window warm-up and formal training. In the warm-up phase, at the start of training, all model weight parameters are frozen, and only the first part of the current sample sequence is input into the network for forward propagation. This allows the Bi-GRU to generate an initial hidden state that matches the current hydrological situation based on this data, replacing the traditional zero-value or random initialization. In the formal training phase, the initialized hidden state is maintained, the model weights are unfrozen, and the complete duration sequence of the sample is used for forward propagation again. The loss is calculated, and backpropagation is performed to update all parameters. In this embodiment, a cosine annealing strategy is used during training, and the learning rate is stabilized at 10. -4 The batch size is set to 16.

[0041] Furthermore, considering that the water level / discharge prediction module needs to simultaneously predict cross-sectional water levels / discharges at three different time scales—6, 24, and 72 hours—a weighted composite loss function is used in this example to guide the model to learn different objectives in a balanced manner. Specifically, the loss is calculated separately for each of the three prediction periods and then summed according to their importance. For the 6-hour prediction, Huber loss is used, which is insensitive to outliers and ensures robustness under complex hydrological fluctuations. For the 24- and 72-hour predictions, smoothed L1 loss is used to capture the medium- to long-term trends in water level changes. In addition, when allocating weights, the weights decrease sequentially from 6 hours to 24 hours and then to 72 hours to guide the model to prioritize short-term warning accuracy.

[0042] In this embodiment, the river network monitoring and early warning system also includes an anomaly early warning module. This module identifies and judges potential anomalies in the river network (such as water levels exceeding warning lines, or pipelines at risk of freezing) based on preset physical thresholds (e.g., safe water levels, critical low temperatures), historical data statistical patterns, time prediction results, and floating object identification information. It then generates corresponding early warning information according to preset risk levels and response mechanisms. Specifically, in this example, the anomaly early warning module includes an instant diagnosis layer, a trend early warning layer, and a predictive diagnosis layer. In the instant diagnosis layer, the anomaly early warning module scans real-time data and performs Boolean logic rule judgments based on predefined physical rules. For example, when the water level sensor reading instantaneously exceeds the red warning line for that point, or when the water temperature is below freezing and turbidity abnormally increases, the diagnosis layer immediately triggers a definite alarm event to ensure the river network monitoring and early warning system's basic response capability to clear and urgent risks. In the trend early warning layer, the anomaly early warning module performs comparisons and early warnings based on the module's internal precursor pattern library. The precursor pattern library is extracted from historical disaster case datasets. Each pattern in the library describes the characteristic shape and coupling relationships of multidimensional data, such as the rate of water level rise and cumulative rainfall curves, in the hours leading up to a disaster. Furthermore, in real-time operation, the anomaly warning module performs a high-precision similarity comparison between the dynamic changes in the current short-term data and the data in the historical precursor pattern library. For example, even if the current absolute water level has not yet exceeded the warning level, if its rapid rise trend highly matches the data trajectory of a previous historical flooding event, the anomaly warning module will determine that there is a precursor to a disaster and issue an early warning of potential flooding risk, thus significantly advancing the warning window and allowing more preparation time for emergency response. In addition, the prediction and diagnosis layer is used to analyze and interpret the future cross-sectional water level prediction sequence output by the water level and flow prediction module, and make warning judgments accordingly. For example, when the water level and flow prediction module predicts that the water level at a certain cross-section will exceed the guaranteed water level in the next 3 hours, the system will generate a high-confidence predictive flood warning even if everything is currently normal.

[0043] For example, in the historical disaster database of this case, there is a typical pattern: water level rises by more than 15 centimeters within 30 minutes, corresponding to rainfall exceeding 20 mm, which is a strong characteristic leading to flooding in a certain area. In an actual monitoring, although the current water level has not yet reached the set maximum safety threshold, the river network monitoring and early warning system detected that "the current water level has risen rapidly by 10 centimeters within 20 minutes, and rainfall is accumulating rapidly." At this time, the matching algorithm of the abnormal early warning module will determine that the current rate of water level change and rainfall combination highly match the historical disaster pattern, thus immediately triggering a potential "flooding warning."

