Urban shallow lake ecological protection and restoration method and system integrated with intelligent management and control

Abstract

The invention discloses an urban shallow lake ecological protection and restoration method and system integrated with intelligent management and control, and relates to the technical field of water environment ecological management and intelligent water affairs, and the method comprises the steps: constructing a three-dimensional monitoring network to synchronously collect water quality, image and meteorological data, and carrying out the fusion to generate a multi-modal data set; a deep learning model is utilized to realize pollution source tracing, water quality prediction and algae bloom early warning in parallel; based on the prediction result, outputting a multi-facility cooperative regulation strategy through a reinforcement learning agent; and converting the strategy into a hierarchical instruction to drive an execution unit, and feeding back the treated environment state to the model and the intelligent agent to form closed-loop optimization. According to the invention, whole-course intelligent management and control from monitoring to execution are realized, systematicness, accuracy and perspectiveness of lake treatment are effectively improved, and water environment risk response capability and ecological restoration effect are significantly enhanced.

Description

Technical Field

[0001] This invention relates to the field of water environment ecological governance and smart water technology, and more specifically, to a method and system for ecological protection and restoration of urban shallow lakes with integrated intelligent management and control. Background Technology

[0002] Urban shallow lakes, as an important component of urban aquatic ecosystems, are characterized by their shallowness, poor water flow, and susceptibility to urban non-point source pollution. They generally face problems such as unstable water quality, high risk of eutrophication, habitat monoculture, ecosystem degradation, and loss of self-purification capacity. Especially against the backdrop of rapid urbanization, a large amount of pollutants enter the lakes through tributaries, leading to frequent exceedances of key water quality indicators such as total phosphorus and chemical oxygen demand, frequent algal blooms, and seriously affecting water body functions and urban landscapes.

[0003] Currently, the treatment of such lakes mostly adopts traditional engineering methods, such as local dredging, artificial aeration, shoreline hardening, and single wetland construction, which have the following prominent problems: (1) Existing methods often focus on a certain link or a single problem, lacking a systematic treatment approach from the whole chain of "tributary-estuary-reservoir area", resulting in poor coordination between various measures and difficulty in achieving lasting treatment effects. (2) The multidimensional coupling relationship between "riverbank-water body-bottom sediment" is often ignored during the restoration process, resulting in incomplete habitat restoration, weak self-recovery capacity of the ecosystem, and difficulty in coping with external disturbances. (3) Management relies on regular manual monitoring, which cannot keep track of water quality dynamic changes in real time, and has insufficient early warning capabilities for sudden ecological risks such as algal blooms. Response measures often lag behind the occurrence of pollution events. (4) After the treatment facilities are built, they are mostly in a fixed operation mode, which cannot be dynamically optimized and controlled according to real-time hydrological and water quality conditions, resulting in low operating efficiency and high costs.

[0004] Although concepts such as "sponge cities" and "smart rivers and lakes" are being promoted, a comprehensive technological system covering the entire chain of pollution source tracing, engineering integration, intelligent feedback, and emergency response has not yet been established, especially lacking customized integrated management solutions for urban shallow lakes. Therefore, there is an urgent need for a systematic solution that can coordinate "source-process-end" and integrate ecological engineering and intelligent management technologies to achieve sustainable ecological restoration and long-term operation and maintenance of urban shallow lakes. Summary of the Invention

[0005] To address the aforementioned technical challenges, this invention proposes an integrated intelligent management and control method and system for the ecological protection and restoration of urban shallow lakes. This system achieves intelligent management and control throughout the entire process, from monitoring to execution, effectively enhancing the systematicness, precision, and foresight of lake governance, and significantly improving the ability to respond to water environment risks and the effectiveness of ecological restoration.

[0006] The first aspect of this invention provides a method for ecological protection and restoration of urban shallow lakes with integrated intelligent management and control, comprising the following steps: A three-dimensional monitoring network was constructed to simultaneously collect water quality parameters, water images, and meteorological data of the target lake. The collected multi-source heterogeneous data were spatiotemporally aligned and fused to generate a multimodal dataset of lake ecological status. The multimodal dataset is analyzed using a deep learning model, and tasks including dynamic pollution source tracing based on graph neural networks, generating water quality distribution maps for preset time periods based on spatiotemporal prediction models, and early warning of algal bloom probability and path based on sequence learning are executed in parallel. Based on the prediction and early warning results, decision optimization is carried out through a reinforcement learning agent. The state space of the reinforcement learning agent includes predicted water quality and algal bloom risk information, and the action space includes control instructions for at least one of aeration equipment, ecological water regulating gate pump, and carbon source dosing device. Through interaction with environmental simulation, the optimal strategy for coordinated regulation of multiple facilities is output. The optimal strategy is converted into hierarchical control instructions and sent to the corresponding execution units. The environmental state changes after execution are collected by the monitoring network and fed back to the deep learning model and reinforcement learning agent for closed-loop optimization.

[0007] This scheme constructs a three-dimensional monitoring network to simultaneously collect water quality parameters, water images, and meteorological data of the target lake. The collected multi-source heterogeneous data are spatiotemporally aligned and fused to generate a multimodal dataset of the lake's ecological status, including: Simulations were conducted using historical hydrological data and hydrodynamic models to identify key hydrological nodes in the target lake. These key hydrological nodes were then used as the basic anchor points of a three-dimensional monitoring network. Integrated intelligent monitoring buoys were deployed at each anchor point to construct intelligent sensing nodes. By establishing communication links between intelligent sensing nodes and weather station nodes through a mobile self-organizing network, a decentralized and self-healing dynamic self-organizing network is generated, thus constructing a three-dimensional monitoring network. Based on the multispectral image data of the target lake acquired by satellite, the satellite downlink synchronization beacon signal is sent to the ground receiving station. The ground receiving station broadcasts the synchronization beacon signal to the three-dimensional monitoring network, triggering nodes to sample and instructing a swarm of drones to conduct coordinated take-off and shooting to acquire three-dimensional laser data and multispectral imaging data. A real-scene 3D model containing lake bottom topography, shoreline structure, and engineering facility locations is generated using the 3D laser data and multispectral imaging data. A digital twin base of the target lake is generated. Visual feature regions are identified and segmented based on the multispectral imaging data and water images. These visual feature regions include algae aggregation areas, turbid water masses, and clear water areas. Information about the visual feature regions is obtained and correlated with the corresponding water quality parameters for verification. A quantitative mapping relationship library between visual features and water quality parameters is established. The water quality parameters, water body images, and meteorological data collected by the nodes are bound with timestamps and coordinates, and integrated into the real-scene 3D model along with data from the quantitative mapping relation library. This constructs a 3D geographic information layer, a spatiotemporally continuous water quality parameter distribution layer, a water body visual feature and event annotation layer, a meteorological driving factor layer, and an ecological engineering setting state layer, generating a structured multimodal dataset of lake ecological status.

