Edge calculation optimization method of seawater parameter monitoring Internet of Things system

By using edge computing optimization methods, an energy consumption evaluation function and a dynamic energy consumption threshold judgment mechanism are constructed. A lightweight model is deployed at the sensing node for preliminary processing, which solves the problem of the imbalance between energy consumption and computing efficiency in the seawater parameter monitoring IoT system, and realizes the optimization of system energy consumption and the improvement of monitoring reliability.

CN120881103APending Publication Date: 2025-10-31青岛道万科技有限公司
View PDF 0 Cites 3 Cited by

Patent Information

Application Number
CN202511103429.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-08-07
Publication Date
2025-10-31

AI Technical Summary

Technical Problem

Existing IoT systems for monitoring seawater parameters face challenges in balancing energy consumption and computing efficiency. The traditional full-data upload mode results in excessive energy consumption of sensor nodes, short battery life, high maintenance costs, and serious latency and packet loss issues during data transmission.

Method used

An edge computing optimization method is adopted to construct an energy consumption evaluation function and a dynamic energy consumption threshold judgment mechanism. A lightweight model is deployed at the sensing node for preliminary processing, and the complete model is executed at the computing center for result verification and fusion. A dynamic balance between energy consumption and computing efficiency is achieved through multi-objective optimization theory.

Benefits of technology

It significantly reduces the energy consumption of sensor nodes, reduces data transmission volume, lowers communication energy consumption, extends the lifespan of sensor networks, and improves the reliability and continuity of seawater parameter monitoring.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120881103A_ABST
    Figure CN120881103A_ABST
Patent Text Reader

Abstract

The invention provides an edge calculation optimization method for a seawater parameter monitoring Internet of Things system, and belongs to the technical field of edge calculation, and the method comprises the steps: constructing a seawater parameter monitoring sensor network, and calculating an energy consumption evaluation reference value based on a data change rate and a disorder degree; performing feature extraction on the seawater parameter data; deploying a double-model collaborative architecture comprising a terminal lightweight model and a center complete model; setting a dynamic energy consumption threshold judgment mechanism to allocate calculation tasks; executing lightweight edge calculation at the sensing nodes; transmitting the processed data through a low-power wide area network; the calculation center verifies and fuses the data; constructing a seawater parameter space-time distribution model; dynamically adjusting an energy consumption evaluation reference value by using a Pareto optimal solution set method; and periodically evaluating system energy consumption efficiency and updating model parameters. According to the method, dynamic balance of energy consumption and calculation efficiency of the seawater parameter monitoring system is realized, and the continuous operation capability and the monitoring reliability of the system are remarkably improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of edge computing technology, and more specifically, relates to an edge computing optimization method for a seawater parameter monitoring Internet of Things system. Background Technology

[0002] Seawater parameter monitoring is a crucial foundation for marine ecological environment protection and marine resource development and utilization. Traditional seawater parameter monitoring systems typically use sensor networks to collect data on parameters such as seawater temperature, salinity, dissolved oxygen, and turbidity, and then transmit all the data to a central server for processing and analysis. This centralized processing architecture is widely used in large-scale application scenarios with high data acquisition frequency and wide monitoring range.

[0003] However, marine environmental monitoring faces unique challenges: sensor nodes are mostly battery-powered, resulting in limited energy resources; monitoring points are widely distributed and far from land, leading to high data transmission costs; and the complex and variable marine environment makes data acquisition and transmission susceptible to interference. In this context, the traditional full-data upload processing model results in excessively high energy consumption, short battery life, and high maintenance costs for sensor nodes, while latency and packet loss during data transmission negatively impact monitoring effectiveness.

[0004] Although some studies have attempted to introduce edge computing technology to optimize marine monitoring systems, most existing solutions adopt a static task partitioning strategy, failing to dynamically adjust the data processing location based on the characteristics of seawater parameter data and network conditions. This makes it difficult to achieve optimal allocation of edge computing and communication resources in energy-constrained environments, resulting in technical problems in balancing energy consumption and computing efficiency in seawater parameter monitoring IoT systems. Summary of the Invention

[0005] In view of this, the present invention provides an edge computing optimization method for a seawater parameter monitoring Internet of Things (IoT) system, which can solve the technical problem of balancing energy consumption and computing efficiency in the existing seawater parameter monitoring IoT system.

[0006] This invention is implemented as follows: It provides an edge computing optimization method for a seawater parameter monitoring IoT system, comprising: constructing a seawater parameter monitoring sensor network, collecting seawater parameter data, calculating edge computing energy consumption and data transmission energy consumption using an energy consumption evaluation function based on the rate of change and disorder of the collected data, and forming an energy consumption evaluation benchmark value; analyzing the collected seawater parameter data, performing data feature extraction based on the data acquisition frequency and data change rate; deploying a seawater parameter intelligent processing edge computing model into the system, deploying a lightweight terminal model at each sensor acquisition node, and deploying a complete central model at the computing center; setting a dynamic energy consumption threshold judgment mechanism, determining whether each type of seawater parameter data is suitable for edge computing processing or suitable for transmission to the computing center based on the energy consumption evaluation benchmark value; for seawater parameter data suitable for edge computing processing, executing a lightweight edge computing task at the sensor acquisition node, and transmitting the processed data and the unprocessed original data to the computing center respectively; the computing center performing data fusion processing, constructing a seawater parameter spatiotemporal distribution model, updating the energy consumption evaluation benchmark value using multi-objective optimization theory, periodically evaluating the system's energy consumption efficiency, and optimizing the computing strategy.

[0007] The seawater parameter data includes seawater temperature, seawater salinity, seawater dissolved oxygen, and seawater turbidity.

[0008] The energy consumption assessment function refers to a mathematical function that calculates edge computing energy consumption and data transmission energy consumption based on the rate of change and disorder of seawater parameter data, as well as the computing power and network status of sensor nodes. The data processing method is determined by comparing the two energy consumption levels.

[0009] The dynamic energy consumption threshold judgment mechanism refers to an algorithm mechanism that automatically adjusts the edge computing and data transmission decision thresholds based on real-time network conditions, battery power, and data importance.

[0010] The aforementioned seawater parameter spatiotemporal distribution model refers to a multidimensional data model constructed by combining geographical location information and time series data to reflect the variation patterns of seawater parameters at different locations and times.

[0011] The multi-objective optimization theory involves constructing an energy consumption calculation delay balance function and using the Pareto optimal solution set method to update the calculation results of edge computing energy consumption and data transmission energy consumption.

[0012] The energy consumption and computation delay balancing function is used to achieve the optimal balance between energy consumption and computation delay during edge computing resource allocation. The inputs include edge computing energy consumption, data transmission energy consumption, data volume, network bandwidth status, and computation task complexity. The outputs are decision values ​​for various seawater parameter data processing strategies and corresponding resource allocation schemes.

[0013] The intelligent edge computing model for seawater parameter processing is a hybrid architecture combining a multilayer perceptron and a convolutional neural network. It includes a feature extraction layer, a temporal correlation analysis layer, a parameter correlation analysis layer, and a decision output layer. An attention mechanism is introduced into the network, and the weight coefficients are dynamically adjusted based on three key parameters: energy consumption assessment benchmark, data change rate, and network transmission status.

[0014] The terminal lightweight model refers to a simplified version of the intelligent edge computing model for water parameter processing deployed on the sensing and acquisition node. It uses depthwise separable convolution to replace the standard convolution operation, removes complex recursive structures, retains the core feature extraction function, and is optimized for energy-constrained scenarios.

[0015] The central complete model refers to the intelligent seawater parameter processing edge computing model deployed in the computing center with complete computing capabilities. It includes all network layer structures and parameters, performs complex data verification, fusion and analysis tasks, and is also responsible for providing parameter updates and optimization guidance to the terminal lightweight model.

[0016] Compared with existing technologies, this invention provides an edge computing optimization method for a seawater parameter monitoring IoT system. This method proposes an edge computing optimization approach for a seawater parameter monitoring IoT system by constructing an energy consumption evaluation function and dynamically determining data processing strategies based on the rate of change and disorder of seawater parameter data, thereby achieving collaborative optimization between edge computing and central computing. This method deploys a dual-model collaborative architecture, deploying a lightweight model on sensor nodes to perform initial data processing, and deploying the complete model on the computing center to perform result verification and fusion.

[0017] Compared to traditional full data transmission, this invention employs a dynamic energy consumption threshold judgment mechanism to intelligently determine which data processing tasks are suitable for execution at the edge, significantly reducing the energy consumption of sensor nodes. After edge computing processing, the amount of data transmitted is reduced, and communication energy consumption is lowered. Simultaneously, the Pareto optimal solution set method is used to dynamically adjust the edge computing strategy, ensuring that the system maintains an optimal balance between energy consumption and computational efficiency under different operating conditions.

[0018] This invention successfully solves the technical problem of balancing energy consumption and computing efficiency in seawater parameter monitoring IoT systems. Through a dynamic computing resource allocation strategy, it achieves overall system energy consumption optimization, extends the lifespan of the sensor network, and improves the reliability and continuity of seawater parameter monitoring. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0021] like Figure 1 The diagram shows a flowchart of an edge computing optimization method for a seawater parameter monitoring IoT system provided by this invention. This method includes the following steps:

[0022] S01. Construct a seawater parameter monitoring sensor network to collect seawater parameter data such as seawater temperature, seawater salinity, seawater dissolved oxygen, and seawater turbidity. Calculate edge computing energy consumption and data transmission energy consumption using an energy consumption assessment function based on the rate of change and disorder of the collected data, and form an energy consumption assessment benchmark value.

[0023] S02. Analyze the collected seawater parameters such as seawater temperature, seawater salinity, seawater dissolved oxygen, and seawater turbidity, and extract data features based on the data collection frequency and data change rate.

[0024] S03. Set up a dynamic energy consumption threshold judgment mechanism to determine whether each type of seawater parameter data is suitable for edge computing processing or suitable for transmission to the computing center based on the energy consumption assessment benchmark value.

