Aquaculture disease prediction method based on internet of things

Abstract

The application discloses an aquaculture disease prediction method based on the Internet of Things and belongs to the technical field of aquaculture disease prediction. The method comprises the following steps: a data multi-source correlation module is used to acquire multi-source data collected by multiple Internet of Things sensors in an aquaculture area, multi-dimensional time series are aligned, and an aligned multi-source time series data set is obtained; a big data analysis module is used to perform time series coupling analysis on the multi-source time series data set, and a disease inducing factor correlation graph is obtained; an online prediction module is used to perform dynamic evolution analysis on the disease inducing factor correlation graph, and a disease risk index at the current moment is obtained; an offline matching module is used to extract features from a historical disease case library, to perform similarity matching between the disease risk index and historical features, and to obtain similar historical cases; an adaptive updating module is used to correct the disease risk index according to a disease confirmation result of the similar historical cases; and a comprehensive generation module is used to perform hierarchical coding on the corrected disease risk index, and to generate a disease early warning signal.

Description

Technical Field

[0001] This invention relates to the field of aquaculture disease prediction technology, specifically to an aquaculture disease prediction method based on the Internet of Things. Background Technology

[0002] In aquaculture, disease outbreaks often cause severe economic losses. Existing disease prediction methods mostly rely on threshold monitoring of single water quality indicators or simple statistical analysis of several environmental parameters. These methods treat each sensor data point as an independent variable, failing to consider the interplay of multiple factors over time, such as water quality, meteorological conditions, feeding management, and biological activity. In the aquaculture environment, phenomena such as decreased dissolved oxygen, sudden temperature changes, and abnormal fish behavior do not occur in isolation but rather exhibit temporally coupled driving relationships. Traditional threshold judgments cannot capture the lagged responses and interconnected changes among these multiple factors, resulting in insufficient early risk identification capabilities and often issuing warnings only after disease has clearly occurred.

[0003] Existing data-driven disease prediction models often operate with static parameters after deployment, lacking a feedback loop between the predicted risk index and the actual disease occurrence. When the predicted risk index deviates from the actual disease intensity or outbreak, the system cannot self-correct using historically confirmed case information. Consequently, the model's adaptability to changes in the aquaculture environment and disease pattern drift decreases, making it difficult to maintain accuracy and stability in the long term.

[0004] To address the above situation, the following issues need to be resolved: Under conditions where various IoT sensor data are highly asynchronous and mutually influential, how to accurately characterize the temporal coupling structure between multiple factors, and dynamically infer the evolution trend of disease risk based on this structure, in order to form a risk metric that more closely reflects the actual disease-inducing mechanisms. Simultaneously, it is also necessary to solve how to effectively correlate the diagnostic results of historical cases with the current risk index, so that the prediction results can be continuously adaptively corrected based on actual disease feedback, avoiding prediction degradation caused by environmental time-varying factors and model rigidity. Summary of the Invention

[0005] This invention aims to provide an IoT-based method for predicting diseases in aquaculture, addressing the shortcomings of existing technologies such as the inability of single-index threshold determination to characterize the temporal coupling driving relationship of multiple factors, and the lack of long-term adaptability due to the absence of a closed loop between static prediction models and actual disease feedback. The invention enables dynamic analysis of the coupled network of disease-inducing factors and real-time risk extrapolation, while simultaneously using the diagnostic results of historical cases to perform closed-loop correction of the prediction index, thereby improving the accuracy of disease early warning and environmental adaptability.

[0006] To achieve the above objectives, this invention provides the following technical solution: This invention provides an IoT-based method for predicting diseases in aquaculture, including a multi-source data association module, a big data analysis module, an online prediction module, an offline matching module, an adaptive update module, and a comprehensive generation module. The multi-source data association module acquires multi-source data collected by multiple IoT sensors within the aquaculture area, performs multi-dimensional time-series alignment on the multi-source data, and obtains an aligned multi-source time-series dataset. This eliminates inconsistencies in acquisition frequency, timescale deviation, and data gaps among multiple sensors, providing a high-quality, spatiotemporally consistent data foundation for subsequent analysis. The big data analysis module acquires the aligned multi-source time-series dataset and performs time-series coupling analysis to uncover the cross-coupling relationships between various influencing factors such as water quality, weather, feeding, and fish behavior, obtaining a disease-inducing factor correlation map, thereby accurately identifying the inducing factors that play a key driving role in disease occurrence. The online prediction module acquires the disease-inducing factor correlation map, performs dynamic evolution analysis on it, deduces the propagation and accumulation process of disease risk, obtains the disease risk index at the current moment, and achieves real-time quantitative assessment of disease risk in aquaculture waters. The offline matching module acquires a historical disease case database, extracts features from the historical multi-source time-series data corresponding to the historical cases to obtain a historical feature database, and performs similarity matching between the disease risk index and historical features to obtain similar historical cases most closely related to the current risk situation, thereby improving the reliability of risk assessment by leveraging historical experience. The adaptive update module acquires the disease confirmation results corresponding to the similar historical cases, and corrects the disease risk index based on the confirmation results to obtain a corrected disease risk index. This forms a closed-loop calibration of the prediction results by actual feedback, enabling the system to autonomously adapt to new trends in the breeding environment and disease evolution. The comprehensive generation module acquires the corrected disease risk index, performs hierarchical encoding on it, and generates disease early warning signals, providing intuitive and hierarchical decision-making reference information for breeding management, which is conducive to timely implementation of targeted prevention and control measures.

[0007] As a technical solution of the present invention, the multi-source data association module, when performing multi-dimensional time series alignment on multi-source data, first obtains the data acquisition timestamp of each IoT sensor, establishes a time reference coordinate system for each sensor based on the data acquisition timestamp, and maps the multi-source data onto the time reference coordinate system to obtain the time series points of each data. Then, linear interpolation is performed on the time series points of each data to form a continuous time series curve. Next, the data length, data frequency, and data sampling interval of the continuous time series curve are normalized to obtain normalized time series data. Finally, time-stamp alignment is performed on the normalized time series data to generate an aligned multi-source time series dataset. Through the above process, the system effectively integrates heterogeneous data at different time granularities, achieving accurate synchronization of multi-source data within a unified time frame.

[0008] Preferably, when aligning the time series data after normalization, the correlation degree between any two sequences in each time series is calculated to obtain a correlation degree matrix. Strongly correlated sequence sets with correlation degrees greater than a preset correlation threshold are then selected. Principal component analysis is performed on this strongly correlated sequence set to extract principal component sequences, which are then used as the aligned multi-source time series dataset. Utilizing correlation degree filtering and principal component extraction can compress data dimensionality while retaining core disease-related information, reducing the computational burden of subsequent analysis, and filtering out irrelevant noise. Furthermore, when calculating the correlation degree, a sliding window is used to divide the normalized time series data. Within each time window segment, the Pearson correlation coefficient between any two time series data is calculated to obtain the in-window correlation coefficient. Then, a weighted average of the in-window correlation coefficients of all time window segments is taken to obtain the correlation degree between them. This processing method can capture the local correlation characteristics of time series data in different time periods, more accurately reflecting the underlying dynamic correlations.

[0009] The big data analysis module, when performing time-series coupling analysis on the aligned multi-source time-series dataset, separates the time series of water quality factors, meteorological factors, feed factors, and biological activity factors from the dataset. It then performs multivariate time-series regulation analysis on these factor time series to obtain the cross-coupling coefficients between each factor, and constructs a time-series coupling network with each factor as a node and the cross-coupling coefficients as edge weights. Next, it identifies key nodes in each node of the time-series coupling network, extracting nodes with the maximum out-degree, maximum in-degree, and maximum betweenness centrality as disease-inducing factor nodes. Finally, it combines all disease-inducing factor nodes and their edges to obtain a disease-inducing factor association map. This approach reveals the implicit coupling patterns between environmental and biological factors, making the identification of disease causes more objective and comprehensive. Preferably, in the multivariate time-series moderating analysis, the time series of each factor are paired to form multiple factor pairs. Lag correlation analysis is performed on each factor pair, calculating the correlation values ​​at different lag orders, and the lag order corresponding to the maximum correlation value is taken as the lag time of that factor pair. Then, the two factor time series are time-aligned using this lag time as the time offset, and the mutual information value of the aligned factor pair time series is calculated. This mutual information value is used as the cross-coupling coefficient of the corresponding factor pair. By introducing time lag correction and mutual information measurement, the nonlinear and asymmetric dependencies between factors can be effectively characterized, making the measurement of coupling strength more accurate.

