Early warning method for disease of carassius auratus based on factory farming in chuzhou

Abstract

The application discloses a kind of early warning method of disease of Chuzhou crucian based on factory farming, belong to aquaculture disease monitoring technical field.The method includes: obtaining the group behavior image sequence of Chuzhou crucian in multiple monitoring points in culture pond and corresponding culture water physicochemical parameter time series data;Group spatial distribution characteristics and individual motion characteristics are extracted from group behavior image sequence and behavior trend atlas representing group stress state is constructed;The behavior trend atlas is spatiotemporally associated with the water physicochemical parameter time series data to generate a multi-modal associated feature set;Disease outbreak risk probability at the current time is inferred based on the multi-modal associated feature set through a pre-constructed disease risk probability graph model, and an early warning signal is issued when the risk probability exceeds the warning threshold.The method fuses the spatiotemporal correlation characteristics of group behavior dynamics and water quality changes, and infers the risk probability through the graph model, which can early warn the disease of Chuzhou crucian bacterial hemorrhagic disease or parasitic disease.

Description

Technical Field

[0001] This invention relates to the field of aquaculture disease monitoring technology, specifically to an early warning method for diseases of Chuzhou crucian carp based on factory farming. Background Technology

[0002] In factory farming, Chuzhou crucian carp are raised in high-density, enclosed or semi-enclosed aquatic environments. Once a disease outbreak occurs, it spreads rapidly and over a wide area, placing high demands on the timeliness and accuracy of early warning systems. Existing disease warning technologies largely rely on threshold judgments of single water quality parameters or manual observation, assessing disease risk by monitoring abnormal changes in physicochemical indicators such as dissolved oxygen and ammonia nitrogen, or by visually observing fish symptoms. Fluctuations in single water quality parameters are affected by multiple factors such as feeding, water changes, and weather, and their changes often lag behind the fish's physiological stress response. Furthermore, the boundary between normal fluctuations and disease precursors is blurred, easily leading to false alarms or missed alarms. Manual observation relies on the observer's experience and is difficult to achieve 24 / 7 coverage, lacking the ability to identify diseases in their incubation period or early stress stages.

[0003] The shortcomings of existing technical solutions are mainly reflected in two aspects. Firstly, the correlation between group behavior dynamics and water quality changes has not been effectively extracted and utilized. In the early stages of pathogen stress, Chuzhou crucian carp exhibit abnormal swimming speed, altered group behavior, and a tendency to move towards the pond walls or corners. These behavioral changes have a specific temporal lag relationship with minor fluctuations in local water physicochemical indicators. Existing technologies treat behavioral observation and water quality monitoring as two isolated dimensions, severing the spatiotemporal causal relationship between them and failing to characterize the precursory patterns of disease occurrence from the perspective of the interaction between behavior and water quality. Secondly, environmental pressures and disease risks in different areas of the aquaculture pond exhibit spatial heterogeneity. Diseases typically start in a localized area and spread to other areas along the water flow direction or through fish contact relationships. Existing methods of overall pond water sampling or single monitoring point deployment cannot capture the spatial distribution differences and spread patterns of disease risk, lacking the ability to perform spatial correlation modeling and probabilistic inference of information from multiple monitoring points. This makes it difficult to accurately identify the source of risk and issue early warnings in the early stages before the disease spreads widely.

[0004] The multimodal correlation feature set formed by mapping behavioral behavior patterns to the spatiotemporal correlation of water quality parameters, and the disease risk probability graphical model based on the spatial topology graph structure to perform confidence propagation inference, are two key paths to solving the above problems. How to align changes in group behavior with fluctuations in physicochemical factors over time, extract the response delay features of behavior to water quality changes and encode them into computable correlation patterns, and how to integrate local anomalies from multiple monitoring points into global risk probability inference through spatial state transmission relationships are urgent technical problems to be solved. Summary of the Invention

[0005] The purpose of this invention is to provide an early warning method for diseases in Chuzhou crucian carp based on factory farming. This method involves simultaneously acquiring images of group behavior and water physicochemical parameters from multiple monitoring points deployed in a spatial grid. A behavioral pattern map is constructed from the spatial distribution characteristics of the group and individual movement characteristics. This map is then spatiotemporally correlated with water quality parameter time-series data to extract a multimodal correlation feature set between behavior and water quality. Finally, a disease risk probability graph model based on a spatial topology is used to infer the probability of disease outbreaks, issuing early warning signals for risk states exceeding the warning threshold. This method aims to capture the spatiotemporal coupling relationship between abnormal group behavior and subtle changes in the aquatic environment in the early stages of disease occurrence, and to perform probability propagation and fusion of risk states from multiple monitoring points in a spatial dimension, thereby achieving accurate early warning and risk source location before the disease spreads on a large scale.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides an early warning method for diseases in Chuzhou crucian carp based on factory farming. This method achieves early inference and warning of disease outbreak risk by performing multimodal spatiotemporal correlation analysis on the group behavior of Chuzhou crucian carp and the physicochemical parameters of the aquaculture water. The method first acquires image sequences of group behavior of Chuzhou crucian carp at multiple monitoring points in the aquaculture pond, along with time-series data of the physicochemical parameters of the aquaculture water aligned with these sequences. Then, it extracts the spatial distribution characteristics of the group and individual movement characteristics from the image sequences to construct a behavioral situation map representing the group's stress state. Next, it performs spatiotemporal correlation mapping between the behavioral situation map and the time-series data of the physicochemical parameters of the aquaculture water to generate a multimodal correlation feature set integrating behavior and water quality. Finally, it infers the probability of disease outbreak risk at the current moment based on a pre-constructed disease risk probability graph model and issues an early warning signal when the risk probability exceeds the warning threshold. This method jointly models changes in fish behavior with responses to aquatic environmental factors within a unified probabilistic graphical model framework. It can capture the spatial heterogeneity and temporal evolution trend of abnormal fish behavior before disease outbreaks, significantly improving the timeliness and accuracy of early warning.

[0007] As a preferred technical solution of the present invention, the process of acquiring the group behavior image sequence of Chuzhou crucian carp at multiple monitoring points in the aquaculture pond and the corresponding time-series data of the physicochemical parameters of the aquaculture water is as follows: Multiple underwater cameras are deployed in a spatial grid layout within the aquaculture pond, simultaneously acquiring swimming images of Chuzhou crucian carp within the coverage area of ​​each underwater camera, forming a multi-view group behavior image sequence covering the entire pond; simultaneously, water quality sensors are installed near each underwater camera, continuously acquiring dissolved oxygen concentration, pH, temperature, and ammonia nitrogen concentration of the aquaculture water at the same sampling frequency as the group behavior image sequence, generating time-series data of the physicochemical parameters of the aquaculture water aligned with the time of the group behavior image sequence. This setup establishes a spatial correspondence between the behavioral data and the water quality data, laying the foundation for subsequent spatiotemporal correlation mapping.

[0008] Furthermore, the process of extracting the spatial distribution features and individual movement features of Chuzhou crucian carp from the group behavior image sequence and constructing a behavioral situation map is as follows: A target detection algorithm is executed on each frame of the group behavior image sequence to identify and locate the individual position and outline of each Chuzhou crucian carp in the image, generating an individual distribution coordinate set within each monitoring point; based on this individual distribution coordinate set, the population density and spatial dispersion of Chuzhou crucian carp within each monitoring point are statistically analyzed as the group spatial distribution features; the positional offset of the same Chuzhou crucian carp individual in adjacent frames is tracked, and the instantaneous swimming speed and turning frequency of each Chuzhou crucian carp are calculated as individual movement features; the group spatial distribution features and individual movement features of each monitoring point at the same time are arranged according to the monitoring point encoding to form a behavior vector reflecting spatial heterogeneity; the behavior vectors of consecutive moments are concatenated along the time axis to construct a behavioral situation map, which contains the evolutionary trajectory of the Chuzhou crucian carp group behavior in the time dimension at each monitoring point. The behavioral pattern map constructed in this way can depict the spatial distribution pattern of fish schools and reflect the dynamic changes in the movement state of individuals.

[0009] Based on this, the process of generating a multimodal correlation feature set integrating behavior and water quality by performing spatiotemporal correlation mapping between the behavior situation map and the time series data of physicochemical parameters of aquaculture water bodies is as follows: The behavior vectors of each monitoring point in the behavior situation map are paired with the corresponding time and physicochemical parameter values ​​of the corresponding monitoring points in the time series data of physicochemical parameters of aquaculture water bodies at the same timestamp, forming a behavior-water quality paired sample for each monitoring point at each time point; Time window sliding aggregation is performed on the behavior-water quality paired sample to calculate the mean change rate of the behavior vector and the mean change rate of the physicochemical parameter values ​​of each monitoring point within the preset time window, generating behavior change features and water quality change features reflecting short-term trends; The behavior change features and water quality change features of the same monitoring point are cross-correlated to calculate the response delay time of behavior changes to water quality changes, generating behavior-water quality response lag features; Finally, the behavior-water quality paired samples, behavior change features, water quality change features, and behavior-water quality response lag features of each monitoring point are uniformly encoded to form a multimodal correlation feature set. This feature set effectively integrates behavioral dynamics, water quality changes, and their interaction time delays, providing high-dimensional time-series feature inputs for disease risk inference.

