Method for predicting and decision-making of spirochetosis based on behavior data of carassius auratus in chuzhou

Abstract

The application discloses a spore disease prediction and decision support method based on behavior data of Chu Zhou crucian, and relates to the technical field of aquaculture disease monitoring. The method comprises the following steps: in response to a trigger signal that the depth data of the swimming trajectory of the Chu Zhou crucian in the aquaculture water environment is lower than a depth threshold value, collecting the back fin oscillation frequency and the gill cover opening and closing period of the marked individual; using a sliding window to extract the power spectral density distribution of the back fin oscillation frequency on the time sequence, and mapping the gill cover opening and closing period into a breathing rhythm array; locating an abnormal peak with an amplitude deviating from the mean value by more than two standard deviations in the breathing rhythm array, and aligning the energy attenuation amount of the corresponding frequency band in the power spectral density distribution with the abnormal peak occurrence time as the reference; after time point alignment, the energy attenuation amount and the amplitude ratio of the abnormal peak are spliced into a behavior feature tensor; the behavior feature tensor is input into a graph attention network, and a spore infection risk probability is output. The application realizes early warning and quantitative risk assessment of spore disease.

Description

Technical Field

[0001] This invention relates to the field of aquaculture disease monitoring and artificial intelligence technology, specifically a method for predicting and supporting sporozoan diseases based on behavioral data of crucian carp from Chuzhou. Background Technology

[0002] In the large-scale farming of crucian carp in Chuzhou, spore disease is one of the most serious parasitic diseases. In the early stages of infection, fish show no obvious external symptoms; by the time cysts are visible on the gills or body surface, the disease has already entered the middle to late stages, leading to high mortality and economic losses. Currently, disease risk is mainly inferred through regular manual sampling and microscopic examination or empirical observation of macroscopic behavioral changes such as slowed swimming and decreased appetite. Existing spore disease monitoring technologies have two limitations: first, manual microscopic examination is inefficient and has a strong lag, making it difficult to cover all fish populations; second, threshold judgment methods based on single behavioral indicators (such as swimming speed) are easily affected by environmental noise and cannot distinguish behavioral fluctuations caused by normal physiological activities such as feeding and stress, resulting in high false alarm or false negative rates. Current technologies lack multimodal coupling analysis of fine-grained behavioral signals, especially failing to uncover the coordinated changes in dorsal fin oscillation frequency and respiratory rhythm over time. Furthermore, conventional machine learning methods treat behavioral data as independent samples, ignoring the temporal proximity and dynamic transmission characteristics between abnormal events. To address the aforementioned issues, this invention proposes a method based on multi-dimensional behavioral feature extraction and graph structure modeling. This method involves real-time acquisition of dorsal fin oscillation frequency and gill cover opening and closing cycles, followed by power spectral density analysis and abnormal respiratory rhythm peak detection. Then, by aligning the abnormal peak times with the behavioral features, a behavioral feature tensor containing the energy attenuation ratio and amplitude ratio is constructed. Finally, a graph attention network is used to capture the temporal correlations between abnormal events, thereby achieving accurate prediction of spore-borne disease infection risk. Summary of the Invention

[0003] The purpose of this invention is to provide a method for predicting and supporting decision-making regarding spore-borne diseases based on behavioral data of crucian carp from Chuzhou. This method triggers signal acquisition by real-time monitoring of fish body depth, simultaneously acquiring the dorsal fin oscillation frequency and gill cover opening and closing cycle. It uses a sliding window transformation to extract the power spectral density distribution and construct a respiratory rhythm array, locating abnormal peaks and aligning the frequency band energy attenuation to form a behavioral feature tensor. Finally, it uses a graph attention network to output the probability of spore-borne disease infection risk, thereby solving the problems of difficulty in early diagnosis of spore-borne diseases, insufficient utilization of behavioral features, and lack of temporal correlation in existing technologies.

[0004] The objective of this invention can be achieved through the following technical solutions:

[0005] This invention provides a method for predicting and supporting decision-making regarding sporozoan diseases based on behavioral data of Chuzhou crucian carp, comprising: responding to a trigger signal that the swimming trajectory depth data of Chuzhou crucian carp in the aquaculture environment is lower than a depth threshold, acquiring the dorsal fin oscillation frequency and gill cover opening and closing cycle of marked individuals; extracting the power spectral density distribution of the dorsal fin oscillation frequency over a time series using a sliding window, and mapping the gill cover opening and closing cycle to a respiratory rhythm array; locating abnormal peaks in the respiratory rhythm array whose amplitude deviates from the mean by more than two standard deviations, and aligning the energy attenuation of the corresponding frequency band in the power spectral density distribution with the occurrence time of the abnormal peak as a reference; concatenating the ratio of the energy attenuation to the amplitude of the abnormal peak in time points, and then splicing them into a behavioral feature tensor; inputting the behavioral feature tensor into a graph attention network, and outputting the probability of sporozoan infection risk. Through the above steps, the subtle behavioral changes in the dorsal fin swing and gill cover opening and closing of Chuzhou crucian carp can be transformed into quantitative features. Graph attention networks can be used to capture the temporal correlation between abnormal respiratory events and the energy decay of dorsal fin swing, thereby accurately predicting the risk of sporozoan infection and providing early warning for aquaculture management.

[0006] As a technical solution of the present invention, the trigger signal that responds to the swimming trajectory depth data of Chuzhou crucian carp in the aquaculture environment is lower than the depth threshold, and the dorsal fin oscillation frequency and gill cover opening and closing cycle of the marked individuals are collected, specifically: multiple underwater binocular cameras are deployed at the bottom of the aquaculture pond, and the coordinate sequence of the body surface markers of each Chuzhou crucian carp in three-dimensional space is continuously acquired through the underwater binocular cameras; the maximum and minimum values ​​of the body surface markers in the vertical direction are extracted from the coordinate sequence, and the difference between the maximum and minimum values ​​is calculated as the swimming trajectory depth data; when the swimming trajectory depth data is lower than the preset depth threshold for three consecutive sampling cycles, the trigger signal is generated, and the dorsal fin edge contour tracking of the same Chuzhou crucian carp is started, and the number of dorsal fin oscillations per unit time is extracted from the dorsal fin edge contour tracking results as the dorsal fin oscillation frequency, and at the same time, the duration from the gill cover to the complete opening and closing of the gill cover is detected from the image sequence of the gill cover area, and the reciprocal of the duration is used as the gill cover opening and closing cycle. This scheme uses stereo vision to accurately locate the vertical movement amplitude of Chuzhou crucian carp, and triggers subsequent high-precision behavior data collection only when the fish exhibit continuous shallow swimming behavior. This effectively reduces the waste of computing resources and ensures that monitoring is only initiated in the early stage of infection (when fish usually swim at shallower depths), thus improving the timeliness of prediction.

