Breeding risk early warning method and system based on environmental data

Abstract

The present application relates to the technical field of breeding risk early warning, and discloses a breeding risk early warning method and system based on environmental data, which comprises the following steps: obtaining water quality historical data, performing time sequence state extraction to obtain water quality time sequence characteristics; extracting fluctuation amplitude and performing frequency domain evolution analysis to obtain environmental time-varying law; performing cross-correlation calculation to obtain correlation evolution matrix; performing path division and evaluation to obtain coupling distortion score; combining distortion determination threshold to generate synergistic stress characteristics; performing warning offset analysis and ecological evaluation to obtain ecological carrying capacity margin, and correcting to obtain dynamic ecological boundary; and performing risk judgment according to real-time numerical value and dynamic ecological boundary to obtain risk early warning information. The present application can realize early and effective early warning of recessive risk under multi-factor synergistic stress.

Description

Technical Field

[0001] This invention relates to the field of aquaculture risk early warning technology, and in particular to an aquaculture risk early warning method and system based on environmental data. Background Technology

[0002] Currently, risk management in aquaculture is transitioning from single water quality threshold alarms to intelligent assessment. Introducing fault prediction and health management concepts into aquaculture environment monitoring, and identifying system-level risks in advance through multi-source data fusion and state evolution modeling, has become an important technological trend in ensuring aquaculture safety. However, water bodies form a highly coupled ecosystem due to biological metabolism and chemical reactions. Multifactorial stress often causes latent damage to organisms even before all indicators exceed limits. This places higher demands on feature extraction and pattern discrimination within the framework of fault prediction and health management.

[0003] In one existing technology, an aquaculture early warning system based on fault prediction and health management acquires historical sequences of dissolved oxygen, ammonia nitrogen, and pH through water sampling and online sensors. It then uses a data-driven model to calculate the health status and remaining effective monitoring window for each parameter. From an ecological coupling mechanism perspective, this approach treats each water quality factor as an independent characteristic that does not interfere with each other, assessing only the deviation trend and decay rate of a single indicator along the time dimension. However, the material cycle in aquaculture water bodies has sequential relationships; for example, the decomposition of uneaten feed first causes an increase in ammonia nitrogen, which then generates nitrite through nitrification. This delayed coupling relationship cannot be characterized by the superposition of isolated parameter health statuses. Simultaneously, the seasonal migration of temperature, algal community, and feeding rhythms alters the interaction strength between factors, leading to drastically different stress effects for the same parameter combination at different times. Existing technologies lack dynamic tracking of factor topological variations and synergistic distortion evolution, reducing the complex nonlinear coupling of multiple factors to a single-dimensional decay curve, failing to capture synergistic stress risks before parameter values ​​cross their respective warning lines.

[0004] In summary, existing technologies suffer from the problem of failing to provide early warnings of latent risks under collaborative stress due to the lack of a time-varying evolution analysis mechanism for multi-factor coupled topology. Summary of the Invention

[0005] This invention provides a method and system for early warning of aquaculture risks based on environmental data, so as to achieve early and effective warning of hidden risks under multi-factor synergistic stress.

[0006] Firstly, in order to solve the above-mentioned technical problems, the present invention provides a method for early warning of aquaculture risks based on environmental data, comprising:

[0007] Historical water quality data is acquired, and time-series state extraction processing is performed on the historical water quality data to obtain water quality time-series features.

[0008] The fluctuation amplitude is extracted based on the water quality time-series characteristics, and frequency domain evolution analysis is performed on the fluctuation amplitude to obtain the environmental time-varying law;

[0009] Based on the time-varying environmental patterns, cross-correlation calculations are performed on the historical water quality data to obtain the correlation evolution matrix;

[0010] The correlation evolution matrix is ​​divided into paths to obtain abnormal isolation paths. The abnormal isolation paths are evaluated to obtain coupling distortion scores. Cooperative stress features are generated based on the coupling distortion scores and a preset distortion judgment threshold.

[0011] Based on the aforementioned collaborative stress characteristics, a warning offset analysis is performed to obtain a warning offset vector. Based on the warning offset vector, an ecological assessment is conducted to obtain the ecological carrying capacity margin. Based on the ecological carrying capacity margin, the warning offset vector is corrected to obtain the dynamic ecological boundary.

[0012] Real-time water quality data are obtained, and risk assessment is performed based on the real-time water quality data and the dynamic ecological boundary to obtain risk warning information.

[0013] Secondly, the present invention provides an aquaculture risk early warning system based on environmental data, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described above.

[0014] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described above.

[0015] Compared with the prior art, the present invention has the following beneficial effects:

[0016] (1) This invention performs time series modeling on the historical values ​​of multiple water quality factors and extracts the fluctuation amplitude. After obtaining the energy spectral density through fast Fourier transform, it constructs a state evolution matrix and extracts the migration trajectory to obtain the environmental time-varying law reflecting the water state migration. This solves the problem that the existing technology can only evaluate the decay trend of a single parameter and cannot track the time-varying evolution of multiple factors, and provides accurate time series characteristics for subsequent correlation analysis.

[0017] (2) This invention calculates the lag time step between factors through cross-correlation function, aligns the sequence and constructs node connection edges by dynamic weight vector weighted mapping, generates the first association evolution matrix through graph convolutional neural network, and then uses the isolated forest algorithm to divide abnormal isolation paths and calculate coupling distortion score. This solves the problem that existing technologies cannot identify collaborative stress when multiple factors are independently evaluated, and effectively captures the multi-factor combination distortion risk when parameters do not exceed the limit.

[0018] (3) This invention extracts the stress response gradient and constructs the superimposed evolution trajectory when the coupled distortion fraction exceeds the threshold, determines the synergistic stress characteristics, calculates the factor superposition weight and metabolic consumption rate, evaluates the ecological carrying capacity margin to dynamically correct the independent warning value, and forms a dynamic ecological boundary that adapts to the current metabolic intensity. This solves the problem that the fixed threshold of the existing technology cannot adapt to environmental changes, realizes the upgrade from static single-point alarm to dynamic synergistic early warning, and significantly improves the timeliness and accuracy of early warning of hidden risks. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the aquaculture risk early warning method based on environmental data provided in the first embodiment of the present invention. Detailed Implementation

[0020] 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.

[0021] Reference Figure 1 The first embodiment of the present invention provides an early warning system for aquaculture risks based on environmental data, including the following steps:

[0022] S11, acquire historical water quality data, perform time-series state extraction processing on the historical water quality data, and obtain water quality time-series features;

[0023] S12, extract the fluctuation amplitude based on the water quality time series characteristics, perform frequency domain evolution analysis on the fluctuation amplitude, and obtain the environmental time-varying law;

[0024] S13, perform cross-correlation calculation processing on the historical water quality data according to the time-varying environmental pattern to obtain the correlation evolution matrix;

[0025] S14, perform path partitioning on the correlation evolution matrix to obtain abnormal isolation paths, perform path evaluation on the abnormal isolation paths to obtain coupling distortion scores; generate cooperative stress features based on the coupling distortion scores and a preset distortion judgment threshold.

[0026] S15, perform warning offset analysis based on the cooperative stress characteristics to obtain a warning offset vector, perform ecological assessment based on the warning offset vector to obtain ecological carrying capacity, and correct the warning offset vector based on the ecological carrying capacity to obtain a dynamic ecological boundary.

