Hydrological flow monitoring system and software products based on edge computing gateway devices

Abstract

This invention discloses a hydrological flow monitoring system and software product based on an edge computing gateway device, including an information acquisition device, an edge computing gateway device, and a central station server. The system deploys an online flow monitoring model on the edge computing gateway device, enabling localized real-time analysis of flow velocity data and calculation of cross-sectional flow. Compared to a centralized cloud computing model, this system solves problems such as high data transmission latency, large network bandwidth consumption, and insufficient real-time response capabilities at the edge. It enables localized real-time analysis of flow velocity data and calculation of cross-sectional flow, effectively reducing dependence on the central station server. It can meet the needs of cross-sectional flow estimation under different water conditions (high, medium, and low water levels), and provides a more convenient method for estimating flow using various flow measurement techniques. This helps to further explore the applicability of online flow measurement equipment in hydrological stations along major rivers and the potential of edge computing, providing technical support for faster and more efficient online flow monitoring and the construction of digital twin hydrological stations.

Description

Technical Field

[0001] This invention belongs to the field of hydrological flow monitoring technology, and relates to a hydrological flow monitoring system and software product based on an edge computing gateway device, specifically a hydrological flow monitoring system and software product based on a dynamic topology adaptive edge computing gateway device. Background Technology

[0002] River flow monitoring is a crucial foundation for hydrological and water resources management, flood control and drought relief decision-making, and water conservancy project scheduling. With the advancement of smart water conservancy construction, traditional manual flow measurement and fixed monitoring equipment are no longer sufficient to meet the demands of high-frequency, real-time, and wide-coverage modern hydrological monitoring. Currently, while online flow measurement technologies based on acoustic Doppler current profilers, radar current meters, and video image current meters are gradually being adopted, their data computation still largely relies on centralized cloud computing platforms. This results in problems such as high data transmission latency, large network bandwidth consumption, and insufficient real-time response capabilities at the edge, which are particularly prominent in remote areas or extreme weather scenarios.

[0003] In recent years, edge computing technology has provided low-latency and highly reliable data processing capabilities for IoT scenarios by pushing computing resources down to network edge nodes. Summary of the Invention

[0004] To address the limitations of the cloud-based centralized computing model in existing online flow monitoring systems, this invention proposes a hydrological flow monitoring system and software product based on a dynamically adaptive edge computing gateway device. By deploying the online flow monitoring model on the edge computing gateway device, it enables localized real-time analysis of flow velocity data and cross-sectional flow calculation, which can adapt to various online flow measurement scenarios at hydrological stations.

[0005] The technical solution adopted by the system of the present invention is: a hydrological flow monitoring system based on an edge computing gateway device, comprising an information acquisition device, an edge computing gateway device, and a central station server;

[0006] The information acquisition device includes a water level sensor, a flow rate sensor, a water quality sensor, a temperature and humidity sensor, and a camera.

[0007] The edge computing gateway device includes a multi-protocol interface data acquisition module, a protocol conversion module, a preprocessing and feature extraction module, a real-time analysis and inference module, and a local storage security encryption module. It is used to periodically acquire data from the information acquisition device and perform real-time processing and analysis, upload abnormal data and periodic summary reports to the central station server, periodically receive lightweight models issued by the central station server, optimize and update the local online flow measurement model, and promptly feed back data features to the central station server to support the training of the local online flow measurement model, thereby achieving collaborative optimization of the local online flow measurement model by the central station server and the edge computing gateway device.

[0008] The central station server includes a visual operation and maintenance management subsystem, a big data storage device, a data warehouse, a digital twin module, a local online flow measurement model, and a model distillation engine. After receiving the periodic summary report from the edge computing gateway, it trains a high-precision local online flow measurement model based on global data, compresses it into a lightweight model that can adapt to the computing power of the edge computing gateway device through the knowledge distillation engine, and finally sends the updated lightweight model to the edge computing gateway device for local traffic calculation.

[0009] Preferably, the lightweight model compressed using a knowledge distillation engine to adapt to the computing power of the edge computing gateway device is specifically implemented as follows:

[0010] By deploying heterogeneous teacher model groups on the central station server, the knowledge output of the heterogeneous teacher model groups is dynamically integrated; based on the hydrological features reported in real time by edge computing gateway devices, the weights of each teacher model are dynamically allocated to generate joint soft tags.

[0011] To optimize the student model based on the computing power of edge computing gateway devices, operator reconstruction, hybrid quantization, and dynamic pruning are employed.

[0012] Edge computing gateway devices use local private data to fine-tune student models and generate hydrological event probability soft labels; central station servers cluster and aggregate multi-node soft labels and update global teacher models; in response to data drift, incremental online distillation is triggered, teacher models generate hydrological event probability soft labels in real time, and student models complete fine-tuning with small samples within a preset time to quickly adapt to distribution changes.

[0013] The student model is deployed on the edge computing gateway device, and a multi-objective optimization of latency, accuracy, and energy consumption is achieved based on the computation offloading strategy of reinforcement learning.

[0014] As a preferred embodiment, the knowledge output of the heterogeneous teacher model group is dynamically fused by deploying a heterogeneous teacher model group on the central station server; the specific implementation process includes multi-model collaborative decision-making, dynamic weighting based on environmental perception, intelligent conflict mediation, generation of fusion instruction packages, and dynamic execution at the edge.

[0015] The multi-model collaborative decision-making involves deploying four types of professional models on the central station server: a physical mechanism model that calculates flow based on hydrophysical laws, an AI time series model that learns water level change patterns from historical data, a visual recognition model that analyzes water level scale images captured by cameras, and a topology analysis model that combines data from other monitoring points within the basin for correlation prediction; each type of model independently outputs flow prediction results and confidence scores.

[0016] The dynamic weighting of environmental perception is achieved by constructing an intelligent weight allocator through real-time collection of environmental parameters; when equipment malfunctions, the weight of the corresponding sensor association model is reduced; and the weights are dynamically refreshed to ensure the optimal combination.

[0017] The intelligent conflict resolution mechanism prioritizes the adoption of predictions that conform to physical rules when the differences between multiple model predictions exceed a threshold, compares the model performance records of similar historical scenarios, and finally makes dynamic arbitration based on the recent accuracy of each model.