[0010] When performing dynamic evolution analysis on the correlation graph of disease-inducing factors, the online prediction module first acquires the topological structure of the graph and determines the predecessor and successor node sets for each node. Then, it aggregates the states of the predecessor node sets of each node to obtain an aggregated predecessor state vector. Using this aggregated predecessor state vector and the input real-time data of each node, it performs covariance analysis to obtain the state propagation probability of each node. Next, it performs time-series propagation deduction on the correlation graph of disease-inducing factors based on the state propagation probability to obtain the deduced node state values. Finally, it weights and sums all the deduced node state values ​​to obtain the disease risk index at the current moment. The state propagation deduction based on the directed graph can realistically simulate the evolutionary dynamics of disease risk, significantly enhancing the real-time performance and sensitivity of the risk index.

[0011] When the offline matching module performs similarity matching between the disease risk index and the historical feature database, it first encodes the disease risk index to obtain a risk feature vector to be matched. Then, it calculates the Euclidean distance and cosine similarity between this risk feature vector and each historical feature vector in the historical feature database. The Euclidean distance is standardized to obtain the distance similarity, and the cosine similarity is normalized to obtain the directional similarity. The distance and directional similarities are weighted and fused to obtain a comprehensive similarity. The historical case corresponding to the historical feature vector with the highest comprehensive similarity is selected as the similar historical case. By fusing both distance and directional similarity measures, the module can more accurately retrieve the case that best matches the current risk pattern from the historical database, providing a reliable basis for subsequent corrections and decisions.

[0012] When the adaptive update module corrects the disease risk index based on the disease confirmation results, it obtains the disease confirmation results corresponding to similar historical cases, including the disease outbreak confirmation identifier and the disease intensity level. Based on the disease outbreak confirmation identifier, it calculates the matching accuracy between the disease risk index and the actual disease, obtaining an accuracy assessment value. Based on this accuracy assessment value, it adjusts the calculation parameters of the disease risk index to obtain the adjusted disease risk index. Then, based on the disease intensity level, it performs intensity mapping on the adjusted disease risk index to obtain the corrected disease risk index. This closed-loop adaptive mechanism allows the system to optimize its parameters after each actual verification, continuously improving the prediction accuracy. Preferably, during parameter adjustment, the accuracy assessment value is compared with a preset accuracy threshold; if it is lower than the threshold, the parameter adjustment process is initiated. In this process, sensitivity analysis is performed on the weights of each factor in the disease risk index calculation model to determine the weight sensitivity of each factor. Then, based on the weight sensitivity, gradient descent optimization is performed on the weights of each factor to obtain the optimized weights of each factor. Finally, the disease risk index is recalculated based on the optimized weights to obtain the adjusted disease risk index. The above method enables automatic optimization of model weights, ensuring that the prediction system has continuous adaptive capability in the face of dynamic changes in the aquaculture environment.

[0013] The integrated generation module performs hierarchical coding based on the corrected disease risk index, generating disease early warning signals that include different warning levels and corresponding prevention and control recommendations. Through the coordinated operation of multi-source data alignment, temporal coupling analysis, dynamic risk extrapolation, historical matching correction, and adaptive closed-loop updates, the entire system significantly improves the accuracy, real-time performance, and stability of aquaculture disease prediction, providing strong technical support for early warning and scientific prevention and control of aquaculture diseases.

[0014] The technical effects and advantages provided by the present invention in the above technical solution are as follows: By employing a big data analytics module to perform temporal coupling analysis on the aligned multi-source time-series dataset, multivariate temporal regulation is applied to the time series data of water quality factors, meteorological factors, feed factors, and biological activity factors. The mutual information values ​​of each factor pair under different lag orders are calculated as cross-coupling coefficients, thereby constructing a temporal coupling network with each factor as a node and the cross-coupling coefficients as edge weights. Nodes with the maximum out-degree, maximum in-degree, and maximum betweenness centrality are identified as disease-inducing factor nodes, generating a disease-inducing factor correlation map. This approach no longer treats multiple sensor parameters as independent time series but explicitly extracts the nonlinear lag-driven relationships between factors and their key positions in the network. This allows disease risk analysis to focus on the core factor combinations that truly have inducing driving effects, overcoming the neglect of multidimensional temporal coupling relationships in traditional methods and making risk identification closer to the actual dynamic processes of disease occurrence in aquatic environments. The online prediction module aggregates the states of the predecessor node set for each node based on the topology of the graph, and performs covariance analysis with the real-time data of the current nodes to obtain the state propagation probability. Based on this, it performs time-series propagation inference to obtain the disease risk index at the current moment. This process combines the structural constraints of cross-coupling between factors and real-time data-driven state propagation reasoning, which can dynamically reflect the risk evolution direction driven by multiple factors and avoid false alarms and false negatives under complex environmental changes caused by static threshold judgment.

[0015] The offline matching module extracts features from multi-source time-series data of each historical case in the historical disease case database to construct a historical feature library. The current disease risk index is encoded as a risk feature vector to be matched, and its Euclidean distance and cosine similarity with each historical feature vector are calculated and fused into a comprehensive similarity score. The historical case with the highest comprehensive similarity score is selected as the similar historical case. The adaptive update module calculates the accuracy assessment value between the disease risk index and the actual disease based on the disease outbreak confirmation marker in the disease confirmation results of similar historical cases. When the accuracy assessment value is lower than a preset accuracy threshold, sensitivity analysis is performed on the weights of each factor in the disease risk index calculation model, and weight optimization is performed along the gradient descent direction. Simultaneously, the adjusted index is intensity-mapped according to the disease intensity level to obtain a corrected disease risk index. Therefore, the system's prediction results do not rely on static parameter settings, but rather gradually eliminate systematic deviations between predictions and reality through continuous comparison with actual disease cases and parameter feedback adjustments. This ensures that the disease risk index maintains high indicative accuracy under time-varying aquaculture environments and disease pattern drift conditions, significantly enhancing the stability and adaptability of long-term online operation. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0017] Figure 1 This is a schematic diagram of a disease prediction method for aquaculture based on the Internet of Things. Figure 2 This is a flowchart for calculating the disease risk index based on the correlation map of disease-inducing factors; Figure 3 This is a flowchart of the adaptive adjustment and intensity correction of the disease risk index; Figure 4 It is a multi-factor time-series normalized curve in the aquaculture disease prediction system; Figure 5 It is a curve comparing the adjustment and correction of the disease risk index. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, 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.

[0019] See Figure 1This invention provides an IoT-based method for predicting diseases in aquaculture, comprising a multi-source data association module, a big data analysis module, an online prediction module, an offline matching module, an adaptive update module, and a comprehensive generation module. The multi-source data association module acquires multi-source data collected by multiple IoT sensors within the aquaculture area, performs multi-dimensional time-series alignment on the multi-source data, and obtains an aligned multi-source time-series dataset. The big data analysis module acquires the aligned multi-source time-series dataset, performs time-series coupling analysis on the aligned multi-source time-series dataset, and obtains a disease-inducing factor correlation map. The online prediction module acquires the disease-inducing factor correlation map, performs dynamic evolution analysis on the disease-inducing factor correlation map, and obtains the disease risk index at the current moment. The offline matching module acquires a historical disease case library, extracts features from the historical multi-source time-series data corresponding to each historical case in the historical disease case library, obtains a historical feature library, and performs similarity matching between the disease risk index and the historical features in the historical feature library to obtain similar historical cases. The adaptive update module obtains the disease confirmation results corresponding to similar historical cases, and corrects the disease risk index based on the disease confirmation results to obtain the corrected disease risk index. The comprehensive generation module obtains the corrected disease risk index, performs hierarchical coding on the corrected disease risk index, and generates disease early warning signals.

