Marine aquaculture shellfish toxin high risk early warning system

Abstract

The present application relates to the technical field of seawater culture environment monitoring and signal alarm, in particular to a high-risk early warning system for shellfish toxins in seawater culture. An environmental parameter collection node and an underwater fluorescence spectrum collection node respectively collect seawater current field vectors, meteorological disturbance data and algal fluorescence spectrum signals and transmit them to a warning server; the warning server constructs a multi-modal signal space-time alignment model, performs time stamp synchronization and space coordinate mapping on the physical field disturbance period and the spectrum characteristic change period to generate space-time alignment data; a long short-term memory network and a time domain convolution coupled feature extraction network are used to extract the lag driving characteristics of the physical field on the proliferation of algae and the nonlinear drift characteristics of the spectrum characteristic peak, and cross-fusion is performed in the hidden space to generate a toxin outbreak risk preposition representation value. The present application overcomes the false alarm defect of single parameter monitoring, triggers an alarm before a large amount of toxins are generated, and improves the preposition of alarm timing and recognition accuracy.

Description

Technical Field

[0001] This invention relates to the field of marine aquaculture environment monitoring and signal alarm technology, specifically a high-risk early warning system for shellfish toxins in marine aquaculture. Background Technology

[0002] Marine shellfish aquaculture is a vital pillar industry of my country's marine fisheries economy, and its scale has been continuously expanding in recent years. However, with the intensification of changes in the marine environment, the frequency of harmful algal blooms (HABs) has increased significantly, and shellfish toxin pollution caused by toxic algae has become a key bottleneck restricting the healthy development of the shellfish aquaculture industry. Shellfish toxins are characterized by high toxicity, short incubation period, and the lack of specific antidotes. Consuming shellfish contaminated with toxins can cause various poisoning symptoms such as paralysis, diarrhea, and amnesia, and in severe cases, can lead to respiratory failure and even death.

[0003] Shellfish toxins are not produced by the shellfish themselves, but rather by filter-feeding shellfish that accumulate and transform secondary metabolites produced by toxic microalgae in their bodies after consuming these algae. Common toxic-producing algae include Alexandrium, Gymnodinium, and Pterygophyta (dinoflagellates), Pseudo-Nyctaginosa (diatoms), and certain species of cyanobacteria. When the concentration of toxic algae in seawater reaches a certain threshold, shellfish can accumulate high concentrations of toxins in their digestive glands, muscles, and other tissues within a short period through continuous filter feeding. Because shellfish have a strong tolerance to most algal toxins, and the toxins are metabolized slowly within their bodies with a half-life of several weeks to months, even after the toxic algae in the aquaculture water have died out, the toxin levels in the shellfish may still remain above safe levels, posing a long-term threat to food safety.

[0004] Currently, early warning systems for shellfish toxins in mariculture mainly rely on two types of technologies. The first type is indirect monitoring technology based on physical environmental parameters. This involves deploying sensors in the aquaculture area to collect data on water temperature, salinity, ocean current vectors, and meteorological disturbances. The monitored values ​​are compared with preset empirical thresholds, and an alarm is triggered when a single parameter exceeds the range. This approach is based on the fundamental principle that the growth of toxin-producing algae is regulated by environmental factors. However, there is a complex nonlinear relationship and a significant time lag effect between changes in the physical environment and the proliferation of toxin-producing algae. A single threshold comparison cannot accurately reflect the intrinsic driving mechanism of algal growth, leading to frequent false alarms and missed alarms.

[0005] The second type is direct detection technology based on underwater fluorescence spectroscopy. This method uses specific wavelengths of excitation light to excite the chlorophyll fluorescence of algae in the water, and identifies the species of toxin-producing algae and estimates their concentration by analyzing the fluorescence spectral characteristics. An alert is issued when the intensity of the characteristic spectral peak of the toxin-producing algae exceeds a safe threshold. This method directly monitors the toxin-producing organism itself, which is more direct than monitoring environmental parameters, but it has significant limitations: First, the spectral signal is easily interfered with by suspended matter, dissolved organic matter, and background light in the water, resulting in unstable detection accuracy; second, it can only obtain the physiological state of the algae at the current moment and cannot predict the driving trend of environmental changes on the subsequent proliferation of algae; third, when the characteristic spectral peak shows a significant shift, the toxin has often already begun to accumulate in large quantities in the shellfish, resulting in a serious delay in the alert. Summary of the Invention