[0007] As one technical solution of the present invention, the marked point on the body surface is the midpoint of the line connecting the origin of the dorsal fin and the origin of the pelvic fin of the Chuzhou crucian carp. This position is located in the middle of the fish's body, has good representativeness of movement, and is not easily obscured by the fins, making it easy for the binocular camera to track stably, thereby obtaining reliable vertical movement data.

[0008] As a technical solution of the present invention, the method of extracting the power spectral density distribution of the dorsal fin oscillation frequency over a time series using a sliding window and mapping the gill cover opening and closing cycle to a respiratory rhythm array specifically involves: arranging the dorsal fin oscillation frequency into a frequency time series according to the acquisition time; setting a rectangular window with a window length of 30 minutes; sliding the rectangular window with a step size of one minute; performing a fast Fourier transform on the frequency time series within each window to obtain the power spectral density distribution corresponding to each window, wherein the power spectral density distribution contains the correspondence between frequency and power spectral density; simultaneously, arranging the gill cover opening and closing cycle into a periodic time series according to the acquisition time; taking the reciprocal of the periodic time series to convert it into an instantaneous respiratory frequency; sampling the instantaneous respiratory frequency according to the median per minute to obtain a respiratory frequency value per minute; and arranging multiple consecutive respiratory frequency values ​​in chronological order to form the respiratory rhythm array. This method, by analyzing the spectral characteristics of the dorsal fin oscillation frequency through a sliding window, can capture energy fluctuation patterns in the short time domain. Simultaneously, converting the periodicity into a respiratory frequency and denoising and resampling yields a stable respiratory rhythm signal, providing a reliable time-domain reference for subsequent abnormal peak detection.

[0009] As a technical solution of the present invention, abnormal peaks whose amplitude deviates from the mean by more than two standard deviations are located in the respiratory rhythm array, and the energy attenuation of the corresponding frequency band in the power spectral density distribution is aligned with the occurrence time of the abnormal peaks as a reference. Specifically, the arithmetic mean of all respiratory frequency values ​​in the respiratory rhythm array is calculated as the rhythm mean; the square root of the average of the squares of the deviations of all respiratory frequency values ​​from the rhythm mean in the respiratory rhythm array is calculated as the rhythm standard deviation; and the deviations of each respiratory frequency value in the respiratory rhythm array from the rhythm mean by more than twice the rhythm standard deviation are traversed. Poor respiratory rate values ​​are marked as abnormal peaks, and the acquisition time corresponding to the abnormal peaks is recorded. Based on the acquisition time corresponding to the abnormal peaks, a rectangular window at the same time is located from the power spectral density distribution. In the power spectral density distribution of the located rectangular window, the frequency where the maximum power spectral density is located is determined as the main frequency band. The average power spectral density of the main frequency band in the current window is calculated, and then the average power spectral density of the main frequency band in the previous adjacent window is calculated. The difference between the two average power spectral density values ​​is taken as the energy attenuation. Specifically, the power spectral density distribution of the current window is denoted as... The power spectral density distribution of the previous adjacent window is denoted as Positioning the main frequency band range ,in This is the frequency corresponding to the maximum power spectral density in the current window. The preset half-bandwidth is used; the average power spectral density of the main band in the current window is calculated according to the following formula. :

[0010]

[0011] in: The total number of discrete frequency points within the main frequency band; calculate the average power spectral density of the main frequency band in the previous adjacent window using the same method. The energy attenuation amount Defined as:

[0012]

[0013] in: This represents the maximum value function, used to ensure that the energy attenuation is non-negative. This scheme uses statistical thresholds to automatically identify abnormal peaks in the respiratory rhythm (such as rapid breathing or intermittent cessation of breathing) and simultaneously extracts the energy attenuation of the dorsal fin oscillation main frequency band at the corresponding moment. This accurately quantifies the coupling relationship between respiratory disturbances and dorsal fin movement fatigue commonly seen in the early stages of Sporozoa infection, providing physically meaningful feature pairs for subsequent feature fusion.

[0014] As a technical solution of the present invention, the energy attenuation amount and the amplitude ratio of the abnormal peak are aligned at time points and then concatenated into a behavioral feature tensor. Specifically, for the acquisition time corresponding to the same abnormal peak, the difference between the respiratory rate value of the abnormal peak and the mean of the rhythm is obtained, and the ratio of the difference to the standard deviation of the rhythm is used as the amplitude ratio of the abnormal peak. The energy attenuation amount and the amplitude ratio are matched one-to-one on the timestamp to form a two-dimensional feature pair. Taking the occurrence time of the abnormal peak as the center of the time axis, all two-dimensional feature pairs generated within 15 minutes before and after are arranged in chronological order to obtain a feature sequence. Each two-dimensional feature pair in the feature sequence is sequentially filled into the first and second dimensions of a three-dimensional tensor. The third dimension of the three-dimensional tensor is used to distinguish between the two feature types: energy attenuation amount and amplitude ratio. This solution organizes multiple abnormal peaks and their associated energy attenuation data into a structured tensor, preserving temporal context information, enabling subsequent networks to model the temporal distribution pattern of abnormal events, and improving the perception ability of behavioral characteristics during the incubation period of spore-forming diseases.

[0015] As a technical solution of the present invention, the behavioral feature tensor is input into a graph attention network to output the risk probability of spore infection. Specifically, each abnormal peak in the three-dimensional tensor is taken as a node, and the reciprocal of the time interval between every two nodes is calculated as the weight of the edge to construct a temporal adjacency graph. The energy decay and amplitude ratio corresponding to each node in the three-dimensional tensor are concatenated to form the initial feature vector of that node. A multi-head graph attention layer is used to transform the initial feature vector of each node, and the attention coefficient between each node and its neighboring nodes is calculated. The features of the neighboring nodes are weighted and summed according to the attention coefficient to update the feature vector of each node. The updated feature vectors of all nodes are input into a global average pooling layer to obtain global graph features. The global graph features are mapped to a scalar through a fully connected layer, and the scalar is converted into the risk probability of spore infection through a normalized exponential function. This scheme automatically learns the temporal dependencies between abnormal peaks through graph attention networks, which can effectively extract global risk features even with a small number of abnormal events. At the same time, it uses a multi-head attention mechanism to enhance the model's ability to represent abnormal patterns at different time scales, and finally outputs a probability value of 0 to 1, which makes it easy for farmers to intuitively judge the level of infection risk.