[0020] Example 1: In practical implementation, multi-source data includes real-time water quality data, meteorological data, feeding data, and fish behavior data. Real-time water quality data is collected by dissolved oxygen, pH, ammonia nitrogen, and water temperature sensors; meteorological data is collected by wind speed and direction sensors, light sensors, and rainfall sensors; feeding data is collected by feeder operation status sensors and feed consumption metering sensors; and fish behavior data is collected by underwater cameras and fish movement trajectory tracking sensors. The data acquisition timestamps of each IoT sensor are obtained, and a time reference coordinate system is established for each sensor based on these timestamps. The method for establishing the time reference coordinate system is as follows: the earliest timestamp from all IoT sensors is set as a unified time origin, and a fixed time step is used as the coordinate axis scale to construct a unified time axis extending from the time origin to the latest timestamp. The multi-source data are then mapped onto the time reference coordinate system to obtain the time series points of each data point. That is, the collected values ​​of each sensor are mapped to the corresponding scale position on the unified time axis according to their timestamps, forming a set of discrete data points on the unified time axis.

[0021] In practice, linear interpolation is performed on the time series points of each data point to obtain a continuous time series curve for each data point. The method of linear interpolation is as follows: for any two adjacent existing data points on a unified time axis, if the time interval between the two existing data points is greater than one time step, then several supplementary data points are inserted between these two existing data points at equal time intervals. The value of each supplementary data point is determined by linear interpolation calculation using the values ​​and time positions of the two preceding and following existing data points. After linear interpolation, the data from each IoT sensor forms a continuously defined time series curve on a unified time axis.

[0022] In practice, the data length, frequency, and sampling interval of each continuous time series curve are uniformized to obtain uniformized time series data. The method for data length uniformization is as follows: a uniform analysis period length is set. If the time span of the time series curve is greater than the analysis period length, a segment within that period is extracted. If the time span is less than the analysis period length, a default value segment is added to the end of the curve to make the total length reach the analysis period length. The default value is the average value of the time series curve. The method for data frequency uniformization is as follows: all time series curves are resampled to a uniform sampling frequency. During resampling, if the original data point time and the resampled point time are completely consistent, the value is directly taken; otherwise, linear interpolation of neighboring data points is used to determine the value of the resampled point. The method for data sampling interval uniformization is as follows: a uniform time step is set, and the discrete point intervals of the time series curves are unified to this time step. Redundant data points are deleted, or interpolation points are generated at locations where no data points exist.

[0023] In practice, time-stamp alignment is performed on the normalized data time series to generate an aligned multi-source time series dataset. During time-stamp alignment, the correlation degree between any two data time series within the normalized dataset is calculated to obtain a correlation matrix. The correlation degree is calculated by dividing the normalized data time series into multiple time window segments using a sliding window. The window length is set to a fixed time length, and the sliding step is set to half the window length. Within each time window segment, the Pearson correlation coefficient between any two data time series is calculated to obtain the intra-window correlation coefficient. The formula for calculating the Pearson correlation coefficient is: ; in, Indicates a sliding window Internal data time series and data time series Pearson correlation coefficient, Representing data time series In the sliding window Inner Values ​​at each point in time. Representing data time series In the sliding window Inner Values ​​at each point in time. Representing data time series In the sliding window All The average value at each time point. Representing data time series In the sliding window All The average value at each time point. Represents a sliding window The total number of time points within the period, For sliding windows The sequence number of the time point within the time frame.

[0024] After calculations are performed on all sliding windows, the data time series is processed. and data time series The time series data is obtained by taking a weighted average of the in-window correlation coefficients across all sliding windows. and data time series The correlation between the data is calculated. The weights of the weighted average are set based on the proximity of the center time point of the sliding window to the current analysis time; the closer the center time point is to the current analysis time, the greater the weight. All correlations are arranged according to the category of the data time series to construct a correlation matrix. The rows and columns of the correlation matrix correspond to the normalized data time series, and the matrix elements are the correlation values ​​between the corresponding two data time series.

[0025] In practice, based on the correlation matrix, several data time series with a correlation greater than a preset correlation threshold are selected from the normalized data time series, and these data time series are considered as a strongly correlated sequence set. The preset correlation threshold is set to 0.75, based on the distribution of the correlation between any two data time series in historical batches of data, determined according to the upper quartiles. This threshold can effectively distinguish data series combinations with strong synergistic relationships. All data time series pairs with values ​​greater than the preset correlation threshold in the correlation matrix are extracted, and all non-repeating data time series involved in these data time series pairs are aggregated to form a strongly correlated sequence set.

[0026] In practice, principal component analysis (PCA) is performed on the strongly correlated sequence set to extract the principal component sequences, which are then used as the aligned multi-source time series dataset. The PCA method involves constructing the covariance matrix of the strongly correlated sequence set, calculating its eigenvalues ​​and eigenvectors, selecting the eigenvector corresponding to the largest eigenvalue as the first principal component direction, and projecting the time series data from each data point in the strongly correlated sequence set onto this first principal component direction to obtain a one-dimensional principal component sequence. This principal component sequence preserves the main variation patterns of the multi-source data in the strongly correlated sequence set, eliminates redundant information, and serves as the output of the aligned multi-source time series dataset.

[0027] Example 2: In practice, water quality factor time series, meteorological factor time series, feed factor time series, and biological activity factor time series are separated from the aligned multi-source time series dataset. The aligned multi-source time series dataset is a collection of data consisting of multiple time series that have undergone normalization, with each time series labeled with a data source category. The method for separating the water quality factor time series is as follows: all time series labeled with real-time water quality data as the data source category are selected from the aligned multi-source time series dataset, including dissolved oxygen sensor time series, pH sensor time series, ammonia nitrogen sensor time series, and water temperature sensor time series; the arithmetic mean of the values ​​of the four selected time series at the same time is calculated to generate a water quality factor time series representing the overall water quality status, with a length consistent with the length of each time series in the aligned multi-source time series dataset. The method for separating meteorological factor time series is as follows: Time series from the aligned multi-source time series dataset, categorized as meteorological data, are selected, including wind speed, light intensity, and rainfall time series. The values ​​of these three time series at the same time point are normalized, and their arithmetic mean is calculated to generate the meteorological factor time series. The method for separating feed factor time series is as follows: Time series from the aligned multi-source time series dataset, categorized as feeding data, are selected, including feeder operating power and feed consumption rate time series. The values ​​of these two time series at the same time point are calculated, and their arithmetic mean is calculated to generate the feed factor time series. The method for separating biological activity factor time series is as follows: Time series from the aligned multi-source time series dataset, categorized as fish behavior data, are selected, including fish swimming speed and fish aggregation density time series. The values ​​of these two time series at the same time point are calculated, and their arithmetic mean is calculated to generate the biological activity factor time series.

[0028] In practice, multivariate time-series regulation analysis was conducted on the time series of water quality factors, meteorological factors, feed factors, and biological activity factors to obtain the cross-coupling coefficients between each factor. The method for multivariate time-series regulation analysis was as follows: the time series of water quality factors, meteorological factors, feed factors, and biological activity factors were paired to obtain six factor pairs. These six factor pairs were: a pair consisting of water quality factor time series and meteorological factor time series; a pair consisting of water quality factor time series and feed factor time series; a pair consisting of water quality factor time series and biological activity factor time series; a pair consisting of meteorological factor time series and feed factor time series; a pair consisting of meteorological factor time series and biological activity factor time series; and a pair consisting of feed factor time series and biological activity factor time series. Lag correlation analysis was performed on the two factor time series in each factor pair to calculate the correlation values ​​at different lag orders. The method for lag correlation analysis is as follows: The time series of the first factor in the factor pair is designated as the reference factor time series, and the time series of the second factor is designated as the response factor time series. A maximum lag order L is set, with L being 7 time steps. This setting is based on the typical physiological response cycle in aquaculture where the delay in the impact of environmental factors on organisms usually does not exceed 7 days. From lag order d equal to 0 to L, the response factor time series is shifted backward by d time steps sequentially. The Pearson correlation coefficient between the shifted response factor time series and the reference factor time series in the overlapping time region is calculated. A correlation value is obtained at each lag order d, and the lag order corresponding to the maximum correlation value is taken as the lag time of this factor pair.