[0010] When inferring the probability of disease outbreak risk at the current moment based on a pre-constructed disease risk probability graph model using a multimodal association feature set, the preferred approach is as follows: Each monitoring point in the multimodal association feature set is considered a node in the graph model, and the spatial adjacency between monitoring points is considered an edge in the graph model, thus constructing a spatial topology graph structure. A local anomaly score is calculated for each node based on the corresponding behavior-water quality paired samples. This local anomaly score quantifies the deviation of the behavior and water quality characteristics at that node from the historical baseline level. The state transmission probability between adjacent nodes is calculated using the behavior-water quality response lag features. This state transmission probability represents the likelihood that an abnormal state of a node will propagate along the spatial edge to adjacent nodes. Using the local anomaly score as observational evidence and the state transmission probability as a state transition constraint, confidence propagation inference is performed on the spatial topology graph structure to iteratively update the posterior probability of disease outbreak for each node. The largest posterior probability of disease outbreak among all nodes is taken as the probability of disease outbreak risk at the current moment. This graph-based reasoning method can utilize spatial adjacency relationships to transmit and integrate abnormal states, effectively identify the spatial spread trend of diseases within the aquaculture pond, and avoid misjudgments caused by fluctuations in single-point observations.

[0011] As a further preferred embodiment of the present invention, an early warning signal is issued when the probability of disease outbreak exceeds a warning threshold. Specifically, this includes: comparing the current probability of disease outbreak with a preset risk probability warning threshold; if the current probability of disease outbreak is greater than or equal to the risk probability warning threshold, then the warning triggering condition is met; after the warning triggering condition is met, the posterior probability of disease outbreak at each node in the spatial topology is further traced back to locate the monitoring point corresponding to the node with the highest posterior probability of disease outbreak as a suspected disease source; based on the spatial location of the suspected disease source in the spatial topology, behavioral change features and water quality change features of the suspected disease source at multiple consecutive times before the triggering time are extracted from the multimodal association feature set to generate a trend evidence chain indicating the direction of disease spread; and an early warning signal containing the location of the suspected disease source and the trend evidence chain is output. Thus, aquaculture managers can not only obtain disease risk warnings but also grasp possible disease sources and spread paths, facilitating timely and targeted intervention measures.

[0012] For the display of behavioral situation maps, it is preferable to use sliding window principal component analysis to reduce the dimensionality of the behavioral situation map over time to a two-dimensional or three-dimensional manifold space, and then display it simultaneously on the aquaculture monitoring terminal in a visual manner. This provides aquaculture managers with an intuitive trend map of the evolution of fish stress status, making it easier to quickly determine the overall health status of the fish population in the pond.

[0013] The preferred implementation method for pre-constructing the disease risk probability graphical model is as follows: Collect multiple sets of historical behavioral image sequences and historical water body physicochemical parameter time-series data within the historical health cycle of Chuzhou crucian carp in a factory farming environment, and simultaneously record whether disease events occurred during the corresponding time periods, forming a historical training dataset labeled with disease. Extract historical multimodal correlation feature sets from this historical training dataset using the same method described above, and use the disease outbreak state at each historical moment as a latent variable and the historical multimodal correlation feature set as an observed variable. Based on the conditional dependency relationship between the observed variable and the latent variable, construct a disease risk probability graphical model that includes the prior distribution of spatial node states, the state transition matrix between nodes, and the observation likelihood function. Use the historical training dataset to perform maximum likelihood estimation on the model parameters of the disease risk probability graphical model. The model parameters include the prior probability of each node's state, the state transition probability between adjacent nodes, and the conditional probability distribution of each observed feature under a given state. Use the disease risk probability graphical model after parameter estimation as the pre-constructed disease risk probability graphical model. The model, trained on a large amount of historical data, can accurately depict the statistical patterns of the normal state and the state before disease outbreak in the aquaculture system, providing a reliable basis for real-time reasoning.

[0014] To adapt to baseline drift caused by seasonal and batch-specific changes in the aquaculture environment, a dynamic update method is preferred for determining the early warning threshold. During normal aquaculture phases without disease events, multimodal correlation feature sets of normal states are continuously collected across multiple normal aquaculture cycles. The probability of disease outbreak at each moment within each normal aquaculture cycle is calculated using a disease risk probability graph model, forming a normal risk probability distribution. The upper limit of this distribution is calculated and used as the baseline early warning threshold. In subsequent aquaculture processes, after each disease-free cycle, the actual multimodal correlation feature set of that cycle is added to the normal sample database, and the updated normal risk probability distribution and its upper limit are recalculated to update the baseline early warning threshold. The updated baseline early warning threshold is compared with a preset minimum safety threshold, and the larger of these two values ​​is taken as the current actual early warning threshold. This method allows the early warning threshold to adaptively adjust with normal fluctuations in the aquaculture system, effectively reducing false alarm rates while maintaining high sensitivity.

[0015] This invention is particularly suitable for early warning of common bacterial hemorrhagic diseases or parasitic diseases in Chuzhou crucian carp farming. The preferred range for the warning threshold is 0.7 to 0.9, and the probability of disease outbreak is a dimensionless probability value between 0 and 1.

[0016] In a preferred embodiment of the present invention, after issuing an early warning signal, a confidence enhancement verification of the early warning is performed. Specifically, after issuing the early warning signal, the system continues to acquire verification behavior image sequences and time-series data of verification water body physicochemical parameters for multiple consecutive time periods. Based on the verification behavior image sequences, the system extracts the spatial distribution characteristics of the verification group and the movement characteristics of the verification individuals to construct a verification behavior situation map. The system performs a spatiotemporal correlation mapping between the verification behavior situation map and the time-series data of verification water body physicochemical parameters to generate a verification multimodal correlation feature set. The verification multimodal correlation feature set is input into the disease risk probability graph model to calculate the verification disease outbreak risk probability for each subsequent time period. If the verification disease outbreak risk probability maintains an upward trend for multiple consecutive time periods and the verification disease outbreak risk probability at each time period is greater than the warning threshold, the early warning signal is confirmed to be effective and the warning output state is maintained. If the verification disease outbreak risk probability drops below the warning threshold within a preset verification time window, the early warning signal is revoked and marked as a false alarm event. The multimodal correlation feature set corresponding to the false alarm event is fed back as a negative sample to the model update queue. Through the continuous verification mechanism after the warning, the authenticity of the initial warning signal can be confirmed a second time, effectively filtering out false alarms caused by momentary interference.

[0017] Furthermore, after feeding the multimodal correlation feature set corresponding to false alarm events as negative samples to the model update queue, incremental model updates are preferentially performed: multiple accumulated negative samples and normal samples confirmed to be free of disease events in the most recent time period are extracted from the model update queue and merged to form an incremental training sample set; for each sample in the incremental training sample set, the prediction error is calculated using the current observation likelihood function of the disease risk probability graph model, and an incremental loss function is constructed based on the prediction error; with the goal of minimizing the incremental loss function, the model parameters of the disease risk probability graph model are incrementally adjusted using an online gradient descent method. The model parameters include the prior state probability of each node, the state transition probability between adjacent nodes, and the conditional probability distribution of each observed feature under a given state; after the incremental adjustment is completed, the updated model parameters replace the original model parameters, and the model update queue is cleared, completing the online incremental update of the disease risk probability graph model. This incremental learning mechanism enables the graph model to continuously learn from false alarms and normal samples, gradually improving its ability to distinguish between non-disease abnormal fluctuations and true disease precursors, achieving continuous optimization of the early warning system in actual operation.

[0018] Through the above technical solution, this invention deeply integrates the spatiotemporal evolution of Chuzhou crucian carp behavior with the dynamic changes of aquatic environmental factors in a factory farming environment. By utilizing the spatial reasoning ability of graph models, it captures the early spatial aggregation and spread characteristics of disease risks. Without relying on traditional pathological detection methods, it can detect the precursor signals of disease outbreaks in advance and dynamically locate suspected disease sources by relying solely on the joint monitoring of fish behavior and environmental parameters, thus gaining critical response time for disease prevention and control in factory farming.

[0019] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0020] By performing spatiotemporal correlation mapping between behavioral behavior maps and time-series data of physicochemical parameters in aquaculture water bodies, and generating a multimodal correlation feature set, a deep coupling between group behavioral dynamics and water quality changes in both time and space was achieved. Behavioral vectors at each monitoring point were paired with corresponding physicochemical parameter values ​​at the same timestamp. The mean change rate of both the behavioral vectors and the physicochemical parameter values ​​was calculated using time window sliding aggregation, generating behavioral change features and water quality change features. Based on this, the behavioral change features and water quality change features at the same monitoring point were cross-correlated to calculate the response delay time of behavioral changes to water quality changes, generating behavior-water quality response lag features. The introduction of this response lag feature changes the traditional paradigm of treating behavioral observation and water quality detection as independent information sources. It enables the system to capture the inherent pattern that fish behavioral responses under disease stress precede significant water quality deterioration from the temporal relationship between behavior and water quality interaction. Thus, potential risks can be identified in the early stage when physicochemical parameters have not yet shown significant abnormalities but behavior has already shown stress characteristics. Compared with the method of relying solely on a single water quality parameter threshold or isolated behavioral indicators, it can significantly advance the warning time window and reduce the false alarm rate for water quality fluctuations caused by non-disease factors.