[0018] The generation of the fusion instruction packet includes generating a transmission packet containing three types of core information: the final predicted value of the weighted average result, the confidence interval reflecting the reliability of the prediction, and key feature indicators marking abnormal states.

[0019] The aforementioned dynamic execution at the edge refers to the direct guidance of the student model to update parameters after the edge computing gateway device receives the fusion instruction packet.

[0020] As a preferred approach, based on the hydrological features reported in real time by the edge computing gateway device, the weights of each teacher's model are dynamically allocated to generate joint soft labels; the specific implementation includes understanding model features, generating state feature vectors, calculating matching degree, dynamically allocating weights, and fusion prediction;

[0021] Before system deployment, the understanding model features are defined by setting a unique capability feature vector for each local online flow measurement model participating in the prediction. This vector implicitly represents the prediction scenarios it is good at, which include different hydrological conditions at medium, high, and low water levels.

[0022] The generated state feature vector is generated when the edge computing gateway device reports the current hydrological data in real time. The system immediately inputs these raw data into a lightweight model, quickly analyzes and extracts the core features of the current data, and generates a state feature vector.

[0023] The matching degree calculation involves comparing the current hydrological status feature vector with the pre-set capability feature vector of each local online flow measurement model, calculating the similarity between the two, and generating a matching degree score for each local online flow measurement model. The higher the score, the more the current hydrological status matches the model's preferred prediction scenario.

[0024] The dynamic weight allocation involves the system collecting the "matching scores" of all local online flow measurement models, converting these scores, and ensuring that the sum is 1.

[0025] In the fusion prediction, each local online flow measurement model independently generates a prediction result based on the current real-time hydrological data, which serves as a node soft label. The system multiplies the prediction result of each local online flow measurement model by its dynamic weight value obtained in the previous step, and adds up the weighted prediction results of all models to obtain the final "joint soft label".

[0026] As a preferred approach, the central station server aggregates multi-node soft labels using KL divergence clustering to update the global teacher model. The specific implementation process is as follows: After collecting node soft labels, the central station server dynamically clusters based on pairwise KL divergence: predictions with differences less than a threshold are automatically clustered into one cluster, reflecting similar hydrological patterns. High-quality "cluster centers" are generated by fusing soft labels within clusters, and a weighted average based on confidence level or intra-cluster KL consistency is used to strengthen reliable and consistent prediction patterns. When updating the teacher model, each cluster center is treated as an independent "teacher," and its KL divergence loss with the global model prediction is calculated. These losses are then aggregated by weighting according to cluster size or confidence level. By minimizing this weighted loss to update the model, the teacher integrates diverse, high-confidence collective knowledge represented by different clusters, rather than a simple mixture of all nodes, thus improving robustness.

[0027] As a preferred option, TensorRT-LLM is used to push and deploy student models on edge computing gateway devices.

[0028] Preferably, the central station server is also equipped with a digital twin module, which is responsible for integrating multi-source data, constructing a virtual mapping of the watershed, supporting dynamic simulation and intelligent decision-making, and realizing the upgrade from edge perception to watershed-level regulation; including data fusion and dynamic modeling, simulation engine and decision-making closed loop, and visualization and collaborative management;

[0029] The data fusion and dynamic modeling integrates pre-processed real-time hydrological data and meteorological and geographic information from edge computing gateway devices to construct a high-precision three-dimensional watershed model, thereby achieving real-time mapping between the physical world and virtual space.

[0030] The simulation engine and decision-making closed loop rely on the hydrodynamic model to pre-simulate the flood peak propagation path and inundation range, and generate scheduling strategies; through reinforcement learning algorithms, the flood discharge and water storage demands are dynamically balanced to form a "monitoring-simulation-decision-execution" closed loop, and the instructions are sent to the edge computing gateway control device in real time.

[0031] The visualization and collaborative management features a 3D "monitoring station map" that overlays real-time data and risk heat maps; it supports cross-departmental command collaboration and pushes work orders via mobile devices to enable rapid handling of emergencies.

[0032] Preferably, the local online flow measurement model is:

[0033] ;

[0034] In the formula: The output of the model is the average flow rate at the measurement section. The independent variables of the model include the flow velocities represented by the points or lines monitored by the online flow measurement equipment; For the model parameter set, refer to the coefficients representing the flow velocity; The model operation rules include linear functions, polynomial functions, exponential functions, logarithmic functions, and power functions. The specific function type is determined after the model parameters are calibrated using a genetic algorithm based on the sample data.

[0035] Preferably, the system adopts a hardware-level encryption and redundancy design structure, including a dual power supply module for the hardware interface layer and multiple communication links, to ensure data privacy and the security and reliability of device operation.

[0036] The technical solution adopted by the software product of the present invention is: a hydrological flow monitoring software product based on an edge computing gateway device, comprising computer program instructions, which drive the system to perform hydrological flow monitoring when the computer program instructions are run on a computer.

[0037] This invention addresses the limitations of existing cloud-based centralized computing models in online flow monitoring systems by proposing an edge computing gateway-based architecture. Compared to existing technologies, this invention deploys a local online flow measurement model on an edge computing gateway device, enabling localized real-time analysis and anomaly warning of flow velocity data. This effectively reduces reliance on cloud-based central server servers and enhances the system's robustness in complex network environments through a lightweight transmission mechanism. Using online flow measurement with a horizontal acoustic Doppler current profiler (H-ADCP) at a hydrological station as a typical application, the technical advantages of the proposed architecture in reducing data transmission volume and improving monitoring timeliness are verified, providing a new technical path for edge computing applications in smart water conservancy scenarios. Attached Figure Description

[0038] The technical solutions of the present invention will be further illustrated below using embodiments and specific implementation methods. In addition, some accompanying drawings are used in the description of the technical solutions. Those skilled in the art can obtain other drawings and the intent of the present invention from these drawings without any creative effort.

[0039] Figure 1 This is a diagram illustrating the architecture of a hydrological flow monitoring system according to an embodiment of the present invention.

[0040] Figure 2This is a graph showing the relationship between index flow velocity and cross-sectional average flow velocity under different water flow characteristic conditions in the embodiments of the present invention;

[0041] Figure 3 This is a graph showing the relationship between the correction index flow velocity and the cross-sectional average flow velocity under different water flow conditions according to embodiments of the present invention.

