River monitoring system based on Internet of Things

Abstract

The invention relates to the technical field of river channel environment monitoring, and discloses a river channel monitoring system based on the Internet of Things. The system comprises a multi-mode sensing module, an environment feature analysis module, a space-time coupling modeling module, a self-adaptive resource scheduling module and an intelligent cooperative control module. The multi-modal sensing module collects multi-source heterogeneous data, the environment feature analysis module extracts various feature vectors, the space-time coupling modeling module generates a coupling space-time feature matrix, and the self-adaptive resource scheduling module constructs a double-layer optimization library and stores related rule strategies. And the intelligent cooperative control module generates a real-time pollution early warning instruction and a hydrological regulation and control decision through a multi-target reinforcement learning framework based on the results. In addition, the system can carry out anomaly detection and emergency treatment on the floating objects and predict sudden change of water quality. According to the system, comprehensive monitoring, intelligent analysis and accurate regulation and control of the river environment are realized, and the efficiency and scientificity of river management are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of river environment monitoring, and particularly to an Internet of Things-based river monitoring system. Background Art

[0002] In today's society, the protection and rational utilization of water resources are of crucial importance. As an important carrier of water resources, the environmental conditions of rivers are directly related to ecological balance, the safety of residents' water use, and the sustainable development of surrounding areas. However, traditional river monitoring methods have many limitations and are difficult to meet the requirements of modern society for refined river management.

[0003] Early river monitoring mainly relied on manual regular inspections. Staff needed to collect water samples and measure hydrological parameters on-site. This method was not only inefficient but also had limited monitoring frequencies and could not obtain river environment information in real time. Once a sudden pollution event or hydrological anomaly occurred, it was difficult to detect and handle in a timely manner, which might lead to consequences such as pollution spread and ecological damage.

[0004] With the development of sensor technology, some simple sensors have been deployed in some rivers for data collection, such as single-point water quality sensors and flow velocity sensors. However, these sensors are often independent of each other, and the data cannot be effectively integrated, forming individual "data islands". Different types of data lack a unified analysis and processing mechanism, making it difficult to grasp the changing trends of the river environment as a whole. Moreover, the monitoring range of a single type of sensor is limited and cannot comprehensively reflect the complex conditions of the river. For example, it cannot monitor key information such as the distribution of underwater obstacles and the movement trajectories of floating objects.

[0005] In terms of data processing and analysis, traditional methods mostly adopt simple statistical analysis means and have insufficient processing capabilities for multi-source heterogeneous data. Facing a large amount of water quality, hydrology, meteorology and other data, it is difficult to extract valuable features and laws from them and cannot provide a scientific and accurate basis for decision-making. For example, when analyzing the impact of rainfall on water quality, the spatio-temporal distribution of rainfall and its lag effect on water quality cannot be considered, resulting in a lack of pertinence when formulating pollution prevention and control measures.

[0006] In terms of resource management and scheduling, traditional monitoring systems lack intelligent means. The allocation of computing resources is unreasonable, with some node resources being idle while some nodes experience data processing delays due to overloading. At the same time, the energy consumption of equipment is relatively high, lacking an effective energy consumption balancing strategy, which increases the operating cost. Moreover, when facing different monitoring tasks, resources cannot be dynamically adjusted according to the priority and actual needs of the tasks, reducing the overall operating efficiency of the system.

[0007] Furthermore, existing river monitoring systems also have shortcomings in decision-making and coordinated control. The system lacks a unified decision-making framework, making it impossible to comprehensively consider multiple factors and generate optimal regulatory decisions. When pollution incidents or hydrological anomalies occur, communication between departments is poor, making collaborative work difficult and making it difficult to quickly implement effective response measures. This significantly reduces the timeliness and effectiveness of problem resolution. Summary of the Invention

[0008] The purpose of the present invention is to provide a river monitoring system based on the Internet of Things to solve the problems raised in the above background technology.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a river monitoring system based on the Internet of Things, the system comprising: Multimodal perception module: used to collect multi-source heterogeneous data of river environment in real time; Environmental feature analysis module: performs cross-modal feature extraction on the multi-source heterogeneous data to generate feature vectors for each data source, including water pollution diffusion characteristics, flow velocity gradient distribution characteristics, rainfall spatiotemporal correlation characteristics, underwater obstacle contour characteristics, and floating object motion trajectory patterns; Spatiotemporal coupling modeling module: constructs a multi-dimensional feature fusion space based on the spatiotemporal graph attention network, maps the feature vectors to a unified spatiotemporal grid coordinate system, and generates a coupled spatiotemporal feature matrix; Adaptive resource scheduling module: Builds a two-layer optimization library of computing resource pool and river status assessment model, which stores dynamic task allocation rules, edge node energy consumption balancing strategy and sensor-buoy equipment topology relationship map; Intelligent collaborative control module: Based on the coupled spatiotemporal feature matrix and the adaptive resource scheduling module, it generates real-time pollution warning instructions and hydrological regulation decisions through a multi-objective reinforcement learning framework.