[0027] S16, Obtain real-time water quality data, and make a risk assessment based on the real-time water quality data and the dynamic ecological boundary to obtain risk warning information.

[0028] In step S11, historical water quality data is acquired, and time-series state extraction processing is performed on the historical water quality data to obtain water quality time-series features, including:

[0029] Obtain historical water quality data, perform interpolation and imputation on the historical water quality data, and perform normalization processing to obtain a normalized matrix;

[0030] Perform window truncation on the normalized matrix to obtain the standard input tensor;

[0031] The hidden layer state is obtained by updating the state of the standard input tensor.

[0032] Based on the hidden layer state, feature extraction is performed at the final time step to obtain water quality time-series features.

[0033] It should be noted that a multi-parameter water quality sensor array deployed in aquaculture ponds collects values ​​for five water quality factors—dissolved oxygen, water temperature, pH, ammonia nitrogen, and nitrite—at a fixed frequency of once every 10 minutes. After each collection, the raw electrical signals undergo analog-to-digital conversion and filtering to obtain initial values. All initial values ​​are timestamped and stored in a local database, with a time span of the past 30 days, thus obtaining historical water quality data. The 10-minute sampling interval is chosen because dissolved oxygen in aquaculture can change drastically during summer nights. In high-density aquaculture ponds, dissolved oxygen can drop from 8 mg / L to 3 mg / L within 2 to 3 hours. A 10-minute sampling interval can capture at least 12 to 18 data points before the risk occurs, providing a sufficiently dense observation sequence for subsequent time-series modeling.

[0034] Next, linear interpolation is performed to impute missing values ​​in the historical water quality data. Each time series is traversed, the missing position is located, and the values ​​of the two valid time points before and after it are read. The increment is calculated according to the time difference and added to the previous value to complete the imputation. After imputation, min-max normalization is performed separately for each water quality factor sequence. During processing, the entire 30-day sequence is scanned to find the minimum and maximum values ​​of the factor. Then, each value is subtracted from the minimum value and divided by the difference between the maximum and minimum values. All values ​​are linearly mapped to the range of 0 to 1, ultimately resulting in a normalized matrix of five water quality factors representing the time step sequence. Scaling factors with different dimensions, such as dissolved oxygen and water temperature, uniformly eliminates gradient update bias caused by differences in numerical ranges, allowing the network to treat each input feature equally during the initial training phase, accelerating convergence and reducing the probability of getting trapped in local minima.

[0035] It's worth noting that the window truncation time step is set to 24, which is equivalent to using the past 4 hours of continuous data as a basic processing unit. Starting from the beginning of the normalized matrix, consecutive segments of length 24 are extracted sequentially. After each extraction, the window is slid forward by one time step, generating several 24-row, 5-column two-dimensional slices, each of which is a standard input tensor. The sliding step is set to 1 instead of being the same length as the window. The purpose is to create a large overlap between adjacent windows, which increases the effective supervision signal density several times while keeping the number of training samples the same. At the same time, it allows the network to repeatedly observe the subtle differences between adjacent windows, thus more stably learning the smooth transition process of water quality changes.

[0036] It should be noted that a Long Short-Term Memory (LSTM) network is used to update the state of the standard input tensor. The network structure consists of two cascaded LSTM hidden layers. The first layer has 64 hidden units, and the second layer has 32 hidden units, both using the tanh function as the activation function. The input dimensions correspond to the five water quality factors. At each time step, the network receives the 5-dimensional normalized vector of the current time step, combines it with the hidden layer state and cell state of the previous time step, and updates the internal cell state through gating calculations using forget gates, input gates, and output gates, and outputs the hidden layer state of the current time step. After the network processes the input for 24 time steps, the hidden layer state of the last time step is taken as the complete temporal encoding of the input tensor. This encoding vector integrates all the fluctuation information of the five water quality factors and their dynamic constraints over the past 4 hours, thus obtaining the water quality temporal features.

[0037] During the model training phase, continuous water quality data from the past 180 days were collected from the historical operation records of the same aquaculture pond to construct a training set. The first 150 days were used as the training set, and the last 30 days as the validation set. Mean squared error was used as the loss function during training, and the Adam algorithm was used as the optimizer. The initial learning rate was set to 0.001, and it decreased to 0.5 every 10 training epochs. Each training iteration used 32 samples as input, and the maximum number of training epochs was set to 200. Early stopping was triggered when the validation set loss did not decrease for 15 consecutive epochs. During the inference phase, the trained weight file was loaded onto the on-site edge computing terminal, and a single forward propagation was performed on the 24-step normalized tensor captured in real time. The hidden layer state output at the last time step was directly used as the water quality time-series feature for subsequent processes.

[0038] In step S12, the fluctuation amplitude is extracted based on the water quality time-series characteristics, and frequency domain evolution analysis is performed on the fluctuation amplitude to obtain the environmental time-varying law, including:

[0039] Extreme value differences are extracted based on the aforementioned water quality time-series characteristics to obtain the fluctuation amplitude;

[0040] Calculate the rate of change of the fluctuation amplitude. If the rate of change is greater than a preset rate determination threshold, then perform a discrete time-domain transformation on the fluctuation amplitude to obtain the frequency domain component.

[0041] Calculate the squared magnitude of the frequency domain component to obtain the energy spectral density;

[0042] A state evolution matrix is ​​constructed based on the energy spectral density, and the state evolution matrix is ​​then optimized by sequence decoding to obtain the time-varying laws of the environment.

[0043] It should be noted that when extracting fluctuation amplitude from water quality time-series features, the largest and smallest elements in the vector are first located, and the difference between them is obtained as a scalar difference. This difference reflects the overall volatility of water quality within the 4-hour window and is denoted as the fluctuation amplitude of the current window. As the window slides forward on the time axis, each 4-hour historical data segment generates a corresponding fluctuation amplitude. 120 consecutive windows can form a fluctuation amplitude sequence describing the trend of water quality fluctuations over the past 48 hours.

[0044] It is worth noting that the rate of change for each element in the fluctuation amplitude sequence is calculated by subtracting the fluctuation amplitude of the previous window from the fluctuation amplitude of the next window, and then dividing by the time interval of 20 minutes between the center points of adjacent windows, resulting in a rate of change sequence in units of change per minute. The preset rate judgment threshold is set at 0.08 fluctuation amplitude units per 20 minutes. This threshold is determined based on the early signs of a dangerous decline in dissolved oxygen content in aquaculture. When phytoplankton communities in the water suddenly stop photosynthesis due to cloudy or rainy weather, the initial stage of the abnormal deflection of dissolved oxygen from normal levels typically sees the rate of change fluctuating between 0.06 and 0.12 per 20 minutes. Selecting 0.08, which is slightly lower than the middle, as the warning threshold can achieve a balance between the false alarm rate and the missed alarm rate. Furthermore, after seasonal changes or adjustments in aquaculture density, the maximum rate of change can be calibrated by rescanning the historical data of 7 consecutive days without abnormalities. If three consecutive points in the rate of change sequence exceed 0.08 and show an increasing trend, it is determined that the water quality fluctuation in the current period has deviated from the range of steady and gradual change, and frequency domain analysis of this fluctuation amplitude sequence needs to be triggered.