[0016] As a technical solution of the present invention, each anomalous peak in the three-dimensional tensor is treated as a node, and the reciprocal of the time interval between any two nodes is calculated as the weight of the edge to construct a temporal adjacency graph. Specifically, the acquisition times corresponding to all anomalous peaks are extracted from the three-dimensional tensor, and the anomalous peaks are sorted according to their chronological order. Each sorted anomalous peak is assigned a node number. For any two sorted nodes, the difference between the acquisition time of the latter node and the acquisition time of the former node is calculated, and the reciprocal of the difference is used as the weight of the edge between the two nodes. Edges are established between each node and its three temporally adjacent nodes before and after it, forming a sparsely connected temporal adjacency graph, where the node set is all anomalous peaks, and the edge set is all established edges and their corresponding weights. This solution effectively reduces the complexity of the graph model through the sparse temporal adjacency graph, and at the same time, by using the reciprocal of the time interval as the weight, it makes temporally adjacent anomalous peaks more correlated, and can express the continuous changing trend of anomalous behavior gradually intensifying or alleviating during the infection process.

[0017] As a technical solution of the present invention, the multi-head graph attention layer transforms the initial feature vector of each node, calculates the attention coefficient of each node and its neighboring nodes, and performs a weighted summation of the features of neighboring nodes based on the attention coefficients to update the feature vector of each node. Specifically, for each node in the temporal adjacency graph, the initial feature vector of the node is concatenated with the initial feature vector of each neighboring node to obtain a concatenated vector. The concatenated vector is input into a single-layer feedforward neural network to obtain the original attention value of the node and each neighboring node. A normalized exponential function operation is performed on the original attention values ​​of all neighboring nodes to obtain the attention coefficient of the node and each neighboring node. The initial feature vector of each neighboring node is multiplied by the attention coefficient and then summed to obtain the aggregated feature vector of the node. The aggregated feature vector is passed through a nonlinear activation function to obtain the updated feature vector of the node. Each node uses four independent attention heads to perform the above process respectively. The four updated feature vectors obtained by the four attention heads are concatenated as the final updated feature vector of the node. This approach utilizes a multi-head attention mechanism to learn interaction patterns between nodes simultaneously from multiple representation subspaces, enhancing the model's ability to capture different attention emphases among anomalous peaks, thereby more comprehensively reflecting the complex behavioral anomalies induced by spore infection.

[0018] As a technical solution of the present invention, the updated feature vectors of all nodes are input into a global average pooling layer to obtain global graph features. The global graph features are mapped to a scalar through a fully connected layer. The scalar is then converted into the spore infection risk probability using a normalized exponential function. Specifically, the following steps are taken: the arithmetic mean of the updated feature vectors of all nodes in each feature dimension is calculated to obtain a one-dimensional vector as the global graph features; the one-dimensional vector is input into a fully connected network with two hidden layers, wherein the first hidden layer uses a rectified linear unit activation function and the second hidden layer does not use an activation function, outputting a real value as the scalar; the scalar is input as an independent variable into a normalized exponential function, and the exponential function value with the natural constant as the base and the scalar as the exponent is calculated. The exponential function value is divided by one plus the sum of the exponential function values ​​to obtain a value between zero and one as the spore infection risk probability. Specifically, the scalar output by the fully connected network is... The probability of Sporozoan infection risk is calculated using the following formula. :

[0019]

[0020] in: It is a natural constant. Indicates a base of the natural constant. The value of the exponential function of the exponent, when When it is determined that there is a risk of sporozoan infection, The system is identified as having a risk of infection with Aspergillus subtilis. This approach integrates information from all abnormal peaks through global average pooling, eliminating the impact of node number fluctuations on the final output dimension. This allows the model to handle monitoring data of different durations. Furthermore, the use of a normalized exponential function to output probability values ​​facilitates the setting of clear decision thresholds, improving the operability and reliability of the prediction results.

[0021] As a technical solution of the present invention, after outputting the probability of spore infection risk, the method further includes: comparing the probability of spore infection risk with a preset risk decision threshold; when the probability of spore infection risk exceeds a first risk decision threshold, generating a first decision instruction to feed an immune enhancer, the first decision instruction including the type and dosage of the immune enhancer to be fed; when the probability of spore infection risk exceeds a second risk decision threshold and the second risk decision threshold is greater than the first risk decision threshold, generating a second decision instruction to start a full water exchange of the circulating water system, the second decision instruction including the water exchange rate and target water temperature; when the probability of spore infection risk is lower than a third risk decision threshold and remains lower than for more than six hours, generating a third decision instruction to terminate all current decision instructions and restore normal aquaculture management, the third decision instruction including the restored feeding amount and water exchange cycle. This solution automatically triggers aquaculture intervention measures according to the risk assessment probability: only immune enhancement is performed at low risk to avoid overtreatment; large-scale water exchange is initiated at medium and high risk to block pathogen transmission; normal management is automatically restored after the risk is eliminated, realizing intelligent phased prevention and control of spore diseases, effectively reducing aquaculture losses and improving management efficiency.

[0022] The beneficial effects of this invention are:

[0023] A sliding window approach was used to extract the power spectral density distribution of the dorsal fin oscillation frequency, and the gill opening and closing cycle was mapped to a respiratory rhythm array. Abnormal peaks with amplitudes deviating from the mean by more than two standard deviations were located within the respiratory rhythm array. Using the occurrence time of these abnormal peaks as a baseline, the energy attenuation of the corresponding frequency band in the power spectral density distribution was aligned. The ratio of the energy attenuation to the amplitude of the abnormal peak was then time-aligned and concatenated into a behavioral feature tensor. This scheme cross-modal alignment of the frequency domain features of the dorsal fin oscillation frequency with the temporal domain abnormal peaks of the respiratory rhythm enables the capture of the temporal causal relationship between "changes in dorsal fin vibration patterns" and "disordered respiratory rhythms" observed in fish during the incubation period of sporozoan infection. This allows for the accurate extraction of synergistic abnormal signals that traditional single indicators cannot reflect, significantly improving the sensitivity of behavioral features to the subclinical infection stage and reducing false alarms caused by environmental disturbances. This approach models temporally discrete anomalous peak events as a graph structure and automatically learns the dependency weights between different anomalous events using an attention mechanism. It overcomes the limitation of conventional algorithms that treat behavioral features as independent samples and ignore temporal proximity. This allows for the inference of the dynamic evolution trend of the current infection risk from historical anomalous events, providing a high-confidence risk probability even in the early stages of anomalous infection when fish show no obvious symptoms, thus providing a reliable basis for subsequent differentiated decision-making. The method employs a multi-head graph attention layer to weight the node features and outputs global graph features, which are then mapped to the spore-borne parasite infection risk probability through a fully connected layer. Attached Figure Description

[0024] The invention will now be further described with reference to the accompanying drawings.