[0042] Figure 4 This is a comparison diagram of the flow process lines of two schemes involved in the embodiments of the present invention. Detailed Implementation

[0043] To facilitate understanding and implementation of the present invention by those skilled in the art, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0044] Please see Figure 1 This embodiment provides a hydrological flow monitoring system based on an edge computing gateway device, including an information acquisition device, an edge computing gateway device, and a central station server; the information acquisition device is connected and communicates with the central station server via wired or wireless means, and the edge computing gateway device is connected and communicates with the central station server via wired or wireless means.

[0045] The information acquisition device includes a water level sensor, a flow rate sensor, a water quality sensor, a temperature and humidity sensor, and a camera; the general installation locations are shown in Table 1.

[0046] Table 1

[0047]

[0048] The edge computing gateway device, in its specific implementation, may include a multi-protocol interface data acquisition module, a protocol conversion module, a preprocessing and feature extraction module, a real-time analysis and multi-model inference module, and a local storage security encryption module. It is used to periodically acquire data from the information acquisition device and perform real-time processing and analysis, upload abnormal data and periodic summary reports to the central station server, periodically receive lightweight models from the central station server to optimize and update the local online flow measurement model, and simultaneously provide timely feedback of data characteristics (such as flow velocity distribution statistics) to the central station server to support the training of the local online flow measurement model. This enables collaborative optimization of the local online flow measurement model by the central station server and the edge computing gateway device.

[0049] The central station server, in its specific implementation, may include a visual operation and maintenance management subsystem, a big data storage device, a data warehouse, a digital twin module, a local online flow measurement model, and a model distillation engine. Upon receiving periodic summary reports from the edge computing gateway, it trains a high-precision local online flow measurement model based on global data (such as full-domain data), compresses it into a lightweight model adaptable to the computing power of the edge computing gateway device using the knowledge distillation engine, and finally distributes the updated lightweight model to the edge computing gateway device to guide local traffic calculation.

[0050] In one implementation, the process of compressing the data into a lightweight model adaptable to the computing power of the edge computing gateway device using a knowledge distillation engine includes:

[0051] By deploying heterogeneous teacher model groups on the central station server, their knowledge outputs are dynamically integrated; based on the hydrological features reported in real time by edge computing gateway devices, the weights of each teacher model are dynamically allocated to generate joint soft labels.

[0052] To optimize the student model based on the computing power of edge computing gateway devices, operator reconstruction, hybrid quantization, and dynamic pruning are employed.

[0053] (1) Operator reconstruction;

[0054] Principle: Reconstruct network layers or computation operators, such as replacing standard convolutions with depthwise separable convolutions to reduce multiply-accumulate operations.

[0055] Edge adaptation: Optimize the computing process for the instruction set of edge chips (such as the ARM Cortex-A series) to improve hardware utilization.

[0056] (2) Hybrid quantization;

[0057] Principle: Different precisions (such as a mix of 8-bit integers and 16-bit floating-point numbers) are used for different layers or weights. For example, sensitive layers (such as classification heads) retain FP16 precision; redundant layers (such as intermediate convolutions) are compressed to INT8.

[0058] Edge adaptation: By combining TensorRT or TFLite quantization toolchains, the model size and memory usage are reduced by 4 times while maintaining less than 2% accuracy loss.

[0059] (3) Dynamic pruning;

[0060] Principle: Dynamically activate some neurons based on input data (such as sparse attention mechanism) to skip redundant calculations. For example, during inference, prune low-contribution weights (‖W‖<θ) according to a threshold.

[0061] Edge adaptation: Utilizes the chip's sparse computing units (such as the Sparse TensorCore of the NVIDIA Ampere architecture) to accelerate sparse matrix operations.

[0062] Edge computing gateway devices use local private data to fine-tune student models and generate hydrological event probability soft labels; central station servers cluster and aggregate multi-node soft labels and update global teacher models; in response to data drift, incremental online distillation is triggered, teacher models generate hydrological event probability soft labels in real time, and student models complete fine-tuning with small samples within a preset time to quickly adapt to distribution changes.

[0063] The student model is deployed on the edge computing gateway device, and based on the computation offloading strategy of reinforcement learning, the network state and task load are used as the state, and the offloading decision is used as the action. The model is trained by integrating the reward function of latency, accuracy and energy consumption, and finally achieves the dynamic optimal balance of the three, realizing the multi-objective optimization of latency, accuracy and energy consumption.

[0064] In one implementation, the knowledge output of the heterogeneous teacher model group is dynamically fused by deploying it on the central station server; the specific implementation process includes multi-model collaborative decision-making, dynamic weighting based on environmental perception, intelligent conflict mediation, generation of fusion instruction packages, and dynamic execution at the edge.

[0065] The multi-model collaborative decision-making involves deploying four types of professional models on the central station server: a physical mechanism model that calculates flow based on hydrophysical laws, an AI time series model that learns water level change patterns from historical data, a visual recognition model that analyzes water level scale images captured by cameras, and a topology analysis model that combines data from other monitoring points within the basin for correlation prediction; each type of model independently outputs flow prediction results and confidence scores.

[0066] The physical mechanism model is based on the fundamental physical laws of fluid mechanics and hydrology (such as Manning's formula and Saint-Venant's equations) and constructs mathematical equations to describe the process of water flow. The input variables include real-time water level, rainfall, flow velocity and other data from monitoring points, as well as river parameters that have been calibrated and built into the model (such as river cross-sectional shape, Manning roughness coefficient, river length, gradient, etc.). The model uses physical formulas to calculate based on the input data and parameters and outputs the predicted results of river cross-sectional flow.

[0067] The AI ​​time series model described is a deep learning model such as Long Short-Term Memory (LSTM) or Transformer, which makes predictions by learning and memorizing long-term patterns and complex nonlinear relationships (such as the changes in water level, flow rate, and rainfall over time) in historical data. During the training phase, the model learns from historical data spanning several years. When making predictions, the model receives recent continuous time series data as input features, performs forward propagation of the input time series sequence within its internal network, and outputs the predicted flow rate at the river cross-section.

[0068] The visual recognition model uses computer vision (CV) technology to automatically identify and analyze water level gauge images captured by cameras deployed along the river, directly read the water level value, and then combine it with river cross-section information to convert it into flow rate, outputting water level prediction results.