[0045] Subsequently, a Fast Fourier Transform (FFT) was used to perform a discrete-time domain transformation on the fluctuation amplitude sequence. Treating the fluctuation amplitude sequence as a time-discrete real-valued signal, a segment of 128 consecutive windows was extracted and decomposed using FFT to obtain a set of complex frequency components. Each component corresponds to a discrete frequency point, and its real and imaginary parts together describe the oscillation intensity and phase shift of the fluctuation signal at that frequency. For aquaculture environments, the effective frequency range was limited to 0.5 to 48 times per day. Components below 0.5 times per day were considered seasonally variable background, and components above 48 times per day were considered random electrical noise from the sensor; both were removed from the component sequence. The remaining mid-frequency components are the frequency components that truly reflect environmental driving events such as diurnal rhythms, feeding cycles, and the passage of meteorological fronts.

[0046] It is worth further explaining that for each selected frequency component, the square of its amplitude is calculated. The real and imaginary parts of the component are squared and then added together to obtain a non-negative real number. This value represents the energy contribution of that frequency component in the entire fluctuation amplitude signal. All energy values ​​are arranged from low to high frequency to directly form the energy spectral density. This transformation removes the positive and negative cancellation interference caused by the phase information in the original frequency components, allowing the contribution of each frequency point to be represented by a single scalar. The entire calculation does not introduce any external hyperparameters and is entirely driven by the data structure of the fluctuation amplitude sequence itself.

[0047] Finally, constructing the state evolution matrix requires observing multiple consecutive 4-hour windows as a group of observation batches. Within each batch, multiple sets of energy spectral densities are generated in chronological order. Each set of energy spectral densities serves as a column vector of the matrix, with rows corresponding to frequency points and columns corresponding to time steps. The matrix values ​​represent the energy values ​​of that frequency at that time step. The distribution of all energy values ​​on this matrix is ​​statistically analyzed, and the 1 / 3 and 2 / 3 quantiles are used as dividing boundaries to classify the energy at each frequency point into three discrete states: high, medium, and low. The transition probabilities between states are directly estimated from the frequency statistics of state changes between adjacent time steps, with the initial state probability taken as the energy distribution ratio of the first column of the matrix. The state with the highest probability at each time step is selected sequentially to form an optimal state transition path. This path fully records the structural switching and continuation of energy components at each frequency along the time axis, i.e., the time-varying environmental patterns of the aquaculture water body under the synergistic influence of multiple factors.

[0048] In step S13, cross-correlation calculations are performed on the historical water quality data according to the time-varying environmental patterns to obtain a correlation evolution matrix, including:

[0049] Based on the time-varying environmental pattern, the lag time step is extracted, and the historical water quality data is time-series aligned based on the lag time step to obtain an alignment factor sequence.

[0050] The alignment factor sequence is cross-calculated to obtain the coupling response characteristics;

[0051] The coupled response features are weighted and mapped according to the pre-acquired dynamic weight vector to obtain the response intensity distribution;

[0052] Nodes are extracted based on the response intensity distribution to obtain evolved nodes, and node connection edges are constructed based on the spatial Euclidean distance between the evolved nodes.

[0053] Based on the feature association analysis of the evolution nodes and the connecting edges between the nodes, the association evolution matrix is ​​obtained.

[0054] It should be noted that the environmental time-varying pattern is the optimal state transition path obtained in the previous step. This path records the complete temporal sequence of energy switching between high, medium, and low states for each frequency component over time. When extracting the lag time step from this path, the time point of state switching is first located, the time interval between every two adjacent switching points is calculated, and the longest interval with the highest frequency is taken as the lag time step. The physical meaning of this step is a complete energy state transition cycle in water quality fluctuations, representing the typical delay required for factors such as dissolved oxygen and water temperature to go from response to rebalancing under the drive of diurnal rhythms or feeding cycles. Next, this lag time step is used to perform temporal alignment on every two water quality factor sequences in the historical water quality data. The two aligned sequences are denoted as the leading sequence and the lag sequence, respectively. The lag sequence is shifted backward by the number of sampling points corresponding to the lag time step, while the leading sequence remains in place. After the shift is completed, the non-corresponding parts caused by misalignment at both ends are truncated, and the remaining aligned segments constitute the aligned factor sequence of that factor pair.

[0055] Next, cross-calculation is performed on the aligned factor sequences generated for every two water quality factors. The values ​​of the leading and lagging sequences are taken separately. First, the mean of each sequence is calculated. Then, the product of the deviations of the leading and lagging values ​​from the mean at each time point is calculated. All products are summed and divided by the quotient of the product of the standard deviations of the two sequences and the sequence length to obtain the coupled response characteristics of the factor under the current lag conditions. The same cross-calculation is performed sequentially on multiple factor combinations, such as those between dissolved oxygen and pH, and between water temperature and ammonia nitrogen. The scalar values ​​obtained from each combination are concatenated into a multidimensional coupled response feature vector. This vector describes the intensity pattern of mutual following and mutual driving among different water quality indicators within the delayed response framework revealed by the time-varying environmental laws.

[0056] It is worth further explaining that the dynamic weight vector originates from the assessment of the contribution of each water quality factor to aquaculture risk events. During the assessment, risk periods marked in the past 180 days of historical data are used. For each of the five water quality factors within a risk period, the area integral of its deviation from the normal range is calculated. The larger the deviation area, the more important the factor's role in risk formation, and its corresponding weight value is higher. The ratio of the deviation area of ​​each factor to the sum of the deviation areas of all factors is used to obtain the normalized weight corresponding to that factor. These five weight values ​​together constitute a dynamic weight vector. This weight vector is updated weekly to adapt to the shift in water quality management focus at different aquaculture stages. Each component in the current coupled response feature vector is multiplied by its corresponding weight value in the dynamic weight vector to obtain a set of weighted response intensity values, which are then arranged in the original order of the factor pairs to form a response intensity distribution.

[0057] Subsequently, nodes are extracted from the response intensity distribution, which is a multi-dimensional vector where each dimension represents the weighted coupling strength of a pair of factor pairs. An intensity screening boundary is set, which is 0.8 times the median of all components in the current response intensity distribution. Factors with response intensities exceeding this boundary are marked as evolutionary nodes. Each evolutionary node corresponds to a pair of water quality factors exhibiting a significant coupling-driven relationship at a specific time delay. The spatial coordinates of the node are taken as the index positions of the two factors in the response intensity distribution vector, and the index values ​​are used as the horizontal and vertical coordinates of the node in a two-dimensional abstract space. The connection edges between nodes are determined by calculating the Euclidean distance between the two evolutionary nodes in the two-dimensional abstract space. When the spatial Euclidean distance between two nodes is less than the preset connection judgment distance of 1.5, a connection edge is established between the two nodes; node pairs with a distance greater than or equal to 1.5 are not connected. The 1.5 connection determination distance is derived from the statistical distribution of response intensity of multiple normal breeding cycles. Under normal conditions, factor pairs with real coupling relationships are often closely distributed in the index space, and the distance usually falls within the range of 1.0 to 2.0. Taking 1.5 can distinguish factor pairs belonging to the same coupling cluster from random correlation factor pairs across clusters.