[0010] Preferably, the multi-source heterogeneous data include water quality sensor parameters, hydrological flow rate monitoring values, meteorological station precipitation data, underwater sonar images and floating object visual detection streams; The cross-modal feature extraction of the multi-source heterogeneous data includes: The water quality sensor parameters are subjected to wavelet transform to decompose multi-band pollution components, and the pollution source diffusion pattern is extracted through time series clustering algorithm; The underwater sonar image is segmented using a residual attention network to segment the underwater obstacle contours, and the obstacle displacement trajectory is predicted in combination with the optical flow method; A multi-scale convolutional network is used to extract the motion direction features of the floating object visual detection flow, and the floating object aggregation trend is matched based on the dynamic time warping algorithm.

[0011] Preferably, the cross-modal feature extraction further includes: Model the river channel terrain dependence relationship for the hydrological flow velocity monitoring values using a graph embedding algorithm, and extract the spatio-temporal correlation features of the flow velocity mutation area through a gradient boosting tree; Model the lagging impact of rainfall events on water quality for the precipitation data of the weather station using a causal convolutional network, and infer the pollution risk propagation path based on a Bayesian network.

[0012] Preferably, the construction of the double-layer optimization library of the computing resource pool and the river channel state evaluation model includes: Construct a distributed computing resource pool using a transfer learning framework based on a dynamic task offloading protocol and an edge node deployment topology; Construct a river channel state evaluation model using a spatio-temporal graph autoencoder according to historical hydrological data and real-time water quality parameter distributions.

[0013] Preferably, the multi-objective reinforcement learning framework adopts a hierarchical decision-making architecture, including: Define the state space as the joint encoding of the coupled spatio-temporal feature matrix and the remaining computing power of the resource pool, and the action space as the decision combination of the monitoring task priority assignment and the data sampling frequency; Balance the conflicts of each decision-making objective through a proximal policy optimization algorithm, and design a multi-dimensional reward function based on the energy consumption constraint.

[0014] Preferably, the multi-objective reinforcement learning framework further includes: Adopt a distributed soft actor-critic algorithm to realize multi-node asynchronous policy updates, and optimize the global convergence speed in combination with the priority experience replay technology; Convert the device energy consumption limit into a dynamic penalty term of the policy network through a constraint relaxation method.

[0015] Preferably, the system further includes: Perform abnormal aggregation detection on the floating object movement trajectory pattern, and use a spectral clustering algorithm to identify the floating object distribution area exceeding the preset threshold; When an abnormal aggregation is detected, trigger the emergency response mechanism of the intelligent collaborative control module, and generate an interception device deployment instruction and a collaborative positioning signal for adjacent buoy devices.

[0016] Preferably, the emergency response mechanism models the interaction of floating objects using a bi-directional graph neural network, including: Extract the floating object migration trend features from the forward propagation path, and extract the interception device influence range features from the reverse propagation path; Generate an optimal deployment plan through an attention mechanism fusion, and update the dynamic risk node weights of the river channel state evaluation model.

[0017] Preferably, the system further includes: A water quality mutation prediction model is constructed based on historical pollution event data and real-time meteorological parameters. A gated recurrent unit network is used to integrate the environmental attenuation factor and dynamically adjust the sampling frequency threshold of the monitoring task.

[0018] Preferably, the water quality mutation prediction model further includes: Based on the variational autoencoder, the uncertainty parameters of multiple sensors are fused to generate a water quality mutation confidence index and embed it into the real-time pollution warning instruction.

[0019] Compared with the prior art, the present invention has the following beneficial effects: In terms of data collection and analysis, the multimodal perception module collects heterogeneous data from multiple sources in real time, including water quality sensor parameters, hydrological flow rate monitoring values, meteorological station precipitation data, underwater sonar images, and visual detection of floating objects, providing a comprehensive and timely picture of the river's environmental status. The environmental feature analysis module utilizes advanced technologies such as wavelet transforms, residual attention networks, and graph embedding algorithms to extract cross-modal features from this data, enabling in-depth analysis of the underlying information. For example, by applying wavelet transforms to water quality sensor parameters to decompose multi-band pollution components and combining them with time series clustering algorithms to extract pollution source diffusion patterns, this module helps accurately locate pollution sources and provides key clues for pollution control. A residual attention network is used to segment obstacle outlines in underwater sonar images and, combined with optical flow methods, predict obstacle displacement trajectories, providing early warning of navigation risks and ensuring river navigation safety. Furthermore, a causal convolutional network is used to model the lagged impact of rainfall events on water quality using meteorological station precipitation data. A Bayesian network is then used to infer the propagation path of pollution risks, enabling managers to proactively identify potential pollution risks and implement targeted prevention and control measures to prevent incidents from escalating.