[0006] The purpose of this invention is to provide a high-risk early warning system for shellfish toxins in marine aquaculture, which can effectively solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A high-risk early warning system for shellfish toxins in marine aquaculture includes environmental parameter acquisition nodes, underwater fluorescence spectroscopy acquisition nodes, and an early warning server. The environmental parameter acquisition node acquires ocean current field vector data and meteorological disturbance data, and the underwater fluorescence spectroscopy acquisition node acquires algal fluorescence spectral signals excited by specific wavelengths underwater, and transmits the ocean current field vector data, the meteorological disturbance data and the algal fluorescence spectral signals to the early warning server; The early warning server constructs a multimodal signal spatiotemporal alignment model, and uses the multimodal signal spatiotemporal alignment model to synchronize the physical field disturbance period with the spectral feature change period with timestamps and spatial coordinate mapping to generate spatiotemporal aligned data; The early warning server uses a feature extraction network based on long short-term memory network and temporal convolution coupling to process the spatiotemporal aligned data, extract the hysteretic driving features of physical field perturbation on algal proliferation and the nonlinear drift features of the characteristic peaks of toxin-producing algae in the fluorescence spectrum, and perform feature cross-fusion of the hysteretic driving features and the nonlinear drift features in the latent space to generate a pre-characteristic value of toxin outbreak risk. The early warning server compares the pre-risk characterization value of the toxin outbreak with the dynamic baseline threshold. When the pre-risk characterization value of the toxin outbreak exceeds the dynamic baseline threshold, it generates an alarm command corresponding to the risk level and sends it to the display terminal in the breeding area.

[0008] Preferably, the early warning server uses the multimodal signal spatiotemporal alignment model to synchronize the physical field disturbance period with the spectral feature change period using timestamps and spatial coordinate mapping to generate spatiotemporal aligned data, including: the early warning server extracts the flow field topology constraint boundary from the ocean current field vector data, and calculates the fluid transport lag time function between the environmental parameter acquisition node and the underwater fluorescence spectroscopy acquisition node based on the flow field topology constraint boundary; The early warning server performs reverse compensation and stretching operations on the meteorological disturbance data and the algal fluorescence spectral signal in the time dimension based on the fluid transport lag time function, aligning the meteorological data response sequence and the spectral data response sequence caused by the same physical disturbance event to the same time axis reference point. The early warning server constructs a spatial coordinate system mapping matrix based on the Lagrange particle trajectories in the ocean current field vector data, and projects the time-aligned meteorological disturbance data and the algal fluorescence spectral signal onto a unified spatial grid to generate the spatiotemporally aligned data.

[0009] Preferably, the early warning server employs a feature extraction network based on long short-term memory network coupled with temporal convolution to extract the lag driving features of physical field disturbance on algal proliferation, including: the early warning server constructs a spatiotemporal graph structure from the ocean current field vector data and the meteorological disturbance data in the spatiotemporal aligned data, wherein the nodes are spatial grid positions and the edges are fluid connectivity relationships; The early warning server applies a spatiotemporal graph convolution operation to the spatiotemporal graph structure to extract the initial propagation features of meteorological disturbances under spatial topological correlation. The early warning server inputs the initial propagation characteristics of the meteorological disturbance into a long short-term memory network configured with an adaptive delay matrix. The adaptive delay matrix dynamically adjusts the delay receptive field scale according to the Reynolds number of the flow field, calculates the state transition correlation matrix with time lag effect, and outputs the lag driving characteristics of the physical field disturbance on algal proliferation.

[0010] Preferably, the early warning server employs a feature extraction network based on long short-term memory network coupled with temporal convolution to extract the nonlinear drift features of the characteristic peaks of toxic algae in the fluorescence spectrum, including: the early warning server performs multi-scale wavelet packet decomposition on the algal fluorescence spectrum signal in the spatiotemporally aligned data to obtain spectral sub-band signals at different frequency band levels; The early warning server calculates the wavelet packet energy entropy of each of the spectral sub-band signals, and constructs a frequency band energy distribution matrix that evolves over time based on the wavelet packet energy entropy; The early warning server inputs the frequency band energy distribution matrix into an extended causal convolutional network, and uses the receptive field corresponding to different expansion rates to extract the nonlinear shift trajectory vectors of the feature peak center frequency at different time scales. The nonlinear shift trajectory vectors at each scale are combined to form the nonlinear drift characteristics of the toxic algae feature peaks in the fluorescence spectrum.