[0008] In this solution, dynamic source tracing of pollution based on graph neural networks includes: The target lake is divided into grids, and each grid is used as a graph node. Key nodes are set up at the river mouths directly flowing into the reservoir, the center of the reservoir area, sensitive areas, and known endogenous pollution points. The water quality parameters, underwater image features, and meteorological data of each grid are used as dynamic attribute vectors. The flow field path is derived from the buoy trajectory. Nodes with hydrodynamic connections are selected based on the flow field path to establish directed edges. Dynamic edge weights are constructed based on real-time wind speed, wind direction, and node distance. The spatial dependency of the constructed graph structure is modeled by a graph attention network. The current wind field and flow field direction are used as prior knowledge to guide attention. Possible pollutant diffusion paths are judged based on the prior knowledge. Neighboring nodes on the pollutant diffusion path are given higher attention weights. The spatial dependency features of pollution are obtained from the output of the graph attention network. Gated temporal convolutional networks are used to analyze the temporal data of each node, obtain the temporal patterns related to endogenous contamination points, and obtain the temporal dependence features of contamination. The spatial and temporal dependence features are used to dynamically simulate the pollution process in the lake. Water quality data of key nodes are used as monitoring signals for training. The trained network is used to calculate the dynamic contribution rate based on the occlusion method. The estuary node and the node group identified as a potential endogenous source area are occluded respectively. The contribution rate of estuary external input and endogenous load in different lake areas is calculated by the change of predicted values.

[0009] In this scheme, a water quality distribution map for a preset time period is generated based on a spatiotemporal prediction model, including: The target lake is divided into grids, and the sensor data, underwater images and meteorological data of the grid blocks are encoded to generate physical feature vectors that include spatial context, visual feature vectors that represent regional image features and meteorological driving feature vectors that represent the meteorological field. Within each grid, different data sources of the same module are weighted according to their importance, and cross-modal attention is introduced to calculate dynamic fusion weights based on information relevance. Data fusion is then performed based on these dynamic fusion weights to obtain high-dimensional features. High-dimensional feature sequences from historical periods are imported into a conditional generative adversarial network. Downsampling is used to obtain the spatiotemporal evolution trend of the lake's state. The data is then imported into the decoder and restored through upsampling and skip connections. A rolling prediction mechanism is used to generate a super-resolution water quality distribution map for each time step in the future preset period.

[0010] In this scheme, the probability and path prediction of algal blooms based on sequence learning include: Several complete algal bloom lifecycle sequences are extracted from historical data, centered on algal bloom events. Scene meta-features, including event type, active driving factors, and event intensity, are labeled based on the algal bloom lifecycle sequence in each algal bloom time. A multi-task learning model based on an encoder-decoder architecture trained on algal bloom events with scene meta-features is used. The input is a sequence of current and past multimodal ecological environments, and the encoder is used to extract a comprehensive state vector containing the potential driving force of algal blooms. The probability of algal blooms is determined by using an occurrence probability decoder, a scale intensity decoder, and a diffusion path decoder to share the comprehensive state vector. The occurrence probability decoder determines the probability of algal blooms at each future time point. The scale intensity decoder outputs the area percentage and peak chlorophyll a concentration that may be affected in the future. The diffusion path decoder combines the predicted wind field and flow field data. The spatiotemporal attention module is used to simulate the dominant drift direction and influence range of the algal bloom.

[0011] In this scheme, based on the prediction and early warning results, decision optimization is performed through a reinforcement learning agent. The state space of the reinforcement learning agent includes predicted water quality and algal bloom risk information, including: The pollution source prediction results, water quality distribution maps for preset time periods, and early warning results of algal bloom probability and path are imported into a reinforcement learning agent to construct the lake state space. The action space is defined as a combination of discretized or segmented continuous control commands for aeration equipment, ecological water regulating gate pumps, and carbon source dosing devices, and a hierarchical structure is set for the control commands in the action space. A multi-objective reward function is constructed based on water quality rewards, ecological risk rewards, economic cost rewards, and engineering resilience rewards. The Dueling DQN architecture is adopted to decompose the Q value into a state value function and an advantage function. The outputs of the state value function and the advantage function are aggregated to obtain the final Q value. The reinforcement learning agent uses the final Q value to determine the severity of the current ecological state of the lake and the governance actions that can bring the most additional improvement. During training, the temporal difference error of each experience is calculated, and the sampling priority is set according to the temporal difference error of the experience samples. By prioritizing the replay of experiences, the reinforcement learning agent learns from erroneous and unexpected experiences, thus accelerating the convergence to the optimal policy. Safety rules are set in the reinforcement learning agent after training, and the generated actions are verified using the safety rules. After verification, the optimal strategy for multi-facility coordinated control is output.

[0012] In this scheme, the optimal strategy is converted into hierarchical control instructions and sent to the corresponding execution units, including: A pre-defined semantic action dictionary maps the original action values ​​in the optimal policy output by the reinforcement learning agent to a predefined standardized operation mode, generates hierarchical control instructions, and introduces fuzzy logic smoothing, defining the critical region of the input variable as the transition region, and performing a smooth weighted transition of the instructions in the transition region. A dynamic multi-level response threshold library is constructed, and the thresholds are dynamically adjusted according to seasonality, hydrological period, and equipment status. The semantic action dictionary is also updated periodically. Hierarchical control commands are sent to each execution unit via the Industrial Internet of Things (IIoT) protocol. After the commands are sent, they are automatically linked to subsequent monitoring data, the changing trends of preset monitoring indicators are analyzed, and the results are compared with the expected results. Anomaly alarms are generated based on the comparison results.