[0020] The spatiotemporal coupling modeling module constructs a multidimensional feature fusion space based on the spatiotemporal graph attention network, mapping various feature vectors to a unified spatiotemporal grid coordinate system to generate a coupled spatiotemporal feature matrix. This enables the system to comprehensively analyze changes in the river environment from a spatiotemporal perspective, accurately grasp environmental evolution trends, and provide a scientific basis for subsequent decision-making. For example, when analyzing the spread of water pollution, it not only understands the spatial distribution of the pollution, but also its spread speed and direction over time, providing strong support for the development of reasonable pollution control plans.

[0021] The establishment of the adaptive resource scheduling module optimizes the system resource management. Based on the dynamic task offloading protocol and the edge node deployment topology, a distributed computing resource pool is constructed using a transfer learning framework, which can dynamically allocate computing resources according to the load conditions and task requirements of each node, improve resource utilization, and reduce task processing latency. At the same time, according to the historical hydrological data and the distribution of real-time water quality parameters, a spatio-temporal graph autoencoder is used to generate the node latent representation of the river channel state evaluation model, realizing the accurate evaluation of the river channel state. On this basis, the stored dynamic task allocation rules, the edge node energy consumption balancing strategy, and the sensor-buoy device topology relationship map can effectively reduce the device energy consumption, extend the device service life, and reduce the system operation cost. For example, when the monitoring tasks in a certain area are heavy, the system automatically offloads some tasks to idle nodes for processing to ensure the efficient completion of tasks; according to the energy consumption balancing strategy, the working states of each node are reasonably adjusted to avoid excessive energy consumption of some nodes.

[0022] The intelligent collaborative control module is based on the coupled spatio-temporal feature matrix and the adaptive resource scheduling module, and generates real-time pollution warning instructions and hydrological regulation decisions through a multi-objective reinforcement learning framework. The multi-objective reinforcement learning framework adopts a hierarchical decision-making architecture, defines a reasonable state space and action space, and balances the conflicts of each decision-making objective through the proximal policy optimization algorithm. A multi-dimensional reward function is designed based on the energy consumption constraint to make the decision-making more scientific and reasonable. For example, in the face of a pollution incident, the system can comprehensively consider factors such as the pollution degree, propagation range, treatment cost, and energy consumption, quickly generate the best pollution warning instructions and hydrological regulation decisions, and timely control the pollution diffusion to ensure the ecological safety of the river channel. At the same time, the distributed soft actor-critic algorithm is used to realize the asynchronous policy update of multiple nodes, the global convergence speed is optimized by combining the priority experience replay technology, and the device energy consumption limit is transformed into a dynamic penalty term of the policy network through the constraint relaxation method, further improving the accuracy and efficiency of the system decision-making.

[0023] In addition, the system also has the functions of detecting abnormal aggregation of floating objects and emergency response, as well as predicting water quality mutations. The spectral clustering algorithm is used to identify the abnormal aggregation areas of floating objects, trigger the emergency response mechanism, and generate the optimal deployment plan of the interception device using a bi-directional graph neural network, and update the dynamic risk node weights of the river channel state evaluation model, which can timely clean up the floating objects and maintain the ecological environment and beauty of the river channel. The water quality mutation prediction model based on the gated recurrent unit network integrates the environmental attenuation factor, dynamically adjusts the sampling frequency threshold of the monitoring task, combines the variational autoencoder to generate the water quality mutation confidence index and embeds it into the real-time pollution warning instructions, which can early warn the water quality mutation risk and provide a strong guarantee for ensuring the safety of residents' water use. Description of the Drawings

[0024] Figure 1 It is the working principle diagram of the river channel monitoring system based on the Internet of Things described in the present invention; Figure 2 Flow chart of multi-source data processing logic for cross-modal feature extraction; Figure 3 Step diagram for linkage control of abnormal aggregation detection and emergency response of floating objects; Figure 4 Step diagram for generating an interception deployment plan of a bidirectional graph neural network. Specific implementation manners

[0025] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0026] Please refer to Figures 1-4 , the present invention provides an Internet of Things-based river monitoring system, and its overall implementation solution is as follows: Multi-modal perception module: This module is responsible for collecting multi-source heterogeneous data of the river environment in real time, which is the basis for the operation of the entire system. By reasonably deploying various sensors and monitoring devices at different positions of the river, rich and diverse data information can be obtained, such as water quality sensor parameters, hydrological flow velocity monitoring values, precipitation data of meteorological stations, underwater sonar images, and visual detection streams of floating objects. These data reflect the state of the river environment from different angles and provide strong support for subsequent analysis and decision-making.

[0027] Environmental feature analysis module: This module performs cross-modal feature extraction on the multi-source heterogeneous data collected by the multi-modal perception module. For different types of data, specific technical methods are used to generate feature vectors of each data source. Specifically, it includes water pollution diffusion characteristics, flow velocity gradient distribution characteristics, rainfall spatio-temporal correlation characteristics, underwater obstacle contour characteristics, and floating object movement trajectory patterns. By extracting these features, the information hidden behind the data can be deeply mined, providing a more valuable basis for subsequent modeling and decision-making.

