A water quality environmental parameter monitoring and early warning method, system and device for deep-sea aquaculture net cages

Abstract

The present application belongs to the technical field of intelligent monitoring and early warning of seawater aquaculture environment, and relates to a water quality environmental parameter monitoring and early warning method, system and device for deep-sea aquaculture net cages. The system comprises: an ecological element three-dimensional monitoring module for real-time monitoring and obtaining water quality parameter data of different water layers; a multi-water quality parameter fusion analysis and prediction module M 3 Net for outputting water quality prediction results; a multi-water quality parameter early warning module for alarming when water quality parameters reach a threshold; a dissolved oxygen prediction review module LA-STGCN-FCI for dissolved oxygen prediction and review of water quality prediction results; and a dissolved oxygen early warning module for alarming to a management end system when results are abnormal. The present application is suitable for a multi-water layer, multi-parameter aquaculture water quality monitoring system and a monitoring and early warning method, can realize real-time and accurate collection of water quality parameters of deep-sea aquaculture net cages, coupled analysis of water quality parameters and accurate early warning, thereby effectively ensuring the safety of aquaculture and reducing economic losses.

Description

Technical Field

[0001] This invention belongs to the field of intelligent monitoring and early warning technology for marine aquaculture environment, and relates to intelligent monitoring of aquaculture water quality parameters, specifically to a method, system and device for monitoring and early warning of water quality environmental parameters for deep-sea aquaculture cages. Background Technology

[0002] With the advancement of aquaculture ecological governance and the development of nearshore tourism, my country's nearshore aquaculture space continues to shrink, making deep-sea cage aquaculture an important development direction for the transformation and upgrading of marine aquaculture. Deep-sea cage aquaculture has advantages such as vast aquaculture space and excellent water exchange conditions, effectively alleviating the environmental pressure brought by nearshore aquaculture. However, extreme weather occurs frequently in my country's deep-sea aquaculture areas, and their open, complex, and highly dynamic marine environment can affect the safety of aquaculture cages and aquatic organisms. Therefore, higher requirements are placed on water quality monitoring and intelligent management technologies. Water quality parameters such as temperature, dissolved oxygen, and pH are important indicators affecting the growth of aquaculture organisms. Once water quality indicators exceed the ecologically suitable range for aquaculture organisms, it will cause stress, disease, and even death, resulting in huge economic losses for aquaculture units. How to achieve precise water quality control in a highly turbulent environment is one of the key scientific and engineering problems currently facing deep-sea cage aquaculture.

[0003] Existing deep-sea aquaculture cage culture water quality monitoring and early warning technologies have three main shortcomings: First, most existing technologies require separate model building and training for different water quality parameters, which is cumbersome and costly to deploy. Second, different water quality parameters are correlated and have spatial distribution differences, and existing models and technologies do not fully consider the correlation and differences between different water quality parameters. Finally, existing deep-sea cage culture water quality parameter prediction technologies lack accuracy and rely on specific thresholds for early warning, leading to frequent false alarms and misleading alerts.

[0004] Therefore, developing a high-performance water quality parameter monitoring module suitable for deep-sea aquaculture cages and establishing a scientific and accurate water quality early warning method has become an urgent technical challenge. This is of great significance for ensuring the biosafety and facility safety of deep-sea cage aquaculture and promoting the healthy and sustainable development of deep-sea cage aquaculture. Summary of the Invention

[0005] The purpose of this invention is to address the problems of existing deep-sea aquaculture cage water quality monitoring and early warning technologies, such as the lack of multi-parameter coupling analysis, weak dynamic early warning capabilities, and insufficient accuracy. This invention proposes a water quality environmental parameter monitoring and early warning method, system, and device for deep-sea aquaculture cages. It is applicable to multi-layer, multi-parameter aquaculture water quality monitoring systems and methods, enabling real-time and accurate acquisition of water quality parameters in deep-sea aquaculture cages, coupled analysis of water quality parameters, and accurate early warning of water quality parameters exceeding limits.

[0006] The technical solution of this invention is: This invention provides a water quality environmental parameter monitoring and early warning system for deep-sea aquaculture cages, comprising: The three-dimensional monitoring module for ecological elements is used to monitor and acquire water quality parameter data of different water layers in real time; The multi-water quality parameter fusion analysis and prediction module provides an adaptive enhanced spatiotemporal graph neural network M. 3 Net is used to extract time-series features across different time spans and correlation features among multiple water quality parameters, outputting water quality prediction results. First, a deep learning model is trained to obtain optimal model weights. Then, these model weights are loaded and used to calculate real-time water quality parameter predictions based on real-time acquired water quality parameter data.

[0007] The multi-water quality parameter early warning module is used to issue an alarm when the predicted water quality parameter reaches the threshold.

[0008] The multi-water quality parameter early warning module adopts an adaptive threshold setting method. Based on empirical thresholds, it can dynamically output the upper and lower limits of various water quality parameters by combining historical stable operation data of aquaculture cages, time period characteristics, stratification characteristics, environmental disturbance characteristics, and prediction model output results.

[0009] The dissolved oxygen prediction verification module uses a hierarchical sensing spatiotemporal graph neural network (LA-STGCN-FCI) for dissolved oxygen prediction and uses a backup model to verify the water quality prediction results.

[0010] The dissolved oxygen early warning module is used to send an alarm to the management system when the verification results are abnormal.

