An aquaculture multivariate water quality parameter prediction method and system

Abstract

This invention belongs to the field of water quality parameter prediction technology, specifically relating to a method and system for predicting multivariate water quality parameters in aquaculture. The method includes: acquiring multivariate water quality monitoring time series; performing adaptive characterization distillation processing on the multivariate water quality monitoring time series to obtain a historical-future fused water quality series; based on the historical-future fused water quality series, employing an attention mechanism that integrates a dynamic Gaussian kernel and dual-path routing, and outputting a weighted fused attention result through a two-stage routing selection strategy; based on the weighted fused attention result, using a cross-granularity sequence interactive perceptron to mine the symmetrical coupling relationship between different time granularities and variables to obtain the final output features; and mapping the final output features to predicted values ​​in the target dimension through fully connected layers or convolutional layers.

Description

Technical Field

[0001] This invention belongs to the field of water quality parameter prediction technology, specifically relating to a method and system for predicting multivariate water quality parameters in aquaculture. Background Technology

[0002] Aquaculture is one of the most dynamic sectors in the global food system. Aquaculture products are not only an important source of protein for human consumption, solving food safety issues, but also play a crucial role in the economy and the environment. The quality of farmed aquatic products is inextricably linked to the aquatic environment. Due to the dynamic changes in aquaculture volume and feed intake, in addition to untimely water purification, excess feed and improper aquatic plant management often affect water quality. Therefore, efficient, real-time, and accurate long-term forecasting of aquaculture water quality is crucial for reducing aquaculture costs and improving fish yield and quality.

[0003] In existing technologies, the prediction of multivariate water quality parameters in aquaculture mainly relies on two types of methods: physical mechanism-based modeling methods and data-driven learning methods. The former characterizes physical processes by establishing system dynamic equations, offering a degree of interpretability. However, when faced with complex nonlinearities, multi-source disturbances, and high-dimensional data, it is prone to problems such as equation distortion, difficulty in parameter identification, and excessive computational complexity, leading to insufficient model generalization performance. The latter, such as traditional machine learning methods (Support Vector Machines, Random Forests, Hidden Markov Models, etc.), while capable of learning patterns from data, relies on manual feature engineering and has limited ability to characterize changes in data distribution and long-term dependencies, making it difficult to adapt to the dynamic changes of multi-scale time-series features.

[0004] In recent years, deep learning methods (such as LSTM, GRU, TCN, etc.) have made some progress in water quality parameter prediction, and can automatically extract features through hierarchical structures. However, these methods still have significant shortcomings when dealing with long or non-stationary sequences: recurrent structure models are prone to gradient vanishing, and the fixed receptive field of convolutional models limits the capture of global dependencies; at the same time, they have poor interpretability, high hyperparameter sensitivity, and are difficult to apply stably in complex multivariate systems.

[0005] The emergence of the Transformer model provides a new approach to water quality parameter prediction based on a global self-attention mechanism. It can capture dependencies between arbitrary time steps and effectively model interactions between multiple variables. However, the standard Transformer suffers from computational complexity that increases quadratically with sequence length, and exhibits problems such as insufficient feature extraction, high information redundancy, and decreased model generalization performance when dealing with multi-scale dynamic changes, seasonal drift, and cross-variable coupling relationships. While existing improved methods (such as Informer, Autoformer, and FEDformer) have achieved some optimization in computational efficiency and feature extraction, a series of technical bottlenecks still need to be addressed.

[0006] (1) Redundancy interference problem in periodic feature extraction: For sequences with strong seasonality or multiple periodicity, the attention mechanism of existing models has difficulty accurately distinguishing key signals from background noise or redundant information when capturing periodic patterns. These redundant features will cause interference in the model representation, not only diluting the weight of key information,

[0007] (2) Insufficient adaptability to dynamic changes in data distribution: Water quality parameter data often experience statistical distribution drift (concept drift). The weight patterns learned by traditional attention mechanisms during the training phase are relatively fixed. Once the distribution characteristics of the test data change, the model cannot adaptively adjust its attention to key time steps, resulting in a decrease in its feature capture ability, significant fluctuations in prediction performance, and poor robustness.

[0008] (3) Weakness in modeling cross-scale and cross-variable coupling relationships: Existing methods often treat short-term fluctuations and long-term trends in isolation, or fail to fully model the complex interaction effects of different variables at different time scales. This lack of ability to jointly model "cross-scale and cross-variable" dynamics makes it difficult for the model to fully grasp the evolution of the system, thus limiting the further improvement of prediction accuracy. Summary of the Invention

[0009] To overcome the problems in the existing technology, this invention proposes a method and system for predicting multivariate water quality parameters in aquaculture.

[0010] The technical solution of the present invention to solve the above-mentioned technical problems is as follows:

[0011] In a first aspect, the present invention provides a method for predicting multivariate water quality parameters in aquaculture, comprising the following steps:

[0012] Step 100: Obtain multivariate water quality monitoring time series;

[0013] Step 200: Perform adaptive characterization distillation on the multivariate water quality monitoring time series to obtain a history-future fused water quality series;

[0014] Step 300: Based on the historical-future fused water quality sequence, an attention mechanism that combines dynamic Gaussian kernel and dual-path routing is adopted. Through a two-stage routing selection strategy, the weighted fused attention result is output.

[0015] Step 400: Based on the attention results after weighted fusion, a cross-granularity sequence interaction perceptron is used to explore the symmetric coupling relationship between variables at different time granularities to obtain the final output features;

[0016] Step 500: Map the final output features to the predicted values ​​of the target dimension through fully connected layers or convolutional layers.

[0017] Furthermore, the multivariate water quality monitoring time series includes conductivity, dissolved oxygen, chlorophyll a, turbidity, temperature, and pH.

[0018] Further, step 200 includes:

[0019] Discrete wavelet decomposition is performed on the input multivariate water quality monitoring time series, which is decomposed into approximation coefficients and detail coefficients;

[0020] Filtering is performed based on approximation coefficients and detail coefficients, and information bottleneck theory is used to retain key information and filter irrelevant features to obtain the processed approximation coefficients and detail coefficients.

[0021] Based on the processed approximation coefficients and detail coefficients, wavelet reconstruction is used to reconstruct the time domain and obtain the estimated value of the reconstructed time series.