[0028] Spatio-temporal coupling modeling module: Based on the spatio-temporal graph attention network, this module constructs a multi-dimensional feature fusion space. The feature vectors generated by the environmental feature analysis module are mapped to a unified spatio-temporal grid coordinate system to generate a coupled spatio-temporal feature matrix. This way of fusing different features in the spatio-temporal dimension can fully consider the spatio-temporal correlation between data and more accurately reflect the dynamic changes of the river environment.

[0029] Adaptive Resource Scheduling Module: Build a two-layer optimization library for the computing resource pool and the river channel state evaluation model. In this two-layer optimization library, dynamic task allocation rules, edge node energy consumption balancing strategies, and the topology relationship map of sensor-buoy devices are stored. Through these rules, strategies, and maps, reasonable allocation and scheduling of computing resources can be achieved, and at the same time, the river channel state can be accurately evaluated, providing guarantee for the efficient operation of the system.

[0030] Intelligent Cooperative Control Module: Based on the coupled spatio-temporal feature matrix and the adaptive resource scheduling module, this module generates real-time pollution warning instructions and hydrological regulation decisions through a multi-objective reinforcement learning framework. The multi-objective reinforcement learning framework can comprehensively consider multiple decision-making objectives, such as the accuracy of monitoring tasks, the reasonable utilization of resources, the control of energy consumption, etc., so as to generate optimal decision-making instructions and realize the intelligent regulation of the river channel environment.

[0031] The following further illustrates the implementation of the present invention in combination with Embodiments 1 to 6.

[0032] Embodiment 1: In this embodiment, the process of collecting multi-source heterogeneous data and cross-modal feature extraction is further elaborated. The multi-modal perception module collects multi-source heterogeneous data through a variety of devices. Water quality sensors are arranged at different depths and positions in the river channel to obtain water quality sensor parameters, which include parameters such as acidity and alkalinity (pH value), dissolved oxygen content, chemical oxygen demand (COD), etc. The hydrological flow velocity monitoring device uses a Doppler flowmeter, which is installed at the bottom or on the bank of the river channel to measure the hydrological flow velocity monitoring value. The precipitation data of the meteorological station is obtained in real time through the data interface with the meteorological department, including information such as rainfall amount and rainfall time. The underwater sonar device uses the principle of acoustic wave reflection to obtain underwater sonar images, which can clearly present the shape and position of underwater obstacles. The floating object visual detection device uses a high-definition camera and cooperates with image recognition technology to obtain the floating object visual detection stream.

[0033] For cross-modal feature extraction, for water quality sensor parameters, wavelet transform is used to decompose multi-band pollution components. Wavelet transform is a time-frequency analysis method, which can decompose a signal into sub-signals of different frequencies. Assume that the water quality sensor parameter signal is , and after wavelet transform, components of different frequency bands ( represents different frequency bands) are obtained. These components are processed through a time series clustering algorithm to extract the pollution source diffusion pattern. The time series clustering algorithm can group similar time series data into one category, so as to find out the law of pollution source diffusion.

[0034] For underwater sonar images, a residual attention network is used for underwater obstacle contour segmentation. By introducing a residual structure and an attention mechanism, the residual attention network can more accurately identify target objects in images. The underwater sonar image is input into the residual attention network, and the network outputs the contour information of the underwater obstacle. Combining with the optical flow method to predict the displacement trajectory of the obstacle, the optical flow method is a method for calculating the object motion based on the brightness change of pixel points in the image. Assume that in two adjacent frames of images, the optical flow vector of the pixel point is . By calculating the optical flow vector, the position change of the obstacle at different times can be obtained, and then its displacement trajectory can be predicted.

[0035] For the visual detection flow of floating objects, a multi-scale convolutional network is used to extract the motion direction features. Through convolutional kernels of different sizes, the multi-scale convolutional network can capture features of different scales in the image. The visual detection flow of floating objects is input into the multi-scale convolutional network, and the network can extract the motion direction information of the floating objects. Based on the dynamic time warping algorithm to match the aggregation trend of floating objects, the dynamic time warping algorithm can measure the similarity of two time series. Assume that there are two time series of the motion trajectories of floating objects and . By calculating the distance between them using the dynamic time warping algorithm, the smaller the distance, the more similar the two time series are, and thus the aggregation trend of floating objects can be judged.