[0011] Furthermore, the M 3 The Net model includes: The backbone convolutional layer is used for feature expansion and extraction; A multi-scale time-series module is used for the extraction and aggregation of time-series features of water quality data. This module employs a multi-branch dilated convolution structure to extract multi-scale features from the time series data, followed by addition and concatenation. The module contains four branches: The first and second branches respectively use dilated convolution to extract temporal features; The third branch uses max pooling to aggregate features within local time windows and further stabilizes the feature distribution through a batch normalization layer. The fourth branch constructs residual connections, which map the input features through 1×1 convolution and batch normalization and then directly pass them to the output. The graph convolution module is used to capture spatial features of the data.

[0012] Furthermore, the graph convolution module includes a graph convolution solution mechanism and an adaptive enhancement mechanism; The graph convolution decomposition mechanism decomposes the graph convolution operation into two parts: static channel attention and dynamic feature mapping, achieving low-dimensional compression and high-dimensional reconstruction of graph features. Its expression is as follows:

[0013] Where H is the channel attention matrix, C0 is the basic weight matrix, Q is the dimension reduction mapping matrix, Φ is the attention coefficient matrix, P is the dimension increase mapping matrix, and g(L)x is the graph feature operation performed on the original input x. The adaptive enhancement mechanism introduces W E Dynamic enhancement of graph structures can be represented by graph convolution computation as follows:

[0014] Among them, W E It is a trainable adaptive enhancement matrix.

[0015] Furthermore, the LA-STGCN-FCI module includes: The temporal convolution module, located at both ends of the feature extraction layer, is used to enhance the learning of temporal information. The FFT amplitude gating module is used to enhance its ability to sense periodic fluctuations in water quality parameters. The multi-scale receptive field feature aggregation module is placed after the first temporal convolution module to enhance the ability to capture temporal features of water quality at different scales. The graph convolution module is used to capture spatial features of the data.

[0016] Furthermore, the temporal convolution module takes the feature tensor X as input and performs three-way convolution calculations on the input features to obtain convolution features Conv1(X), Conv2(X), and Conv3(X); The second convolution result, after Sigmoid activation, is fused with the results from the other two paths and then subjected to ReLU nonlinear mapping to obtain the temporal convolution output H. time :

[0017] The FFT amplitude gating module first performs mean-centering on the input features, then performs a real-valued Fast Fourier Transform along the time dimension, subsequently extracts the amplitude from the spectrum, and obtains the amplitude features through logarithmic compression.

[0018] Among them, X f Represents the frequency domain; Based on this, the frequency domain amplitude characteristics are averaged and pooled to obtain the frequency domain statistics M. pool The gating weights G are generated by 1×1 convolution and sigmoid activation; these G are then convolved with the temporal convolution to output H. time Element-wise residual modulation is performed to obtain the module output H. out The calculation formula is: .

[0019] Furthermore, the three-dimensional monitoring module for ecological elements includes a water quality sensor group, a TCP communication interface and a standard data interface connected to the water quality sensor group for data transmission and communication; the water quality sensor group includes sensors arranged in layers within the aquaculture cage at different water layers to be measured, and the sensors collect water quality parameter data of different water layers in real time and transmit them to the data storage unit.

[0020] Furthermore, it also includes an integrated control board, which encapsulates a parallel multi-water quality parameter fusion analysis and prediction module and a dissolved oxygen prediction verification module, and is externally connected to an ecological element three-dimensional monitoring module through a standard data interface, and connected to an early warning system through an early warning module.

[0021] This invention also provides a method for monitoring and early warning of water quality environmental parameters for deep-sea aquaculture cages. The method is based on the monitoring and early warning system described above and includes the following steps: (1) Acquire time-series data of water quality sensor nodes; (2) By converting the time series channel and splitting the data, the collected water quality sensor node time series data is converted into a standard prediction format to achieve data standardization; (3) Input the dataset into the pre-trained multi-water quality parameter fusion analysis and prediction module and dissolved oxygen prediction verification module, and use the edge computing unit to make real-time predictions by using the prediction algorithm with the optimal weight; (4) Compare the prediction results with the warning threshold, input the warning / silence command into the warning module, and drive the warning module to issue an alarm.

[0022] Furthermore, in step (4), the predicted value is compared with the corresponding threshold to determine whether there is a risk of exceeding the limit:

[0023] Where represents the predicted value at the next k time steps, T max T represents the upper limit threshold. min E represents the lower threshold, and E=1 indicates that the prediction result has the risk of exceeding the threshold. When a predicted value is detected to exceed a threshold, the warning level is classified according to the degree of exceedance and the duration: Level 1 Warning: Slight exceedance, please pay attention; Level 2 warning: Moderate exceedance, intervention measures recommended; Level 3 warning: Severe over-limit, immediately alarm and trigger control terminal.

[0024] This invention also provides a water quality environmental parameter monitoring and early warning device for deep-sea aquaculture cages. The device encapsulates a multi-water quality parameter fusion analysis and prediction module, a multi-water quality parameter early warning module, a dissolved oxygen prediction and verification module, and a dissolved oxygen early warning module, as well as a data storage unit and an edge computing unit. It is externally connected to an ecological element three-dimensional monitoring module through a standard data interface and externally connected to a system early warning terminal through an early warning module.