[0044] To integrate the early warning information from each layer of the anomaly early warning module, this example also includes a comprehensive layer. This comprehensive layer is used to deduplicatize and merge simultaneous and related early warnings, thereby generating an early warning report that includes the risk type, expected development timeline, diagnostic criteria, and the specific spatial range affected.

[0045] Furthermore, to achieve visualization of early warnings in this example, the river network monitoring and early warning system also includes a 3D digital module. This module is based on high-precision Geographic Information System (GIS) data, a 3D model of the river network topology, pipelines, pumping stations, and other facilities, and utilizes professional 3D engines such as Unity3D to construct a digital twin scene highly consistent with the physical river network. Additionally, by real-time input of monitoring data, prediction results, and early warning information to the 3D digital module, see [link to documentation]. Figure 1 The system dynamically renders the operational status of the river network (such as water flow animation, water level changes, and floating debris distribution) through data-driven methods, and intuitively displays early warning areas and event details in a 3D visualization, enabling virtual inspections by management personnel. Furthermore, in this example, the 3D digital module can also simulate the impact of different scheduling schemes (such as starting a water pump or adjusting gate opening) on ​​the river network's water conditions to evaluate scheduling effectiveness and optimize emergency response plans.

[0046] Furthermore, this example also discloses a river network monitoring and early warning method, which is implemented based on the aforementioned river network monitoring and early warning system, and specifically includes the following steps: (1) The monitoring data output by the sensor array is processed and calibrated based on the data processing module.

[0047] Specifically, the river network monitoring and early warning system periodically collects sensor data deployed throughout the river network via an IoT gateway. The data processing module checks the continuity of the collected data. If a sensor's data packet is lost, linear interpolation or forward padding is used to complete the data. Then, for each sensor's data stream, a sliding window of length 100 is maintained. Outliers within the window are calculated and identified in real time, and outliers are capped by replacing the original values ​​with boundary values. The appropriate filtering algorithm is selected based on the data type, and all sensor data is normalized to the [0, 1] interval after filtering. Simultaneously, the system continuously analyzes the long-term trend of each sensor's data. When the trend exceeds the allowable drift threshold, a reverse compensation coefficient is applied at the software level, or a hardware maintenance work order is generated.

[0048] (2) The floating objects in the water body are identified in real time by the water body floating object identification module, and the water body floating objects are identified, and the river network accessibility coefficient matrix is ​​generated at the same time.

[0049] Video streams are acquired using network cameras deployed under key bridges, at drainage outlets, and at river bends. Inference is performed on each frame of the video stream to output the bounding boxes, categories, and confidence scores of all floating objects in the images. Using pre-calibrated camera parameters, the pixel coordinates in the images are transformed to the actual geographic coordinate system, and the target river network area is divided into a 64×64 regular geographic grid. For each detected floating object, an initial local accessibility coefficient is generated at the corresponding grid location based on its category and size. Then, a Gaussian kernel is used to spatially diffuse the influence of each object, simulating the local range of the blockage effect. The minimum influence coefficient is taken for multiple objects within the same grid. Finally, a 64×64 river network accessibility coefficient matrix is ​​generated, where each element has a value between [0, 1], representing the drainage capacity of that grid cell, where 0 represents complete blockage and 1 represents complete unobstructed flow.

[0050] (3) Construct a multi-channel spatiotemporal tensor containing a static spatial feature field, a dynamic monitoring field, and a river network accessibility coefficient matrix. Input the multi-channel spatiotemporal tensor into the water level and flow prediction module, and output the flow and water level time series prediction values ​​of key sections of the river network in a specific future time period.

[0051] Specifically, the spatiotemporal baseline for prediction is first established. In this example, the current moment is used as the time reference, and snapshots are taken every 10 minutes over the past 12 hours, totaling 72 time steps. Spatially, a 64×64 grid, identical to the river network accessibility matrix, is used. For each time step, a 6-channel feature map is constructed, including dynamic water level depth, real-time cumulative rainfall, digital elevation model height, river network topology, water turbidity, and the river network accessibility matrix. The 6-channel feature maps from the 72 time steps are then stacked chronologically and input into the water level and flow prediction module for prediction. The water level and flow prediction module outputs corresponding water level and flow prediction sequences for the next 6 hours, 24 hours, and 72 hours. In some other embodiments, a 64×64 probability map of water accumulation risk for the next hour is also output simultaneously.

