Marine ranch water quality parameter real-time correction and compensation method and system of multi-source sensor

Abstract

The invention provides a marine ranch water quality parameter real-time correction and compensation method and system for a multi-source sensor, and relates to the technical field of multi-source sensors, and the method comprises the steps: constructing a double-layer edge computing network, and connecting a sensor through a micro-service architecture to collect water quality data. And carrying out data preprocessing in combination with wavelet transform. And establishing a sensor digital twinborn model, and calculating the real-time credibility. Establishing a multi-dimensional sensor association network, optimizing a weight coefficient by adopting federal learning, and establishing a self-evolution correction parameter matrix; and fusing the sensor data by using a multi-task deep learning model to generate an initial correction value. And calculating a theoretical reference value through a space-time sequence prediction model. A compensation coefficient is adaptively adjusted by adopting a fuzzy decision tree, hierarchical water quality parameter correction is realized, and a closed-loop self-optimization intelligent correction system is formed through verification of a digital twin model. The accuracy and reliability of marine ranch water quality monitoring data are effectively improved.

Description

Technical Field

[0001] The present invention relates to multi-source sensor technology, and in particular to a real-time correction and compensation method and system for marine ranch water quality parameters using multi-source sensors. Background Art

[0002] Marine ranching is a comprehensive artificial fishery constructed in specific sea areas based on marine ecological principles and utilizing modern engineering technology. It integrates environmental protection, resource conservation, and fishery economic development. Water quality parameters are important indicators of the healthy development of marine ranching. Real-time and accurate monitoring of water quality parameters is crucial for protecting the living environment of marine life and improving production and quality. Traditional marine ranch water quality monitoring relies on manual sampling and laboratory analysis, which is inefficient, costly, and lacks real-time performance. With the development of the Internet of Things and sensor technology, more and more sensors are being deployed in marine ranches, enabling real-time collection of water quality parameters. However, due to the complexity of the marine environment and the limitations of sensor performance, the data collected by sensors often suffers from noise, outliers, and drift, which affects the accuracy of water quality monitoring.

[0003] The quality of sensor data varies: Due to differences in performance between different types of sensors, different deployment locations, and the complexity of the marine environment, the quality of data collected by sensors varies, and direct use will lead to inaccurate monitoring results.

[0004] Lack of multi-source data fusion mechanism: Most existing water quality monitoring systems rely solely on data from a single sensor and lack an effective multi-source data fusion mechanism, making it impossible to comprehensively utilize information from different sensors to improve monitoring accuracy.

[0005] Insufficient system adaptability: The marine environment changes dynamically, and the performance of sensors will also change over time. The existing water quality monitoring system lacks adaptability and is unable to cope with the complex and changing marine environment. Summary of the Invention

[0006] The embodiments of the present invention provide a method and system for real-time correction and compensation of marine ranch water quality parameters using multi-source sensors, which can solve the problems in the prior art.

[0007] According to a first aspect of the embodiments of the present invention,

[0008] Provides a real-time correction and compensation method for marine ranch water quality parameters using multiple source sensors, including:

[0009] A two-layer edge computing network, consisting of the surface layer and the water layer, was constructed in the marine ranch, with edge computing nodes deployed in each layer. Based on a microservices architecture, these edge computing nodes connected to temperature sensors, salinity sensors, and turbidity sensors to collect raw water quality data. These edge computing nodes used lightweight blockchain technology to distribute and store the raw water quality data, identified outliers through an annealing algorithm, and combined with multi-scale wavelet transforms to perform data denoising and feature extraction to generate a preprocessed data set. A sensor digital twin model was established to monitor the sensor's status parameters in real time, which served as the basis for subsequent credibility assessments.

[0010] The method comprises receiving the preprocessed data set and the state parameters of the sensor, constructing a dynamic reputation evaluation model for the sensor using a graph neural network, and calculating a real-time reputation based on the state parameters of the sensor; establishing a multidimensional sensor association network with vertical hierarchical association, horizontal spatial association, and parameter cross-association based on the real-time reputation, and optimizing the weight coefficients of the multidimensional sensor association network using a federated learning method; sharing the optimized weight information through incremental learning between edge nodes, and establishing a self-evolving correction parameter matrix, wherein the self-evolving correction parameter matrix dynamically adjusts the data contribution of sensors at each layer to achieve cross-layer collaborative computing;

[0011] Based on the self-evolving correction parameter matrix, a multi-task deep learning model is used to perform weighted fusion of sensor data, and an initial correction value is generated in combination with the real-time credibility; a spatiotemporal sequence prediction model is constructed to calculate the theoretical reference value, and an error analysis is performed between the initial correction value and the theoretical reference value; according to the error analysis results, a fuzzy decision tree is used to adaptively adjust the compensation coefficient in the self-evolving correction parameter matrix to achieve hierarchical water quality parameter correction; the correction result is input into the sensor digital twin model for verification, and the verification result is fed back to the sensor dynamic credibility evaluation model, and the real-time credibility and the weight coefficient are dynamically updated to form a closed-loop self-optimizing intelligent correction system.

[0012] Receiving the preprocessed data set and the state parameters of the sensor, constructing a dynamic sensor reputation evaluation model using a graph neural network, and calculating real-time reputation based on the state parameters of the sensor; establishing a multidimensional sensor association network with vertical hierarchical association, horizontal spatial association, and parameter cross-association based on the real-time reputation, and optimizing the weight coefficients of the multidimensional sensor association network using a federated learning method, including:

[0013] Receiving the preprocessed data set and the state parameters of the sensor, forming a feature vector from the preprocessed data set and the state parameters of the sensor; constructing a graph neural network for dynamic sensor reputation evaluation, inputting the feature vector into the graph neural network, the graph neural network performing spatial feature aggregation through a first convolutional layer to obtain sensor spatial correlation features, extracting temporal features through a second convolutional layer to obtain sensor performance change features, performing cross-parameter feature fusion through a third convolutional layer to obtain sensor correlation features, and calculating the real-time reputation of the sensor based on the sensor spatial correlation features, the sensor performance change features, and the sensor correlation features;

[0014] Establishing a multidimensional sensor association network based on the real-time reputation of the sensors, the multidimensional sensor association network comprising a vertical hierarchical association matrix, a horizontal spatial association matrix, and a parameter cross-correlation matrix, wherein the vertical hierarchical association matrix is constructed based on the vertical distance between sensors, the reputation difference, and the water mass mixing coefficient; the horizontal spatial association matrix is constructed based on the spatial correlation and flow field information of the Kriging interpolation algorithm; and the parameter cross-correlation matrix is constructed based on the physical and chemical relationship of water quality parameters and the influence of environmental factors;

[0015] A federated learning method is used to optimize the weight coefficients of the multidimensional sensor association network. The federated learning method uses the Adam optimizer to train sub-models at edge nodes and designs an adaptive learning rate adjustment strategy. The model parameters of each node are aggregated through a weighted averaging method to obtain a global model. The weights in the weighted averaging method are determined based on the node data quality and computing power. A differential privacy mechanism is introduced to protect the data. A multi-objective loss function is designed to train the global model. The global model is optimized using a dynamic pruning strategy and knowledge distillation technology, and the optimized weight coefficients of the multidimensional sensor association network are output.

[0016] By sharing optimized weight information through incremental learning among edge nodes, a self-evolving correction parameter matrix is established. The self-evolving correction parameter matrix dynamically adjusts the data contribution of sensors at each layer to achieve cross-layer collaborative computing, including:

[0017] An edge node incremental learning model is constructed based on a recurrent neural network. The edge node incremental learning model extracts the temporal features of sensor data through a state gate and a forget gate, calculates feature weights using an attention mechanism, and fuses sensor working state information and data credibility to generate a feature vector reflecting sensor performance.

[0018] A weight sharing mechanism is constructed between edge nodes. The local weight matrix of each edge node is calculated based on the characteristic vector of the sensor performance. A communication topology is established according to the physical distance between the nodes. The local weight matrix is propagated in the communication topology using a distributed consistency protocol. Important weight information is screened using compressed sensing technology. The influence of communication delay is eliminated through an asynchronous compensation mechanism to obtain optimized weight information shared among edge nodes.

[0019] Based on the shared optimization weight information, a self-evolution correction parameter matrix is established. The self-evolution correction parameter matrix uses a three-dimensional tensor structure to map the relationship between water layers, spatial positions, and parameter types. The shared optimization weight information is used as an initialization parameter. A fitness evaluation function is designed to calculate the importance of parameters. The matrix parameters are iteratively optimized through a genetic algorithm, and the optimization experience stored in the knowledge base is used to guide parameter updates.

[0020] The self-evolving correction parameter matrix calculates the inter-layer influence according to the hydrological dynamics model, and dynamically adjusts the data contribution of sensors at each layer in combination with the correction coefficients of environmental factors such as the thermocline. Based on the data contribution, the sensor sampling strategy is determined, and a sensor collaborative sampling scheme is established. A multi-objective optimization method is used to balance sampling accuracy and resource consumption, and cross-layer collaborative computing is performed.

[0021] The self-evolving correction parameter matrix calculates the inter-layer influence based on the hydrodynamic model and combines the correction coefficients of environmental factors such as the thermocline to dynamically adjust the data contribution of sensors at each layer. Based on the data contribution, the sensor sampling strategy is determined, and a sensor collaborative sampling scheme is established. A multi-objective optimization method is used to balance sampling accuracy and resource consumption. The cross-layer collaborative calculation is performed, including:

[0022] Constructing a hydrodynamic model, wherein the self-evolution correction parameter matrix uses the hydrodynamic model to describe the layered structure of the water body, uses density gradient to characterize interlayer exchange characteristics, calculates material migration flux through the vertical turbulent diffusion equation, and generates an initial interlayer influence parameter set;

[0023] A thermocline characteristic model is constructed based on the initial interlayer influence parameter set. The thermocline characteristic model identifies the location and strength of the thermocline through temperature profile data, analyzes the influence of the thermocline on sensor performance, calculates a measurement error compensation coefficient based on salinity and pressure changes, and outputs an environmental factor correction coefficient matrix.