[0025] The beneficial effects of this invention are: (1) This invention provides a multi-layer, multi-parameter aquaculture water quality monitoring and multi-step prediction system suitable for deep-sea aquaculture cages. It utilizes an integrated control board to drive a control algorithm to control dissolved oxygen, temperature, pH, COD, and chlorophyll a sensors to acquire deep-sea aquaculture water quality parameters in real time, and uses a current meter to collect the flow direction and velocity of the aquaculture water. After acquisition, the collected water quality environmental parameters are analyzed using an edge high-performance computing unit, and based on M... 3 The -Net model predicts water quality parameters for multiple water layers, and then the LA-STGCN-FCI network is used to accurately predict dissolved oxygen. When the predicted water quality parameter value exceeds the warning threshold, the control panel will trigger an alarm.

[0026] (2) The system of the present invention acquires aquaculture water quality parameters in real time through the onboard sensors and uses a deep learning model to provide real-time early warning of aquaculture water quality parameters at the edge. Using the system of the present invention, the water quality environment in deep-sea aquaculture cages can be predicted and warned, ensuring aquaculture safety and reducing economic losses.

[0027] (3) This invention proposes a water quality parameter monitoring module and early warning method for deep-sea aquaculture cages, which has excellent prediction accuracy for complex multi-parameter, multi-water-layer, and long-term prediction scenarios; and can effectively ensure the safety of water quality environment in deep-sea aquaculture scenarios. Attached Figure Description

[0028] Figure 1 The overall flowchart of the water quality parameter monitoring and early warning system provided by the present invention; Figure 2 M, the water quality parameter monitoring and early warning system provided by the present invention 3 A schematic diagram of the Net multi-water quality parameter, multi-water layer, multi-step prediction network; Figure 3 M, the water quality parameter monitoring and early warning system provided by the present invention 3A schematic diagram of the multi-scale timing module structure in Net; Figure 4 M, the water quality parameter monitoring and early warning system provided by the present invention 3 A schematic diagram of adaptive augmentation and decomposition embedding graph convolutional structures in Net; Figure 5 A schematic diagram of the LA-STGCN-FCI multi-scale receptive field fusion spatiotemporal graph convolutional network structure for the water quality parameter monitoring and early warning system provided by the present invention; Figure 6 A schematic diagram of the temporal convolution and frequency domain amplitude gating structure in the LA-STGCN-FCI water quality parameter monitoring and early warning system provided by this invention; Figure 7 A schematic diagram of the multi-scale receptive field fusion structure in the LA-STGCN-FCI water quality parameter monitoring and early warning system provided by the present invention; Figure 8 A schematic diagram of the multi-decoder output structure in the LA-STGCN-FCI water quality parameter monitoring and early warning system provided by the present invention; Figure 9 A schematic diagram of the multi-water quality parameter early warning module of the water quality parameter monitoring and early warning system provided by the present invention. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] To further understand the present invention, it will be further described in conjunction with the accompanying drawings and embodiments.

[0031] Example 1 like Figure 1 As shown, this invention provides a water quality environmental parameter monitoring and early warning system for deep-sea aquaculture cages. It is a multi-feature fusion-based water quality parameter monitoring and early warning system, comprising: The system includes a three-dimensional monitoring module for ecological elements, a multi-water quality parameter fusion analysis and prediction module, a multi-water quality parameter early warning module, a dissolved oxygen prediction and verification module, a dissolved oxygen early warning module, early warning equipment, and an integrated control board.

[0032] The integrated control board encapsulates a multi-water quality parameter fusion analysis and prediction module, a multi-water quality parameter early warning module, a dissolved oxygen prediction verification module, and a dissolved oxygen early warning module, as well as a data storage unit and an edge computing unit, and connects to water quality sensors and early warning devices through a standard data interface.

[0033] The three-dimensional ecological element monitoring module includes a water quality sensor array (including sensors for dissolved oxygen, temperature, pH, COD, and chlorophyll a), a TCP communication interface, a power supply module, and a standard data interface. The water quality sensor array is used to acquire real-time hydrological information within the aquaculture cages, including dissolved oxygen, temperature, pH, COD, chlorophyll a, and ocean current velocity. The standard data interface and TCP communication interface connect to the water quality sensor array for data transmission and communication, and can acquire water quality parameter data for different water layers through layered deployment.

[0034] Secure the water quality sensor to the aquaculture cage frame or other rigid structure, ensuring the integrated sensor probe is completely submerged in water. If data collection is required across multiple water layers, ensure the probe is fixed to the desired water layer to obtain key water quality parameters from various water layers within the aquaculture water body. Standard data and TCP communication interfaces are located on the integrated control board, ensuring connection to the water quality sensor array for data transmission and communication.

[0035] In the actual process of water quality parameter acquisition and prediction, the water quality sensor group collects key water quality parameters in real time and transmits them to the data storage unit through a standard data interface.

[0036] like Figure 2 As shown, the multi-water quality parameter fusion analysis and prediction module, namely the adaptive enhanced spatiotemporal graph neural network M, 3 Net consists of stacked backbone convolutional layers, multi-scale temporal modules, and graph convolutional modules. 3 In Net, the backbone convolutional layer is used for feature expansion and extraction; the multi-scale temporal module is used for the extraction and aggregation of temporal features of water quality data; and the graph convolutional module is used to capture spatial features of the data.