[0011] Preferably, the early warning server performs feature cross-fusion of the hysteresis driving feature and the nonlinear drift feature in the latent space to generate a pre-characterization value of toxin outbreak risk, including: the early warning server performs orthogonal projection decomposition on the hysteresis driving feature and the nonlinear drift feature respectively to obtain the principal component projection matrix and orthogonal residual matrix of the corresponding physical field and the principal component projection matrix and orthogonal residual matrix of the corresponding spectral field; The early warning server performs a Hadamard product interaction between the principal component projection matrix of the physical field and the orthogonal residual matrix of the spectral field to generate a first cross feature, and performs a Hadamard product interaction between the principal component projection matrix of the spectral field and the orthogonal residual matrix of the physical field to generate a second cross feature. The early warning server concatenates and fully connects the first cross feature and the second cross feature to generate the pre-characteristic value of the toxin outbreak risk.

[0012] Preferably, the comparison between the pre-risk characterization value of the toxin outbreak and the dynamic baseline threshold by the early warning server includes: the early warning server extracting historical pre-risk characterization values ​​of the toxin outbreak based on a sliding time window to construct a historical risk feature distribution sequence; The early warning server performs kernel density estimation on the historical risk feature distribution sequence to obtain a nonparametric probability density function; The early warning server selects the feature value whose cumulative probability reaches a preset quantile on the nonparametric probability density function as the initial baseline reference point. The early warning server calculates the environmental fluctuation correction coefficient based on the current ocean current field vector data velocity scalar and the air pressure gradient change rate of the meteorological disturbance data, and multiplies the initial baseline reference point with the environmental fluctuation correction coefficient to obtain the dynamic baseline threshold.

[0013] Preferably, the early warning server extracts the flow field topological constraint boundary from the ocean current field vector data, and calculates the fluid transport lag time function between the environmental parameter acquisition node and the underwater fluorescence spectroscopy acquisition node based on the flow field topological constraint boundary, including: the early warning server calculates a finite-time Lyapunov exponential field based on the ocean current field vector data, and extracts ridges with exponent values ​​greater than a preset threshold in the finite-time Lyapunov exponential field as Lagrange coherent structures; The early warning server divides the Lagrange coherent structure into topological boundaries that separate water transport, and calculates the shortest fluid trajectory path of fluid particles from the environmental parameter acquisition node to the underwater fluorescence spectrum acquisition node along the topological boundaries. The early warning server integrates the flow velocity vector along the shortest fluid trajectory path to generate the fluid transport lag time function, which varies with spatial position and time.

[0014] Preferably, the adaptive delay matrix dynamically adjusts the delay sensing field scale based on the flow field Reynolds number, which includes: the early warning server calculating the local flow field Reynolds number based on the ocean current field vector data, and mapping the local flow field Reynolds number to a delay control factor; The early warning server constructs a delay memory unit containing multiple parallel delay branches, each of which is configured with a different fixed delay step size; The early warning server uses the delay control factor to weight and fuse the output states of each of the parallel delay branches. When the local flow field Reynolds number increases, the weight of the parallel delay branch with a small delay step size is increased to shrink the delay receptive field scale to extract recent sudden disturbance correlation features. When the local flow field Reynolds number decreases, the weight of the parallel delay branch with a large delay step size is increased to expand the delay receptive field scale to extract long-term lag correlation features.

[0015] Preferably, the early warning server calculates the wavelet packet energy entropy of each of the spectral sub-band signals and constructs a frequency band energy distribution matrix that evolves over time based on the wavelet packet energy entropy, including: for each of the spectral sub-band signals, the early warning server calculates the proportion of signal energy to the total energy of the entire frequency band as a relative energy entropy value; The early warning server moves an analysis window of a fixed length along the time axis, extracts the relative energy entropy values ​​of all the spectral sub-band signals within each analysis window, and constructs a column vector characterizing the frequency band energy distribution. The early warning server concatenates the column vectors obtained in continuous time steps in chronological order to form the frequency band energy distribution matrix, in which the row direction represents time evolution and the column direction represents the frequency band level. The early warning server performs a difference operation on the frequency band energy distribution matrix to extract the transient migration vector of the characteristic peak energy between different frequency bands.