[0025] S04. For seawater parameter data suitable for edge computing processing, perform lightweight edge computing tasks such as downsampling, mean filtering, and outlier detection at the sensing nodes.

[0026] S05. Pack the feature data processed by edge computing and the unprocessed raw data separately, and transmit them to the computing center using a low-power wide area network protocol.

[0027] S06. After receiving the data, the computing center first verifies the integrity of the edge computing processing results, and then performs fusion processing with the unprocessed original collected data.

[0028] S07. Construct a spatiotemporal distribution model of seawater parameters based on the fused data, and generate a global view of the water quality status of the monitored sea area.

[0029] S08. Construct an energy consumption calculation delay balance function using multi-objective optimization theory, update the calculation results of edge computing energy consumption and data transmission energy consumption through the Pareto optimal solution set method, realize dynamic adjustment of energy consumption assessment benchmark value, and continuously optimize edge computing strategy.

[0030] S09. Periodically assess the overall energy efficiency of the system, and reclassify the seawater parameter data categories suitable for edge computing and center computing based on the assessment results. Apply the classification results to the dynamic energy consumption threshold judgment mechanism.

[0031] Among them, the energy consumption assessment function is a mathematical function that calculates the edge computing energy consumption and data transmission energy consumption based on the rate of change and disorder of seawater parameter data, as well as the computing power and network status of sensor nodes. The data processing method is determined by comparing the magnitude of the two energy consumptions.

[0032] Among them, the energy consumption assessment benchmark value refers to the benchmark standard of electrical energy consumed by sensor nodes in data processing and data transmission under different network conditions and computing loads, which is used to determine the data processing method.

[0033] Among them, data feature extraction refers to extracting key features from raw seawater parameter data through statistical analysis, time-frequency domain transformation and other methods, so as to reduce data redundancy and reduce the burden of subsequent processing.

[0034] Among them, the dynamic energy consumption threshold judgment mechanism refers to the algorithm mechanism that automatically adjusts the decision thresholds for edge computing and data transmission based on real-time network conditions, battery power, and data importance.

[0035] Among them, Low Power Wide Area Network (LPWAN) protocol refers to long-distance, low-power communication technology designed for Internet of Things (IoT) applications to ensure data transmission stability in marine environments.

[0036] Among them, the seawater parameter spatiotemporal distribution model refers to a multidimensional data model constructed by combining geographical location information and time series data, which reflects the variation pattern of seawater parameters at different locations and times.

[0037] The energy consumption and computational latency balancing function is used to achieve the optimal balance between energy consumption and computational latency during edge computing resource allocation. Inputs include edge computing energy consumption, data transmission energy consumption, data volume, network bandwidth, and computational task complexity. Outputs are processing strategy decision values ​​for various seawater parameter data and corresponding resource allocation schemes. The edge computing energy consumption and data transmission energy consumption are derived from the calculation results of the energy consumption evaluation function. The data volume is derived from the collected seawater parameter data. The network bandwidth is derived from real-time monitoring of the Low Power Wide Area Network (LPWAN) protocol. The computational task complexity is derived from the resource consumption assessment of lightweight edge computing tasks. The processing strategy decision values ​​are used for a dynamic energy consumption threshold judgment mechanism, and the resource allocation scheme is used for scheduling computational resources of sensor nodes.

[0038] The specific structure of the intelligent seawater parameter processing network is a hybrid architecture combining a multilayer perceptron and a convolutional neural network. It includes a feature extraction layer, a temporal correlation analysis layer, a parameter correlation analysis layer, and a decision output layer. The feature extraction layer uses lightweight convolutional kernels to achieve data dimensionality reduction. The temporal correlation analysis layer uses long short-term memory units to capture the temporal variation patterns of parameters. The parameter correlation analysis layer extracts the intrinsic correlation between different seawater parameters through residual connections. The decision output layer determines the optimal edge computing strategy based on the aforementioned analysis results. An attention mechanism is introduced into the network, and the weight coefficients are dynamically adjusted through three key parameters: energy consumption assessment benchmark, data change rate, and network transmission status.

[0039] The steps for establishing the training dataset in the pre-training process of the intelligent seawater parameter processing network specifically include: collecting historical seawater parameter data covering different seasons and weather conditions from multiple marine monitoring stations; performing standardization processing on the collected data to eliminate dimensional differences; labeling the optimal processing strategy for each type of data under different environmental conditions based on expert knowledge; generating abnormal samples in edge cases through data augmentation technology; and finally constructing a paired training sample set containing input features and optimal decision labels.

[0040] The pre-training steps of the intelligent seawater parameter processing network specifically include: first, pre-training the network using unsupervised learning to learn the inherent distribution characteristics of seawater parameter data; then, performing supervised fine-tuning training based on labeled data to improve decision accuracy; next, implementing adversarial training to enhance the model's robustness in response to environmental fluctuations; then, performing knowledge distillation to compress the model size to meet the deployment requirements of edge devices; and finally, continuously optimizing model performance through online learning in real-world scenarios. Throughout the training process, an early stopping strategy is used to avoid overfitting and to save the model parameters that perform best on the validation set.

[0041] Among them, the Pareto optimal solution set method refers to finding a set of solutions in multi-objective optimization problems such that any improvement in any objective will inevitably lead to the deterioration of other objectives, thereby finding the optimal balance between edge computing energy consumption and computational efficiency.

[0042] Among them, the data change rate refers to the magnitude of change of seawater parameter data per unit time, which is used to evaluate the dynamic characteristics of the data and the processing priority.

[0043] Among them, disorder refers to the degree of irregularity and randomness of seawater parameter data, reflecting the complexity and processing difficulty of the data.

[0044] The specific implementation methods of the above steps are described in detail below.

[0045] The specific implementation of step S01 involves deploying a multi-layered seawater parameter monitoring sensor network in the monitored sea area to form a densely covered monitoring array. This sensor network consists of four types of sensors: a temperature sensor using a platinum resistance temperature sensor, with a measurement range of -5℃ to 40℃ and an accuracy of ±0.1℃; a salinity sensor using the conductivity method, with a measurement range of 0 to 40‰ and an accuracy of ±0.1‰; a dissolved oxygen sensor using the fluorescence quenching method, with a measurement range of 0 to 20 mg / L and an accuracy of ±0.2 mg / L; and a turbidity sensor using the scattered light measurement method, with a measurement range of 0 to 100 NTU and an accuracy of ±1 NTU. The sensor nodes are deployed in a layered manner, with several nodes arranged in the surface, middle, and bottom layers, and the horizontal spacing between adjacent nodes ranging from 100m to 500m. Each sensor node is equipped with an ARM Cortex-M4 microprocessor with a main frequency of 80MHz, 4MB of built-in Flash storage and 512KB of RAM, and runs a real-time operating system. The energy consumption assessment function uses a weighted summation method, calculating the total energy consumption as a weighted sum of edge computing energy consumption and data transmission energy consumption, with a weighting coefficient sum of 1. Edge computing energy consumption is calculated based on data volume, data complexity, and processing time, multiplied by the device energy efficiency coefficient; data transmission energy consumption is calculated based on data transmission volume and network conditions, multiplied by the network transmission energy efficiency coefficient. The data change rate is calculated as a percentage based on the difference in values ​​between adjacent sampling points, with a threshold set at 2%; data disorder is determined using an entropy calculation method, with a threshold set at 0.6. The above calculations yield a baseline value for energy consumption assessment, which serves as the basis for subsequent processing decisions.

[0046] The specific implementation of step S02 involves feature extraction processing of the collected seawater parameter data. First, based on the physical characteristics of different parameters, differentiated sampling frequencies are set: temperature data is sampled every 10 minutes, salinity data every 15 minutes, dissolved oxygen data every 5 minutes, and turbidity data every 3 minutes. The data change rate is calculated using the sliding window method, with a window size of 12 sampling points. For parameters with slow changes, such as temperature and salinity, wavelet transform is applied for time-frequency domain analysis to extract low-frequency features; for parameters with rapid changes, such as dissolved oxygen and turbidity, fast Fourier transform is applied to extract spectral features. Principal component analysis is used for dimensionality reduction, retaining feature vectors with a contribution rate of 95%. For periods where the data change rate exceeds the threshold (temperature 2%, salinity 1.5%, dissolved oxygen 5%, turbidity 8%), the sampling frequency is increased and fine-grained feature extraction is performed. Statistical feature extraction includes calculating statistics such as mean, variance, skewness, kurtosis, and quartiles; time-domain feature extraction includes trend components, periodic components, and autocorrelation coefficients; frequency-domain feature extraction includes power spectral density and dominant frequency components. These methods extract key features from the original data, reducing data redundancy by over 96%, laying the foundation for subsequent processing.

[0047] The specific implementation of step S03 involves deploying a pre-trained intelligent edge computing model for seawater parameter processing into the system. A lightweight terminal model is deployed at the sensor acquisition nodes, using quantization compression technology to convert model parameters from 32-bit floating-point numbers to 8-bit integers, keeping the model size within 200KB. Model deployment employs an incremental update approach, dividing the model into a fixed base layer and an updatable top layer. The base layer is stored in ROM, and the top layer is stored in RAM for dynamic updates. The central complete model is deployed at the computing center, using a distributed computing architecture that supports GPU acceleration. The dual-model collaborative workflow is as follows: the lightweight terminal model is responsible for initial data screening, feature extraction, and emergency anomaly detection, transmitting the processing results to the computing center via a low-power wide-area network; the central complete model receives the processing results and performs deep analysis, cross-parameter correlation mining, and global state evaluation. The models complement each other through differentiated training objectives: the terminal model optimizes energy efficiency and real-time response, while the central model optimizes analysis accuracy and global view. A feedback mechanism is established between the two, with the central model periodically pushing parameter updates to the terminal model to improve its adaptability.