[0024] Performing parameter correction on the environmental factor correction coefficient matrix and the initial inter-layer influence parameter set, dynamically calculating the data contribution of sensors at each layer based on the correction results using the self-evolving correction parameter matrix, evaluating data value using the information entropy criterion, determining data redundancy through mutual information analysis, and generating a real-time data contribution vector based on the data contribution, the data value, and the data redundancy;

[0025] constructing a sensor sampling strategy based on the real-time data contribution vector, allocating sampling resources based on the real-time data contribution vector, adjusting the measurement frequency using an adaptive sampling mechanism, optimizing data acquisition efficiency through triggered sampling control, and outputting an initial sampling scheme matrix;

[0026] Inputting the initial sampling scheme matrix into a collaborative sampling optimizer, the collaborative sampling optimizer establishes a sensor collaborative sampling scheme, adopts a hierarchical Markov decision process to coordinate sensor working states, optimizes the inter-layer collaboration mechanism based on a game theory method, and generates a collaborative sampling optimization framework;

[0027] Based on the collaborative sampling optimization framework, multi-objective optimization is performed, the optimization objective function of sampling accuracy and resource consumption is constructed, energy constraints and data quality thresholds are set, the particle swarm algorithm is used to solve the optimal sampling parameters, the optimized sampling instruction set is output, and the optimized sampling instruction set is executed to complete cross-level collaborative computing.

[0028] Based on the self-evolving correction parameter matrix, a multi-task deep learning model is used to perform weighted fusion on sensor data, and an initial correction value is generated in combination with the real-time credibility; a spatiotemporal sequence prediction model is constructed to calculate a theoretical reference value, and an error analysis is performed between the initial correction value and the theoretical reference value, including:

[0029] receiving a sensor feature representation vector, and constructing a multi-layer data fusion network based on weight coefficients in the self-evolving correction parameter matrix, wherein the multi-layer data fusion network includes a feature mapping layer and a weight calculation layer, wherein the feature mapping layer projects the sensor feature representation vector into a unified feature space, and the weight calculation layer generates an adaptive fusion weight in combination with the real-time reputation;

[0030] An adaptive normalization layer is used to balance the distribution of different sensor data. An attention mechanism is built to calculate the correlation scores between features. The fusion weights are dynamically adjusted based on the correlation scores. A multi-layer perceptron network is used to perform nonlinear transformation on the weighted features. The original feature information is retained through residual connections to generate initial correction values.

[0031] Constructing a spatiotemporal sequence prediction model, wherein the spatiotemporal sequence prediction model establishes a spatial association graph of sensor nodes based on the sensor feature representation vector, uses a graph convolutional network to extract topological features between nodes, aggregates local features using a diffuse convolution operator, and retains spatial information in combination with an attention pooling layer;

[0032] The spatiotemporal sequence prediction model uses a decoder and an encoder to process the sensor historical data sequence. The encoder captures long-term dependencies through a multi-head self-attention mechanism. The decoder fuses historical information and spatial features based on a cross-attention mechanism, predicts parameter change trends, outputs theoretical reference values, and compares and analyzes the initial correction value with the theoretical reference value.

[0033] Based on the error analysis results, a fuzzy decision tree is used to adaptively adjust the compensation coefficients in the self-evolving correction parameter matrix to achieve hierarchical water quality parameter correction. The correction results are input into the sensor digital twin model for verification, and the verification results are fed back to the sensor dynamic reputation evaluation model to dynamically update the real-time reputation and the weight coefficient, forming a closed-loop self-optimizing intelligent correction system including:

[0034] An error analysis network is constructed to process sensor data. The error analysis network uses wavelet transform to decompose the frequency components of the error sequence, combines variational inference to model the error distribution characteristics, and uses Gaussian process regression to predict the error propagation trend, generating an error feature vector containing error amplitude, variance, and trend characteristics.

[0035] A fuzzy decision tree is constructed based on the error feature vector. The fuzzy decision tree uses a Gaussian membership function to map the error feature vector to a fuzzy language space. A mapping relationship between the error feature vector and a compensation coefficient is established through an adaptive neuro-fuzzy inference system. A multi-level decision structure is constructed using an information gain criterion to output hierarchical correction rules.

[0036] The self-evolving correction parameter matrix is adjusted according to the hierarchical correction rule, the self-evolving correction parameter matrix includes initial compensation coefficients of water quality parameters, the compensation coefficients are updated based on the reasoning results of the fuzzy decision tree, a recursive feature elimination algorithm is used to optimize the feature selection strategy, and an adaptive boosting algorithm is used to enhance the compensation effect to obtain corrected water quality parameters;

[0037] Inputting the corrected water quality parameters into a sensor digital twin model, the sensor digital twin model integrates a physical model and a deep learning network, simulates the working mechanism of the sensor based on the physical model, uses the deep learning network to fit the dynamic response characteristics, compares the correction results with the predicted values through a state predictor, and generates verification features including correction deviation and stability assessment;

[0038] updating the sensor dynamic reputation assessment model based on the verification feature, wherein the sensor dynamic reputation assessment model uses multi-dimensional indicators to quantify sensor reliability, calculates real-time accuracy based on the correction deviation in the verification feature, determines a time decay coefficient based on stability assessment, and outputs a real-time reputation reflecting the current state of the sensor;

[0039] The real-time credibility is fed back into the self-evolving correction parameter matrix to dynamically adjust the weight coefficients of different sensor nodes. The weight coefficients are positively correlated with the real-time credibility. The parameter update speed is controlled by an adaptive learning rate. A multi-objective optimization algorithm is used to balance correction accuracy and system stability to achieve closed-loop optimization of the correction system.

[0040] The sensor dynamic reputation evaluation model is updated based on the verification feature. The sensor dynamic reputation evaluation model uses multi-dimensional indicators to quantify sensor reliability, calculates real-time accuracy based on the correction deviation in the verification feature, determines the time decay coefficient based on the stability evaluation, and outputs a real-time reputation reflecting the current state of the sensor, including:

[0041] Calculate the correction deviation based on the verification feature, the correction deviation is obtained by comparing the correction result with the standard value, use the root mean square error to evaluate the measurement accuracy, combine the relative error index to analyze the correction effect under different ranges, use the standard deviation to calculate the dispersion of the measurement result, and output the correction deviation feature vector;

[0042] Inputting the correction deviation feature vector into the sensor dynamic reputation evaluation model, the sensor dynamic reputation evaluation model uses multi-dimensional indicators to quantify sensor reliability, the multi-dimensional indicators including measurement accuracy indicator, response stability indicator and environmental adaptability indicator, and constructing an indicator fusion module through a multi-layer neural network to generate the real-time accuracy of the sensor;

[0043] A stability assessment is performed based on the real-time accuracy of the sensor. The stability assessment uses a time series analysis method to extract the sensor performance fluctuation characteristics, uses an autoregressive model to predict the performance change trend, evaluates the repeatability of the measurement results through variance analysis, and calculates a stability assessment score.

[0044] A time decay coefficient is determined based on the stability assessment score. The time decay coefficient increases as the stability assessment score decreases. An exponential decay function is used to construct a time weight model. The historical data weight is updated through a sliding time window to establish a dynamic decay mechanism. The time decay coefficient is weightedly fused with the real-time accuracy of the sensor, and a fuzzy comprehensive evaluation method is used to calculate the real-time credibility of the current state of the sensor.

[0045] According to a second aspect of the embodiments of the present invention,

[0046] Provides a real-time correction and compensation system for marine ranch water quality parameters using multiple source sensors, including:

[0047] The first unit is used to build a two-layer edge computing network in the marine ranch, with edge computing nodes deployed in each layer. The edge computing nodes are connected to temperature sensors, salinity sensors, and turbidity sensors based on a microservices architecture to collect raw water quality data. The edge computing nodes use lightweight blockchain technology to distribute and store the raw water quality data, identify outliers through an annealing algorithm, and perform data noise reduction and feature extraction in combination with multi-scale wavelet transform to generate a preprocessed data set. A sensor digital twin model is established to monitor the sensor's status parameters in real time. The sensor's status parameters serve as the basis for subsequent credibility assessment.

[0048] The second unit is configured to receive the preprocessed data set and the state parameters of the sensor, construct a dynamic reputation evaluation model for the sensor using a graph neural network, and calculate real-time reputation based on the state parameters of the sensor; establish a multidimensional sensor association network with vertical hierarchical association, horizontal spatial association, and parameter cross-association based on the real-time reputation, and optimize the weight coefficients of the multidimensional sensor association network using a federated learning method; establish a self-evolving correction parameter matrix by sharing the optimized weight information through incremental learning between edge nodes, and dynamically adjust the data contribution of sensors at each layer to achieve cross-layer collaborative computing;

[0049] The third unit is used to perform weighted fusion of sensor data using a multi-task deep learning model based on the self-evolution correction parameter matrix, and generate an initial correction value in combination with the real-time credibility; construct a spatiotemporal sequence prediction model to calculate the theoretical reference value, and perform error analysis between the initial correction value and the theoretical reference value; according to the error analysis results, use a fuzzy decision tree to adaptively adjust the compensation coefficient in the self-evolution correction parameter matrix to achieve hierarchical water quality parameter correction; input the correction result into the sensor digital twin model for verification, and feed the verification result back to the sensor dynamic credibility evaluation model, dynamically update the real-time credibility and the weight coefficient, and form a closed-loop self-optimizing intelligent correction system.