[0037] The multi-water quality parameter fusion analysis and prediction module receives time-series water quality parameter data from various sensors in the data storage unit, performs missing value imputation, standardization, and data reconstruction on the water quality parameter features, and scientifically divides the dataset, as follows: First, the time-series water quality parameter data in the data storage unit are concatenated into a three-dimensional tensor: (1) Where N is the feature dimension, T is the time step, and F is the number of water quality parameters; Then, outliers were detected using the 3σ principle, and missing values ​​were filled using the mean of the previous 10 time steps via a sliding window averaging method. (2) Then, a sliding window is used to generate the dataset. First, a rolling time window strategy is applied to the original continuous time series data to generate samples. Specifically, the length L of the historical input window is set. inWith prediction window length L out Extract a length of L from the beginning of the time series. in The historical sequence is used as the model input, and the subsequent continuous L... out The data from each time step is used as the prediction target. Then, a window slides forward along the time axis with a fixed step size, continuously generating new input-output sample pairs. Through this rolling Zurier generation strategy, the original long-term series can be transformed into a large number of supervised learning samples, while maintaining the continuity and temporal dependencies of the time series, thereby improving the model training efficiency and prediction accuracy.

[0038] To ensure the rationality of data partitioning during the training of time-series prediction models and to avoid future information leakage, this invention adopts a rolling time-series cross-validation strategy to construct the model training set, validation set, and test set.

[0039] First, the original time series data is sorted chronologically to form a continuous time series dataset of length T. Then, the entire time series is divided into multiple continuous time data blocks along the time axis, with each block maintaining its original chronological order.

[0040] During model training, multiple rounds of data partitioning are performed by progressively expanding the training data range. Specifically, in the k-th experiment, the first n parts of the initial time series are selected. k The training set is constructed by using a percentage of the data as the training data, the subsequent continuous time data as the validation set, and then selecting data from later time periods as the test set. In the next round of experiments, the time range of the training set is further expanded while maintaining the temporal order of the validation and test sets.

[0041] In each round of experiments, the prediction model's parameters are trained using the training set, its performance is evaluated and parameters are tuned using the validation set, and finally, the model's predictive performance is evaluated on the test set. After completing multiple rounds of experiments, the test results from each round are statistically analyzed and averaged to obtain the final performance evaluation result of the model.

[0042] By adopting a rolling time-series cross-validation strategy, we can make full use of limited samples while ensuring the temporal causal relationship of time series data, improve the stability of model training and evaluation results, and effectively avoid the adverse effects of future information leakage on the performance evaluation of prediction models.

[0043] like Figure 3 As shown, M in this invention 3The Net multi-scale temporal series module employs a multi-branch dilated convolutional structure to extract multi-scale features from time series data, capturing temporal dependencies across different time spans. This module processes input features in parallel using dilated convolutional kernels with varying dilation rates, and then concatenates and fuses the outputs of each branch to achieve comprehensive modeling of multi-scale temporal information.

[0044] Specifically, the module includes four parallel branch structures: First, the input features are fed into four branches. Each branch performs a 1×1 convolution to linearly map the input features, adjusting the feature channel dimension and reducing the computational cost of subsequent convolutions. Then, in each branch, batch normalization and ReLU activation functions are sequentially applied for feature normalization and non-linear mapping. To ensure consistent feature dimensions across different scales during subsequent fusion, each branch performs padding before the convolution operation, ensuring a consistent temporal dimension in the output feature map after convolution.

[0045] Based on this, different branches use different structures for multi-scale feature extraction: the first branch and the second branch respectively use dilated convolution for temporal feature extraction, where the kernel size is 3 and the dilation rate is set to 4 and 6 respectively to expand the receptive field of the convolution, thereby capturing the temporal dependencies under different time spans.

[0046] The third branch uses max pooling to aggregate features within the local time window and further stabilizes the feature distribution through a batch normalization layer.

[0047] The fourth branch constructs residual connections to improve training efficiency and model performance. After mapping the input features through 1×1 convolution and batch normalization, the features are directly passed to the output to preserve the original temporal feature information and alleviate the gradient vanishing problem during the training of deep networks.

[0048] Finally, the features output from each branch are concatenated along the channel dimension and integrated through feature fusion operations to obtain a fused feature representation containing multi-scale temporal information, which is then used as input to subsequent network modules.

[0049] like Figure 4 As shown, M in this invention 3The graph convolution module of Net consists of gated units and adaptive enhancement and decomposition embedding (AE&DE) graph convolution units (including graph embedding, graph convolution decomposition mechanism, and adaptive enhancement mechanism) stacked sequentially. The gated unit contains two parallel dilated convolution kernels, which are connected to the sigmoid and tanh activation functions respectively, and are used to output the retention ratio of different feature information and the nonlinear features after compression activation.

[0050] In the graph convolutional embedding process, the input graph structure is first represented using graph embedding. Assume a directed graph is represented as... Where V is the set of nodes, E is the set of edges in the graph, and A is the adjacency matrix. A ij It is a node i and nodes j The edge weights between nodes are given, where N is the number of nodes in the graph. By introducing trainable embedding matrices A1 and A2, the original adjacency relationships are embedded and mapped, constructing a new graph structure (the graph's adjacency matrix) as follows: (3) in, and As a trainable matrix, this method enables the embedding of graph structure information from source nodes to target nodes, thereby enhancing the expressive power of potential relationships between nodes. The ReLU function is used to ensure the non-negativity of the correlation matrices, while the Softmax function can regularize and sparsify the weight matrix.

[0051] Subsequently, graph convolutional feature learning is performed based on the embedded graph structure. Let the input node features be... x Then the standard graph convolution operation can be expressed as: (4) in, g (·) represents the graph convolution filter function, and L is the Laplacian matrix calculated from the adjacency matrix. W For learnable weight parameters, , E It is the projection dimension. The degree matrix is ​​defined as follows: ,in .therefore, .