[0048] The specific implementation of step S04 involves establishing a dynamic energy consumption threshold judgment mechanism. This mechanism, based on energy consumption assessment benchmarks, sets initial thresholds for each type of seawater parameter: temperature at 3 mJ / bit, salinity at 2.5 mJ / bit, dissolved oxygen at 3.2 mJ / bit, and turbidity at 3.8 mJ / bit. The judgment process is implemented using a fuzzy logic controller. Input variables include data change rate, data disorder, the proportion of remaining node power, and network signal strength. The output is the processing decision value. The fuzzy rule design principles are: when the node's remaining power is higher than 70%, computational tasks are tended to be performed at the edge; when the network signal strength is higher than -80 dBm, data is tended to be transmitted to the computing center; when the data change rate is higher than the set threshold, processing is prioritized at the edge to reduce communication load; when the data disorder is higher than 0.6, data is prioritized for transmission to the computing center for complex analysis. The judgment mechanism employs an adaptive adjustment strategy, updating the threshold parameters hourly. The new threshold is calculated using exponential smoothing, combining the old threshold with the currently calculated optimal threshold, with a smoothing coefficient of 0.3.

[0049] The specific implementation of step S05 involves performing lightweight edge computing tasks at the sensor acquisition node. For seawater parameter data deemed suitable for edge computing processing, three types of processing tasks are performed through the terminal's lightweight model. First, downsampling is performed using an adaptive sampling algorithm based on Shannon's sampling theorem. The sampling rate is dynamically adjusted according to the signal spectrum characteristics. For parameters with slow changes, such as temperature and salinity, the sampling rate can be reduced to 25% of the original sampling rate; for parameters with rapid changes, such as dissolved oxygen and turbidity, the sampling rate is maintained at over 60% of the original sampling rate. Next, mean filtering is performed using a weighted moving average algorithm. The window size is set to 5 to 15 sampling points depending on the parameter type, and the weight coefficients are distributed exponentially, with the newest data having the highest weight. Finally, outlier detection is performed using a modified Z-Score algorithm. Outliers are identified as outliers when the absolute value of the standardized score is greater than 3, and local linear interpolation is used to correct them. All of the above processing is completed on the node's microprocessor, and the processing results and the original data are packaged and stored separately, awaiting transmission.

[0050] The specific implementation of step S06 involves data packaging and transmission. The edge computing-processed feature data and the unprocessed raw collected data are packaged separately. The feature data packet structure includes: header (8 bytes), node ID (4 bytes), timestamp (8 bytes), parameter type (2 bytes), feature value (32 bytes), processing flag (2 bytes), and checksum (4 bytes). The raw data packet structure includes: header (8 bytes), node ID (4 bytes), timestamp (8 bytes), parameter type (2 bytes), original value (variable length), and checksum (4 bytes). Data compression uses differential coding combined with Huffman coding, achieving a compression rate of over 60%. Transmission uses low-power wide-area network protocols, including LoRaWAN, NB-IoT, or Sigfox, selecting the most suitable protocol based on the deployment environment. LoRaWAN is suitable for offshore monitoring, with a communication range of up to 15km and a data rate of 0.3-50kbps; NB-IoT is suitable for nearshore monitoring, with a communication range of 1-10km and a data rate of 60-100kbps; Sigfox is suitable for ultra-low power scenarios, with a communication range of up to 50km, but a data rate of only 100bps. A differentiated transmission strategy is adopted: emergency data (such as sudden anomalies) is transmitted immediately, routine data is transmitted in batches according to a scheduling plan, and non-critical data is transmitted when network conditions are good. Forward error correction coding is used to enhance transmission reliability, with the error correction capability set to detect 4-bit errors and correct 2-bit errors.

[0051] Step S07 is specifically implemented as the processing flow after the computing center receives the data. The computing center first verifies the integrity of the received data using a Cyclic Redundancy Check (CRC-32) algorithm. Data failing verification requires retransmission. For feature data processed by edge computing, a triple verification is performed through the center's complete model: consistency verification, checking whether the data conforms to physical laws and historical trends; rationality verification, checking whether the data is within a reasonable range (e.g., temperature -5℃ to 40℃); and correlation verification, checking whether different parameters satisfy known correlations (e.g., the negative correlation between temperature and dissolved oxygen). The verified edge processing results are fused with the original collected data using a Kalman filter algorithm. The edge processing results are used as prior estimates, and the original data as observations. The optimal estimate is obtained through recursive calculation. For complex features that edge computing fails to handle, such as nonlinear correlations between parameters, these are extracted in the center's machine learning model. The fusion process considers spatiotemporal consistency; data from adjacent time points and spatial points are smoothed using Kriging interpolation.

[0052] The specific implementation of step S08 involves constructing a spatiotemporal distribution model of seawater parameters. Based on the fused data, a hierarchical modeling method is used to construct this model. First, a spatial interpolation model is established, employing an improved Kriging interpolation algorithm combined with a geographic weighted regression method, considering the influence of environmental factors such as topography and ocean currents, with a spatial resolution of 100m × 100m. Then, a time series model is established, using a seasonal autoregressive integral moving average (SARIMA) model to capture the periodic changes and long-term trends of the parameters. Next, a multi-parameter correlation model is constructed, using a Bayesian network to express the probabilistic dependencies between parameters. Finally, the above models are integrated to form a spatiotemporally integrated distribution model, and high-dimensional data is processed using tensor decomposition. Based on this model, a global view of the water quality status of the monitored sea area is generated, updated hourly, including visualizations such as parameter distribution heatmaps, time-varying trend maps, and anomaly area markers. The model prediction accuracy is evaluated using the root mean square error (RMSE), which requires a temperature prediction RMSE of less than 0.3℃, a salinity prediction RMSE of less than 0.2‰, a dissolved oxygen prediction RMSE of less than 0.5mg / L, and a turbidity prediction RMSE of less than 2NTU.

[0053] The specific implementation of step S09 involves constructing an energy consumption calculation delay balance function and updating the energy consumption assessment benchmark value. The energy consumption calculation delay balance function adopts a weighted multi-objective optimization form, comprehensively considering the energy consumption objective function, delay objective function, and accuracy objective function, with the sum of their weight coefficients being 1. The weights are adjusted according to the system operation stage: in the initial stage, the energy consumption weight is 0.5, the delay weight is 0.3, and the accuracy weight is 0.2; in the stable stage, the energy consumption weight is 0.4, the delay weight is 0.3, and the accuracy weight is 0.3; in emergency situations (such as when the battery level is below 20%), the energy consumption weight is 0.7, the delay weight is 0.2, and the accuracy weight is 0.1. The ε-constrained multi-objective solution method is applied to transform it into a single-objective optimization problem, which is solved using a genetic algorithm. The population size is set to 50, the number of iterations is 100, the crossover probability is 0.8, and the mutation probability is 0.1. Using the Pareto optimal solution set method, a set of non-dominated solutions is obtained, and the solution that best matches the current system state is selected as the optimization result. The above calculation updates the edge computing energy consumption and data transmission energy consumption calculation results, realizing the dynamic adjustment of the energy consumption assessment benchmark value. The old benchmark value is weighted and averaged with the current calculated optimal benchmark value, and the adjustment coefficient is set to 0.7.

[0054] The specific implementation of step S10 involves system evaluation and optimization. The overall energy efficiency of the system is periodically evaluated, with an evaluation period of 24 hours. Evaluation indicators include: average energy consumption per unit of data (mJ / bit), system uptime (hours), data processing latency (seconds), and prediction accuracy (%). The evaluation process employs a hierarchical evaluation method: first, an evaluation is conducted for individual sensor nodes, calculating the node's energy efficiency as total power consumption divided by the amount of data processed; then, an evaluation is conducted for the node cluster, calculating the uniformity of cluster energy consumption distribution; finally, an evaluation is conducted for the entire system, calculating the system energy efficiency ratio as the service quality index divided by the total system power consumption. Based on the evaluation results, the seawater parameter data categories suitable for edge computing and suitable for central computing are reclassified using a classification decision tree algorithm. The decision tree construction uses the ID3 algorithm, feature selection is based on information gain, and the decision threshold is determined through cross-validation. The partitioning results are applied to a dynamic energy consumption threshold judgment mechanism to update the energy consumption thresholds of each parameter. Simultaneously, the lightweight model parameters of the sensor acquisition node terminal are updated using a federated learning method, with an update frequency of once a week and the parameter change rate controlled within 20% to ensure model stability. The overall optimization goal of the system is to reduce energy consumption by more than 30% while maintaining data analysis accuracy of no less than 95%.

[0055] The intelligent edge computing model for seawater parameter processing employs a hybrid architecture combining a multilayer perceptron and a convolutional neural network. First, in the feature extraction layer, a 3×3 lightweight convolutional kernel combined with the ReLU activation function is used, with three convolutional layers containing 16, 32, and 64 kernels respectively, achieving data dimensionality reduction and feature extraction. The temporal correlation analysis layer uses Long Short-Term Memory (LSTM) units with a hidden layer size of 128 and a time step of 24 to capture the patterns of parameter changes over time. The parameter correlation analysis layer establishes correlation mappings between different seawater parameters through residual connection structures. The residual blocks consist of two 3×3 convolutional layers plus a skip connection, with a total of four blocks. The decision output layer comprises fully connected layers with 256, 128, 64, 32, and 16 neurons respectively, with the final layer outputting the processing strategy decision value. An attention mechanism is introduced throughout the network, with attention weights dynamically adjusted through three key parameters: a weight of 0.4 for the energy consumption assessment baseline, a weight of 0.3 for the data change rate, and a weight of 0.3 for the network transmission status.

[0056] The lightweight terminal model is a simplified version of the edge computing model. It replaces standard convolutions with depthwise separable convolutions, decomposing each standard convolution into a depthwise convolution and a pointwise convolution, reducing computational cost by approximately 75%. The simplified network structure includes: the input layer directly connected to two depthwise separable convolutional layers (8 and 16 kernels respectively), followed by a global average pooling layer, and then three fully connected layers (32, 16, and 8 neurons respectively). Complex recursive structures are removed, and a simple echo-state network replaces LSTM, with a state vector size of 64. Residual connections are simplified to linear skip connections. Model quantization uses 8-bit integer quantization, reducing the number of weight parameters by 93%, from 4.6MB to 320KB, and reducing computational complexity by 87%, from 2.4 GFLOPS to 0.31 GFLOPS, ensuring real-time operation on an 80MHz processor.