[0069] The aforementioned topology analysis model constructs a graph neural network (GNN) around all monitoring points within the watershed, using data from upstream and neighboring stations to predict the flow at the target station. The model models the entire watershed as a graph, where each monitoring point is a node and the river channel is the edge connecting the nodes. The model learns the spatial dynamics of water flow propagation within the river network, allowing the target station node to "receive" information from its upstream and downstream neighbors (for example, a 10% increase in flow at an upstream station will, after one hour of propagation, lead to an approximately 8% increase in flow at the downstream target station). After integrating information from all associated stations, the model outputs the predicted flow value for the target station.

[0070] The central server runs these four models simultaneously, each outputting a predicted value and a confidence score. The final collaborative decision typically employs a confidence-weighted average. The advantage of this approach is that it allows the system to dynamically rely on the most reliable model under different conditions. For example, during periods of heavy rain and flooding (extreme events), physical and topological models may be more reliable than AI models learned solely from historical data; while at night, the visual model may have lower confidence due to dark images, and the system will automatically rely on other models. This multi-model collaboration significantly improves the accuracy, robustness, and reliability of the prediction system.

[0071] The dynamic weighting of environmental perception is achieved by constructing an intelligent weight allocator through real-time collection of environmental parameters such as water level, flow velocity, and equipment health. For example, in a rainstorm scenario, the weight of the physical model is automatically increased (it is more adaptable to extreme water flow changes); when equipment malfunctions, the weight of the corresponding sensor-related model is reduced; and the weight is dynamically refreshed every 5 minutes to ensure the optimal combination; a fault self-isolation mechanism is adopted here, and when several sensors fail, their data input links are automatically cut off; after the equipment is repaired, the weight is gradually restored (from 0.1 to 0.8 within 2 hours).

[0072] The intelligent conflict resolution mechanism prioritizes adopting predictions that conform to physical rules when the differences between multiple model predictions exceed a threshold, compares the model performance records of similar historical scenarios, and finally makes dynamic arbitration based on the recent accuracy of each model.

[0073] The fusion instruction packet generation generates a transmission packet containing three types of core information: the final predicted value of the weighted average result, the confidence interval reflecting the reliability of the prediction, and the key feature indication marking the abnormal state;

[0074] The dynamic execution at the edge directly guides the student model to update parameters after the edge computing gateway device receives the fusion instruction packet. When the network is interrupted, a local caching strategy is activated (reusing the effective weights of the most recent hour). When data is abnormal, a degradation mode is automatically triggered, using only physical rules and a simple AI model. The physical rules are based on the device's built-in sensor data and preset thresholds to make "if-then" logical judgments. The simple AI model uses extremely lightweight models (such as small decision trees, linear regression, or very small-scale neural networks). These models are pre-deployed and stored in the fixed storage area of ​​the edge computing gateway device, relying only on a few key features for rapid inference. Physical rules are used to handle deterministic and interpretable anomalies; the simple AI model handles scenarios that require slightly more complex pattern recognition (such as simple classification and regression prediction). A lightweight transmission protocol is used here, and a dedicated compression encoding scheme is designed. For weight parameters, only the change amount is transmitted (e.g., "physical model +0.2, visual model -0.3"); for environmental data, it is represented by 8-bit codes (e.g., code A7 = high water / device normal).

[0075] The student model in this embodiment adopts a "learning path of imitation before innovation" and strictly follows the fusion results for training in the initial stage; after stable operation, it allows for autonomous fine-tuning within the framework of physical rules.

[0076] In one implementation, based on the hydrological features reported in real time by the edge computing gateway device, the weights of each teacher model are dynamically allocated to generate joint soft labels; the specific implementation includes understanding model features, generating state feature vectors, calculating matching degree, dynamically allocating weights, and fusion prediction;

[0077] Before system deployment, the understanding model features are defined by setting a unique capability feature vector for each local online flow measurement model participating in the prediction. This vector implicitly represents the prediction scenarios it is good at, including different hydrological conditions such as medium, high, and low water levels.

[0078] The generated state feature vector is generated when the edge computing gateway device reports the current hydrological data such as water level and flow rate in real time. The system immediately inputs these raw data into a lightweight model, quickly analyzes and extracts the core features of the current data, and generates a state feature vector.

[0079] The matching degree calculation involves comparing the current hydrological status feature vector with the pre-set capability feature vector of each local online flow measurement model, calculating the similarity between the two, and generating a matching degree score for each local online flow measurement model. The higher the score, the more the current hydrological status matches the model's preferred prediction scenario.

[0080] The dynamic weight allocation involves the system collecting the "matching score" of all local online flow measurement models, converting these scores to ensure that the sum is 1 (i.e., the sum of the weights assigned to all models is 100%).

[0081] The fusion prediction involves each local online flow measurement model independently generating a prediction result based on the current real-time hydrological data, which serves as a node soft label. The system multiplies the prediction result of each local online flow measurement model by the dynamic weight value obtained in the previous step, and adds up the weighted prediction results of all models to obtain the final "joint soft label". This result is not a static average, but a comprehensive prediction dominated by the model that best adapts to the current hydrological conditions.

[0082] In one implementation, the cloud aggregates multi-node soft labels through KL divergence clustering to update the global teacher model. Specifically, after collecting node soft labels, the cloud dynamically clusters based on pairwise KL divergence: predictions with differences less than a threshold are automatically clustered into a single cluster, reflecting similar hydrological patterns. High-quality "cluster centers" are generated by fusing soft labels within clusters—using a weighted average based on confidence level or intra-cluster KL consistency to strengthen reliable and consistent prediction patterns. When updating the teacher model, each cluster center is treated as an independent "teacher," and its KL divergence loss with the global model prediction is calculated. These losses are then aggregated by weighting according to cluster size or confidence level. By minimizing this weighted loss to update the model, the teacher integrates diverse, high-confidence collective knowledge represented by different clusters, rather than a simple mixture of all nodes, thus improving robustness.

[0083] In one implementation, TensorRT-LLM is used to push and deploy student models on an edge smart gateway.

[0084] TensorRT-LLM is irreplaceable in hydrological flow monitoring systems based on edge computing gateway devices, mainly due to the following three advantages:

[0085] (1) Low latency, high throughput, and strong memory utilization. TensorRT-LLM features low latency and high throughput, significantly improving model inference speed through kernel fusion, dynamic batch processing, and quantization techniques. For example, on Jetson Orin edge devices, latency is reduced to 1 / 6 of that of traditional frameworks; it can meet the strong real-time requirements of hydrological flow monitoring, which needs to complete data cleaning, anomaly detection, and early warning decisions within milliseconds. It supports paging caching technology, reducing memory usage of long texts (such as 32k context) by 60%, enabling large models to run on resource-constrained edge computing gateway devices.