[0036] For the hydrological flow velocity monitoring values, a graph embedding algorithm is used to model the dependence relationship of the river channel terrain. The graph embedding algorithm can map graph-structured data into a low-dimensional vector space for subsequent analysis and processing. Assume that the river channel terrain can be represented as a graph , where is the set of nodes, representing different positions in the river channel, and is the set of edges, representing the connection relationship between nodes. The graph is mapped into a low-dimensional vector space through the graph embedding algorithm to obtain the vector representation of the nodes. The spatio-temporal correlation features of the flow velocity mutation region are extracted by the gradient boosting tree, and the gradient boosting tree is an ensemble learning algorithm based on decision trees. Taking the hydrological flow velocity monitoring values as the input, the gradient boosting tree can automatically learn the spatio-temporal correlation features of the flow velocity mutation region.

[0037] For the precipitation data of the meteorological station, a causal convolutional network is used to model the lag effect of rainfall events on water quality. The causal convolutional network is a special convolutional neural network that can process time series data with causal relationships. Assume that the precipitation data is , and the water quality data is , the causal convolutional network can learn the lagging impact relationship of precipitation events on water quality. Based on the Bayesian network to infer the pollution risk propagation path, the Bayesian network is a graphical model based on probabilistic inference. According to the historical precipitation data and water quality data, a Bayesian network is constructed, and the pollution risk propagation path can be obtained through network inference.

[0038] Embodiment 2: This embodiment details the construction process of the computing resource pool and the river channel state evaluation model in the adaptive resource scheduling module.

[0039] Based on the dynamic task offloading protocol and the edge node deployment topology, a distributed computing resource pool is constructed using the transfer learning framework. The dynamic task offloading protocol decides whether to offload a task to other nodes for processing according to the computing power, load condition, and task urgency of the edge nodes. Assume there are multiple edge nodes , and the computing power of each edge node is , the current load is , and the task has a computing requirement of . When the of a certain edge node , according to the dynamic task offloading protocol, the task is offloaded to other nodes with sufficient computing power and lower load.

[0040] The transfer learning framework then utilizes the existing knowledge and models to quickly adapt to new tasks and data. When constructing the distributed computing resource pool, some basic models are first trained on some edge nodes, and these models can be models for common river channel monitoring tasks, such as water quality prediction models, flow velocity analysis models, etc. When a new task arrives, through the transfer learning framework, these basic models are fine-tuned to adapt to the requirements of the new task, thereby improving the computing efficiency.

[0041] According to the historical hydrological data and the real-time water quality parameter distribution, a spatio-temporal graph autoencoder is used to generate the node latent representation of the river channel state evaluation model. The historical hydrological data includes information such as water level, flow rate, and flow velocity over the years, and the real-time water quality parameter distribution is the water quality data at different positions of the river channel at the current moment. The spatio-temporal graph autoencoder is a neural network model that can process spatial and temporal information simultaneously. Assume the river channel can be represented as a spatio-temporal graph , where is the node set, representing different positions in the river channel, is the edge set, representing the connection relationship between nodes, is the time dimension. Taking historical hydrological data and real-time water quality parameters as inputs, the spatio-temporal graph autoencoder generates the latent representations of nodes through the encoding and decoding processes. These latent representations contain the key information of the river channel state and can be used to evaluate the current state of the river channel, such as whether the water quality has deteriorated, whether the water flow is abnormal, etc. By continuously updating the historical data and real-time data, the spatio-temporal graph autoencoder can dynamically adjust the latent representations of nodes and improve the accuracy of river channel state assessment.

[0042] Embodiment 3: This embodiment introduces the hierarchical decision-making architecture of the multi-objective reinforcement learning framework. The multi-objective reinforcement learning framework adopts a hierarchical decision-making architecture and defines the state space as the joint encoding of the coupled spatio-temporal feature matrix and the remaining computing power of the resource pool. The coupled spatio-temporal feature matrix is generated by the spatio-temporal coupling modeling module, which contains the fusion features of multi-source heterogeneous data of the river channel environment in the spatio-temporal dimension. Suppose the coupled spatio-temporal feature matrix is , and the remaining computing power of the resource pool is , then the state space can be expressed as . This joint encoding method can comprehensively reflect the current state of the system and provide an accurate basis for decision-making.

[0043] The action space is the decision combination of the monitoring task priority assignment and the data sampling frequency. The monitoring task priority assignment determines which monitoring tasks need to be executed first. For example, when water quality anomalies are detected, the priority of tasks related to water quality monitoring should be increased. Suppose there are multiple monitoring tasks , and the priority of each task can be represented by a numerical value , the larger the , the higher the priority of task . The decision of the data sampling frequency determines the frequency at which the sensor collects data. For example, in areas where the water quality changes rapidly, the sampling frequency can be appropriately increased. Suppose the value range of the data sampling frequency is . The appropriate sampling frequency

[0044] can be selected according to the actual situation.

[0045] Design a multi - dimensional reward function based on energy consumption constraints. The reward function is used to guide the decisions of the agent and make it develop in the optimal direction. Considering the energy consumption constraints, the reward function should include terms related to energy consumption. Assume the energy consumption of the device is , set an energy consumption threshold . When , give a certain reward; when , give a penalty. At the same time, the reward function should also consider factors such as the completion of monitoring tasks and the accuracy of data. For example, if the monitoring task can be accurately completed and the collected data has high accuracy, a higher reward is given. By designing such a multi - dimensional reward function, the agent will comprehensively consider various factors when making decisions and achieve the optimal operation of the system.