[0016] Preferably, the early warning server performs Hadamard product interaction between the principal component projection matrix of the physical field and the orthogonal residual matrix of the spectral field to generate a first cross feature, and performs Hadamard product interaction between the principal component projection matrix of the spectral field and the orthogonal residual matrix of the physical field to generate a second cross feature, including: the early warning server performs singular value decomposition on the hysteresis driving feature and the nonlinear drift feature respectively, extracts the left singular vectors corresponding to the preceding preset number of singular values ​​to construct the principal component projection matrix, and constructs the orthogonal residual matrix based on the left singular vectors corresponding to the remaining singular values; The early warning server performs element-wise multiplication of the row vectors of the principal component projection matrix of the physical field with the row vectors of the orthogonal residual matrix of the spectral field to generate the first cross feature, and performs element-wise multiplication of the row vectors of the principal component projection matrix of the spectral field with the row vectors of the orthogonal residual matrix of the physical field to generate the second cross feature. The early warning server applies polarization regularization constraints to the first cross feature and the second cross feature, thereby maximizing the projection modulus of the cross feature vector in the orthogonal subspace and generating an orthogonalized cross feature vector.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs a multimodal signal spatiotemporal alignment model to synchronize the physical field disturbance period with the spectral feature change period through timestamp synchronization and spatial coordinate mapping. Based on the flow field topological constraint boundary, the fluid transport lag time function is calculated, and inverse compensation and stretching operations are performed. Spatial grid projection is then performed under the Lagrange particle trajectory, eliminating the spatiotemporal misalignment between physical environment changes and algal spectral responses. By combining a long short-term memory network with a temporal convolution-coupled feature extraction network, the lag-driven features of physical field disturbances on algal proliferation are extracted. This overcomes the limitation of single environmental parameter threshold comparisons in reflecting complex nonlinear driving mechanisms, achieving an accurate characterization of environmental-driven lag effects.

[0018] 2. This invention performs multi-scale wavelet packet decomposition on algal fluorescence spectral signals and constructs a frequency band energy distribution matrix. By using an extended causal convolutional network, it extracts the nonlinear shift trajectory vectors of the characteristic peak center frequencies at different time scales, thus obtaining the nonlinear drift characteristics of the characteristic peaks. After orthogonally projecting the hysteresis-driven features and the nonlinear drift features, it performs Hadamard product interaction between the principal component projection matrix and the orthogonal residual matrix, achieving deep feature cross-fusion of the physical field and the spectral field in the latent space. This cross-fusion method reveals the intrinsic coupling law between environmental drivers and biological responses, generating a pre-emergence characterization value for toxin outbreak risk, triggering an alarm before the actual large-scale generation of toxins.

[0019] 3. This invention extracts historical risk precursor values ​​based on a sliding time window, performs kernel density estimation to obtain a nonparametric probability density function, and selects the feature value whose cumulative probability reaches a preset quantile as the initial baseline reference point. An environmental fluctuation correction coefficient is calculated based on the current ocean current velocity scalar and the rate of change of atmospheric pressure gradient due to meteorological disturbances, and the initial baseline reference point is corrected to obtain a dynamic baseline threshold. This threshold generation method adaptively adjusts with the dynamic fluctuations of the marine environment, avoiding misjudgments under complex and variable sea conditions caused by fixed thresholds, and improving the robustness of risk comparison and judgment and the stability of alarm output. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the overall operation of the high-risk early warning system for shellfish toxins in marine aquaculture according to the present invention. Figure 2 This is a flowchart of the multimodal signal spatiotemporal alignment data generation process of the present invention; Figure 3 This is a flowchart illustrating the hysteresis-driven feature extraction of algal proliferation by physical field perturbation according to the present invention. Figure 4 This is a flowchart of the nonlinear drift feature extraction process for the characteristic peaks of toxin-producing algae in the fluorescence spectrum of this invention; Figure 5 This is a flowchart of the process for generating pre-characterization values ​​of toxin outbreak risk through the cross-fusion of latent spatial features in this invention; Figure 6 This is a flowchart of the dynamic baseline threshold calculation and risk alarm instruction generation process of the present invention. Detailed Implementation

[0021] 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, not all, of the embodiments of the present invention. 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.

[0022] Please refer to Figure 1This embodiment provides a high-risk early warning system for shellfish toxins in marine aquaculture, including environmental parameter acquisition nodes, underwater fluorescence spectroscopy acquisition nodes, and an early warning server. The environmental parameter acquisition nodes are deployed at different hydrological locations within the aquaculture area to collect ocean current field vector data and meteorological disturbance data. The ocean current field vector data includes the magnitude and direction components of the current velocity, while the meteorological disturbance data includes time-series data of air pressure, air temperature, wind speed, and wind direction. The underwater fluorescence spectroscopy acquisition nodes are deployed at different depths within the aquaculture water body, emitting excitation light of specific wavelengths to collect underwater algal fluorescence spectral signals excited at those specific wavelengths. The wavelength range of the fluorescence spectral signals covers 400 nm to 700 nm, with a spectral resolution of 1 nm. The environmental parameter acquisition nodes and the underwater fluorescence spectroscopy acquisition nodes transmit the ocean current field vector data, meteorological disturbance data, and algal fluorescence spectral signals to the early warning server via a wireless communication network.