[0013] In this scheme, the changes in the environmental state after execution are collected through a monitoring network and fed back to the deep learning model and reinforcement learning agent for closed-loop optimization, including: An intervention event log is created based on the content of the executed instructions, the start and end times, the geographical scope of the effect, and the expected governance goals. Centered on each intervention event, the pre-intervention state sequence, post-intervention response sequence, and control area data are extracted as environmental feedback data. The intervention events and environmental feedback data within a preset period are used to construct a fine-tuning task set. The deep learning models for different tasks are then fine-tuned. During the fine-tuning process, the importance of the model parameters is evaluated, and elastic constraints are applied to model parameters that exceed the importance threshold to prevent drastic changes in the parameters. The agent model in the simulator is fine-tuned using a fine-tuning task set. The intervention event is taken as input, and the environmental response sequence output by the simulator is required to be close to the actual monitored response sequence. The simulator is then calibrated. Furthermore, in the experience replay library of the reinforcement learning agent, real decision-making and execution processes are stored as special experiences and their corresponding priorities are set to the highest for periodic retraining of the reinforcement learning agent.

[0014] The second aspect of this invention provides an integrated intelligent management and control system for the ecological protection and restoration of urban shallow lakes. The system includes: a three-dimensional monitoring network unit, a multimodal data fusion and preprocessing unit, a multi-task deep learning prediction unit, a reinforcement learning intelligent decision-making unit, and a hierarchical execution and control unit. The three-dimensional monitoring network unit integrates water quality monitoring buoys for collecting water temperature, dissolved oxygen, total phosphorus, and chlorophyll a, camera units for acquiring surface and underwater images of the water body, and meteorological stations for collecting wind speed and light intensity, thereby collecting water quality parameters, water body images, and meteorological data of the target lake. The multimodal data fusion and preprocessing unit is responsible for spatiotemporal alignment and fusion of the collected multi-source heterogeneous data to generate a multimodal dataset of lake ecological status. The multi-task deep learning prediction unit analyzes the multimodal dataset and executes tasks in parallel, including dynamic pollution source tracing based on graph neural networks, generating water quality distribution maps for preset time periods based on spatiotemporal prediction models, and early warning of algal bloom probability and path based on sequence learning; the unit as a whole supports periodic optimization using an online learning mechanism. The reinforcement learning intelligent decision-making unit optimizes decisions based on prediction and early warning results through a deep Q-network reinforcement learning agent. The state space of the reinforcement learning agent contains predicted water quality and algal bloom risk information, and the action space contains control instructions for at least one of aeration equipment, ecological water regulating gate pumps, and carbon source dosing devices. Through interaction with environmental simulation, it outputs the optimal strategy for coordinated regulation of multiple facilities. The hierarchical execution and control unit transforms the generated intelligent decision-making strategy into specific executable physical actions, and performs refined control of the execution unit based on preset multi-level risk thresholds.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention overcomes the fragmented shortcomings of traditional pollution control measures. Through an integrated monitoring network, it achieves a holistic understanding of the lake's ecological state. Furthermore, by employing technologies such as dynamic spatiotemporal neural networks, it enables dynamic and precise tracing of pollution sources and accurate depiction of pollution migration processes within the lake. This precise targeting of key pollution sources and ecologically vulnerable areas significantly enhances the relevance and synergistic benefits of engineering measures.

[0016] By employing a deep learning model that integrates multimodal data, a high spatiotemporal resolution water quality distribution map for future periods is generated, enabling precise prediction and early warning of the probability, intensity, and diffusion paths of ecological risks such as algal blooms. This provides advance notice for management decisions. Combined with optimized control strategies generated by reinforcement learning agents, the system can automatically initiate tiered and collaborative response measures, effectively containing ecological risks at their inception and significantly improving the level of water environment safety. The intelligent decision engine comprehensively considers multiple objectives such as water quality improvement, risk prevention and control, operating costs, and equipment lifespan, outputting the optimal strategy that achieves a balance between ecological and economic benefits, ensuring the sustainability and long-term effectiveness of the governance project. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments or examples of the present invention, the drawings used in the embodiments or examples will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained according to these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating an integrated intelligent management and control method for the ecological protection and restoration of urban shallow lakes is shown. Figure 2 A flowchart illustrating decision optimization through reinforcement learning agents is shown. Figure 3 A flowchart illustrating the conversion of the optimal strategy into hierarchical control instructions is shown. Figure 4 A block diagram of an integrated intelligent management and control system for the ecological protection and restoration of urban shallow lakes is shown. Detailed Implementation

[0019] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0020] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0021] like Figure 1 As shown, this embodiment provides a method for ecological protection and restoration of urban shallow lakes with integrated intelligent management and control, including: A three-dimensional monitoring network was constructed to simultaneously collect water quality parameters, water images, and meteorological data of the target lake. The collected multi-source heterogeneous data were spatiotemporally aligned and fused to generate a multimodal dataset of lake ecological status. The multimodal dataset is analyzed using a deep learning model, and tasks including dynamic pollution source tracing based on graph neural networks, generating water quality distribution maps for preset time periods based on spatiotemporal prediction models, and early warning of algal bloom probability and path based on sequence learning are executed in parallel. Based on the prediction and early warning results, decision optimization is carried out through a reinforcement learning agent. The state space of the reinforcement learning agent includes predicted water quality and algal bloom risk information, and the action space includes control instructions for at least one of aeration equipment, ecological water regulating gate pump, and carbon source dosing device. Through interaction with environmental simulation, the optimal strategy for coordinated regulation of multiple facilities is output. The optimal strategy is converted into hierarchical control instructions and sent to the corresponding execution units. The environmental state changes after execution are collected by the monitoring network and fed back to the deep learning model and reinforcement learning agent for closed-loop optimization.