[0050] According to a third aspect of the embodiments of the present invention,

[0051] An electronic device is provided, comprising:

[0052] processor;

[0053] a memory for storing processor-executable instructions;

[0054] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0055] According to a fourth aspect of the embodiments of the present invention,

[0056] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0057] The beneficial effects of this application are as follows:

[0058] 1. Improved the accuracy and reliability of water quality parameter monitoring: This invention realizes real-time correction and compensation of raw water quality data by constructing a two-layer edge computing network and a sensor digital twin model, combining multi-scale wavelet transform, graph neural network and other technologies, effectively reducing the impact of sensor errors and environmental noise, and improving the accuracy and reliability of water quality parameter monitoring.

[0059] 2. Realized intelligent and adaptable water quality parameter monitoring: This invention uses intelligent algorithms such as federated learning, multi-task deep learning and fuzzy decision trees to construct a self-evolving correction parameter matrix and a spatiotemporal series prediction model, realizing dynamic adjustment of the contribution of different sensor data and intelligent prediction of water quality parameters, enabling the system to adapt to different marine environments and sensor states, and improving the intelligent level of water quality monitoring.

[0060] 3. Enhanced stability and maintainability of the water quality parameter monitoring system: This invention uses lightweight blockchain technology for distributed data storage, improving the security of the system; through incremental learning and sharing of optimized weight information among edge nodes, the system's self-evolution and continuous optimization are achieved; the closed-loop self-optimizing intelligent correction system design reduces system maintenance costs and improves the long-term stability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 This is a flow chart of a method for real-time correction and compensation of marine ranch water quality parameters using a multi-source sensor according to an embodiment of the present invention;

[0062] Figure 2 This is a structural diagram of a real-time correction and compensation system for marine ranch water quality parameters using multi-source sensors according to an embodiment of the present invention. DETAILED DESCRIPTION

[0063] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0064] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0065] Figure 1 FIG. 1 is a flow chart of a method for real-time correction and compensation of water quality parameters of a marine ranch using a multi-source sensor according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0066] S11. Construct a two-layer edge computing network at the marine ranch, with edge computing nodes deployed in each layer. These edge computing nodes connect to temperature, salinity, and turbidity sensors based on a microservices architecture to collect raw water quality data. These edge computing nodes utilize lightweight blockchain technology for distributed storage of this raw water quality data, identify outliers using an annealing algorithm, and perform data denoising and feature extraction in conjunction with a multi-scale wavelet transform to generate a preprocessed dataset. A sensor digital twin model is established to monitor sensor status parameters in real time, with these sensor status parameters serving as the basis for subsequent reputation assessments.

[0067] S12. Receive the preprocessed dataset and the sensor state parameters, construct a dynamic sensor reputation assessment model using a graph neural network, and calculate real-time reputation based on the sensor state parameters; establish a multidimensional sensor association network based on the real-time reputation, with vertical hierarchical associations, horizontal spatial associations, and parameter cross-associations; optimize the weight coefficients of the multidimensional sensor association network using a federated learning method; share the optimized weight information through incremental learning between edge nodes, and establish a self-evolving correction parameter matrix. The self-evolving correction parameter matrix dynamically adjusts the data contribution of sensors at each layer to achieve cross-layer collaborative computing;

[0068] S13. Based on the self-evolution correction parameter matrix, a multi-task deep learning model is used to perform weighted fusion of sensor data, and an initial correction value is generated in combination with the real-time credibility; a spatiotemporal sequence prediction model is constructed to calculate a theoretical reference value, and an error analysis is performed between the initial correction value and the theoretical reference value; according to the error analysis results, a fuzzy decision tree is used to adaptively adjust the compensation coefficient in the self-evolution correction parameter matrix to achieve hierarchical water quality parameter correction; the correction result is input into the sensor digital twin model for verification, and the verification result is fed back to the sensor dynamic credibility evaluation model, and the real-time credibility and the weight coefficient are dynamically updated to form a closed-loop self-optimizing intelligent correction system.

[0069] In an optional embodiment, receiving the preprocessed data set and the state parameters of the sensor, constructing a dynamic sensor reputation evaluation model using a graph neural network, and calculating real-time reputation based on the state parameters of the sensor; establishing a multidimensional sensor association network with vertical hierarchical association, horizontal spatial association, and parameter cross-association based on the real-time reputation, and optimizing weight coefficients of the multidimensional sensor association network using a federated learning method include:

[0070] Receiving the preprocessed data set and the state parameters of the sensor, forming a feature vector from the preprocessed data set and the state parameters of the sensor; constructing a graph neural network for dynamic sensor reputation evaluation, inputting the feature vector into the graph neural network, the graph neural network performing spatial feature aggregation through a first convolutional layer to obtain sensor spatial correlation features, extracting temporal features through a second convolutional layer to obtain sensor performance change features, performing cross-parameter feature fusion through a third convolutional layer to obtain sensor correlation features, and calculating the real-time reputation of the sensor based on the sensor spatial correlation features, the sensor performance change features, and the sensor correlation features;

[0071] Establishing a multidimensional sensor association network based on the real-time reputation of the sensors, the multidimensional sensor association network comprising a vertical hierarchical association matrix, a horizontal spatial association matrix, and a parameter cross-correlation matrix, wherein the vertical hierarchical association matrix is constructed based on the vertical distance between sensors, the reputation difference, and the water mass mixing coefficient; the horizontal spatial association matrix is constructed based on the spatial correlation and flow field information of the Kriging interpolation algorithm; and the parameter cross-correlation matrix is constructed based on the physical and chemical relationship of water quality parameters and the influence of environmental factors;

[0072] A federated learning method is used to optimize the weight coefficients of the multidimensional sensor association network. The federated learning method uses the Adam optimizer to train sub-models at edge nodes and designs an adaptive learning rate adjustment strategy. The model parameters of each node are aggregated through a weighted averaging method to obtain a global model. The weights in the weighted averaging method are determined based on the node data quality and computing power. A differential privacy mechanism is introduced to protect the data. A multi-objective loss function is designed to train the global model. The global model is optimized using a dynamic pruning strategy and knowledge distillation technology, and the optimized weight coefficients of the multidimensional sensor association network are output.

[0073] First, data preprocessing and feature vector construction. Collect observation data from different sensors, such as temperature, pH value, dissolved oxygen, etc. Preprocess the collected raw data, including data cleaning, missing value filling, and outlier processing. For example, use linear interpolation to fill missing values and use the 3σ criterion to eliminate outliers. Combine the preprocessed data set and the sensor's state parameters (such as sensor operating time, battery voltage, etc.) into a feature vector. Assuming that a sensor's data set contains three parameters: temperature, pH value, and dissolved oxygen, and each parameter has 10 time steps of data, the dimension of the sensor's feature vector is 33 (3 parameters * 10 time steps + 3 state parameters).

[0074] Next, a graph neural network is constructed to perform dynamic sensor reputation assessment. A three-layer convolutional graph neural network model is constructed. The first convolutional layer aggregates sensor spatial information, such as the distance and connectivity between sensors. The second convolutional layer extracts sensor performance variation characteristics, such as the time series trend of sensor readings. The third convolutional layer integrates correlation information between different parameters, such as the relationship between temperature and dissolved oxygen. The constructed feature vector is input into the graph neural network. Through three layers of convolution, the graph neural network obtains sensor spatial correlation features, sensor performance variation features, and sensor correlation features, respectively. Assuming the output dimension of the first convolutional layer is 16, the output dimension of the second convolutional layer is 8, and the output dimension of the third convolutional layer is 4, these three feature vectors are concatenated into a 28-dimensional vector to represent the comprehensive sensor features. Based on this comprehensive feature, the real-time sensor reputation is calculated. For example, a softmax function is used to convert the feature vector into a probability value between 0 and 1, representing the sensor reputation.

[0075] Subsequently, a multidimensional sensor association network is established based on the real-time sensor credibility. Based on the calculated real-time sensor credibility, a vertical hierarchical correlation matrix, a horizontal spatial correlation matrix, and a parameter cross-correlation matrix are constructed to form the multidimensional sensor association network. The vertical hierarchical correlation matrix considers the vertical distance between sensors, the credibility difference, and the water mass mixing coefficient. For example, the credibility correlation strength between sensors at different water depths is related to the water mass mixing coefficient. The horizontal spatial correlation matrix considers the spatial correlation between sensors and flow field information. For example, the spatial correlation between sensors is calculated using the Kriging interpolation algorithm and combined with the flow field information to construct the horizontal spatial correlation matrix. The parameter cross-correlation matrix considers the physicochemical relationships between water quality parameters and the influence of environmental factors. For example, there is a certain negative correlation between temperature and dissolved oxygen, and a parameter cross-correlation matrix can be constructed based on this relationship.

[0076] Finally, a federated learning approach is used to optimize the weight coefficients of the multidimensional sensor association network. On each edge node, a local model is trained using the Adam optimizer, and an adaptive learning rate adjustment strategy is designed. For example, the learning rate is dynamically adjusted based on the training loss of the local model. The model parameters trained at each edge node are uploaded to a central server. The central server aggregates the model parameters of each node using a weighted average method to generate a global model. Weights are determined based on the node's data quality and computing power, with nodes with higher data quality receiving greater weights. To protect data privacy, a differential privacy mechanism is introduced. For example, a certain amount of noise is added before uploading the model parameters. The global model is trained using a designed multi-objective loss function, for example, considering both model accuracy and robustness. Dynamic pruning strategies and knowledge distillation techniques are used to optimize the global model. For example, model parameters are pruned based on their importance and knowledge distillation is used to transfer knowledge from complex models to simpler ones. Finally, the optimized weight coefficients of the multidimensional sensor association network are output.

[0077] The solution of this application can:

[0078] Improving the accuracy of sensor reputation assessment: By fusing multi-source information through graph neural networks, we can more accurately assess the real-time reputation of sensors, thereby identifying faulty sensors and improving data quality. Enhancing the reliability of sensor association analysis: By building a multi-dimensional sensor association network, we can more comprehensively analyze the associations between sensors, thereby better understanding the temporal and spatial variations of water quality parameters. Protecting data privacy and security: Using federated learning methods, we can perform model training without leaking the original data, thereby protecting data privacy and security.

