A water body carbon flux robot observation system for algal community analysis

Abstract

The application discloses a water body carbon flux robot observation system for algal community analysis, comprising a cooperative observation module, an edge calculation and modeling module, a dynamic coupling carbon flux prediction module and a digital twin feedback optimization module. The system collects algal images and environmental parameters through multiple robots in a distributed manner, and generates a global feature set using federated learning at an edge aggregation node; the edge module ensures the real-time inference efficiency of the core model through model lightweight and calculation scheduling; the dynamic coupling module extracts features using a multi-branch network and adaptively fuses them, and outputs real-time carbon flux data and future trends; the digital twin module generates an optimized observation strategy based on the predicted trend and feeds it back to the robot, dynamically adjusting its observation path and sampling density. The application realizes a closed-loop autonomous operation from data perception, in-situ calculation, dynamic prediction to strategy optimization, significantly improving the real-time performance, adaptability and resource utilization efficiency of carbon flux monitoring.

Description

Technical Field

[0001] This invention relates to the fields of environmental monitoring, ecological analysis and robotic automatic control technology, and in particular to a robotic observation system for water carbon flux analysis for algal community analysis. Background Technology

[0002] Carbon flux in water bodies is a crucial link in the global carbon cycle, and algae, as the most important primary producers in water bodies, have a decisive influence on the processes of carbon fixation, transformation, and release in water through their community structure and physiological state. Therefore, how to achieve precise correlation and dynamic characterization between algal community characteristics and changes in water carbon flux under complex aquatic environmental conditions has always been an important technical issue in aquatic environment monitoring and ecological process research.

[0003] In practical applications, the observation of carbon flux in water bodies often requires the simultaneous acquisition of algal morphology information, physiological activity information, and multidimensional environmental parameters, followed by comprehensive analysis. However, due to factors such as the large spatial scale of water bodies, rapid changes in environmental conditions, and highly non-uniform distribution of algal communities, existing observation and analysis methods struggle to achieve a balance between accuracy, timeliness, and system synergy.

[0004] On the one hand, existing technologies typically employ separate or static analysis methods to model the correlation between algal community characteristics and carbon flux. These methods are insufficient in responding to dynamic changes in algal community structure and environmental conditions, making it difficult to characterize the nonlinear evolution of carbon flux at different time and spatial scales, thus limiting the accuracy and stability of prediction results.

[0005] On the other hand, in multi-regional or large-scale water body observation scenarios, the efficiency of data collaboration and information fusion between observation nodes is low, the data utilization methods are singular, and it is difficult to achieve efficient collaborative analysis while ensuring data security, thereby affecting the overall system's ability to characterize the spatiotemporal changes of carbon flux in water bodies.

[0006] In addition, existing systems often rely on remote centralized processing in their computing architecture, resulting in long data transmission links and slow processing response, making it difficult to meet the application requirements for real-time analysis and rapid response to carbon flux in water bodies. This problem of insufficient system adaptability is even more prominent in scenarios with rapid succession or sudden changes in algal communities.

[0007] Meanwhile, existing technologies mostly focus on output results and lack a feedback mechanism to dynamically adjust observation behavior based on analysis and prediction results. This makes it difficult to form a closed-loop operation mode of "observation-analysis-prediction-optimization", resulting in unreasonable allocation of observation resources and limited ability to capture carbon flux changes in key areas and key periods.

[0008] In summary, existing technologies in the fields of algal community analysis and water carbon flux observation still face problems such as insufficient dynamic coupling modeling capabilities, limited collaborative observation and computational efficiency, and insufficient real-time performance and adaptive optimization capabilities. There is an urgent need for a systematic technical solution that can integrate multi-node collaborative observation, edge-side intelligent analysis, and dynamic prediction feedback to improve the overall performance of water carbon flux observation and analysis. Summary of the Invention

[0009] To address the shortcomings of existing technologies, the present invention aims to provide a robotic observation system for water carbon flux analysis in algal community analysis, which significantly improves the accuracy of carbon flux prediction, the efficiency of observation resource utilization, and the overall adaptive capability of the system.

[0010] To achieve the above objectives, the present invention provides the following technical solution: a robotic observation system for water carbon flux analysis in algal communities, comprising:

[0011] The collaborative observation module includes at least three mobile observation robots and an edge aggregation node; each robot is used to collect and preprocess algae images and environmental parameters of the target water area to form a local feature vector, and upload it to the edge aggregation node; the edge aggregation node is used to aggregate the local feature vectors based on a federated learning algorithm to generate a global feature set.

[0012] The edge computing and modeling module, deployed at the edge aggregation node and connected to the collaborative observation module, includes a modeling and inference unit and a model lightweighting unit. The modeling and inference unit is used to model and infer a preset deep learning model based on the global feature set and the acquired real-time environmental parameters. The model lightweighting unit, connected to the modeling and inference unit, is used to prune and quantize the deep learning model to obtain a lightweight model.

[0013] The dynamically coupled carbon flux prediction module, connected to the edge computing and modeling module, includes a feature extraction unit and an adaptive fusion unit. The feature extraction unit is used to extract algal morphological and physiological features from the global feature set according to the lightweight model, and to extract environmental temporal features from the real-time environmental parameters. The adaptive fusion unit is connected to the feature extraction unit and is used to dynamically fuse the algal morphological and physiological features and the environmental temporal features through a gating mechanism to output real-time carbon flux data and future change trends.

[0014] The digital twin feedback optimization module, connected to the dynamically coupled carbon flux prediction module, includes a strategy optimization unit and a feedback execution unit. The strategy optimization unit is used to generate an optimized observation strategy based on the future trend. The feedback execution unit, connected to the strategy optimization unit, is used to distribute the optimized observation strategy to the corresponding robot to adjust its observation behavior.

[0015] Furthermore, in the process of the edge aggregation node aggregating each of the local feature vectors based on the federated learning algorithm to generate the global feature set, the following aggregation process is specifically executed:

[0016] The federated learning scheduling subunit is used to coordinate the execution of multiple rounds of global model aggregation iteration. In each round of iteration, it distributes the current global feature extraction model to each robot and receives the local model parameters updated by each robot after it has trained the global feature extraction model locally based on its local feature vector.

[0017] The dynamic weight allocation subunit is used to allocate corresponding aggregate weights to the local model parameters based on the real-time richness assessment value of the algal community in the observation sub-region corresponding to each robot.

[0018] The model weighted aggregation subunit connects the federated learning scheduling subunit and the dynamic weight allocation subunit, and is used to perform a weighted average of the received local model parameters according to the allocated aggregation weights, and update the global feature extraction model.

[0019] The global feature set is generated by the global feature extraction model, which performs unified feature extraction on the local feature vector after multiple rounds of iterative aggregation.

[0020] Furthermore, the modeling and inference unit includes a heterogeneous processor and a real-time scheduling subunit;

[0021] The heterogeneous processor consists of a main control processor and a programmable logic device, and is used to carry out the computational tasks of the lightweight model.

[0022] The real-time scheduling subunit is connected to the heterogeneous processor and is used to assign the highest priority to the inference computing tasks of the lightweight model and the secondary priority to the environmental parameter acquisition and data upload tasks, so as to ensure the real-time processing of the global feature set and the real-time environmental parameters.

[0023] Furthermore, the edge computing and modeling module also includes a global task scheduling unit, which connects the modeling and inference unit and the model lightweighting unit;

[0024] The global task scheduling unit is used for:

[0025] The modeling and inference task of the modeling and inference unit on the global feature set and the real-time environmental parameters is set as the highest priority task.

[0026] Set the robot data upload and regular communication tasks in the collaborative observation module as secondary priority tasks;

[0027] Based on the resource requirements of the highest priority task, computing resources are dynamically scheduled to ensure that the time taken for the modeling and inference unit to output the real-time carbon flux data in a single calculation does not exceed a preset time threshold.

[0028] Furthermore, the model lightweighting unit specifically includes:

[0029] The structured pruning subunit is used to perform structured pruning on the convolutional layers of the model based on channel importance assessment, removing channels with importance below a predetermined threshold.

[0030] A fixed-point quantization subunit, connected to the structured pruning subunit, is used to convert floating-point parameters in the pruned model into low-bit-width fixed-point integer representations.