[0022] It should be noted that a three-dimensional monitoring network is constructed, comprising a ground-based monitoring unit, an image monitoring unit, and a meteorological monitoring unit. The ground-based monitoring unit deploys multi-parameter intelligent water quality monitoring buoys to collect real-time water physicochemical parameters, including water temperature, pH, dissolved oxygen, ammonia nitrogen, total phosphorus, and chlorophyll a. The image monitoring unit includes underwater cameras and visible light and multispectral cameras deployed along the shoreline and on drones to periodically acquire image data on underwater biological status, surface geology, and shoreline vegetation coverage. The meteorological monitoring unit deploys small weather stations to collect meteorological data on wind speed, wind direction, light intensity, and precipitation. The aforementioned multi-source, heterogeneous monitoring data are transmitted in real-time and spatiotemporally aligned via an IoT gateway to form a multimodal fusion data cube describing the state of the lake ecosystem. Simulations using historical hydrological data and hydrodynamic models identify key hydrological nodes in the target lake, such as inflow estuaries, outflow estuaries, reservoir center, areas of slow flow, and areas of abrupt changes in water depth. These key hydrological nodes serve as the foundational anchor points for a three-dimensional monitoring network. Integrated intelligent monitoring buoys are deployed at each anchor point to construct intelligent sensing nodes. Communication links are established between the intelligent sensing nodes and meteorological station nodes through a mobile ad hoc network, generating a decentralized and self-healing dynamic ad hoc network to construct the three-dimensional monitoring network. When a node fails or communication is disrupted, data can be relayed through other paths, enhancing the network's robustness.

[0023] Based on the multispectral image data of the target lake acquired by satellite, when the Gaofen series satellite passes over and takes a multispectral image of the target lake, the satellite downlink synchronization beacon signal is sent to the ground receiving station. The ground receiving station broadcasts the synchronization beacon signal to the three-dimensional monitoring network. The signal is broadcast to the trigger nodes of the entire monitoring network for sampling, and instructs the UAV swarm to take off and take pictures in a coordinated manner to acquire three-dimensional laser data and multispectral imaging data, ensuring that the data is aligned at the millisecond level in terms of acquisition time. A realistic 3D model containing lakebed topography, shoreline structure, and engineering facility locations is generated using the aforementioned 3D laser data and multispectral imaging data. This model includes information such as topography, shoreline, water depth, and terrestrial vegetation. This model serves as a digital twin of the target lake. Based on the multispectral imaging data and water images, a pre-defined deep learning model identifies and segments visual feature regions, including algae aggregation areas, turbid water masses, and clear water areas. Information from these visual feature regions is correlated and verified with water quality parameters such as chlorophyll a and turbidity measured by buoys, establishing a quantitative mapping relationship library between visual features and water quality parameters. Water quality parameters, water images, and meteorological data collected from nodes are bound to timestamps and coordinates, and integrated with data from the quantitative mapping relationship library into the realistic 3D model. This constructs a 3D geographic information layer, a spatiotemporally continuous water quality parameter distribution layer, a water visual feature and event annotation layer, a meteorological driving factor layer, and an ecological engineering setting state layer, generating a structured multimodal dataset of lake ecological status.

[0024] It should be noted that the lake internal pollution source tracing based on dynamic spatiotemporal graph neural networks identifies and quantifies the contribution rate of different regions within the lake and direct inflow river estuaries to the overall pollution of the lake area. The target lake is divided into grids, with each grid serving as a graph node. Key nodes are deployed at direct inflow river estuaries, the center of the reservoir area, sensitive areas, and known endogenous pollution points. Water quality parameters (total phosphorus, ammonia nitrogen, dissolved oxygen, etc.), underwater image features (turbidity, algae density visual indicators, etc.), and meteorological data (wind speed, wind direction, etc.) of each grid are used as dynamic attribute vectors. The flow field path is inverted based on the buoy trajectory. Nodes with hydrodynamic connections are selected based on the flow field path to establish directed edges. Dynamic edge weights are constructed based on real-time wind speed, wind direction, and node distance to simulate the impact of wind direction and wind speed on pollutant diffusion. The spatial dependency of the constructed graph structure is modeled using a graph attention network. The current wind and flow directions are used as prior knowledge to guide attention. Based on this prior knowledge, possible pollutant diffusion paths are determined, and neighboring nodes along these paths are assigned higher attention weights. The spatial dependency features of the pollution are obtained from the graph attention network output. For example, when the wind is from the south, a node will assign higher attention weights to its northern neighbors because pollutants may diffuse from the south. This mechanism can more physically simulate the diffusion paths of pollutants within the lake. A gated temporal convolutional network is used to analyze the temporal data of each node, obtaining temporal patterns related to endogenous pollution points and acquiring the temporal dependency features of the pollution. For example, if dissolved oxygen drops sharply during high temperatures, the total phosphorus concentration at the bottom sediment pollution node will subsequently rise, delaying its impact on other areas. The spatial and temporal dependency features are used to dynamically simulate the pollution process within the lake. For example, a southerly wind pushes pollutants from estuary A towards the lake center, while high temperatures cause phosphorus release from the bottom sediment in the northern region. If the wind direction changes to easterly, the two pollution plumes will converge near the intake. Water quality data from key nodes were used as monitoring signals for training. The trained network was then used to calculate the dynamic contribution rate based on the occlusion method. The estuary node and the node group identified as a potential endogenous source area were occluded respectively. The data of the estuary node were occluded, and the changes in the predicted water quality values ​​of the target points were observed to calculate the real-time contribution rate of the estuary external source input. The node group identified as a potential endogenous source area was occluded, and the contribution rate of the endogenous load in different lake areas was quantified by the changes in the predicted values.

[0025] It should be noted that the water quality parameter super-resolution prediction based on the multimodal attention mechanism generates a high spatiotemporal resolution water quality parameter distribution map covering the entire lake within a preset future time period. The target lake is divided into grids, and sensor data from each grid block is processed by a spatial interpolation encoder to generate physical feature vectors containing spatial context information. A pre-trained convolutional neural network is used to extract deep features from image blocks within each grid, generating visual feature vectors representing the regional image features. Meteorological field data is directly mapped to each grid, generating meteorological driving feature vectors representing the meteorological field.