[0058] Finally, a feature label is constructed for each evolution node, containing the names of the two water quality factors represented by the node and the current response intensity value of the node. Nodes with connections are grouped into the same coupled subgraph based on the connecting edges. All node feature labels within each coupled subgraph are arranged in descending order of response intensity and concatenated into a feature description string. The feature description strings between different coupled subgraphs are then arranged sequentially along the rows of a matrix according to the number of nodes within the subgraph, ultimately forming an association evolution matrix with coupled subgraphs as rows and feature description strings as content. This matrix organizes the discrete factor pair coupling relationships into a structured association topology, directly presenting the complete pattern of mutual co-evolution among multiple water quality factors under the constraints of time-varying environmental patterns, providing a complete data foundation for subsequent anomaly path segmentation and cooperative stress identification.

[0059] In step S14, the correlation evolution matrix is ​​divided into paths to obtain abnormal isolation paths, and the abnormal isolation paths are evaluated to obtain coupling distortion scores, including:

[0060] The correlation evolution matrix is ​​decomposed into eigenvalues ​​to obtain the topological features and edge weight variation values.

[0061] Based on the topological features and the variation values ​​of edge weights, spatial recursion is performed to obtain abnormal isolation paths;

[0062] Based on the abnormal isolation path, hierarchical numerical extraction is performed to obtain a path depth set, and the average depth value of the path depth set is calculated.

[0063] The difference between the average depth value and the pre-acquired evaluation benchmark value is calculated to obtain the state deviation magnitude;

[0064] The deviation magnitude of the state is mapped to obtain the coupling distortion score.

[0065] It should be noted that the association evolution matrix is ​​a structured matrix with coupled subgraphs as rows and feature description strings as content. Before eigenvalue decomposition, this matrix is ​​first converted into a graph Laplacian matrix in the form of an adjacency matrix. The conversion process involves traversing all evolved nodes within each coupled subgraph of the association evolution matrix, marking each pair of nodes within the same subgraph as having a connection edge, and taking the average of the response intensities of the two nodes as the edge weight. This constructs an adjacency matrix with evolved nodes as row and column indices and the average edge weights as elements. Next, the degree matrix of this adjacency matrix is ​​calculated. The degree matrix is ​​a diagonal matrix, and each element on the diagonal is the sum of the weights of all connecting edges to the corresponding node. Subtracting the adjacency matrix from the degree matrix yields the Laplacian matrix. Eigenvalue decomposition is performed on this Laplace matrix, and the eight smallest non-zero eigenvalues ​​are taken as topological features. From the perspective of graph theory, this set of eigenvalues ​​describes the overall connectivity and modularity of the current multi-factor coupled network. Smaller eigenvalues ​​with more uniform spacing indicate a more stable and cooperative coupling relationship between factors in the network. Sudden jumps in eigenvalues ​​suggest abnormal shifts in local coupling strength within the network. Simultaneously, the weights of all connecting edges are extracted from the Laplace matrix, and the coefficient of variation of these weights is calculated. This coefficient of variation reflects the degree of difference in the strength of each coupled edge; a larger coefficient of variation indicates a more significant difference in response strength between different factor pairs, and a poorer balance in the coupled network.

[0066] Next, based on this set of topological features and the variation values ​​of edge weights, a spatial recursive partition is performed, recursively executing a bisection. First, the feature coordinates of all nodes are extracted from the eigenvector corresponding to the second smallest eigenvalue of the Laplacian matrix. The median of these values ​​is taken as the split point, with nodes whose values ​​are greater than the median assigned to the left subset and those less than or equal to the median assigned to the right subset, completing the first layer of partitioning. Then, within each subset, the local Laplacian matrix is ​​recalculated, and the same bisection process is repeated. The termination condition for the recursive partitioning is set at the difference in the coefficient of variation of edge weights between the left and right subsets after a single split not exceeding 8%. A difference exceeding 8% means that the edge weights within one subset fluctuate wildly while the edge weights within the other subset tend to be uniform. This asymmetric structure itself is a topological manifestation of abnormal coupling, and the recursive path from the root node to the current cutting plane is recorded as an abnormal isolation path. Each abnormal isolation path fully preserves the spatial nesting hierarchy relationship from the global coupled network layer by layer to the abnormal locality.

[0067] It's worth noting that each recursive partition is recorded as a level depth, and the total number of cuts from the root node to the termination node of the anomaly isolation path is the path's depth. The average depth is obtained by averaging the depths of all anomaly isolation paths. A larger average depth indicates that more recursive stripping is required to isolate the anomalous substructure in the current coupled network, suggesting that the anomalous coupling source is buried deeper, and the collaborative imbalance mechanism between multiple factors is more complex and the transmission chain is longer. A smaller average depth indicates that the anomalous substructure is superficial and may be triggered by a sudden imbalance in a single factor pair.

[0068] It is worth further explaining that the evaluation baseline value was obtained by selecting eight equally spaced time points each day during a continuous 30-day normal aquaculture period to generate an association evolution matrix and performing the same recursive partitioning and average depth calculation, accumulating 7200 average depth samples under normal conditions. The arithmetic mean of these samples was taken as the evaluation baseline value. The absolute value of the difference between the current average depth value and this baseline value is the state deviation magnitude. The state deviation magnitude quantifies the difference between the degree of anomaly of the multi-factor coupled network at the current moment and its normal stable operating state; the larger the difference, the more significant the deviation.

[0069] It should be noted that a piecewise mapping method is used when mapping the deviation magnitude to convert the deviation magnitude into a coupling distortion score with a value between 0 and 1. First, the standard deviation of 7200 normal state depth mean samples is calculated. The deviation magnitude divided by three times the standard deviation is used as the mapping input. When the input is less than or equal to 1, it is directly multiplied by 0.3 to obtain the mapped output. When the input is greater than 1 but less than 2, the increment of 0.3 from the first part of the output is added to the increment of 0.3 from the excess part. When the input is greater than or equal to 2, the upper limit of 1 is taken. The basis for this piecewise mapping is that when the deviation magnitude falls within three times the standard deviation, the water quality coupling network is still within the normal random fluctuation range, and the growth rate of the distortion score remains constant and slow. After exceeding three times the standard deviation, the contribution of each additional unit of deviation to the overall risk decreases progressively, indicating that the anomaly has entered a warning range that cannot be ignored. The final output coupling distortion score fully reflects the degree of distortion of the synergistic relationship of multiple water quality factors in the current aquaculture water body under the driving force of environmental time-varying laws, providing a quantifiable input basis for the subsequent generation of synergistic stress characteristics.

[0070] In one possible implementation, a cooperative stress feature is generated based on the coupled distortion score and a preset distortion determination threshold, including:

[0071] Obtain a pre-recorded stress response sequence, perform differential calculation on the stress response sequence, and obtain the response gradient;

[0072] The response gradients are clustered to obtain gradient cluster centers;

[0073] The dispersion is calculated based on the gradient cluster centers to obtain the cluster dispersion.

[0074] If the clustering dispersion is less than a preset threshold, an abnormal behavior sequence of the organism is obtained. Then, feature fusion and splicing are performed based on the stress response sequence and the abnormal behavior sequence to obtain a superimposed evolutionary trajectory.

[0075] The overlap index of the superimposed evolution trajectory is extracted, and a cooperative stress feature is generated based on the overlap index.