[0031] The lightweight model, through structured pruning and fixed-point quantization, has fewer parameters and lower computational cost than the initial model before optimization.

[0032] Furthermore, the feature extraction unit includes an image feature extraction subunit and a temporal feature extraction subunit configured in parallel;

[0033] The image feature extraction subunit employs a convolutional neural network with an encoder-decoder structure to extract spatial feature vectors representing the individual morphology and community structure of algae from the algae image data contained in the global feature set.

[0034] The temporal feature extraction subunit employs a temporal convolutional network to extract feature vectors representing the dynamic changes of environmental factors from the temporal data of the real-time environmental parameters.

[0035] The image feature extraction subunit is provided with an attention selection module between the encoder and the decoder to enhance the focus on the target algal species region in the algal image.

[0036] Furthermore, the gating mechanism is implemented through a dynamic weight decision network based on reinforcement learning;

[0037] The dynamic weight decision network is trained and operates in the following manner:

[0038] Construct a system state space defined by historical and real-time algal community state time-series data and corresponding environmental parameter state time-series data;

[0039] The multidimensional fusion weight configuration between the algal morphological and physiological characteristics and the environmental temporal characteristics is defined as the executable action space.

[0040] The error between the real-time carbon flux data and the corresponding true value is mapped as a reward signal to drive learning. This drives the dynamic weight decision network to learn by interacting with the environment, so as to adaptively output a dynamic fusion weight vector, thereby realizing the dynamic coupling of the algal physiological characteristics and the environmental temporal characteristics.

[0041] Furthermore, the adaptive fusion unit includes:

[0042] The coupling state sensing subunit is used to generate a coupling state vector representing the current environment-organism interaction strength based on the currently input algal morphological and physiological characteristics and the environmental temporal characteristics.

[0043] The dynamic gating subunit, connected to the coupling state sensing subunit, is used to receive the coupling state vector and output the dynamic fusion weight vector that matches the current coupling state in real time.

[0044] The feature fusion execution subunit, connected to the dynamic gating subunit, is used to perform nonlinear weighted fusion of the input algal morphological and physiological features and the environmental temporal features based on the dynamic fusion weight vector, to generate high-dimensional coupled features for carbon flux calculation.

[0045] The dynamic gating subunit is configured to trigger the dynamic weight decision network to perform forward inference and update the dynamic fusion weight vector in real time when the rate of change of the coupled state vector exceeds an adaptive adjustment threshold.

[0046] Furthermore, the strategy optimization unit specifically includes:

[0047] The enrichment zone identification subunit is used to identify algal enrichment zones and non-enrichment zones based on the predicted algal community density and distribution in the future change trend.

[0048] The dynamic path planning subunit, connected to the enrichment area identification subunit, is used to plan an observation path using a spiral search for robots deployed in the algae-rich area, and to plan an observation path using a straight-line cruise for robots deployed in the non-rich area.

[0049] The sampling density adjustment subunit, connected to the enrichment area identification subunit and the dynamic path planning subunit, is used to increase the density of robot sampling points in the algae enrichment area to a first predetermined density, and adjust the density of robot sampling points in the non-enrichment area to a second predetermined density lower than the first predetermined density.

[0050] Furthermore, the feedback execution unit specifically includes:

[0051] The strategy verification subunit is used to deduce the strategy based on the virtual scenario constructed by the digital twin feedback optimization module and evaluate its expected observation efficiency before the optimized observation strategy is issued.

[0052] The collaborative instruction generation subunit is connected to the strategy verification subunit and is used to deconstruct the optimized observation strategy into an independent set of control instructions that matches the state of each robot after the verification is passed.

[0053] The instruction distribution and synchronization subunit is connected to the collaborative instruction generation subunit and each of the robots. It is used to distribute the independent control instruction set to the corresponding robot and receive status confirmation feedback from each robot to complete the synchronization.

[0054] The beneficial effects of this invention are:

[0055] Compared with the prior art, the present invention has at least the following beneficial effects:

[0056] 1. Significantly improve the accuracy of coupled analysis of algal community and water carbon flux: By jointly modeling the characteristics of multi-source algal communities and environmental parameters, and introducing a dynamic feature fusion mechanism, the calculation and prediction of carbon flux can adaptively reflect the changes in the structure and physiological state of algal communities under different water environment conditions, effectively overcoming the problem of insufficient adaptability of traditional static models, thereby improving the accuracy and stability of carbon flux analysis results.

[0057] 2. Achieve efficient data utilization and security under multi-regional collaborative observation: By conducting collaborative observation with multiple robots and performing feature-level aggregation and model collaboration at the edge, the system can expand the coverage of water body observation while avoiding centralized transmission and storage of raw observation data, reducing the risk of data leakage, and improving the representativeness and credibility of multi-regional observation data. It is suitable for water monitoring scenarios with high data security requirements.

[0058] 3. Improve the overall real-time performance and on-site response capability of the system: By completing model inference and data analysis at edge nodes, the dependence on remote computing platforms is reduced, the data transmission burden is lowered, and carbon flux calculation and prediction results can be generated quickly at the observation site. This enhances the system's ability to respond to rapid changes in water carbon flux and meets the needs of real-time monitoring and emergency analysis.

[0059] 4. Construct a closed-loop operation mechanism for observation, analysis, prediction, and strategy optimization: By predicting the trend of carbon flux change in water bodies, and dynamically generating and feeding back optimization observation strategies based on the prediction results, adaptive adjustment of observation behavior can be achieved, thereby improving the efficiency of observation resource allocation and enhancing the system's ability to capture carbon flux changes in key areas and key periods.

[0060] 5. Enhance the system's adaptability and long-term operation capability to complex aquatic environments: Through the coordinated design of the observation, calculation and prediction processes, the system can maintain stable operation in water bodies with large changes in environmental conditions, while taking into account energy consumption control and computational efficiency, making it suitable for long-term, continuous water carbon flux observation and analysis applications. Attached Figure Description

[0061] Figure 1 This is a schematic diagram of the water carbon flux robotic observation system used for algal community analysis in this invention;

[0062] Figure 2 This is a schematic diagram of the edge aggregation node in this invention;

[0063] Figure 3 This is a schematic diagram of the structure of the modeling and reasoning unit in this invention;

[0064] Figure 4 This is a schematic diagram of the lightweight model unit in this invention;

[0065] Figure 5 This is a schematic diagram of the feature extraction unit in this invention;

[0066] Figure 6 This is a schematic diagram of the adaptive fusion unit in this invention;

[0067] Figure 7 This is a schematic diagram of the strategy optimization unit in this invention;

[0068] Figure 8 This is a schematic diagram of the feedback execution unit in this invention.

[0069] Figure labels: 1. Collaborative observation module; 11. Edge aggregation node; 111. Federated learning scheduling subunit; 112. Dynamic weight allocation subunit; 113. Model weighted aggregation subunit; 2. Edge computing and modeling module; 21. Modeling and inference unit; 211. Heterogeneous processor; 212. Real-time scheduling subunit; 22. Model lightweighting unit; 221. Structured pruning subunit; 222. Fixed-point quantization subunit; 23. Global task scheduling unit; 3. Dynamically coupled carbon flux prediction module; 31. Feature extraction unit; 311. Graph The following are sub-units: 312 Feature Extraction Subunit; 32 Adaptive Fusion Unit; 321 Coupled State Awareness Subunit; 322 Dynamic Gating Subunit; 323 Feature Fusion Execution Subunit; 4 Digital Twin Feedback Optimization Module; 41 Strategy Optimization Unit; 411 Rich Region Identification Subunit; 412 Dynamic Path Planning Subunit; 413 Sampling Density Adjustment Subunit; 42 Feedback Execution Unit; 421 Strategy Verification Subunit; 422 Cooperative Instruction Generation Subunit; 423 Instruction Distribution Synchronization Subunit. Detailed Implementation

[0070] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Identical components are denoted by the same reference numerals. It should be noted that the terms "front," "rear," "left," "right," "upper," and "lower" used in the following description refer to directions in the accompanying drawings, and the terms "bottom surface," "top surface," "inner," and "outer" refer to directions toward or away from the geometric center of a specific component, respectively.