[0057] The central complete model retains the original network structure and parameters, performing complex data validation, fusion, and analysis tasks. The model structure adds the following to the original architecture: a cross-node attention layer to capture data correlations between different monitoring locations; a multi-scale feature extraction module to extract features in parallel using convolutional kernels of different sizes; a deep residual network module containing 16 residual blocks; and a conditional random field layer to model the spatial dependencies of seawater parameters. The central model has 42MB of parameters, a computational complexity of 15 GFLOPS, and is deployed on a GPU-accelerated server.

[0058] The pre-training dataset establishment process for the intelligent edge computing model for seawater parameter processing includes the following steps: Two years of historical data were collected from 25 monitoring stations in the East China Sea, South China Sea, and Bohai Sea, covering seawater parameter changes under different seasons and meteorological conditions; the collected data were standardized using the Z-Score standardization method to eliminate dimensional differences between different parameters; optimal data processing strategies under different environmental conditions were labeled based on oceanographic expert knowledge, including scenario division suitable for edge processing and suitable for center processing; edge cases and abnormal samples were generated through data augmentation techniques, including adding Gaussian noise (signal-to-noise ratio of 10-20 dB), random missing values ​​(missing rate of 5-15%), and extreme value injection (15-25% exceeding the normal range); finally, a paired training sample set containing input features and optimal decision labels was constructed, with a total of 1.8 million samples, divided into training, validation, and test sets in a 7:2:1 ratio.

[0059] The model pre-training process employs a multi-stage learning strategy: First, unsupervised pre-training is performed using an autoencoder to learn the inherent distribution characteristics of seawater parameter data, with the encoding dimension set to 40% of the original dimension. Then, supervised fine-tuning training is conducted based on labeled data, using the cross-entropy loss function, a learning rate of 0.001, a batch size of 64, and 200 training epochs. Next, adversarial training is implemented to build a discriminator that distinguishes between real and simulated environment data, enhancing the model's robustness to environmental changes. Subsequently, knowledge distillation is performed, using the central model as the teacher model and the terminal model as the student model, achieving knowledge transfer by minimizing the KL divergence of their output distributions, with the temperature parameter set to 3. Finally, the model is deployed in a real-world scenario, with continuous performance optimization through collaborative online learning between the central full model and the lightweight terminal model. The central model aggregates terminal model parameter updates weekly, using a federated averaging algorithm, with update weights dynamically allocated based on node data quality. Throughout the training process, an early stopping strategy is applied to avoid overfitting; training is stopped if there is no improvement in validation set performance after five consecutive epochs, and the model parameters with the best performance on the validation set are saved.

[0060] The mathematical model or calculation process involved in this invention will be described in detail below.

[0061] The energy consumption assessment function mentioned in step S01 is specifically expressed as follows:

[0062] E total =α·E edge +β·E trans ;

[0063] In the formula, E total Total energy consumption; E edge Energy consumption for edge computing; E trans α represents the energy consumption for data transmission; α and β are weighting coefficients and α+β=1.

[0064] The formula for calculating the energy consumption of edge computing is:

[0065] E edge =k1·D·C·T;

[0066] In the formula, E edge The unit is edge computing power consumption, expressed in joules (J); D is the data size, expressed in bits; C is the data complexity, a dimensionless number ranging from 0.1 to 10; T is the processing time, expressed in seconds (s); k1 is the device energy efficiency coefficient, expressed in joules per (bit·complexity·second), the value of which is determined based on device performance, and for ARM Cortex-M4 processors, a typical value is 1 × 10⁻⁶. -9 Up to 5×10 -9 .

[0067] The formula for calculating data transmission energy consumption is:

[0068] E trans =k2·S·N;

[0069] In the formula, E trans S is the data transmission energy consumption, measured in joules (J); S is the data transmission amount, measured in bits (bit); N is the network condition coefficient, a dimensionless number ranging from 1 to 10, with a larger value indicating a worse network condition; k2 is the network transmission energy efficiency coefficient, measured in joules per (bit·network condition coefficient), with a typical value of 1 × 10 for low-power wide area networks. -7 Up to 5×10 -7 .

[0070] The formula for calculating the rate of change of data is:

[0071]

[0072] In the formula, R represents the rate of change of data, expressed as a percentage (%); V current V represents the value at the current sampling point. previous This is the value of the previous sampling point.

[0073] The formula for calculating data disorder is:

[0074] Q=-∑(p i ·log(p i ));

[0075] In the formula, Q represents the data disorder degree, a dimensionless number; p i The probability that a data point falls within the i-th interval is obtained by statistically analyzing the frequency of the value falling within each interval in historical data. The interval division method involves dividing the data range into 10 equal intervals and calculating the proportion of data points in each interval to the total number of data points as the probability value.

[0076] The formula for calculating the benchmark value for energy consumption assessment is:

[0077]

[0078] In the formula, B is the energy consumption assessment benchmark value, and the unit is joules per bit (J / bit); E total D represents total energy consumption in joules (J); D represents data size in bits.

[0079] These formulas are constructed considering the following principles: total energy consumption is expressed as a weighted sum, reflecting the proportional contribution of edge computing and data transmission energy consumption; edge computing energy consumption is proportional to data volume, complexity, and processing time, reflecting the impact of computing load on energy consumption; data transmission energy consumption is proportional to the amount of data transmitted and network conditions, reflecting communication costs; data change rate is calculated using relative change, eliminating the influence of differences in the magnitude of different parameters; data disorder is expressed using the concept of entropy, quantifying the uncertainty and complexity of the data. The energy consumption assessment benchmark represents the energy consumption per unit of data processing, providing a benchmark for comparison between different processing strategies.

[0080] The membership function and fuzzy rules of the fuzzy logic controller mentioned in step S04 are specifically represented as follows:

[0081] For the input variable "rate of change of data", its membership function is defined as:

[0082]

[0083] In the formula, μ 变化率_低 (R) represents the "low" membership degree of the data change rate; μ 变化率_高 (R) represents the "high" membership degree of the data change rate; R is the data change rate; R min As the lower limit of the rate of change, 1% is taken for the temperature parameter; R max As the upper limit of the rate of change, the temperature parameter is set to 3%.

[0084] A similar trapezoidal membership function is used for other input variables (data disorder, node remaining power ratio, network signal strength).

[0085] The fuzzy rule output is defuzzified using the centroid method:

[0086]

[0087] In the formula, D is the processing decision value, ranging from 0 to 1, where 0 indicates transmission to central processing and 1 indicates edge processing; μ i d represents the trigger strength of the i-th rule; i is the output value of the i-th rule; n is the total number of rules, which is 16 rules in this case (2^4 input combinations).

[0088] Rule trigger strength is calculated using the minimum number of operations:

[0089]

[0090] In the formula, μ i Let be the trigger strength of the i-th rule; Let the first input variable x1 correspond to the set A in the i-th rule. i The degree of membership, and others are similar.

[0091] Threshold adaptive adjustment uses exponential smoothing:

[0092] T new =(1-γ)·T old +γ·T calc ;

[0093] In the formula, T new The updated threshold; T old The original threshold; T calc γ is the optimal threshold obtained from the current calculation; γ is the smoothing coefficient with a value of 0.3.

[0094] These formulas are constructed based on the following principles: the fuzzy logic controller uses a trapezoidal membership function to achieve fuzzification of input variables, making the decision-making process smoother; the rule triggering strength uses minimum operation, which conforms to the "AND" logic of fuzzy control; the output defuzzification uses the centroid method to calculate the weighted average, which improves the accuracy of decision-making; the threshold adaptive adjustment uses the exponential smoothing method, which preserves historical information while introducing new information, avoiding drastic fluctuations in the threshold.

[0095] The Z-Score outlier detection algorithm mentioned in step S05 is specifically represented as follows:

[0096]

[0097] In the formula, Z is the standardized score; X is the current observation; μ is the moving average; and σ is the moving standard deviation.

[0098] Moving average calculation formula:

[0099]

[0100] In the formula, μ t X is the moving average at time t; i is the observed value at time i; w is the window size, with temperature parameter set to 15, salinity parameter set to 12, dissolved oxygen parameter set to 8, and turbidity parameter set to 6.

[0101] Formula for calculating moving standard deviation:

[0102]

[0103] In the formula, σ t is the moving standard deviation at time t; other parameters are the same as above.

[0104] When |Z|>3, it is identified as an outlier and corrected using local linear interpolation.

[0105]

[0106] In the formula, X corrected The corrected value; X t-1 The value of X at the previous moment; t+1 This is the value at the next moment.

[0107] These formulas are constructed based on the following principles: Z-Score standardization transforms data of different magnitudes to a unified standard for easy comparison; moving average and moving standard deviation can adapt to dynamic changes in data; the anomaly detection threshold is set to 3, based on the 3-sigma rule of normal distribution, and the probability of exceeding this range is only 0.3%; local linear interpolation assumes that parameter changes are approximately linear in a short period of time, and estimates the true value of anomalies by the average rate of change of adjacent normal points.

[0108] The Kalman filter algorithm mentioned in step S07 is specifically represented as follows:

[0109] State prediction equation:

[0110]

[0111] In the formula, The prior state estimate at time t; Let A be the posterior state estimate at time t-1; A is the state transition matrix; B is the control input matrix; u t For controlling input.

[0112] Prediction error covariance equation:

[0113]

[0114] In the formula, Let P be the prior estimation error covariance matrix at time t; t-1 Let be the posterior estimation error covariance matrix at time t-1; Q is the process noise covariance matrix.

[0115] Kalman gain calculation:

[0116]

[0117] In the formula, K tH is the Kalman gain; H is the observation matrix; R is the observation noise covariance matrix.