[0022] The reconstructed time series estimates are divided into multiple subsequences of length, and the number of placeholders is calculated to initialize the future placeholder sequence.

[0023] Multiple subsequences are concatenated with placeholder sequences to obtain a historical-future integrated water quality sequence.

[0024] Further, step 300 includes:

[0025] Based on the historical-future integrated water quality sequence, an adaptive Gaussian kernel is constructed using a learnable standard deviation parameter. The output of the Gaussian kernel is subjected to causal mask offset, and then multiplied with the value vector of the location correlation branch to generate a location bias signal.

[0026] Based on the historical-future integrated water quality sequence, a linear mapping is performed to calculate the query vector, key vector, and value vector of the content relevance branch, and then the attention score is calculated.

[0027] The location bias signal and attention score are weighted and fused through a routing mechanism to obtain the fused attention score.

[0028] Based on the fused attention score, the output is a weighted fused attention result through the collaborative work of the shared header and the routing header.

[0029] Furthermore, based on the fused attention score, through the collaborative work of the shared header and the routing header, the output is a weighted fused attention result, including:

[0030] The routing score is calculated by adding a score to each individual header and a score associated with the header type. Defined as:

[0031] ;

[0032] In the above formula, Indicates the number of shared headers; and These represent the projection matrices of the shared header and the routing header, respectively. Indicates the first in the input Input markers;

[0033] Among them, the balance coefficient and The contribution of the shared header and the routing header is balanced and defined as follows:

[0034] ;

[0035] In the above formula, It is a trainable projection matrix; yes The hidden size;

[0036] The weighted fusion attention result is calculated based on the routing score and the attention result after weighted fusion.

[0037] Further, step 400 includes:

[0038] Local context features of different granularities are extracted by temporal multi-scale convolution, and the feature expression is enhanced by cross-scale splicing and channel recombination to obtain the recombined features;

[0039] Based on the recombined features, a dual-path pooling mechanism is introduced, which generates channel-adaptive scaling factors and offsets through dimensionality reduction, nonlinear activation, and dimensionality increase operations, respectively.

[0040] Based on channel-adaptive scale factors and offsets, a dynamic gating mechanism is used to adaptively fuse local convolutional features and global statistical features.

[0041] By modeling the symmetric coupling relationship between local and global features through bidirectional outer product operations, and combining residual connections to preserve the original temporal structure, the final output features are obtained.

[0042] Furthermore, based on local features, global statistical information, scale fusion gating, and bias fusion gating, a symmetric coupling relationship between local and global features is modeled through bidirectional outer product operations. This, combined with residual connections to preserve the original temporal structure, yields the final output features. ,include:

[0043] ;

[0044] ;

[0045] In the above formula, This represents the result of a two-way outer product operation; Indicates the first i Individual channel scale integrated gating; Indicates the first i Local features of each channel; Indicates the first j Offset fusion gating for each channel; Indicates the first j Global information for each channel; This is the interaction weight matrix for adaptive learning; This represents the attention result after weighted fusion; d Indicates the number of channel dimensions.

[0046] Secondly, a multivariate water quality parameter prediction system for aquaculture is provided. The system includes a data acquisition and input module, an adaptive characterization distillation module, a dual-path routing attention mechanism module, a cross-granularity sequence interaction multilayer perceptron module, and an output module.

[0047] The input module is used to acquire and input multivariate water quality monitoring time series;

[0048] The adaptive characterization distillation module is used to process the input multivariate water quality monitoring time series to obtain a history-future fused water quality series;

[0049] The dual-path routing attention mechanism module is used to output the weighted fusion attention result based on the historical-future fused water quality sequence, using a fusion dynamic Gaussian kernel and dual-path routing attention mechanism. Through a two-stage routing selection strategy, it outputs the weighted fusion attention result.

[0050] The cross-granularity sequence interaction multilayer perceptron module is used to mine the symmetrical coupling relationship between variables at different time granularities based on the attention results after weighted fusion, and obtain the final output features.

[0051] The output module is used to map the final output features to predicted values ​​in the target dimension through fully connected or convolutional layers.

[0052] Compared with the prior art, the present invention has the following technical effects:

[0053] (1) Existing multivariate water quality parameter prediction models for aquaculture require a large number of parameters in multi-scale dynamic modeling, and are prone to redundant information interference and noise amplification when processing features at different time scales, resulting in high model complexity, unstable training, and poor generalization performance. This invention proposes an adaptive representation distillation method based on the information bottleneck theory. First, wavelet decomposition is used to divide the original time series into different frequency components to extract multi-scale features; then, an information bottleneck mechanism is introduced in the feature layer to perform redundant compression and key information screening of the feature channels; then, the residual connection structure is used to ensure the continuous transmission and fusion of effective information; finally, wavelet inverse transform is used to recover the time domain signal to achieve high-fidelity prediction feature reconstruction. This method can effectively suppress redundant and noise information while maintaining the main dynamic features, reduce the number of model parameters and computational complexity; significantly improve the adaptability and stability of the model to dynamics at different time scales, and achieve high-precision, low-complexity multi-scale time series modeling.

[0054] (2) Traditional multi-head attention mechanisms in aquaculture multivariate water quality parameter prediction generally suffer from sensitivity to data distribution changes and difficulty in adaptively adjusting attention patterns, leading to performance degradation and insufficient feature extraction in environments with distribution drift or non-stationarity. This invention proposes an attention mechanism that integrates dynamic Gaussian kernels and dual-path routing. This mechanism utilizes dynamic Gaussian kernels to perform neighborhood adaptive modeling of the input sequence, enhancing the ability to perceive distribution changes over different time periods; it employs a global and local dual-path attention structure to simultaneously capture long-range dependencies and local changes; and through a dynamic selection mechanism of shared heads and routing heads, it achieves adaptive feature selection and fusion for multiple distribution patterns. When faced with time-distribution drift, non-stationarity, and heterogeneous data, this mechanism can dynamically adjust the attention weight distribution, effectively enhancing the model's ability to capture key features and its robustness, thereby improving the model's prediction accuracy and generalization performance in complex environments.