[0021] Based on a multimodal association feature set, a pre-constructed disease risk probability graph model is used to infer the probability of a disease outbreak at the current moment. A spatial topology graph is constructed using monitoring points as graph nodes and spatial adjacency relationships as graph edges. The state transfer probability between adjacent nodes is calculated using behavior-water quality response lag features. Local anomaly scores of nodes are used as observational evidence, and confidence propagation inference is performed on the spatial topology graph to iteratively update the posterior probability of disease outbreak for each node. This probability inference mechanism connects independent monitoring information from different areas within the aquaculture pond into a holistic probability network through a spatial state transfer model. It can characterize the dynamic process of disease risk spreading from a local area to neighboring areas. The global risk level can be inferred in the early stages when the disease is still at a local point and has not yet spread to the entire pond. Furthermore, suspected disease sources can be directly located based on the spatial distribution of the posterior probability of disease outbreak at each node. Compared to relying solely on average water quality indicators across the entire pond or fish behavior at a single monitoring point for overall assessment, this approach can identify the direction of risk transmission and track the chain of trend evidence in a spatial dimension, providing directional information for timely and precise intervention measures such as isolation and localized medication. Attached Figure Description

[0022] 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.

[0023] Figure 1 This is a flowchart of an early warning method for diseases in Chuzhou crucian carp based on factory farming.

[0024] Figure 2 This is a flowchart of the synchronous acquisition process of images of the behavior of crucian carp in Chuzhou and the physicochemical parameters of the aquaculture water.

[0025] Figure 3 This is a flowchart for constructing a behavioral situation map of the stress state of crucian carp in Chuzhou;

[0026] Figure 4 This is a flowchart of the process for generating behavior-water quality multimodal correlation feature sets;

[0027] Figure 5 This is a flowchart for inferring disease outbreak risk and providing early warning based on multimodal correlation features and probabilistic graphical models;

[0028] Figure 6 This is a curve showing the spatial distribution characteristics of the Chuzhou crucian carp population over time.

[0029] Figure 7 This is the evolutionary trajectory of the behavior pattern of crucian carp in Chuzhou using two-dimensional principal component analysis. Detailed Implementation

[0030] 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.

[0031] See Figure 1 This invention provides an early warning method for diseases in Chuzhou crucian carp based on factory farming. This method uses Chuzhou crucian carp in a factory farming environment as the object, and simultaneously acquires a series of group behavior images of Chuzhou crucian carp at multiple monitoring points within the farming pond, along with time-series data of the physicochemical parameters of the farming water corresponding to these group behavior image sequences, to achieve synchronous perception of the farming status. After obtaining the group behavior image sequences, the spatial distribution characteristics and individual movement characteristics of the Chuzhou crucian carp are extracted. The spatial distribution characteristics reflect the aggregation or dispersion of the fish population at each monitoring point, while the individual movement characteristics reflect the fish's behavior in dimensions such as instantaneous swimming speed and turning frequency. Then, the spatial distribution characteristics and individual movement characteristics are used together to construct a behavioral situation map characterizing the stress state of the Chuzhou crucian carp population. This behavioral situation map retains the behavioral heterogeneity information of each monitoring point in the spatial dimension and continues to depict the dynamic process of behavioral evolution in the temporal dimension. After constructing the behavioral situation map, it is spatiotemporally correlated and mapped with the time-series data of physicochemical parameters of aquaculture water. Under the dual constraints of timestamps and monitoring points, behavioral vectors are paired with physicochemical parameter values ​​and features are aggregated to generate a multimodal correlation feature set that integrates behavior and water quality. Based on this multimodal correlation feature set, a pre-constructed disease risk probability graph model is used to infer the probability of disease outbreak at the current moment. When the inferred probability of disease outbreak exceeds the warning threshold, an early warning signal is issued.

[0032] Example 1:

[0033] In specific implementation, please refer to Figure 2 The process of acquiring group behavior image sequences of Chuzhou crucian carp at multiple monitoring points in the aquaculture pond, as well as the time series data of the physicochemical parameters of the aquaculture water corresponding to the group behavior image sequences, is achieved through the following methods.

[0034] When deploying underwater cameras in the aquaculture pond, a spatial grid layout is pre-defined based on the pond's wall and bottom geometry. This grid divides the pond's water area into several equally or unequally spaced grid units on a horizontal plane. The intersections or centers of these grid units serve as the installation points for the underwater cameras, ensuring that the fields of view of each camera are interconnected across the entire pond without any blind spots. During deployment, the lens optical axis of each underwater camera can be adjusted for pitch and azimuth angles according to the pond's bottom topography and depth at the monitoring point, ensuring the integrity and identifiability of the Chuzhou crucian carp swimming images within the coverage area. All underwater cameras are equipped with a unified internal clock source or receive the same external clock synchronization signal, synchronously acquiring Chuzhou crucian carp swimming images within their respective coverage areas. Each underwater camera continuously captures images at a fixed frame rate, set according to the typical swimming speed of the Chuzhou crucian carp, for example, 25 frames per second or 30 frames per second. Multiple video streams, simultaneously acquired by various underwater cameras, are aligned frame by frame in time to form a multi-view group behavior image sequence covering the entire pool. Each frame in the multi-view group behavior image sequence contains image information from multiple monitoring points, and the images from each monitoring point can be spatially stitched together or used independently for subsequent processing.

[0035] When installing water quality sensors near underwater cameras, the sensors are fixed at a known spatial offset relative to the cameras, ensuring that each sensor belongs to the same monitoring point within the spatial grid layout as the neighboring cameras. The water quality sensors include dissolved oxygen, pH, temperature, and ammonia nitrogen sensors, each type integrated as an independent probe or a composite probe. The sensors are configured with the same sampling trigger period as the underwater cameras, and the sampling frequency matches the frame rate of the group behavior image sequence. That is, the number of times water quality parameters are collected per second is the same as the number of frames captured per second by the underwater camera. When the underwater camera captures images at 25 frames per second, the water quality sensors also collect dissolved oxygen concentration, pH, temperature, and ammonia nitrogen concentration values ​​at a frequency of 25 times per second. At each sampling time, the sensors record the collected dissolved oxygen concentration, pH, temperature, and ammonia nitrogen concentration values ​​with a timestamp, generating time-series data of the aquaculture water's physicochemical parameters. Each record in the time-series data of physicochemical parameters of the aquaculture water contains a timestamp that is strictly aligned with each frame of the herd behavior image sequence, as well as the corresponding monitoring point identifier under that timestamp. This achieves synchronous alignment between the herd behavior image sequence and the time-series data of physicochemical parameters of the aquaculture water in the time dimension. Through the above arrangement and synchronous acquisition mechanism, herd behavior image sequences of Chuzhou crucian carp at multiple monitoring points, as well as time-series data of physicochemical parameters of the aquaculture water aligned with the herd behavior image sequence, were obtained.

[0036] Example 2:

[0037] In specific implementation, please refer to Figure 3 The process of extracting the spatial distribution characteristics and individual movement characteristics of Chuzhou crucian carp from the group behavior image sequence, and constructing a behavioral situation map representing the stress state of the Chuzhou crucian carp group using the group spatial distribution characteristics and individual movement characteristics, is achieved through the following methods.

[0038] When performing object detection on each frame of the group behavior image sequence, the object detection algorithm adopts a two-stage detection architecture based on convolutional neural networks. The core architecture of the object detection algorithm consists of three modules connected in series: a backbone feature extraction network, a region proposal network, and a classification and regression network. The backbone feature extraction network uses a deep residual network structure, containing multiple cascaded residual convolutional blocks. Each residual convolutional block consists of two convolutional layers connected by identity skip connections. The input of the backbone feature extraction network is a three-channel image tensor from a single frame of the group behavior image sequence, and the output is a set of multi-scale feature maps. The region proposal network is connected after the backbone feature extraction network. The region proposal network slides anchor boxes of a preset size on each layer of the multi-scale feature map set output by the backbone feature extraction network. The set of anchor box size ratios includes three aspect ratios and three scales. For each anchor box, the region proposal network simultaneously outputs a foreground / background binary classification score and a bounding box coordinate regression offset, and retains a preset number of candidate region proposals after non-maximum suppression processing. The classification and regression network is connected after the region proposal network. It utilizes candidate region proposals to perform region-of-interest pooling on the multi-scale feature map set output by the backbone feature extraction network. This pooling process uniformly samples the feature regions corresponding to each candidate region proposal to a fixed-size feature representation. The classification and regression network then performs classification of the Chuzhou crucian carp category against the background category and refines the bounding box coordinates of individual Chuzhou crucian carp on this fixed-size feature representation. The model parameters of the object detection algorithm are obtained through supervised training on a labeled dataset of individual Chuzhou crucian carp in a factory farming environment. This labeled dataset includes manually labeled bounding box coordinates and keypoint coordinates of individual Chuzhou crucian carp contours.

[0039] During the execution of the target detection algorithm, the image region corresponding to each monitoring point in each frame of the group behavior image sequence is input into the backbone feature extraction network of the target detection algorithm. After forward propagation calculation, the confidence score of the Chuzhou crucian carp category and the bounding box coordinates of the individual Chuzhou crucian carp are obtained at the output of the classification and regression network. For each detection result of the Chuzhou crucian carp category confidence score exceeding the preset detection confidence threshold, the bounding box coordinates of the individual Chuzhou crucian carp are extracted. The center coordinates of the bounding box of the individual Chuzhou crucian carp are taken as the individual position of the individual Chuzhou crucian carp in the image plane, and the closed contour formed by curve interpolation of the four corner coordinates of the bounding box of the individual Chuzhou crucian carp is taken as the contour of the individual Chuzhou crucian carp. The individual position coordinates and contour information of all detected individuals Chuzhou crucian carp within the same monitoring point are collected to form the individual distribution coordinate set within each monitoring point. The individual distribution coordinate set is organized in the form of a list or matrix. Each row corresponds to one individual Chuzhou crucian carp within the monitoring point. Each row contains the individual's identification number, the horizontal and vertical coordinates of the individual's location in the image plane coordinate system, and the sequence of contour points.