[0118] State update equation:

[0119]

[0120] In the formula, z is the posterior state estimate at time t; t Let be the observed value at time t.

[0121] Posterior error covariance update:

[0122]

[0123] In the formula, P t Let be the posterior estimation error covariance matrix at time t; I is the identity matrix.

[0124] In seawater parameter fusion, the state vector It includes four parameters: temperature, salinity, dissolved oxygen, and turbidity; the observation vector z t Includes raw collected data; prior estimates Results from edge processing. The noise covariance matrices Q and R were obtained statistically from historical data; typical values ​​are:

[0125]

[0126] These formulas are constructed based on the following principles: the Kalman filter algorithm achieves optimal estimation through two stages: prediction and update; the state prediction equation predicts the next state based on the system dynamics model; the prediction error covariance equation quantifies the uncertainty of the prediction; the Kalman gain balances the reliability of the prediction and observation; the state update integrates the predicted and observed values ​​to obtain the optimal estimate; and the posterior error covariance update reflects the improvement in estimation accuracy. In this way, the algorithm can obtain optimal estimates in noisy environments and is suitable for fusing edge processing results with the original data.

[0127] The Kriging spatial interpolation algorithm mentioned in step S08 is specifically represented as follows:

[0128]

[0129] In the formula, Z * (x0) is the interpolated estimate of position x0; Z(x i ) represents the position x i The known observations; λ i is the weighting coefficient; n is the number of known points involved in the interpolation.

[0130] Weighting coefficient λ i This is obtained by solving the following system of equations:

[0131]

[0132] In the formula, γ(x) i x j ) represents the position x i With x j The semivariance between them; μ is the Lagrange multiplier; γ(x) i x0) is the position x i The semivariance between x0 and x0.

[0133] The semivariance function adopts a spherical model:

[0134]

[0135] In the formula, γ(h) is the semivariance between two points at a distance h; c0 is the nugget effect, representing variability at a small scale; c is the sill value, representing the variability of the spatially correlated structure; and a is the range, representing the maximum distance of the spatial correlation. These parameters are obtained by fitting experimental semivariograms. Specifically, the semivariance of sample point pairs at different distance intervals is calculated, a scatter plot is drawn, and then a spherical model is fitted using the least squares method.

[0136] The SARIMA time series model is specifically represented as follows:

[0137] φ p (B)(1-B) d (1-B s ) D φ P (B s )X t =θ q (B)Θ Q (B s )ε t ;

[0138] In the formula, X t It is a time series; ε t B represents white noise; B is the hysteresis operator, i.e., BX. t =X t-1 ;φ p (B)=1-φ1B-φ2B 2 -…-φ p B p It is a non-seasonal autoregressive polynomial; θ q (B)=1-θ1B-θ2B 2 -…-θ q B q It is a non-seasonal moving average polynomial; Φ P (B s )=1-Φ1B s-Φ2B 2s -…-Φ P B Ps It is a seasonal autoregressive polynomial; Θ Q (B s )=1-Θ1B s -Θ2B 2s -…-Θ Q B Qs is the seasonal moving average polynomial; d is the non-seasonal difference order; D is the seasonal difference order; s is the seasonal cycle.

[0139] Parameters (p, d, q)(P, D, Q) s The values ​​were determined through autocorrelation function (ACF) and partial autocorrelation function (PACF) analysis, and the optimal model was selected using the Akaike Information Criterion (AIC). Typical values ​​for the temperature parameters are (2, 1, 1) and (1, 1, 1). 24 This indicates that it includes a second-order non-seasonal autoregressive term, a first-order non-seasonal difference term, a first-order non-seasonal moving average term, a first-order seasonal autoregressive term, a first-order seasonal difference term, and a first-order seasonal moving average term, with a seasonal period of 24 hours.

[0140] These formulas are constructed based on the following principles: Kriging interpolation considers spatial autocorrelation, and weight allocation is based on spatial location and data variation structure; the semivariogram function describes the impact of distance on data similarity, and the spherical model can reflect the spatial variation characteristics of most seawater parameters; the SARIMA model comprehensively considers the trend, seasonality, and randomness of time series, and can capture the periodic changes and long-term trends of seawater parameters. Through these methods, the constructed spatiotemporal distribution model can accurately reflect the variation patterns of seawater parameters at different locations and times.

[0141] The multi-objective optimization function mentioned in step S09 is specifically represented as follows:

[0142] F(x) = w1·f energy (x)+w2·f delay (x)+w3·f accuracy (x);

[0143] In the formula, F(x) is the comprehensive objective function; f energy (x) is the energy consumption objective function; f delay (x) is the delay objective function; f accuracy (x) is the accuracy objective function; w1, w2, and w3 are weight coefficients and w1+w2+w3=1; x is a decision variable vector, which includes the edge computing task allocation ratio and resource allocation scheme.

[0144] Energy consumption objective function:

[0145]

[0146] In the formula, E total (x) represents the total energy consumption; E min and E max These are the historical minimum and maximum energy consumption, used for normalization.

[0147] Delay objective function:

[0148]

[0149] In the formula, T total (x) represents the total processing delay; T min and T max These are the historical minimum and maximum delays, used for normalization.

[0150] Precision objective function:

[0151]

[0152] In the formula, ACC(x) represents the prediction accuracy; ACC min and ACC max These are the historical lowest and highest accuracy rates, used for normalization; "1-" is used because the objective function needs to be minimized.

[0153] The ε-constraint method transforms a multi-objective problem into a single-objective problem:

[0154]

[0155] In the formula, ε1 and ε2 are the constraint thresholds for delay and accuracy, and the Pareto front is obtained by systematically changing these thresholds.

[0156] The Pareto optimal solution set is obtained using a genetic algorithm. Individual populations are encoded as real-number vectors, representing the marginal computation proportions and resource allocation schemes for various seawater parameter data. Simulated binary crossover (SBX) is used for crossover, and polynomial mutation is employed. The fitness function is the objective function described above. Population evolution maintains the diversity of the solution set through non-dominated sorting and crowding distance.

[0157] Energy consumption assessment benchmark update formula:

[0158] B new =η·B old +(1-η)·B calc ;

[0159] In the formula, B new The updated energy consumption assessment baseline; B old B is the original baseline value; calc η is the currently calculated optimal baseline value; η is the adjustment coefficient, with a value of 0.7.

[0160] These formulas are constructed based on the following principles: the multi-objective optimization function adopts a weighted sum form, achieving comprehensive optimization of energy consumption, latency, and accuracy; each objective function is normalized to eliminate dimensional differences; the ε-constraint method systematically explores the Pareto optimal solution set by adjusting constraints; the genetic algorithm is suitable for solving complex nonlinear optimization problems; and the energy consumption assessment benchmark is updated using exponential smoothing, which preserves historical information while adapting to system changes. In this way, an optimal balance can be found between energy consumption, computational latency, and processing accuracy.

[0161] The system evaluation metrics mentioned in step S10 are specifically represented as follows:

[0162] Node energy efficiency:

[0163]

[0164] In the formula, E node P represents node energy efficiency, measured in joules per bit (J / bit). total Total power consumption, in joules (J); D processed The unit for processing data volume is bits.

[0165] Cluster energy consumption distribution uniformity:

[0166]

[0167] In the formula, σ 2 Variance represents the uniformity of energy consumption distribution; E represents the energy efficiency of the i-th node. avg is the average node energy efficiency; n is the total number of nodes.

[0168] System energy efficiency ratio:

[0169]

[0170] In the formula, E sys Q represents the system's energy efficiency ratio. service As a service quality indicator, it is defined as follows: Where ACC is the prediction accuracy, T is the average response time, COV is the monitoring coverage, and α1, α2, and α3 are weighting coefficients, with α1 + α2 + α3 = 1; P sys This represents the total power consumption of the system.

[0171] The ID3 decision tree algorithm uses information gain to select the optimal splitting feature:

[0172]

[0173] In the formula, Gain(S, A) is the information gain of feature A on sample set S; Entropy(S) is the entropy of sample set S; Values(A) are all possible values ​​of feature A; S v Let S be a subset of samples whose feature A takes the value v; |S| is the number of samples in sample set S; |S v |For the sample subset S v The number of samples.

[0174] Optional, the formula for calculating entropy:

[0175]

[0176] In the formula, Entropy(S) is the entropy of the sample set S; p i represents the proportion of samples of class i in the sample set S; m is the number of classes. In this system, m = 2, corresponding to the two classes "suitable for edge computing" and "suitable for center computing".

[0177] Optionally, the mean filtering algorithm mentioned in step S05 is specifically represented as follows:

[0178]

[0179] In the formula, Y t X is the filtered output value at time t; t-i W represents the original sampled value at time ti; i is the i-th weight coefficient; w is the size of the filter window.

[0180] Optionally, the weighting coefficients can be distributed using an exponentially decreasing method:

[0181]

[0182] In the formula, W i λ is the i-th weighting coefficient; λ is the attenuation factor, ranging from 0 to 1, usually set to 0.8 to 0.9; w is the filter window size, with temperature parameter set to 15, salinity parameter set to 12, dissolved oxygen parameter set to 8, and turbidity parameter set to 6.

[0183] Specifically, the principle of this invention is based on a multi-objective optimization concept that balances energy consumption in edge computing and data transmission. The core of this invention lies in constructing a dynamic decision-making mechanism to evaluate and adjust data processing strategies in real time. First, the invention establishes a model relating data characteristics to energy consumption. By analyzing the rate of change and disorder of seawater parameter data, it quantifies the energy cost under different processing strategies, providing a theoretical basis for decision-making. Data with low rates of change and low disorder is suitable for processing at edge nodes, significantly reducing the amount of data transmitted; while data with drastic changes or high disorder requires more complete analysis and is suitable for transmission to the computing center.