[0029] In practice, the lag time is used as a time offset to align the two factor time series, resulting in an aligned factor pair time series. The time alignment process is as follows: based on the obtained lag time, if the lag time is greater than 0, the response factor time series is shifted forward by several time steps, aligning the reference factor time series with the start time of the response factor time series; if the lag time is 0, no shift is needed. After shifting, the overlapping time intervals of the two factor time series are extracted, resulting in two factor time series of equal length, which are then used as the aligned factor pair time series. The mutual information value of the aligned factor pair time series is calculated and used as the cross-coupling coefficient of this factor pair. The calculation formula is: ; in, Indicates the first The and the first The sequence after discretizing the first factor in the time series of the factor pairs after aligning the factors. Indicates the first The and the first The sequence after discretizing the second factor in the time series of the factor pair after aligning the first factor. and The factor category number has a value range of 1 to 4. 1 corresponds to water quality factors, 2 corresponds to meteorological factors, 3 corresponds to feed factors, and 4 corresponds to biological activity factors. This represents the total number of intervals divided when the numerical values ​​of the two factor time series are discretized with equal width. The value is set to 5. The rationale is that when the continuous numerical domain is divided into 5 equal intervals, the sample size in each interval can be used to estimate the probability distribution relatively stably, while ensuring low information loss. This represents the index of the value range of the first factor after time series discretization. This represents the index of the time series discretization interval for the second factor. and All from 1 to ; This indicates that the time series value of the first factor falls within the first... The interval and the time series value of the second factor falls within the interval. The joint probability of each interval is given by the time points corresponding to the two factors after time series alignment (intervals). , interval The frequency of occurrence is calculated by dividing the total number of time points. This indicates that the time series value of the first factor falls within the first... The marginal probability of each interval is determined by the first factor time series falling into the interval alone. The frequency is calculated by dividing the total number of time points; This indicates that the time series value of the second factor falls within the first... The marginal probability of each interval is determined by the second factor time series falling into the interval alone. The frequency is calculated by dividing the total number of time points.

[0030] In practical implementation, a temporal coupling network is constructed based on the cross-coupling coefficient, with each factor as a node and the cross-coupling coefficient as the edge weight. The node set of the temporal coupling network is {water quality factor node, meteorological factor node, feed factor node, biological activity factor node}. The direction of the edges in the network is determined based on the lag time of each factor pair: if the lag time between the reference factor time series and the response factor time series in the factor pair indicates that the change in the response factor time series lags behind the reference factor time series, then a directed edge is established from the reference factor node to the response factor node; if the lag time is 0, then a bidirectional edge is established. The weight of each directed edge is set to the corresponding cross-coupling coefficient. This results in a temporal coupling network containing four nodes and several directed edges.

[0031] In practical implementation, key nodes are identified for each node in the temporally coupled network, and nodes with the maximum out-degree, maximum in-degree, and maximum betweenness centrality are extracted as disease-inducing factor nodes. The out-degree is calculated as follows: for a node in the temporally coupled network, the weights of all directed edges pointing to that node are summed, and the sum is taken as the out-degree of that node. The in-degree is calculated as follows: for a node in the temporally coupled network, the weights of all directed edges pointing to that node are summed, and the sum is taken as the in-degree of that node. The betweenness centrality is calculated as follows: for a node in the temporally coupled network, the total number of shortest paths between all other node pairs in the network and the number of shortest paths passing through that node are calculated. The ratio of the number of shortest paths passing through that node to the total number of shortest paths between all other node pairs is taken as the betweenness centrality of that node. The shortest path is measured using the reciprocal of the edge weight as the path length; that is, the larger the cross-coupling coefficient, the shorter the path length. Extract the node with the largest out-degree, the node with the largest in-degree, and the node with the largest betweenness centrality from all nodes. After removing duplicates from these three nodes, use them as disease-inducing factor nodes.

[0032] In practice, all disease-inducing factor nodes and the edges between them are combined to obtain a disease-inducing factor association graph. The combination method is as follows: extract a subset of nodes consisting of all disease-inducing factor nodes from the temporal coupling network, extract the directed edges between any two nodes in this subset of nodes in the temporal coupling network, retain the direction of the edges and the weight of the cross-coupling coefficients, and form a subgraph composed of disease-inducing factor nodes and the edges between them. This subgraph is the disease-inducing factor association graph.

[0033] See Figure 4 The figure shows the trend of normalized factor values ​​over time for the four types of factor time series separated from the aligned multi-source time series dataset in Example 2. The horizontal axis represents time in hours, ranging from approximately 0 to 500 hours; the vertical axis represents the normalized factor values, ranging from 0 to 1. The legend identifies the four curves as representing the time series of water quality factors (blue solid line), meteorological factors (orange dashed line), feed factors (green dotted line), and biological activity factors (red dotted line), respectively.

[0034] From the overall trend of the curves, all four types of factor time series exhibit obvious periodic fluctuations, with multiple peaks and troughs, showing the dynamic changes of each factor within the monitoring period. The periodic peaks of each factor time series are mostly concentrated around 50 hours, 270 hours, and 420 hours, with the normalized factor value in the peak interval approaching or even reaching 1, indicating relatively active factor activity during this period. The troughs are mostly distributed around 150 hours and 380 hours, with the normalized factor value dropping to near 0, reflecting a significant decrease in factor activity.

[0035] In terms of specific numerical changes, the fluctuation range of meteorological factors was slightly larger than that of other factors, with more significant peak-to-trough differences, exhibiting higher peaks and lower troughs. The time-series curves of water quality factors, feed factors, and biological activity factors showed relatively consistent trends and a high degree of overlap, reflecting a strong synchronicity and correlation among these three types of factors in their temporal evolution. In particular, during peak periods, the peak values ​​of water quality factors and biological activity factors almost overlapped, indicating a strong coupling relationship between water quality changes and fish behavior.

[0036] Example 3: In specific implementation, please refer to Figure 2 The topological structure of the disease-inducing factor association graph is obtained, and the predecessor and successor node sets of each node in the graph are determined based on the topological structure. The disease-inducing factor association graph is a directed weighted graph, and all nodes contained in the graph are denoted as the node set. ,in Let be the total number of nodes in the disease-inducing factor correlation graph, with each node corresponding to a disease-inducing factor. The topological structure of the disease-inducing factor correlation graph contains a set of directed edges, each directed edge pointing from a starting node to an ending node, and carrying a cross-coupling coefficient as its weight. Traverse the set of nodes. Each node in Determine the node The method for finding the precursor node set is as follows: Query all nodes in the disease induction factor association graph. For directed edges that terminate at a node, the set of the starting nodes of these directed edges is taken as the node. The set of predecessor nodes, denoted as Determine the node. The method for obtaining the successor node set is as follows: Query all nodes in the disease induction factor association graph. For each directed edge starting from a node, the set of the terminating nodes of these directed edges is considered as a node. The set of successor nodes, denoted as .

[0037] In practice, the state of each node's predecessor node set is aggregated to obtain the aggregated predecessor state vector for each node. Before state aggregation, the sensor monitoring values ​​corresponding to each disease-inducing factor node at the current moment are extracted from the multi-source real-time data stream and used as the node's current state value. Once the corresponding factor type is determined, the real-time data point of that factor type at the current moment is obtained from the aligned multi-source time series dataset and denoted as a node. Current state value For nodes If the predecessor node set If the set is non-empty, then traverse the set of predecessor nodes. Each predecessor node in Obtain the predecessor node Current state value and obtain from node Pointing to node Cross-coupling coefficient on directed edges Calculate the weighted average of the current state values ​​of all predecessor nodes, and use this average as the node's state value. Aggregation precursor state value The calculation method is to calculate the cross-coupling coefficient of each predecessor node. The proportion of the total cross-coupling coefficients of all predecessor nodes is used as the weight, which is then multiplied by the current state value of the corresponding predecessor node. And sum over all predecessor nodes. If the set of predecessor nodes... If the set is empty, then the nodes will be... Aggregation precursor state value Set to zero. At this point, each node... A clustered predecessor state value is obtained, and the clustered predecessor state values ​​of all nodes constitute the clustered predecessor state vector.