[0071] Example 1:

[0072] Reference Figure 1 This is an embodiment of the present invention, which provides a robotic observation system for water carbon flux analysis for algal community analysis.

[0073] I. Overall System Structure;

[0074] This embodiment provides a robotic observation system for water carbon flux analysis for algal community analysis. The system includes a collaborative observation module 1, an edge computing and modeling module 2, a dynamically coupled carbon flux prediction module 3, and a digital twin feedback optimization module 4. The modules are connected and interact with each other through data and control commands to form a closed-loop operation system for observation, analysis, prediction, and optimization.

[0075] In the specific deployment, a freshwater lake was selected as the target observation area, with an area of ​​approximately [area missing]. Within this observation area, at least three mobile observation robots are deployed; in this embodiment, five robots are specifically deployed, and an edge aggregation node 11 is set on a floating platform at the center of the observation area.

[0076] The edge aggregation node 11 is powered by a combination of solar panels and lithium batteries. The solar panels have a rated power of 200W and the lithium batteries have a capacity of 100Ah, which is used to ensure the long-term continuous operation of the edge aggregation node 11 in outdoor aquatic environments.

[0077] Each robot is evenly distributed at the edge of the observation area. The initial observation path is set to a rectangular cruise mode, and the sampling point density is set to one sampling point every 30m. This sampling point density serves as the basic threshold for the initial observation strategy of the system and is dynamically adjusted based on the prediction results.

[0078] II. Composition and working principle of collaborative observation module 1;

[0079] The collaborative observation module 1 includes at least three mobile observation robots and an edge aggregation node 11.

[0080] (a) Robot end structure and function;

[0081] Each robot is equipped with the following devices and operating environment:

[0082] Underwater or near-surface navigation platforms for autonomous movement within target waters;

[0083] The lightweight sensing unit includes an algae image acquisition device, a chlorophyll fluorescence sensor, a dissolved oxygen sensor, a water temperature sensor, and a pH sensor.

[0084] Local computing units are used to perform data preprocessing and feature extraction;

[0085] The communication unit supports dual-mode switching between 5G and LoRa communication;

[0086] The energy management unit coordinates battery and solar power supply.

[0087] During operation, each robot synchronously collects algae image data and environmental parameter data of the target water area at a sampling frequency of 1Hz.

[0088] Environmental parameters include chlorophyll fluorescence intensity, dissolved oxygen, water temperature, and pH value.

[0089] (ii) Local preprocessing and feature vector generation;

[0090] Each robot preprocesses the collected data in its local computing unit, including:

[0091] Denoising and image enhancement processes were performed on algae images to improve the recognizability of algal morphological information in the images;

[0092] Based on a preset feature extraction network structure, a 128-dimensional algae image feature vector is extracted from each frame of algae image;

[0093] Environmental parameter data are time-aligned to ensure joint analysis of data from different sensors on the same time scale.

[0094] Outlier removal is performed on the environmental parameter sequence to reduce the impact of sensor noise on subsequent modeling.

[0095] After the above processing, each robot forms a local feature vector set for the corresponding observation area, which includes algae image features and temporal features of environmental parameters.

[0096] III. Federated Learning Aggregation and Edge Computing Modeling Mechanism;

[0097] (a) The principle of data aggregation in federated learning;

[0098] Each robot uploads its local feature vector set to the edge aggregation node 11 through the communication unit, but does not upload the original image data, thereby avoiding the centralized exposure of the original data of sensitive water areas.

[0099] Edge aggregation node 11 aggregates local feature vectors from each robot based on a federated learning algorithm. In this embodiment, a federated averaging algorithm is used as the aggregation strategy, specifically including:

[0100] Edge aggregation node 11 sends initial model parameters to each robot;

[0101] Each robot updates its model locally based on local data;

[0102] Each robot only uploads model update parameters or feature statistics;

[0103] Edge aggregation node 11 performs a weighted average of the parameters from each robot.

[0104] After 20 iterations of the above aggregation process, a global feature set reflecting the characteristics of the algal community in the entire observed water area is formed. The "20 iterations" is used as an empirical threshold to ensure the convergence stability of the model in this embodiment, and is clearly stated in the specification.

[0105] (II) Working principle of edge computing and modeling module 2;

[0106] The edge computing and modeling module 2 is deployed on the edge aggregation node 11 and includes a modeling inference unit 21 and a model lightweighting unit 22.

[0107] The modeling and reasoning unit 21 models and reasons about a preset deep learning model based on the global feature set and the environmental parameters acquired in real time.

[0108] Deep learning models are used to characterize the nonlinear relationship between algal community features and environmental parameters, and output the calculation results of carbon flux in water bodies.

[0109] The model lightweight unit 22 is connected to the modeling and inference unit 21 and is used to prune and quantize the deep learning model to reduce the model parameter size and computational complexity, so as to adapt it to the low-power heterogeneous computing architecture of the edge aggregation node 11, thereby enabling the model to run in situ in real time at the edge.

[0110] IV. Working principle of dynamic coupling carbon flux prediction module 3;

[0111] The dynamically coupled carbon flux prediction module 3 is connected to the edge computing and modeling module 2, and includes a feature extraction unit 31 and an adaptive fusion unit 32.

[0112] (a) Feature extraction mechanism;

[0113] Feature extraction unit 31 is based on a lightweight model:

[0114] Morphological and physiological characteristics of algae are extracted from the global feature set to reflect the structure and activity status of algal communities.

[0115] Environmental temporal characteristics are extracted from real-time environmental parameters to reflect the dynamic changes in the physicochemical conditions of water bodies.

[0116] (ii) Adaptive fusion and prediction mechanism;

[0117] The adaptive fusion unit 32 dynamically fuses algal morphological and physiological characteristics with environmental temporal characteristics through a gating mechanism.

[0118] The gating mechanism assigns dynamic weights to different features, enabling the system to automatically adjust the contribution ratio of features according to changes in the aquatic environment, thereby improving the adaptability of carbon flux prediction to different environmental conditions.

[0119] The fused features are used to output real-time carbon flux data for the current moment and further generate information on future trends.

[0120] V. Working principle of digital twin feedback optimization module 4;

[0121] The digital twin feedback optimization module 4 is connected to the dynamically coupled carbon flux prediction module 3, and includes a strategy optimization unit 41 and a feedback execution unit 42.

[0122] (a) Digital twin simulation and trend prediction;

[0123] The strategy optimization unit 41 constructs a digital twin corresponding to the actual carbon metabolism process in water bodies based on real-time carbon flux data and future trends.

[0124] In this embodiment, an improved long short-term memory network is used to predict the trend of carbon flux change over the next 24 hours. The "24 hours" is defined as the prediction time window threshold to meet the needs of ecological monitoring, and is specified in the specification.

[0125] (ii) Generation and execution of observation strategies;

[0126] When the prediction results show that the algae density in a certain area reaches or exceeds When this occurs, the system identifies the area as a high-interest area and triggers an optimization of the observation strategy:

[0127] The sampling point density in this area was increased from one every 30m to one every 10m.

[0128] The observation path of the corresponding robot was adjusted from rectangular cruising to a spiral search path.

[0129] The feedback execution unit 42 sends the above-mentioned optimized observation strategy to the corresponding robot, adjusts its observation behavior, and thus achieves refined observation of key areas.

[0130] VI. Technical Effects and Performance Verification Description;

[0131] Through the system structure and working principle of this embodiment, the present invention can achieve the following technical effects:

[0132] By combining multi-robot collaborative observation with federated learning aggregation, the spatial representativeness of algal community characteristics can be improved.

[0133] High-precision, low-latency calculation of water carbon flux is achieved through dynamic coupling modeling on the edge side.

[0134] A closed-loop control system of observation, calculation, prediction, and optimization is formed through a digital twin feedback mechanism.

[0135] In comparative tests with single-robot and shore-based static modeling systems, this embodiment demonstrated higher carbon flux calculation accuracy, better real-time response capability, and higher utilization efficiency of observation resources, while significantly reducing the overall power consumption of the system, verifying the effectiveness and feasibility of the technical solution of this invention in practical water environment applications.