[0184] At the implementation level, this invention adopts a dual-model collaborative architecture. The lightweight terminal model focuses on low-computational-complexity feature extraction and anomaly detection tasks, reducing computational complexity through techniques such as depthwise separable convolution, enabling it to run efficiently on energy-constrained sensing nodes. The central complete model is responsible for complex analysis and verification tasks, while also guiding the optimization of the lightweight model. This hierarchical design allows the system to flexibly allocate computational tasks based on data characteristics and network conditions.

[0185] The innovation of this invention lies in the introduction of a dynamic energy consumption threshold judgment mechanism and an energy consumption calculation delay balance function, which continuously optimizes the decision boundary through the Pareto optimal solution set method. Unlike static task partitioning methods, this system can adaptively adjust according to environmental changes. When network conditions are good, it tends to transmit more raw data; when energy is scarce or network is limited, it increases the proportion of edge computing. This dynamic balancing strategy ensures that the system achieves an optimal balance between energy consumption and computational efficiency under various operating conditions, thereby effectively solving the problem of continuous operation of the seawater parameter monitoring IoT system in energy-constrained environments.

[0186] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0187] The specific implementation of step S01 involves deploying a multi-layered seawater parameter monitoring sensor network in the monitored sea area to form a densely covered monitoring array. An energy consumption assessment function is then used to calculate the edge computing energy consumption and data transmission energy consumption, forming an energy consumption assessment benchmark value. The energy consumption assessment function adopts a weighted summation form, specifically expressed as follows:

[0188] E total =α·E edge +β·E trans ;

[0189] In the formula, E total Total energy consumption; E edge Energy consumption for edge computing; E trans α represents the energy consumption for data transmission; α and β are weighting coefficients and α+β=1.

[0190] The formula for calculating the energy consumption of edge computing is:

[0191] E edge =k1·D·C·T;

[0192] In the formula, E edgeThe unit is edge computing power consumption, expressed in joules (J); D is the data size, expressed in bits; C is the data complexity, a dimensionless number ranging from 0.1 to 10; T is the processing time, expressed in seconds (s); k1 is the device energy efficiency coefficient, expressed in joules per (bit·complexity·second), the value of which is determined based on device performance, and for ARM Cortex-M4 processors, a typical value is 1 × 10⁻⁶. -9 Up to 5×10 -9 .

[0193] The formula for calculating data transmission energy consumption is:

[0194] E trans =k2·S·N;

[0195] In the formula, E trans S is the data transmission energy consumption, measured in joules (J); S is the data transmission amount, measured in bits (bit); N is the network condition coefficient, a dimensionless number ranging from 1 to 10, with a larger value indicating a worse network condition; k2 is the network transmission energy efficiency coefficient, measured in joules per (bit·network condition coefficient), with a typical value of 1 × 10 for low-power wide area networks. -7 Up to 5×10 -7 .

[0196] The formula for calculating the rate of change of data is:

[0197]

[0198] In the formula, R represents the rate of change of data, expressed as a percentage (%); V current V represents the value at the current sampling point. previous This is the value from the previous sampling point. The data change rate threshold is set to 2%.

[0199] The formula for calculating data disorder is:

[0200] Q=-∑(p i ·log(p i ));

[0201] In the formula, Q represents the data disorder degree, a dimensionless number; p i The probability of a data point falling into the i-th interval is obtained by statistically analyzing the frequency of historical data falling into each interval. The interval division method involves dividing the data range into 10 equal intervals and calculating the proportion of data points in each interval to the total number of data points as the probability value. The data disorder threshold is set to 0.6.

[0202] The formula for calculating the benchmark value for energy consumption assessment is:

[0203]

[0204] In the formula, B is the energy consumption assessment benchmark value, and the unit is joules per bit (J / bit); E total D represents total energy consumption in joules (J); D represents data size in bits. This baseline value serves as the basis for subsequent processing decisions.

[0205] The specific implementation methods for steps S02-S03 are the same as those described above, and will not be repeated here.

[0206] The specific implementation of step S04 involves establishing a dynamic energy consumption threshold judgment mechanism. This mechanism, based on energy consumption assessment benchmarks, sets initial thresholds for each type of seawater parameter: temperature threshold is 3 mJ / bit, salinity threshold is 2.5 mJ / bit, dissolved oxygen threshold is 3.2 mJ / bit, and turbidity threshold is 3.8 mJ / bit. The judgment process is implemented using a fuzzy logic controller. Input variables include data change rate, data disorder, node remaining power ratio, and network signal strength; the output is the processing decision value. For the input variable "data change rate," its membership function is defined as:

[0207]

[0208] In the formula, μ 变化率_低 (R) represents the "low" membership degree of the data change rate; μ 变化率_高 (R) represents the "high" membership degree of the data change rate; R is the data change rate; R min As the lower limit of the rate of change, 1% is taken for the temperature parameter; R max As the upper limit of the rate of change, the temperature parameter is set to 3%.

[0209] For other input variables (data disorder, node remaining power percentage, network signal strength), a similar trapezoidal membership function is used. The design principles of the fuzzy rules are as follows: when the node's remaining power is higher than 70%, computational tasks are tended to be performed at the edge; when the network signal strength is higher than -80dBm, data is tended to be transmitted to the computing center; when the data change rate is higher than a set threshold, processing is prioritized at the edge to reduce communication load; and when the data disorder is higher than 0.6, data is prioritized for transmission to the computing center for complex analysis.

[0210] The fuzzy rule output is defuzzified using the centroid method:

[0211]

[0212] In the formula, D is the processing decision value, ranging from 0 to 1, where 0 indicates transmission to central processing and 1 indicates edge processing; μ i d represents the trigger strength of the i-th rule; i This represents the output value of the i-th rule; n is the total number of rules, which is 16 rules in this case (2^n). 4(Types of input combinations).

[0213] Rule trigger strength is calculated using the minimum number of operations:

[0214]

[0215] In the formula, μ i Let be the trigger strength of the i-th rule; Let the first input variable x1 correspond to the set A in the i-th rule. i The degree of membership, and others are similar.

[0216] Threshold adaptive adjustment uses exponential smoothing:

[0217] T new =(1-γ)·T old +γ·T calc ;

[0218] In the formula, T new The updated threshold; T old The original threshold; T calc γ is the optimal threshold obtained from the current calculation; γ is the smoothing coefficient with a value of 0.3.

[0219] The specific implementation of step S05 involves performing lightweight edge computing tasks at the sensor acquisition node. For seawater parameter data deemed suitable for edge computing processing, three types of processing tasks are performed through the terminal's lightweight model. First, downsampling processing is performed using an adaptive sampling algorithm based on Shannon's sampling theorem. The sampling rate is dynamically adjusted according to the signal spectrum characteristics. For parameters with slow changes, such as temperature and salinity, the sampling rate can be reduced to 25% of the original sampling rate; for parameters with rapid changes, such as dissolved oxygen and turbidity, the sampling rate is maintained at over 60% of the original sampling rate. Then, mean filtering is performed using a weighted moving average algorithm. The window size is set to 5 to 15 sampling points depending on the parameter type, and the weight coefficients are distributed using an exponentially decreasing method. The mean filtering algorithm is specifically expressed as follows:

[0220]

[0221] In the formula, Y t X is the filtered output value at time t; t-i W represents the original sampled value at time ti; i is the i-th weight coefficient; w is the size of the filter window.

[0222] The weighting coefficients are allocated using an exponentially decreasing method:

[0223]

[0224] In the formula, W iλ is the i-th weighting coefficient; λ is the attenuation factor, ranging from 0 to 1, usually set to 0.8 to 0.9; w is the filter window size, with temperature parameter set to 15, salinity parameter set to 12, dissolved oxygen parameter set to 8, and turbidity parameter set to 6.

[0225] Finally, outlier detection is performed using a modified Z-Score algorithm:

[0226]

[0227] In the formula, Z is the standardized score; X is the current observation; μ is the moving average; and σ is the moving standard deviation.

[0228] Moving average calculation formula:

[0229]

[0230] In the formula, μ t X is the moving average at time t; i is the observed value at time i; w is the window size, with temperature parameter set to 15, salinity parameter set to 12, dissolved oxygen parameter set to 8, and turbidity parameter set to 6.

[0231] Formula for calculating moving standard deviation:

[0232]

[0233] In the formula, σ t is the moving standard deviation at time t; other parameters are the same as above.

[0234] When |Z|>3, it is identified as an outlier and corrected using local linear interpolation.

[0235]

[0236] In the formula, X corrected The corrected value; X t-1 The value of X at the previous moment; t+1 This represents the value at the next time step. The above processing is performed on the node's microprocessor, and the processing result and the original data are packaged and stored separately, awaiting transmission.

[0237] The specific implementation method of step S06 is the same as described above, and will not be repeated here.

[0238] Step S07 is specifically implemented as the processing flow after the computing center receives the data. The computing center first verifies the integrity of the received data using a Cyclic Redundancy Check (CRC-32) algorithm. Data failing the verification is retransmitted. For the feature data after edge computing processing, a triple verification is performed through the center's complete model: consistency verification, checking whether the data conforms to physical laws and historical trends; rationality verification, checking whether the data is within a reasonable range (e.g., temperature -5℃ to 40℃); and correlation verification, checking whether different parameters satisfy known correlations (e.g., the negative correlation between temperature and dissolved oxygen). The verified edge processing results are fused with the original acquired data using a Kalman filter algorithm. The Kalman filter algorithm is specifically expressed as follows:

[0239] State prediction equation:

[0240]

[0241] In the formula, The prior state estimate at time t; Let A be the posterior state estimate at time t-1; A is the state transition matrix; B is the control input matrix; u t For controlling input.

[0242] Prediction error covariance equation:

[0243]

[0244] In the formula, Let P be the prior estimation error covariance matrix at time t; t-1 Let be the posterior estimation error covariance matrix at time t-1; Q is the process noise covariance matrix.

[0245] Kalman gain calculation:

[0246]

[0247] In the formula, K t H is the Kalman gain; H is the observation matrix; R is the observation noise covariance matrix.