[0038] In practice, covariance analysis is performed on the aggregated predecessor state vector and the input real-time data of each current node to obtain the state propagation probability of each node. The input real-time data of each current node represents the current state value of all nodes. The vector formed by the covariance analysis is as follows: During system operation, a fixed-length time sliding window is continuously maintained, with a time sliding window length of [value missing]. Each time step The value is set to 30, based on the fact that 30 time steps cover the data fluctuation range within a typical incubation period of aquaculture diseases. The time sliding window stores the most recent... Each node at each time point Aggregated predecessor state value history sequence and the current state value history sequence The superscript indicates the time number within the time sliding window. For nodes... The aggregated predecessor state value history sequence Historical sequence of current state values Perform Pearson correlation analysis, calculate the absolute value of the Pearson correlation coefficient between the two, and use the calculated absolute value of the Pearson correlation coefficient as the node. The state transit probability, denoted as State transit probability The value of is a real number between 0 and 1. The larger the value, the more likely the state change of the predecessor node is to affect the node. The higher the probability of transmission.

[0039] In practice, based on the state propagation probability, a time-series propagation deduction is performed on the correlation graph of disease-inducing factors to obtain the deduced node state values. The input to the time-series propagation deduction is the current state value of all nodes at the current moment. and the corresponding node state propagation probability The process of time-series propagation is as follows: Record the current moment as time. By extrapolating a time step, we can obtain the time. The node state value. For each node in the disease induction factor correlation graph. The derived node state values It consists of two superimposed parts: the first part is a node. At any moment Current state value The second part is the propagation increment from all predecessor nodes. The propagation increment is calculated by traversing the nodes. Predecessor node set Each predecessor node in , the predecessor node State transit probability Predecessor node To the node Directed edge cross coupling coefficient and predecessor nodes Current state value With nodes Current state value The difference Multiply, sum, and then multiply by a propagation step size factor. , The value is set to 0.1, which is based on limiting the amplitude of state changes within a single time step to avoid numerical oscillations. All state propagation probabilities involved are the state propagation probabilities corresponding to the predecessor node.

[0040] In practice, all the derived node state values ​​are weighted and summed to obtain the disease risk index at the current moment. The formula for the weighted summation is: ; in, This represents the disease risk index at the current moment. It is a non-negative real number, and the larger the value, the higher the risk of disease outbreak. This represents the total number of nodes in the disease-inducing factor correlation graph; The node index is taken from 1 to... ; Represents a node The deduced node state values; Represents a node The weighted weighting coefficients. According to the node The determination of betweenness centrality in the disease-inducing factor association map is calculated as follows: The betweenness centrality value calculated for each node when establishing the disease-inducing factor association map is obtained, and the nodes are... The betweenness centrality value divided by all The sum of the betweenness centrality values ​​of each node is used to obtain the normalized betweenness centrality, which is then used as the weighting coefficient. After weighted summation, the disease risk index is output. This serves as a quantitative result of the disease risk at the current moment.

[0041] Example 4: In practice, the construction of the historical feature database was completed offline. A historical disease case database was acquired, containing multiple historical cases. Each historical case consisted of a historical disease outbreak confirmation marker, a historical disease intensity level, and a historical multi-source time-series dataset collected within a specified time period prior to the disease outbreak. The historical multi-source time-series dataset corresponding to each historical case in the database was collected over a continuous period of 168 hours (7 calendar days) prior to the confirmed outbreak. This duration was based on the fact that the incubation period for common bacterial and viral diseases in aquaculture is typically between 3 and 10 days, and the 7-day collection window covers the complete process from accumulation to the onset of clinical symptoms for most pathogenic factors. The historical disease case database was maintained at a size of no less than 200 historical cases, including both outbreak and non-outbreak cases, with the ratio of the two types controlled between 1:1 and 1:2 to ensure a balance between positive and negative samples. The criteria for case inclusion are as follows: each historical case must be accompanied by a disease outbreak confirmation marker and disease intensity level confirmed by laboratory pathogen detection or on-site diagnosis by a professional veterinarian; the completeness of the historical multi-source time-series dataset must be no less than 90%, meaning the total duration of missing data within 168 hours must not exceed 16.8 hours, and missing portions must be completed through linear interpolation before inclusion; a maximum of three historical cases are allowed to be included in the same aquaculture pond within the same calendar year to avoid feature bias caused by duplicate cases. The historical data collection period is consistent with the alignment period of the online real-time data, meaning that when the online prediction module performs dynamic evolution analysis, it also uses multi-source time-series data from the previous 168 hours as input, thereby ensuring that the offline constructed historical feature vector and the online generated risk feature vector to be matched are comparable in terms of time span and data granularity. For each historical case, the historical multi-source time-series dataset is processed in the same manner as the online disease risk index calculation. This involves multi-dimensional time series alignment, temporal coupling analysis to obtain a historical disease-inducing factor correlation graph, and dynamic evolution analysis to calculate the historical disease risk index for that case. Simultaneously, the node state values ​​and state propagation probabilities of each node generated during the historical disease risk index calculation are recorded. Feature encoding is then applied to the historical disease risk index for each historical case. The encoding method is as follows: Let the total number of nodes in the historical disease-inducing factor correlation graph be... Using the historical disease risk index as a scalar, The derived node state values ​​are arranged into a historical derived node state vector. The state transmission probabilities are arranged into a historical transmission probability vector. The historical disease risk index, the historical projection node state vector, and the historical transmission probability vector are then concatenated sequentially to form a vector with dimension [missing information]. The historical feature vectors are used to construct a historical feature library. Each historical feature vector in the library corresponds one-to-one with a historical case.

[0042] In practical implementation, the disease risk index calculated by the online prediction module at the current moment is feature-encoded to obtain a risk feature vector to be matched. The feature encoding method is consistent with the method used when constructing historical feature vectors. The disease risk index at the current moment is obtained from the online prediction module, and simultaneously, the deduced node state values ​​and state propagation probabilities of all nodes in the disease inducing factor correlation graph obtained during the dynamic evolution analysis are obtained. The total number of nodes in the disease inducing factor correlation graph is denoted as... , The specific values ​​are determined by the actual scale of the disease-inducing factor correlation map. The current inferred node state vector is composed of the inferred node state values. The current propagation probability vector is composed of the state propagation probabilities. The disease risk index at the current moment is taken as the first component of the vector, and the current inferred node state vector... The components are used as the second to third elements of the vector. Each component will be a part of the current transmitted probability vector. The nth component is the vector's nth component. To the Each component is combined to obtain a dimension. The vector is the risk feature vector to be matched.

[0043] In practice, the Euclidean distance and cosine similarity between the risk feature vector to be matched and each historical feature vector in the historical feature database are calculated. The risk feature vector to be matched is denoted as... The first in the historical feature database The historical feature vectors are denoted as ,in This refers to the index of the historical feature vector in the historical feature database. The risk feature vector to be matched. and each historical feature vector All dimensions are Euclidean distance The calculation method is as follows: the risk feature vector to be matched With historical feature vectors Calculate the difference between the components in each corresponding dimension, square the differences, sum them up, and take the square root of the sum to obtain the vector. with vector Euclidean distance Cosine similarity The calculation method is as follows: calculate the vector with vector The dot product, divided by the vector. Magnitude and vector The cosine similarity is obtained by multiplying the moduli of the product of the moduli. .