[0079] In an optional embodiment, the optimized weight information is shared through incremental learning between edge nodes to establish a self-evolving correction parameter matrix. The self-evolving correction parameter matrix dynamically adjusts the data contribution of sensors at each layer to achieve cross-layer collaborative computing, including:

[0080] An edge node incremental learning model is constructed based on a recurrent neural network. The edge node incremental learning model extracts the temporal features of sensor data through a state gate and a forget gate, calculates feature weights using an attention mechanism, and fuses sensor working state information and data credibility to generate a feature vector reflecting sensor performance.

[0081] A weight sharing mechanism is constructed between edge nodes. The local weight matrix of each edge node is calculated based on the characteristic vector of the sensor performance. A communication topology is established according to the physical distance between the nodes. The local weight matrix is propagated in the communication topology using a distributed consistency protocol. Important weight information is screened using compressed sensing technology. The influence of communication delay is eliminated through an asynchronous compensation mechanism to obtain optimized weight information shared among edge nodes.

[0082] Based on the shared optimization weight information, a self-evolution correction parameter matrix is established. The self-evolution correction parameter matrix uses a three-dimensional tensor structure to map the relationship between water layers, spatial positions, and parameter types. The shared optimization weight information is used as an initialization parameter. A fitness evaluation function is designed to calculate the importance of parameters. The matrix parameters are iteratively optimized through a genetic algorithm, and the optimization experience stored in the knowledge base is used to guide parameter updates.

[0083] The self-evolving correction parameter matrix calculates the inter-layer influence according to the hydrological dynamics model, and dynamically adjusts the data contribution of sensors at each layer in combination with the correction coefficients of environmental factors such as the thermocline. Based on the data contribution, the sensor sampling strategy is determined, and a sensor collaborative sampling scheme is established. A multi-objective optimization method is used to balance sampling accuracy and resource consumption, and cross-layer collaborative computing is performed.

[0084] First, an incremental learning model is built on each edge node. This model uses a recurrent neural network structure with state gates and forget gates to extract time series features from sensor data. For example, for a temperature sensor, the model can learn temperature trends over time and periodic fluctuations. The model also incorporates an attention mechanism to assign different weights based on feature importance. For example, if the temperature changes suddenly, the model assigns a higher weight. Furthermore, the model incorporates sensor operating status information and data reliability. For example, if a sensor is faulty or the data is abnormal, the model reduces its weight. Ultimately, the model generates a feature vector that reflects the overall performance of the sensor.

[0085] Next, a weight sharing mechanism is established between edge nodes. Based on the sensor performance feature vector generated by each edge node, its local weight matrix is calculated. Assume there are three edge nodes, each monitoring water temperature, pH, and dissolved oxygen. Each node's local weight matrix reflects the importance of the parameter it monitors. Next, a communication topology is established based on the physical distance between nodes. For example, if the three nodes are close together, a fully connected graph can be constructed. Within this communication topology, a distributed consensus protocol is used to propagate the local weight matrices. For example, each node can send its local weight matrix to its neighboring nodes and receive their weight matrices. To reduce communication overhead, compressed sensing technology is used to filter important weight information. For example, only elements with large absolute weight values are transmitted. Furthermore, an asynchronous compensation mechanism is used to mitigate the impact of communication delays. For example, the received weight information is weighted averaged based on the delay time. Ultimately, optimized weight information shared among the edge nodes is obtained.

[0086] Then, a self-evolving correction parameter matrix is established based on the shared optimization weight information. This matrix uses a three-dimensional tensor structure to map the relationships between water layers, spatial locations, and parameter types. For example, an element of the matrix can represent the correction coefficient for the temperature parameter at a location in the surface waters. The shared optimization weight information is used as the initialization parameter. A fitness evaluation function is designed to calculate the importance of the parameters. For example, the importance of a parameter can be evaluated based on its impact on the model's prediction accuracy. A genetic algorithm is used to iteratively optimize the matrix parameters. For example, new parameter combinations are generated through operations such as crossover and mutation, and the parameter combinations with higher fitness are selected. A knowledge base is used to store optimization experience to guide parameter updates. For example, parameter combinations that perform well can be recorded and prioritized in subsequent optimization processes.

[0087] Finally, a self-evolving correction parameter matrix calculates interlayer influences based on the hydrodynamic model and, incorporating correction coefficients for environmental factors such as thermoclines, dynamically adjusts the data contribution of sensors at each layer. For example, if the bottom water temperature has a greater impact on the surface water temperature, the data contribution of the bottom sensors will be increased. Based on the data contribution, a sensor sampling strategy is determined, and a sensor collaborative sampling scheme is established. For example, data from sensors with higher data contributions can be prioritized. A multi-objective optimization method is used to balance sampling accuracy and resource consumption. For example, while maintaining a certain level of sampling accuracy, the sampling frequency can be minimized to reduce energy consumption. Finally, cross-layer collaborative computing is performed.

[0088] The solution of this application can:

[0089] Improved data quality: Through incremental learning and weight sharing mechanisms, data from different sensors can be effectively fused, and noise and outliers can be removed, thereby improving data accuracy and reliability. For example, in one case, this method reduced the root mean square error of water quality prediction by 20%. Reduced computational cost: Through compressed sensing and asynchronous compensation mechanisms, communication overhead and computational burden can be reduced, thereby reducing the energy consumption and cost of the system. For example, in one case, this method reduced the energy consumption of edge nodes by 15%. Enhanced system adaptability: Through the self-evolution correction parameter matrix, the data contribution and sampling strategy of the sensor can be dynamically adjusted according to environmental changes, thereby enhancing the adaptability and robustness of the system. For example, in one case, even when some sensors failed, the method was still able to maintain a high prediction accuracy.

[0090] In an optional embodiment, the self-evolving correction parameter matrix calculates the inter-layer influence based on the hydrodynamic model, combines the correction coefficients of environmental factors such as the thermocline, dynamically adjusts the data contribution of sensors at each layer, determines the sensor sampling strategy based on the data contribution, establishes a sensor collaborative sampling scheme, and uses a multi-objective optimization method to balance sampling accuracy and resource consumption. Executing cross-layer collaborative computing includes:

[0091] Constructing a hydrodynamic model, wherein the self-evolution correction parameter matrix uses the hydrodynamic model to describe the layered structure of the water body, uses density gradient to characterize interlayer exchange characteristics, calculates material migration flux through the vertical turbulent diffusion equation, and generates an initial interlayer influence parameter set;

[0092] A thermocline characteristic model is constructed based on the initial interlayer influence parameter set. The thermocline characteristic model identifies the location and strength of the thermocline through temperature profile data, analyzes the influence of the thermocline on sensor performance, calculates a measurement error compensation coefficient based on salinity and pressure changes, and outputs an environmental factor correction coefficient matrix.

[0093] Performing parameter correction on the environmental factor correction coefficient matrix and the initial inter-layer influence parameter set, dynamically calculating the data contribution of sensors at each layer based on the correction results using the self-evolving correction parameter matrix, evaluating data value using the information entropy criterion, determining data redundancy through mutual information analysis, and generating a real-time data contribution vector based on the data contribution, the data value, and the data redundancy;

[0094] constructing a sensor sampling strategy based on the real-time data contribution vector, allocating sampling resources based on the real-time data contribution vector, adjusting the measurement frequency using an adaptive sampling mechanism, optimizing data acquisition efficiency through triggered sampling control, and outputting an initial sampling scheme matrix;

[0095] Inputting the initial sampling scheme matrix into a collaborative sampling optimizer, the collaborative sampling optimizer establishes a sensor collaborative sampling scheme, adopts a hierarchical Markov decision process to coordinate sensor working states, optimizes the inter-layer collaboration mechanism based on a game theory method, and generates a collaborative sampling optimization framework;

[0096] Based on the collaborative sampling optimization framework, multi-objective optimization is performed, the optimization objective function of sampling accuracy and resource consumption is constructed, energy constraints and data quality thresholds are set, the particle swarm algorithm is used to solve the optimal sampling parameters, the optimized sampling instruction set is output, and the optimized sampling instruction set is executed to complete cross-level collaborative computing.

[0097] First, a hydrodynamic model is constructed. This model describes the layered structure of the water body and uses density gradients to characterize the interlayer material exchange characteristics. By simulating the vertical turbulent diffusion process of substances in the water body, the material migration flux between different water layers is calculated, thereby generating an initial set of interlayer influence parameters. For example, if the water body is divided into three layers: the surface layer, the middle layer, and the bottom layer, the model calculation shows that the influence of the surface layer on the middle layer is 0.8, the influence of the surface layer on the bottom layer is 0.2, and the influence of the middle layer on the bottom layer is 0.6. And so on, forming an initial set of interlayer influence parameters.

[0098] Next, a thermocline characteristic model is constructed based on the initial set of interlayer influence parameters. Temperature profile data acquired by the temperature sensor is used to identify the location and strength of the thermocline. For example, by analyzing the temperature data, a region with a large temperature gradient at a depth of 5 meters is identified as a thermocline. The influence of the thermocline on sensor performance is then analyzed. For example, the thermocline may affect the measurement accuracy of the acoustic sensor. Combining changes in environmental factors such as salinity and pressure, measurement error compensation coefficients are calculated, ultimately outputting a matrix of environmental factor correction coefficients. For example, the influence coefficient of the thermocline on the acoustic sensor is 0.9, the influence coefficient of salinity changes is 0.85, and the influence coefficient of pressure changes is 0.95. These coefficients are then combined into a matrix.