[0040] In the process of statistically analyzing the population density and spatial dispersion of *Crucian carp* at each monitoring point based on the individual distribution coordinate set, the population density is calculated by dividing the number of *Crucian carp* individuals in the same monitoring point's individual distribution coordinate set by the area of ​​the observation water covered by the underwater camera at that monitoring point. The observation water area is pre-calibrated based on the underwater camera's field of view parameters and deployment depth through geometric conversion. The population density is expressed as the number of *Crucian carp* individuals per square meter.

[0041] The spatial dispersion is calculated by treating the individual position coordinates of all Chuzhou crucian carp individuals within the same monitoring point's distribution coordinate set as a point set in a two-dimensional space, and calculating the sum of the variances of the abscissa and ordinate values ​​of all individual position coordinates within the point set. The numerical value of the spatial dispersion is expressed by the following formula:

[0042]

[0043] in, Indicates the monitoring point number. The range of values ​​is Total number of monitoring points Integers between; Indicates the sampling time number. The range of values ​​is Total number of frames in the group behavior video sequence Integers between; Indicates the number is The monitoring points were numbered at the sampling time. The spatial dispersion of time; Indicates the number is The monitoring points were numbered at the sampling time. The total number of Chuzhou crucian carp individuals within the time-distribution coordinate set. The number of detection results whose confidence level for the Chuzhou crucian carp category output by the target detection algorithm exceeds a preset detection confidence threshold is determined; This represents the index of an individual Chuzhou crucian carp within the set of individual distribution coordinates. The range of values ​​is to Integers between; Indicates the number is The monitoring points were numbered at the sampling time. The time sequence number index is The x-coordinate value of the individual Chuzhou crucian carp's location in the image plane coordinate system. The x-coordinate of the center of the bounding box of the individual Chuzhou crucian carp is determined by the target detection algorithm; Indicates the number is The monitoring points were numbered at the sampling time. The arithmetic mean of the x-coordinates of all individual Chuzhou crucian carp within the time-distribution coordinate set; Indicates the number is The monitoring points were numbered at the sampling time. The time sequence number index is The vertical coordinate of the individual Chuzhou crucian carp in the image plane coordinate system. The center ordinate of the bounding box of the individual Chuzhou crucian carp is determined by the target detection algorithm. Indicates the number is The monitoring points were numbered at the sampling time. The arithmetic mean of the ordinate values ​​of all individual Chuzhou crucian carp within the time-varying coordinate set. Population density and spatial dispersion together constitute the spatial distribution characteristics of the population.

[0044] In the process of tracking the positional offset of the same individual Chuzhou crucian carp in adjacent frames of imagery to calculate instantaneous swimming speed and turning frequency, a multi-target tracking algorithm based on Kalman filtering is used to establish the correspondence between the same Chuzhou crucian carp in adjacent frames. The multi-target tracking algorithm models the motion state of each Chuzhou crucian carp individual as a state vector, which contains four state variables: the horizontal coordinate of the individual's position, the vertical coordinate of the individual's position, the velocity component in the horizontal direction, and the velocity component in the vertical direction. In each frame, the multi-target tracking algorithm predicts the prior estimate of the state vector using Kalman filtering, and uses the center coordinates of the Chuzhou crucian carp individual's bounding box output by the target detection algorithm as the observation value to update the state vector posteriorly. Finally, the Hungarian algorithm is used to allocate the minimum cost between the predicted position and the observed position corresponding to the predicted state vector, achieving correlated tracking of the same Chuzhou crucian carp individual between adjacent frames. The instantaneous swimming speed of each individual Chuzhou crucian carp is calculated as follows: The individual's position coordinates, updated via Kalman filtering posterior time between two consecutive frames, are taken. The Euclidean distance between the individual's position coordinates in the two frames is calculated. This Euclidean distance is divided by the time interval between the two frames to obtain the instantaneous swimming speed, expressed in centimeters per second (cm / s). The turning frequency of each individual Chuzhou crucian carp is calculated as follows: Within a preset turning statistics time window, the absolute value of the change in the swimming direction angle of the same individual Chuzhou crucian carp between adjacent frames is accumulated. The swimming direction angle is calculated based on the arctangent angle of the displacement vectors of the individual's position coordinates in the image plane coordinate system between adjacent frames. Each turn is counted when the cumulative absolute value of the change in the swimming direction angle exceeds a preset turning angle threshold. The number of turns within the turning statistics time window is divided by the length of the turning statistics time window to obtain the turning frequency, expressed in times per second (times / s). The preset turning statistics time window is set to 2 seconds, and the preset turning angle threshold is set to 45 degrees. Instantaneous swimming speed and turning frequency together constitute the characteristics of individual movement.

[0045] In the process of arranging the population spatial distribution characteristics and individual movement characteristics of each monitoring point at the same time into a behavior vector according to the monitoring point code, a unique monitoring point code is pre-assigned to all monitoring points in the aquaculture pond. The monitoring point codes are arranged sequentially according to the row-to-column order of the spatial grid layout using integer serial numbers. At the same sampling time, the points are numbered as follows: At any given time, all monitoring points are traversed. For each monitoring point, the population spatial distribution characteristics of that point, including population density and spatial dispersion, as well as the average instantaneous swimming speed and average turning frequency of all individuals of Chuzhou crucian carp within that point—a total of four feature values—are concatenated into a four-dimensional feature vector in a fixed order of population density, spatial dispersion, average instantaneous swimming speed, and average turning frequency. The four-dimensional feature vectors of all monitoring points at the same sampling time are then concatenated end-to-end according to the ascending order of the monitoring point codes, forming a behavior vector with a dimension four times the total number of monitoring points. Dimensions 1 to 4 of the behavior vector correspond to monitoring points coded as 1, dimensions 5 to 8 correspond to monitoring points coded as 2, and so on. The values ​​in each dimension of the behavior vector are real numbers; the unit for population density is cells / square meter, the unit for spatial dispersion is square pixels, the unit for average instantaneous swimming speed is centimeters / second, and the unit for average turning frequency is times / second.

[0046] In constructing a behavioral situation map by concatenating behavioral vectors from consecutive moments along the time axis, starting from the moment corresponding to the first frame of the group behavior image sequence, behavioral vectors are extracted sequentially for each sampling moment. These vectors are then stacked row-by-row along the time axis in ascending order of sampling moment number, forming a two-dimensional numerical matrix. The number of rows in the two-dimensional numerical matrix of the behavioral situation map corresponds to the total number of frames in the group behavior image sequence. The total number of monitoring points is four times the number of columns, and the second column in the two-dimensional numerical matrix of the behavior pattern map is... The corresponding sampling time number is The action vector at time , the first Line 1 The elements of the column represent the first element in the action vector at that moment. The behavioral situation map contains the evolutionary trajectory of the behavior of the Chuzhou crucian carp population at each monitoring point over time. The sequence of numerical changes in the four dimensions corresponding to any monitoring point in the behavioral situation map between consecutive rows constitutes a complete record of the evolution of the population behavior state at that monitoring point over time.

[0047] See Figure 6 The curve in the figure shows the numbering of a monitoring point at different sampling times in Example 2. The changing trends of population density and spatial dispersion in the spatial distribution characteristics of crucian carp in Xiachuzhou. The horizontal axis represents the sampling time number. The range is from 0 to 500, reflecting continuous time series sampling points. The left vertical axis corresponds to the population density represented by the solid line, with the unit being cells / square meter, and the value fluctuates between 6 and 18; the right vertical axis corresponds to the spatial dispersion represented by the dashed line, with the unit being square pixels, and the value ranges from approximately 1600 to 2300.

[0048] The population density curve exhibits significant periodic fluctuations. Throughout the sampling window, the population density experienced multiple peaks and troughs, with the highest peak reaching nearly 18 individuals / m² and the lowest trough around 6 individuals / m², indicating dynamic changes in the aggregation degree of Chuzhou crucian carp at this monitoring site. The spatial dispersion curve also shows fluctuating changes, but its fluctuation trend has a certain negative correlation with the population density: when the population density reaches a local peak, the spatial dispersion often experiences a local trough; conversely, when the spatial dispersion peaks, the population density is often at a lower level, indicating an increased tendency for the fish to disperse spatially.

[0049] Between approximately sampling time 100 and 200, the population density generally showed a decreasing trend, while the spatial dispersion increased, reflecting the transition of the Chuzhou crucian carp population from a relatively aggregated state to a dispersed state during this period. Around time 300 to 400, the population density increased significantly again, reaching multiple peaks, while the spatial dispersion decreased accordingly, exhibiting the behavior characteristic of the fish re-aggregating. Overall, population density and spatial dispersion together reflect the dynamic characteristics of the spatial distribution of the Chuzhou crucian carp population at this monitoring point in the temporal dimension, consistent with the technical content of Example 2 where population density and spatial dispersion jointly constitute the spatial distribution characteristics of the population.

[0050] Example 3:

[0051] In specific implementation, please refer to Figure 4 The process of generating a multimodal correlation feature set that integrates behavior patterns and water quality by spatiotemporally mapping behavioral patterns with time-series data of physicochemical parameters of aquaculture water bodies is achieved through the following methods.