[0136] Example 2:

[0137] This embodiment is based on the water carbon flux robotic observation system for algal community analysis in Embodiment 1, and further elaborates on the internal structure, scheduling process and weight allocation mechanism of the edge aggregation node 11 in the federated learning aggregation stage.

[0138] The overall system architecture, collaborative observation module 1, edge computing and modeling module 2, dynamically coupled carbon flux prediction module 3, and digital twin feedback optimization module 4 involved in this embodiment are consistent with those in Embodiment 1. The difference lies in:

[0139] This embodiment focuses on defining the working principle and technical effect of dynamic weighted aggregation in federated learning during the global feature set generation process.

[0140] II. Federated Learning-Driven Multi-Robot Distributed Observation Layer Structure;

[0141] In this embodiment, the system includes N mobile observation robots, wherein The specific number is dynamically configured based on the observed water area and spatial complexity.

[0142] Each robot includes the following functional units and hardware devices:

[0143] (a) Lightweight sensing unit;

[0144] The lightweight sensing unit is integrated onto the robot, including:

[0145] A high-resolution optical microscope with an effective pixel count of no less than 12 million pixels is used to acquire microscopic images of algal communities.

[0146] The autofocus module is used to adjust the focal length in real time based on the turbidity of the water and the sharpness of the image, so that the system can achieve optimal focusing even in turbidity conditions. Clear images of algae can still be obtained in the aquatic environment;

[0147] Chlorophyll fluorescence sensor for acquiring parameters related to the physiological activity of algae;

[0148] Dissolved oxygen sensors, water temperature sensors, and pH sensors are used to collect water environment parameters.

[0149] The above configuration enables the simultaneous acquisition of original images, physiological parameters, and environmental parameters of algal communities, providing basic data for subsequent multi-feature fusion modeling.

[0150] (ii) Local preprocessing module;

[0151] Each robot is equipped with a local preprocessing module, which is implemented using a low-power MCU chip. In this embodiment, the STM32H7 series chip is used as an example, but it is not limited to this.

[0152] The local preprocessing module performs the following processing flow on the robot:

[0153] Algae image processing: Median filtering algorithm is used to denoise the acquired algae images in order to suppress random noise caused by suspended particles in the water.

[0154] An adaptive histogram equalization algorithm was used to enhance the image after noise reduction, thereby improving the contrast between the edges and internal structures of algal cells.

[0155] Algal feature extraction: Extracting algal morphological features from the enhanced image, including cell area, perimeter, and roundness;

[0156] Physiological characteristics reflecting the physiological state of algae were extracted by combining chlorophyll fluorescence sensor data.

[0157] Environmental parameter preprocessing: Time-series alignment of environmental parameters such as water temperature, dissolved oxygen, and pH;

[0158] use The criteria are used to remove outliers in order to reduce the impact of instantaneous sensor anomalies on the modeling results.

[0159] After the above processing, each robot generates a 128-dimensional local feature vector containing algal morphological features, physiological features, and environmental parameters. This dimension serves as the system's unified feature representation threshold and is specified in the instruction manual.

[0160] (iii) Communication module;

[0161] Each robot is equipped with a communication module, which uses a dual-mode communication method combining 5G and LoRa communication to interact with the edge aggregation node 11.

[0162] During communication:

[0163] Only upload locally preprocessed feature vectors and environmental parameter time series data;

[0164] Do not upload raw image data and raw sensor data;

[0165] This allows for the expansion of the collaborative observation range of multiple robots while ensuring the privacy and security of the observation data.

[0166] III. The working principle of the federated learning aggregation process in edge aggregation node 11;

[0167] In the process of generating the global feature set, edge aggregation node 11 performs multiple rounds of model aggregation based on the federated learning algorithm, referring to... Figure 2 Its internal components specifically include:

[0168] The federated learning scheduling subunit 111, the dynamic weight allocation subunit 112, and the model weighted aggregation subunit 113 are included.

[0169] (a) Federated Learning Scheduling Subunit 111;

[0170] Federated learning scheduling subunit 111 is used to coordinate multiple rounds of global model aggregation iteration, and its workflow is as follows:

[0171] During the initialization phase, edge aggregation node 11 generates the initial global feature extraction model;

[0172] In each round of global iteration, the current global feature extraction model is distributed to each robot;

[0173] Each robot trains the global feature extraction model locally based on its own collected local feature vectors.

[0174] In this embodiment, the following is set:

[0175] Global iteration count ;

[0176] Number of local iterations in each round ;

[0177] Model training learning rate .

[0178] The above parameters, as key thresholds for ensuring stable convergence of the model and balancing edge computing resources, are clearly stated in the specification.

[0179] After each robot completes its local training, it only uploads the updated local model parameters to the edge aggregation node 11.

[0180] (ii) Dynamic weight allocation subunit 112;

[0181] The dynamic weight allocation subunit 112 is used to allocate different aggregate weights to different local model parameters according to the algal community status in the corresponding observation sub-region of each robot.

[0182] Specifically, the weights of this sub-unit are calculated based on the following principles:

[0183] Algal richness assessment was performed on the observation sub-region corresponding to each robot.

[0184] The algal abundance assessment value is positively correlated with the weight of the robot's local model parameters in the global aggregation;

[0185] When the algal distribution in a certain sub-region is more complex or denser, its corresponding model parameters have a higher weight in the global model update.

[0186] By introducing a dynamic weight allocation mechanism, the global feature extraction model can more fully learn the algal community characteristics of regions that contribute more significantly to changes in carbon flux.

[0187] (iii) Model weighted aggregation subunit 113;

[0188] The model weighted aggregation subunit 113 connects the federated learning scheduling subunit 111 and the dynamic weight allocation subunit 112, and its main functions are:

[0189] Receive local model parameters from each robot;

[0190] Based on the aggregate weights determined by the dynamic weight allocation subunit 112, a weighted average is performed on the parameters of each local model.

[0191] Update the global feature extraction model.

[0192] After completing T rounds of global iterations, a global feature extraction model with multiple rounds of weighted aggregation is obtained.

[0193] IV. Working principle of global feature set generation and subsequent modeling;

[0194] The global feature set is formed by a global feature extraction model that has undergone multiple rounds of iterative aggregation, and then performs unified feature extraction on the local feature vectors generated by each robot.

[0195] This global feature set serves as the input to edge computing and modeling module 2, and is used for subsequent algal morphological and physiological feature extraction, environmental temporal feature analysis, and dynamic coupling carbon flux prediction.

[0196] Through the aforementioned federated learning-driven dynamic weighted aggregation mechanism, the system can achieve high-quality modeling of algal community characteristics in multiple regions without centralizing the original data, providing a reliable data foundation for the accurate calculation and prediction of water carbon flux.

[0197] V. The technical effects produced by this embodiment;

[0198] By employing the technical solution of Embodiment 2, compared to a system that does not introduce a dynamic weighted federated learning mechanism, at least the following technical effects can be achieved:

[0199] 1. Enhance the ability of the global feature set to represent the distribution of complex algal communities: By introducing a dynamic weight allocation mechanism based on algal richness, the global model pays more attention to the observation areas that contribute more to carbon flux changes, thereby improving the representativeness of the feature set.

[0200] 2. Improve the efficiency of multi-robot collaborative modeling while ensuring data privacy: Only transmit model parameters and feature vectors to avoid the security risks caused by centralized storage of raw data.

[0201] 3. Enhance the stability and convergence speed of edge-side modeling: By defining the global iteration count, local iteration count, and learning rate threshold, the model can still converge stably under the condition of limited edge computing resources.

[0202] 4. Provide high-quality input for subsequent dynamic coupling prediction of carbon flux: Provide more reliable data support for the dynamic coupling carbon flux prediction module 3 and the digital twin feedback optimization module 4.

[0203] Example 3:

[0204] This embodiment further limits the implementation of Embodiments 1 and 2, with the following key points:

[0205] The hardware architecture and task scheduling mechanism inside the edge computing and modeling module 2;

[0206] The collaborative working principle of the heterogeneous processor 211 and the real-time scheduling subunit 212 in the modeling and inference unit 21;

[0207] The model lightweight unit 22 achieves low-power real-time inference through structured pruning and fixed-point quantization, and the specific methods and technical effects thereof.