[0052] To enhance feature extraction capabilities, this invention further introduces a graph convolution decomposition mechanism. By decomposing the graph convolution operation into two parts—static channel attention and dynamic feature mapping—low-dimensional compression and high-dimensional reconstruction of graph features are achieved. Its expression is as follows: (5) Where H is the channel attention matrix. This is used to implement static feature weight learning, where C0 is the basic weight matrix and Q is the dimensionality reduction mapping matrix. Used to compress graph features, compressing graph features into a low-dimensional space M. This is the attention coefficient matrix. g(L)x represents the graph feature operation performed on the original input x. P is the up-dimensional mapping matrix. This method is used to recover the output feature dimension, expanding the feature dimension to the output dimension. This decomposition method can reduce computational complexity while maintaining the expressive power of graph convolution.

[0053] Building upon this, the present invention further proposes an adaptive enhancement mechanism, which introduces an adaptive weight matrix W. E Dynamic enhancement of graph structures can be represented by graph convolution computation as follows: (6) Among them, W E This is a trainable adaptive weight matrix used to dynamically adjust the relationships between nodes in the graph structure. Through this matrix, the embedded graph structure can be adaptively enhanced to prevent over-aggregation of node information and imbalance problems caused by the sparse graph structure during optimization. This mechanism can adaptively strengthen the connections between important nodes during model training while suppressing information propagation from irrelevant nodes, thereby avoiding the over-aggregation problem caused by the sparse graph structure.

[0054] During the model training phase, the embedded matrices A1, A2, and the adaptive weight matrix W are used. E All updates are performed using the backpropagation algorithm, thereby achieving dynamic optimization learning of the graph structure. Through the aforementioned adaptive enhancement and decomposition of the embedded graph convolutional structure, the ability of graph neural networks to model complex spatial correlations in multi-parameter temporal data can be effectively improved, and the feature representation ability and prediction accuracy in multivariate prediction tasks can be enhanced.

[0055] To improve the efficiency of feature aggregation between neighboring nodes, the graph convolution module uses the K nearest neighbor algorithm to select and aggregate the top K most relevant node features. Then, the values ​​of these top K neighboring nodes are subtracted from the node mean to capture the correlation information between nodes. (7) Among them, F k This represents the K-nearest neighbor algorithm, where the number of nearest neighbors is k. After feature aggregation, a 1×1 convolution is used for feature extraction, and the feature with the largest activation value is selected as the input to the next layer of the network.

[0056] Furthermore, to address the imbalance between different water quality parameters in different water layers, the model uses a composite loss function that includes weighted MSE, focal weighted MSE, and Cauchy loss function: (8) Where a1, a2, and a3 are the three coefficients of the loss function, L wmse It is the weighted MSE loss function, L fwmse It is the focus-weighted MSE loss function, L cauchy It is the Cauchy loss function: (9) Where w is the weight vector of multiple water quality parameters, It is the mean-bias loss function; (10) Where β and γ are hyperparameters, the tanh function smooths the loss and maps the loss value to the range of (-1, 1), and the soft thresholding method balances hard and easy samples to reduce the impact of extreme values ​​on imbalanced data. (11) s is the scaling factor. A larger value of this factor can reduce the impact of outliers.

[0057] like Figure 5 As shown, the dissolved oxygen prediction verification module provided by this invention, namely the hierarchical perceptual spatiotemporal graph neural network LA-STGCN-FCI, includes a temporal convolution module, an FFT amplitude gating module, a multi-scale receptive field feature aggregation module, and a graph convolution module.

[0058] like Figure 6 As shown, the present invention provides the structure of the temporal convolution module and amplitude gating module in the LA-STGCN-FCI network. This module takes the feature tensor X as input and consists of two parts: a temporal convolution branch and a frequency domain amplitude gating branch. The temporal convolution branch performs three-way convolution calculations on the input features to obtain convolutional features Conv1(X), Conv2(X), and Conv3(X). The result of the second convolution is activated by Sigmoid and then fused with the results of the other two convolutions, and finally subjected to ReLU nonlinear mapping to obtain the temporal convolution output H. time : (12) Temporal convolutional branches are deployed at the beginning and end of the network feature extraction structure, serving as core components for capturing temporal features: the first temporal convolutional module fuses temporal feature information through temporal convolution operations, enhancing the learning of temporal patterns; the second temporal convolutional module performs temporal convolution on the topological features output by the graph convolutional neural network, capturing the dynamic dependencies between different time steps. The three convolutional outputs can be written as: (13) Further, the time-domain features are obtained as follows: (14) Where * represents two-dimensional convolution, σ represents Sigmoid, the kernel size is (1, k), and convolution is performed only along the time dimension.

[0059] The frequency domain amplitude-gated branch first performs mean-neutralization on the input features to obtain... X c Then, perform a real fast Fourier transform (rFFT) along the time dimension to obtain the frequency domain representation. X f Subsequently, the amplitude value of the spectrum was obtained. And the amplitude characteristics are obtained through logarithmic compression: (15) The FFT amplitude gating module enhances the perception of periodic fluctuations in water quality parameters by building upon the temporal convolution module. The module first performs mean-removal processing on the time series data. (16) Then, a real-valued Fast Fourier Transform is performed along the time dimension to map the time-domain signal to the frequency domain: (17) Frequency domain amplitude is defined as: (18) The logarithmic compression of the amplitude is performed to suppress the problem of excessive dynamic range. Then, average pooling is performed on the frequency dimension to obtain the frequency domain summary. (19) at this time: (20) Based on this, 1×1 convolutions are used to generate gated weights: (twenty one) Where f gate The output dimension is: (twenty two) The frequency domain amplitude features are then averaged using pooling to obtain the frequency domain statistics. M pool Gated weights are generated by 1×1 convolution and Sigmoid activation. G Final frequency domain gating weights G Convolutional output with time H time Element-wise residual modulation is performed to obtain the module output. H out The formula for its calculation is: (twenty three) This module introduces frequency domain amplitude statistics during the temporal convolution process, enabling the network to adaptively focus on key frequency band features in different samples, thereby enhancing its ability to model periodic fluctuations and temporal dynamic patterns and improving the accuracy of dissolved oxygen prediction.