[0052] The behavioral vectors of each monitoring point in the behavioral trend map are paired with the corresponding physicochemical parameter values ​​of the monitoring points at the same time in the time series data of aquaculture water physicochemical parameters, using the same timestamp. The behavioral trend map is a two-dimensional numerical matrix, with each row corresponding to a sampling time. Each row stores the behavioral vectors of each monitoring point in segments according to the monitoring point coding order. The time series data of aquaculture water physicochemical parameters is organized in a time series format. Each record includes the timestamp of the sampling time, the monitoring point code, and the values ​​of four physicochemical parameters: dissolved oxygen concentration, pH, temperature, and ammonia nitrogen concentration. At each sampling time, all monitoring points are traversed. For each monitoring point, a four-dimensional behavioral sub-vector corresponding to that monitoring point is extracted from the behavioral trend map, and the four physicochemical parameter values ​​corresponding to the same monitoring point code at the same time are extracted from the time series data of aquaculture water physicochemical parameters. The four-dimensional behavioral sub-vector and the four physicochemical parameter values ​​are concatenated into an eight-dimensional vector in a fixed order: behavior vector first, physicochemical parameter values ​​second, forming a behavior-water quality pairing sample. In the behavior-water quality paired samples, the first four dimensions are, in order, population density, spatial dispersion, mean instantaneous swimming speed, and mean turning frequency; the latter four dimensions are, in order, dissolved oxygen concentration, pH, temperature, and ammonia nitrogen concentration. The behavior-water quality paired samples formed by all monitoring points at all sampling times constitute a complete behavior-water quality paired sample set.

[0053] A time-window sliding aggregation was performed on the behavior-water quality paired samples to calculate the mean rate of change of the behavior vector and the mean rate of change of the physicochemical parameters for each monitoring point within a preset time window. The length of the preset time window was set in units of the number of sampling moments. The length of the preset time window was determined based on the known short-time response period between the behavior response and water quality fluctuations of Chuzhou crucian carp in a factory farming environment, and was set to 60 sampling moments. When the sampling frequency was 25 frames / second, this corresponded to a time span of 2.4 seconds. During the sliding aggregation process, for each monitoring point, a complete preset time window was slid along the time axis from the start time to the end time with a fixed sliding step size. After each slide, the window contained behavior-water quality paired samples from 60 consecutive sampling moments. For each window position, the first four dimensions of the 60 behavior-water quality paired samples within the window were extracted to form the window behavior matrix, and the last four dimensions of the 60 behavior-water quality paired samples within the window were extracted to form the window water quality matrix. Calculate the arithmetic mean of all row vectors in the window behavior matrix to obtain the behavior mean vector at that window location. Similarly, calculate the arithmetic mean of all row vectors in the window water quality matrix to obtain the water quality mean vector at that window location. The rate of change of the Euclidean distance between the behavior mean vectors of two adjacent window locations is taken as the rate of change of the behavior vector mean, and its calculation formula is as follows:

[0054]

[0055] in, Indicates the monitoring point code. The range of values ​​is Total number of monitoring points deployed in the aquaculture ponds Integers between; The index representing the position of the sliding window. The range of values ​​is Total number of sliding windows Integers between; The monitoring point code is indicated as The monitoring point is at the The rate of change of the mean of the behavior vector at each sliding window position It is a dimensionless, non-negative real number; The monitoring point code is indicated as The monitoring point is at the The behavior mean vector at each sliding window position. A four-dimensional real vector obtained by taking the arithmetic mean of the first four dimensions of the behavior-water quality paired samples within the window; The monitoring point code is indicated as The monitoring point is at the The mean vector of behavior at each sliding window position; The Euclidean norm of a vector; A smoothing constant is introduced to avoid the denominator being zero. The value is set to .

[0056] The calculation of the mean water quality variation rate uses the same operational structure as the mean variation rate of the behavior vector. The rate of change of the Euclidean distance between the mean water quality vectors of two adjacent window positions is taken as the mean variation rate of the physicochemical parameter values. The monitoring point code is indicated as The monitoring point is at the The rate of change of the mean water quality at each sliding window position is calculated in the same way as... The only difference is that the behavior mean vector is replaced with the water quality mean vector at the corresponding window position. The sequence of the rate of change of the behavior vector mean constitutes the behavior change feature, and the sequence of the rate of change of the physicochemical parameter values ​​means constitutes the water quality change feature. Both the behavior change feature and the water quality change feature form a one-dimensional time series along the direction of the sliding window in chronological order.

[0057] The behavioral change characteristics and water quality change characteristics of the same monitoring point are cross-correlated to calculate the response delay time of behavioral change to water quality change. For each monitoring point, a one-dimensional time series of behavioral change characteristics and a one-dimensional time series of water quality change characteristics are extracted, and the cross-correlation function between the two series is calculated. The cross-correlation function is calculated by fixing the behavioral change characteristic series and shifting the water quality change characteristic series forward and backward along the time axis by multiple lag steps, and calculating the sum of the dot products of the two series at each lag step. A curve of cross-correlation function value changing with lag step is generated, and the lag step corresponding to the maximum cross-correlation function value in the cross-correlation function curve is searched. The lag step is multiplied by the time interval corresponding to the sliding window step size to obtain the response delay time of behavioral change to water quality change, with the unit of response delay time being seconds. When the cross-correlation function curve has multiple peaks, the lag step corresponding to the peak with the largest absolute value is selected. If the cross-correlation function value corresponding to the maximum peak is lower than the preset correlation threshold, it is determined that there is no statistically significant response relationship between the behavioral change and water quality change at the current monitoring point, and the response delay time is marked as null. The response delay time of behavioral changes to water quality changes is used as a behavior-water quality response hysteresis characteristic.

[0058] The behavior-water quality paired samples, behavior change features, water quality change features, and behavior-water quality response lag features of each monitoring point are uniformly encoded to form a multimodal correlation feature set. During the uniform encoding process, for each monitoring point at each effective sliding window position, an encoded feature vector is constructed. The encoded feature vector consists of four parts concatenated sequentially: the first part is the standardized eight-dimensional vector of the behavior-water quality paired sample corresponding to the midpoint of the window; the second part consists of two values: the mean change rate of the behavior vector at the current window position and the mean change rate of the behavior vector at the previous window position; the third part consists of two values: the mean change rate of the water quality at the current window position and the mean change rate of the water quality at the previous window position; and the fourth part consists of two values: the maximum cross-correlation function value corresponding to the behavior-water quality response lag time and the response delay time at the monitoring point. If the behavior-water quality response lag time of a certain monitoring point is marked as null, the fourth part is filled with two preset invalid value identifier constants. The encoded feature vectors constructed from all monitoring points at all effective sliding window positions are aggregated into a two-dimensional matrix. Each row of the two-dimensional matrix corresponds to an encoded feature vector, and each column corresponds to a feature dimension of the encoded feature vector. This two-dimensional matrix is ​​the multimodal association feature set. The multimodal association feature set simultaneously contains instantaneous behavior-water quality pairing information, short-term behavior trend information, short-term water quality trend information, and time-series response relationship information between behavior and water quality.

[0059] Example 4:

[0060] In specific implementation, please refer to Figure 5The process of inferring the probability of disease outbreak risk at the current moment based on a multimodal association feature set and a pre-constructed disease risk probability graph model is achieved in the following way.

[0061] Each monitoring point in the multimodal correlation feature set is considered a node in the graph model, and a unique node identifier is assigned to each node, with a one-to-one correspondence between the node identifier and the monitoring point code. Spatial adjacency between monitoring points is considered as edges in the graph model, determined based on the physical adjacency between monitoring points in the spatial grid layout within the aquaculture pond. In the spatial grid layout, if two monitoring points share a grid boundary or grid vertex on the horizontal plane, or if two monitoring points have a predetermined proportion of overlapping area within the underwater camera's field of view, then a spatial adjacency is determined between the two monitoring points, and an undirected edge is established between the corresponding two nodes in the graph model. All nodes and edges are combined to form a spatial topology graph structure, stored in the form of an adjacency matrix. The row and column indices in the adjacency matrix correspond to node identifiers; if an edge exists between two nodes, the matrix element has a value of 1; otherwise, the matrix element has a value of 0.

[0062] The local anomaly score for each node is calculated based on the behavior-water quality paired samples corresponding to each node. For each node, an eight-dimensional vector is extracted from the behavior-water quality paired samples corresponding to that node at the current moment from the multimodal association feature set. The first four dimensions of the eight-dimensional vector are behavior sub-vectors, including population density, spatial dispersion, mean instantaneous swimming speed, and mean turning frequency. The last four dimensions are water quality sub-vectors, including dissolved oxygen concentration, pH, temperature, and ammonia nitrogen concentration. The behavior baseline level and water quality baseline level of the node within the historical health cycle have been pre-established through statistical modeling. The behavior baseline level is represented by the mean vector and covariance matrix of the behavior sub-vectors of the node at all sampling moments within the historical health cycle, and the water quality baseline level is represented by the mean vector and covariance matrix of the water quality sub-vectors of the node at all sampling moments within the historical health cycle. The local anomaly score is calculated as follows: the behavior subvector at the current moment is substituted into the Mahalanobis distance calculation formula of the behavior baseline level to obtain the behavior anomaly item; the water quality subvector at the current moment is substituted into the Mahalanobis distance calculation formula of the water quality baseline level to obtain the water quality anomaly item; the behavior anomaly item and the water quality anomaly item are weighted and summed, with weight coefficients of 0.5 and 0.5 respectively. The result of the weighted sum is used as the local anomaly score of the node. The local anomaly score is a dimensionless non-negative real number. The larger the local anomaly score, the greater the deviation of the behavior characteristics and water quality characteristics from the historical baseline level.