[0055] (3) Existing multivariate water quality parameter prediction models for aquaculture are insufficient in terms of cross-scale correlation and cross-variable dependency modeling, making it difficult to simultaneously capture the dynamic coupling relationship between short-term fluctuations and long-term trends, resulting in a decrease in prediction accuracy and structural interpretability. This invention proposes a cross-granularity sequence-aware MLP structure. This structure introduces multi-scale convolutional branches in the time domain to extract local contextual features, and achieves multi-granularity information fusion through feature concatenation and channel recombination; further, it adopts a dual-path pooling mechanism to aggregate global statistical features, and combines a bottleneck MLP with parameter decoupling to explicitly model high-order coupling between variables; finally, it achieves adaptive fusion of local and global features through dynamic gating and bidirectional outer product operations. This method can comprehensively characterize the dynamic dependencies between different variables and different time scales, and achieve collaborative modeling of short-term changes and long-term trends; while improving prediction accuracy, it effectively enhances the structural expressiveness and feature interpretability of the model. Attached Figure Description

[0056] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0057] Figure 1 This is a schematic diagram of the overall system structure of the present invention;

[0058] Figure 2 This is a schematic diagram of the adaptive characterization distillation structure of the present invention;

[0059] Figure 3 This is the overall framework of the dual-path routing attention mechanism of the present invention;

[0060] Figure 4 This invention describes the structure and calculation process of a single attention head in the dual-path routing attention mechanism.

[0061] Figure 5 This is a curve showing the fitted conductivity.

[0062] Figure 6 This is a scatter plot of conductivity regression.

[0063] Figure 7 This is a temperature fitting curve;

[0064] Figure 8 This is a scatter plot of temperature regression.

[0065] Figure 9 This is a curve fitting diagram for dissolved oxygen.

[0066] Figure 10This is a scatter plot of dissolved oxygen regression.

[0067] Figure 11 This is a pH fitting curve;

[0068] Figure 12 This is a scatter plot of pH regression.

[0069] Figure 13 A fitted curve for chlorophyll a;

[0070] Figure 14 A scatter plot of chlorophyll a regression;

[0071] Figure 15 This is a turbidity fitting curve.

[0072] Figure 16 This is a turbidity regression scatter plot. Detailed Implementation

[0073] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the specific implementation methods, structures, features, and effects of the technical solutions proposed according to the present invention are described in detail below with reference to the accompanying drawings and preferred embodiments. Specific features, structures, or characteristics in one or more embodiments may be combined in any suitable form. Unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0074] In one embodiment, reference is made to Figures 1-16 This paper provides a method for predicting multivariate water quality parameters in aquaculture, including the following steps:

[0075] Step 100: Obtain multivariate water quality monitoring time series;

[0076] Step 200: Perform adaptive characterization distillation on the multivariate water quality monitoring time series to obtain a history-future fused water quality series;

[0077] Step 300: Based on the historical-future fused water quality sequence, an attention mechanism that combines dynamic Gaussian kernel and dual-path routing is adopted. Through a two-stage routing selection strategy, the weighted fused attention result is output.

[0078] Step 400: Based on the attention results after weighted fusion, a cross-granularity sequence interaction perceptron is used to explore the symmetric coupling relationship between variables at different time granularities to obtain the final output features;

[0079] Step 500: Map the final output features to the predicted values ​​of the target dimension through fully connected layers or convolutional layers.

[0080] The following is a detailed explanation of each of the above steps:

[0081] Step 100: Obtain and preprocess the multivariate water quality monitoring time series, which includes water quality monitoring indicators, namely conductivity, dissolved oxygen, chlorophyll a, turbidity, temperature, and pH.

[0082] The preprocessing includes filling missing values, detecting and correcting outliers, suppressing noise, and performing necessary normalization on the raw water quality monitoring data to eliminate the impact of sensor errors, environmental interference, and data acquisition discontinuity on subsequent modeling.

[0083] The time series of multivariate water quality monitoring data for aquaculture water bodies includes water quality indicators such as conductivity, dissolved oxygen (DO), chlorophyll a (Chl-a), turbidity, temperature, and pH. Each water quality monitoring indicator serves as a variable dimension in the input sequence; no pre-set manual weights are required, and their importance is adaptively learned by the model during training.

[0084] Step 200: Process the input multivariate water quality monitoring time series to retain key information while suppressing redundant features.

[0085] Multivariate water quality monitoring time series generally exhibit significant multi-scale characteristics. Existing methods typically rely on additional parameterization to capture these multi-scale features, which increases the computational burden on the models. Furthermore, due to significant differences between different datasets, models often heavily depend on cumbersome hyperparameter tuning in practical applications, thus limiting their generalization ability in diverse scenarios. Particularly when dealing with data exhibiting strong seasonal patterns, many models struggle to effectively distinguish key signals from redundant features. These redundant features not only interfere with the model's learning process and increase prediction bias but may also mask potential true patterns in the data, leading to decreased reliability of prediction results.

[0086] Based on this, this embodiment proposes an adaptive characterization distillation based on the information bottleneck theory. After the multivariate water quality parameter series enters the aquaculture multivariate water quality parameter prediction model, it first passes through the adaptive characterization distillation module based on the information bottleneck theory. This module aims to extract the most critical and representative information from the multivariate water quality monitoring time series, while removing redundant and irrelevant information as much as possible.

[0087] like Figure 2 As shown, wavelet decomposition is used to decompose the time series into components of different frequencies to identify patterns and trends at different time resolutions, thereby comprehensively capturing feature information. Then, the information bottleneck theory is used to filter out redundant features that may be contained in the multi-scale information, and residual connections are used to ensure that valuable information is fully preserved during the filtering process. Wavelet reconstruction is used to reconstruct the processed data back to the time domain. Specifically, the following steps are included:

[0088] Step 210: Perform discrete wavelet decomposition on the input multivariate water quality monitoring time series, decomposing it into approximation coefficients and detail coefficients.

[0089] For the input multivariate water quality monitoring time series matrix ,in, Indicates the length of the backtracking window; The number of variables representing water quality parameters includes water quality indicators such as dissolved oxygen (DO), chlorophyll a (Chl-a), water temperature, turbidity, salinity, and pH. The observation values ​​of each variable over time constitute the column vector of the matrix.