[0052] (4) Risk prediction is made based on real-time monitoring data, flow and water level time-series prediction values ​​and historical disaster precursor pattern database, and the risk prediction results are input into the three-dimensional digital twin scene constructed by the three-dimensional digital module for display.

[0053] (5) Control the river network regulation facilities according to the risk prediction, such as the start and stop of the antifreeze heating device, the switch of the drainage pump, and the adjustment of the opening of the shut-off valve, thereby realizing the intervention of the river network status.

[0054] Furthermore, this example also discloses a river network monitoring and early warning device, which includes a data acquisition unit, a data processing unit, and an execution unit. In this example, the execution unit includes sensor arrays and image acquisition devices deployed under key bridges, at drainage outlets, and at river bends, such as temperature sensors, non-contact ultrasonic water level sensors, turbidity sensors, flow velocity sensors, and rain gauges. The data processing unit includes a cloud control system that communicates with each sensor, and the aforementioned river network monitoring and early warning system runs within this cloud control system. The execution unit is used to control river network facilities according to instructions from the cloud control system, such as starting and stopping antifreeze heating devices, switching drainage pumps on and off, and adjusting the opening of shut-off valves.

[0055] Considering the risk of communication interruptions due to extreme weather, geological disasters, or network failures, which could prevent effective handling of emergencies, this embodiment deploys lightweight edge computing devices (such as NVIDIA Jetson series modules) at key control nodes (e.g., important pumping stations, sluice gates, and flood-prone monitoring stations) within urban river and drainage networks. In this example, the edge computing device embeds the aforementioned lightweight waterborne object identification module and water level / flow prediction module, enabling it to perform short-term predictions and identifications based on local sensor data (water level gauges, cameras, etc.) even with limited computing power at the edge. Furthermore, the edge computing device also embeds an emergency decision rule base, which defines the highest-priority safety operation logic that is autonomously executed based on local analysis results and sensor readings during communication interruptions.

[0056] When the network connection is normal, the edge nodes corresponding to each edge computing device act as distributed sensing and execution terminals, uploading data to the cloud control system and receiving scheduling instructions from the cloud control system. When the edge computing device detects an interruption in the communication link with the cloud, each device immediately switches to offline mode. In offline mode, each node relies on its local lightweight module to perform real-time analysis of the collected data (e.g., predicting the water level change trend upstream of the node in the next 30-60 minutes) and makes decisions based on a built-in emergency rule base. For example, the rule in the rule base is set as: "IF (local water level > warning threshold) AND (water level and flow prediction module predicts that the water level will continue to rise in the next 30 minutes) THEN (automatically start the drainage pump associated with this node and gradually increase it to the preset safe power)". This allows for a reasonable response to local emergencies even when the connection with the cloud control unit is lost, thereby preventing the accident from escalating.

[0057] In addition, in this embodiment, the data collected locally during offline periods, the prediction results, and the decision logs executed are stored in the edge device. After the network is restored, the above data is synchronized to the cloud control system for subsequent analysis and model optimization.

[0058] To verify the effectiveness of the aforementioned river network monitoring and early warning system, this case study describes the deployment of the system near Jingsi Garden in the Niyang River basin. Jingsi Garden is located along the Niyang River and is situated in a relatively low-lying area. This area faces dual flood risks: firstly, the risk of external flooding, where heavy rains or snowmelt in the upper reaches of the Niyang River cause a surge in water levels, directly threatening the riverbanks; secondly, the risk of internal flooding, where rising river levels exceed municipal drainage outlets, causing river water to backflow into the drainage network surrounding Jingsi Garden (i.e., "external backflow"). Furthermore, the Niyang River has a rapid current and often carries large amounts of floating debris such as branches and other contaminants, easily clogging drainage outlets and preventing local rainwater from draining, exacerbating internal flooding.