[0060] like Figure 7 As shown, the LA-STGCN-FCI network provided by this invention includes a Multi-Receptive Field Aggregation (MRFA) module. The MRFA module receives the temporal feature tensor information output from the first temporal convolutional block, divides the resulting feature tensor channels into four equal parts, performs convolution using different kernels, and controls its fusion with the vector from the previous layer using Gaussian gating units. (twenty four) Where F local Local features are extracted using 3×3 depthwise convolution, F context The receptive field of the convolution kernel is gradually expanded by 1×7, 1×9, and 1×11, and context modulation is achieved through a gating mechanism.

[0061] The module first performs channel normalization on the input features: (25) (26) After normalization, the channel is divided into four equal blocks:

[0062] Each sub-block contains a C / 4 channel, and then the receptive field is gradually expanded through three stages: In the first stage, local contextual temporal features are captured using a 1×7 convolutional kernel: (27) DWConv stands for depthwise convolution.

[0063] Simultaneously, the Gaussian gating unit generates gating weights and performs gating modulation: (28) (29) Simultaneously, features are captured using 3×3 depthwise convolutions, and concatenated with the gated and modulated large-kernel convolution result as the first layer output: (30) (31) That is, the output of the first layer is: (32) Similar to the first layer, the second layer expands the large kernel convolution into a 1×9 convolution, that is: (33) (34) (35) The block division result is: (36) (37) Then its output is: (38) Similarly, the third layer further enlarges the convolution kernel to a 1×11 convolution, that is: (39) Then its output is: (40) The graph convolution module uses the Laplacian matrix to model spatial relationships, given a graph adjacency matrix of... Through Laplace normalization: (41) Where D is the degree matrix, I is the identity matrix, and I is the standardized Laplacian matrix. Used for graph convolution operations: (42) Subsequently, the parameter matrix Perform feature mapping: (43) Where C s For spatial feature dimensions.

[0064] Furthermore, the multi-decoder output module improves the model's predictive ability for long-term features through block decoding and fusion. This is achieved by dividing the features output from the last temporal convolution layer into blocks according to the node dimension: (44) Each block interacts with other blocks through a cross-attention mechanism to enhance feature complementarity: (45) Where (i, j, k) is a cyclic combination of block indices, d k is the scaling factor for the feature dimension. Finally, each decoder outputs its predictions through a fully connected layer, and these predictions are concatenated to form the final output: (46) Among them, T out To predict the time step.

[0065] In practical applications, model weights are pre-trained, and sensor data is captured in real time and fed into a neural network to predict future dissolved oxygen changes.

[0066] like Figure 8 As shown, the structure of the Multi-Directional Joint Prediction (MDJP) module in the LA-STGCN-FCI network provided by this invention is illustrated. This module is used to model the vertical coupling relationship between different water layers in a marine aquaculture environment. The model first divides the node features into three layers based on water depth: (47) Among them, H U H represents the characteristics of the upper water layer. M H represents the characteristics of the middle layer of water. L It indicates the characteristics of the lower water layer.

[0067] To achieve cross-layer information interaction, taking the prediction of the upper layer as an example, the interaction score is first calculated using features from the middle and lower layers: (48) The interaction scores were then normalized using Softmax to obtain the gating weights. (49) The gating weights are used to modulate the target layer features: (50) Finally, the enhanced features are concatenated with the original features: (51) The same interaction process applies to both the middle and lower layer features: (52) (53) This cross-layer interaction mechanism can explicitly model the vertical coupling relationship between different water layers, thereby improving the accuracy of dissolved oxygen prediction.

[0068] like Figure 9 As shown, the multi-water quality parameter early warning module and dissolved oxygen early warning module in the system of this invention include a command encoder, a downstream control circuit, and an alarm device. The downstream control circuit includes a command decoder and a PWM signal generator. Its workflow includes threshold management, threshold determination, early warning classification, alarm decision, terminal alarm, and terminal response steps, used to send an alarm signal to the control terminal in advance when it is predicted that water quality parameters are about to exceed the safe range. After receiving the early warning command, the early warning module converts it into an RS485 digital signal and transmits it to the downstream control circuit. The downstream control circuit parses the status code and generates a control command, which is then transmitted to the PWM signal generator, outputting a pulse waveform to trigger the early warning module alarm.

[0069] To achieve the early warning function, the system establishes corresponding early warning thresholds for different water quality parameters, including: Static threshold: A fixed threshold is set based on industry standards, regulatory requirements, or experience range; Dynamic threshold: The threshold is adaptively adjusted according to the season, water layer, aquaculture species, or historical statistical characteristics; Hysteresis buffer zone: A buffer zone is set near the threshold to avoid frequent alarms when parameters fluctuate repeatedly near the boundary.