[0025] Figure 1 This is a flowchart of the sporozoan disease prediction and decision support method based on Chuzhou crucian carp behavioral data as described in this invention;

[0026] Figure 2 This is a flowchart of the swimming depth triggering and physiological parameter acquisition of Chuzhou crucian carp;

[0027] Figure 3 This is a flowchart of the generation of the dorsal fin swing power spectral density and respiratory rhythm array. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] See Figure 1 This invention provides a method for predicting and supporting decision-making regarding sporozoan diseases based on behavioral data of Chuzhou crucian carp. The method includes: responding to a trigger signal that the swimming trajectory depth data of Chuzhou crucian carp in the aquaculture environment is below a depth threshold; collecting the dorsal fin oscillation frequency and gill cover opening and closing cycle of marked individuals; extracting the power spectral density distribution of the dorsal fin oscillation frequency over a time series using a sliding window, and mapping the gill cover opening and closing cycle to a respiratory rhythm array; locating abnormal peaks in the respiratory rhythm array whose amplitude deviates from the mean by more than two standard deviations, and aligning the energy attenuation of the corresponding frequency band in the power spectral density distribution with the occurrence time of the abnormal peak as a reference; concatenating the ratio of the energy attenuation to the amplitude of the abnormal peak over time to form a behavioral feature tensor; and inputting the behavioral feature tensor into a graph attention network to output the probability of sporozoan infection risk.

[0030] Example 1: In specific implementation, in response to a trigger signal indicating that the swimming trajectory depth data of Chuzhou crucian carp in the aquaculture environment is below a depth threshold, the dorsal fin oscillation frequency and gill cover opening and closing cycle of the marked individuals are collected. (See also...) Figure 2 Multiple underwater binocular cameras are deployed at the bottom of the breeding pond to continuously acquire the coordinate sequence of surface markers for each Chuzhou crucian carp in three-dimensional space. The surface markers are located at the midpoint of the line connecting the origin of the dorsal fin and the origin of the pelvic fin. The maximum and minimum values ​​of the surface markers in the vertical direction are extracted from the coordinate sequence, and the difference between the maximum and minimum values ​​is calculated as the swimming trajectory depth data. When the swimming trajectory depth data is below a preset depth threshold for three consecutive sampling periods, a trigger signal is generated. After the trigger signal is generated, dorsal fin edge contour tracking of the same Chuzhou crucian carp is initiated, and the number of dorsal fin swings per unit time is extracted from the dorsal fin edge contour tracking results as the dorsal fin swing frequency. Simultaneously, the duration from complete opening to complete closing of the gill cover is detected from the image sequence of the gill cover area, and the reciprocal of this duration is taken as the gill cover opening and closing cycle.

[0031] In practice, the underwater binocular cameras are pre-calibrated with internal and external parameters. Each camera, in conjunction with adjacent cameras, forms a stereoscopic vision covering the entire bottom area of ​​the aquaculture pond. Images are synchronously acquired through the underwater binocular cameras, and a binocular stereo matching algorithm is used to calculate the coordinate values ​​of the surface markers of the Chuzhou crucian carp in three-dimensional space in each frame. The coordinate values ​​of multiple consecutive frames are arranged in chronological order to form the coordinate sequence. The surface markers are located at the midpoint of the line connecting the origin of the dorsal fin and the origin of the pelvic fin of the Chuzhou crucian carp. This position has a relatively fixed geometric relationship on the fish's body, facilitating binocular camera identification and tracking.

[0032] In specific implementation, the maximum value of the surface marker point in the vertical direction is extracted from the coordinate sequence, i.e., the maximum value of the z-axis component of all coordinates; the minimum value of the surface marker point in the vertical direction is also extracted, i.e., the minimum value of the z-axis component of all coordinates. The difference between the maximum and minimum values ​​is calculated, and this difference is the swimming trajectory depth data. The depth threshold is a value preset according to the normal swimming depth range of the Chuzhou crucian carp. The sampling period is the frame interval time of the underwater binocular camera acquiring images, and the three consecutive sampling periods represent three consecutive frame interval times. When the swimming trajectory depth data is less than the depth threshold in all three consecutive sampling periods, the trigger signal is generated.

[0033] In specific implementation, when initiating dorsal fin edge contour tracking, image processing algorithms are used to perform edge detection on the dorsal fin region of the Chuzhou crucian carp in the images acquired by the underwater binocular camera, obtaining a set of contour points of the dorsal fin edge. The number of times the dorsal fin edge contour swings vertically within a unit of time (e.g., one minute) is counted; that is, a complete swing is defined as the dorsal fin edge swinging from one side to the other and back. This number of swings is recorded as the dorsal fin swing frequency. Simultaneously, from the image sequence of the gill cover region, the opening and closing states of the gill cover edge are detected, and the duration of the change from a fully open state to a fully closed state is recorded. The reciprocal of this duration is taken to obtain the gill cover opening and closing period. Both the dorsal fin swing frequency and the gill cover opening and closing period are continuously acquired data for the same marked individual Chuzhou crucian carp.

[0034] Example 2: In specific implementation, refer to Figure 3 A sliding window method was used to extract the power spectral density distribution of the dorsal fin oscillation frequency over a time series, and the gill cover opening and closing cycle was mapped to a respiratory rhythm array. The dorsal fin oscillation frequencies were arranged into a frequency time series according to the acquisition time. A rectangular window with a window length of 30 minutes was set, and the window was slid in a step of one minute. A Fast Fourier Transform was performed on the frequency time series within each window to obtain the power spectral density distribution corresponding to each window. The power spectral density distribution contained the correspondence between frequency and power spectral density. Simultaneously, the gill cover opening and closing cycle was arranged into a periodic time series according to the acquisition time. The reciprocal of the periodic time series was taken to convert it into an instantaneous respiratory frequency. The instantaneous respiratory frequency was sampled according to the median per minute to obtain a respiratory frequency value per minute. Multiple consecutive respiratory frequency values ​​were arranged in chronological order to form the respiratory rhythm array.

[0035] In practice, the frequency time series is composed of the dorsal fin oscillation frequency values ​​corresponding to each acquisition moment, arranged chronologically. The window length of the rectangular window is set to thirty minutes, which controls the data time span covered by one Fast Fourier Transform (FFT). The step size of the rectangular window is set to one minute, meaning the rectangular window slides forward once on the frequency time series every minute, resulting in a twenty-nine-minute overlap between adjacent windows. A FFT is performed on the frequency time series data within each window, and the transformation result is the power spectral density distribution corresponding to that window, with frequency as the independent variable and power spectral density as the dependent variable. The power spectral density distribution of each window is stored independently.

[0036] In practice, the gill opening and closing cycles are arranged into a periodic time series according to the collection time. The reciprocal of each cycle value in the periodic time series is taken to obtain the instantaneous respiratory rate corresponding to that moment. Since the unit of the gill opening and closing cycle is seconds / cycle, the unit of its reciprocal is cycles / second, i.e., the instantaneous respiratory rate. All instantaneous respiratory rate values ​​within each minute are sorted, and the median of that minute is taken as the respiratory rate value for that minute. In this way, the temporally continuous instantaneous respiratory rates are converted into a respiratory rate value per minute, i.e., one respiratory rate value is obtained per minute. The respiratory rate values ​​obtained from multiple consecutive minutes are arranged in chronological order to form the respiratory rhythm array. Each element in the respiratory rhythm array corresponds to a respiratory rate value for one minute, and the array length is determined by the total number of minutes of data collection.