[0090] Wavelet transform decomposes multivariate water quality monitoring time series into approximate coefficients. and detail coefficient Among them, the approximation coefficient The component representing the overall trend or smoothing of a multivariate water quality monitoring time series reflects low-frequency information; while the detail coefficients... This captures fine-grained high-frequency information, highlighting local fluctuations and details in the sequence. Discrete wavelet decomposition can be specifically represented as:

[0091] ;

[0092] ;

[0093] in, Indicates the original water quality parameters over time. The possible values ​​of ; This is a scaling function, corresponding to a low-pass filter, used to extract the low-frequency trend components of the signal; This is a wavelet function, corresponding to a high-pass filter, used to extract high-frequency local variation features of a signal; and These are the filter coefficients when expanding the scaling function and wavelet function, respectively; and These represent the number of coefficients for the low-pass and high-pass filters, respectively; This is the filter coefficient index, used to control the translation position of the function; and Let represent the scaling and translation forms of the scaling function and wavelet function, respectively, where "2t" represents the scaling change. k " indicates translation; integral symbol" This indicates that a convolution operation is performed on the time axis to extract signal components at the corresponding scale.

[0094] Through the wavelet decomposition described above, the model can obtain the trend features and local detail features of the signal at different scales, thereby achieving fine modeling of complex time series dynamics.

[0095] Step 220: Filter based on approximation coefficients and detail coefficients, and use the information bottleneck theory to retain key information and filter irrelevant features to obtain the processed approximation coefficients and detail coefficients.

[0096] After completing wavelet decomposition, approximate coefficients are obtained. and detail coefficient Then, the two are further processed through filtering.

[0097] Assumption ( () represents the original coefficients, i.e., the approximate coefficients obtained from wavelet decomposition. and detail coefficient ; This represents the filtered result. At this point, and Treated as random variables, their joint distribution is .

[0098] Mutual information between two random variables Quantified and The amount of information shared between them. The mutual information between joint distributions is defined as follows:

[0099] ;

[0100] in, It is the KL divergence, used to quantify joint distributions. Product with marginal distribution The differences between them; express Value and Value The joint probability, It is a set of possible values; express The marginal probability distribution; express The marginal probability distribution; express The entropy represents its uncertainty or information content; Indicates a given hour The conditional entropy represents the conditional entropy under given conditions. The remaining uncertainty.

[0101] Define random variables in the intermediate hidden layer It follows the Markov chain principle. And mutual information satisfies .

[0102] Although it is passed to the intermediate layer Information loss is inevitable during the process, but Predictions were retained. The most critical information is obtained while irrelevant features are filtered out. Therefore, it is necessary to maximize... and Mutual information between them, while minimizing and Mutual information between them:

[0103] ;

[0104] in, It is a trade-off parameter that balances information retention with redundancy filtering:

[0105] maximize : Ensure Z retains the prediction Key information required, such as water quality trends or anomalies;

[0106] minimize : Inhibit Z-interaction Irrelevant redundant information, such as noise or minor fluctuations.

[0107] Through this optimization, the intermediate layer Z achieves adaptive compression and purification of the original coefficients: preserving the prediction target. The core information is filtered out, and the interference features are filtered out, thereby improving the model's efficiency and robustness in modeling complex water quality dynamics.

[0108] Step 230: Based on the processed approximation coefficients and detail coefficients, reconstruct the time domain using wavelets to obtain the estimated value of the time series.

[0109] After performing discrete wavelet decomposition and filtering on the multivariate water quality time series, the processed approximation coefficients and detail coefficients are recombined using inverse wavelet transform (IWT) to reconstruct the estimated values ​​of the multivariate water quality monitoring time series. :

[0110] ;

[0111] In the above formula, This represents the approximation coefficients after filtering. This represents the detail factor after filtering.

[0112] Step 240: Estimated values ​​of the reconstructed time series Divided into n subsequences .

[0113] The number of subsequences, n, is determined by the subsequence window length W and the step size S, i.e. , This represents the floor function.

[0114] Step 250: Calculate the number of placeholders and initialize the future placeholder sequence.

[0115] For the future data to be predicted, initialize learnable placeholders to represent future subsequences of length w. Then, for a prediction range of H time steps, m placeholders are needed. Indicating the future. In this context, .

[0116] Step 260: Concatenate the subsequence with the placeholder sequence to obtain the historical-future integrated water quality sequence.

[0117] The n subsequences obtained in step 240 and the m placeholders initialized in step 250 By piecing them together sequentially, we get:

[0118] ;

[0119] In the above formula, This represents a historical-future integrated water quality sequence.

[0120] Step 300: Based on the historical-future fused water quality sequence, an attention mechanism that combines dynamic Gaussian kernel and dual-path routing is adopted. Through a two-stage routing selection strategy, the weighted fused attention result is output.

[0121] In time series prediction of aquaculture water quality parameters, data distribution often shifts over time, meaning the statistical characteristics of the time series change significantly across different periods. This distribution shift weakens the model's generalization ability on new data, causing models that perform well during a particular training period to experience decreased prediction accuracy, increased errors, and unstable output when faced with changes in the period or external environment. Especially in long-term prediction tasks, the cumulative effect of distribution shift further amplifies model bias, making it difficult to consistently and accurately capture future dynamic trends.

[0122] Traditional multi-head attention mechanisms (MHA) have limitations in handling distribution drift. Because the weights of all attention heads are fixed during training, they cannot flexibly adjust their attention patterns according to dynamic changes in data distribution. This static configuration may result in good model performance on training data, but when the distribution of test data shifts, some attention heads may become ineffective, causing the overall attention distribution to deviate from key features and thus reducing prediction accuracy. Furthermore, static attention mechanisms struggle to effectively select feature representations suitable for different time periods, easily introducing information redundancy or causing the loss of crucial information, thus limiting the model's adaptability to complex environmental changes.