[0070] Furthermore, to address the threshold differences caused by different seasons, water layers, and aquaculture species, this invention designs an adaptive threshold calculation method. Based on historical stable operating data, and combined with current time-period characteristics, stratification characteristics, environmental disturbance characteristics, and prediction model outputs, it dynamically generates upper or lower warning thresholds for various water quality parameters, and achieves stable and reliable early warning through hysteresis and risk grading mechanisms.

[0071] First, based on the species being farmed and their growth stage, and according to farming experience, a safe range is set for each water quality indicator:

[0072] Based on this, grouped baselines are established according to the performance of aquaculture water quality at different times, seasons, and water layers. For example, under condition combination c, the historical reference mean and fluctuation scale of the i-th parameter are denoted as:

[0073] Where 'c' can be considered a combination of season, time period, and water layer.

[0074] For each water quality parameter, under the current condition c, a dynamic threshold is calculated based on historical data. Taking dissolved oxygen as an example:

[0075] Where k i This is the sensitivity coefficient; the larger the coefficient, the more sensitive the warning.

[0076] Based on this, using the sampling frequency as a reference, the predicted value is compared with the corresponding threshold to determine whether there is a risk of exceeding the limit: (54) in, This represents the predicted value at the next k time steps. T max and T min These represent the upper and lower thresholds, respectively, with E=1 indicating that the prediction result has a risk of exceeding the threshold.

[0077] When a predicted value is detected to exceed a threshold, the warning level is classified according to the degree of exceedance and the duration: Level 1 Warning: Mild exceedance, requiring attention; Level 2 Warning: Moderate exceedance, intervention measures recommended; Level 3 Warning: Severe exceedance, immediate alarm and activation of control terminals. Warning Level Judgment: (55) Where Δ represents the magnitude by which the predicted value exceeds the threshold, τ represents the predicted duration of the exceedance, and L represents the warning level.

[0078] To improve system reliability and reduce false alarm rate, this invention introduces an alarm decision and deduplication mechanism, including: only when the predicted value exceeds the limit for multiple consecutive time steps is it considered a valid warning. (56) in m K represents the number of prediction steps for continuous observation; K represents the minimum number of over-limit attempts required to trigger an alarm.

[0079] Confirmation is achieved by combining predictions from multiple future steps to avoid false alarms caused by single-step anomalies. A "hysteresis buffer" is set up so that when the water quality parameter recovers from the warning value to a safe value, the alarm is only lifted after it has left the threshold by a certain margin, thus avoiding boundary oscillations.

[0080] If the same parameter has already issued an alarm of the same level within a short period of time, the same alarm will not be sent repeatedly within a preset suppression time window, thereby preventing terminal alarm overload.

[0081] When the alarm triggering conditions are met, the system sends an alarm message to the control terminal. The alarm methods include, but are not limited to: audible and visual alarms, platform pop-ups, and remote monitoring interface prompts; the alarm content includes parameter name, predicted value, threshold, warning level, occurrence time, and suggested handling measures.

[0082] After receiving an alarm message, the control terminal can execute one or more of the following response methods: manually confirm the alarm; execute control actions in conjunction with the alarm (start the aeration equipment, adjust the water exchange or circulation device, prompt to stop feeding); and send the terminal processing results back to the monitoring platform to form a closed-loop feedback.

[0083] Example 2 This invention proposes a method for monitoring and early warning of water quality parameters in deep-sea aquaculture cages. It exhibits excellent prediction accuracy for complex application scenarios involving multiple parameters, multiple water layers, and long prediction time intervals, effectively ensuring the safety of the water environment in deep-sea aquaculture. The system of this invention can predict and warn of water quality conditions in deep-sea aquaculture cages, ensuring aquaculture safety and reducing economic losses.

[0084] The monitoring and early warning method provided by this invention is applied to the aforementioned water quality environmental parameter monitoring and early warning system for deep-sea aquaculture cages, and includes the following steps: Acquire time-series data from water quality sensor nodes; By converting time-series channel data and segmenting data, the collected time-series data of water quality sensor nodes are transformed into a standard prediction format, thereby achieving data standardization. The above dataset is input into a pre-trained multi-water quality parameter fusion analysis and prediction module and a dissolved oxygen prediction verification module. Real-time prediction is performed using an edge computing unit through a prediction algorithm with optimal weights. The prediction results are compared with the warning threshold, and the warning / silence command is input into the warning module to drive the warning module to issue an alarm.

[0085] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, alterations, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A water quality environmental parameter monitoring and early warning system for deep-sea aquaculture cages, characterized in that, include: The three-dimensional monitoring module for ecological elements is used to monitor and acquire water quality parameter data of different water layers in real time; The multi-water quality parameter fusion analysis and prediction module provides an adaptive enhanced spatiotemporal graph neural network M. 3 Net is used to extract time-series features across different time spans and correlation features among multiple water quality parameters, and output water quality prediction results. The multi-water quality parameter early warning module is used to issue an alarm when the predicted water quality parameter reaches the threshold. The dissolved oxygen prediction and verification module is a hierarchical sensing spatiotemporal graph neural network LA-STGCN-FCI, used for dissolved oxygen prediction and verification of water quality prediction results; The dissolved oxygen early warning module sends an alarm to the management system when the verification results are abnormal.