[0046] Example 4: This example further elaborates on the implementation methods of multi - node asynchronous policy update and energy consumption limit conversion in the multi - objective reinforcement learning framework.

[0047] Use the distributed soft actor - critic algorithm to achieve multi - node asynchronous policy update. In a distributed system, there are multiple edge nodes, and each edge node can independently perform policy updates. The distributed soft actor - critic algorithm realizes multi - node asynchronous policy update by sharing experiences and parameters among different nodes. Assume there are edge nodes , and each node has its own policy network and value network ( ). Each node samples and learns in its local environment, and then uploads the experience data to the shared experience pool. Other nodes can obtain the experience data from the experience pool and update their own policy networks and value networks. In this way, each node can update its policy asynchronously without interfering with each other, improving the learning efficiency.

[0048] Combine the priority experience replay technique to optimize the global convergence speed. The priority experience replay technique samples experience data according to the importance of the experience data. In the river monitoring system, some experience data is more important for learning the optimal policy, such as the experience data when water quality mutation or flow velocity anomaly is detected. By assigning higher priorities to these important experience data and using these data more frequently during training, the convergence speed of the policy network can be accelerated and the performance of the system can be improved. Assume the priority of the experience data is , during sampling, select experience data from the experience pool for learning according to the priority .

[0049] The device energy consumption limit is converted into a dynamic penalty term for the policy network through the constraint relaxation method. The device energy consumption limit is an important constraint condition in system operation. To incorporate the energy consumption limit into the learning process of the policy network, the constraint relaxation method is adopted. Assume the device energy consumption limit is and, through the constraint relaxation method, the energy consumption limit is converted into a dynamic penalty term . When the device energy consumption is close to or exceeds the energy consumption limit , the value of the penalty term will increase, thus affecting the output of the policy network and making the agent avoid excessive energy consumption as much as possible when making decisions. For example, the penalty term can be expressed as , where is a parameter for adjusting the penalty strength. In this way, the policy network will automatically consider the energy consumption limit during the learning process and achieve reasonable utilization of energy.

[0050] Example 5: This example details the implementation process of the abnormal aggregation detection and emergency response mechanism for the floating object movement trajectory pattern.

[0051] For the abnormal aggregation detection of the floating object movement trajectory pattern, the spectral clustering algorithm is used to identify the distribution area of floating objects that exceeds the preset threshold. The spectral clustering algorithm is a clustering algorithm based on graph theory. It realizes data clustering by performing eigen-decomposition on the similarity matrix between data points. In this system, the position information of floating objects is regarded as nodes in the graph, and the distance between nodes is used as the weight of the edge to construct the floating object position graph. Assume the position set of floating objects is , and by calculating the similarity matrix between nodes, where represents the similarity between floating object and floating object . For the similarity matrix Perform eigen - decomposition to obtain eigen - vectors and eigenvalues. According to the eigenvalues and a preset threshold, select appropriate eigen - vectors for clustering, thereby identifying the distribution area of floating objects that exceed the preset threshold. When an abnormal aggregation is detected, trigger the emergency response mechanism of the intelligent collaborative control module to generate an interception device deployment instruction and a collaborative positioning signal for adjacent buoy devices. The emergency response mechanism uses a bidirectional graph neural network to model the interactions of floating objects. The bidirectional graph neural network can propagate information both forward and backward, capturing the mutual relationships between floating objects more comprehensively. Extract the migration trend features of floating objects from the forward propagation path. By analyzing the movement trajectories and speed information of floating objects, obtain features such as the migration direction and speed magnitude of floating objects. Extract the influence range features of the interception device from the backward propagation path. Considering factors such as the position, size, and function of the interception device, determine the range of floating objects that the interception device can affect.

[0052] Generate an optimal deployment plan through attention mechanism fusion. The attention mechanism can weight - fuse features according to the importance of different features. In this system, take the migration trend features of floating objects and the influence range features of the interception device as inputs, calculate the weights of each feature through the attention mechanism, and then perform weighted fusion to obtain the optimal interception device deployment plan. For example, for a floating object aggregation area near the riverbank with a slow migration speed, small - sized interception devices can be preferentially deployed; for a floating object aggregation area in the center of the river channel with a fast migration speed, large - sized and efficient interception devices are required. At the same time, according to the deployment plan, generate a collaborative positioning signal for adjacent buoy devices, enabling adjacent buoy devices to cooperate with the interception device to achieve effective interception of floating objects. During the implementation process, continuously adjust the deployment plan and collaborative positioning signal according to the actual situation to ensure the effectiveness of the emergency response.