[0123] Based on this, this embodiment proposes a dual-path routing attention mechanism that integrates a dynamic Gaussian kernel and a dual-branch attention mechanism. Its overall framework is as follows: Figure 3 As shown, this model can efficiently model multi-scale dependencies of time series data globally, employing a two-stage routing strategy. In the first stage, attention heads that may be of significant importance are initially selected based on the characteristics of the input features and task requirements. In the second stage, the initially selected attention heads are further refined, adaptively activating some of them. In this way, the model can focus on specific distribution patterns and key regions. For example, under certain water quality change conditions, the model can accurately identify which regions have the most significant changes in water quality indicators and which distribution patterns are closely related to water quality anomalies, thus more accurately capturing the patterns and characteristics of water quality changes.

[0124] The two-branch attention mechanism includes a position-dependent branch and a content-dependent branch. To more intuitively reveal its internal workings, the structure and computational flow of a single attention head are as follows: Figure 4 As shown, the historical-future integrated water quality sequence is first simultaneously fed into both the content-related branch and the location-related branch. The location-related branch, upon receiving the input, constructs an adaptive Gaussian kernel using a learnable standard deviation parameter to generate a location bias signal. The content-related branch directly performs a linear mapping on the input features to obtain Q, K, and V representations. The dynamic location bias signal generated by the location-related branch and the attention score calculated by the content-related branch are aligned at the same time step and then fed into the routing module. The routing module weights the two attention results based on the routing weights generated from the input features, thus obtaining the final output of the attention head.

[0125] Step 310: Based on the historical-future fused water quality sequence, construct an adaptive Gaussian kernel using a learnable standard deviation parameter, perform causal mask offset on the Gaussian kernel output, and then multiply it with the value vector of the location correlation branch to generate a location bias signal.

[0126] The positional dependency branch, guided by a Gaussian kernel, better describes the dynamic dependency structure between time steps. This is achieved through a learnable standard deviation parameter. An adaptive Gaussian kernel is constructed to introduce a dynamic position bias for attention allocation. This Gaussian kernel can adaptively adjust its position based on sequence features. With position The strength of the correlation between them enables the model to capture local neighborhood features while maintaining its ability to characterize global patterns.

[0127] ;

[0128] In the above formula, This represents the time series position index, indicating the correlation between different time steps; It is the numerical stability constant; Indicates position and The dynamic correlation weights are adaptively adjusted by using a Gaussian kernel to balance the local neighborhood and global patterns. From input features Position is controlled by the standard deviation parameter dynamically generated by the MLP. The attenuation range of surrounding correlation; this formula, through local sensitivity adjustment, enables the model to maintain high-resolution modeling of nearest-neighbor relationships under distribution drift, while suppressing long-distance noise interference. When When, focus attention on nearest neighbors (to deal with mutation distribution); when When the weights tend to be evenly distributed (applicable to the steady phase), the weights tend to be evenly distributed.

[0129] Causal mask offset is applied to the Gaussian kernel output g, and then the value vector of the position-dependent branch is used. Multiply to generate a position offset signal:

[0130] ;

[0131] ;

[0132] In the above formula, The matrix is ​​a linear transformation matrix; Implement causal mask offset to prevent future information leakage; This indicates the generation of a position offset signal.

[0133] Step 320: Based on the historical-future fused water quality sequence, perform linear mapping to calculate the query vector, key vector, and value vector of the content relevance branch, and then calculate the attention score.

[0134] The content relevance branch focuses on extracting significant pattern features from time-series signals, performing a linear mapping on the historical-future fused water quality sequence to obtain the query vector Q, the key vector K, and the value vector of the content relevance branch. :

[0135] ;

[0136] ;

[0137] ;

[0138] In the above formula, This indicates a query for a linear transformation matrix; Represents the linear transformation matrix of the bond; The matrix represents a linear transformation of the value.

[0139] Calculate the attention score based on the query vector, key vector, and value vector:

[0140] ;

[0141] In the above formula, This represents the dimension of the key vector, used to scale the dot product result and prevent gradient vanishing. The normalization function is used to convert attention scores into probability distributions, highlighting significant pattern features such as sudden changes in water quality and periodic fluctuations. A value vector representing the content relevance branch.

[0142] Step 330: Through the routing mechanism, the position bias signal and the attention score are weighted and fused to obtain the fused attention score.

[0143] The two types of attention results are adaptively fused using a routing mechanism, and the specific calculation formula is as follows:

[0144] ;

[0145] ;

[0146] In the above formula, This represents the attention score after output fusion; This represents dynamic gating weights, which are calculated using a learnable matrix to balance the contributions of the content relevance branch and the location relevance branch. and It is a learnable matrix; This represents an activation function that introduces nonlinearity to enhance expressive power.

[0147] When the input is stable, such as when the water quality is stable over a long period of time, the global dependency is strengthened and the content-related branch dominates.

[0148] When mutations are detected, such as algal blooms, local smoothing is enhanced, and location-dependent branches dominate.

[0149] Step 340: Based on the fused attention score, the weighted fused attention result is output through the collaborative work of the shared head and the routing head.

[0150] By incorporating the concept of MoE (Hybrid Experts), each head is treated as an expert. Through the collaborative work of the shared head and the routing head, combined with a two-stage routing strategy, efficient and stable feature extraction is achieved in scenarios with distributional drift. The shared head ensures the capture of basic dependencies, the routing head is dynamically activated for specific patterns, and the balancing coefficient enables adaptive adjustment of weights. The final output is a weighted fusion of attention results, significantly improving the model's ability to model complex and dynamic scenes.

[0151] In this routing strategy, the routing score is determined by the score of each individual header and a score associated with the header type. Specifically, given... The first in Input markers Routing score Defined as:

[0152] ;

[0153] in, Indicates the number of shared headers; and These represent the projection matrices of the shared header and the routing header, respectively. i Indicates the first i Individual attention; balance coefficient and The contribution of the shared header and the routing header is balanced and defined as follows:

[0154] ;

[0155] in, It is a trainable projection matrix; yes The hidden size.

[0156] make sure + =1, dynamically adjust the contributions of the shared header and routing header:

[0157] When the input is stable (e.g., water quality is stable over a long period of time), →1. Enhance the extraction of general features from shared heads.

[0158] When mutations (such as algal blooms) are detected, →1, Enhances the ability to capture specific patterns in the routing header.