[0026] Within each grid, different data sources within the same module are weighted by importance. For example, buoy data within the current grid has higher confidence than data obtained through interpolation at a greater distance. Cross-modal attention is introduced to calculate dynamic fusion weights based on information relevance. For instance, in algal bloom prediction scenarios, the model will pay more attention to chlorophyll a data from physical sensors in clear lake areas. Based on the dynamic fusion weights, specific spatiotemporal context dynamic data fusion is performed to obtain high-dimensional features. The high-dimensional feature sequences from historical periods are imported into a conditional generative adversarial network (GAN). The generator is a U-Net-structured spatiotemporal sequence prediction model. Downsampling is used to obtain the spatiotemporal evolution trend of the lake state. This trend is then imported into the decoder, where upsampling and skip connections are used to restore spatial details. The discriminator judges whether the generated distribution map is consistent with the actual observations, making the generator output more reasonable and avoiding ambiguous predictions. The prediction results of the previous time step are used as one of the inputs for the next time step. A rolling prediction mechanism is used to generate super-resolution water quality distribution maps for each time step in the future preset period. Monte Carlo Dropout is used to provide uncertainty intervals for the predicted values ​​of each grid at future time points and to label the prediction confidence.

[0027] It should be noted that this is a sequence learning-based early warning system for algal bloom probability and path. Several complete algal bloom lifecycle sequences are segmented from historical data, centered on the algal bloom event. Within each algal bloom time period, scene meta-features are labeled based on the algal bloom lifecycle sequence, including event type (post-typhoon type, sustained high-temperature type, etc.), active driving factors (temperature-driven, light-driven, etc.), and event intensity. A meta-task is constructed, randomly selecting a subset of events from the meta-knowledge base as the support set, and another subset as the query set. This allows the model to learn to quickly adjust its internal parameters after only seeing the support set, and to make accurate predictions on the query set. It also enables the model to quickly identify and adapt to new algal bloom patterns based on a small amount of new data. A multi-task learning model based on an encoder-decoder architecture trained on algal bloom events with scene-specific features is used. It takes current and past multimodal ecological environment sequences as input, understands and encodes the current ecological state context of the lake, and uses the encoder to extract a comprehensive state vector containing the potential driving forces of algal blooms. This comprehensive state vector is shared by an occurrence probability decoder, a scale-intensity decoder, and a diffusion path decoder. The occurrence probability decoder determines the probability of algal blooms at future time points, the scale-intensity decoder outputs the potential area affected and peak chlorophyll a concentration, and the diffusion path decoder combines predicted wind and flow field data. A spatiotemporal attention module simulates the dominant drift direction and impact range of the algal bloom. The model ultimately outputs a comprehensive early warning map integrating the results of the three decoders.

[0028] When the model is deployed on a new lake, it only needs to be input as a small amount of observation data in the new scenario as a support set. The model parameters can be fine-tuned quickly through the meta-learning framework, and the general knowledge learned in the meta-knowledge base can be adapted to the current specific environment.

[0029] It should be noted that a hybrid simulator is constructed to model mechanistic and data processes, simulating water physical processes and basic biochemical reactions such as flow velocity and diffusion, ensuring the physical plausibility of the simulation results. The aforementioned deep learning prediction model is embedded into the simulator to simulate complex nonlinear ecological processes such as algal growth and pollutant transformation, accelerating the simulation process.

[0030] like Figure 2 As shown, the pollution source tracing prediction results, water quality distribution maps for preset time periods, and early warning results of algal bloom occurrence probability and path are imported into a reinforcement learning agent to construct a lake state space, including algal bloom risk status, water quality health, equipment operating status, and economic costs. The action space is defined as a combination of discrete or segmented continuous control commands for aeration equipment, ecological water regulating gate pumps, and carbon source dosing devices. A hierarchical structure is set for the control commands in the action space. The upper-level decision selects the dominant governance mode, and the lower-level decision adjusts the operating parameters of each device continuously under the selected mode.

[0031] A multi-objective reward function is constructed based on water quality rewards, ecological risk rewards, economic cost rewards, and engineering resilience rewards. The total reward = W1 * water quality reward + W2 * ecological risk reward + W3 * economic cost reward + W4 * engineering resilience reward, where W1, W2, W3, and W4 represent weights. The water quality reward is positively correlated with the reduction in key water quality parameters (such as chlorophyll a and total phosphorus) and the area meeting the standards. The ecological risk reward is correlated with the reduction in the probability of algal blooms. The economic cost reward is negatively correlated with energy consumption costs and reagent consumption costs. A time-of-use electricity price factor is introduced to encourage the operation of high-energy-consuming equipment during off-peak hours. The engineering resilience reward penalizes frequent start-ups and shutdowns of equipment, encourages stable operation, and extends equipment life.

[0032] The Dueling DQN architecture enables the agent to better distinguish between the inherent goodness or badness of the current lake state and the additional benefits of performing a specific action relative to other actions. For example, in a high-risk state of algal bloom, the agent will focus on selecting advantageous actions that significantly reduce risk. The Q-value is decomposed into a state value function and an advantage function. The state value function measures the intrinsic goodness or badness of the state, regardless of the specific action taken; the advantage function measures the additional benefits of taking a specific action in a state relative to the average action in that state. The outputs of the state value function and the advantage function are aggregated to obtain the final Q-value. The reinforcement learning agent uses this final Q-value to determine the severity of the current lake ecological state and the governance action that can bring the most additional improvement; the output of the output layer is then aggregated. Represented as: , in, It is in the state of a lake. For a specific action, These are shared parameters for the main network entities. These are the unique parameters for the advantage stream and the value stream, respectively. The output of the value stream represents the value of state s. The output of the dominant flow has the same dimensions as the action space, with each dimension representing the dominant value of the corresponding action. The size of the action space, For possible actions, This is the average of the dominance values.

[0033] During training, the time difference error of each experience sample is calculated. Sampling priority is set based on the time difference error of the experience samples, and the time difference error of each experience sample is calculated. Time difference error This measures the difference between the current network's prediction of the Q-value and the improved target Q-value. The larger the absolute value, the more likely the current network's prediction is unexpectedly wrong or correct, and the greater the potential for empirical results. It is expressed as: , in, For the agent in time step At that time, in a state Execute action Then, the immediate reward value from environmental feedback, As a discount factor, For estimating the maximum Q-value of the next state in the objective Q-value, To perform the action After that, the environment transitions to the next state. For the target network parameters, The Q value calculated for the target network. The current parameters of the main network.