[0053] Update the dynamic risk node weights of the river channel state assessment model. When an abnormal aggregation of floating objects is detected and processed, it is necessary to update the dynamic risk node weights of the river channel state assessment model. Because the abnormal aggregation of floating objects may have an impact on the river channel ecological environment and water quality, etc., the corresponding risk node weights should be adjusted. Suppose a certain risk node in the river channel state assessment model is related to the aggregation of floating objects. According to the degree and influence range of the abnormal aggregation of floating objects, increase or decrease the weight of this risk node. If the floating object aggregation area is large and may cause serious pollution to the water quality, increase the weight of the risk node ; if the floating object aggregation is timely controlled and has little impact on the river channel environment, appropriately reduce the weight of the risk node . By continuously updating the dynamic risk node weights, the river channel state assessment model can more accurately reflect the actual risk situation of the river channel.

[0054] Example 6: This embodiment elaborates in detail the construction and application process of the water quality mutation prediction model. When constructing the water quality mutation prediction model, historical pollution event data is first collected and sorted out. These data record in detail the time and location of past pollution events, the numerical changes of various water quality indicators at that time, such as acidity and alkalinity (pH value), dissolved oxygen (DO), chemical oxygen demand (COD), ammonia nitrogen (NH3-N), etc., as well as information such as the type and source of pollutants. At the same time, meteorological parameters are obtained in real time, including air temperature, air pressure, humidity, wind speed, wind direction, precipitation, etc. These meteorological factors have varying degrees of influence on the water quality of the river course. For example, heavy rainfall may wash surface pollutants into the river course and change the water quality situation.

[0055] The gated recurrent unit network (GRU) is used to process the data. GRU can effectively capture the long-term dependence relationships in time series data and is very suitable for analyzing the changing patterns of historical pollution event data and real-time meteorological parameters over time. Assume that the time series composed of historical pollution event data is , where represents the data vector related to historical pollution events at the -th time step; the time series of real-time meteorological parameters is , represents the real-time meteorological parameter vector at the -th time step. The environmental attenuation factor is denoted as , which comprehensively considers factors such as the water flow velocity of the river course, the water body's self-purification ability, and light intensity, and is used to reflect the natural degradation or dilution effect of the environment on pollutants. When the water flow velocity is fast, the water body's self-purification ability is strong, and the light is sufficient, takes a larger value, meaning that the pollutants decay faster; otherwise takes a smaller value.

[0056] Taking , and as the inputs of the gated recurrent unit network, the network learns the complex relationships between the data through its own structure and training process, and outputs the prediction results of water quality mutation. The prediction results include whether the water quality will mutate in a certain future time period, as well as the approximate time and degree when the mutation may occur.

[0057] Dynamically adjust the sampling frequency threshold of the monitoring task according to the prediction results. If it is predicted that the water quality may mutate, in order to more timely and accurately grasp the changes in water quality, reduce the sampling frequency threshold, and increase the sampling frequency of the monitoring task. For example, originally, a water sample was collected every 2 hours for water quality detection. When it is predicted that there is a risk of water quality mutation, the sampling frequency is increased to once every 30 minutes. On the contrary, if the water quality is predicted to be relatively stable, appropriately increase the sampling frequency threshold, reduce unnecessary sampling operations, which can not only reduce the loss of equipment such as sensors, but also relieve the pressure of data processing. For example, adjust the sampling frequency from once an hour to once every 3 hours.

[0058] The water quality mutation prediction model also uses a variational autoencoder to fuse the uncertainty parameters of multiple sensors. During the data collection process of multiple sensors, due to factors such as equipment accuracy and environmental interference, there is a certain degree of uncertainty. For example, water quality sensors may be affected by impurities in the water, electrode aging, etc., resulting in errors in the measured data. The variational autoencoder can model and fuse these uncertainty parameters. Let the set of uncertainty parameters of multiple sensors be , represent the uncertainty parameter of the th sensor. Through the variational autoencoder, is processed to generate a water quality mutation confidence index that comprehensively reflects the uncertainty of multiple sensors. The value range of

[0059] is between 0 and 1. The closer it is to 1, the higher the confidence in the water quality mutation prediction result, and the closer it is to 0, the lower the reliability of the prediction result. Embed the generated water quality mutation confidence index into the real-time pollution warning instruction. When a pollution warning is issued, not only inform relevant personnel of the information that the water quality may mutate, but also provide this confidence index at the same time. In this way, the personnel receiving the warning can arrange response measures more reasonably according to the confidence index. If the confidence index is relatively high, close to 1, it means that the possibility of water quality mutation is very high, and relatively strict prevention and control measures need to be taken immediately, such as strengthening the investigation of pollution sources and starting emergency treatment equipment; if the confidence index is relatively low, hovering around 0.5, it is necessary to further observe the changes in water quality and at the same time appropriately increase the monitoring frequency to more accurately judge the water quality status.It should be noted that in this text, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device.