[0059] For this complex risk, see [link / reference] Figure 2 This example deploys an unattended pipeline guardian terminal running the aforementioned river network monitoring and early warning system. Node A is used for external flood monitoring. This node is deployed on the bank of the Niyang River outside Jingsi Garden, using an HC-SR04 ultrasonic module to monitor the real-time water level of the main channel of the Niyang River and a MaixCAM Pro camera aimed at the river surface. Node B is used for internal flooding monitoring. This node is deployed at the main stormwater outfall connecting to the Niyang River in the Jingsi Garden area, using an HC-SR04 to monitor the water level inside the pipeline network to determine if there is backflow of river water. Both devices are edge-computed locally using STM32 microcontrollers and connected in real-time to the Unity simulation platform in the cloud via a 4G module.

[0060] In a spring snowmelt event combined with short-duration heavy rainfall, the system demonstrated its effectiveness in multidimensional diagnosis: At T=0 minutes, the system, combining upstream hydrological data from the Linzhi Hydrological Bureau and meteorological radar data, predicted that the Jingsi Garden section of the Niyang River would experience a flood peak at T+60 minutes, with the water level exceeding the warning line. The system immediately and automatically issued a Level 1 warning to the management: risk of external flooding.

[0061] At T=30min, the MaixCAM Pro camera at node A detected a sharp increase in the density of floating debris (mainly broken tree branches) as the water level rose, and this debris was drifting towards the drain outlet at node B. See also Figure 3 The water body floating object identification module locks onto these floating objects in real time and reports them as "blockage risk".

[0062] At T=40min, the ultrasonic sensor at node A showed that the water level in the main channel of the Niyang River had exceeded the elevation of the drainage outlet at node B. The system immediately determined that there was a risk of "external water backflow".

[0063] At T=45min, the ultrasonic sensor at node B confirmed an abnormal rise in the water level within the pipe network, confirming backflow. Simultaneously, the floating debris identification module at node A confirmed that a large amount of floating debris had accumulated at the drain outlet, forming a physical blockage. At this point, the system triggered a complex high-risk alarm and alerted the management department via the Unity simulation interface: "Emergency—'External water backflow' has occurred at the Jingsiyuan drain outlet, and the floating debris identification module confirms 'outlet blockage.' The water level and flow prediction module predicts the flood peak will arrive in 15 minutes. It is recommended to immediately close the backflow prevention gate at the drain outlet and activate the park's internal forced drainage pumping station!" Relying solely on traditional water level gauges, management could only respond passively after backflow occurred at T=45 minutes. However, the river network monitoring and early warning system provided a 1-hour predictive warning at T=0 minutes, a diagnostic warning (identifying blockages) at T=30 minutes, and a comprehensive decision-making basis at T=45 minutes. This successfully shifted the emergency response time from reactive to proactive, gaining valuable time to prevent severe flooding in the Jingsiyuan area.

[0064] Although some preferred embodiments of the invention 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 the invention.

[0065] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this application and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A river network monitoring and early warning system, characterized in that, The system includes a data processing module for processing river network data collected by a sensor array; a water body floating object identification module for identifying the type and location of floating objects in the water; a physical conversion module for converting the identification information output by the water body floating object identification module into a corresponding river network accessibility coefficient matrix; a water level and flow prediction module that corresponds to the fusion of a U-net encoder-decoder architecture and a bidirectional gated loop unit and corrects the pipeline flow capacity parameters based on the river network accessibility coefficient matrix; an abnormal early warning module that generates risk classification early warning information based on the real-time data of the sensor array, the prediction results of the water level and flow prediction module, and the identification results of the water body floating object identification module; and a three-dimensional digital module for constructing a digital twin scene and accessing real-time monitoring data and prediction and early warning information. The network accessibility coefficient, as one of the input feature channels of the water level and flow prediction module, is used to quantitatively characterize the attenuation state of drainage flow capacity at each spatial grid location in the gridded river network geospatial space.

2. The river network monitoring and early warning system according to claim 1, characterized in that, The data processing module includes a data missing layer, an anomaly correction layer, a specific filtering layer, and a normalization processing layer; The data gap detection layer is used to check the river network data according to the time series and fill in the missing data using linear interpolation or pre- and post-filling; the anomaly correction layer constructs a corresponding data statistics window and identifies outliers using 3-Sigma or interquartile range methods according to the data distribution within the data statistics window, correcting outliers exceeding the upper and lower bounds to the corresponding boundary values; the specific filtering layer performs exponential moving average filtering or Gaussian filtering on water temperature and water level data, and median filtering or Savitzky-Golay filtering on turbidity data after identifying the data type; the normalization processing layer is used for normalizing the filtered data.