[0037] Example 3: In a specific implementation, abnormal peaks whose amplitude deviates from the mean by more than two standard deviations are located in the respiratory rhythm array. The energy attenuation of the corresponding frequency band in the power spectral density distribution is aligned with the time of occurrence of the abnormal peak. The arithmetic mean of all respiratory frequency values ​​in the respiratory rhythm array is calculated as the rhythm mean. The square root of the average of the squared deviations of all respiratory frequency values ​​from the rhythm mean is calculated as the rhythm standard deviation. Each respiratory frequency value in the respiratory rhythm array is traversed, and respiratory frequency values ​​deviating from the rhythm mean by more than twice the rhythm standard deviation are marked as abnormal peaks. The acquisition time corresponding to the abnormal peak is recorded. Based on the acquisition time corresponding to the abnormal peak, a rectangular window at the same time is located in the power spectral density distribution. In the power spectral density distribution of the located rectangular window, the frequency where the maximum power spectral density is located is determined as the main frequency band. The average power spectral density of the main frequency band in the current window is calculated, and then the average power spectral density of the main frequency band in the previous adjacent window is calculated. The difference between the two average power spectral density values ​​is taken as the energy attenuation. Let the power spectral density distribution of the current window be denoted as The power spectral density distribution of the previous adjacent window is denoted as Positioning the main frequency band range ,in This is the frequency corresponding to the maximum power spectral density in the current window. The preset half-bandwidth is used. The average power spectral density of the main band in the current window is calculated using the following formula. :

[0038]

[0039] in: This represents the total number of discrete frequency points within the main frequency band. The average power spectral density of the main frequency band for the previous adjacent window is calculated using the same method. The energy attenuation amount Defined as:

[0040]

[0041] in: This represents the maximum value function, used to ensure that the energy decay is non-negative.

[0042] In specific implementation, the rhythm mean is obtained by summing all respiratory rate values ​​in the respiratory rhythm array and dividing by the total number of respiratory rate values. The calculation steps for the rhythm standard deviation are as follows: for each respiratory rate value in the respiratory rhythm array, calculate the difference between it and the rhythm mean, square the difference, sum the results, divide by the total number of respiratory rate values, and take the square root of the result. When traversing the respiratory rhythm array, for each respiratory rate value, determine whether the respiratory rate value satisfies the following condition: the absolute value of the difference between the respiratory rate value and the rhythm mean is greater than twice the rhythm standard deviation. If the condition is met, the respiratory rate value is marked as an abnormal peak, and the acquisition time corresponding to the respiratory rate value is recorded. The acquisition time is the timestamp of the element in the respiratory rhythm array.

[0043] In specific implementation, based on the acquisition time corresponding to the abnormal peak, a rectangular window is located from the power spectral density distribution set. The start time of this rectangular window includes the acquisition time, meaning the acquisition time falls within the time range of this rectangular window. Within the located rectangular window's power spectral density distribution, the frequency point with the highest power spectral density value is searched and recorded as... The half-bandwidth This is a preset value, ranging from 0.1Hz to 0.5Hz. Within the main frequency band... Within the current window, calculate the average power spectral density. The specific calculation method is as follows: calculate the power spectral density values ​​at all discrete frequency points within the main frequency band. Summing is performed, and then divided by the total number of discrete frequency points within the main frequency band. . The value of depends on the frequency resolution and is determined by the Fast Fourier Transform parameters. For the preceding adjacent window, i.e., the rectangular window that is temporally immediately preceding the current window, the same dominant frequency band is used. Calculate its average power spectral density The main frequency band range is based on the current window. This is confirmed and will not change depending on the window. Finally, calculate. minus The difference is taken as 0 if the difference is negative. If the difference is positive, then the difference is taken as the energy attenuation amount. By using the maximum value function Perform this operation.

[0044] Example 4: In specific implementation, the energy attenuation amount and the amplitude ratio of the abnormal peak are aligned by time points and then concatenated into a behavioral feature tensor. For the same acquisition time corresponding to the same abnormal peak, the difference between the respiratory rate value of the abnormal peak and the mean of the rhythm is obtained, and the ratio of the difference to the standard deviation of the rhythm is taken as the amplitude ratio of the abnormal peak. The energy attenuation amount and the amplitude ratio are matched one-to-one on the timestamp to form a two-dimensional feature pair. Taking the occurrence time of the abnormal peak as the center of the time axis, all two-dimensional feature pairs generated within 15 minutes before and after are arranged in chronological order to obtain a feature sequence. Each two-dimensional feature pair in the feature sequence is sequentially filled into the first and second dimensions of a three-dimensional tensor. The third dimension of the three-dimensional tensor is used to distinguish between the two feature types: energy attenuation amount and amplitude ratio.

[0045] In practice, for respiratory rate values ​​marked as abnormal peaks, the acquisition time corresponding to the abnormal peak is obtained. The respiratory rate value at that time is read from the respiratory rhythm array and recorded as follows. The mean value of the rhythm is obtained from the calculation results of the respiratory rhythm array. and the rhythm standard deviation The amplitude ratio of the abnormal peaks is calculated according to the following formula. :

[0046]

[0047] in: The respiratory rate value of the abnormal peak. This is the arithmetic mean of all respiratory rate values ​​in the respiratory rhythm array. This represents the rhythm standard deviation of all respiratory rate values ​​in the respiratory rhythm array. The amplitude ratio... The degree to which the abnormal peak deviates from the rhythm mean is expressed as a multiple of the rhythm standard deviation.

[0048] In specific implementation, the energy attenuation amount The ratio of the amplitude They are paired according to the same acquisition time. Each acquisition time corresponds to a two-dimensional feature pair, which consists of two values: the first value is the energy attenuation, and the second value is the amplitude ratio. The energy attenuation and the amplitude ratio are in one-to-one correspondence on the timestamp, that is, each acquisition time corresponding to an abnormal peak uniquely determines an energy attenuation and an amplitude ratio, which together constitute a two-dimensional feature pair.