[0086] (2) Edge adaptation capability. Deeply optimized NVIDIA Jetson series edge hardware to fully utilize computing power potential. Through INT4 / AWQ quantization technology, the model size is reduced by 75% and the accuracy loss is controlled to <1%, which is suitable for the needs of real-time processing of sensor time series data in hydrological monitoring.

[0087] (3) Localization and convenient management of complex models. Hydrological analysis requires the integration of multimodal data (such as video water level recognition + sensor flow monitoring). TensorRT-LLM supports edge deployment of multimodal models and can be integrated with Triton Inference Server to realize multi-gateway cluster management, meet the edge deployment of multimodal models, and realize end-to-end integrated analysis and unified scheduling of distributed hydrological monitoring nodes.

[0088] While TensorRT-LLM offers superior performance and is the optimal solution for processing high-frequency sensor data with stringent real-time requirements, other solutions can be flexibly selected based on the specific scenario. See Table 2 below for a detailed analysis of their advantages and limitations.

[0089] Table 2

[0090]

[0091] In one embodiment, the system's hardware layer adopts a modular design, including a power supply module, a switching module, a self-test module, and multiple types of interfaces (such as RS485, RS232, SDI, CAN, AI / DI, etc.) to support various hydrological sensor devices. The power supply module ensures the stability of the device in the field environment through redundant power supply design (such as dual power supply switching); the self-test module monitors the hardware interface status in real time, automatically triggering alarms and switching to backup interfaces when abnormalities occur.

[0092] In one implementation, the digital twin module is responsible for integrating multi-source data, constructing a virtual watershed mapping, supporting dynamic simulation and intelligent decision-making, and realizing an upgrade from edge perception to watershed-level regulation; including data fusion and dynamic modeling, simulation engine and decision-making closed loop, and visualization and collaborative management;

[0093] (1) Data fusion and dynamic modeling. By integrating real-time hydrological data (such as water level and flow velocity) preprocessed by edge computing gateway devices and multi-source data such as meteorological and geographic information, a high-precision three-dimensional watershed model is constructed. Hydrological response units (HRUs) are divided using GIS and SWAT models, and parameters are dynamically optimized based on a meta-learning mechanism (such as automatically increasing the grid resolution by 50% when water level changes abruptly) to achieve real-time mapping between the physical world and the virtual space.

[0094] (2) Simulation Engine and Decision-Making Closed Loop. Based on hydrodynamic models such as the Saint-Venant equations, the propagation path and inundation range of the flood peak are simulated in the digital twin platform to generate scheduling strategies. Through reinforcement learning algorithms (such as CEM-MASAC), the demand for flood discharge and water storage is dynamically balanced to form a "monitoring-simulation-decision-execution" closed loop, and the instructions are sent to the edge computing gateway device in real time.

[0095] (3) Visualization and Collaborative Management. A 3D "single-site map" is rendered using engines such as Cesium or Unity, overlaid with real-time data and risk heat maps. Cross-departmental command collaboration is supported, and work orders can be pushed via mobile devices to enable rapid handling of emergencies.

[0096] In one implementation, the online flow measurement model is:

[0097] ;

[0098] In the formula: The output of the model is the average flow rate at the measurement section. The independent variables of the model include the flow velocities represented by the points or lines monitored by the online flow measurement equipment; For the model parameter set, refer to the coefficients representing the flow velocity; The model operation rules include linear functions, polynomial functions, exponential functions, logarithmic functions, and power functions. The specific function type is determined after the model parameters are calibrated using a genetic algorithm based on the sample data.

[0099] In one implementation, the system employs a hardware-level encryption and redundancy design structure, including a dual-power module for the hardware interface layer and multiple communication links (5G + BeiDou + short-range transmission + optical fiber) to ensure data privacy and the security and reliability of device operation.

[0100] This embodiment provides a hydrological flow monitoring software product based on an edge computing gateway device, which includes computer program instructions. When the computer program instructions are run on a computer, they drive the system to perform hydrological flow monitoring.

[0101] The technical solution adopted in this invention will be further illustrated by specific experiments below.

[0102] This experiment selected the Gaobazhou Hydrological Station, the outlet control station of the Qingjiang River, a tributary of the Yangtze River, as the subject of this example. The station has a catchment area of ​​15,650 km². 2 The tested river section is straight. Two kilometers upstream of the basic water gauge section is the Gaobazhou Hydropower Project, and approximately 1.1 kilometers upstream is the Gaobazhou Bridge. About 0.7 kilometers downstream is a bend, and about 10 kilometers downstream is the confluence of the Qingjiang River and the Yangtze River. The tested section is W-shaped, with a river width of approximately 340 meters. Hydrological changes at this station are mainly influenced by upstream hydropower projects and the backwater effect of downstream Yangtze River fluctuations, resulting in significant variations in the elevation difference between the upstream and downstream sections and a complex relationship between water level and flow rate. The station monitors the flow velocity at the H-ADCP monitoring section and the real-time water level using a float level gauge. Input data includes measured hydrological elements from 2017 to 2020 during medium-to-high water conditions, corresponding to the H-ADCP data. In one specific embodiment, this data represents the measured flow rate from a rotor-type current meter at the river section, totaling 75 samples, with water levels ranging from 38.58 to 50.24 meters and flow rates ranging from 23.6 to 6540 cubic meters per second. 3 / s.

[0103] In this experiment, the Gaobazhou hydrological station employed an edge computing gateway algorithm model to determine the settings of relevant measurement parameters for the H-ADCP (Horizontal Acoustic Doppler Profiler) at the station. Representative velocity ranges were selected, the correlation between representative velocity and cross-sectional average velocity was established, and a flow estimation scheme was determined. The edge computing gateway device adopted a water conservancy intelligent edge computing gateway, which incorporates large-section data, monitoring section information such as starting point distance, and an online flow monitoring model to achieve localized flow calculation and visualization. The collected hydrological measurement environment information included cross-sectional shape, riverbed elevation, and the installation location of the flow measuring equipment.