[0099] Next, the environmental factor correction coefficient matrix and the initial inter-layer influence parameter set are subjected to parameter correction to obtain a self-evolving correction parameter matrix. This matrix dynamically calculates the data contribution of sensors at each layer. Data value is assessed using the information entropy criterion; for example, higher information entropy indicates greater data value. Data redundancy is determined through mutual information analysis; for example, higher mutual information indicates greater data redundancy. Based on data contribution, data value, and data redundancy, a real-time data contribution vector is generated. For example, the real-time data contributions of surface, mid-layer, and bottom-layer sensors are 0.7, 0.5, and 0.3, respectively.

[0100] Subsequently, a sensor sampling strategy is constructed based on the real-time data contribution vector. Sampling resources are allocated based on the real-time data contribution vector, for example, sensors with high contribution rates are allocated more sampling time. An adaptive sampling mechanism is used to adjust the measurement frequency, for example, increasing the sampling frequency when water quality parameters fluctuate dramatically. Data collection efficiency is optimized through triggered sampling control, for example, triggering a sensor to sample when a parameter exceeds a preset threshold. Finally, an initial sampling plan matrix is output, for example, specifying that surface sensors sample once an hour, mid-layer sensors sample every two hours, and bottom-layer sensors sample every four hours.

[0101] The initial sampling plan matrix is input into the collaborative sampling optimizer. This optimizer establishes a sensor collaborative sampling plan and uses a hierarchical Markov decision process to coordinate sensor operating states. For example, it selects the appropriate sensor combination for sampling based on the current water state. Game theory is used to optimize the inter-layer coordination mechanism, for example, by simulating the game process between different sensors to find the optimal coordination strategy. Ultimately, a collaborative sampling optimization framework is generated.

[0102] Finally, a multi-objective optimization solution is performed based on the collaborative sampling optimization framework. An optimization objective function is constructed for sampling accuracy and resource consumption, for example, maximizing sampling accuracy and minimizing energy consumption. Energy constraints and data quality thresholds are set, for example, limiting total energy consumption to no more than 100 watt-hours and requiring data accuracy to be higher than 90%. Optimization algorithms such as the particle swarm optimization algorithm are used to solve for the optimal sampling parameters. For example, the optimal sampling frequency and sensor combination are found through iterative calculations. Finally, an optimized sampling instruction set is output, for example, instructing surface sensors to sample at specific times and transmit the data to the data center. The optimized sampling instruction set is executed to complete cross-level collaborative computing.

[0103] The solution of this application can:

[0104] Improved water monitoring accuracy: By dynamically adjusting sensor data contributions and collaborative sampling strategies, changes in water parameters can be effectively captured, improving the accuracy and reliability of monitoring data. Reduced resource consumption: By optimizing sampling frequency and sensor combinations, energy consumption and data transmission volume can be reduced, lowering monitoring costs. Enhanced system adaptability: This method can dynamically adjust sampling strategies based on water environment and monitoring requirements, improving system adaptability and robustness.

[0105] In an optional embodiment, based on the self-evolving correction parameter matrix, a multi-task deep learning model is used to perform weighted fusion on the sensor data, and an initial correction value is generated in combination with the real-time credibility; a spatiotemporal sequence prediction model is constructed to calculate a theoretical reference value, and an error analysis is performed between the initial correction value and the theoretical reference value, including:

[0106] receiving a sensor feature representation vector, and constructing a multi-layer data fusion network based on weight coefficients in the self-evolving correction parameter matrix, wherein the multi-layer data fusion network includes a feature mapping layer and a weight calculation layer, wherein the feature mapping layer projects the sensor feature representation vector into a unified feature space, and the weight calculation layer generates an adaptive fusion weight in combination with the real-time reputation;

[0107] An adaptive normalization layer is used to balance the distribution of different sensor data. An attention mechanism is built to calculate the correlation scores between features. The fusion weights are dynamically adjusted based on the correlation scores. A multi-layer perceptron network is used to perform nonlinear transformation on the weighted features. The original feature information is retained through residual connections to generate initial correction values.

[0108] Constructing a spatiotemporal sequence prediction model, wherein the spatiotemporal sequence prediction model establishes a spatial association graph of sensor nodes based on the sensor feature representation vector, uses a graph convolutional network to extract topological features between nodes, aggregates local features using a diffuse convolution operator, and retains spatial information in combination with an attention pooling layer;

[0109] The spatiotemporal sequence prediction model uses a decoder and an encoder to process the sensor historical data sequence. The encoder captures long-term dependencies through a multi-head self-attention mechanism. The decoder fuses historical information and spatial features based on a cross-attention mechanism, predicts parameter change trends, outputs theoretical reference values, and compares and analyzes the initial correction value with the theoretical reference value.

[0110] First, acquire sensor data and convert it into a feature vector. For example, a temperature sensor can provide information such as temperature, humidity, and timestamp, which can be combined into a feature vector. Suppose there are two sensors. Sensor 1's feature vector is [25, 80, 1678886400], and sensor 2's feature vector is [26, 75, 1678886400], representing temperature, humidity, and timestamp, respectively.

[0111] Next, multi-layer data fusion is performed based on a self-evolving correction parameter matrix. This matrix stores the weight coefficients for each sensor, reflecting its reliability. The initial matrix can be set to the identity matrix and subsequently dynamically adjusted based on sensor performance. The multi-layer data fusion network consists of a feature mapping layer and a weight calculation layer. The feature mapping layer projects different types of sensor data into a unified feature space. For example, temperature and humidity are converted to numerical values between 0 and 1. The weight calculation layer combines the real-time reputation of the sensors to generate adaptive fusion weights. Assuming that the real-time reputation of sensor 1 is 0.9 and that of sensor 2 is 0.8, the adaptive fusion weights for sensor 1 and sensor 2 are calculated to be 0.55 and 0.45, respectively, based on the parameter matrix.

[0112] Next, data preprocessing and feature fusion are performed. An adaptive normalization layer is used to balance the distribution of data from different sensors. For example, all sensor data is normalized to a standard normal distribution with a mean of 0 and a variance of 1. An attention mechanism is implemented to calculate correlation scores between features. For example, if temperature and humidity are likely to have a positive correlation, their correlation scores will be high. Fusion weights are dynamically adjusted based on the correlation scores. For example, if temperature and humidity are highly correlated, the weight of one feature is reduced to avoid information redundancy. A multilayer perceptron network is used to perform a nonlinear transformation on the weighted features. Residual connections are used to preserve the original feature information, ultimately generating an initial correction value. Assume that after the above steps, the initial correction temperature value is 25.5 degrees Celsius.

[0113] Next, a spatiotemporal series prediction model is constructed to calculate theoretical reference values. This model establishes a spatial correlation graph of sensor nodes based on sensor feature representation vectors. For example, connections exist between geographically close sensor nodes. A graph convolutional network (GCN) is used to extract topological features between nodes, and the diffuse convolution operator is used to aggregate local features. This is combined with an attention pooling layer to preserve spatial information. Assuming that sensor 1 and sensor 2 are geographically close, their topological features are extracted through the GCN and aggregated to obtain local features.

[0114] The spatiotemporal series prediction model uses an encoder and decoder to process historical sensor data sequences. The encoder uses a multi-head self-attention mechanism to capture long-term dependencies, such as temperature trends over the past few days. The decoder uses a cross-attention mechanism to fuse historical information and spatial features, predict parameter trends, and output theoretical reference values. For example, based on historical data and spatial features, it predicts temperature trends over a period of time and outputs a theoretical reference temperature value of 25.2 degrees Celsius.

[0115] Finally, the initial correction value is compared and analyzed with the theoretical reference value. For example, the difference between the initial correction value of 25.5 degrees and the theoretical reference value of 25.2 degrees is calculated to evaluate the correction effect and adjust the self-evolution correction parameter matrix to provide more accurate weight coefficients for the next round of correction.

[0116] The solution of this application can:

[0117] Improved sensor data accuracy: A multi-task deep learning model and spatiotemporal series prediction model effectively identify and correct errors in sensor data, improving data accuracy and reliability. Enhanced system robustness: Adaptive fusion weights and a real-time credibility assessment mechanism effectively reduce the impact of abnormal sensor data on the system, enhancing its robustness and stability. Intelligent correction: A self-evolving correction parameter matrix dynamically adjusts weight coefficients based on sensor performance, enabling intelligent correction of sensor data and reducing the need for manual intervention.

[0118] In an optional embodiment, based on the error analysis results, a fuzzy decision tree is used to adaptively adjust the compensation coefficients in the self-evolving correction parameter matrix to achieve hierarchical water quality parameter correction; the correction results are input into the sensor digital twin model for verification, and the verification results are fed back to the sensor dynamic reputation evaluation model to dynamically update the real-time reputation and the weight coefficient, forming a closed-loop self-optimizing intelligent correction system including:

[0119] An error analysis network is constructed to process sensor data. The error analysis network uses wavelet transform to decompose the frequency components of the error sequence, combines variational inference to model the error distribution characteristics, and uses Gaussian process regression to predict the error propagation trend, generating an error feature vector containing error amplitude, variance, and trend characteristics.

[0120] A fuzzy decision tree is constructed based on the error feature vector. The fuzzy decision tree uses a Gaussian membership function to map the error feature vector to a fuzzy language space. A mapping relationship between the error feature vector and a compensation coefficient is established through an adaptive neuro-fuzzy inference system. A multi-level decision structure is constructed using an information gain criterion to output hierarchical correction rules.