[0208] Except for the content explicitly defined in this embodiment, the structure, data source and functional flow of the remaining modules are consistent with those in Embodiment 1.

[0209] II. Overall structure of the low-power edge computing in-situ modeling layer;

[0210] In this embodiment, the edge computing and modeling module 2 is deployed within the edge aggregation node 11 to form a low-power edge computing in-situ modeling layer.

[0211] This layer uses Heterogeneous computing architectures of programmable logic devices are used to undertake modeling and inference computation tasks of global feature sets and real-time environmental parameters.

[0212] Specifically, the heterogeneous processor 211 includes:

[0213] One main control processor, using an Intel Celeron N5105 CPU, is used for model scheduling, task management, and some general-purpose computing.

[0214] A programmable logic device, selected It is used to accelerate high-parallelism computation operations in lightweight models.

[0215] Through the above heterogeneous configuration, the system is able to complete in-situ real-time inference of complex models under limited power consumption conditions.

[0216] III. Composition and working principle of modeling and reasoning unit 21;

[0217] (a) Cooperative computing mechanism of heterogeneous processor 211;

[0218] Reference Figure 3 The modeling and reasoning unit 21 includes a heterogeneous processor 211 and a real-time scheduling subunit 212.

[0219] In the actual operation process:

[0220] The CPU is responsible for executing tasks such as model control logic, feature scheduling, and inference result integration.

[0221] The FPGA is responsible for performing computationally intensive convolution operations, matrix multiplication, and feature mapping operations in the lightweight model.

[0222] By mapping different types of computing tasks to the most suitable hardware units for execution, lightweight models can achieve efficient and low-latency operation at the edge.

[0223] (II) Priority allocation principle of real-time scheduling subunit 212;

[0224] The real-time scheduling subunit 212 is connected to the heterogeneous processor 211 and is used for priority management of different tasks. Its scheduling principle is as follows:

[0225] The modeling and inference computation tasks performed by the lightweight model on the global feature set and real-time environment parameters are set as the highest priority tasks.

[0226] Set the environmental parameter acquisition task and data upload task as secondary priority tasks.

[0227] By prioritizing tasks as described above, we can ensure that carbon flux modeling and inference tasks always have priority access to computing resources when system resources are limited or tasks are concurrent, thereby guaranteeing real-time output capabilities.

[0228] IV. Real-time scheduling mechanism of edge computing and modeling module 2;

[0229] In this embodiment, the edge computing and modeling module 2 further includes a global task scheduling unit 23, which connects the modeling and inference unit 21 and the model lightweighting unit 22.

[0230] (a) Task priority management mechanism;

[0231] Global task scheduling unit 23 specifically executes the following scheduling strategies:

[0232] Set the model modeling and reasoning tasks executed by modeling and reasoning unit 21 to the highest priority;

[0233] Set the data upload and routine communication tasks of each robot in the collaborative observation module 1 to a secondary priority;

[0234] The ratio of CPU and FPGA computing resources is dynamically allocated based on the real-time resource requirements of the highest priority task.

[0235] (ii) Mechanism for guaranteeing the threshold of time consumed in a single calculation;

[0236] To meet the requirements for real-time monitoring of carbon flux in water bodies, this embodiment presupposes:

[0237] The time threshold for a single inference computation of the lightweight model does not exceed 500ms.

[0238] The global task scheduling unit 23 continuously monitors the model inference time during operation. When it approaches the time threshold, it prioritizes scheduling computing resources to ensure that the inference task is completed on time, thereby ensuring the stable output of real-time carbon flux data.

[0239] V. Structure and working principle of the lightweight unit 22 in the model;

[0240] Reference Figure 4The model lightweight unit 22 includes a structured pruning subunit 221 and a fixed-point quantization subunit 222, which are used to reduce the number of model parameters and computational load without significantly reducing the modeling accuracy.

[0241] (a) Structured pruning subunit 221;

[0242] The structured pruning subunit 221 performs pruning on the convolutional layers in the deep learning model based on channel importance assessment.

[0243] In this embodiment:

[0244] The absolute value of the weight parameters in the convolutional channels is used as an indicator of channel importance.

[0245] When the absolute value of the weight of a certain channel is lower than a predetermined threshold of 0.01, the channel is judged to have a low contribution to the model output.

[0246] Channels that meet the above conditions will be removed entirely.

[0247] This structured pruning strategy reduces the number of model parameters by more than 45% compared to before pruning, thereby significantly reducing computational complexity.

[0248] (ii) Fixed-point quantization subunit 222;

[0249] The fixed-point quantization subunit 222 is connected to the structured pruning subunit 221 and is used to quantize the pruned model parameters.

[0250] Specifically:

[0251] The 32-bit floating-point parameters in the model are uniformly converted to 8-bit fixed-point integer representation (INT8 quantization).

[0252] While ensuring that the model inference accuracy meets the requirements of carbon flux analysis, the model storage requirements and computational power consumption are significantly reduced.

[0253] The model after pruning and quantization constitutes a lightweight model, with fewer parameters and less computational cost than the initial model before optimization.

[0254] VI. Power Consumption Control and Edge Operating Environment;

[0255] To meet the long-term operation requirements of outdoor water areas, this embodiment includes a power consumption control module in the edge aggregation node 11.

[0256] This module employs dynamic voltage and frequency regulation technology.

[0257] When the system is idle, reduce the operating frequency of the CPU and FPGA to 500MHz;

[0258] In model inference computing mode, the operating frequency is increased to 1.6GHz;

[0259] The average power consumption of the system is controlled below 15W.

[0260] Through the aforementioned power consumption control mechanism, the edge aggregation node 11 can operate stably in a combined solar and lithium battery power supply environment.

[0261] VII. The technical effects produced by this embodiment;

[0262] The technical solution of Embodiment 3 can achieve at least the following technical effects:

[0263] 1. Ensure the real-time performance of carbon flux modeling and inference: Through heterogeneous computing architecture and multi-level real-time scheduling mechanism, ensure that the time consumption of a single carbon flux calculation is stably controlled within the preset threshold.

[0264] 2. Significantly reduce edge computing and energy consumption pressure: By using structured pruning and fixed-point quantization, the number of model parameters and computational load are reduced, enabling complex models to run stably on low-power edge devices.

[0265] 3. Improve the engineering feasibility of the system in complex aquatic environments: Through power consumption control and priority scheduling mechanisms, the system is suitable for long-term deployment in outdoor aquatic environments.

[0266] 4. Provide stable computing power support for dynamic coupling carbon flux prediction: Provide reliable and real-time modeling and inference capabilities for the dynamic coupling carbon flux prediction module 3 and the digital twin feedback optimization module 4.

[0267] Example 4:

[0268] This embodiment is used to specifically illustrate the parallel structure and working principle of the feature extraction unit 31 in the dynamic coupling carbon flux prediction module 3 of Embodiment 1, the modeling method of dynamic fusion based on reinforcement learning for the gating mechanism, and the real-time response mechanism of the adaptive fusion unit 32 under different environmental-biological coupling states.

[0269] The collaborative observation module 1, edge computing and modeling module 2, and digital twin feedback optimization module 4 involved in this embodiment are consistent with those in embodiments one to three, with the following differences:

[0270] This embodiment provides a detailed explanation of the dynamic adaptive coupling process of "algal community characteristics - environmental parameters - carbon flux" from the perspectives of algorithm structure and working mechanism.

[0271] II. Structure and working principle of feature extraction unit 31;

[0272] In this embodiment, refer to Figure 3The feature extraction unit 31 includes an image feature extraction subunit 311 and a temporal feature extraction subunit 312 arranged in parallel, which are used to process input data of different modalities respectively.

[0273] (a) Image feature extraction subunit 311;

[0274] The image feature extraction subunit 311 employs a convolutional neural network with an encoder-decoder structure to extract spatial features from algae image data contained in the global feature set.

[0275] In this embodiment, the convolutional neural network adopts an improved U-Net structure, and its specific working process is as follows:

[0276] Input data format: The input is algae image data aggregated by the collaborative observation module 1 and federated learning. The size of a single frame image is 256×256×3, which serves as the unified image input specification for the system.

[0277] Encoding process: The encoder part includes a 4-layer downsampling structure. Each layer uses a 3×3 convolution kernel and a stride of 2 for convolution operation to extract low-dimensional spatial features of algal individual morphology and community structure layer by layer.