[0044] In practice, the calculated Euclidean distances for each historical feature vector are standardized to obtain the corresponding distance similarity. The standardization method is as follows: among the Euclidean distances between all historical feature vectors and the risk feature vectors to be matched in the historical feature database, the largest Euclidean distance is found and denoted as [the largest distance is denoted as '[']']. Find the minimum Euclidean distance, denoted as . For the first Each historical feature vector corresponds to a distance similarity. The calculation formula is When the Euclidean distances of all historical feature vectors in the historical feature database are equal, that is... equal At that time, the distance similarity corresponding to all historical feature vectors is calculated. Set all to 1. Distance similarity The value of is a real number between 0 and 1. The smaller the Euclidean distance of a historical feature vector, the closer its distance similarity value is to 1.

[0045] In practice, the cosine similarity corresponding to each historical feature vector is normalized to obtain the corresponding directional similarity. (Cosine similarity) The value range is a real number between -1 and 1, and the normalization operation uses... By linearly mapping the cosine similarity to the interval between 0 and 1, we obtain the directional similarity. Orientation similarity The larger the value, the higher the degree of agreement between the risk feature vector to be matched and the corresponding historical feature vector in the vector direction.

[0046] In practice, the distance similarity and directional similarity corresponding to each historical feature vector are weighted and fused to obtain a comprehensive similarity. The formula for weighted fusion is: ; in, Indicates the first The comprehensive similarity between historical feature vectors and risk feature vectors to be matched. The value of is a real number between 0 and 1; Indicates the first Distance similarity corresponding to each historical feature vector; Indicates the first The directional similarity corresponding to each historical feature vector; The fusion weight coefficients for distance similarity are... The value of is a real number between 0 and 1. The specific value is set to 0.5. The basis for this setting is to ensure that the numerical amplitude similarity represented by Euclidean distance and the distribution direction similarity represented by cosine similarity contribute equally to the overall similarity, without favoring either metric.

[0047] In practice, all historical feature vectors in the historical feature database are traversed, and a comprehensive similarity calculation is performed on each historical feature vector to obtain a comprehensive similarity set. ,in This represents the total number of historical feature vectors in the historical feature database. The maximum value is found in the comprehensive similarity set, and the historical feature vector corresponding to this maximum value is determined as the most similar historical feature vector. The historical case corresponding to the most similar historical feature vector is then searched in the historical disease case database, and this searched historical case is taken as the similar historical case. If the maximum comprehensive similarity value corresponds to multiple historical feature vectors with the same value, the distance similarity of these historical feature vectors is compared, and the historical case corresponding to the historical feature vector with the largest distance similarity is selected as the similar historical case.

[0048] Example 5: In specific implementation, please refer to Figure 3 The system retrieves disease confirmation results corresponding to similar historical cases. These results include a disease outbreak confirmation flag and a disease intensity level. Similar historical cases are obtained from the output of the offline matching module. These cases are those with the highest comprehensive similarity in feature space to the current disease risk index in the historical disease case database. Each historical case in the database undergoes manual review or laboratory testing upon entry and is accompanied by authoritative disease confirmation results. The disease outbreak confirmation flag is a binary variable; it is set to 1 when an aquaculture disease actually occurred in the historical case, and 0 when no such disease occurred. The disease intensity level is an ordered categorical variable, classified according to the mortality rate of aquaculture organisms caused by the disease in the historical cases: level 1 for a mortality rate less than 5%, level 2 for a mortality rate between 5% and 15%, and level 3 for a mortality rate above 15%.

[0049] In practice, based on the disease outbreak confirmation markers, the matching accuracy between the disease risk index and the actual disease is calculated to obtain an accuracy assessment value. The method for calculating the matching accuracy is as follows: the current disease risk index is compared with a preset risk judgment threshold, which is set to 0.5. This threshold is set based on maximizing the harmonic mean of the precision and recall rates of disease warnings during historical data backtesting. When the disease risk index is greater than or equal to 0.5, the system outputs "Predicted disease presence"; when the disease risk index is less than 0.5, the system outputs "Predicted no disease presence". The system's output of "predicted disease" or "predicted no disease" is compared with the disease outbreak confirmation flags corresponding to similar historical cases. If the disease outbreak confirmation flag of a similar historical case is 1 and the system output is "predicted disease", or if the disease outbreak confirmation flag of a similar historical case is 0 and the system output is "predicted no disease", then the matching accuracy evaluation value is recorded as 1. If the disease outbreak confirmation flag of a similar historical case is 1 but the system output is "predicted no disease", or if the disease outbreak confirmation flag of a similar historical case is 0 but the system output is "predicted disease", then the matching accuracy evaluation value is recorded as 0.

[0050] In practice, the calculation parameters of the disease risk index are adjusted based on the accuracy assessment value to obtain the adjusted disease risk index. The accuracy assessment value is compared with a preset accuracy threshold, which is set to 1. When the accuracy assessment value equals the preset accuracy threshold of 1, it indicates that the calculation parameters of the current disease risk index match the actual disease occurrence, and no adjustment is made to the calculation parameters of the disease risk index; the current disease risk index is directly used as the adjusted disease risk index. When the accuracy assessment value is lower than the preset accuracy threshold of 1, i.e., the accuracy assessment value equals 0, the parameter adjustment process is initiated.

[0051] In practical implementation, during the parameter adjustment process, sensitivity analysis is performed on the weights of each factor in the disease risk index calculation model to determine the sensitivity of each factor's weight. The disease risk index calculation model is expressed as a weighted summation formula. ,in This is a disease risk index. This represents the total number of nodes in the correlation graph of disease-inducing factors. For nodes The deduced node state values, For nodes The weighted weighting coefficients, i.e., the weights of each factor. The sensitivity analysis method involves sequentially analyzing each node in the disease-inducing factor correlation graph. Weighting coefficients Apply a small perturbation, the magnitude of which is set to the current weighting coefficient. Five percent of the original value was applied, and after applying a small perturbation, the weighting coefficients of all other nodes remained unchanged. The disease risk index was recalculated, and the absolute value of the change in the recalculated disease risk index relative to the unperturbed disease risk index was recorded. The absolute value of the change was divided by the magnitude of the small perturbation to obtain the node... Weighting coefficients For the local gradient approximation of the disease risk index, this local gradient approximation is used as a node. The weight sensitivity. For all factors in the disease-inducing factor correlation map. Each node performs the above operation one by one to obtain the weight sensitivity of each node.

[0052] In the parameter adjustment process, a continuously differentiable loss function is first constructed to guide weight optimization. The purpose of this loss function is to measure the deviation between the current disease risk index and an "expected risk index" determined based on similar historical cases. Specifically, the disease intensity levels of similar historical cases are obtained, and a target value for the expected risk index is determined based on these levels. When the disease severity level is 1, The value is 0.3; when the disease intensity level is 2, The value is 0.6; when the disease intensity level is 3, The target value is set to 0.9. This target value is set based on mapping different levels of disease severity to a continuous risk scale from 0 to 1, providing a clear optimization direction for subsequent gradient descent. Then, the loss function is defined. Current disease risk index With expected risk index The mean square error between them is calculated using the following formula: ; in, This represents the disease risk index at the current moment. This is the target value of the expected risk index, determined based on the severity levels of similar historical cases. (This loss function...) For weighting coefficients It is differentiable.

[0053] Next, calculate the loss function. Weights for each factor The gradient. Due to the loss function. Through disease risk index Indirectly dependent on weight According to the chain rule, gradient It can be represented as: ; in, , is the derivative of the loss function with respect to the risk index; It is the risk index relative to the weight. The partial derivatives. According to the risk index calculation formula... , can be obtained , i.e., node The derived node state values. Therefore, the loss function affects the weights. The gradient is: ; Once the accurate gradients are obtained, standard gradient descent optimization can be performed. For each node... Weighting coefficients Adjust along the negative gradient direction; the adjustment formula is: ; in, Indicates the first The weighted weight coefficients of each node after adjustment; Indicates the first The weighted coefficients of each node before adjustment, which are the values ​​currently used in the calculation of the disease risk index; The learning rate is set to 0.01. The rationale for this setting is to make the step size of each parameter update smaller in order to ensure the stability of the parameter adjustment process. The loss function calculated in the previous step is applied to the weights. The gradient. According to the above formula, for all... After updating the weighted weight coefficients of each node one by one, the updated weighted weight coefficients are normalized so that the sum of all weighted weight coefficients is still 1. The normalization method is to divide each updated weighted weight coefficient by the sum of all updated weighted weight coefficients to obtain the optimized weights of each factor.