[0121] The self-evolving correction parameter matrix is adjusted according to the hierarchical correction rule, the self-evolving correction parameter matrix includes initial compensation coefficients of water quality parameters, the compensation coefficients are updated based on the reasoning results of the fuzzy decision tree, a recursive feature elimination algorithm is used to optimize the feature selection strategy, and an adaptive boosting algorithm is used to enhance the compensation effect to obtain corrected water quality parameters;

[0122] Inputting the corrected water quality parameters into a sensor digital twin model, the sensor digital twin model integrates a physical model and a deep learning network, simulates the working mechanism of the sensor based on the physical model, uses the deep learning network to fit the dynamic response characteristics, compares the correction results with the predicted values through a state predictor, and generates verification features including correction deviation and stability assessment;

[0123] updating the sensor dynamic reputation assessment model based on the verification feature, wherein the sensor dynamic reputation assessment model uses multi-dimensional indicators to quantify sensor reliability, calculates real-time accuracy based on the correction deviation in the verification feature, determines a time decay coefficient based on stability assessment, and outputs a real-time reputation reflecting the current state of the sensor;

[0124] The real-time credibility is fed back into the self-evolving correction parameter matrix to dynamically adjust the weight coefficients of different sensor nodes. The weight coefficients are positively correlated with the real-time credibility. The parameter update speed is controlled by an adaptive learning rate. A multi-objective optimization algorithm is used to balance correction accuracy and system stability to achieve closed-loop optimization of the correction system.

[0125] First, an error analysis network is constructed. This network processes sensor data in three steps: First, wavelet transform is used to decompose the original error sequence into subsequences of different frequencies, separating high-frequency noise and low-frequency trends. Second, variational inference is used to fit the probability distribution of each frequency subsequence and extract the error amplitude and variance characteristics. Third, a Gaussian process regression model is used to predict the future trend of each frequency subsequence. The prediction results are integrated to generate an error feature vector containing the error amplitude, variance, and trend characteristics. For example, if the error sequence between a sensor measurement and the true value is [-0.1, 0.2, -0.3, 0.1, 0.2], wavelet transform decomposition results in a high-frequency component [-0.05, 0.1, -0.15, 0.05, 0.1] and a low-frequency component [-0.05, 0.1, -0.15, 0.05, 0.1]. Variational inference is used to estimate the amplitude and variance of these components, and Gaussian process regression is used to predict the future trend. Ultimately, the components are integrated into a single error feature vector.

[0126] Next, a fuzzy decision tree is constructed based on the error feature vector. First, a Gaussian membership function is used to map the numerical features in the error feature vector to a fuzzy language space. For example, the error amplitude is divided into three fuzzy sets: "small", "medium", and "large". Then, an adaptive neuro-fuzzy inference system is used to establish a mapping relationship between the error feature vector and the compensation coefficient, and to learn fuzzy rules, such as "if the error amplitude is 'large' and the trend is 'increasing', then the compensation coefficient is 0.8". Finally, the optimal features and split points are selected according to the information gain criterion, and a multi-level decision tree structure is constructed to output hierarchical correction rules. For example, the root node of the decision tree is divided according to the error amplitude, with the left subtree corresponding to "small" and the right subtree corresponding to "large". It is further divided into finer levels according to the error trend, and finally the compensation coefficients for different situations are output.

[0127] Subsequently, the self-evolving correction parameter matrix is adjusted according to the hierarchical correction rules. This matrix contains the initial compensation coefficients for each water quality parameter. Based on the inference results of the fuzzy decision tree, the corresponding compensation coefficients are updated. Simultaneously, a recursive feature elimination algorithm is used to evaluate the contribution of each feature to the correction effect, remove redundant features, and optimize the feature selection strategy. Furthermore, an adaptive boosting algorithm is used to iteratively adjust the compensation coefficients, gradually improving the correction accuracy and ultimately obtaining the corrected water quality parameters. For example, if the initial compensation coefficient is 0.5 and is adjusted to 0.8 based on the correction rules output by the fuzzy decision tree, the final compensation coefficient might be 0.75 after recursive feature elimination and adaptive boosting.

[0128] The corrected water quality parameters are then input into the sensor's digital twin model for verification. This model integrates a physical model and a deep learning network. The physical model simulates the sensor's operating mechanism, such as the redox reaction of an electrochemical sensor. The deep learning network fits the sensor's dynamic response characteristics under different operating conditions, such as the influence of factors such as temperature and pressure. A state predictor compares the correction results with the model's predicted values, generating validation features that include correction deviation and stability assessment. For example, the corrected dissolved oxygen concentration is input into the digital twin model to predict its future trend and compare it with the actual measured value to evaluate the correction effect.

[0129] Finally, the sensor's dynamic reputation assessment model is updated based on the verification features. This model uses multi-dimensional metrics to quantify sensor reliability, such as accuracy, stability, and response time. Real-time accuracy is calculated based on the correction deviation in the verification features, and a time decay coefficient is determined based on the stability assessment. The output reflects the sensor's current state. The real-time reputation is fed back into the self-evolving correction parameter matrix to dynamically adjust the weight coefficients of different sensor nodes. The weight coefficients are positively correlated with the real-time reputation. An adaptive learning rate is used to control the parameter update speed, and a multi-objective optimization algorithm is employed to balance correction accuracy and system stability, achieving closed-loop optimization of the correction system. For example, if a sensor's real-time reputation decreases, its weight coefficient is reduced to minimize its impact on the final correction result.

[0130] The solution of this application can:

[0131] Improved calibration accuracy: The error analysis network accurately captures error characteristics, and combined with a fuzzy decision tree, hierarchical calibration effectively improves the calibration accuracy of water quality parameters. Enhanced system stability: The linkage between the sensor's dynamic reputation assessment model and the self-evolving calibration parameter matrix achieves closed-loop self-optimization, enhancing the system's long-term stability and reliability. Reduced maintenance costs: Intelligent calibration technology reduces the need for manual intervention and calibration, reducing sensor maintenance costs.

[0132] In an optional embodiment, the sensor dynamic reputation assessment model is updated based on the verification feature. The sensor dynamic reputation assessment model uses multi-dimensional indicators to quantify sensor reliability, calculates real-time accuracy in combination with the correction deviation in the verification feature, determines a time decay coefficient based on a stability assessment, and outputs a real-time reputation reflecting the current state of the sensor, including:

[0133] Calculate the correction deviation based on the verification feature, the correction deviation is obtained by comparing the correction result with the standard value, use the root mean square error to evaluate the measurement accuracy, combine the relative error index to analyze the correction effect under different ranges, use the standard deviation to calculate the dispersion of the measurement result, and output the correction deviation feature vector;

[0134] Inputting the correction deviation feature vector into the sensor dynamic reputation evaluation model, the sensor dynamic reputation evaluation model uses multi-dimensional indicators to quantify sensor reliability, the multi-dimensional indicators including measurement accuracy indicator, response stability indicator and environmental adaptability indicator, and constructing an indicator fusion module through a multi-layer neural network to generate the real-time accuracy of the sensor;

[0135] A stability assessment is performed based on the real-time accuracy of the sensor. The stability assessment uses a time series analysis method to extract the sensor performance fluctuation characteristics, uses an autoregressive model to predict the performance change trend, evaluates the repeatability of the measurement results through variance analysis, and calculates a stability assessment score.

[0136] A time decay coefficient is determined based on the stability assessment score. The time decay coefficient increases as the stability assessment score decreases. An exponential decay function is used to construct a time weight model. The historical data weight is updated through a sliding time window to establish a dynamic decay mechanism. The time decay coefficient is weightedly fused with the real-time accuracy of the sensor, and a fuzzy comprehensive evaluation method is used to calculate the real-time credibility of the current state of the sensor.

[0137] First, sensor data is collected and preprocessed. The raw data collected by the sensor may contain noise and outliers, requiring preprocessing to improve data quality. Preprocessing steps include data cleaning, outlier removal, and smoothing. For example, median filtering can be used to remove impulse noise, and the Laida criterion can be used to remove data that deviates from the normal range. For example, suppose a temperature sensor collects the data sequence of 25.1°C, 25.3°C, 25.2°C, 50°C, and 25.4°C. Using the Laida criterion, the value 50°C is identified as an outlier and is therefore removed.

[0138] Next, the calibration deviation is calculated based on the verification features. The calibration deviation is obtained by comparing the sensor measurement value with a known standard value. Measurement accuracy is assessed using a variety of metrics, such as the square root of the average of the sum of the squares of the differences between the measured value and the standard value. The ratio of the difference between the measured value and the standard value to the standard value is used to analyze the calibration effect under different ranges, and the fluctuation range of the measurement results is calculated to assess the dispersion of the measurement results. These metrics are combined into a single feature vector, called the calibration deviation feature vector. For example, assuming the standard value is 25°C and the measured value is 25.2°C, the square root of the average of the sum of the squares of the differences is 0.2, resulting in a relative error of 0.8%. Assuming the standard deviation of multiple measurements is 0.1°C, the calibration deviation feature vector is [0.2%, 0.8%, 0.1].

[0139] The corrected deviation feature vector is then input into the sensor's dynamic reputation assessment model. This model quantifies sensor reliability using multiple metrics, including measurement accuracy, response stability, and environmental adaptability. A multi-layer neural network is used to construct a metric fusion module to generate the sensor's real-time accuracy. For example, the corrected deviation feature vector, response time, and temperature change metrics are input into a three-layer neural network. The output value of the output layer node is the sensor's real-time accuracy, assuming the output value is 0.95.

[0140] Next, stability assessment is performed. Time series analysis methods are used to extract sensor performance fluctuation characteristics. For example, a time series model is used to predict sensor performance trends. The repeatability of measurement results is analyzed, for example, by calculating the fluctuations in multiple measurement results to assess repeatability. Ultimately, a stability assessment score is calculated. For example, if the sensor's accuracy over the past period was 0.95, 0.96, 0.94, and 0.95, respectively, and the predicted accuracy for the next moment is 0.95 with a variance of 0.0001, the stability assessment score is high.

[0141] The time decay coefficient is determined based on the stability assessment score. The time decay coefficient increases as the stability assessment score decreases. An exponential decay function is used to construct a time weight model. Historical data weights are updated using a sliding time window, establishing a dynamic decay mechanism. For example, if the stability assessment score is 0.9, the time decay coefficient is 0.1, and the weight of historical data decays exponentially. For example, the weight from one day ago is 0.9, the weight from two days ago is 0.81, and so on.