[0159] The overall output is a combination of the shared header and the routing header activated through the gating mechanism:

[0160] ;

[0161] In the above formula, This represents the attention result after weighted fusion; Indicates the first i The attention after fusion of outputs (obtained through content-location dual-branch calculation).

[0162] The shared header ensures stable extraction of basic features, avoiding the loss of basic information due to distribution drift. The routing header uses Top-K sparse activation, activating only when a specific distribution pattern is detected, reducing computational redundancy and improving response speed to sudden changes.

[0163] Step 400: Based on the attention results after weighted fusion, a cross-granularity sequence interactive perceptron is used to explore the symmetric coupling relationship between variables at different time granularities to obtain the final output features.

[0164] Modeling cross-channel correlations plays a crucial role in improving the accuracy and stability of aquaculture water quality parameter predictions. Besides direct dependencies between variables, the interactions of different variables at different time scales (such as the coupling of short-term fluctuations and long-term trends) also influence the evolution patterns of the series. If this cross-scale correlation is not fully captured, the model often struggles to accurately characterize complex time-series dynamics. While multilayer perceptrons (MLPs) offer lightweight advantages in modeling multivariate dependencies due to their lower computational complexity and fewer parameters, their fully connected structure makes them insensitive to input order, hindering their ability to effectively model temporal dependencies. Furthermore, the lack of inductive bias for time series makes it difficult for MLPs to adaptively extract long-term dependencies and local patterns at different scales, thus limiting their application in complex time-series prediction tasks.

[0165] To further enhance the model's ability to model complex dynamic dependencies, a cross-granularity sequence interaction MLP was introduced. In water quality monitoring of aquaculture water bodies, complex coupling relationships exist between water quality changes at different time granularities and between different variables. Cross-granularity sequence interaction MLP utilizes bidirectional outer product operations to deeply explore these symmetrical coupling relationships between variables at different time granularities. Bidirectional outer product operations can process and analyze data from multiple perspectives, considering not only the mutual influence between different granularities over time but also the interactions between different variables. Through this operational method, the model can more comprehensively and deeply understand the complex dynamic dependencies in water quality data, thereby more accurately predicting water quality change trends and providing more reliable support for the management and decision-making of aquaculture water bodies.

[0166] Step 410: Based on the attention results after weighted fusion, local context features of different granularities are extracted through temporal multi-scale convolution, and the feature expression is enhanced by cross-scale splicing and channel recombination to obtain the recombined features.

[0167] In the time domain dimension, using different scales The convolutional branches extract the attention results after weighted fusion. Local contextual features are used to capture short-term temporal fluctuations and multi-granularity local patterns. The representational capability is enhanced through cross-scale feature concatenation and hierarchical reorganization strategies, enabling the model to perceive local dynamic relationships between variables. The specific process is shown in the following formula:

[0168] ;

[0169] ;

[0170] in, Represented as a channel reorganization operator, Indicates tensor splicing; Indicates the kernel size; Represents features at different scales; This indicates the characteristics after recombination.

[0171] Step 420: Based on the recombined features, a dual-path pooling mechanism is introduced. Through dimensionality reduction, nonlinear activation, and dimensionality increase operations, adaptive scaling factors and offsets for each channel are generated, and global statistics are calculated.

[0172] A dual-path pooling mechanism is introduced to aggregate global statistical information over time, thereby integrating long-term dependencies across variables and enhancing the model's ability to perceive overall temporal patterns. A parameter-decoupled, two-layer bottleneck MLP structure is designed. Through dimensionality reduction, nonlinear activation, and dimensionality increase operations, channel-adaptive scaling factors and offsets are generated respectively to explicitly model higher-order coupling relationships between different variables and strengthen the interaction of cross-scale information.

[0173] ;

[0174] ;

[0175] ;

[0176] in, Indicates feature splicing, For parameter-sharing dimensionality reduction networks: ;

[0177] Indicates global statistical characteristics; Indicates local statistical characteristics; , These represent the adaptive scaling factor and offset of the generated channels, respectively; This indicates global statistical information.

[0178] Step 430: Based on the channel-adaptive scale factor and offset, the local convolutional features and global statistical features are adaptively fused using a dynamic gating mechanism to obtain scale fusion gating and offset fusion gating.

[0179] The dynamic gating mechanism adaptively fuses local convolutional features and global statistical features, enabling the model to accurately capture multi-scale interaction patterns between variables while preserving the original temporal dynamic structure.

[0180]

[0181] ;

[0182] In the above formula, Indicates scale fusion gating; This indicates bias fusion gating; the Sigmoid function maps the input to the (0,1) interval, thus acting as a gating mechanism.

[0183] Step 440: Based on local features, global statistical information, scale fusion gating, and bias fusion gating, the symmetric coupling relationship between local and global features is modeled through bidirectional outer product operations. Residual connections are then used to preserve the original temporal structure, achieving efficient modeling of complex spatiotemporal dependencies and obtaining the final output features. .

[0184] This process can be represented as:

[0185] ;

[0186] ;

[0187] In the above formula, This represents the result of a two-way outer product operation; Indicates the first i Individual channel scale integrated gating; Indicates the first i Local features of each channel; Indicates the first j Offset fusion gating for each channel; Indicates the first j Global information for each channel; This is the interaction weight matrix for adaptive learning; d Indicates the number of channel dimensions.

[0188] To enhance the expressive power of bidirectional feature interactions, this invention defines symbols. The bidirectional outer product operation is represented as follows:

[0189] ;

[0190] in, This represents the standard outer product operation. This definition ensures the symmetry of information interaction between features, which helps to capture synergistic relationships across variables and scales.

[0191] By explicitly modeling the high-order interaction between two types of features through bidirectional outer product operations, local patterns and global trends can participate in the final representation construction in a symmetrical manner, thus overcoming the deficiency of traditional MLPs in capturing cross-scale correlations. Because this structure can simultaneously enhance local sensitivity and global dependency modeling capabilities, and avoid redundant feature interference under dynamic gating, it directly leads to a significant improvement in the prediction accuracy and stability of the model when dealing with complex, non-stationary multivariate time series.

[0192] Step 500: Map the final output features to the predicted values ​​of the target dimension through fully connected layers or convolutional layers.