[0076] It should be noted that when the coupling distortion score exceeds the preset distortion judgment threshold, the generation process of collaborative stress features is triggered. The distortion judgment threshold is set based on the statistical distribution of coupling distortion scores calculated at 8 equally spaced time points per day during a continuous 30-day normal aquaculture period. The distortion score corresponding to the 95th percentile is rounded up to 0.35 as the judgment threshold. A coupling distortion score exceeding 0.35 means that the current anomaly of the multi-factor coupling network has exceeded the random range of normal fluctuations, and the generation process of collaborative stress features is triggered at this time. If the coupling distortion score is less than or equal to 0.35, it is determined that the current water quality coupling network is within the normal collaborative fluctuation range, and subsequent stress response analysis is not triggered. The process directly returns to step S11 to continue collecting the next round of historical water quality data. The pre-recorded stress response sequence is stored in the local database, with a time span of the past 72 hours, and indirect indicators representing the biological stress state in the water body are recorded at a frequency of once every 10 minutes. The specific contents of the stress response sequence include the diurnal fluctuation shift rate of dissolved oxygen in water, the hourly drift rate of pH in water, and the instantaneous change amplitude of the ratio of total ammonia nitrogen to ionized ammonia nitrogen in water. These three indicators together constitute the physiological stress signal chain of organisms when facing the synergistic deterioration of water quality.

[0077] It is worth noting that the differential calculation calculates each index in the stress response sequence separately, subtracting the value at each time point from the value at its immediately preceding time point. The difference sequence records the instantaneous increase or decrease of that stress index within adjacent sampling intervals. After aligning the difference sequences of the three indices along the time axis, they are concatenated in the order of collection to form a set of multidimensional response gradients. The physical meaning of the response gradient is the rate of change of stress intensity experienced by the organism when it perceives the coordinated distortion of external water quality. Positive values ​​represent increased stress, negative values ​​represent reduced stress, and the larger the absolute value, the more severe the stress impact experienced by the organism within a unit time window.

[0078] Subsequently, clustering was performed on the multidimensional response gradients using a cohesive hierarchical clustering algorithm with splitting rules. Initially, the multidimensional response gradient at each time point was treated as an independent cluster. The Euclidean distance between each pair of clusters was calculated, and the two clusters with the smallest distance were merged. After merging, the center position of the new cluster was recalculated, and the merging process continued iteratively until the minimum distance between the centers of any two clusters exceeded a preset clustering threshold of 0.05. This clustering threshold of 0.05 was determined based on the fluctuation characteristics of the stress response gradient during normal breeding periods. Over seven consecutive days without stress events, the point-by-point variation range of the response gradient values ​​stabilized between -0.03 and +0.03. Setting the threshold at approximately 1.5 times this range avoids misclassifying normal physiological fluctuations as independent clusters. After clustering, the geometric center of each cluster is a gradient cluster center. Each gradient cluster center represents the characteristic rate level of the organism's fast, medium, or slow response when performing stress regulation.

[0079] Then, the cluster dispersion is calculated by taking the Euclidean distance between each cluster center and the geometric centroid of all cluster centers, summing the squares of these distances, and dividing by the total number of cluster centers. The resulting value is the cluster dispersion. Cluster dispersion measures the degree of separation between different stress rate levels in the feature space. The larger the dispersion, the more the organism exhibits a multi-level, distinct stress regulation strategy when facing coordinated water quality distortions. The smaller the dispersion, the more the response rates of various stress indicators tend to assimilate, and the organism has entered a single global stress mode.

[0080] It should be noted that when the cluster dispersion is less than a preset threshold, the organism is considered to have entered a state of coordinated stress. The threshold is set at 0.15. This value is determined by calculating the cluster dispersion daily during a 30-day normal aquaculture period without stress events. The mean of all normal state dispersion values ​​plus three times the standard deviation is used as the upper limit, and the boundary value is rounded down to 0.15. A dispersion below 0.15 indicates that the response rates of various stress indicators have become highly convergent. The diurnal fluctuations of dissolved oxygen, pH drift, and ammonia nitrogen ratio in the water no longer exhibit their own independent rhythms, but instead enter a stress-induced stagnation state forced into synchronization by external water quality coupling distortions. At this point, abnormal behavior sequences of the organisms are acquired. These sequences are obtained through frame-by-frame analysis of video streams captured by underwater cameras above the aquaculture water body. For each frame, the swimming speed and turning frequency of individual fish are extracted. Abnormal behavior events are recorded when the swimming speed consistently exceeds four times the fish's body length per second for six consecutive frames, or when the turning frequency sharply increases to more than five times per 10 seconds. The density of these abnormal behavior events is statistically analyzed every 10 minutes to form a biological abnormal behavior sequence aligned with the stress response sequence. If the cluster dispersion is greater than or equal to 0.15, the organism is considered to still maintain normal stress regulation capabilities at multiple levels, and stress response data continues to be collected and the response gradient updated at the normal monitoring frequency.

[0081] The stress response sequence and the abnormal biological behavior sequence were fused and spliced ​​along a unified time axis. The splicing method involved arranging the three-dimensional stress response values ​​and the one-dimensional abnormal behavior density values ​​in parallel at each time point to form a set of four-dimensional feature tuples. Concatenating all feature tuples in chronological order yielded the superimposed evolutionary trajectory. This superimposed evolutionary trajectory fully recorded the synchronous evolution of the physiological stress signal chain and biological behavioral response to water quality co-distortion at the same time density.

[0082] Finally, an overlap index was extracted from the superimposed evolutionary trajectories. This was achieved by comparing the direction of change in the stress response sequence with the direction of change in the density of anomalous biological behavior at each time point, and then calculating the proportion of time points where both changes in the same direction to the total number of time points in the entire sequence. This proportion is the overlap index. The overlap index measures the directional consistency between water quality stress factors and biological behavioral responses over time. A higher overlap index indicates a closer causal relationship between water quality synergistic distortion and anomalous biological behavior. The overlap index is directly output as a synergistic stress feature for subsequent warning migration analysis.

[0083] In step S15, a warning offset analysis is performed based on the cooperative stress characteristics to obtain a warning offset vector. An ecological assessment is then performed based on the warning offset vector to obtain the ecological carrying capacity margin. Finally, the warning offset vector is corrected based on the ecological carrying capacity margin to obtain the dynamic ecological boundary, including:

[0084] The feature variance contribution rate is calculated based on the collaborative stress characteristics to obtain the factor superposition weights;

[0085] A linear mapping is performed on the preset independent warning values ​​based on the superimposed weights of the factors to obtain the warning offset vector;

[0086] The metabolic intensity of the water body is obtained, and a fitting and extrapolation calculation is performed based on the metabolic intensity of the water body and the warning offset vector to obtain the metabolic consumption rate.

[0087] If the metabolic consumption rate is greater than the preset metabolic baseline threshold, the ecological carrying capacity is calculated based on the pre-acquired characteristics of biological stock and total dissolved volume.

[0088] A floating correction range is set according to the ecological carrying capacity, and a smoothing correction process is performed on the warning offset vector according to the floating correction range to obtain the dynamic ecological boundary.