[0142] Finally, the time decay coefficient and the sensor's real-time accuracy are weighted and fused, and the real-time credibility of the sensor's current state is calculated using a fuzzy comprehensive evaluation method. For example, if the current real-time accuracy of 0.95 is weighted and fused with the time decay coefficient of 0.1, the final calculated real-time credibility of the sensor is 0.945.

[0143] The solution of this application can:

[0144] Improving sensor data credibility: By evaluating sensor reliability in real time, we identify and eliminate unreliable sensor data, improving sensor data quality and credibility, and providing a reliable basis for subsequent data analysis and decision-making. Enhancing system robustness: By dynamically adjusting sensor weights, we reduce the impact of unreliable sensor data on the system, enhance the system's fault tolerance and robustness, and ensure the system can continue to operate normally under abnormal conditions. Optimizing resource allocation: Based on the real-time sensor credibility, we dynamically adjust sensor usage policies. For example, we prioritize the use of sensors with high credibility and reduce the frequency of use of sensors with low credibility, thereby optimizing resource allocation and improving system efficiency.

[0145] Figure 2 This is a structural diagram of a real-time correction and compensation system for marine ranch water quality parameters using a multi-source sensor according to an embodiment of the present invention. Figure 2 As shown, the system includes:

[0146] The first unit is used to build a two-layer edge computing network in the marine ranch, with edge computing nodes deployed in each layer. The edge computing nodes are connected to temperature sensors, salinity sensors, and turbidity sensors based on a microservices architecture to collect raw water quality data. The edge computing nodes use lightweight blockchain technology to distribute and store the raw water quality data, identify outliers through an annealing algorithm, and perform data noise reduction and feature extraction in combination with multi-scale wavelet transform to generate a preprocessed data set. A sensor digital twin model is established to monitor the sensor's status parameters in real time. The sensor's status parameters serve as the basis for subsequent credibility assessment.

[0147] The second unit is configured to receive the preprocessed data set and the state parameters of the sensor, construct a dynamic reputation evaluation model for the sensor using a graph neural network, and calculate real-time reputation based on the state parameters of the sensor; establish a multidimensional sensor association network with vertical hierarchical association, horizontal spatial association, and parameter cross-association based on the real-time reputation, and optimize the weight coefficients of the multidimensional sensor association network using a federated learning method; establish a self-evolving correction parameter matrix by sharing the optimized weight information through incremental learning between edge nodes, and dynamically adjust the data contribution of sensors at each layer to achieve cross-layer collaborative computing;

[0148] The third unit is used to perform weighted fusion of sensor data using a multi-task deep learning model based on the self-evolution correction parameter matrix, and generate an initial correction value in combination with the real-time credibility; construct a spatiotemporal sequence prediction model to calculate the theoretical reference value, and perform error analysis between the initial correction value and the theoretical reference value; according to the error analysis results, use a fuzzy decision tree to adaptively adjust the compensation coefficient in the self-evolution correction parameter matrix to achieve hierarchical water quality parameter correction; input the correction result into the sensor digital twin model for verification, and feed the verification result back to the sensor dynamic credibility evaluation model, dynamically update the real-time credibility and the weight coefficient, and form a closed-loop self-optimizing intelligent correction system.

[0149] According to a third aspect of the embodiments of the present invention,

[0150] An electronic device is provided, comprising:

[0151] processor;

[0152] a memory for storing processor-executable instructions;

[0153] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0154] According to a fourth aspect of the embodiments of the present invention,

[0155] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0156] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0157] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A real-time correction and compensation method for marine ranch water quality parameters using multiple source sensors, characterized in that: include: A two-layer edge computing network, consisting of the surface layer and the water layer, was constructed in the marine ranch, with edge computing nodes deployed in each layer. Based on a microservices architecture, these edge computing nodes connected to temperature sensors, salinity sensors, and turbidity sensors to collect raw water quality data. These edge computing nodes used lightweight blockchain technology to distribute and store the raw water quality data, identified outliers through an annealing algorithm, and combined with multi-scale wavelet transforms to perform data denoising and feature extraction to generate a preprocessed data set. A sensor digital twin model was established to monitor the sensor's status parameters in real time, which served as the basis for subsequent credibility assessments. The method comprises receiving the preprocessed data set and the state parameters of the sensor, constructing a dynamic reputation evaluation model for the sensor using a graph neural network, and calculating a real-time reputation based on the state parameters of the sensor; establishing a multidimensional sensor association network with vertical hierarchical association, horizontal spatial association, and parameter cross-association based on the real-time reputation, and optimizing the weight coefficients of the multidimensional sensor association network using a federated learning method; sharing the optimized weight information through incremental learning between edge nodes, and establishing a self-evolving correction parameter matrix, wherein the self-evolving correction parameter matrix dynamically adjusts the data contribution of sensors at each layer to achieve cross-layer collaborative computing; Based on the self-evolving correction parameter matrix, a multi-task deep learning model is used to perform weighted fusion on the sensor data, and an initial correction value is generated in combination with the real-time credibility; A spatiotemporal sequence prediction model is constructed to calculate the theoretical reference value, and an error analysis is performed between the initial correction value and the theoretical reference value. Based on the error analysis results, a fuzzy decision tree is used to adaptively adjust the compensation coefficient in the self-evolution correction parameter matrix to achieve hierarchical water quality parameter correction. The correction result is input into the sensor digital twin model for verification, and the verification result is fed back to the sensor dynamic credibility evaluation model to dynamically update the real-time credibility and the weight coefficient to form a closed-loop self-optimizing intelligent correction system.

2. The method according to claim 1, characterized in that Receiving the preprocessed data set and the state parameters of the sensor, constructing a dynamic sensor reputation evaluation model using a graph neural network, and calculating real-time reputation based on the state parameters of the sensor; establishing a multidimensional sensor association network with vertical hierarchical association, horizontal spatial association, and parameter cross-association based on the real-time reputation, and optimizing the weight coefficients of the multidimensional sensor association network using a federated learning method, including: Receiving the preprocessed data set and the state parameters of the sensor, forming a feature vector from the preprocessed data set and the state parameters of the sensor; constructing a graph neural network for dynamic sensor reputation evaluation, inputting the feature vector into the graph neural network, the graph neural network performing spatial feature aggregation through a first convolutional layer to obtain sensor spatial correlation features, extracting temporal features through a second convolutional layer to obtain sensor performance change features, performing cross-parameter feature fusion through a third convolutional layer to obtain sensor correlation features, and calculating the real-time reputation of the sensor based on the sensor spatial correlation features, the sensor performance change features, and the sensor correlation features; Establishing a multidimensional sensor association network based on the real-time reputation of the sensors, the multidimensional sensor association network comprising a vertical hierarchical association matrix, a horizontal spatial association matrix, and a parameter cross-correlation matrix, wherein the vertical hierarchical association matrix is constructed based on the vertical distance between sensors, the reputation difference, and the water mass mixing coefficient; the horizontal spatial association matrix is constructed based on the spatial correlation and flow field information of the Kriging interpolation algorithm; and the parameter cross-correlation matrix is constructed based on the physical and chemical relationship of water quality parameters and the influence of environmental factors; A federated learning method is used to optimize the weight coefficients of the multidimensional sensor association network. The federated learning method uses the Adam optimizer to train sub-models at edge nodes and designs an adaptive learning rate adjustment strategy. The model parameters of each node are aggregated through a weighted averaging method to obtain a global model. The weights in the weighted averaging method are determined based on the node data quality and computing power. A differential privacy mechanism is introduced to protect the data. A multi-objective loss function is designed to train the global model. The global model is optimized using a dynamic pruning strategy and knowledge distillation technology, and the optimized weight coefficients of the multidimensional sensor association network are output.

3. The method according to claim 1, characterized in that By sharing optimized weight information through incremental learning among edge nodes, a self-evolving correction parameter matrix is established. The self-evolving correction parameter matrix dynamically adjusts the data contribution of sensors at each layer to achieve cross-layer collaborative computing, including: An edge node incremental learning model is constructed based on a recurrent neural network. The edge node incremental learning model extracts the temporal features of sensor data through a state gate and a forget gate, calculates feature weights using an attention mechanism, and fuses sensor working state information and data credibility to generate a feature vector reflecting sensor performance. A weight sharing mechanism is constructed between edge nodes. The local weight matrix of each edge node is calculated based on the characteristic vector of the sensor performance. A communication topology is established according to the physical distance between the nodes. The local weight matrix is propagated in the communication topology using a distributed consistency protocol. Important weight information is screened using compressed sensing technology. The influence of communication delay is eliminated through an asynchronous compensation mechanism to obtain optimized weight information shared among edge nodes. Based on the shared optimization weight information, a self-evolution correction parameter matrix is established. The self-evolution correction parameter matrix uses a three-dimensional tensor structure to map the relationship between water layers, spatial positions, and parameter types. The shared optimization weight information is used as an initialization parameter. A fitness evaluation function is designed to calculate the importance of parameters. The matrix parameters are iteratively optimized through a genetic algorithm, and the optimization experience stored in the knowledge base is used to guide parameter updates. The self-evolving correction parameter matrix calculates the inter-layer influence according to the hydrological dynamics model, and dynamically adjusts the data contribution of sensors at each layer in combination with the correction coefficients of environmental factors such as the thermocline. Based on the data contribution, the sensor sampling strategy is determined, and a sensor collaborative sampling scheme is established. A multi-objective optimization method is used to balance sampling accuracy and resource consumption, and cross-layer collaborative computing is performed.