[0104] The velocity distribution at the H-ADCP test section is affected by the backwater effect of the Yangtze River and the water level difference between the test section and the Yangtze River estuary, resulting in a complex water level-discharge relationship. The original index velocity measured by H-ADCP has a non-uniform relationship with the average velocity of the section. Therefore, the index velocity is not suitable for calculating the average discharge of the section. It is necessary to construct a "hypothetical" corrected index velocity to establish a uniform relationship with the average velocity of the section. For example... Figure 2 As shown, taking the flood discharge period of June 28, 2020 as the node, the H-ADCP index flow velocity was... Average flow velocity at measured cross section By establishing two relationship lines, the average flow velocity at the same cross-section corresponds to two index flow velocities. and .consider Influenced by factors such as the degree of backwater in the Yangtze River, the head difference between the measuring section and the outlet, and the rate of rise and fall of the Yangtze River water level, each factor is compared with... The mathematical relationship is established as shown in Equation (1), and the calculation method for the corrected index flow rate is shown in Equation (2). The functional form is determined by experience and trial and error. The corrected index flow rate consists of the original index flow rate and the index flow rate increment caused by various factors.

[0105]

[0106]

[0107] In the formula: The increase in index flow velocity is due to the backwater effect of the Yangtze River. The increment of the index velocity caused by the head difference between the flow measurement section and the outlet; The increase in flow velocity due to the rise and fall of the Yangtze River water level; The original index flow rate; The calculated correction index flow rate; The water level at the H-ADCP test section; The water level at Yidu Station (located on the main stream of the Yangtze River near the confluence of the Qingjiang River and the Yidu Station); The water level at Zhicheng Station (located 20km downstream of the confluence of the Qingjiang River and the Yangtze River). The baseline water level at Zhicheng Station is the value corresponding to the moment when the backwater of the Yangtze River just appears at the H-ADCP flow measurement section. These are the weighting coefficients; , , The different functional relationships, such as linear, exponential, logarithmic, or exponential, are determined by trial and error.

[0108] According to the rate, The value is 0.396. The value is taken as 38.4 m. The increase in the index velocity caused by the rise and fall of the Yangtze River water level is small and can be ignored. In summary, the formula for calculating the single-valued correction index flow velocity at the Gaobazhou hydrological station is as follows: .

[0109] Calculate the calibration index flow rate for each measurement. , dot painting and Relationship diagram, correlation R 2 It is 0.987. For example... Figure 3 As shown, the data points are densely distributed in a band-like pattern with no significant deviation. The systematic error is -0.5%, and the random uncertainty is 13.8%. The accuracy of each test item meets the requirements for accuracy evaluation of flow measurement results in the "Specifications for River Flow Measurement" (GB50179-2015) and the "Specifications for Hydrological Data Compilation" (SLT247-2020). and The correlation can be obtained by estimating the average flow velocity of the cross section from the H-ADCP correction index flow velocity, and then multiplying the average flow velocity of the cross section by the cross section area to obtain the real-time flow rate. The flow rate calculation model is shown in equation (3).

[0110]

[0111] In the formula: The cross-sectional area of ​​the water passage is m. 2 ; The cross-sectional average flow rate is m 3 / s.

[0112] Edge computing gateway devices are deployed near monitoring points to integrate sensor data acquisition, real-time computing and decision-making functions, and support dynamic switching and collaborative computing of multiple models. The hardware structure of the edge computing gateway devices is implemented through software-defined hardware (SDH) technology to realize data acquisition and protocol parsing, edge computing and real-time response, and adopts hardware-level encryption and redundancy design structure (such as dual power supply modules and multiple communication links (5G+BeiDou+short transmission+fiber) in the hardware interface layer). The hardware layer adopts a modular design, including power supply module, switching module, self-test module and multiple types of interfaces (such as RS485, RS232, SDI, CAN, AI / DI, etc.), supporting various hydrological sensor devices. In addition, the edge computing gateway includes three key circuits: (1) Power management circuit. Dual DC-DC power input (3.3-12V) is adopted and equipped with overvoltage / reverse connection protection circuit. Each hardware interface is equipped with an independent LDO regulator to ensure stable power supply to the sensor. (2) Multi-protocol interface circuit. The RS485 interface uses an ADM2483 isolated transceiver, combined with a TVS diode array (SMBJ12CA) to achieve ±15kV ESD protection. The analog input channel is configured with an ADG5408 multiplexer, and hydrological data acquisition with an accuracy of 0.01% is achieved through an ADS1262 32-bit Σ-Δ ADC. (3) Multi-mode communication circuit. While retaining the basic 5G / 4G communication capabilities, the gateway is equipped with the ability to access heterogeneous networks covering the entire space-ground system and the intelligent selection function of multi-mode fusion by introducing Beidou-3 short message, Bluetooth 5.3 and LoRa spread spectrum technologies, thereby realizing the automatic switching of the optimal communication mode in different scenarios.

[0113] The edge computing gateway's computing power design includes the following features: (1) AI acceleration unit. It integrates a dedicated NPU, supporting 21 TOPS NPU (supporting INT8 / FP16) computing power. (2) General computing unit. It adopts a Rockchip RK3568 quad-core A55 processor, is configured with 8GB LPDDR4x memory, and runs a real-time Linux system. (3) Dynamic power management. It achieves adaptive power adjustment from 0.5 to 12W through an intelligent power allocation algorithm, with power consumption of only 3W in the rainstorm monitoring mode. The advantages of the edge computing power design are shown in Table 3.

[0114] Table 3 Advantages of Edge Computing Design

[0115]

[0116] The flow velocity data collected by H-ADCP is filtered, interpolated, parsed and compiled using an edge computing gateway. A multi-objective optimization algorithm is used to quickly calibrate the parameters of the online flow measurement model in real time, and then the average flow rate of the test section is calculated.

[0117] Finally, a comparative analysis and accuracy evaluation of the online traffic monitoring results based on edge computing gateways were conducted.