[0049] In specific implementation, the occurrence time of each anomalous peak is taken as the center point of the time axis, and a time window of 30 minutes is defined, consisting of 15 minutes before and after the center point. This time window may include acquisition times corresponding to other anomalous peaks, as well as acquisition times corresponding to non-anomalous peaks. All two-dimensional feature pairs located within this time window are arranged in chronological order of their acquisition times to form a feature sequence. The length of the feature sequence is equal to the number of two-dimensional feature pairs contained within the time window. For each two-dimensional feature pair in the feature sequence, its first value (energy attenuation) and second value (amplitude ratio) are filled into corresponding positions in a three-dimensional tensor. The first dimension of the three-dimensional tensor corresponds to the index order (i.e., chronological order) of the two-dimensional feature pairs in the feature sequence, the second dimension corresponds to the two values ​​within each two-dimensional feature pair (i.e., energy attenuation and amplitude ratio occupy two different positions), and the third dimension is used to distinguish between the two feature types: energy attenuation and amplitude ratio. Specifically, the shape of the three-dimensional tensor is as follows: ,in The first dimension index is the number of two-dimensional feature pairs in the feature sequence. Corresponding to the Two-dimensional feature pairs, second-dimensional index exist The corresponding energy attenuation amount, Corresponding to the amplitude ratio, the third dimension is used for redundant representation of feature types (for example, third dimension indices 0 and 1 correspond to two feature types respectively, but the second dimension has already distinguished them during actual filling). Alternatively, the shape of the three-dimensional tensor is... The first dimension index corresponds to the time order, the second dimension index corresponds to the two numerical values ​​of each two-dimensional feature pair, and the third dimension index corresponds to the two feature label dimensions. The specific padding method is as follows: for the first... A pair of two-dimensional features is used to assign the energy decay amount to a tensor element. and The amplitude ratio is assigned to the tensor element. and This ensures that the two positions in the third dimension have the same value. After filling all two-dimensional feature pairs, the behavioral feature tensor is obtained. This behavioral feature tensor is then used as input to the graph attention network.

[0050] Example 5: In a specific implementation, the behavioral feature tensor is input into a graph attention network to output the probability of spore infection risk. Each anomalous peak in the three-dimensional tensor is treated as a node, and the reciprocal of the time interval between any two nodes is used as the weight of the edge to construct a temporal adjacency graph. The acquisition times corresponding to all anomalous peaks are extracted from the three-dimensional tensor, and the anomalous peaks are sorted chronologically, with each sorted anomalous peak assigned a node number. For any two sorted nodes, the difference between the acquisition time of the latter node and the acquisition time of the former node is calculated, and the reciprocal of this difference is used as the weight of the edge between these two nodes. Edges are established between each node and its three temporally adjacent nodes before and after it, forming a sparsely connected temporal adjacency graph, where the node set represents all anomalous peaks, and the edge set represents all established edges and their corresponding weights. The energy attenuation and amplitude ratio corresponding to each node in the three-dimensional tensor are concatenated to form the initial feature vector of that node. A multi-head graph attention layer transforms the initial feature vector of each node, calculates the attention coefficient between each node and its neighboring nodes, and weights and sums the features of neighboring nodes based on the attention coefficients to update the feature vector of each node. For each node in the temporal adjacency graph, the initial feature vector of the node is concatenated with the initial feature vectors of each neighboring node to obtain a concatenated vector. This concatenated vector is input into a single-layer feedforward neural network to obtain the original attention value between the node and each neighboring node. A normalized exponential function operation is performed on the original attention values ​​of all neighboring nodes to obtain the attention coefficient between the node and each neighboring node. The initial feature vector of each neighboring node is multiplied by the attention coefficient and then summed to obtain the aggregated feature vector of the node. The aggregated feature vector is passed through a non-linear activation function to obtain the updated feature vector of the node. Each node uses four independent attention heads to perform the above process, and the four updated feature vectors obtained from the four attention heads are concatenated as the final updated feature vector of the node. The updated feature vectors of all nodes are input into a global average pooling layer to obtain global graph features. These global graph features are then mapped to a scalar through a fully connected layer. The scalar is then converted into the spore infection risk probability using a normalized exponential function. The arithmetic mean of the updated feature vectors of all nodes across each feature dimension is calculated to obtain a one-dimensional vector as the global graph feature. This one-dimensional vector is input into a fully connected network with two hidden layers. The first hidden layer uses a rectified linear unit activation function, while the second hidden layer does not use an activation function, outputting a real value as the scalar. The scalar is then input as an independent variable into a normalized exponential function. The exponential function value, with the natural constant as the base and the scalar as the exponent, is calculated. The exponential function value is divided by one plus the sum of the exponential function values, yielding a value between zero and one as the spore infection risk probability.Let the scalar output of the fully connected network be denoted as . The probability of Sporozoan infection risk is calculated using the following formula. :

[0051]

[0052] in: It is a natural constant. Indicates a base of the natural constant. The value of the exponential function of the exponent, when When it is determined that there is a risk of sporozoan infection, The system is determined to be free of sporozoan infection risk. After the sporozoan infection risk probability is output, it is compared with a preset risk decision threshold. When the sporozoan infection risk probability exceeds a first risk decision threshold, a first decision instruction to feed an immune enhancer is generated, which includes the type and dosage of the immune enhancer to be fed. When the sporozoan infection risk probability exceeds a second risk decision threshold and the second risk decision threshold is greater than the first risk decision threshold, a second decision instruction to start a full water change of the circulating water system is generated, which includes the water change rate and target water temperature. When the sporozoan infection risk probability is lower than a third risk decision threshold and remains below it for more than six hours, a third decision instruction to terminate all current decision instructions and restore normal aquaculture management is generated, which includes restoring the feeding amount and water change cycle.

[0053] In specific implementation, the construction process of the temporal adjacency graph is as follows: The acquisition time corresponding to each anomalous peak is extracted from the three-dimensional tensor, and the acquisition time is recorded in the form of a timestamp. All anomalous peaks are sorted according to the order of their timestamps, and each anomalous peak is assigned an integer starting from 1 as a node number. For any two different nodes, the time difference between the acquisition time of the latter node and the acquisition time of the former node is calculated, in seconds. The reciprocal of the time difference is used as the weight of the edge between these two nodes, defined as... ,in and Let be the acquisition times of node i and node j, respectively, and The edge connection rule is as follows: For each node, establish a bidirectional edge between it and its three preceding and following nodes in time. For node i, connect the preceding nodes... (If it exists) Connect the nodes backward. (If they exist). All established edges constitute the edge set, with each edge having a corresponding weight. The node set consists of all anomalous peaks, and the edge set consists of all established edges and their weights, forming the temporal adjacency graph.

[0054] In practice, the initial feature vector of each node is formed by concatenating the energy attenuation and amplitude ratio corresponding to that node in the three-dimensional tensor. For the anomalous peak corresponding to a node, the two-dimensional feature pair corresponding to that node, namely the energy attenuation value and the amplitude ratio value, is extracted from the first and second dimensions of the three-dimensional tensor. These two values ​​are then concatenated in sequence to form a one-dimensional vector of length 2, which serves as the initial feature vector of that node. .