[0023] After receiving multi-source data, the early warning server constructs a multimodal signal spatiotemporal alignment model. The input to this model is the raw data from environmental parameter acquisition nodes and underwater fluorescence spectroscopy acquisition nodes, and the output is spatiotemporally aligned data after time synchronization and spatial mapping. The early warning server uses this model to timestamp and spatially map the physical field disturbance period and the spectral feature change period, generating spatiotemporally aligned data. The physical field disturbance period refers to the time interval during which significant changes occur in the ocean current field and meteorological field, while the spectral feature change period refers to the time interval during which detectable changes occur in the fluorescence spectral characteristics of algae. Since a certain fluid transport time is required after a physical disturbance event to affect the algae at the location of the underwater fluorescence spectroscopy acquisition node, there is a time lag between the physical field disturbance period and the spectral feature change period. Furthermore, because different acquisition nodes are located in different spatial locations, the impact of physical disturbances propagates differently in space, necessitating spatial coordinate mapping.

[0024] The early warning server employs a feature extraction network based on a long short-term memory (LSTM) network coupled with temporal convolution to process spatiotemporally aligned data. This feature extraction network comprises two parallel branches: the first branch processes ocean current field vector data and meteorological disturbance data from the spatiotemporally aligned data, extracting the hysteretic driving features of physical field disturbances on algal proliferation; the second branch processes algal fluorescence spectral signals from the spatiotemporally aligned data, extracting the nonlinear drift features of toxin-producing algal characteristic peaks in the fluorescence spectra. The LSM network is used to capture long-term dependencies in the time series, while temporal convolution is used to extract feature changes at local time scales; the coupling of these two technologies enables the simultaneous capture of feature information at different time scales.

[0025] The early warning server performs feature cross-fusion of hysteresis-driven features and nonlinear drift features in the latent space to generate a preliminary characterization value for the risk of toxin outbreaks. This feature cross-fusion process is achieved through the interaction of orthogonal projection decomposition and Hadamard product, revealing the intrinsic coupling relationship between physical field perturbations and algal spectral responses. The preliminary characterization value for the risk of toxin outbreaks is a scalar value, the magnitude of which reflects the likelihood of a shellfish toxin outbreak in the near future.

[0026] The early warning server compares the pre-risk indicator value for toxin outbreaks with a dynamic baseline threshold. The dynamic baseline threshold is dynamically generated based on historical risk characteristics and current environmental fluctuations, rather than a fixed empirical value. When the pre-risk indicator value for a toxin outbreak exceeds the dynamic baseline threshold, the early warning server generates an alarm command corresponding to the risk level and sends it to the display terminal in the aquaculture area via a wireless communication network. Upon receiving the alarm command, the display terminal in the aquaculture area displays the risk level and early warning information to the aquaculture personnel using a combination of sound and light.

[0027] In this embodiment, the deployment density of environmental parameter acquisition nodes and underwater fluorescence spectroscopy acquisition nodes is determined based on the area and hydrological conditions of the aquaculture area. For aquaculture areas with large areas and complex hydrological conditions, the number of acquisition nodes is increased to improve the spatial resolution of data acquisition. The data sampling frequency of the acquisition nodes is determined according to monitoring needs: the sampling frequency for ocean current field vector data and meteorological disturbance data is once every 10 minutes, and the sampling frequency for algal fluorescence spectral signals is once every 30 minutes. The early warning server preprocesses the received raw data, including outlier removal and missing value imputation. Outlier removal adopts the 3σ criterion, and data exceeding the mean ± 3 times the standard deviation are considered outliers and removed; missing value imputation adopts linear interpolation, and missing values ​​are calculated based on data from adjacent time points. The preprocessing parameter settings for multimodal raw data are shown in Table 1.

[0028] Table 1. Preprocessing parameter settings for multimodal raw data

[0029] Table 1 shows the preprocessing parameter settings for the multimodal raw data. These preprocessing operations remove noise and outliers from the raw data, ensuring the accuracy of subsequent data processing. The preprocessed data is stored in the database of the early warning server for use in subsequent model calls and analysis.