[0248] State update equation:

[0249]

[0250] In the formula, z is the posterior state estimate at time t; t Let be the observed value at time t.

[0251] Posterior error covariance update:

[0252]

[0253] In the formula, P t Let be the posterior estimation error covariance matrix at time t; I is the identity matrix.

[0254] In seawater parameter fusion, the state vector It includes four parameters: temperature, salinity, dissolved oxygen, and turbidity; the observation vector z t Includes raw collected data; prior estimates Results from edge processing. The noise covariance matrices Q and R were obtained statistically from historical data; typical values ​​are:

[0255]

[0256] For complex features that edge computing cannot handle, such as nonlinear relationships between parameters, these are extracted in the central machine learning model. The fusion process considers spatiotemporal consistency, and data from adjacent time points and spatial points are smoothed using Kriging interpolation.

[0257] The specific implementation of step S08 involves constructing a spatiotemporal distribution model of seawater parameters. Based on the fused data, a hierarchical modeling method is used to construct the spatiotemporal distribution model of seawater parameters. First, a spatial interpolation model is established, using an improved Kriging interpolation algorithm, specifically expressed as follows:

[0258]

[0259] In the formula, Z * (x0) is the interpolated estimate of position x0; Z(x i ) represents the position x i The known observations; λ i is the weighting coefficient; n is the number of known points involved in the interpolation.

[0260] Weighting coefficient λ i This is obtained by solving the following system of equations:

[0261]

[0262] In the formula, γ(x) i x j ) represents the position x i With x j The semivariance between them; μ is the Lagrange multiplier; γ(x) i x0) is the position x i The semivariance between x0 and x0.

[0263] The semivariance function adopts a spherical model:

[0264]

[0265] In the formula, γ(h) is the semivariance between two points at a distance h; c0 is the nugget effect, representing variability at a small scale; c is the sill value, representing the variability of the spatially correlated structure; and a is the range, representing the maximum distance of the spatial correlation.

[0266] Then, a time series model is established, using a seasonal autoregressive integral moving average (SARIMA) model, specifically expressed as follows:

[0267] φ p (B)(1-B) d (1-B s ) D Φ P (B s )X t =θ q (B)Θ Q (B s )ε t ;

[0268] In the formula, X t It is a time series; ε t B represents white noise; B is the hysteresis operator, i.e., BX. t =X t-1 ;φ p (B)=1-φ1B-φ2B 2 -…-φ p B p It is a non-seasonal autoregressive polynomial; θ q (B)=1-θ1B-θ2B 2 -…-θ q B q It is a non-seasonal moving average polynomial; Φ P (B s )=1-Φ1B s -φ2B 2s -…-Φ P B Ps It is a seasonal autoregressive polynomial; Θ Q (B s )=1-Θ1B s -Θ2B 2s -…-Θ Q B Qs is the seasonal moving average polynomial; d is the non-seasonal difference order; D is the seasonal difference order; s is the seasonal cycle.

[0269] Next, a multi-parameter correlation model was constructed, using a Bayesian network to express the probabilistic dependencies between parameters. Finally, the above models were integrated to form a spatiotemporal integrated distribution model, and high-dimensional data was processed using tensor decomposition. Based on this model, a global view of the monitored sea area's water quality was generated, updated hourly, including visualizations such as parameter distribution heatmaps, temporal trend maps, and anomaly area markers. The model's prediction accuracy was evaluated using root mean square error (RMSE), requiring a predicted RMSE of less than 0.3℃ for temperature, less than 0.2‰ for salinity, less than 0.5 mg / L for dissolved oxygen, and less than 2 NTU for turbidity.

[0270] The specific implementation of step S09 involves constructing an energy consumption calculation delay balance function and updating the energy consumption assessment benchmark value. The energy consumption calculation delay balance function adopts a weighted multi-objective optimization form:

[0271] F(x) = w1·f energy (x)+w2·f delay (x)+w3·f accuracy (x);

[0272] In the formula, F(x) is the comprehensive objective function; f energy (x) is the energy consumption objective function; f delay (x) is the delay objective function; f accuracy (x) is the accuracy objective function; w1, w2, and w3 are weight coefficients and w1+w2+w3=1; x is a decision variable vector, which includes the edge computing task allocation ratio and resource allocation scheme.

[0273] Energy consumption objective function:

[0274]

[0275] In the formula, E total (x) represents the total energy consumption; E min and E max These are the historical minimum and maximum energy consumption, used for normalization.

[0276] Delay objective function:

[0277]

[0278] In the formula, T total (x) represents the total processing delay; T min and T max These are the historical minimum and maximum delays, used for normalization.

[0279] Precision objective function:

[0280]

[0281] In the formula, ACC(x) represents the prediction accuracy; ACC min and ACC max These are the historical lowest and highest accuracy rates, used for normalization; "1-" is used because the objective function needs to be minimized.

[0282] The weights are adjusted according to the system's operational phases: Initial phase: w1 = 0.5, w2 = 0.3, w3 = 0.2; Stable phase: w1 = 0.4, w2 = 0.3, w3 = 0.3; Emergency situations (e.g., battery level below 20%): w1 = 0.7, w2 = 0.2, w3 = 0.1. The ε-constrained multi-objective solution method is then applied to transform it into a single-objective optimization problem.

[0283]

[0284] In the formula, ε1 and ε2 are the constraint thresholds for delay and accuracy. The Pareto front is obtained by systematically changing these thresholds. A genetic algorithm is used to solve for the Pareto optimal solution set, with a population size of 50, 100 iterations, a crossover probability of 0.8, and a mutation probability of 0.1. The solution that best matches the current system state is selected as the optimization result. The energy consumption calculation results for edge computing and data transmission are updated through the above calculations, achieving dynamic adjustment of the energy consumption assessment benchmark value.

[0285] B new =η·B old +(1-η)·B calc ;

[0286] In the formula, B new The updated energy consumption assessment baseline; B old B is the original baseline value; calc η is the currently calculated optimal baseline value; η is the adjustment coefficient, with a value of 0.7.

[0287] The specific implementation of step S10 is system evaluation and optimization. The overall system energy efficiency is periodically evaluated, with the evaluation period set to 24 hours. Evaluation indicators include: average energy consumption per unit data (mJ / bit), system uptime (hours), data processing latency (seconds), and prediction accuracy (%). The evaluation process adopts a hierarchical evaluation method: first, an evaluation is conducted on individual sensor nodes, calculating the node's energy efficiency.

[0288]

[0289] In the formula, E node P represents node energy efficiency, measured in joules per bit (J / bit). total Total power consumption, in joules (J); D processed The unit for processing data volume is bits.

[0290] Then, for the node cluster, the uniformity of cluster energy consumption distribution is calculated:

[0291]

[0292] In the formula, σ 2 Variance represents the uniformity of energy consumption distribution; E represents the energy efficiency of the i-th node. avg is the average node energy efficiency; n is the total number of nodes.

[0293] Finally, the system energy efficiency ratio is calculated for the entire system evaluation:

[0294]

[0295] In the formula, E sys Q represents the system's energy efficiency ratio. service As a service quality indicator, it is defined as follows: Where ACC is the prediction accuracy, T is the average response time, COV is the monitoring coverage, and α1, α2, and α3 are weighting coefficients, with α1 + α2 + α3 = 1; P sys This represents the total power consumption of the system.

[0296] Based on the evaluation results, the seawater parameter data categories suitable for edge computing and those suitable for center computing were reclassified using a classification decision tree algorithm. The ID3 decision tree algorithm uses information gain to select the optimal splitting feature.

[0297]

[0298] In the formula, Gain(S, A) is the information gain of feature A on sample set S; Entropy(S) is the entropy of sample set S; Values(A) are all possible values ​​of feature A; S v Let S be a subset of samples whose feature A takes the value v; |S| is the number of samples in sample set S; |S v |For the sample subset S v The number of samples.

[0299] The formula for calculating entropy:

[0300]

[0301] In the formula, Entropy(S) is the entropy of the sample set S; p i represents the proportion of samples of class i in the sample set S; m is the number of classes. In this system, m = 2, corresponding to the two classes "suitable for edge computing" and "suitable for center computing".

[0302] The decision tree is constructed using the ID3 algorithm, feature selection is based on information gain, and the decision threshold is determined through cross-validation. The partitioning results are applied to a dynamic energy consumption threshold judgment mechanism to update the energy consumption thresholds of each parameter. Simultaneously, the parameters of the lightweight model of the sensor acquisition node terminal are updated using a federated learning method, with an update frequency of once a week and the parameter change rate controlled within 20% to ensure model stability. The overall system optimization goal is to reduce energy consumption by more than 30% while maintaining data analysis accuracy of no less than 95%.

[0303] In this embodiment, the energy consumption evaluation function considers both computational and transmission energy consumption. The fuzzy logic controller enables intelligent decision-making under multiple conditions. The Kalman filter algorithm effectively integrates edge processing results with raw data. Kriging interpolation and the SARIMA model accurately capture the spatiotemporal variation characteristics of seawater parameters. The multi-objective optimization method balances the three objectives of energy consumption, latency, and accuracy. The collaborative working mechanism of the lightweight terminal model and the central complete model fully leverages the respective advantages of edge computing and central computing. Through the organic combination of these technologies and methods, this system achieves edge computing optimization for seawater parameter monitoring IoT, significantly reducing system energy consumption, extending system runtime, and improving monitoring efficiency while ensuring data analysis accuracy.

[0304] To better understand and implement this invention, a specific application scenario is provided below as Example 2: In a certain marine research project, researchers deployed a seawater parameter monitoring IoT system in the northern South China Sea and applied edge computing optimization methods. This sea area covers approximately 1000 square kilometers and is a typical subtropical marine environment with a water depth ranging from 100 to 300 meters, current speeds of 0.3 to 1.5 meters per second, annual seawater temperature variations of 15 to 30°C, and annual salinity variations of 29 to 35‰.