[0063] The state transfer probability between adjacent nodes is calculated using behavior-water quality response lag features. For any pair of adjacent nodes connected by an edge in the spatial topology graph, the behavior-water quality response lag times of each adjacent node are extracted from the multimodal association feature set. If the behavior-water quality response lag time of either node is marked as null, the state transfer probability between the two adjacent nodes is directly set to a preset baseline transfer probability value, which is set to 0.1. If both behavior-water quality response lag times of the two adjacent nodes are valid values, the absolute value of the difference between the two response delay times is calculated. This absolute value is then input into an exponential decay function, and the expression for the state transfer probability is:

[0064]

[0065] in, This represents the identifier of the starting node of an edge in a spatial topological graph. The range of values ​​is Total number of nodes Integers between; This represents the identifier of the terminating node of the same edge in a spatial topological graph structure. The range of values ​​is Total number of nodes Integers between 1 and 2, and the node identifier is 1. The nodes and node identifiers are There are edges connecting the nodes; The identifier of the slave node is The node-to-node identifier is The probability of state propagation of nodes. The range of values ​​is within to between; The node identifier is The behavior of the node - water quality response lag time, in seconds; The node identifier is The behavior of the node - water quality response lag time, in seconds; To transmit probability amplitude coefficients, The range of values ​​is to real numbers between Set as The setting is based on the empirical proportion of attenuation of abnormal states propagating along the direction of water flow in a factory farming environment. The decay rate coefficient, Set as The setting is based on the fact that for every 2-second increase in response latency difference, the probability of state propagation decreases to approximately [value missing]. The exponential decay rate is times that of the previous one.

[0066] Using local anomaly scores as observational evidence and state transition probabilities as state transition constraints, confidence propagation inference is performed on a spatial topology graph structure. The confidence propagation inference employs a factor graph-based belief propagation algorithm, transforming the spatial topology graph structure into a factor graph representation. The factor graph contains variable nodes and factor nodes; each monitoring point corresponds to a variable node, and the state space of the variable node is binary, including both outbreak and non-outbreak states. Each variable node is associated with an observation factor, determined by the local anomaly score. The observation factor defines the observed likelihood probability of observing the local anomaly score when the variable node is in an outbreak state and the observed likelihood probability of observing the local anomaly score when the variable node is in a non-outbreak state. Each pair of adjacent variable nodes is associated with a transition factor, determined by the state transition probability, which defines the spatial transition probability matrix of outbreak states between adjacent variable nodes. The belief propagation algorithm performs iterative message passing on the factor graph, with 20 iterations. In each iteration, each variable node sends a variable-to-factor message to its neighboring factor nodes, and each factor node sends a factor-to-variable message to its neighboring variable nodes. The message update rule is implemented in the logarithmic field using a sum-product algorithm. After 20 iterations converge, the posterior probability of the disease outbreak state is extracted from the final marginal beliefs of each variable node. The posterior probabilities of disease outbreaks for all variable nodes are summed, and the maximum value among all posterior probabilities is taken as the disease outbreak risk probability at the current moment.

[0067] The process of issuing an early warning signal when the probability of a disease outbreak exceeds the warning threshold is implemented as follows: The current probability of a disease outbreak is compared with a preset risk probability warning threshold, which is between 0.7 and 0.9. If the current probability of a disease outbreak is greater than or equal to the risk probability warning threshold, the warning triggering condition is determined to be met.

[0068] After the early warning triggering conditions are met, the posterior probability of disease outbreak obtained by each variable node in the spatial topology structure during the final iteration of the belief propagation algorithm is traced back. The posterior probabilities of disease outbreak of each variable node are sorted in descending order of numerical value, and the variable node with the highest posterior probability of disease outbreak is located. The monitoring point corresponding to this variable node is determined as the suspected disease source point. Based on the spatial location of the suspected disease source point in the spatial topology structure, the behavioral change features and water quality change features of the suspected disease source point for multiple consecutive time periods before the triggering time are extracted from the multimodal association feature set. The duration of the multiple consecutive time periods is set to the time period corresponding to the 120 sliding window positions before the triggering time. The mean change rate sequence of the behavioral vector and the mean change rate sequence of water quality of the suspected disease source point are extracted within this time period. The mean change rate sequence of the behavioral vector and the mean change rate sequence of water quality are organized into two synchronous change curves in chronological order to generate a trend evidence chain to indicate the direction of disease spread. The trend evidence chain contains the synchronous evolution of the behavioral change trend curve and the water quality change trend curve on the time axis. The output includes an early warning signal containing the location of suspected disease sources and a trend evidence chain. The specific form of the early warning signal includes the monitoring point code of the suspected disease source, the two-dimensional or three-dimensional spatial coordinates of the suspected disease source in the spatial grid layout of the aquaculture pond, the time series value list of the mean change rate of the behavior vector of the trend evidence chain, the time series value list of the mean change rate of the water quality of the trend evidence chain, and the probability value of the disease outbreak risk at the current moment.

[0069] Example 5:

[0070] In practice, the evolution trajectory of the behavioral situation map over time is reduced to a two-dimensional or three-dimensional manifold space using sliding window principal component analysis (PCA), and then simultaneously displayed on the aquaculture monitoring terminal in a visual manner. The PCA process involves extracting a sliding window along the time axis from the two-dimensional numerical matrix of the behavioral situation map. The length of the sliding window is set to 120 sampling times, and the sliding step size is set to 6 sampling times. For the window behavior matrix at each sliding window position, all column vectors of the window behavior matrix are centered to make the mean of each column zero. The covariance matrix of the centered matrix is ​​calculated, and eigenvalue decomposition is performed on the covariance matrix to extract the eigenvectors corresponding to the two largest eigenvalues, forming a two-dimensional projection matrix. The behavior vector at the last sampling time within the sliding window is multiplied by the two-dimensional projection matrix to obtain a coordinate point in the two-dimensional manifold space. If a three-dimensional manifold space needs to be displayed, the eigenvectors corresponding to the three largest eigenvalues ​​are extracted to form a three-dimensional projection matrix. The window is slid along the time axis sequentially, and the above projection operation is repeated to connect all projected coordinate points in chronological order, forming the continuous evolution trajectory of the behavioral situation map in the low-dimensional manifold space. The evolutionary trajectory is rendered in real time on the display interface of the aquaculture monitoring terminal as a dynamic curve or trajectory animation. The horizontal and vertical axes correspond to the scores of the first and second principal components, respectively. If it is a three-dimensional display, a third principal component score is added as a depth dimension for color or height mapping. The aquaculture monitoring terminal simultaneously marks the projection point position corresponding to the current moment in the evolutionary trajectory, and uses color changes to represent the fluctuation amplitude of the group spatial distribution characteristics and individual movement characteristics in the behavioral situation map.

[0071] The specific construction method of the disease risk probability graph model is as follows: Multiple sets of historical behavioral image sequences and time-series data of historical water physicochemical parameters were collected from Chuzhou crucian carp in a factory farming environment during their historical health cycles. The time span of the historical health cycles covers at least two complete Chuzhou crucian carp farming cycles, including seasonal changes and water quality fluctuations, and other common farming variations. Disease events were recorded synchronously within the corresponding time periods. Disease events were recorded by aquaculture technicians during the Chuzhou crucian carp farming process based on symptoms of bleeding on the carp's body surface or positive microscopic examination results of parasites. The time of diagnosis and the period tracing back to the start of the disease incubation period before the diagnosis time were marked as disease occurrence periods. All sampling times within the disease occurrence period were marked as disease outbreak states, and other sampling times were marked as non-disease states, forming a historical training dataset labeled with disease.

[0072] Historical multimodal correlation feature sets were extracted from the historical training dataset, using the same extraction method as that used to generate multimodal correlation feature sets from group behavior image sequences and time-series data of aquaculture water physicochemical parameters. The disease outbreak state at each historical moment was used as a latent variable, a binary variable with a value of 1 representing a disease outbreak state and a value of 0 representing a non-disease state. The historical multimodal correlation feature sets were used as observed variables. Based on the conditional dependencies between the observed and latent variables, a disease risk probability graphical model was constructed, including the spatial node state prior distribution, the inter-node state transition matrix, and the observation likelihood function. The core framework of the disease risk probability graphical model is a combination of conditional random fields and hidden Markov models. Spatially, pairwise conditional random fields are used to model the interdependence of disease states between nodes, while in the temporal dimension, hidden Markov chains are used to model the evolution of the disease state of the same node along the time axis. The spatial node state prior distribution defines the initial probability of each node being in a disease outbreak state under the condition of no observation information. The spatial node state prior distribution is stored in vector form, and the length of the vector is equal to the total number of nodes in the spatial topology graph. Each element in the vector is the prior probability value of the corresponding node being in a disease outbreak state. The node state transition matrix defines the spatial propagation probability of disease states between adjacent nodes in the spatial topology graph. The node state transition matrix is ​​a two-dimensional matrix. The row index corresponds to the disease state of the starting node, and the column index corresponds to the disease state of the ending node. The matrix elements represent the conditional probability that the ending node is in a certain disease state given that the starting node is in a certain disease state. The observation likelihood function defines the conditional probability distribution of the corresponding encoded feature vectors observed in the historical multimodal association feature set under a given latent variable state. The observation likelihood function is represented by a Gaussian mixture model. The number of Gaussian components in the Gaussian mixture model is set to 3, and each Gaussian component corresponds to a typical association pattern between disease states and multimodal association feature sets.