[0089] It should be noted that the synergistic stress feature is an overlap index with a value between 0 and 1. This scalar value itself can only reflect the directional consistency between water quality stress and abnormal biological behavior in the time dimension, and is insufficient to decompose the independent contributions of each water quality factor in synergistic stress. In order to obtain the superposition weight of each water quality factor in the stress process, it is necessary to trace back to the superposition evolution trajectory before the generation of the synergistic stress feature. The superposition evolution trajectory stores a set of four-dimensional feature tuples every 10 minutes. The first three terms are the diurnal fluctuation offset rate of dissolved oxygen, the hourly drift rate of pH, and the instantaneous change amplitude of the ratio of total ammonia nitrogen to ionized ammonia nitrogen corresponding to the stress response sequence. The fourth term is the density of abnormal biological behavior. Principal component analysis was performed on these four-dimensional feature tuples along the entire trajectory. The eigenvalues ​​corresponding to each principal component were calculated and arranged from largest to smallest. The loading matrix was truncated when the cumulative eigenvalues ​​of the first three principal components accounted for more than 85%. Then, the sum of squares of the loading coefficients corresponding to each original dimension was calculated column-by-column in the truncated loading matrix. The proportion of this sum of squares in the total sum of squares of all dimensions was used as the feature variance contribution rate of that dimension. The contribution rates of the three stress factors represent the relative importance of dissolved oxygen fluctuations, pH drift, and ammonia-nitrogen ratio changes in synergistic stress. After normalization, the three contribution rate values ​​directly constitute a set of factor superposition weights.

[0090] Next, the preset independent warning values ​​are safety thresholds set individually for each water quality factor. These thresholds are set according to common standards in the aquaculture industry. For dissolved oxygen in aquaculture ponds, the independent warning value is 3 mg / L, which is the critical dissolved oxygen level at which warm-water fish exhibit surfacing stress. For pH, the upper limit is 8.5 and the lower limit is 6.5, corresponding to the alkaline boundary where ammonia toxicity increases sharply and the acidic boundary where nitrification is hindered, respectively. For the ratio of total ammonia nitrogen to ionized ammonia nitrogen, the independent warning value is 1.8. Exceeding this value means that the proportion of free ammonia in total ammonia nitrogen has entered the risk range of irreversible damage to fish gill epithelial cells. The specific operation of linear mapping is to multiply the independent warning value of each water quality factor by the factor's corresponding superposition weight. The three products are arranged in the order of dissolved oxygen, pH, and ammonia nitrogen ratio into a three-dimensional vector, which is the warning offset vector. Each component value in the warning offset vector represents the extent to which the individual warning lines of each factor are compressed or amplified under the actual weight allocation of the coordinated stress. The higher the weight of a factor, the lower its warning line is compressed, which means that the factor needs to trigger an early warning in the coordinated stress scenario.

[0091] It should be further explained that the water metabolic intensity is estimated by the difference between the nighttime decline rate and the daytime rise rate of dissolved oxygen in the aquaculture pond over a continuous 48-hour period. At night, the linear slope of the dissolved oxygen sequence during the 3-hour period from 2:00 AM to 5:00 AM is calculated as the respiratory consumption intensity. During the day, the linear slope of the dissolved oxygen sequence during the 4-hour period from 10:00 AM to 2:00 PM is calculated as the photosynthetic oxygen production intensity. The average of the respiratory consumption intensity and the photosynthetic oxygen production intensity is taken as the comprehensive water metabolic intensity for the current period, expressed in mg / (L·h). The dissolved oxygen component in the warning offset vector is used as the starting point for oxygen demand under stress response. Using the water metabolic intensity as the consumption rate base, the consumption curve of dissolved oxygen content decreasing over time in the next 6 hours is fitted and extrapolated according to the exponential decay law. During the extrapolation, the decay coefficient of dissolved oxygen dissipation is set to 1.2 times the metabolic intensity to account for the additional consumption due to the increased respiratory rate of the fish under stress. The final metabolic consumption rate sequence is obtained, where each time point in the sequence represents the predicted net consumption of dissolved oxygen within a 10-minute interval.

[0092] It should be noted that the calculation of the ecological carrying capacity is triggered when the value at any point in the metabolic consumption rate sequence exceeds the preset metabolic baseline threshold. The metabolic baseline threshold is set at 0.25 mg / L every 10 minutes. This value is determined based on the fact that the fluctuation range of the metabolic consumption rate during the normal aquaculture period is concentrated between 0.08 and 0.18 mg / L every 10 minutes. When the consumption rate exceeds 0.25 mg / L, it means that the rate of decrease in dissolved oxygen has exceeded the range that the fish population can adapt to through its own physiological regulation. At this time, the pre-recorded biomass and total dissolved oxygen characteristics are obtained. The biomass is taken from the total mass of the fish population per unit water volume in the aquaculture pond recorded in the most recent aquaculture log, in kg / m³. The total dissolved oxygen characteristic is taken from the product of the real-time dissolved oxygen value at the current moment and the total water volume of the pond. The ecological carrying capacity is calculated by dividing the total dissolved oxygen by the product of the biomass and the peak metabolic consumption rate. The quotient obtained is the continuous consumption time that the remaining dissolved oxygen can support under the current biological load, in hours. If the values ​​at all time points in the metabolic consumption rate sequence are less than or equal to 0.25 mg per L per 10 minutes, it is determined that the current dissolved oxygen consumption rate is still within the range of physiological regulation and adaptation of the fish population. The calculation of the ecological carrying capacity is skipped, and the warning offset vector is directly used as the dynamic ecological boundary output at this moment without performing floating correction processing.

[0093] The upper and lower boundaries of the floating correction range are set in segments. When the ecological carrying capacity is greater than or equal to 24 hours, the fluctuation range is ±10%, indicating that the aquaculture water body is in a state of sufficient safety redundancy. The warning line can fluctuate flexibly within a certain range to reduce the frequency of false alarms. When the ecological carrying capacity is between 12 and 24 hours, the fluctuation range tightens to ±7%, indicating that the water body has entered a medium load range. The fluctuation space of the warning line needs to be appropriately compressed to maintain sensitivity to changes. When the ecological carrying capacity is less than 12 hours, the fluctuation range further narrows to ±4%. At this point, the remaining dissolved oxygen reserve is insufficient to cope with a new round of synergistic stress impact, and the warning line must be kept within a narrow range of high vigilance. The correction operation processes the three components of the warning offset vector item by item according to the floating correction range. For each component, its current value is multiplied by its corresponding floating correction coefficient. The floating correction coefficient is determined by the factor superposition weight of the current component and the fluctuation range corresponding to the carrying capacity level. After the correction, a dynamic ecological boundary with a fine-tuned position in three-dimensional space is obtained. The three components of the dynamic ecological boundary respectively provide the real-time safe operating upper or lower limits of dissolved oxygen, pH and ammonia nitrogen ratio under the dual constraints of the current coordinated stress situation and ecological carrying capacity, directly providing a quantifiable comparison benchmark for subsequent risk warning judgment.

[0094] In step S16, real-time water quality values ​​are acquired, and a risk assessment is performed based on the real-time water quality values ​​and the dynamic ecological boundary to obtain risk warning information, including:

[0095] Obtain real-time water quality values, calculate the difference between the real-time water quality values ​​and the dynamic ecological boundary, and obtain the boundary deviation magnitude;

[0096] A state sequence is constructed based on the boundary deviation magnitude, and the transition probability is deduced from the state sequence to obtain a risk evolution sequence;

[0097] The risk distribution matrix is ​​obtained by performing a two-dimensional mapping between the risk evolution sequence and a pre-defined spatial grid.