[0118] (1) Analysis of results over a specific period;

[0119] Based on the requirements for the accuracy of sequential flow measurement and alignment, the flow data from 8:00 to 24:00 on June 28, 2020, were selected for analysis. During this period, the Gaobazhou Hydropower Station gradually increased its downstream discharge, from 1500 m³ / h. 3 / s rapidly rose to 7700m 3 During this flood discharge period, the flow rate was measured seven times using a cableway current meter method at approximately / s. The H-ADCP flow measurement results from the online flow measurement model based on an edge computing gateway (hereinafter referred to as the new scheme) during the same period were compiled. The flow process curves for water level-flow alignment and the new scheme of this invention are shown below. Figure 4 As shown in the figure, the flow process of the two methods is consistent, but there is a certain deviation. Specifically, the value of the flow calculated by the water level-flow relationship at the same time is larger than that of the new scheme. The characteristic flow deviation statistics within the time period are shown in Table 4. The maximum flow of the two methods is basically the same, the minimum flow deviation is 5.7%, and the average flow deviation within the time period is -4.0%. The reason is that the flow changes rapidly and the measurement frequency is insufficient when measuring the flood peak.

[0120] Table 4 Comparison of Reorganization Results by Time Period

[0121]

[0122] (2) Annual results analysis;

[0123] The flow data of the Gaobazhou hydrological station from May to October 2019, measured using the H-ADCP method, were compiled into a daily hydrological element result table according to the new scheme. This data was compared with the flow measurement results of the H-ADCP online flow measurement model (hereinafter referred to as the original scheme) during the same period, which did not consider the backwater effect of the Yangtze River and the operation of the Gaobazhou hydropower station. The error statistics are shown in Table 5. It can be seen that: 1) the relative error of the average flow in May, June, September, and October is -4.6% to 2.8%, which is a reasonable range; 2) the average flow in July and August is 12.4% and 11.3% higher in the new scheme than in the original scheme, respectively; 3) the relative error of the maximum flow from May to October is -15.9% to 9.9%. The analysis suggests that the frequent scheduling of the Gaobazhou Hydropower Project upstream of the station in July and August led to an increase in surface flow velocity during flood discharge. The new scheme of this invention takes into account factors such as backwater backflow and hydropower station operation to make the correction and estimation of index flow velocity more reasonable. This indicates that the single-valued water level-flow relationship during the dry season is not applicable to the flow propagation during the wet season. The constructed online flow measurement system based on edge computing gateway has a better application effect in river sections and time periods with disordered water level-flow relationship and complex hydrological conditions.

[0124] Table 5 Monthly Results Comparison

[0125]

[0126] The annual runoff and annual average flow of the new scheme and the original scheme are shown in Table 6. When calculating the annual results, the results of the new scheme from May to October are calculated based on the monthly compilation results, while the results of the new scheme from January to April and November to December during the dry season are calculated using the original scheme (i.e., the scheme of single-valued relationship between H-ADCP index velocity and cross-sectional average velocity). The relative error of annual runoff and annual average flow of both schemes is 1.3%, which meets the requirement of no more than 2% in the "Specifications for Compilation of Hydrological Data" (SLT 247-2020).

[0127] Table 6 Comparison of Annual Results

[0128]

[0129] This invention provides an online flow monitoring system based on an edge computing gateway. Compared to the cloud-based centralized computing mode of existing online flow monitoring systems, it solves problems such as high data transmission latency, large network bandwidth consumption, and insufficient real-time response capability at the edge. It can realize local real-time parsing of flow velocity data and cross-sectional flow calculation, effectively reducing dependence on the central station server. It can meet the needs of cross-sectional flow estimation under different water conditions (high, medium, and low water levels), and provide a more convenient method for estimating flow using various non-contact flow measurement methods. It helps to further explore the applicability of online flow measurement equipment in hydrological stations along major rivers and the potential of edge computing, and provides technical support for achieving faster and more efficient online flow monitoring and the construction of digital twin hydrological stations.

[0130] It should be understood that the embodiments described above are only some, not all, of the embodiments of the present invention. Furthermore, the technical features of the various embodiments or individual embodiments provided by the present invention can be arbitrarily combined to form feasible technical solutions. Such combinations are not constrained by the order of steps and / or structural composition patterns, but must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0131] It should be understood that the above description of the preferred embodiments is quite detailed, but it should not be considered as a limitation on the scope of protection of this invention. Those skilled in the art, under the guidance of this invention, can make substitutions or modifications without departing from the scope of protection of the claims of this invention, and all such substitutions or modifications fall within the scope of protection of this invention. The scope of protection of this invention should be determined by the appended claims.

Claims

1. A hydrological flow monitoring system based on an edge computing gateway device, characterized in that: This includes information collection devices, edge computing gateway devices, and central station servers; The information acquisition device includes a water level sensor, a flow rate sensor, a water quality sensor, a temperature and humidity sensor, and a camera. The edge computing gateway device includes a multi-protocol interface data acquisition module, a protocol conversion module, a preprocessing and feature extraction module, a real-time analysis and inference module, and a local storage security encryption module; It is used to periodically acquire data from the information acquisition device and perform real-time processing and analysis, and upload abnormal data and periodic summary reports to the central station server. It also periodically receives lightweight models issued by the central station server to optimize and update the local online flow measurement model, and promptly feeds back data characteristics to the central station server to support the training of the local online flow measurement model. The central station server includes a visual operation and maintenance management subsystem, a big data storage device, a data warehouse, a digital twin module, a local online flow measurement model, and a model distillation engine; After receiving the periodic summary report from the edge computing gateway, the model is used to train a high-precision local online flow measurement model based on global data, compress it into a lightweight model that can adapt to the computing power of the edge computing gateway device through a knowledge distillation engine, and finally send the updated lightweight model to the edge computing gateway device for local traffic calculation, thereby realizing global-local model collaborative optimization. The specific implementation process of compressing the knowledge distillation engine into a lightweight model that can adapt to the computing power of the edge computing gateway device includes: By deploying heterogeneous teacher model groups on the central station server, their knowledge outputs are dynamically integrated; based on the hydrological features reported in real time by edge computing gateway devices, the weights of each teacher model are dynamically allocated to generate joint soft labels. To optimize the student model based on the computing power of edge computing gateway device chips, operator reconstruction, hybrid quantization, and dynamic pruning are employed. Edge computing gateway devices use local private data to fine-tune student models and generate hydrological event probability soft labels; cloud clustering aggregates multi-node soft labels and updates the global teacher model; in response to data drift, incremental online distillation is triggered—the teacher model generates hydrological event probability soft labels in real time, and the student model completes fine-tuning with a small sample within a preset time to quickly adapt to distribution changes. The student model is pushed and deployed on the edge intelligent gateway, and the computation offloading strategy based on reinforcement learning is used to achieve multi-objective optimization of latency, accuracy and energy consumption. The process involves deploying a heterogeneous group of teacher models on a central server and dynamically integrating their knowledge outputs. The specific implementation process includes multi-model collaborative decision-making, dynamic weighting based on environmental perception, intelligent conflict resolution, generation of fusion instruction packages, and dynamic execution at the edge. The multi-model collaborative decision-making involves deploying four types of professional models on the central station server: a physical mechanism model that calculates flow based on hydrophysical laws, an AI time series model that learns water level change patterns from historical data, a visual recognition model that analyzes water level scale images captured by cameras, and a topology analysis model that combines data from other monitoring points within the basin for correlation prediction; each type of model independently outputs flow prediction results and confidence scores. The dynamic weighting of environmental perception is achieved by constructing an intelligent weight allocator through real-time collection of environmental parameters; when equipment malfunctions, the weight of the corresponding sensor association model is reduced; and the weights are dynamically refreshed to ensure the optimal combination. The intelligent conflict resolution mechanism prioritizes the prediction results that conform to physical rules when the difference between the predictions of multiple models exceeds a threshold, compares the performance records of models in similar historical scenarios, and finally makes dynamic arbitration based on the recent accuracy of each model. The fusion instruction packet generation generates a transmission packet containing three types of core information: the final predicted value of the weighted average result, the confidence interval reflecting the reliability of the prediction, and the key feature indication marking the abnormal state; The dynamic execution at the edge directly guides the student model to update parameters after the edge computing gateway device receives the fusion instruction packet.