[0193] Features after interaction X out It contains complex dependencies across scales and variables; these are handled through fully connected layers or convolutional layers. X out Predicted values ​​mapped to the target dimension:

[0194] ;

[0195] in, Let Y represent the flattening operator, and Y represent the predicted time series of multivariate water quality parameters.

[0196] To ensure the stability and reproducibility of the experimental results, the model used a uniform input sequence length and training configuration in all experimental settings. Specifically, the initial learning rate was set to 0.001, the batch size to 24, and the maximum number of training epochs to 100. An early stopping strategy was employed during training based on the validation set performance to mitigate the risk of overfitting. The experiments were conducted on an NVIDIA RTX 4060 (8GB) GPU, and the mean squared error (MSE) was used as the loss function during training.

[0197] Figures 5-16This paper demonstrates the predictive performance of the present invention on various key water quality parameters (conductivity, dissolved oxygen, chlorophyll a, turbidity, temperature, and pH) in water quality datasets (including Burnet, etc.), aiming to comprehensively evaluate the model's generalization ability. To more intuitively reflect the model's adaptability and stability under different scenarios, the present invention uses time series plots and scatter plots to visually compare and analyze the observed and predicted values ​​for each dataset. The visualization results of the time series plots show that the model's predicted curves at multiple monitoring stations and for various water quality parameters highly match the actual observed values, fully demonstrating the model's excellent generalization ability.

[0198] Specifically, the model not only accurately fits long-term trends, such as periodic temperature fluctuations, but also effectively responds to short-term abrupt changes and anomalies, such as the rapid rise and fall of high-frequency oscillating variables like dissolved oxygen and turbidity. Furthermore, the model's fitting performance remains consistent across different regions and water environment conditions, without significant regional overfitting or underfitting, indicating strong spatiotemporal transferability. Overall, the trend fidelity and fitting consistency reflected in the time series plots fully validate the model's generalization adaptability to diverse input features and time-series patterns. From the visualization results of the scatter plots, the model exhibits highly consistent fitting quality and prediction accuracy across different regions and various water quality parameter prediction tasks, further validating its excellent generalization ability. Specifically, the predicted values ​​and actual values ​​of each parameter show a strong linear relationship, with the scatter points closely distributed near the ideal diagonal in the plot, indicating that the model can effectively capture the true distribution characteristics of data across different variables and regions. The coefficient of determination for most indicators... With a slope exceeding 0.85 and approaching 1, the model not only possesses high-precision predictive ability but also exhibits a uniform error distribution without significant systematic bias or outliers. Overall, whether analyzed from the perspective of the fit to the predicted trend or the error distribution, the model demonstrates excellent stability and effectiveness in long-term water quality prediction tasks. In comparison with other models, the proposed model exhibits stronger generalization ability and predictive performance.

[0199] Based on the same inventive concept, this invention also provides a multivariate water quality parameter prediction system for aquaculture. The solution provided by this system is similar to the solution described in the above method; therefore, the specific limitations in one or more system embodiments provided below can be found in the limitations of the multivariate water quality parameter prediction method for aquaculture described above, and will not be repeated here.

[0200] Reference Figures 1-16This embodiment provides a multivariate water quality parameter prediction system for aquaculture. The system includes a data acquisition and input module, an adaptive characterization distillation module, a dual-path routing attention mechanism module, a cross-granularity sequence interaction multilayer perceptron module, and an output module.

[0201] The input module is used to acquire and input multivariate water quality monitoring time series. This data covers observed values ​​of water quality indicators such as conductivity, dissolved oxygen, chlorophyll a, turbidity, temperature, and pH at different time points, providing a fundamental data source for the entire prediction system.

[0202] Reference Figure 2 The adaptive characterization distillation module processes the input multivariate water quality monitoring time series, suppressing redundant features while retaining key information to obtain a historical-future fused water quality sequence. Specifically, the adaptive characterization distillation module includes a discrete wavelet decomposition submodule, an adaptive distillation network submodule, and a discrete wavelet reconstruction module. The discrete wavelet decomposition submodule decomposes the multivariate water quality monitoring time series into components of different frequencies to identify patterns and trends at different time resolutions, thereby comprehensively capturing feature information. The adaptive distillation network submodule uses information bottleneck theory to filter out redundant features that may be contained in multi-scale information and uses residual connections to ensure that valuable information is fully retained during the filtering process. The discrete wavelet reconstruction module uses wavelet reconstruction to reconstruct the processed data back to the time domain, obtaining a historical-future fused water quality sequence.

[0203] Reference Figures 3-4 The dual-path routing attention mechanism module is used to extract attention results based on historical-future fused water quality sequences. It employs a dynamic Gaussian kernel and a dual-path routing attention mechanism, using a two-stage routing selection strategy to output a weighted fused attention result. Specifically, the dual-path routing attention mechanism module includes an input layer, a dual-branch attention mechanism, a MoE routing strategy, and a fusion and output layer. The input layer takes the historical-future fused water quality sequences as input, which are fed in parallel into a shared head and a routing head, each head representing an independent expert. The dual-branch attention mechanism is the core feature extractor of each head, including a positional relevance branch and a content relevance branch. The positional relevance branch constructs an adaptive Gaussian kernel using a learnable standard deviation parameter after receiving the input to generate a positional bias signal. The content relevance branch directly performs a linear mapping on the input features to obtain query, key, and value vector representations and calculates the attention score. In the MoE routing strategy, the shared head is always active to extract general features, such as long-term trends. The routing head is activated through a Top-K mechanism, activating when a sudden change is detected, and the contributions of the shared head and the routing head are adjusted by a balancing coefficient. The fusion and output layer is used to weighted sum the outputs of all heads to generate the final representation.

[0204] A cross-granularity sequence interactive multilayer perceptron module is used to mine the symmetrical coupling relationship between variables at different temporal granularities based on the attention results after weighted fusion, and obtain the final output features. Specifically, the cross-granularity sequence interactive multilayer perceptron includes a temporal multi-scale convolution submodule, a dual-path pooling submodule, a dynamic gating submodule, and a bidirectional outer product operation modeling submodule. The temporal multi-scale convolution submodule introduces multi-scale convolution branches in the temporal domain to extract local contextual features, and achieves multi-granularity information fusion through feature concatenation and channel recombination; the dual-path pooling submodule generates channel-adaptive scale factors and offsets based on the recombined features through dimensionality reduction-nonlinear activation-dimensionality increase operations; the dynamic gating submodule adaptively fuses local convolutional features and global statistical features using a dynamic gating mechanism; the bidirectional outer product operation modeling submodule achieves efficient modeling of complex spatiotemporal dependencies by preserving the original temporal structure through the symmetrical coupling relationship between local and global features and residual connections, thus obtaining the final output features.