[0098] Risk analysis is performed based on the risk distribution matrix to obtain risk warning information.

[0099] It should be noted that real-time water quality data are continuously collected every 10 minutes by the same multi-parameter water quality sensor array as in step S11, reading the current measured values ​​of dissolved oxygen, pH, and the ratio of total ammonia nitrogen to ionized ammonia nitrogen. The dynamic ecological boundary is a three-dimensional vector output from step S15, corresponding to the safe lower limit of dissolved oxygen, the safe upper / lower limits of pH, and the danger threshold of the ammonia nitrogen ratio. When calculating the boundary deviation, for dissolved oxygen, the measured value is subtracted from the safe lower limit of dissolved oxygen in the dynamic ecological boundary; a positive difference indicates it remains within the safe range, while a negative difference indicates it has fallen below the lower limit and entered the hypoxia risk range. For pH, the difference between the measured value and both the upper and lower limit components is calculated, and the smaller absolute value is taken as the deviation magnitude; the direction of deviation is determined by which end the measured value is closer to. For the ammonia nitrogen ratio, the danger threshold in the dynamic ecological boundary is subtracted from the measured value; a positive difference indicates it has exceeded the safe threshold under synergistic stress conditions. The three differences form a boundary deviation magnitude vector, with the sign indicating the direction of deviation and the absolute value representing the severity of deviation.

[0100] Next, a state sequence was constructed based on the boundary deviation amplitude vectors from multiple consecutive sampling times. The calculated three-dimensional deviation amplitude vectors were stored chronologically in a sliding buffer of length 24, corresponding to the monitoring window of the past 4 hours. Each deviation amplitude vector within the buffer was discretized into a state: a negative deviation amplitude exceeding the risk threshold was mapped to state 2, representing a dangerous deviation; a deviation amplitude within the safety threshold was mapped to state 0, representing normal; and a positive deviation amplitude exceeding the redundancy threshold was mapped to state 1, representing safe redundancy. Risk thresholds were defined as dangerous deviations when the absolute value of a negative deviation in dissolved oxygen exceeded 0.5 mg / L, the pH deviation exceeded 0.3 units, and the ammonia-nitrogen ratio deviation exceeded 0.2, referencing the half-lethal dissolved oxygen concentration range and the acute toxicity threshold of free ammonia for warm-water fish, respectively. The safety threshold was uniformly set to an absolute deviation of no more than 0.1, and the redundancy threshold was set to a positive deviation exceeding 0.3. The three-dimensional state discretization values ​​of the 24 consecutive sampling points were arranged chronologically to form the state sequence.

[0101] It is worth further explaining that the transition probability extrapolation of the state sequence uses a first-order Markov chain model. First, the frequency of transitions between any two states in all state sequences during the past 90 days of normal farming is counted. The frequency is divided by the total number of times that state occurs to obtain a one-step transition probability matrix. The rows and columns of the matrix correspond to states 0, 1, and 2, respectively. The element in the i-th row and j-th column of the matrix represents the probability of transitioning to state j at the next moment, given that the current state is i. Using the latest state in the current state sequence as the starting point, the matrix is ​​repeatedly multiplied by the one-step transition probability matrix. Each multiplication represents extrapolating forward by one sampling interval (10 minutes). This process is repeated for 12 steps to obtain the probability distribution of the three states at each time step within the next 2 hours. This set of probability distributions, arranged in time step order, constitutes the risk evolution sequence. The reason for setting the simulation step count to 12 is that emergency response in aquaculture requires prediction at least 2 hours before the risk occurs. This time is sufficient to complete standard procedures such as starting oxygenation equipment, stopping feed feeding, and water exchange scheduling. Furthermore, if the simulation step count exceeds 12, the spread of the state probability distribution will increase significantly, leading to a decrease in the usability of the simulation results.

[0102] Then, a pre-defined spatial grid covers the entire water surface area of ​​the aquaculture pond, dividing the space into 10m x 10m square units. Each unit independently deploys a set of water quality sensors and calculates its own dynamic ecological boundary and risk evolution sequence. During two-dimensional mapping, each spatial unit is traversed, and the probability value of the dangerous deviation state corresponding to the 6th time step in its risk evolution sequence (i.e., the next hour) is extracted. This probability value is multiplied by 100 and rounded down to obtain the risk intensity score for that unit. The scores of all spatial units are filled into the corresponding number of rows and columns of a matrix according to the physical location of the pond, resulting in a risk distribution matrix that presents the distribution pattern of risk intensity in the next hour in a spatial dimension.

[0103] It should be noted that when conducting risk analysis based on the risk distribution matrix, the proportion of grid cells with a risk intensity score exceeding 60 in the entire matrix is ​​first calculated to obtain the high-risk coverage ratio. Simultaneously, the average physical distance between each pair of the top 5% of grid cells with the highest scores is calculated as the spatial cluster density; the smaller the distance, the more concentrated and contiguous the high-risk areas. Multiplying the high-risk coverage ratio by the cluster density yields the comprehensive risk index. The risk warning information includes the comprehensive risk index, a spatiotemporal distribution trend map of the high-risk coverage ratio, and a list of specific grid cell numbers with scores exceeding 80. A blue alert is issued when the comprehensive risk index is less than 0.15; a yellow alert is issued when it is between 0.15 and 0.35, triggering the aeration equipment to pre-increase to 60% of its rated power; and a red alert is issued when it is greater than 0.35, triggering full-load operation of the aeration equipment, suspension of feed feeding, and initiation of the drainage and water exchange process. The dividing values ​​for the three levels were determined through a retrospective analysis of all risk events in the past 180 days. The lowest comprehensive risk index on the day when a fish surfacing event actually occurred in the aquaculture pond was taken as the red warning dividing value, and 40% and 80% of this value were taken as the blue and yellow dividing values, respectively.

[0104] In summary, this invention discloses a method for early warning of aquaculture risks based on environmental data, comprising: acquiring historical water quality data; performing time-series state extraction processing on the historical water quality data to obtain water quality time-series features; extracting fluctuation amplitudes based on the water quality time-series features; performing frequency domain evolution analysis processing on the fluctuation amplitudes to obtain environmental time-varying patterns; performing cross-correlation calculation processing on the historical water quality data based on the environmental time-varying patterns to obtain a correlation evolution matrix; performing path partitioning on the correlation evolution matrix to obtain abnormal isolation paths; performing path evaluation on the abnormal isolation paths to obtain coupling distortion scores; generating cooperative stress features based on the coupling distortion scores and a preset distortion judgment threshold; performing warning offset analysis based on the cooperative stress features to obtain a warning offset vector; performing ecological assessment based on the warning offset vector to obtain ecological carrying capacity; correcting the warning offset vector based on the ecological carrying capacity to obtain a dynamic ecological boundary; acquiring real-time water quality values; and performing risk judgment based on the real-time water quality values ​​and the dynamic ecological boundary to obtain risk warning information. This invention integrates multidimensional temporal characteristics of water quality with time-varying environmental patterns, extracts multi-factor coupled distortion and synergistic stress characteristics, and dynamically corrects warning boundaries based on ecological carrying capacity, thereby achieving early and effective warning of hidden risks under multi-factor synergistic stress.