[0034] Prioritize this experience Set as , Let be a very small positive number, storing the priorities of all experiences. Sampling is performed according to the priority ratio. When sampling experiences for training, sampling is performed according to the proportion of each experience's priority in the total priority. The loss function for each sample is multiplied by the importance sampling weight. To avoid introducing bias, , For the size of the experience replay library, Let be the probability that sample i is selected, which is proportional to the priority. These are hyperparameters. Use them. Update network parameters to ensure the unbiasedness and stability of the learning process. ，By prioritizing experience replay, the reinforcement learning agent learns from erroneous and unexpected experiences, accelerating convergence to the optimal policy. Safety rules are set in the trained reinforcement learning agent, and the generated actions are verified using the safety rules. After verification, the optimal policy for multi-facility coordinated control is output.

[0035] It should be noted that, as Figure 3 As shown, a pre-defined semantic action dictionary maps the original action values ​​from the optimal strategy output by the reinforcement learning agent to predefined standardized operating modes, generating hierarchical control commands. For example, if the agent suggests a 65% opening degree for the water gate pump, it is interpreted and mapped to Mode 3: Mode 3: Medium-intensity ecological water diversion, based on the current water level and safety rules, and associated with the recommended duration and expected flow range. Fuzzy logic smoothing is introduced to avoid continuous equipment start-stop due to small fluctuations around thresholds. The critical region of the input variable is defined as a transition zone, within which commands are smoothly weighted and transitioned; for example, a risk probability of 0.4-0.6 is a medium-risk transition zone. When the risk probability increases from 0.45 to 0.55, the command smoothly transitions from executing a low-level response with 80% weight and a medium-level response with 20% weight to a low-level response with 20% weight and a medium-level response with 80% weight. A dynamic multi-level response threshold library is constructed, dynamically adjusting thresholds according to seasonality, hydrological periods, and equipment status, and periodically updating the semantic action dictionary; higher thresholds are used in winter to avoid unnecessary equipment operation. If a certain execution unit is under maintenance, the relevant threshold will be temporarily increased, or the response action will be automatically switched to other available devices. Hierarchical control commands are sent to each execution unit via the Industrial Internet of Things (IIoT) protocol. For complex strategies involving multiple devices, time-sequential work orders are generated. After the commands are sent, subsequent monitoring data is automatically correlated, the changing trends of preset monitoring indicators are analyzed, and the results are compared with the expected effects. Anomaly alarms are generated based on the comparison results.

[0036] An intervention event log is created based on the executed instructions, start and end times, geographical scope of effect, and expected governance objectives. For each intervention event, pre-intervention state sequences, post-intervention response sequences, and control area data (data from other similar areas not directly affected by the intervention within the same timeframe) are extracted as environmental feedback data. The intervention events and environmental feedback data within a preset period are used to construct a fine-tuning task set. Deep learning models for different tasks are fine-tuned, and the importance of model parameters is evaluated during fine-tuning. Elastic constraints are applied to model parameters exceeding the importance threshold to prevent drastic parameter changes. The fine-tuning task set is used to fine-tune the surrogate model in the simulator, using intervention events as input. The simulator's output environmental response sequence is required to closely approximate the actual monitored response sequence. The simulator is calibrated, and the simulated physical laws increasingly resemble the response characteristics of real lakes. Real decision-making and execution processes are stored as special experiences in the experience replay library of the reinforcement learning agent, with the corresponding priorities set to the highest. The reinforcement learning agent is periodically retrained, continuously revising the decision-making model by learning directly from real-world feedback, ultimately evolving the optimal control strategy for the specific lake.

[0037] like Figure 4 As shown, the second embodiment of the present invention provides an integrated intelligent management and control system for the ecological protection and restoration of urban shallow lakes. The system includes: a three-dimensional monitoring network unit, a multimodal data fusion and preprocessing unit, a multi-task deep learning prediction unit, a reinforcement learning intelligent decision-making unit, and a hierarchical execution and control unit. The three-dimensional monitoring network unit integrates water quality monitoring buoys for collecting water temperature, dissolved oxygen, total phosphorus, and chlorophyll a, camera units for acquiring surface and underwater images of the water body, and meteorological stations for collecting wind speed and light intensity, thereby collecting water quality parameters, water body images, and meteorological data of the target lake. The multimodal data fusion and preprocessing unit is responsible for spatiotemporal alignment and fusion of the collected multi-source heterogeneous data to generate a multimodal dataset of lake ecological status. The multi-task deep learning prediction unit analyzes the multimodal dataset and executes tasks in parallel, including dynamic pollution source tracing based on graph neural networks, generating water quality distribution maps for preset time periods based on spatiotemporal prediction models, and early warning of algal bloom probability and path based on sequence learning; the unit as a whole supports periodic optimization using an online learning mechanism. The reinforcement learning intelligent decision-making unit optimizes decisions based on prediction and early warning results through a deep Q-network reinforcement learning agent. The state space of the reinforcement learning agent contains predicted water quality and algal bloom risk information, and the action space contains control instructions for at least one of aeration equipment, ecological water regulating gate pumps, and carbon source dosing devices. Through interaction with environmental simulation, it outputs the optimal strategy for coordinated regulation of multiple facilities. The hierarchical execution and control unit transforms the generated intelligent decision-making strategy into specific executable physical actions, and performs refined control of the execution unit based on preset multi-level risk thresholds.

[0038] The third embodiment of the present invention provides a computer-readable storage medium, which includes a program for an integrated intelligent management and control method for the ecological protection and restoration of urban shallow lakes. When the program is executed by a processor, it implements the steps of the integrated intelligent management and control method for the ecological protection and restoration of urban shallow lakes.