[0073] Maximum likelihood estimation was performed on the model parameters of the disease risk probabilistic graphical model using a historical training dataset. The model parameters include the prior state probabilities of each node, the state transition probabilities between adjacent nodes, and the conditional probability distribution of each observed feature under a given state. Maximum likelihood estimation was performed iteratively using the expectation-maximization algorithm, which terminated when the relative change in the log-likelihood function value was less than a certain threshold. Or the number of iterations reaches 100. In the expectation step of the expectation-maximization algorithm, the posterior probability distribution of each latent variable under the historical training dataset is calculated using the current model parameters. In the maximization step of the expectation-maximization algorithm, the prior state probability of each node, the state transition probability between adjacent nodes, and the mean vector, covariance matrix, and mixing coefficients of each Gaussian component in the Gaussian mixture model are updated according to the posterior probability distribution. The disease risk probability graphical model after parameter estimation is completed is used as a pre-constructed disease risk probability graphical model, saving the spatial node state prior distribution vector, the state transition matrix between nodes, and the complete parameter set of the Gaussian mixture model.

[0074] The warning threshold is determined using a dynamic update method. During the normal breeding phase without disease events, multimodal correlation feature sets of the normal state are continuously collected across multiple normal breeding cycles, with at least three complete breeding cycles. The probability of a disease outbreak in the normal state at each moment within each normal breeding cycle is calculated using a pre-constructed disease risk probability graph model. The probability of a disease outbreak in the normal state at all moments within all normal breeding cycles is aggregated into a one-dimensional array, forming a normal risk probability distribution. The statistical upper limit of the normal risk probability distribution is calculated using the 95th percentile. All normal state disease outbreak probability values ​​in the normal risk probability distribution are sorted from smallest to largest, and the value at the 95th percentile after sorting is taken as the statistical upper limit. This statistical upper limit is used as the baseline warning threshold.

[0075] In subsequent breeding processes, after each disease-free breeding cycle is completed, the actual multimodal association feature set collected during that cycle is added to the normal sample library. The normal sample library accumulates and stores the normal state multimodal association feature sets of all historical disease-free breeding cycles. The probability of disease outbreak under normal conditions is recalculated for all samples in the updated normal sample library using a pre-constructed disease risk probability graph model, forming an updated normal risk probability distribution. The 95th percentile of this updated normal risk probability distribution is used as the updated statistical upper limit to update the baseline warning threshold. The updated baseline warning threshold is compared with a preset minimum safety threshold, which is set to 0.7. The larger of the updated baseline warning threshold and the minimum safety threshold is taken as the current actual warning threshold.

[0076] The types of diseases include bacterial hemorrhagic septicemia and parasitic diseases. The warning threshold ranges from 0.7 to 0.9, and the probability of disease outbreak is a dimensionless probability value between 0 and 1. When the disease type is bacterial hemorrhagic septicemia, the warning threshold is 0.8; when the disease type is parasitic disease, the warning threshold is 0.75. The specific value of the warning threshold is determined based on the incubation period and transmission speed of the disease type. Bacterial hemorrhagic septicemia has a shorter incubation period and a faster transmission speed than parasitic diseases; therefore, the warning threshold for bacterial hemorrhagic septicemia is set slightly higher than that for parasitic diseases.

[0077] After issuing the early warning signal, a confidence enhancement verification of the warning is performed. Following the issuance of the early warning signal, subsequent consecutive time-series data of verification behavior images and water body physicochemical parameters are acquired, with 180 sampling times corresponding to the time required for manual patrol response in aquaculture management. Based on the verification behavior image sequences, spatial distribution characteristics of the verification population and movement characteristics of individual verification individuals are extracted to construct a verification behavior situation map. This map is then spatiotemporally mapped to the water body physicochemical parameter time-series data to generate a verification multimodal correlation feature set. The extraction and construction methods for the verification population spatial distribution characteristics, individual movement characteristics, behavior situation map, and multimodal correlation feature set are completely consistent with the generation methods for the corresponding population spatial distribution characteristics, individual movement characteristics, behavior situation map, and multimodal correlation feature set. The verification multimodal correlation feature set is then input into a disease risk probability graph model to calculate the probability of disease outbreak risk at subsequent time points.

[0078] If the probability of a disease outbreak is verified to maintain an upward trend over multiple consecutive time periods and the probability of a disease outbreak at each time period is greater than the warning threshold, the criteria for judging multiple consecutive time periods are set to 120 consecutive sampling time periods. The criteria for judging the upward trend are that the slope of the linear fitting sequence of the probability of a disease outbreak is positive and the p-value of the hypothesis test is less than 0.05. Under the condition that the probability of a disease outbreak is verified to be greater than the warning threshold at each of the 120 consecutive sampling time periods and the slope of the linear fitting is significantly positive, the early warning signal is confirmed to be effective and the warning output status is maintained. The warning mark and the location information of the suspected disease source are continuously displayed on the aquaculture monitoring terminal.

[0079] If the probability of a disease outbreak decreases below the warning threshold within a preset verification time window (which is 300 sampling moments after the early warning signal is issued), the early warning signal is immediately revoked. The warning flag is removed from the aquaculture monitoring terminal, the multimodal association feature set corresponding to this false alarm event is marked as a negative sample, and the marked negative sample is added to the model update queue. The model update queue stores the 100 most recent negative samples in a first-in-first-out manner.

[0080] After feeding the multimodal correlation feature set corresponding to false alarm events as negative samples to the model update queue, incremental model updates are performed. Multiple accumulated negative samples and normal samples confirmed to be free of disease events within the most recent time period are extracted from the model update queue. The most recent time period is defined as the time length seven calendar days prior to the current moment. Normal samples are extracted from the portion of the normal sample library whose sampling time falls within the most recent time period. The extracted negative and normal samples are merged to form an incremental training sample set. The ratio of negative samples to normal samples in the incremental training sample set is controlled between 1:5 and 1:10. If the number of negative samples is insufficient, normal samples are randomly undersampled to maintain a balanced sample ratio.

[0081] For each sample in the incremental training sample set, the prediction error under the current disease risk probability graphical model is calculated according to the observation likelihood function in the disease risk probability graphical model. The prediction error is expressed as a negative log-likelihood loss. The incremental loss function is constructed by summing the negative log-likelihood losses of all samples in the incremental training sample set. The expression is:

[0082]

[0083] in, This represents the set of model parameters for the disease risk probability graph model. The set of model parameters includes the prior state probability of each node, the state transition probability between adjacent nodes, and the mean vector, covariance matrix, and mixing coefficients of each Gaussian component in the Gaussian mixture model. This represents the total number of samples in the incremental training sample set; This represents the index of a sample in the incremental training sample set. The range of values ​​is to Integers between; Indicates the index of the incremental training sample set. The observed variables corresponding to the samples are the encoded feature vectors of those samples. Indicates the index of the incremental training sample set. The latent variables corresponding to the samples are the known disease outbreak status labels corresponding to the samples, which are the disease labels or negative sample labels of false alarm events from the historical training dataset. This indicates that the set of model parameters in the disease risk probability diagram model is... Under the condition that the latent variable is Observed variable at time The conditional probability is calculated from the observation likelihood function defined by the Gaussian mixture model.

[0084] To minimize the incremental loss function, an online gradient descent method is used to incrementally adjust the model parameters of the disease risk probability graph model. The learning rate for online gradient descent is set to 0.001. In each iteration, a batch of 32 samples is randomly selected from the incremental training sample set. The gradient of the incremental loss function for each sample within the batch is calculated, multiplied by the learning rate, and subtracted from the current model parameters, completing one parameter update. This iteration is repeated until the incremental loss function value decreases by less than 0.001 over 10 consecutive iterations. Or the total number of iterations reaches 500. After completing the incremental adjustment, the updated model parameters replace the original model parameters. All newly received early warning inference requests use the updated model parameters to calculate the probability of disease outbreak risk. The model update queue is then cleared, releasing the storage space of the model update queue, thus completing the online incremental update of the disease risk probability map model.

[0085] See Figure 7 The figure shows the evolution trajectory of the behavioral situation map after dimensionality reduction to a two-dimensional manifold space using sliding window principal component analysis. The horizontal axis represents the score of the first principal component, and the vertical axis represents the score of the second principal component. The trajectory consists of multiple closed-loop curves, reflecting the dynamic changes of the behavioral vectors at different times in the low-dimensional manifold space. Each closed-loop curve exhibits certain periodic fluctuation characteristics. The score of the first principal component ranges from approximately -8 to 2, and the score of the second principal component ranges from approximately 1 to 7.5, indicating that the temporal characteristics of the behavioral situation map have a relatively stable but diverse distribution in the spatial dimension. The fluctuation amplitude and shape changes of the curves indicate that the spatial distribution characteristics and individual movement characteristics of the Chuzhou crucian carp population underwent multi-stage dynamic evolution during the monitoring period. The continuous connections between multiple nodes in the trajectory represent continuous sampling moments of the behavioral situation map in the time dimension. The changing trend and closed shape of the trajectory reflect the periodic changes in the behavioral state of the fish in the aquaculture pond and their possible stress responses. The two-dimensional manifold spatial evolution trajectory is displayed synchronously on the aquaculture monitoring terminal in the form of a dynamic curve, which helps to intuitively observe the spatiotemporal evolution of the behavior of Chuzhou crucian carp and provides morphological feature support for the early warning of subsequent diseases.