[0060] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An Internet of Things-based river monitoring system, characterized in that, It includes: A multi-modal perception module: used to collect multi-source heterogeneous data of the river environment in real time; An environmental feature analysis module: performing cross-modal feature extraction on the multi-source heterogeneous data to generate feature vectors of each data source, including water pollution diffusion characteristics, flow velocity gradient distribution characteristics, rainfall spatio-temporal correlation characteristics, underwater obstacle contour characteristics, and floating object movement trajectory patterns; A spatio-temporal coupling modeling module: constructing a multi-dimensional feature fusion space based on a spatio-temporal graph attention network, mapping the feature vectors to a unified spatio-temporal grid coordinate system, and generating a coupled spatio-temporal feature matrix; An adaptive resource scheduling module: constructing a two-layer optimization library of a computing resource pool and a river state evaluation model, where the two-layer optimization library stores dynamic task allocation rules, edge node energy consumption balancing strategies, and a sensor-buoy device topology relationship map; An intelligent collaborative control module: generating real-time pollution warning instructions and hydrological regulation decisions through a multi-objective reinforcement learning framework based on the coupled spatio-temporal feature matrix and the adaptive resource scheduling module.

2. The river monitoring system according to claim 1, wherein The multi-source heterogeneous data includes water quality sensor parameters, hydrological flow velocity monitoring values, meteorological station precipitation data, underwater sonar images, and floating object visual detection flows; The cross-modal feature extraction of the multi-source heterogeneous data includes: Using wavelet transform to decompose multi-band pollution components of the water quality sensor parameters, and extracting the pollution source diffusion pattern through a time series clustering algorithm; Using a residual attention network to segment the underwater obstacle contour of the underwater sonar image, and predicting the obstacle displacement trajectory in combination with the optical flow method; Using a multi-scale convolutional network to extract the motion direction feature of the floating object visual detection flow, and matching the floating object aggregation trend based on the dynamic time warping algorithm.

3. The river channel monitoring system according to claim 2, wherein The cross-modal feature extraction further includes: Using a graph embedding algorithm to model the river terrain dependence relationship of the hydrological flow velocity monitoring values, and extracting the spatio-temporal correlation characteristics of the flow velocity mutation area through a gradient boosting tree; Using a causal convolutional network to model the lagging effect of rainfall events on water quality for the meteorological station precipitation data, and inferring the pollution risk propagation path based on a Bayesian network.

4. The river channel monitoring system according to claim 1, characterized in that, The construction of the two-layer optimization library of the computing resource pool and the river state evaluation model includes: Constructing a distributed computing resource pool using a transfer learning framework based on a dynamic task offloading protocol and an edge node deployment topology; Constructing a river state evaluation model using a spatio-temporal graph autoencoder according to historical hydrological data and real-time water quality parameter distributions.

5. The river channel monitoring system according to claim 1, characterized in that The multi-objective reinforcement learning framework adopts a hierarchical decision-making architecture, including: Defining the state space as the joint encoding of the coupled spatio-temporal feature matrix and the remaining computing power of the resource pool, and the action space as the decision combination of monitoring task priority allocation and data sampling frequency; Balancing the conflicts of each decision-making objective through a proximal policy optimization algorithm, and designing a multi-dimensional reward function based on energy consumption constraints.

6. The river channel monitoring system according to claim 5, wherein The multi-objective reinforcement learning framework further includes: Implementing multi-node asynchronous policy updates using a distributed soft actor-critic algorithm, and optimizing the global convergence speed in combination with the priority experience replay technique; Converting the device energy consumption limit into a dynamic penalty term of the policy network through a constraint relaxation method.

7. The river channel monitoring system according to claim 1, wherein, The system further includes: Perform anomaly clustering detection on the floating object movement trajectory pattern, and use the spectral clustering algorithm to identify the distribution area of floating objects exceeding the preset threshold; When an abnormal aggregation is detected, trigger the emergency response mechanism of the intelligent collaborative control module to generate an interception device deployment instruction and a collaborative positioning signal for adjacent buoy devices.

8. The river channel monitoring system according to claim 7, wherein The emergency response mechanism uses a bidirectional graph neural network to model the interaction of floating objects, including: Extract the floating object migration trend characteristics from the forward propagation path, and extract the influence range characteristics of the interception device from the reverse propagation path; Generate an optimal deployment plan through the attention mechanism and update the dynamic risk node weights of the river channel state evaluation model.

9. The river channel monitoring system according to claim 1, characterized in that The system also includes: Construct a water quality mutation prediction model, based on historical pollution event data and real-time meteorological parameters, use a gated recurrent unit network to fuse the environmental attenuation factor, and dynamically adjust the sampling frequency threshold of the monitoring task.

10. The river monitoring system according to claim 9, wherein, The water quality mutation prediction model also includes: Based on the variational autoencoder, fuse the multi-sensor uncertainty parameters, generate a water quality mutation confidence index and embed it into the real-time pollution warning instruction.

Images