[0105] The second embodiment of the present invention provides an aquaculture risk early warning system based on environmental data, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the method described above.

[0106] It should be noted that the aquaculture risk early warning system based on environmental data provided in this embodiment of the invention is used to execute all the process steps of the aquaculture risk early warning method based on environmental data in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0107] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0108] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for early warning of aquaculture risks based on environmental data, characterized in that, include: Historical water quality data is acquired, and time-series state extraction processing is performed on the historical water quality data to obtain water quality time-series features. The fluctuation amplitude is extracted based on the water quality time-series characteristics, and frequency domain evolution analysis is performed on the fluctuation amplitude to obtain the environmental time-varying law; Based on the time-varying environmental patterns, cross-correlation calculations are performed on the historical water quality data to obtain the correlation evolution matrix; The correlation evolution matrix is ​​divided into paths to obtain abnormal isolation paths. The abnormal isolation paths are evaluated to obtain coupling distortion scores. Cooperative stress features are generated based on the coupling distortion scores and a preset distortion judgment threshold. Based on the aforementioned collaborative stress characteristics, a warning offset analysis is performed to obtain a warning offset vector. Based on the warning offset vector, an ecological assessment is conducted to obtain the ecological carrying capacity margin. Based on the ecological carrying capacity margin, the warning offset vector is corrected to obtain the dynamic ecological boundary. Real-time water quality data are obtained, and risk assessment is performed based on the real-time water quality data and the dynamic ecological boundary to obtain risk warning information.

2. The aquaculture risk early warning method based on environmental data according to claim 1, characterized in that, The process of acquiring historical water quality data and performing time-series state extraction processing on the historical water quality data to obtain water quality time-series features includes: Obtain historical water quality data, perform interpolation and imputation on the historical water quality data, and perform normalization processing to obtain a normalized matrix; Perform window truncation on the normalized matrix to obtain the standard input tensor; The hidden layer state is obtained by updating the state of the standard input tensor. Based on the hidden layer state, feature extraction is performed at the final time step to obtain water quality time-series features.

3. The aquaculture risk early warning method based on environmental data according to claim 1, characterized in that, The step of extracting the fluctuation amplitude based on the water quality time-series characteristics and performing frequency domain evolution analysis on the fluctuation amplitude to obtain the environmental time-varying law includes: Extreme value differences are extracted based on the aforementioned water quality time-series characteristics to obtain the fluctuation amplitude; Calculate the rate of change of the fluctuation amplitude. If the rate of change is greater than a preset rate determination threshold, then perform a discrete time-domain transformation on the fluctuation amplitude to obtain the frequency domain component. Calculate the squared magnitude of the frequency domain component to obtain the energy spectral density; A state evolution matrix is ​​constructed based on the energy spectral density, and the state evolution matrix is ​​then optimized by sequence decoding to obtain the time-varying laws of the environment.

4. The aquaculture risk early warning method based on environmental data according to claim 1, characterized in that, The step of performing cross-correlation calculations on the historical water quality data based on the time-varying environmental patterns to obtain a correlation evolution matrix includes: Based on the time-varying environmental pattern, the lag time step is extracted, and the historical water quality data is time-series aligned based on the lag time step to obtain an alignment factor sequence. The alignment factor sequence is cross-calculated to obtain the coupling response characteristics; The coupled response features are weighted and mapped according to the pre-acquired dynamic weight vector to obtain the response intensity distribution; Nodes are extracted based on the response intensity distribution to obtain evolved nodes, and node connection edges are constructed based on the spatial Euclidean distance between the evolved nodes. Based on the feature association analysis of the evolution nodes and the connecting edges between the nodes, the association evolution matrix is ​​obtained.

5. The aquaculture risk early warning method based on environmental data according to claim 1, characterized in that, The process of partitioning the correlation evolution matrix to obtain abnormal isolation paths, and evaluating these abnormal isolation paths to obtain coupling distortion scores, includes: The correlation evolution matrix is ​​decomposed into eigenvalues ​​to obtain the topological features and edge weight variation values. Based on the topological features and the variation values ​​of edge weights, spatial recursion is performed to obtain abnormal isolation paths; Based on the abnormal isolation path, hierarchical numerical extraction is performed to obtain a path depth set, and the average depth value of the path depth set is calculated. The difference between the average depth value and the pre-acquired evaluation benchmark value is calculated to obtain the state deviation magnitude; The deviation magnitude of the state is mapped to obtain the coupling distortion score.

6. The aquaculture risk early warning method based on environmental data according to claim 1, characterized in that, The step of generating cooperative stress features based on the coupled distortion score and a preset distortion determination threshold includes: Obtain a pre-recorded stress response sequence, perform differential calculation on the stress response sequence, and obtain the response gradient; The response gradients are clustered to obtain gradient cluster centers; The dispersion is calculated based on the gradient cluster centers to obtain the cluster dispersion. If the clustering dispersion is less than a preset threshold, an abnormal behavior sequence of the organism is obtained. Then, feature fusion and splicing are performed based on the stress response sequence and the abnormal behavior sequence to obtain a superimposed evolutionary trajectory. The overlap index of the superimposed evolution trajectory is extracted, and a cooperative stress feature is generated based on the overlap index.

7. The aquaculture risk early warning method based on environmental data according to claim 1, characterized in that, The process involves performing a warning offset analysis based on the collaborative stress characteristics to obtain a warning offset vector, conducting an ecological assessment based on the warning offset vector to obtain an ecological carrying capacity margin, and correcting the warning offset vector based on the ecological carrying capacity margin to obtain a dynamic ecological boundary, including: The feature variance contribution rate is calculated based on the collaborative stress characteristics to obtain the factor superposition weights; A linear mapping is performed on the preset independent warning values ​​based on the superimposed weights of the factors to obtain the warning offset vector; The metabolic intensity of the water body is obtained, and a fitting and extrapolation calculation is performed based on the metabolic intensity of the water body and the warning offset vector to obtain the metabolic consumption rate. If the metabolic consumption rate is greater than the preset metabolic baseline threshold, the ecological carrying capacity is calculated based on the pre-acquired characteristics of biological stock and total dissolved volume. A floating correction range is set according to the ecological carrying capacity, and a smoothing correction process is performed on the warning offset vector according to the floating correction range to obtain the dynamic ecological boundary.

8. The aquaculture risk early warning method based on environmental data according to claim 1, characterized in that, The process of acquiring real-time water quality data, assessing risks based on these data and the dynamic ecological boundary, and obtaining risk warning information includes: Obtain real-time water quality values, calculate the difference between the real-time water quality values ​​and the dynamic ecological boundary, and obtain the boundary deviation magnitude; A state sequence is constructed based on the boundary deviation magnitude, and the transition probability is deduced from the state sequence to obtain a risk evolution sequence; The risk distribution matrix is ​​obtained by performing a two-dimensional mapping between the risk evolution sequence and a pre-defined spatial grid. Risk analysis is performed based on the risk distribution matrix to obtain risk warning information.

9. A livestock farming risk early warning system based on environmental data, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the aquaculture risk early warning method based on environmental data as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the aquaculture risk early warning method based on environmental data as described in any one of claims 1 to 8.

Images