[0205] The output module is used to map the final output features to predicted values ​​of the target dimension through fully connected layers or convolutional layers.

[0206] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for predicting multivariate water quality parameters in aquaculture, characterized in that, Includes the following steps: Step 100: Obtain multivariate water quality monitoring time series; Step 200: Perform adaptive characterization distillation processing on the multivariate water quality monitoring time series to obtain a history-future fused water quality series, including: Discrete wavelet decomposition is performed on the input multivariate water quality monitoring time series, decomposing it into approximation coefficients and detail coefficients. Based on these approximation and detail coefficients, filtering is performed, and information bottleneck theory is used to retain key information and filter out irrelevant features, resulting in processed approximation and detail coefficients. Based on these processed approximation and detail coefficients, wavelet reconstruction is used to reconstruct the time series, obtaining an estimate of the reconstructed time series. The estimated reconstructed time series is divided into multiple subsequences of length, and the number of placeholders is calculated to initialize a future placeholder sequence. The multiple subsequences are concatenated with the placeholder sequence to obtain a history-future fused water quality series. Step 300: Based on the historical-future fused water quality sequence, an attention mechanism combining dynamic Gaussian kernel and dual-path routing is adopted. Through a two-stage routing selection strategy, the weighted fused attention result is output, including: Based on the historical-future fused water quality sequence, an adaptive Gaussian kernel is constructed using a learnable standard deviation parameter. A causal mask offset is applied to the Gaussian kernel output, which is then multiplied with the value vector of the location-related branch to generate a location bias signal. Based on the historical-future fused water quality sequence, a linear mapping is performed to calculate the query vector, key vector, and value vector of the content-related branch, and then the attention score is calculated. Through a routing mechanism, the location bias signal and the attention score are weighted and fused to obtain the fused attention score. Based on the fused attention score, through the collaborative work of the shared header and the routing header, the weighted fused attention result is output. Step 400: Based on the weighted fusion attention results, a cross-granularity sequence interaction perceptron is used to explore the symmetric coupling relationship between variables at different time granularities, obtaining the final output features, including: Based on the attention results after weighted fusion, local contextual features of different granularities are extracted through temporal multi-scale convolution. These features are then enhanced through cross-scale concatenation and channel recombination to obtain recombined features. Based on these recombined features, a dual-path pooling mechanism is introduced. Through dimensionality reduction, nonlinear activation, and dimensionality increase operations, adaptive scale factors and offsets are generated for each channel, and global statistical information is calculated. Based on the adaptive scale factors and offsets, a dynamic gating mechanism is used to adaptively fuse local convolutional features and global statistical features, resulting in scale fusion gating and bias fusion gating. Based on local features, global statistical information, scale fusion gating, and bias fusion gating, a bidirectional outer product operation is used to model the symmetric coupling relationship between local and global features. Residual connections are then combined to preserve the original temporal structure, yielding the final output features. Step 500: Map the final output features to the predicted values ​​of the target dimension through fully connected layers or convolutional layers.

2. The method for predicting multivariate water quality parameters in aquaculture according to claim 1, characterized in that, The multivariate water quality monitoring time series includes conductivity, dissolved oxygen, chlorophyll a, turbidity, temperature, and pH.

3. The method for predicting multivariate water quality parameters in aquaculture according to claim 2, characterized in that, Based on the fused attention score, through the collaborative work of the shared header and the routing header, the output is a weighted fused attention result, including: The routing score is calculated by adding a score to each individual header and a score associated with the header type. Defined as: ; In the above formula, Indicates the number of shared headers; and These represent the projection matrices of the shared header and the routing header, respectively. Indicates the first in the input Input markers; Among them, the balance coefficient and The contribution of the shared header and the routing header is balanced and defined as follows: ； In the above formula, It is a trainable projection matrix; yes The hidden size; The weighted fusion attention result is calculated based on the routing score and the attention result after weighted fusion.

4. The method for predicting multivariate water quality parameters in aquaculture according to claim 3, characterized in that, Based on local features, global statistical information, scale fusion gating, and bias fusion gating, a symmetric coupling relationship between local and global features is modeled through bidirectional outer product operations. Residual connections are then used to preserve the original temporal structure, yielding the final output features. ,include: ； ； In the above formula, This represents the result of a two-way outer product operation; Indicates the first i Individual channel scale integrated gating; Indicates the first i Local features of each channel; Indicates the first j Offset fusion gating for each channel; Indicates the first j Global information for each channel; This is the interaction weight matrix for adaptive learning; This represents the attention result after weighted fusion; d Indicates the number of channel dimensions; This indicates a two-way outer product operation; The bidirectional outer product operation is as follows: ； in, This indicates the standard outer product operation.

5. A multivariate water quality parameter prediction system for aquaculture, employing the multivariate water quality parameter prediction method for aquaculture as described in any one of claims 1-4, characterized in that, The system includes a data acquisition and input module, an adaptive characterization distillation module, a dual-path routing attention mechanism module, a cross-granularity sequence interaction multilayer perceptron module, and an output module. The input module is used to acquire and input multivariate water quality monitoring time series; The adaptive characterization distillation module is used to process the input multivariate water quality monitoring time series to obtain a history-future fused water quality series; The dual-path routing attention mechanism module is used to output the weighted fusion attention result based on the historical-future fused water quality sequence, using a fusion dynamic Gaussian kernel and dual-path routing attention mechanism. Through a two-stage routing selection strategy, it outputs the weighted fusion attention result. The cross-granularity sequence interaction multilayer perceptron module is used to mine the symmetrical coupling relationship between variables at different time granularities based on the attention results after weighted fusion, and obtain the final output features. The output module is used to map the final output features to predicted values ​​in the target dimension through fully connected or convolutional layers.

Images