[0055] In specific implementation, the multi-head graph attention layer comprises four independent attention heads. For each attention head, the following operation is performed: for each node in the temporal adjacency graph... Get the set of all its neighboring nodes. For each adjacent node , will node initial feature vector With nodes initial feature vector Concatenate the vectors to obtain the concatenated vector. ,in This represents a vector concatenation operation, resulting in a concatenated vector of length 4 (because each initial feature vector has a length of 2). The concatenated vector is then input into a single-layer feedforward neural network, which consists of a linear transformation layer and a non-linear activation function. The linear transformation layer maps the concatenated vector of length 4 to a scalar, resulting in nodes. With nodes raw attention values ​​between The original attention value The calculation formula is: ,in The weight matrix is ​​trainable and has dimensions of . (Transform the input features from 2D to 2D). The attention vector is trainable and has a dimension of 4. LeakyReLU is a leaky rectified linear unit activation function with a negative slope parameter set to 0.2. For nodes... All neighboring nodes its original attention value Perform the normalized exponential function (softmax) operation to obtain the nodes. With each adjacent node Attention coefficient :

[0056]

[0057] in: This represents an exponential function. Then, each adjacent node... initial feature vector Multiply by the corresponding attention coefficient Summing the products of all adjacent nodes yields the node. Aggregated feature vectors The aggregated feature vector is passed through a non-linear activation function (ELU activation function) to obtain the nodes. Feature vector updated under this attention head Repeat the above process, and each of the four attention heads generates four updated feature vectors, each with a dimension of 2. Concatenate these four feature vectors to obtain the node. Final updated feature vector The dimension is 8 (2 times 4).

[0058] In practice, the updated feature vectors of all nodes are... The input is fed into a global average pooling layer. The global average pooling layer calculates the arithmetic mean of all nodes across each feature dimension. Assume the total number of nodes is... For the first Each feature dimension ( ), calculate the average of the feature values ​​of all nodes in this dimension: Combine the average values ​​of all dimensions into a one-dimensional vector. , as a feature of the global graph.

[0059] In its implementation, the fully connected network has two hidden layers. The first hidden layer contains 16 neurons and uses a Rectified Linear Unit (ReLU) as the activation function; the second hidden layer contains one neuron and does not use an activation function (i.e., linear output). The global graph features are then used... The input is fed into the first hidden layer to obtain a 16-dimensional intermediate feature vector, which is then mapped to a real-valued scalar through the second hidden layer. .Will Input it into the normalized exponential function (sigmoid function), according to the formula. Calculate the probability of Sporozoan infection risk ,in is a natural constant, approximately equal to 2.71828. When When it is determined that there is a risk of sporozoan infection, It was determined to be at risk of Asporidium infection.

[0060] In specific implementation, the first risk decision threshold is a preset value, ranging from 0.6 to 0.7. When the probability of infection by the sporozoites exceeds the first risk decision threshold, a first decision instruction to administer an immune enhancer is generated. This first decision instruction includes the type and dosage of the immune enhancer to be administered. The immune enhancer is selected from one of vitamin C, β-glucan, or yeast culture, and the dosage is calculated based on the volume of the aquaculture pond and the weight of the Chuzhou crucian carp. The second risk decision threshold is a preset value, ranging from 0.8 to 0.9, and is greater than the first risk decision threshold. When the probability of infection by the sporozoites exceeds the second risk decision threshold, a second decision instruction to initiate a full water exchange of the circulating water system is generated. This second decision instruction includes the water exchange rate and the target water temperature. The water exchange rate is set to a total water exchange volume equal to 1 / 3 of the aquaculture pond volume per hour, and the target water temperature is set to be 2 degrees Celsius lower than the current water temperature. The third risk decision threshold is a preset value, ranging from 0.3 to 0.4. When the probability of infection by the sporozoites is lower than the third risk decision threshold for more than six hours, a third decision instruction is generated to terminate all current decision instructions and restore normal aquaculture management. The third decision instruction includes the restored feeding amount and water change cycle. The restored feeding amount is the normal daily feeding amount (e.g., 2% of the fish's body weight), and the restored water change cycle is one water change every three days, with each water change being 1 / 5 of the aquaculture pond volume.

[0061] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for predicting and supporting decision support for sporozoan diseases based on behavioral data of crucian carp in Chuzhou, characterized in that, include: In response to the trigger signal that the swimming trajectory depth data of Chuzhou crucian carp in the aquaculture environment is lower than the depth threshold, the dorsal fin oscillation frequency and gill cover opening and closing cycle of the marked individuals were collected. A sliding window was used to extract the power spectral density distribution of the dorsal fin oscillation frequency over time, and the gill cover opening and closing cycle was mapped to a respiratory rhythm array. Locate abnormal peaks in the respiratory rhythm array whose amplitudes deviate from the mean by more than two standard deviations, and align the energy attenuation of the corresponding frequency band in the power spectral density distribution with the occurrence time of the abnormal peaks as a reference. After aligning the ratio of the energy attenuation to the amplitude of the abnormal peak at specific times, the two values ​​are concatenated into a behavioral feature tensor. The behavioral feature tensor is input into a graph attention network, which outputs the probability of spore infection risk.

2. The method for predicting and supporting sporozoan diseases based on Chuzhou crucian carp behavioral data according to claim 1, characterized in that, The trigger signal, which responds to the swimming trajectory depth data of Chuzhou crucian carp in the aquaculture environment being lower than a depth threshold, collects the dorsal fin oscillation frequency and gill cover opening and closing cycle of the marked individuals, specifically: Multiple underwater binocular cameras are deployed at the bottom of the breeding pond to continuously acquire the coordinate sequence of the body surface markers of each Chuzhou crucian carp in three-dimensional space. Extract the maximum and minimum values ​​of the body surface markers in the vertical direction from the coordinate sequence, and calculate the difference between the maximum and minimum values ​​as the swimming trajectory depth data; When the swimming trajectory depth data is lower than the preset depth threshold for three consecutive sampling periods, the trigger signal is generated, and the dorsal fin edge contour tracking of the same Chuzhou crucian carp is started. The number of dorsal fin swings per unit time is extracted from the dorsal fin edge contour tracking results as the dorsal fin swing frequency. At the same time, the duration from when the gill cover is fully open to when it is fully closed is detected from the image sequence of the gill cover area, and the reciprocal of the duration is taken as the gill cover opening and closing cycle.

3. The method for predicting and supporting sporozoan diseases based on Chuzhou crucian carp behavioral data according to claim 2, characterized in that, The marked point on the body surface is the midpoint of the line connecting the origin of the dorsal fin and the origin of the pelvic fin of the Chuzhou crucian carp.