[0030] In a preferred embodiment, reference Figure 2The early warning server utilizes a multimodal signal spatiotemporal alignment model to synchronize the physical field disturbance period with the spectral feature change period using timestamps and spatial coordinate mapping, generating spatiotemporally aligned data through the following processes: The early warning server extracts the flow field topological constraint boundaries from the ocean current field vector data and calculates the fluid transport lag time function between the environmental parameter acquisition node and the underwater fluorescence spectroscopy acquisition node based on these boundaries. The flow field topological constraint boundaries refer to the boundaries separating different water body transport regions; these boundaries are determined by the dynamic characteristics of the flow field and can limit the fluid transport path. The fluid transport lag time function describes the time required for a fluid particle to travel from the location of the environmental parameter acquisition node to the location of the underwater fluorescence spectroscopy acquisition node; this time varies with spatial location and time.

[0031] The early warning server performs reverse compensation and stretching operations in the time dimension on meteorological disturbance data and algal fluorescence spectral signals based on the fluid transport lag time function. Reverse compensation involves shifting the timestamp of the algal fluorescence spectral signal forward by the corresponding fluid transport lag time, aligning the meteorological data response sequence and the spectral data response sequence triggered by the same physical disturbance event to the same time axis reference point. Stretching involves stretching or compressing the spectral data sequence along the time axis as the fluid transport lag time changes, to match the time scale of the meteorological data sequence. Through these reverse compensation and stretching operations in the time dimension, the time lag between the physical disturbance event and the algal spectral response can be eliminated, ensuring they remain synchronized on the time axis.

[0032] The early warning server constructs a spatial coordinate system mapping matrix based on the Lagrange particle trajectories in the ocean current field vector data. The Lagrange particle trajectories describe the motion of fluid particles in the flow field; by tracking the trajectories of a large number of Lagrange particles, the fluid connectivity between different spatial locations can be determined. The spatial coordinate system mapping matrix projects the original spatial coordinates of environmental parameter acquisition nodes and underwater fluorescence spectroscopy acquisition nodes onto a unified spatial grid. The unified spatial grid uses a regular rectangular grid, and the grid resolution is determined based on the area of ​​the aquaculture area and monitoring requirements. Time-aligned meteorological disturbance data and algal fluorescence spectral signals are projected onto the unified spatial grid to generate spatiotemporally aligned data. The spatiotemporally aligned data is a four-dimensional data structure, including a time dimension, the x-dimensional of the spatial grid, the y-dimensional of the spatial grid, and a feature dimension.

[0033] In this embodiment, the early warning server extracts the topological constraint boundary of the ocean current field vector data, and calculates the fluid transport lag time function between the environmental parameter acquisition node and the underwater fluorescence spectroscopy acquisition node based on the topological constraint boundary. This process includes the following steps: The early warning server calculates the finite-time Lyapunov exponent field based on the ocean current field vector data. The finite-time Lyapunov exponent describes the separation rate between adjacent fluid particles in the flow field and can characterize the mixing properties of the flow field. The formula for calculating the finite-time Lyapunov exponent is:

[0034] in, This represents the initial position of the fluid particles. The initial time, For integration time, The distance between adjacent fluid particles at the initial moment. For the time elapsed The distance between adjacent fluid particles.

[0035] The early warning server extracts ridges in the finite-time Lyapunov exponential field where the exponential value exceeds a preset threshold as Lagrange coherent structures. A Lagrange coherent structure is a curve or surface in the flow field with special dynamic characteristics, capable of separating different fluid regions and restricting cross-regional fluid transport. The preset threshold is determined based on the characteristics of the flow field, typically twice the average value of the finite-time Lyapunov exponential field. The early warning server divides the Lagrange coherent structure into topological boundaries separating water transport and calculates the shortest fluid trajectory path for fluid particles from the environmental parameter acquisition node to the underwater fluorescence spectroscopy acquisition node along these topological boundaries. The shortest fluid trajectory path refers to the path that takes the shortest time for a fluid particle to travel from its starting point to its destination under the influence of the flow field.

[0036] The early warning server integrates the velocity vector along the shortest fluid trajectory path to generate a fluid transport lag time function that varies with both spatial location and time. The formula for calculating the fluid transport lag time function is:

[0037] in, and For spatial coordinates, For time, The length of the shortest fluid trajectory path. Let arc length be the parameter along the shortest fluid trajectory path. For time Time arc length position The flow velocity at the specified location. The parameter settings for flow field topology analysis and fluid transport hysteresis calculation are shown in Table 2.