[0054] In practice, the disease risk index is recalculated based on the optimized factor weights to obtain the adjusted disease risk index. The method for recalculating the disease risk index is as follows: maintain the derived node state values ​​of each node at the current moment. The weighting coefficients in the weighted summation formula remain unchanged. Replace the optimized factor weights and perform a weighted summation calculation again. The result is the adjusted disease risk index.

[0055] In practice, the adjusted disease risk index is mapped to the disease severity level to obtain the corrected disease risk index. The method for intensity mapping is as follows: obtain the disease severity level corresponding to similar historical cases, and determine an intensity mapping coefficient based on the disease severity level. When the disease intensity level is 1, the intensity mapping coefficient is... The value is 1.0; when the disease intensity level is 2, the intensity mapping coefficient is... The value is 1.3; when the disease intensity level is 3, the intensity mapping coefficient is... The value is set to 1.6. The intensity mapping coefficient is set based on the ratio of mortality rates of farmed organisms under different disease intensity levels. Using the median mortality rate of 2.5% for level 1 diseases as a benchmark, the median mortality rate of 10% for level 2 diseases is approximately four times the benchmark, and the median mortality rate of 30% for level 3 diseases is approximately twelve times the benchmark. After cubic compression, the coefficient is mapped to the amplification factor of the risk index, ensuring that the mapped corrected risk index corresponds reasonably in magnitude to the severity of the disease consequences. The adjusted disease risk index is then multiplied by the intensity mapping coefficient. The product obtained is the corrected disease risk index. If the product result is greater than 1.0, the corrected disease risk index is truncated to 1.0; if the product result is less than 0, the corrected disease risk index is set to 0, ensuring that the corrected disease risk index takes the value within the closed interval of 0 to 1.

[0056] See Figure 5 In the graph, the horizontal axis represents time in hours, and the vertical axis represents the disease risk index, ranging from 0 to 1. The legend shows three curves: the "original disease risk index (online prediction)" is represented by a solid blue line, the "adjusted disease risk index (weight optimization)" is represented by a dashed green line, and the "corrected disease risk index (intensity mapping)" is represented by a dotted red line.

[0057] The curves show that the blue original disease risk index and the green adjusted risk index curves are generally consistent and closely overlap. This indicates that after weight optimization, the risk index values ​​are highly consistent with the online prediction results in terms of overall level and trend. This reflects the fine-tuning effect of weight optimization on the risk index and enhances the accuracy and stability of the model.

[0058] The red, corrected disease risk index curve is generally above the blue and green curves, with values ​​typically higher than the other two risk indices and greater fluctuations. Within timeframes of approximately 0–100 hours, 190–280 hours, and 380–480 hours, the red curve repeatedly reaches or approaches the maximum value of 1 on the ordinate, indicating that after intensity mapping, the risk index shows a higher risk of disease outbreaks during these periods, reflecting the amplifying effect of disease intensity levels on the risk index.

[0059] Between approximately 100 and 190 hours and between 280 and 380 hours, the red curve, although declining somewhat, remains higher than the blue and green curves and exhibits significant peak-and-trough fluctuations, indicating that the revised risk index can capture dynamic changes in disease risk and respond more sensitively.

[0060] In practical implementation, the comprehensive generation module receives the corrected disease risk index output by the adaptive update module. This corrected disease risk index is a real number within a closed interval of 0 to 1. The comprehensive generation module first obtains the corrected disease risk index, then performs hierarchical encoding on it to generate a disease early warning signal. The hierarchical encoding process includes: pre-setting a set of continuous numerical intervals as hierarchical thresholds, mapping the intervals into which the corrected disease risk index falls to a discrete early warning level, retrieving the corresponding prevention and control suggestion text from a pre-configured prevention and control suggestion database based on the early warning level, and combining the early warning level and the prevention and control suggestion into a structured disease early warning signal for output.

[0061] In practice, the grading thresholds are set at three critical points, dividing the range of the corrected disease risk index into four continuous intervals. The values ​​of the three critical points are 0.25, 0.50, and 0.75, respectively. When the corrected disease risk index is greater than or equal to 0 and less than 0.25, it is classified into the first warning level, defined as a green warning, indicating that the disease risk in the aquaculture area is at a low level, and no immediate intervention is required, but it is recommended to maintain the daily monitoring frequency. When the corrected disease risk index is greater than or equal to 0.25 and less than 0.50, it is classified into the second warning level, defined as a yellow warning, indicating that the disease risk in the aquaculture area is at a moderately low level, and it is recommended that aquaculture managers increase the frequency of inspections, focusing on the changing trends of dissolved oxygen and ammonia nitrogen concentrations in the water, and checking the operating status of aeration equipment. When the revised disease risk index is greater than or equal to 0.50 and less than 0.75, it is classified as Level 3, which is defined as an orange alert. This indicates that the disease risk in the aquaculture area is at a high level. It is recommended that aquaculture managers immediately activate the emergency monitoring plan, increase the water quality sampling frequency from once a day to once every four hours, prepare disinfectants and anti-stress agents, reduce feed intake to 70% of normal levels, and turn on backup aeration equipment to increase dissolved oxygen levels in the water. When the revised disease risk index is greater than or equal to 0.75 and less than or equal to 1.00, it is classified as Level 4, which is defined as a red alert. This indicates that the disease risk in the aquaculture area is at an extremely high level. It is recommended that aquaculture managers immediately implement the emergency response plan for diseases, including but not limited to: spraying chlorine dioxide disinfectant throughout the pond at a dosage of 200 grams per acre of water surface; stopping feed feeding; turning on all aeration equipment and using sodium percarbonate chemical oxygenator at a dosage of 500 grams per acre of water surface; notifying surrounding aquaculture farmers to take isolation and protective measures; and contacting the local aquatic technology extension station for on-site technical support.

[0062] In practical implementation, the specific operation process of hierarchical coding is as follows: The comprehensive generation module maintains a hierarchical mapping table, which is stored in key-value pair format. The key is a numerical range object, each of which is defined by a lower limit and an upper limit. The ranges are left-closed and right-open, with the last range being left-closed and right-closed. The value is a structured record containing two fields: a warning level field and a prevention and control recommendation field. The warning level field stores an enumeration value, including green, yellow, orange, and red warnings. The prevention and control recommendation field stores a pre-generated text string. The comprehensive generation module takes the corrected disease risk index as input and compares it sequentially with the lower and upper limits of each numerical range object in the hierarchical mapping table. It finds the first numerical range object to which the corrected disease risk index belongs, retrieves the corresponding structured record, and extracts the warning level field and prevention and control recommendation field from the structured record. The extracted warning level and prevention and control recommendation are concatenated into a string according to a predetermined format, which is the disease warning signal. The predetermined format is: first, write the name of the warning level, followed by a colon and a space, and then write the prevention and control recommendation text. For example, if the revised disease risk index is 0.68, the classification code result is "Orange Alert: The disease risk in the aquaculture water area is at a high level. It is recommended that aquaculture managers immediately activate the emergency monitoring plan, increase the water quality sampling frequency from once a day to once every four hours, prepare disinfectants and anti-stress agents, reduce the amount of feed to 70% of the normal level, and turn on the backup aeration equipment to increase the dissolved oxygen in the water." In practice, after the integrated generation module generates a disease early warning signal, it pushes the signal to the front-end display interface of the aquaculture management platform. The front-end display interface presents the disease early warning signal in two forms: pop-up messages and audible and visual alarms. Pop-up messages appear on the monitoring screen of the aquaculture management platform, displaying a color icon corresponding to the warning level and complete prevention and control recommendations in the center of the pop-up. The audible and visual alarm device triggers different alarm modes according to the warning level: no audible and visual alarm is triggered for a green warning; a low-frequency flashing yellow indicator light is triggered for a yellow warning; a medium-frequency flashing orange indicator light is triggered for an orange warning, accompanied by intermittent buzzing; and a high-frequency flashing red indicator light is triggered for a red warning, accompanied by a continuous siren. Upon receiving the disease early warning signal, aquaculture managers can implement corresponding management measures according to the operation guidelines in the prevention and control recommendations, thereby achieving early intervention and effective control of aquaculture diseases.