4. The method for predicting and supporting sporozoan diseases based on Chuzhou crucian carp behavioral data according to claim 2, characterized in that, The process involves extracting the power spectral density distribution of the dorsal fin oscillation frequency over a time series using a sliding window, and mapping the gill cover opening and closing cycle to a respiratory rhythm array. Specifically: The dorsal fin oscillation frequency is arranged into a frequency time series according to the acquisition time. A rectangular window with a window length of 30 minutes is set. The rectangular window is slid with a step size of one minute. A fast Fourier transform is performed on the frequency time series in each window to obtain the power spectral density distribution corresponding to each window. The power spectral density distribution contains the correspondence between frequency and power spectral density. Simultaneously, the gill cover opening and closing cycle is arranged into a periodic time series according to the collection time. The reciprocal of the periodic time series is taken to convert it into an instantaneous respiratory rate. The instantaneous respiratory rate is sampled according to the median per minute to obtain a respiratory rate value per minute. Multiple consecutive respiratory rate values ​​are arranged in chronological order to form the respiratory rhythm array.

5. The method for predicting and supporting sporozoan diseases based on Chuzhou crucian carp behavioral data according to claim 4, characterized in that, Locate abnormal peaks in the respiratory rhythm array whose amplitudes deviate from the mean by more than two standard deviations, and align the energy attenuation of the corresponding frequency band in the power spectral density distribution with the occurrence time of the abnormal peaks as a reference. Specifically: The arithmetic mean of all respiratory rate values ​​in the respiratory rhythm array is calculated as the rhythm mean. The square root of the average of the squares of the deviations of all respiratory rate values ​​from the rhythm mean is calculated as the rhythm standard deviation. Each respiratory rate value in the respiratory rhythm array is traversed, and respiratory rate values ​​that deviate from the rhythm mean by more than twice the rhythm standard deviation are marked as abnormal peaks. The acquisition time corresponding to the abnormal peak is recorded. Based on the acquisition time corresponding to the abnormal peak, a rectangular window at the same time is located from the power spectral density distribution. In the power spectral density distribution of the located rectangular window, the frequency where the maximum power spectral density is located is determined as the main frequency band. The average power spectral density of the main frequency band in the current window is calculated, and then the average power spectral density of the main frequency band in the previous adjacent window is calculated. The difference between the two average power spectral density values ​​is taken as the energy attenuation.

6. The method for predicting and supporting sporozoan diseases based on Chuzhou crucian carp behavioral data according to claim 5, characterized in that, After aligning the ratio of the energy attenuation to the amplitude of the abnormal peak over time, they are concatenated into a behavioral feature tensor, specifically: For the same abnormal peak at the same acquisition time, the difference between the respiratory rate value of the abnormal peak and the mean of the rhythm is obtained, and the ratio of the difference to the standard deviation of the rhythm is used as the amplitude ratio of the abnormal peak. The energy attenuation and the amplitude ratio are mapped one-to-one on the timestamp to form a two-dimensional feature pair; Using the occurrence time of the abnormal peak as the center of the time axis, all two-dimensional feature pairs generated within 15 minutes before and after are arranged in chronological order to obtain a feature sequence. Each two-dimensional feature pair in the feature sequence is then filled into the first and second dimensions of a three-dimensional tensor. The third dimension of the three-dimensional tensor is used to distinguish between two feature types: energy attenuation and amplitude ratio.

7. The method for predicting and supporting sporozoan diseases based on Chuzhou crucian carp behavioral data according to claim 6, characterized in that, The behavioral feature tensor is input into a graph attention network, which outputs the probability of spore infection risk, specifically: Each anomalous peak in the three-dimensional tensor is treated as a node, and the reciprocal of the time interval between every two nodes is calculated as the weight of the edge to construct a temporal adjacency graph. The energy decay and amplitude ratio corresponding to each node in the three-dimensional tensor are concatenated to form the initial feature vector of that node; A multi-head graph attention layer is used to transform the initial feature vector of each node, calculate the attention coefficient between each node and its neighboring nodes, and perform a weighted summation of the features of the neighboring nodes based on the attention coefficients to update the feature vector of each node. The updated feature vectors of all nodes are input into a global average pooling layer to obtain global graph features. The global graph features are then mapped to a scalar through a fully connected layer. The scalar is then converted into the probability of infection risk of the spore-forming organism using a normalized exponential function.

8. The method for predicting and supporting sporozoan diseases based on Chuzhou crucian carp behavioral data according to claim 7, characterized in that, Each anomalous peak in the three-dimensional tensor is treated as a node, and the reciprocal of the time interval between any two nodes is calculated as the weight of the edge. A temporal adjacency graph is constructed as follows: Extract the acquisition time corresponding to all abnormal peaks from the three-dimensional tensor, sort the abnormal peaks in chronological order, and assign a node number to each sorted abnormal peak. For any two sorted nodes, calculate the difference between the acquisition time of the latter node and the acquisition time of the former node, and use the reciprocal of the difference as the weight of the edge between the two nodes. For each node, edges are established between it and the three nodes that are temporally adjacent to it, forming a sparsely connected temporal adjacency graph. The node set consists of all abnormal peaks, and the edge set consists of all established edges and their corresponding weights.

9. The method for predicting and supporting sporozoan diseases based on Chuzhou crucian carp behavioral data according to claim 8, characterized in that, The process involves using a multi-head graph attention layer to transform the initial feature vector of each node, calculating the attention coefficient between each node and its neighboring nodes, and then weighted summing the features of neighboring nodes based on these attention coefficients to update the feature vector of each node. Specifically: For each node in the temporal adjacency graph, the initial feature vector of the node is concatenated with the initial feature vector of each neighboring node to obtain a concatenated vector. The concatenated vector is then input into a single-layer feedforward neural network to obtain the original attention values ​​of the node and each neighboring node. Perform a normalized exponential function operation on the original attention values ​​of all neighboring nodes to obtain the attention coefficients between the node and each of its neighboring nodes; The initial feature vector of each adjacent node is multiplied by the attention coefficient and then summed to obtain the aggregated feature vector of that node. The aggregated feature vector is then passed through a non-linear activation function to obtain the updated feature vector of that node. Each node uses four independent attention heads to perform the above process respectively. The four updated feature vectors obtained by the four attention heads are concatenated to obtain the final updated feature vector of that node.

10. The method for predicting and supporting sporozoan diseases based on Chuzhou crucian carp behavioral data according to claim 7, characterized in that, The updated feature vectors of all nodes are input into a global average pooling layer to obtain global graph features. These global graph features are then mapped to a scalar through a fully connected layer. Finally, the scalar is converted into the probability of spore infection risk using a normalized exponential function. Calculate the arithmetic mean of the updated feature vectors of all nodes in each feature dimension to obtain a one-dimensional vector as the global graph feature; The one-dimensional vector is input into a fully connected network with two hidden layers, where the first hidden layer uses the rectified linear unit activation function and the second hidden layer does not use an activation function, and outputs a real value as the scalar. The scalar is input as an independent variable into the normalized exponential function, and the exponential function value with the natural constant as the base and the scalar as the exponent is calculated. The exponential function value is divided by one plus the sum of the exponential function values ​​to obtain a value between zero and one as the probability of infection risk of the spore-forming parasite.

Images