2. The hydrological flow monitoring system based on an edge computing gateway device according to claim 1, characterized in that: Based on the hydrological features reported in real time by edge computing gateway devices, the attention mechanism is used to dynamically allocate the weights of each teacher model and generate joint soft labels; the specific implementation includes understanding model features, generating state feature vectors, calculating matching degree, dynamically allocating weights and fusion prediction; Before system deployment, the understanding model features are defined by setting a unique capability feature vector for each local online flow measurement model participating in the prediction. This vector implicitly represents the prediction scenarios it is good at, which include different hydrological conditions at medium, high, and low water levels. The generated state feature vector is generated when the edge computing gateway device reports the current hydrological data in real time. The system immediately inputs these raw data into a lightweight model, quickly analyzes and extracts the core features of the current data, and generates a state feature vector. The matching degree calculation involves comparing the current hydrological status feature vector with the pre-set capability feature vector of each local online flow measurement model, calculating the similarity between the two, and generating a matching degree score for each local online flow measurement model. The higher the score, the more the current hydrological status matches the model's preferred prediction scenario. The dynamic weight allocation involves the system collecting the "matching scores" of all local online flow measurement models, converting these scores, and ensuring that the sum is 1. In the fusion prediction, each local online flow measurement model independently generates a prediction result based on the current real-time hydrological data, which serves as a node soft label. The system multiplies the prediction result of each local online flow measurement model by the dynamic weight value obtained in the previous step, and adds up the weighted prediction results of all models to obtain the final "joint soft label".

3. The hydrological flow monitoring system based on an edge computing gateway device according to claim 1, characterized in that: The cloud uses KL divergence clustering to aggregate multi-node soft labels and update the global teacher model; The specific implementation process is as follows: After collecting node soft tags in the cloud, dynamic clustering is performed based on pairwise KL divergence: predictions with differences less than the threshold are automatically clustered into a cluster to reflect similar hydrological patterns. Integrating soft labels within clusters to generate high-quality "cluster centers"—using a weighted average based on confidence or intra-cluster KL consistency to strengthen reliable and consistent prediction patterns; when updating the teacher model, each cluster center is treated as an independent "teacher," calculating its KL divergence loss with the global model prediction, and aggregating these losses by cluster size or confidence; updating the model by minimizing the weighted loss enables teachers to integrate diverse, high-confidence collective knowledge represented by different clusters, rather than a simple mixture of all nodes, thus improving robustness.

4. The hydrological flow monitoring system based on an edge computing gateway device according to claim 1, characterized in that: Use TensorRT-LLM to push and deploy student models on edge computing gateway devices.

5. The hydrological flow monitoring system based on an edge computing gateway device according to claim 1, characterized in that: The digital twin module is responsible for integrating multi-source data, constructing a virtual mapping of the watershed, supporting dynamic simulation and intelligent decision-making, and realizing an upgrade from edge perception to watershed-level regulation; This includes data fusion and dynamic modeling, simulation engines and decision-making loops, and visualization and collaborative management; The data fusion and dynamic modeling integrates pre-processed real-time hydrological data and meteorological and geographic information from edge computing gateway devices to construct a high-precision three-dimensional watershed model, thereby achieving real-time mapping between the physical world and virtual space. The simulation engine and decision-making closed loop rely on the hydrodynamic model to pre-simulate the propagation path and inundation range of the flood peak and generate scheduling strategies; through reinforcement learning algorithms, the demand for flood discharge and water storage is dynamically balanced to form a "monitoring-simulation-decision-execution" closed loop, and the instructions are sent to the edge computing gateway control device in real time. The visualization and collaborative management features a 3D "monitoring station map" that overlays real-time data and risk heat maps; it supports cross-departmental command collaboration and pushes work orders via mobile devices to enable rapid handling of emergencies.

6. The hydrological flow monitoring system based on an edge computing gateway device according to claim 1, characterized in that, The local online flow measurement model is as follows: ； In the formula: The output of the model is the average flow rate at the measurement section. The independent variables of the model include the flow velocities represented by the points or lines monitored by the online flow measurement equipment; For the model parameter set, refer to the coefficients representing the flow velocity; The model operation rules include linear functions, polynomial functions, exponential functions, logarithmic functions, and power functions. The specific function type is determined after the model parameters are calibrated using a genetic algorithm based on the sample data.

7. The hydrological flow monitoring system based on an edge computing gateway device according to any one of claims 1-6, characterized in that: The system adopts a hardware-level encryption and redundancy design structure, including a dual power supply module for the hardware interface layer and multiple communication links, to ensure data privacy and the security and reliability of device operation.

8. A hydrological flow monitoring software product based on an edge computing gateway device, comprising computer program instructions, characterized in that: When the computer program instructions are executed on a computer, they drive the system according to any one of claims 1-7 to perform hydrological flow monitoring.

Images