[0278] Attention Selection Module: An attention selection module is set between the encoder and decoder to perform weighted enhancement on the target algal species region in the algae image.

[0279] The attention selection module automatically adjusts the weights of different spatial regions based on the intensity of feature response, making the model pay more attention to the dominant algal species regions that contribute more to carbon flux.

[0280] Decoding process and output: The decoder part includes a 4-layer upsampling structure, which uses transposed convolution to gradually restore the spatial resolution of the feature map, and finally outputs a 256-dimensional comprehensive feature vector of algae, which is used to characterize the morphological and physiological comprehensive state of algal communities.

[0281] (ii) Temporal feature extraction subunit 312;

[0282] The temporal feature extraction subunit 312 employs a temporal convolutional network (TCN) to extract dynamic environmental change features from the temporal data of real-time environmental parameters.

[0283] Its specific working method is as follows:

[0284] Input data format: The input is time-series data of environmental parameters collected in the past hour, with a sampling frequency of 1Hz and a total of 360 time points. This time window is used as the time threshold for system modeling and is clearly stated in the manual.

[0285] Temporal convolutional structure: TCN consists of 3 temporal convolutional layers, each with a kernel size of 3×1 and dilation rates of 1, 2 and 4, respectively, to capture the variation patterns of environmental parameters at different time scales.

[0286] Output features: Through the above temporal convolution processing, a 128-dimensional environmental temporal feature vector is generated to characterize the dynamic evolution trend of environmental factors such as water temperature, dissolved oxygen, and pH.

[0287] III. Gating Mechanism and Dynamic Weight Decision Based on Reinforcement Learning In this embodiment, the gating mechanism is implemented through a dynamic weight decision network based on reinforcement learning, which is used to adaptively adjust the fusion weight between algal morphological and physiological characteristics and environmental temporal characteristics.

[0288] (a) Definition of elements for reinforcement learning modeling;

[0289] The training and operation of the dynamic weight decision network are based on the following reinforcement learning elements:

[0290] State space: Defined by historical and real-time algal community state time series data and corresponding environmental parameter state time series data, it is used to reflect the comprehensive state of the algal-environment system under different water layers and different seasons.

[0291] Action space: The multidimensional fusion weight configuration between algal morphological and physiological characteristics and environmental temporal characteristics is defined as an executable action, i.e., the dynamic fusion weight vector output by the network.

[0292] Reward signal: The error between the real-time carbon flux data output by the system and the corresponding true value is used as the input of the reward function. When the prediction error decreases, a positive reward is given; otherwise, a negative reward is given.

[0293] The above definition enables the dynamic weight decision network to continuously learn the optimal feature fusion strategy through continuous interaction with the environment.

[0294] (II) The working principle of the gating mechanism;

[0295] During system operation, the dynamic weight decision network performs forward inference based on the reinforcement learning algorithm (DQN structure is used in this embodiment) and outputs the dynamic fusion weight vector at the current time.

[0296] The dynamic fusion weights are normalized using the sigmoid activation function to ensure that the weights of each feature are within a reasonable range, thereby achieving a balanced fusion of algal features and environmental features.

[0297] IV. Structure and Real-time Response Mechanism of Adaptive Fusion Unit 32;

[0298] In this embodiment, refer to Figure 6 The adaptive fusion unit 32 includes:

[0299] Coupled state perception subunit 321, dynamic gating subunit 322 and feature fusion execution subunit 323.

[0300] (a) Coupled state sensing subunit 321;

[0301] The coupled state sensing subunit 321 generates a coupled state vector representing the current environment-organism interaction strength based on the current input algal morphological and physiological characteristics and environmental temporal characteristics.

[0302] This coupled state vector is used to characterize the degree of coupling between changes in algal communities and changes in environmental factors, providing a basis for subsequent gating decisions.

[0303] (ii) Dynamic gating subunit 322;

[0304] The dynamic gating subunit 322 embeds a dynamic weight decision network to receive the coupling state vector and output a dynamic fusion weight vector that matches the current coupling state in real time.

[0305] Specifically, when the rate of change of the coupled state vector exceeds an adaptive adjustment threshold, the dynamic weight decision network is triggered to perform forward inference and update the fusion weights.

[0306] The adaptive adjustment threshold is used by the system to determine whether there has been a significant change in the environment-biological coupling relationship, and it is clearly defined in the instruction manual.

[0307] (iii) Feature fusion execution subunit 323;

[0308] The feature fusion execution subunit 323 performs nonlinear weighted fusion of algal morphological and physiological characteristics and environmental temporal characteristics based on the dynamic fusion weight vector to generate a high-dimensional coupled feature vector for carbon flux calculation.

[0309] This high-dimensional coupled eigenvector serves as the direct input for carbon flux modeling and prediction, with an output unit of mgC / (m²). 2 The carbon flux calculation results for h).

[0310] V. The technical effects produced by this embodiment;

[0311] The technical solution of Embodiment 4 can achieve at least the following technical effects:

[0312] 1. Enhance the expressive power of coupled modeling of algal communities and environmental parameters: Through parallel feature extraction and attention mechanisms, the model can accurately depict algal morphology, physiological characteristics and dynamic changes in the environment.

[0313] 2. Achieve adaptive dynamic adjustment of carbon flux prediction weights: Through a reinforcement learning-driven gating mechanism, the system can adjust the feature fusion strategy in real time for different water layers, seasons, and sudden changes.

[0314] 3. Reduce the impact of human experience rules on model accuracy: Through reward-driven learning, the feature fusion weights are adaptively determined by the data and environmental state.

[0315] 4. Provide high-quality predictive input for subsequent digital twin simulation and observation strategy optimization: Provide more accurate and stable carbon flux prediction results for digital twin feedback optimization module 4.

[0316] VI. Model formula and working mechanism of dynamic weighted decision network:

[0317] In Example 4, the gating mechanism is implemented through a dynamic weighted decision network based on reinforcement learning.

[0318] (I) The core mapping formula of dynamic weighted decision network;

[0319] During system operation, the dynamic weight decision network outputs a dynamic fusion weight vector for feature fusion based on the current coupling state vector. Its calculation process can be represented as follows: ;

[0320] in, This represents the dynamic fusion weight vector output at time t, used for weighted fusion of algal morphological and physiological characteristics with environmental temporal characteristics;

[0321] The coupled state vector input at time t is composed of the current algal morphological and physiological characteristics and environmental temporal characteristics, and corresponds to the output of the coupled state sensing subunit.

[0322] Indicates by parameters The nonlinear policy network function is used to perform feature mapping and decision reasoning on the coupled state vector;

[0323] This represents the Sigmoid activation function, used to normalize the output of the policy network.

[0324] Range of values: Since the output range of the Sigmoid function is (0,1), the dynamically fused weight vector... The range of values ​​for each element in the array is: This range constraint ensures that the weight allocation of different features in the fusion process is interpretable and stable, and avoids the unbounded amplification of the weight of a certain type of feature.

[0325] (II) Explanation of the nonlinear structure function of the policy network;

[0326] Policy network function Employing a multi-layer nonlinear mapping structure, its internal structure can be represented as:

[0327] ;

[0328] in, , These represent the weight matrices of the first and second layers, respectively. , These represent the bias vectors of the corresponding layers; This represents the ReLU activation function, used to enhance the model's ability to express nonlinear coupling relationships; This represents the hyperbolic tangent function, used to compress the numerical range of intermediate feature maps and improve training stability; This represents the complete set of trainable parameters for the policy network. By introducing two different nonlinear functions, ReLU and tanh, the policy network can simultaneously possess local feature activation capabilities and global numerical stability, thereby adapting to the complex nonlinear variation characteristics of the algal community-environment system.

[0329] (III) Reinforcement learning reward function and error-driven mechanism;

[0330] To enable the dynamic weight decision network to adaptively learn the optimal feature fusion strategy, this embodiment maps the carbon flux prediction error to a reinforcement learning reward signal, defined as follows:

[0331] ;

[0332] in, This represents the instantaneous reward value at time t; This represents the real-time carbon flux prediction value output by the system based on the current dynamic fusion weight vector, in units of... ; This represents the true or high-confidence reference carbon flux value at the corresponding time. By taking the negative of the absolute value of the prediction error as the reward signal, the reward value increases when the prediction error decreases, thereby driving the dynamic weight decision network to update its strategy in the direction of minimizing the carbon flux prediction error.