[0039] In the several embodiments provided in this application, it should be understood that the disclosed methods and systems can be implemented in other ways. The system embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, and can be electrical, mechanical, or other forms. Furthermore, in the various embodiments of the present invention, all functional units can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0040] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0041] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for ecological protection and restoration of urban shallow lakes with integrated intelligent management and control, characterized in that, Includes the following steps: A three-dimensional monitoring network was constructed to simultaneously collect water quality parameters, water images, and meteorological data of the target lake. The collected multi-source heterogeneous data were spatiotemporally aligned and fused to generate a multimodal dataset of lake ecological status. The multimodal dataset is analyzed using a deep learning model, and tasks including dynamic pollution source tracing based on graph neural networks, generating water quality distribution maps for preset time periods based on spatiotemporal prediction models, and early warning of algal bloom probability and path based on sequence learning are executed in parallel. Based on the prediction and early warning results, decision optimization is carried out through a reinforcement learning agent. The state space of the reinforcement learning agent includes predicted water quality and algal bloom risk information, and the action space includes control instructions for at least one of aeration equipment, ecological water regulating gate pump, and carbon source dosing device. Through interaction with environmental simulation, the optimal strategy for coordinated regulation of multiple facilities is output. The optimal strategy is converted into hierarchical control instructions and sent to the corresponding execution units. The environmental state changes after execution are collected by the monitoring network and fed back to the deep learning model and reinforcement learning agent for closed-loop optimization.

2. The integrated intelligent management and control method for ecological protection and restoration of urban shallow lakes according to claim 1, characterized in that, A three-dimensional monitoring network was constructed to simultaneously collect water quality parameters, water images, and meteorological data of the target lake. The collected multi-source heterogeneous data were spatiotemporally aligned and fused to generate a multimodal dataset of the lake's ecological status, including: Simulations were conducted using historical hydrological data and hydrodynamic models to identify key hydrological nodes in the target lake. These key hydrological nodes were then used as the basic anchor points of a three-dimensional monitoring network. Integrated intelligent monitoring buoys were deployed at each anchor point to construct intelligent sensing nodes. By establishing communication links between intelligent sensing nodes and weather station nodes through a mobile self-organizing network, a decentralized and self-healing dynamic self-organizing network is generated, thus constructing a three-dimensional monitoring network. Based on the multispectral image data of the target lake acquired by satellite, the satellite downlink synchronization beacon signal is sent to the ground receiving station. The ground receiving station broadcasts the synchronization beacon signal to the three-dimensional monitoring network, triggering nodes to sample and instructing a swarm of drones to conduct coordinated take-off and shooting to acquire three-dimensional laser data and multispectral imaging data. A real-scene 3D model containing lake bottom topography, shoreline structure, and engineering facility locations is generated using the 3D laser data and multispectral imaging data. A digital twin base of the target lake is generated. Visual feature regions are identified and segmented based on the multispectral imaging data and water images. These visual feature regions include algae aggregation areas, turbid water masses, and clear water areas. Information about the visual feature regions is obtained and correlated with the corresponding water quality parameters for verification. A quantitative mapping relationship library between visual features and water quality parameters is established. The water quality parameters, water body images, and meteorological data collected by the nodes are bound with timestamps and coordinates, and integrated into the real-scene 3D model along with data from the quantitative mapping relation library. This constructs a 3D geographic information layer, a spatiotemporally continuous water quality parameter distribution layer, a water body visual feature and event annotation layer, a meteorological driving factor layer, and an ecological engineering setting state layer, generating a structured multimodal dataset of lake ecological status.

3. The integrated intelligent management and control method for ecological protection and restoration of urban shallow lakes according to claim 1, characterized in that, Dynamic source tracing of pollution based on graph neural networks includes: The target lake is divided into grids, and each grid is used as a graph node. Key nodes are set up at the river mouths directly flowing into the reservoir, the center of the reservoir area, sensitive areas, and known endogenous pollution points. The water quality parameters, underwater image features, and meteorological data of each grid are used as dynamic attribute vectors. The flow field path is derived from the buoy trajectory. Nodes with hydrodynamic connections are selected based on the flow field path to establish directed edges. Dynamic edge weights are constructed based on real-time wind speed, wind direction, and node distance. The spatial dependency of the constructed graph structure is modeled by a graph attention network. The current wind field and flow field direction are used as prior knowledge to guide attention. Possible pollutant diffusion paths are judged based on the prior knowledge. Neighboring nodes on the pollutant diffusion path are given higher attention weights. The spatial dependency features of pollution are obtained from the output of the graph attention network. Gated temporal convolutional networks are used to analyze the temporal data of each node, obtain the temporal patterns related to endogenous contamination points, and obtain the temporal dependence features of contamination. The spatial and temporal dependence features are used to dynamically simulate the pollution process in the lake. Water quality data of key nodes are used as monitoring signals for training. The trained network is used to calculate the dynamic contribution rate based on the occlusion method. The estuary node and the node group identified as a potential endogenous source area are occluded respectively. The contribution rate of estuary external input and endogenous load in different lake areas is calculated by the change of predicted values.

4. The integrated intelligent management and control method for ecological protection and restoration of urban shallow lakes according to claim 1, characterized in that, A water quality distribution map for a preset time period is generated based on a spatiotemporal prediction model, including: The target lake is divided into grids, and the sensor data, underwater images and meteorological data of the grid blocks are encoded to generate physical feature vectors that include spatial context, visual feature vectors that represent regional image features and meteorological driving feature vectors that represent the meteorological field. Within each grid, different data sources of the same module are weighted according to their importance, and cross-modal attention is introduced to calculate dynamic fusion weights based on information relevance. Data fusion is then performed based on these dynamic fusion weights to obtain high-dimensional features. High-dimensional feature sequences from historical periods are imported into a conditional generative adversarial network. Downsampling is used to obtain the spatiotemporal evolution trend of the lake's state. The data is then imported into the decoder and restored through upsampling and skip connections. A rolling prediction mechanism is used to generate a super-resolution water quality distribution map for each time step in the future preset period.