[0086] 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 early disease warning of Chuzhou crucian carp based on factory farming, characterized in that, Includes the following steps: Acquire the group behavior image sequence of Chuzhou crucian carp at multiple monitoring points in the aquaculture pond, and the time series data of the physicochemical parameters of the aquaculture water corresponding to the group behavior image sequence; Based on the group behavior image sequence, the spatial distribution characteristics and individual movement characteristics of Chuzhou crucian carp are extracted, and the spatial distribution characteristics and individual movement characteristics are used to construct a behavioral situation map characterizing the stress state of the Chuzhou crucian carp group; The behavioral situation map is spatiotemporally correlated with the time series data of the physicochemical parameters of the aquaculture water body to generate a multimodal correlation feature set that integrates behavior and water quality. Based on the multimodal association feature set, the probability of disease outbreak at the current moment is inferred through a pre-constructed disease risk probability graph model, and an early warning signal is issued when the probability of disease outbreak exceeds the warning threshold.

2. The method for early disease warning of Chuzhou crucian carp based on factory farming as described in claim 1, characterized in that, Acquire group behavior image sequences of Chuzhou crucian carp at multiple monitoring points within the aquaculture pond, and corresponding time-series data of the physicochemical parameters of the aquaculture water. Specifically: Multiple underwater cameras are deployed in a spatial grid layout within the breeding pond to simultaneously collect images of Chuzhou crucian carp swimming within the coverage area of ​​each underwater camera, forming a multi-view group behavior image sequence covering the entire pond; Simultaneously, water quality sensors are installed near each of the underwater cameras. The water quality sensors continuously collect dissolved oxygen concentration, pH, temperature, and ammonia nitrogen concentration in the aquaculture water at the same sampling frequency as the group behavior image sequence, generating time-series data of the physicochemical parameters of the aquaculture water that are time-aligned with the group behavior image sequence.

3. The method for early disease warning of Chuzhou crucian carp based on factory farming as described in claim 2, characterized in that, Based on the aforementioned group behavior image sequence, the spatial distribution characteristics and individual movement characteristics of the Chuzhou crucian carp were extracted. Then, using these characteristics, a behavioral pattern map representing the stress state of the Chuzhou crucian carp group was constructed. Specifically: For each frame of the group behavior image sequence, a target detection algorithm is executed to identify and locate the individual position and outline of each Chuzhou crucian carp in the image, and generate a set of individual distribution coordinates within each monitoring point; The population density and spatial dispersion of Chuzhou crucian carp at each monitoring point are statistically analyzed based on the individual distribution coordinate set, which serves as the spatial distribution characteristics of the population. Track the positional offset of the same Chuzhou crucian carp in adjacent frames of imagery, and calculate the instantaneous swimming speed and turning frequency of each Chuzhou crucian carp as the individual's motion characteristics; The spatial distribution characteristics of the population and the individual movement characteristics of each monitoring point at the same time are arranged according to the monitoring point code to form a behavior vector reflecting spatial heterogeneity; The behavioral vectors at consecutive moments are concatenated along the time axis to construct the behavioral situation map, which contains the evolution trajectory of the behavior of the Chuzhou crucian carp population at each monitoring point in the time dimension.

4. The method for early disease warning of Chuzhou crucian carp based on factory farming as described in claim 3, characterized in that, The behavioral pattern map is spatiotemporally correlated with the time-series data of the physicochemical parameters of the aquaculture water body to generate a multimodal correlation feature set that integrates behavior and water quality. Specifically: The behavior vectors of each monitoring point in the behavior pattern map are paired with the corresponding time and the corresponding physicochemical parameter values ​​of the aquaculture water body in the time series data of the physicochemical parameters of the monitoring point according to the same timestamp, so as to form a behavior-water quality pairing sample of each monitoring point at each time. The behavior-water quality paired samples are subjected to time window sliding aggregation to calculate the mean change rate of the behavior vector and the mean change rate of the physicochemical parameter values ​​of each monitoring point within the preset time window, thereby generating behavior change features and water quality change features that reflect short-term trends. Cross-correlation is performed between the behavioral change characteristics and the water quality change characteristics at the same monitoring point to calculate the response delay time of behavioral change to water quality change, and to generate behavioral-water quality response hysteresis characteristics. The behavior-water quality paired samples, behavior change features, water quality change features, and behavior-water quality response hysteresis features of each monitoring point are uniformly encoded to form the multimodal association feature set.

5. The method for early disease warning of Chuzhou crucian carp based on factory farming as described in claim 4, characterized in that, Based on the aforementioned multimodal correlation feature set, the probability of disease outbreak at the current moment is inferred through a pre-constructed disease risk probability graph model, specifically as follows: Each monitoring point in the multimodal correlation feature set is regarded as a node in the graph model, and the spatial adjacency between monitoring points is regarded as an edge in the graph model, thus constructing a spatial topology graph structure. The local anomaly score for each node is calculated based on the behavior-water quality paired samples corresponding to each node. The local anomaly score is used to quantify the degree of deviation between the behavioral characteristics and water quality characteristics at that node and the historical baseline level. The state propagation probability between adjacent nodes is calculated using the behavior-water quality response hysteresis characteristics. The state propagation probability represents the possibility that an abnormal state of a node will propagate to adjacent nodes along the spatial edge. Using the local anomaly score as observational evidence and the state transmission probability as state transition constraint, confidence propagation reasoning is performed on the spatial topology graph structure to iteratively update the posterior probability of disease outbreak for each node. The largest posterior probability of disease outbreak among all nodes is taken as the disease outbreak risk probability at the current moment.

6. The method for early disease warning of Chuzhou crucian carp based on factory farming as described in claim 5, characterized in that, An early warning signal is issued when the probability of disease outbreak exceeds the warning threshold, specifically: The probability of disease outbreak at the current moment is compared with a preset risk probability warning threshold. If the probability of disease outbreak at the current moment is greater than or equal to the risk probability warning threshold, the warning triggering condition is determined to be met. After the warning triggering conditions are met, the posterior probability of disease outbreak of each node in the spatial topology graph structure is further traced back, and the monitoring point corresponding to the node with the highest posterior probability of disease outbreak is located as the suspected disease source point. Based on the spatial location of the suspected disease source in the spatial topology, behavioral change features and water quality change features of the suspected disease source at multiple consecutive times before the triggering time are extracted from the multimodal association feature set to generate a trend evidence chain to indicate the direction of disease spread. The output includes the location of the suspected disease source and the trend evidence chain, forming an early warning signal.

7. The method for early disease warning of Chuzhou crucian carp based on factory farming as described in claim 6, characterized in that, The evolution trajectory of the behavioral situation map in the time dimension is reduced to a two-dimensional or three-dimensional manifold space using sliding window principal component analysis, and then displayed synchronously on the aquaculture monitoring terminal in a visual manner.

8. The method for early disease warning of Chuzhou crucian carp based on factory farming as described in claim 1, characterized in that, The specific method for constructing the disease risk probability graph model is as follows: Multiple sets of historical behavioral image sequences and time series data of historical water body physicochemical parameters were collected from Chuzhou crucian carp during their historical health cycles under factory farming conditions. At the same time, it was recorded whether disease events occurred during the corresponding time periods, forming a historical training dataset with disease labels. The historical multimodal correlation feature set is extracted from the historical training dataset according to the method described in claims 3 to 4, and the disease outbreak status at each historical moment is used as a latent variable, and the historical multimodal correlation feature set is used as an observation variable. Based on the conditional dependency between the observed variables and the latent variables, a disease risk probability graphical model is constructed, which includes the prior distribution of spatial node states, the state transition matrix between nodes, and the observation likelihood function. The model parameters of the disease risk probability graph model are estimated by maximum likelihood using the historical training dataset. The model parameters include the prior state probability of each node, the state transition probability between adjacent nodes, and the conditional probability distribution of each observed feature under a given state. The disease risk probability graph model after parameter estimation is used as the pre-constructed disease risk probability graph model.

9. A method for early disease warning of Chuzhou crucian carp based on factory farming, as described in claim 1, characterized in that, The warning threshold is determined using a dynamic update method, specifically: During the normal breeding phase without disease events, multimodal association feature sets of normal state are continuously collected for multiple normal breeding cycles, and the probability of disease outbreak in normal state at each moment in each normal breeding cycle is calculated through the disease risk probability graph model to form a normal risk probability distribution. Calculate the statistical upper limit of the normal risk probability distribution, and use the statistical upper limit as the benchmark early warning threshold; In subsequent breeding processes, after each breeding cycle without disease events is completed, the actual multimodal association feature set of the breeding cycle is added to the normal sample library, and the updated normal risk probability distribution and its statistical upper limit are recalculated to update the benchmark early warning threshold. The updated baseline warning threshold is compared with the preset minimum safety threshold, and the larger value between the updated baseline warning threshold and the minimum safety threshold is taken as the current actual warning threshold.

10. A method for early disease warning of Chuzhou crucian carp based on factory farming, as described in claim 1, characterized in that, The disease is a bacterial hemorrhagic disease or a parasitic disease, the warning threshold ranges from 0.7 to 0.9, and the probability of disease outbreak is a dimensionless probability value between 0 and 1.

Images