2. The water quality environmental parameter monitoring and early warning system for deep-sea aquaculture cages according to claim 1, characterized in that, The M 3 The Net model includes: The backbone convolutional layer is used for feature expansion and extraction; A multi-scale time-series module is used for the extraction and aggregation of time-series features of water quality data. This module employs a multi-branch dilated convolution structure to extract multi-scale features from the time series data, followed by addition and concatenation. The module contains four branches: The first and second branches respectively use dilated convolution to extract temporal features; The third branch uses max pooling to aggregate features within local time windows and further stabilizes the feature distribution through a batch normalization layer. The fourth branch constructs residual connections, which directly pass the input features to the output after mapping them through 1×1 convolution and batch normalization. The graph convolution module is used to capture spatial features of the data.

3. The water quality environmental parameter monitoring and early warning system for deep-sea aquaculture cages according to claim 2, characterized in that, The graph convolution module includes a graph convolution solution mechanism and an adaptive enhancement mechanism; The graph convolution decomposition mechanism decomposes the graph convolution operation into two parts: static channel attention and dynamic feature mapping, achieving low-dimensional compression and high-dimensional reconstruction of graph features. Its expression is as follows: ; Where H is the channel attention matrix, C0 is the basic weight matrix, Q is the dimension reduction mapping matrix, Φ is the attention coefficient matrix, P is the dimension increase mapping matrix, and g(L)x is the graph feature operation performed on the original input x; The adaptive enhancement mechanism introduces W E Dynamic enhancement of graph structures can be represented by graph convolution computation as follows: ; Among them, W E It is a trainable adaptive enhancement matrix.

4. The water quality environmental parameter monitoring and early warning system for deep-sea aquaculture cages according to claim 1, characterized in that, The LA-STGCN-FCI network includes: The temporal convolution module is used to enhance the learning of temporal information and capture the dynamic dependencies between different time steps; The FFT amplitude gating module is used to enhance its ability to sense periodic fluctuations in water quality parameters. A multi-scale receptive field feature aggregation module is used to enhance the ability to capture water quality temporal features at different scales; The graph convolution module is used to capture spatial features of the data.

5. The water quality environmental parameter monitoring and early warning system for deep-sea aquaculture cages according to claim 4, characterized in that, The temporal convolution module takes the feature tensor X as input and performs three-way convolution calculations on the input features to obtain convolution features Conv1(X), Conv2(X), and Conv3(X). The second convolution result, after Sigmoid activation, is fused with the results from the other two paths and then subjected to ReLU nonlinear mapping to obtain the temporal convolution output H. time : ; The FFT amplitude gating module first performs mean-neutralization on the input features, then performs a real-valued Fast Fourier Transform along the time dimension, subsequently extracts the amplitude from the spectrum, and obtains the amplitude features through logarithmic compression. ; Among them, X f Represents the frequency domain; Then, average pooling is performed on the frequency domain amplitude characteristics to obtain the frequency domain statistics M. pool The gating weights G are generated by 1×1 convolution and Sigmoid activation; Convolve it with time to output H time Element-wise residual modulation is performed to obtain the module output H. out The calculation formula is: 。 6. The water quality environmental parameter monitoring and early warning system for deep-sea aquaculture cages according to claim 1, characterized in that, The three-dimensional monitoring module for ecological elements includes a water quality sensor group, a TCP communication interface for data transmission and communication connected to the water quality sensor group, and a standard data interface. The water quality sensor group includes sensors arranged in layers within the aquaculture cages at different water layers to be measured. The sensors collect water quality parameter data of different water layers in real time and transmit them to the data storage unit.

7. The water quality environmental parameter monitoring and early warning system for deep-sea aquaculture cages according to claim 1, characterized in that, It also includes an integrated control board, which encapsulates a parallel multi-water quality parameter fusion analysis and prediction module and a dissolved oxygen prediction verification module, and is connected to an ecological element three-dimensional monitoring module through a standard data interface, and connected to an early warning system through an early warning module.

8. A method for monitoring and early warning of water quality environmental parameters for deep-sea aquaculture cages, the method being based on the monitoring and early warning system described in any one of claims 1-7, characterized in that, Includes the following steps: (1) Acquire time-series data of water quality sensor nodes; (2) By converting the time series channel and splitting the data, the collected water quality sensor node time series data is converted into a standard prediction format to achieve data standardization; (3) Input the dataset into the pre-trained multi-water quality parameter fusion analysis and prediction module and dissolved oxygen prediction verification module, and use the edge computing unit to make real-time predictions by using the prediction algorithm with the optimal weight; (4) Compare the prediction results with the warning threshold, input the warning / silence command into the warning module, and drive the warning module to issue an alarm.

9. The method for monitoring and early warning of water quality environmental parameters for deep-sea aquaculture cages according to claim 8, characterized in that, In step (4), the predicted value is compared with the corresponding threshold to determine whether there is a risk of exceeding the limit: ； in, T represents the predicted value at the next k time steps. max T represents the upper limit threshold. min E represents the lower threshold, and E=1 indicates that the prediction result has the risk of exceeding the threshold. When a predicted value is detected to exceed a threshold, the warning level is classified according to the degree of exceedance and the duration: Level 1 Warning: Slight exceedance, please pay attention; Level 2 warning: Moderate exceedance, intervention measures recommended; Level 3 warning: Severe over-limit, immediately alarm and trigger control terminal.

10. A water quality environmental parameter monitoring and early warning device for deep-sea aquaculture cages, characterized in that, The device encapsulates a multi-water quality parameter fusion analysis and prediction module, a multi-water quality parameter early warning module, a dissolved oxygen prediction and verification module, and a dissolved oxygen early warning module, as well as a data storage unit and an edge computing unit. It is also connected to an ecological element three-dimensional monitoring module through a standard data interface and to a system early warning terminal through an early warning module.

Images