[0038] Table 2 Parameter Settings for Flow Field Topology Analysis and Fluid Transport Hysteresis Calculation

[0039] Table 2 shows the parameter settings for flow field topology analysis and fluid transport lag calculation. The fluid transport lag time function calculated through the above process can accurately reflect the transport time characteristics of physical disturbances in the flow field, providing a basis for subsequent time dimension alignment.

[0040] In a preferred embodiment, reference Figure 3 The early warning server employs a feature extraction network based on a long short-term memory network coupled with temporal convolution to extract the lagged driving features of physical field perturbations on algal proliferation, including the following process: The early warning server constructs a spatiotemporal graph structure from ocean current field vector data and meteorological perturbation data in the spatiotemporally aligned data. The nodes of the spatiotemporal graph structure are each grid location in a unified spatial grid, and the edges represent the fluid connectivity between adjacent grid locations. Fluid connectivity is determined by Lagrange particle trajectories; if a fluid particle is transported from one grid location to another, an edge connects these two grid locations. The weight of the edge is the probability of fluid transport; the higher the transport probability, the greater the edge weight.

[0041] The early warning server applies spatiotemporal graph convolution operations to the spatiotemporal graph structure to extract the initial propagation features of meteorological disturbances under spatial topological correlations. Spatiotemporal graph convolution operations consider both spatial topological correlations and temporal sequence correlations, enabling the capture of both the spatial propagation patterns and temporal evolution patterns of meteorological disturbances. The formula for calculating spatiotemporal graph convolution is:

[0042] in, For the first The input feature matrix of the layer, For the first The output feature matrix of the layer, To add self-loops to the adjacency matrix, for The degree matrix, For the first The learnable weight matrix of the layer, This is the activation function.

[0043] The early warning server inputs the initial propagation characteristics of meteorological disturbances into a Long Short-Term Memory (LSTM) network configured with an adaptive delay matrix. The adaptive delay matrix dynamically adjusts the delay receptive field scale based on the Reynolds number of the flow field, adapting to the hysteresis driving effect under different flow field conditions. The basic unit of the LSM network includes an input gate, a forget gate, an output gate, and a cell state, effectively capturing long-term dependencies in time series. The LSM network configured with the adaptive delay matrix adds multiple parallel delay branches to the traditional LSM network, each with a different fixed delay step size.

[0044] The early warning server calculates a state transition correlation matrix with a time lag effect, outputting the lagged driving characteristics of physical field disturbances on algal proliferation. The state transition correlation matrix describes the degree of influence of meteorological disturbance characteristics at different time steps on the current algal proliferation state. Through dynamic adjustment of the adaptive delay matrix, the Long Short-Term Memory network can automatically select an appropriate delay receptive field scale according to the flow field conditions, accurately extracting driving characteristics with different time lags.

[0045] In this embodiment, the adaptive delay matrix dynamically adjusts the delay receptive field scale based on the flow field Reynolds number, including the following process: The early warning server calculates the local flow field Reynolds number based on ocean current field vector data. The flow field Reynolds number is a dimensionless number characterizing the fluid flow state and reflecting the degree of turbulence in the flow field. The formula for calculating the local flow field Reynolds number is:

[0046] in, The density of seawater, For the magnitude of the local flow velocity, For characteristic length, The dynamic viscosity of seawater.

[0047] The early warning server maps the local flow field Reynolds number to a delay control factor. The delay control factor ranges from 0 to 1, and the mapping is achieved using the sigmoid function.

[0048] in, The mapping slope, The threshold is the Reynolds number.

[0049] The early warning server constructs a delay memory unit containing multiple parallel delay branches, each configured with a different fixed delay step size. The delay step size is determined based on the actual application scenario, for example, set to 1 hour, 3 hours, 6 hours, 12 hours, and 24 hours. The early warning server uses a delay control factor to weight and fuse the output states of each parallel delay branch. The formula for weighted fusion is:

[0050] in, For the current moment's fusion output, The number of parallel delay branches, For the first The weights of each delayed branch, For the first Fixed delay step size for each delay branch For the first The output status of each delayed branch.

[0051] When the Reynolds number of the local flow field increases, the turbulence of the flow field intensifies, the transmission speed of physical disturbances accelerates, and the lag time shortens. In this case, increasing the weight of the parallel delay branch with a small delay step size shrinks the delay receptive field scale to extract recent sudden disturbance correlation features. Conversely, when the Reynolds number of the local flow field decreases, the turbulence of the flow field weakens, the transmission speed of physical disturbances slows down, and the lag time lengthens. In this case, increasing the weight of the parallel delay branch with a large delay step size expands the delay receptive field scale to extract long-term lag correlation features. The parameter settings of the adaptive delay long short-term memory network are shown in Table 3.