3. The river network monitoring and early warning system according to claim 2, characterized in that, The data processing module also includes a data calibration layer, which calculates the baseline drift trend and signal-to-noise ratio of the sensor array based on normalized data, and calibrates or compensates the corresponding sensors when the baseline drift trend and signal-to-noise ratio exceed a set threshold.

4. The river network monitoring and early warning system according to claim 1, characterized in that, The water body floating object recognition module includes a backbone network for extracting feature maps from images, a neck network for feature fusion and outputting fused feature maps, and a prediction head for interpreting the fused feature maps and outputting detection boxes and categories.

5. The river network monitoring and early warning system according to claim 4, characterized in that, The backbone network is a cross-stage local connection structure, and redundant feature channels are removed by a channel pruning algorithm. The neck network is a path aggregation network architecture, and channel and / or spatial attention blocks are integrated at its feature fusion nodes for adaptive weighted fusion of features from different layers or channels. The convolutional layers of the backbone network and / or the neck network use depth-separable convolution operators. The prediction head separates the target location regression and category classification tasks and outputs target information within different pixel ranges synchronously through multi-scale detection layers.

6. The river network monitoring and early warning system according to claim 1, characterized in that, The water level and flow prediction module includes an encoder for extracting multi-scale spatial features from the input multi-channel spatiotemporal tensor and storing output feature maps at each level; a bottleneck layer for serializing the features output by the encoder and learning their cross-temporal dependencies; a decoder for fusing the feature maps stored at the corresponding level of the encoder through a skip connection mechanism and reconstructing spatiotemporal features; and an output layer for generating future time-series prediction values ​​of flow and water level at key cross-sections of the river network based on the reconstructed features output by the decoder.

7. The river network monitoring and early warning system according to claim 6, characterized in that, The encoder includes a four-layer cascaded encoding processing block comprising a 3D convolutional layer and a spatial downsampling layer; the bottleneck layer includes a bidirectional gated loop block for dependency modeling of sensor features via a bidirectional temporal loop mechanism; the decoder includes a four-layer cascaded decoding processing block comprising a spatial upsampling layer and a 3D convolutional layer.

8. The river network monitoring and early warning system according to claim 1, characterized in that, The anomaly warning module includes an instant diagnosis layer that makes real-time anomaly judgments based on preset physical thresholds and Boolean logic rules; a trend warning layer that identifies potential risks based on a historical disaster precursor pattern database and the matching similarity between real-time data change rate and combination characteristics and historical patterns; and a prediction diagnosis layer that provides risk warnings based on the time-series prediction values ​​output by the water level and flow prediction module.

9. A method for monitoring and early warning of river networks, characterized in that, The implementation based on the river network monitoring and early warning system of claim 1 specifically includes the following steps: (1) The monitoring data output by the sensor array is processed and calibrated based on the data processing module; (2) The floating objects in the water body are identified in real time by the water body floating object identification module, and the water body floating objects are identified, and the river network accessibility coefficient matrix is ​​generated at the same time. (3) Construct a multi-channel spatiotemporal tensor containing a static spatial feature field, a dynamic monitoring field and the river network accessibility coefficient matrix, input the multi-channel spatiotemporal tensor into the water level and flow prediction module, and output the flow and water level time series prediction values ​​of key sections of the river network in a specific future time period. (4) Risk prediction is made based on real-time monitoring data, the time-series predicted values ​​of flow and water level and the corresponding database of historical disaster precursors, and the risk prediction results are input into the three-dimensional digital twin scene constructed by the three-dimensional digital module for display. (5) Control the river network regulation facilities according to the risk prediction.

10. A river network monitoring and early warning device, characterized in that, The system includes a data acquisition unit for collecting river network hydrological data, a data processing unit that is communicatively connected to the data acquisition unit and embedded with the river network monitoring and early warning system, an execution unit for controlling river network facilities, and an edge computing device deployed at the corresponding control nodes of the river network and communicatively connected to the data processing unit. The data acquisition unit includes a sensor array and an image acquisition device. The edge computing device embeds the water body floating object identification module, the water level and flow prediction module, and an emergency decision rule base.

Images