[0305] Researchers deployed 45 sensor nodes, forming a three-layer monitoring network: 15 nodes at the surface (5 meters underwater), 15 nodes at the middle layer (50 meters underwater), and 15 nodes at the bottom layer (10 meters above the seabed). Each node is equipped with an ARM Cortex-M4 microprocessor with an 80MHz clock speed, 4MB of built-in Flash memory and 512KB of RAM, running the FreeRTOS real-time operating system, and powered by a 3.7V / 10000mAh lithium battery with a solar-assisted charging system. The types and frequencies of data collected by the sensors at each node are shown in Table 1.

[0306] Table 1 Sensor Data Acquisition Frequency Table

[0307] Parameter type sampling frequency Single data volume (bytes) Data change rate threshold Disorder threshold temperature 10 minutes / session 4 2.0% 0.6 salinity 15 minutes / session 4 1.5% 0.6 Dissolved oxygen 5 minutes / time 4 5.0% 0.6 Turbidity 3 minutes / time 4 8.0% 0.6

[0308] The energy consumption assessment function was constructed, and the processing methods for various parameter data were determined. Taking a surface node as an example, the actual measured energy consumption assessment benchmark values ​​are shown in Table 2.

[0309] Table 2 Energy Consumption Assessment Benchmark Values

[0310]

[0311] During system operation, a surface node performs edge computing processing on temperature data. The raw collected data is processed using a lightweight terminal model, including downsampling, mean filtering, and outlier detection. Downsampling reduces the sampling rate from 10 minutes / time to 25 minutes / time; a weighted moving average filter with a window size of 15 and a decay factor of 0.85 is used; and outliers are detected and corrected using the Z-Score method. Table 3 shows a comparison of temperature data before and after edge processing on a given day.

[0312] Table 3 Comparison of Temperature Data Before and After Edge Processing

[0313]

[0314]

[0315] The temperature reading of 23.05℃ at 01:20 was noted as a clear outlier, which was successfully detected and corrected using the Z-Score algorithm. After edge processing, the original and processed data were packaged separately and transmitted to the computing center using the LoRaWAN protocol. Compared to the traditional method of transmitting all data to the center, this method significantly reduces the amount of data transmitted and lowers energy consumption.

[0316] After the data arrives at the computing center, it is fused using the Kalman filter algorithm. For surface temperature data, the state transition matrix A = [0.98, 0.1; 0, 0.95], the process noise covariance matrix Q = [0.05, 0; 0, 0.08], and the observation noise covariance matrix R = 0.1 are set. The fused data is used to construct a spatiotemporal distribution model and generate a global view of the water quality status of the monitored sea area.

[0317] After one month of system operation, the system was optimized using the constructed energy consumption calculation delay balance function. Table 4 shows a comparison of the system performance before and after optimization.

[0318] Table 4 Performance Comparison Before and After System Optimization

[0319] Performance indicators Before optimization After optimization Improvement rate (%) Average energy consumption per unit of data (mJ / bit) 3.85 2.62 31.9 Average daily system uptime (hours) 18.5 23.2 25.4 Average data processing latency (seconds) 2.46 1.85 24.8 Temperature forecast RMSE (°C) 0.42 0.28 33.3 Salinity prediction RMSE (‰) 0.31 0.19 38.7 Predicted RMSE (mg / L) for dissolved oxygen 0.63 0.48 23.8 Turbidity prediction RMSE (NTU) 2.33 1.85 20.6 Average battery life (days) 83.5 114.3 36.9

[0320] The key to optimization lies in dynamically adjusting the allocation strategy between edge computing and central computing. For example, when seawater temperature changes slowly (rate of change <1%) and network signal is good (signal strength > -70dBm), the system chooses to transmit data directly to the center; when seawater temperature changes significantly (rate of change > 3%) or network signal is poor (signal strength < -90dBm), data processing is prioritized at the edge. By periodically evaluating the overall system performance, the decision tree algorithm continuously optimizes the data processing strategy, and the parameters of the lightweight terminal model are updated weekly.

[0321] Traditional seawater parameter monitoring systems typically employ two approaches: one is to transmit all data back to a central processing center, resulting in high energy consumption, large bandwidth requirements, and short system lifespan; the other is simple edge processing, such as fixed data filtering rules, lacking intelligence and adaptability. In contrast, this invention employs a dual-model collaborative edge computing optimization method, achieving the following improvements: 1) Dynamically determining data processing location through an energy consumption evaluation function reduces energy consumption by 31.9%; 2) The lightweight terminal model and the central complete model work together, balancing real-time response and analysis accuracy; 3) A dynamic energy consumption threshold judgment mechanism adaptively adjusts based on network conditions and battery power, extending battery life by 36.9%; 4) The Pareto optimal solution set method achieves a balance between energy consumption and computational latency, reducing processing latency by 24.8%; 5) An energy consumption and computational latency balance function constructed using multi-objective optimization theory achieves dynamic optimization of overall system performance, improving prediction accuracy by 20.6%-38.7%. These advancements enable the seawater parameter monitoring IoT system to achieve low energy consumption, high accuracy, and long lifespan, providing a more efficient and reliable technical solution for marine environmental monitoring.

[0322] It should be noted that the variables involved in this invention are explained in detail in Tables 5, 6, and 7 below.

[0323] Table 5. Variable Explanation Table (Part 1)

[0324]

[0325] Table 6. Variable Explanation Table (Part Two)

[0326]

[0327]

[0328] Table 7. Variable Explanation Table (Part 3)

[0329]

[0330] 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. An edge computing optimization method for a seawater parameter monitoring Internet of Things (IoT) system, characterized in that, include: A seawater parameter monitoring sensor network is constructed to collect seawater parameter data. Based on the rate of change and disorder of the collected data, the energy consumption of edge computing and data transmission is calculated using an energy consumption assessment function to form an energy consumption assessment benchmark value. The collected seawater parameter data is analyzed, and data feature extraction is performed based on the data acquisition frequency and data change rate. Deploy the intelligent processing edge computing model for seawater parameters into the system, deploy the lightweight terminal model at each sensor acquisition node, and deploy the complete central model at the computing center; A dynamic energy consumption threshold judgment mechanism is set up to determine whether each type of seawater parameter data is suitable for edge computing processing or suitable for transmission to the computing center based on the energy consumption assessment benchmark value. For seawater parameter data suitable for edge computing processing, a lightweight edge computing task is executed at the sensor acquisition node, and the processed data and the unprocessed original data are transmitted to the computing center respectively. The computing center performs data fusion processing, constructs a spatiotemporal distribution model of seawater parameters, updates the energy consumption assessment benchmark value using multi-objective optimization theory, periodically evaluates the system's energy consumption efficiency, and optimizes the computing strategy.

2. The edge computing optimization method for the seawater parameter monitoring IoT system according to claim 1, characterized in that, The seawater parameter data includes seawater temperature, seawater salinity, seawater dissolved oxygen, and seawater turbidity.

3. The edge computing optimization method for the seawater parameter monitoring IoT system according to claim 2, characterized in that, The energy consumption assessment function refers to a mathematical function that calculates edge computing energy consumption and data transmission energy consumption based on the rate of change and disorder of seawater parameter data, as well as the computing power and network status of sensor nodes. The data processing method is determined by comparing the two energy consumption levels.

4. The edge computing optimization method for the seawater parameter monitoring IoT system according to claim 3, characterized in that, The dynamic energy consumption threshold judgment mechanism refers to an algorithm mechanism that automatically adjusts the decision thresholds for edge computing and data transmission based on real-time network conditions, battery power, and data importance.

5. The edge computing optimization method for the seawater parameter monitoring IoT system according to claim 4, characterized in that, The aforementioned seawater parameter spatiotemporal distribution model refers to a multidimensional data model constructed by combining geographical location information and time series data to reflect the variation patterns of seawater parameters at different locations and times.

6. The edge computing optimization method for the seawater parameter monitoring IoT system according to claim 5, characterized in that, The multi-objective optimization theory is to construct an energy consumption calculation delay balance function and use the Pareto optimal solution set method to update the calculation results of edge computing energy consumption and data transmission energy consumption.

7. The edge computing optimization method for the seawater parameter monitoring IoT system according to claim 6, characterized in that, The energy consumption and computation delay balancing function is used to achieve the optimal balance between energy consumption and computation delay during edge computing resource allocation. The inputs include edge computing energy consumption, data transmission energy consumption, data volume, network bandwidth status, and computation task complexity. The outputs are decision values ​​for various seawater parameter data processing strategies and corresponding resource allocation schemes.

8. The edge computing optimization method for the seawater parameter monitoring IoT system according to claim 7, characterized in that, The intelligent edge computing model for seawater parameter processing is a hybrid architecture combining a multilayer perceptron and a convolutional neural network. It includes a feature extraction layer, a temporal correlation analysis layer, a parameter correlation analysis layer, and a decision output layer. An attention mechanism is introduced into the network, and the weight coefficients are dynamically adjusted based on three key parameters: energy consumption assessment benchmark, data change rate, and network transmission status.

9. The edge computing optimization method for the seawater parameter monitoring IoT system according to claim 8, characterized in that, The terminal lightweight model refers to a simplified version of the intelligent edge computing model for water parameter processing deployed on the sensing and acquisition node. It uses depthwise separable convolution to replace the standard convolution operation, removes complex recursive structures, retains the core feature extraction function, and is optimized for energy-constrained scenarios.

10. The edge computing optimization method for the seawater parameter monitoring IoT system according to claim 9, characterized in that, The central complete model refers to the intelligent seawater parameter processing edge computing model deployed in the computing center with complete computing capabilities. It includes all network layer structures and parameters, performs complex data verification, fusion and analysis tasks, and is also responsible for providing parameter updates and optimization guidance to the terminal lightweight model.

Citation Information

Cited By

  • Internet of things edge computing dynamic load prediction data processing method and system

    CN121984884A

  • Communication parameter dynamic optimization method, device and system based on spatio-temporal data driving

    CN122317686A

  • Space-time data driven communication parameter dynamic optimization method, device and system

    CN122317686B