[0333] (iv) The technical effects of the model formula;

[0334] 1. The formula uses a strategy network function. By combining reinforcement learning reward mechanisms, dynamic decision-making is achieved by integrating the weights of algal morphological and physiological characteristics with environmental temporal characteristics.

[0335] 2. Dynamically fused weight vector This is the output of the dynamic gating subunit, used to guide the feature fusion execution subunit to complete nonlinear weighted fusion.

[0336] 3. By introducing the above-mentioned dynamic weight decision formula based on reinforcement learning, the system can break through the limitations of traditional static weights or artificial experience weights, and adaptively adjust the contribution ratio of algal characteristics and environmental characteristics under different water layers, seasons and sudden algal changes, thereby significantly improving the accuracy of water carbon flux prediction and environmental adaptability.

[0337] I. Overall description of Example 5;

[0338] This fifth embodiment is used to specifically illustrate the implementation of the digital twin feedback optimization module 4 in the system operation, focusing on: the principle of algae enrichment area identification based on carbon flux prediction results; the dynamic optimization mechanism of robot observation path and sampling density under different observation areas; and the digital twin inference verification and synchronous execution process of the optimized observation strategy before it is issued.

[0339] In this embodiment, the structure and function of the collaborative observation module 1, the edge computing and modeling module 2, and the dynamically coupled carbon flux prediction module 3 are consistent with those of Embodiments 1 to 4.

[0340] II. Construction and working principle of the digital twin inference and prediction layer;

[0341] In this embodiment, the digital twin feedback optimization module 4 constructs a digital twin of the algal carbon metabolism process based on the real-time carbon flux data and algal community characteristic data output by the edge aggregation node 11.

[0342] (a) Methods for constructing digital twins;

[0343] Digital twins are constructed using three-dimensional virtual modeling, and specifically include:

[0344] A 3D virtual scene of the observed water body was constructed using the Unity3D engine;

[0345] Import the terrain data of the observation area; the terrain data resolution should not exceed 1m.

[0346] Import hydrological data, including water flow velocity and direction information;

[0347] Real-time collected algal community characteristic data (including algal species and density), carbon flux data, and environmental parameter data are mapped onto the virtual scene.

[0348] Through the above mapping method, the digital twin can replicate the carbon absorption and release behavior of algae in the process of photosynthesis and respiration in real time, thereby forming a virtual operating environment that is highly consistent with the real water body.

[0349] (ii) Carbon flux change trend prediction module;

[0350] A prediction module is set up in the digital twin to predict future trends in carbon flux.

[0351] The prediction module employs an improved long short-term memory network that incorporates an attention mechanism, and its operation is as follows:

[0352] Input data definition: The input consists of carbon flux time series data, algal community characteristic data, and environmental parameter data for the past 6 hours. This time window serves as the historical data threshold for the system to make trend predictions.

[0353] Predicted output: Outputs the trend of carbon flux change and the direction of algal community succession in the next 24–48 hours, with the prediction error controlled within 15%.

[0354] The role of the attention mechanism: By dynamically strengthening the weight of key environmental factors such as light intensity and water temperature through the attention mechanism, the prediction results are more in line with the real ecological process.

[0355] III. Structure and working principle of strategy optimization unit 41;

[0356] In this embodiment, refer to Figure 7 The strategy optimization unit 41 includes:

[0357] The rich region identification subunit 411, the dynamic path planning subunit 412, and the sampling density adjustment subunit 413 are included.

[0358] (a) Enriched region identification subunit 411;

[0359] The enrichment zone identification subunit 411 divides the observation area into regions based on the predicted algal community density and spatial distribution information in the future change trend.

[0360] When the algal community density in the predicted area meets the following condition: algal density ≥ 10 6 If the number of cells / L is less than 1, the area is considered an algae-rich area; otherwise, it is considered a non-rich area.

[0361] The aforementioned algae density threshold serves as the basis for determining the differentiated allocation of observation resources in the system, and is clearly stated in the instruction manual.

[0362] (ii) Dynamic path planning sub-unit 412;

[0363] The dynamic path planning subunit 412 is connected to the rich area identification subunit 411 and is used to plan differentiated observation paths for the robot according to the area type.

[0364] Its working principle is as follows:

[0365] For robots deployed in algae-rich areas, a spiral search path is planned to improve coverage of local high-density algae areas;

[0366] For robots deployed in non-rich areas, a straight cruising path is planned to reduce redundant observations and improve overall cruising efficiency.

[0367] (iii) Sampling density adjustment subunit 413;

[0368] The sampling density adjustment subunit 413 dynamically configures the robot sampling point density based on the rich region identification results and path planning results:

[0369] Within the algae-rich area, the density of robot sampling points is increased to a first predetermined density, i.e., one sampling point is set up every 10m.

[0370] In non-rich areas, the density of robot sampling points is adjusted to a second predetermined density, which is lower than the first predetermined density, i.e., one sampling point is set every 50m.

[0371] The above methods enable the allocation of observation resources towards key areas.

[0372] IV. The structure and synchronous execution mechanism of feedback execution unit 42;

[0373] In this embodiment, refer to Figure 8 The feedback execution unit 42 includes:

[0374] The strategy verification subunit 421, the cooperative instruction generation subunit 422, and the instruction distribution synchronization subunit 423 are included.

[0375] (a) Strategy verification subunit 421;

[0376] Before the optimized observation strategy is distributed to the robot, the strategy verification subunit 421 constructs a virtual operation scenario based on a digital twin to simulate and verify the optimized observation strategy.

[0377] By simulating the robot's execution of optimized paths and sampling strategies in a virtual scenario, its expected observation efficiency, coverage, and resource consumption are evaluated. Only when the evaluation results meet the preset efficiency requirements are the instructions allowed to be issued.

[0378] (ii) Cooperative instruction generation subunit 422;

[0379] After the strategy verification is successful, the cooperative instruction generation subunit 422 deconstructs the optimized observation strategy into an independent set of control instructions that matches the current state of each robot.

[0380] The control instruction set includes, but is not limited to: path type instructions (spiral search or straight cruise); sampling point density parameters; path start point and execution order information.

[0381] (iii) Instruction distribution synchronization subunit 423;

[0382] The instruction distribution and synchronization subunit 423 distributes the independent control instruction set to the corresponding robot through the communication module and receives status confirmation feedback from each robot.

[0383] Once it is confirmed that all target robots have completed receiving instructions and entered the execution state, the observation strategy for this round is updated synchronously, thereby achieving system-level collaborative observation optimization.

[0384] V. The technical effects produced by this embodiment;

[0385] The technical solution in Embodiment 5 can achieve at least the following technical effects:

[0386] 1. Achieve forward-looking optimization of observation strategies based on prediction results: Through digital twin simulation and trend prediction, the observation strategy is transformed from a passive response to a forward-looking adjustment.

[0387] 2. Significantly improve the observation hit rate of algae-rich areas: Improve the observation accuracy of high-value areas through spiral search path and high-density sampling configuration.

[0388] 3. Reduce observation resource consumption in non-critical areas: Reduce redundant observations and improve overall resource utilization efficiency through straight-line cruise and low-density sampling.

[0389] 4. Ensure the security and executability of the observation strategy: Through digital twin verification and synchronous execution mechanisms, avoid unreasonable strategies from being directly applied to the real system.

[0390] 5. Establish a complete "prediction-optimization-execution" closed-loop control system: providing continuous self-optimization capability for the water carbon flux robotic observation system.