[0063] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for predicting aquatic diseases based on the Internet of Things, characterized in that, include: The data multi-source association module acquires multi-source data collected by multiple IoT sensors in the aquaculture area, performs multi-dimensional time series alignment on the multi-source data, and obtains the aligned multi-source time series dataset. The big data analysis module obtains the aligned multi-source time series dataset output by the data multi-source association module, performs time series coupling analysis on the aligned multi-source time series dataset, and obtains the disease inducing factor association map. The online prediction module obtains the disease-inducing factor correlation map output by the big data analysis module, performs dynamic evolution analysis on the disease-inducing factor correlation map, and obtains the disease risk index at the current moment. The offline matching module obtains a historical disease case library, extracts features from the historical multi-source time series data corresponding to each historical case in the historical disease case library to obtain a historical feature library, and performs similarity matching between the disease risk index output by the online prediction module and the historical features in the historical feature library to obtain similar historical cases. The adaptive update module obtains the disease confirmation results corresponding to similar historical cases output by the offline matching module, and corrects the disease risk index according to the disease confirmation results to obtain the corrected disease risk index. The comprehensive generation module obtains the corrected disease risk index output by the adaptive update module, performs hierarchical encoding on the corrected disease risk index, and generates a disease early warning signal.

2. The aquaculture disease prediction method based on the Internet of Things as described in claim 1, characterized in that, The multi-source data association module performs multi-dimensional time series alignment on the multi-source data to obtain an aligned multi-source time series dataset, including: The multi-source data includes: real-time water quality data, meteorological data, feeding data, and fish behavior data; Obtain the data acquisition timestamp of each IoT sensor, establish a time reference coordinate system for each sensor based on the data acquisition timestamp, and map the multi-source data onto the time reference coordinate system to obtain the time series points of each data. Linear interpolation is performed on the time series points of each data point to obtain continuous time series curves for each data point; The data length, data frequency, and data sampling interval of each continuous time series curve are uniformized to obtain the uniformized time series of each data. The time series data after homogenization are time-stamped to generate an aligned multi-source time series dataset.

3. The aquaculture disease prediction method based on the Internet of Things as described in claim 2, characterized in that, The multi-source data association module performs time-stamp alignment on the normalized data time series to generate an aligned multi-source time series dataset, including: Calculate the correlation degree between any two data time series in each of the normalized data time series to obtain the correlation degree matrix; Based on the correlation matrix, several data time series with a correlation greater than a preset correlation threshold are selected from the normalized data time series, and these several data time series are used as a set of strongly correlated sequences. Principal component analysis is performed on the strongly correlated sequence set to extract the principal component sequences, which are then used as an aligned multi-source time series dataset.

4. The aquaculture disease prediction method based on the Internet of Things as described in claim 3, characterized in that, The multi-source data correlation module calculates the correlation degree between any two data time series in each of the normalized data time series, obtaining a correlation degree matrix, including: The time series of the homogenized data are divided into multiple time window segments by a sliding window. Within each time window segment, calculate the Pearson correlation coefficient between any two data time series to obtain the in-window correlation coefficient; The correlation between any two data time series is obtained by taking a weighted average of the correlation coefficients within all time window segments. Arrange all the aforementioned correlations according to the categories of the data time series to construct a correlation matrix.

5. The aquaculture disease prediction method based on the Internet of Things as described in claim 1, characterized in that, The big data analysis module performs time-series coupling analysis on the aligned multi-source time-series dataset to obtain a correlation map of disease-inducing factors, including: The time series of water quality factors, meteorological factors, feed factors, and biological activity factors are separated from the aligned multi-source time series dataset. Multivariate time series regulation analysis was performed on the time series of water quality factors, meteorological factors, feed factors, and biological activity factors to obtain the cross-coupling coefficients between the factors. Based on the cross-coupling coefficient, a temporal coupling network is constructed with each factor as a node and the cross-coupling coefficient as the edge weight. For each node in the temporally coupled network, key node identification is performed, and the node with the maximum out-degree, maximum in-degree, and maximum betweenness centrality is extracted as the disease inducing factor node. By combining all disease-inducing factor nodes and the edges between them, a disease-inducing factor association graph is obtained.

6. The aquaculture disease prediction method based on the Internet of Things as described in claim 5, characterized in that, The big data analysis module performs multivariate time-series regulation analysis on the time series of water quality factors, meteorological factors, feed factors, and biological activity factors to obtain the cross-coupling coefficients between each factor, including: The time series of water quality factors, meteorological factors, feed factors, and biological activity factors are combined in pairs to obtain multiple factor pairs. For each factor pair, perform lag correlation analysis on the time series of the two factors, calculate the correlation value under different lag orders, and take the lag order corresponding to the maximum correlation value as the lag time of the factor pair. Using the lag time as a time offset, the two factor time series are time-aligned to obtain the aligned factor pair time series. Calculate the mutual information value of the aligned factor pair time series, and use the mutual information value as the cross-coupling coefficient of the factor pair.

7. The aquaculture disease prediction method based on the Internet of Things as described in claim 1, characterized in that, The online prediction module performs dynamic evolution analysis on the correlation map of disease-inducing factors to obtain the disease risk index at the current moment, including: Obtain the topological structure of the disease-inducing factor association map, and determine the predecessor node set and successor node set of each node in the map based on the topological structure; Aggregate the states of the predecessor node set for each node to obtain the aggregated predecessor state vector for each node; The covariance analysis is performed on the aggregated predecessor state vector and the input real-time data of each current node to obtain the state transmission probability of each node. Based on the state propagation probability, a time-series propagation deduction is performed on the disease inducing factor correlation map to obtain the deduced node state values. The disease risk index at the current moment is obtained by summing the weighted values ​​of all the derived node states.

8. The aquaculture disease prediction method based on the Internet of Things as described in claim 1, characterized in that, The offline matching module performs similarity matching between the disease risk index and historical features in the historical feature database to obtain similar historical cases, including: The disease risk index is feature-encoded to obtain the risk feature vector to be matched; Calculate the Euclidean distance and cosine similarity between the risk feature vector to be matched and each historical feature vector in the historical feature database. Standardize the Euclidean distance to obtain the distance similarity. Normalize the cosine similarity to obtain the directional similarity. The distance similarity and direction similarity are weighted and fused to obtain a comprehensive similarity. The historical case corresponding to the historical feature vector with the highest comprehensive similarity in the historical feature library is taken as a similar historical case.

9. The method for predicting aquaculture diseases based on the Internet of Things as described in claim 1, characterized in that, The adaptive update module corrects the disease risk index based on the disease confirmation result, obtaining a corrected disease risk index, including: Obtain the disease confirmation results corresponding to the similar historical cases, including the disease outbreak confirmation identifier and the disease intensity level; Based on the disease outbreak confirmation marker, the matching accuracy between the disease risk index and the actual disease is calculated to obtain the accuracy assessment value. Based on the accuracy assessment value, the calculation parameters of the disease risk index are adjusted to obtain the adjusted disease risk index; Based on the disease intensity level, the adjusted disease risk index is intensity-mapped to obtain the corrected disease risk index.

10. The method for predicting aquaculture diseases based on the Internet of Things as described in claim 9, characterized in that, The adaptive update module adjusts the calculation parameters of the disease risk index based on the accuracy assessment value to obtain the adjusted disease risk index, including: The accuracy evaluation value is compared with a preset accuracy threshold. If the accuracy evaluation value is lower than the preset accuracy threshold, the parameter adjustment process is initiated. In the parameter adjustment process, sensitivity analysis is performed on the weights of each factor in the calculation model of the disease risk index to determine the weight sensitivity of each factor. Based on the weight sensitivity, the weights of each factor are optimized by gradient descent to obtain the optimized weights of each factor. The disease risk index is recalculated based on the optimized factor weights to obtain the adjusted disease risk index.

Images