5. The integrated intelligent management and control method for ecological protection and restoration of urban shallow lakes according to claim 1, characterized in that, Algal bloom probability and path prediction based on sequence learning include: Several complete algal bloom lifecycle sequences are extracted from historical data, centered on algal bloom events. Scene meta-features, including event type, active driving factors, and event intensity, are labeled based on the algal bloom lifecycle sequence in each algal bloom time. A multi-task learning model based on an encoder-decoder architecture trained on algal bloom events with scene meta-features is used. The input is a sequence of current and past multimodal ecological environments, and the encoder is used to extract a comprehensive state vector containing the potential driving force of algal blooms. The probability of algal blooms is determined by using an occurrence probability decoder, a scale intensity decoder, and a diffusion path decoder to share the comprehensive state vector. The occurrence probability decoder determines the probability of algal blooms at each future time point. The scale intensity decoder outputs the area percentage and peak chlorophyll a concentration that may be affected in the future. The diffusion path decoder combines the predicted wind field and flow field data. The spatiotemporal attention module is used to simulate the dominant drift direction and influence range of the algal bloom.

6. The integrated intelligent management and control method for ecological protection and restoration of urban shallow lakes according to claim 1, characterized in that, Based on the prediction and early warning results, decision optimization is performed through a reinforcement learning agent. The state space of the reinforcement learning agent includes predicted water quality and algal bloom risk information, including: The pollution source prediction results, water quality distribution maps for preset time periods, and early warning results of algal bloom probability and path are imported into a reinforcement learning agent to construct the lake state space. The action space is defined as a combination of discretized or segmented continuous control commands for aeration equipment, ecological water regulating gate pumps, and carbon source dosing devices, and a hierarchical structure is set for the control commands in the action space. A multi-objective reward function is constructed based on water quality rewards, ecological risk rewards, economic cost rewards, and engineering resilience rewards. The Dueling DQN architecture is adopted to decompose the Q value into a state value function and an advantage function. The outputs of the state value function and the advantage function are aggregated to obtain the final Q value. The reinforcement learning agent uses the final Q value to determine the severity of the current ecological state of the lake and the governance actions that can bring the most additional improvement. During training, the temporal difference error of each experience is calculated, and the sampling priority is set according to the temporal difference error of the experience samples. By prioritizing the replay of experiences, the reinforcement learning agent learns from erroneous and unexpected experiences, thus accelerating the convergence to the optimal policy. Safety rules are set in the reinforcement learning agent after training, and the generated actions are verified using the safety rules. After verification, the optimal strategy for multi-facility coordinated control is output.

7. The integrated intelligent management and control method for ecological protection and restoration of urban shallow lakes according to claim 1, characterized in that, The optimal strategy is converted into hierarchical control instructions and sent to the corresponding execution units, including: A pre-defined semantic action dictionary maps the original action values ​​in the optimal policy output by the reinforcement learning agent to a predefined standardized operation mode, generates hierarchical control instructions, and introduces fuzzy logic smoothing, defining the critical region of the input variable as the transition region, and performing a smooth weighted transition of the instructions in the transition region. A dynamic multi-level response threshold library is constructed, and the thresholds are dynamically adjusted according to seasonality, hydrological period, and equipment status. The semantic action dictionary is also updated periodically. Hierarchical control commands are sent to each execution unit via the Industrial Internet of Things (IIoT) protocol. After the commands are sent, they are automatically linked to subsequent monitoring data, the changing trends of preset monitoring indicators are analyzed, and the results are compared with the expected results. Anomaly alarms are generated based on the comparison results.

8. The integrated intelligent management and control method for ecological protection and restoration of urban shallow lakes according to claim 1, characterized in that, The changes in the environmental state after execution are collected through a monitoring network and fed back to the deep learning model and reinforcement learning agent for closed-loop optimization, including: An intervention event log is created based on the content of the executed instructions, the start and end times, the geographical scope of the effect, and the expected governance goals. Centered on each intervention event, the pre-intervention state sequence, post-intervention response sequence, and control area data are extracted as environmental feedback data. The intervention events and environmental feedback data within a preset period are used to construct a fine-tuning task set. The deep learning models for different tasks are then fine-tuned. During the fine-tuning process, the importance of the model parameters is evaluated, and elastic constraints are applied to model parameters that exceed the importance threshold to prevent drastic changes in the parameters. The agent model in the simulator is fine-tuned using a fine-tuning task set. The intervention event is taken as input, and the environmental response sequence output by the simulator is required to be close to the actual monitored response sequence. The simulator is then calibrated. Furthermore, in the experience replay library of the reinforcement learning agent, real decision-making and execution processes are stored as special experiences and their corresponding priorities are set to the highest for periodic retraining of the reinforcement learning agent.

9. An integrated intelligent management and control system for the ecological protection and restoration of urban shallow lakes, characterized in that, The system is used to implement the integrated intelligent management and control method for ecological protection and restoration of urban shallow lakes as described in any one of claims 1-8. The system includes: a three-dimensional monitoring network unit, a multimodal data fusion and preprocessing unit, a multi-task deep learning prediction unit, a reinforcement learning intelligent decision-making unit, and a hierarchical execution and control unit. The three-dimensional monitoring network unit integrates water quality monitoring buoys for collecting water temperature, dissolved oxygen, total phosphorus, and chlorophyll a, camera units for acquiring surface and underwater images of the water body, and meteorological stations for collecting wind speed and light intensity, thereby collecting water quality parameters, water body images, and meteorological data of the target lake. The multimodal data fusion and preprocessing unit is responsible for spatiotemporal alignment and fusion of the collected multi-source heterogeneous data to generate a multimodal dataset of lake ecological status. The multi-task deep learning prediction unit analyzes the multimodal dataset and executes tasks in parallel, including dynamic pollution source tracing based on graph neural networks, generating water quality distribution maps for preset time periods based on spatiotemporal prediction models, and early warning of algal bloom probability and path based on sequence learning; the unit as a whole supports periodic optimization using an online learning mechanism. The reinforcement learning intelligent decision-making unit optimizes decisions based on prediction and early warning results through a deep Q-network reinforcement learning agent. The state space of the reinforcement learning agent contains predicted water quality and algal bloom risk information, and the action space contains control instructions for at least one of aeration equipment, ecological water regulating gate pumps, and carbon source dosing devices. Through interaction with environmental simulation, it outputs the optimal strategy for coordinated regulation of multiple facilities. The hierarchical execution and control unit transforms the generated intelligent decision-making strategy into specific executable physical actions, and performs refined control of the execution unit based on preset multi-level risk thresholds.

Images