[0391] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principle of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A robotic observation system for water carbon flux analysis in algal communities, characterized in that, include: The collaborative observation module includes at least three mobile observation robots and an edge aggregation node; Each robot is used to collect and preprocess algae images and environmental parameters of the target water area to form local feature vectors, and upload them to the edge aggregation node; the edge aggregation node is used to dynamically weight and aggregate each local feature vector based on a federated learning algorithm to generate a global feature set; The edge computing and modeling module, deployed at the edge aggregation node and connected to the collaborative observation module, includes a modeling and inference unit and a model lightweighting unit. The modeling and inference unit is used to input the global feature set and the acquired real-time environmental parameters into a preset deep learning model for inference to obtain an intermediate feature representation for carbon flux prediction. The model lightweighting unit, connected to the modeling and inference unit, is used to prune and quantize the deep learning model to obtain a lightweight model. A dynamically coupled carbon flux prediction module, connected to the edge computing and modeling module, includes a feature extraction unit and an adaptive fusion unit. The feature extraction unit comprises a parallel image feature extraction subunit and a temporal feature extraction subunit. The image feature extraction subunit extracts algal morphological and physiological features from the global feature set based on the lightweight model. The temporal feature extraction subunit extracts environmental temporal features from the real-time environmental parameters. The adaptive fusion unit includes a coupling state perception subunit, a dynamic gating subunit, and a feature fusion execution subunit. The dynamic gating subunit is implemented through a reinforcement learning-based dynamic weight decision network and outputs a dynamic fusion weight vector based on the coupling state vector generated by the coupling state perception subunit. The feature fusion execution subunit performs nonlinear weighted fusion of the algal morphological and physiological features and the environmental temporal features based on the dynamic fusion weight vector, outputting real-time carbon flux data and future trends. The digital twin feedback optimization module, connected to the dynamically coupled carbon flux prediction module, includes a strategy optimization unit and a feedback execution unit. The strategy optimization unit is used to generate an optimized observation strategy based on the future trend. The feedback execution unit, connected to the strategy optimization unit, is used to distribute the optimized observation strategy to the corresponding robot to adjust its observation behavior.

2. The water carbon flux robotic observation system for algal community analysis according to claim 1, characterized in that, In the process of the edge aggregation node aggregating the local feature vectors based on the federated learning algorithm to generate the global feature set, the following aggregation process is specifically executed: The federated learning scheduling subunit is used to coordinate the execution of multiple rounds of global model aggregation iteration. In each round of iteration, it distributes the current global feature extraction model to each robot and receives the local model parameters updated by each robot after it has trained the global feature extraction model locally based on its local feature vector. The dynamic weight allocation subunit is used to allocate corresponding aggregate weights to the local model parameters based on the real-time richness assessment value of the algal community in the observation sub-region corresponding to each robot. The model weighted aggregation subunit connects the federated learning scheduling subunit and the dynamic weight allocation subunit, and is used to perform a weighted average of the received local model parameters according to the allocated aggregation weights, and update the global feature extraction model. The global feature set is generated by the global feature extraction model, which performs unified feature extraction on the local feature vector after multiple rounds of iterative aggregation.

3. The water carbon flux robotic observation system for algal community analysis according to claim 1, characterized in that, The modeling and inference unit includes a heterogeneous processor and a real-time scheduling subunit; The heterogeneous processor consists of a main control processor and a programmable logic device, and is used to carry out the computational tasks of the lightweight model. The real-time scheduling subunit is connected to the heterogeneous processor and is used to assign the highest priority to the inference computing tasks of the lightweight model and the secondary priority to the environmental parameter acquisition and data upload tasks, so as to ensure the real-time processing of the global feature set and the real-time environmental parameters.

4. The water carbon flux robotic observation system for algal community analysis according to claim 1, characterized in that, The edge computing and modeling module also includes a global task scheduling unit, which connects the modeling and inference unit and the model lightweighting unit; The global task scheduling unit is used for: The modeling and inference task of the modeling and inference unit on the global feature set and the real-time environmental parameters is set as the highest priority task. Set the robot data upload and regular communication tasks in the collaborative observation module as secondary priority tasks; Based on the resource requirements of the highest priority task, computing resources are dynamically scheduled to ensure that the time taken for the modeling and inference unit to output the real-time carbon flux data in a single calculation does not exceed a preset time threshold.

5. The water carbon flux robotic observation system for algal community analysis according to claim 1, characterized in that, The lightweight model unit specifically includes: The structured pruning subunit is used to perform structured pruning on the convolutional layers of the model based on channel importance assessment, removing channels with importance below a predetermined threshold. A fixed-point quantization subunit, connected to the structured pruning subunit, is used to convert floating-point parameters in the pruned model into low-bit-width fixed-point integer representations. The lightweight model, through structured pruning and fixed-point quantization, has fewer parameters and lower computational cost than the initial model before optimization.

6. The water carbon flux robotic observation system for algal community analysis according to claim 1, characterized in that, The feature extraction unit includes a parallel image feature extraction subunit and a temporal feature extraction subunit; The image feature extraction subunit employs a convolutional neural network with an encoder-decoder structure to extract spatial feature vectors representing the individual morphology and community structure of algae from the algae image data contained in the global feature set. The temporal feature extraction subunit employs a temporal convolutional network to extract feature vectors representing the dynamic changes of environmental factors from the temporal data of the real-time environmental parameters. The image feature extraction subunit is provided with an attention selection module between the encoder and the decoder to enhance the focus on the target algal species region in the algal image.

7. The water carbon flux robotic observation system for algal community analysis according to claim 1, characterized in that, The dynamic gating subunit is implemented through a dynamic weight decision network based on reinforcement learning; The dynamic weight decision network is trained and operates in the following manner: Construct a system state space defined by historical and real-time algal community state time-series data and corresponding environmental parameter state time-series data; The multidimensional fusion weight configuration between the algal morphological and physiological characteristics and the environmental temporal characteristics is defined as the executable action space. The error between the real-time carbon flux data and the corresponding true value is mapped as a reward signal to drive learning. This drives the dynamic weight decision network to learn by interacting with the environment, so as to adaptively output a dynamic fusion weight vector, thereby realizing the dynamic coupling of the algal physiological characteristics and the environmental temporal characteristics.

8. The water carbon flux robotic observation system for algal community analysis according to claim 7, characterized in that, The adaptive fusion unit includes: The coupling state sensing subunit is used to generate a coupling state vector representing the current environment-organism interaction strength based on the currently input algal morphological and physiological characteristics and the environmental temporal characteristics. The dynamic gating subunit, connected to the coupling state sensing subunit, is used to receive the coupling state vector and output the dynamic fusion weight vector that matches the current coupling state in real time. The feature fusion execution subunit, which is embedded with the dynamic weight decision network and connected to the dynamic gating subunit, is used to perform nonlinear weighted fusion of the input algal morphological and physiological features and the environmental temporal features based on the dynamic fusion weight vector, so as to generate high-dimensional coupled features for carbon flux calculation. The dynamic gating subunit is configured to trigger the dynamic weight decision network to perform forward inference and update the dynamic fusion weight vector in real time when the rate of change of the coupled state vector exceeds an adaptive adjustment threshold.

9. The water carbon flux robotic observation system for algal community analysis according to claim 1, characterized in that, The strategy optimization unit specifically includes: The enrichment zone identification subunit is used to identify algal enrichment zones and non-enrichment zones based on the predicted algal community density and distribution in the future change trend. The dynamic path planning subunit, connected to the enrichment area identification subunit, is used to plan an observation path using a spiral search for robots deployed in the algae-rich area, and to plan an observation path using a straight-line cruise for robots deployed in the non-rich area. The sampling density adjustment subunit, connected to the enrichment area identification subunit and the dynamic path planning subunit, is used to increase the density of robot sampling points in the algae enrichment area to a first predetermined density, and adjust the density of robot sampling points in the non-enrichment area to a second predetermined density lower than the first predetermined density.

10. The robotic observation system for water carbon flux analysis for algal community analysis according to claim 1 or 9, characterized in that, The feedback execution unit specifically includes: The strategy verification subunit is used to deduce the strategy based on the virtual scenario constructed by the digital twin feedback optimization module and evaluate its expected observation efficiency before the optimized observation strategy is issued. The collaborative instruction generation subunit is connected to the strategy verification subunit and is used to deconstruct the optimized observation strategy into an independent set of control instructions that matches the state of each robot after the verification is passed. The instruction distribution and synchronization subunit is connected to the collaborative instruction generation subunit and each of the robots. It is used to distribute the independent control instruction set to the corresponding robot and receive status confirmation feedback from each robot to complete the synchronization.

Images