4. The method according to claim 3, characterized in that The self-evolving correction parameter matrix calculates the inter-layer influence based on the hydrodynamic model and combines the correction coefficients of environmental factors such as the thermocline to dynamically adjust the data contribution of sensors at each layer. Based on the data contribution, the sensor sampling strategy is determined, and a sensor collaborative sampling scheme is established. A multi-objective optimization method is used to balance sampling accuracy and resource consumption. The cross-layer collaborative calculation is performed, including: Constructing a hydrodynamic model, wherein the self-evolution correction parameter matrix uses the hydrodynamic model to describe the layered structure of the water body, uses density gradient to characterize interlayer exchange characteristics, calculates material migration flux through the vertical turbulent diffusion equation, and generates an initial interlayer influence parameter set; A thermocline characteristic model is constructed based on the initial interlayer influence parameter set. The thermocline characteristic model identifies the location and strength of the thermocline through temperature profile data, analyzes the influence of the thermocline on sensor performance, calculates a measurement error compensation coefficient based on salinity and pressure changes, and outputs an environmental factor correction coefficient matrix. Performing parameter correction on the environmental factor correction coefficient matrix and the initial inter-layer influence parameter set, dynamically calculating the data contribution of sensors at each layer based on the correction results using the self-evolving correction parameter matrix, evaluating data value using the information entropy criterion, determining data redundancy through mutual information analysis, and generating a real-time data contribution vector based on the data contribution, the data value, and the data redundancy; constructing a sensor sampling strategy based on the real-time data contribution vector, allocating sampling resources based on the real-time data contribution vector, adjusting the measurement frequency using an adaptive sampling mechanism, optimizing data acquisition efficiency through triggered sampling control, and outputting an initial sampling scheme matrix; Inputting the initial sampling scheme matrix into a collaborative sampling optimizer, the collaborative sampling optimizer establishes a sensor collaborative sampling scheme, adopts a hierarchical Markov decision process to coordinate sensor working states, optimizes the inter-layer collaboration mechanism based on a game theory method, and generates a collaborative sampling optimization framework; Based on the collaborative sampling optimization framework, multi-objective optimization is performed, the optimization objective function of sampling accuracy and resource consumption is constructed, energy constraints and data quality thresholds are set, the particle swarm algorithm is used to solve the optimal sampling parameters, the optimized sampling instruction set is output, and the optimized sampling instruction set is executed to complete cross-level collaborative computing.

5. The method according to claim 1, wherein Based on the self-evolving correction parameter matrix, a multi-task deep learning model is used to perform weighted fusion on the sensor data, and an initial correction value is generated in combination with the real-time credibility; Constructing a spatiotemporal sequence prediction model to calculate a theoretical reference value, and performing error analysis between the initial correction value and the theoretical reference value includes: receiving a sensor feature representation vector, and constructing a multi-layer data fusion network based on weight coefficients in the self-evolving correction parameter matrix, wherein the multi-layer data fusion network includes a feature mapping layer and a weight calculation layer, wherein the feature mapping layer projects the sensor feature representation vector into a unified feature space, and the weight calculation layer generates an adaptive fusion weight in combination with the real-time reputation; An adaptive normalization layer is used to balance the distribution of different sensor data. An attention mechanism is built to calculate the correlation scores between features. The fusion weights are dynamically adjusted based on the correlation scores. A multi-layer perceptron network is used to perform nonlinear transformation on the weighted features. The original feature information is retained through residual connections to generate initial correction values. Constructing a spatiotemporal sequence prediction model, wherein the spatiotemporal sequence prediction model establishes a spatial association graph of sensor nodes based on the sensor feature representation vector, uses a graph convolutional network to extract topological features between nodes, aggregates local features using a diffuse convolution operator, and retains spatial information in combination with an attention pooling layer; The spatiotemporal sequence prediction model uses a decoder and an encoder to process the sensor historical data sequence. The encoder captures long-term dependencies through a multi-head self-attention mechanism. The decoder fuses historical information and spatial features based on a cross-attention mechanism, predicts parameter change trends, outputs theoretical reference values, and compares and analyzes the initial correction value with the theoretical reference value.

6. The method according to claim 1, characterized in that Based on the error analysis results, a fuzzy decision tree is used to adaptively adjust the compensation coefficients in the self-evolving correction parameter matrix to achieve hierarchical water quality parameter correction. The correction results are input into the sensor digital twin model for verification, and the verification results are fed back to the sensor dynamic reputation evaluation model to dynamically update the real-time reputation and the weight coefficient, forming a closed-loop self-optimizing intelligent correction system including: An error analysis network is constructed to process sensor data. The error analysis network uses wavelet transform to decompose the frequency components of the error sequence, combines variational inference to model the error distribution characteristics, and uses Gaussian process regression to predict the error propagation trend, generating an error feature vector containing error amplitude, variance, and trend characteristics. A fuzzy decision tree is constructed based on the error feature vector. The fuzzy decision tree uses a Gaussian membership function to map the error feature vector to a fuzzy language space. A mapping relationship between the error feature vector and a compensation coefficient is established through an adaptive neuro-fuzzy inference system. A multi-level decision structure is constructed using an information gain criterion to output hierarchical correction rules. The self-evolving correction parameter matrix is adjusted according to the hierarchical correction rule, the self-evolving correction parameter matrix includes initial compensation coefficients of water quality parameters, the compensation coefficients are updated based on the reasoning results of the fuzzy decision tree, a recursive feature elimination algorithm is used to optimize the feature selection strategy, and an adaptive boosting algorithm is used to enhance the compensation effect to obtain corrected water quality parameters; Inputting the corrected water quality parameters into a sensor digital twin model, the sensor digital twin model integrates a physical model and a deep learning network, simulates the working mechanism of the sensor based on the physical model, uses the deep learning network to fit the dynamic response characteristics, compares the correction results with the predicted values through a state predictor, and generates verification features including correction deviation and stability assessment; updating the sensor dynamic reputation assessment model based on the verification feature, wherein the sensor dynamic reputation assessment model uses multi-dimensional indicators to quantify sensor reliability, calculates real-time accuracy based on the correction deviation in the verification feature, determines a time decay coefficient based on stability assessment, and outputs a real-time reputation reflecting the current state of the sensor; The real-time credibility is fed back into the self-evolving correction parameter matrix to dynamically adjust the weight coefficients of different sensor nodes. The weight coefficients are positively correlated with the real-time credibility. The parameter update speed is controlled by an adaptive learning rate. A multi-objective optimization algorithm is used to balance correction accuracy and system stability to achieve closed-loop optimization of the correction system.

7. The method according to claim 6, characterized in that The sensor dynamic reputation evaluation model is updated based on the verification feature. The sensor dynamic reputation evaluation model uses multi-dimensional indicators to quantify sensor reliability, calculates real-time accuracy based on the correction deviation in the verification feature, determines the time decay coefficient based on the stability evaluation, and outputs a real-time reputation reflecting the current state of the sensor, including: Calculate the correction deviation based on the verification feature, the correction deviation is obtained by comparing the correction result with the standard value, use the root mean square error to evaluate the measurement accuracy, combine the relative error index to analyze the correction effect under different ranges, use the standard deviation to calculate the dispersion of the measurement result, and output the correction deviation feature vector; Inputting the correction deviation feature vector into the sensor dynamic reputation evaluation model, the sensor dynamic reputation evaluation model uses multi-dimensional indicators to quantify sensor reliability, the multi-dimensional indicators including measurement accuracy indicator, response stability indicator and environmental adaptability indicator, and constructing an indicator fusion module through a multi-layer neural network to generate the real-time accuracy of the sensor; A stability assessment is performed based on the real-time accuracy of the sensor. The stability assessment uses a time series analysis method to extract the sensor performance fluctuation characteristics, uses an autoregressive model to predict the performance change trend, evaluates the repeatability of the measurement results through variance analysis, and calculates a stability assessment score. A time decay coefficient is determined based on the stability assessment score. The time decay coefficient increases as the stability assessment score decreases. An exponential decay function is used to construct a time weight model. The historical data weight is updated through a sliding time window to establish a dynamic decay mechanism. The time decay coefficient is weightedly fused with the real-time accuracy of the sensor, and a fuzzy comprehensive evaluation method is used to calculate the real-time credibility of the current state of the sensor.

8. A multi-source sensor real-time correction and compensation system for marine ranching water quality parameters, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to build a two-layer edge computing network in the marine ranch, with edge computing nodes deployed in each layer. The edge computing nodes are connected to temperature sensors, salinity sensors, and turbidity sensors based on a microservices architecture to collect raw water quality data. The edge computing nodes use lightweight blockchain technology to distribute and store the raw water quality data, identify outliers through an annealing algorithm, and perform data noise reduction and feature extraction in combination with multi-scale wavelet transform to generate a preprocessed data set. A sensor digital twin model is established to monitor the sensor's status parameters in real time. The sensor's status parameters serve as the basis for subsequent credibility assessment. The second unit is configured to receive the preprocessed data set and the state parameters of the sensor, construct a dynamic reputation evaluation model for the sensor using a graph neural network, and calculate real-time reputation based on the state parameters of the sensor; establish a multidimensional sensor association network with vertical hierarchical association, horizontal spatial association, and parameter cross-association based on the real-time reputation, and optimize the weight coefficients of the multidimensional sensor association network using a federated learning method; establish a self-evolving correction parameter matrix by sharing the optimized weight information through incremental learning between edge nodes, and dynamically adjust the data contribution of sensors at each layer to achieve cross-layer collaborative computing; A third unit is configured to perform weighted fusion of sensor data using a multi-task deep learning model based on the self-evolving correction parameter matrix, and generate an initial correction value in combination with the real-time credibility; A spatiotemporal sequence prediction model is constructed to calculate the theoretical reference value, and an error analysis is performed between the initial correction value and the theoretical reference value. Based on the error analysis results, a fuzzy decision tree is used to adaptively adjust the compensation coefficient in the self-evolution correction parameter matrix to achieve hierarchical water quality parameter correction. The correction result is input into the sensor digital twin model for verification, and the verification result is fed back to the sensor dynamic credibility evaluation model to dynamically update the real-time credibility and the weight coefficient to form a closed-loop self-optimizing intelligent correction system.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images