[0052] Table 3. Parameter settings for the adaptive delayed long short-term memory network

[0053] Table 3 shows the parameter settings for the adaptively delayed Long Short-Term Memory (LSTM) network. Through the aforementioned adaptive delay mechanism, the LSM network can automatically adjust the time scale of feature extraction according to the dynamic changes in the flow field, accurately capturing the hysteretic driving effect of physical field disturbances on algal proliferation.

[0054] In a preferred embodiment, reference Figure 4 The early warning server employs a feature extraction network based on a long short-term memory network coupled with temporal convolution to extract the nonlinear drift features of the characteristic peaks of toxic algae in the fluorescence spectrum. The process includes the following steps: The early warning server performs multi-scale wavelet packet decomposition on the algal fluorescence spectral signals in the spatiotemporally aligned data to obtain spectral sub-band signals at different frequency levels. Multi-scale wavelet packet decomposition can decompose the original spectral signal into multiple sub-band signals with different frequency ranges, each containing spectral feature information at different scales. The wavelet packet decomposition uses the Daubechies wavelet as the mother wavelet, and the number of decomposition levels is determined based on the characteristics of the spectral signal.

[0055] The early warning server calculates the wavelet packet energy entropy of each spectral sub-band signal and constructs a frequency band energy distribution matrix that evolves over time based on the wavelet packet energy entropy. The wavelet packet energy entropy characterizes the energy distribution properties of the spectral signal in different frequency bands, reflecting changes in spectral features. The rows of the frequency band energy distribution matrix correspond to time steps, the columns correspond to different frequency band levels, and the matrix elements are the wavelet packet energy entropy values ​​for the corresponding time step and frequency band level.

[0056] The early warning server inputs the frequency band energy distribution matrix into a dilated causal convolutional network (DCR). It then uses the receptive field corresponding to different dilation rates to extract nonlinear offset trajectory vectors of the feature peak center frequencies at different time scales. DCR expands the receptive field without increasing the number of parameters, while ensuring the causality of the convolution operation; that is, the output at the current time depends only on the input at the current time and previous time steps. Convolutional layers with different dilation rates can capture feature changes at different time scales, and combining the outputs of different dilation rates yields multi-scale feature information.

[0057] The early warning server combines nonlinear offset trajectory vectors at various scales to construct the nonlinear drift characteristics of the characteristic peaks of toxin-producing algae in the fluorescence spectrum. These nonlinear drift characteristics reflect the time-varying frequency of the characteristic peak center, a pattern closely related to the physiological state and toxin synthesis process of the toxin-producing algae.

[0058] In this embodiment, the early warning server calculates the wavelet packet energy entropy of each spectral sub-band signal and constructs a frequency band energy distribution matrix that evolves over time based on the wavelet packet energy entropy, including the following process: For each spectral sub-band signal, the early warning server calculates the proportion of the signal energy to the total energy of the entire frequency band as the relative energy entropy value. The formula for calculating the energy of a spectral sub-band signal is:

[0059] in, For the first The first spectral subband signal One sampling point, This represents the number of sampling points.

[0060] The total energy of the entire frequency band is the sum of the signal energies of all spectral subbands:

[0061] in, This represents the number of spectral subband signals.

[0062] No. The formula for calculating the relative energy entropy of a spectral sub-band signal is:

[0063] The early warning server moves a fixed-length analysis window along the time axis, extracting the relative energy entropy values ​​of all spectral sub-band signals within each window to construct a column vector characterizing the frequency band energy distribution. The length of the analysis window is determined based on monitoring requirements, for example, set to 24 hours. The early warning server concatenates the column vectors acquired at consecutive time steps in chronological order to form a frequency band energy distribution matrix, with rows representing temporal evolution and columns representing frequency band levels. The dimension of the frequency band energy distribution matrix is... ,in For the number of time steps, This represents the number of frequency band levels.

[0064] The early warning server performs a difference operation on the frequency band energy distribution matrix to extract the transient migration vectors of characteristic peak energy between different frequency bands. The difference operation can highlight the changes in frequency band energy distribution, and the transient migration vectors reflect the trend of characteristic peak energy migrating from one frequency band to another. The settings of fluorescence spectroscopy signal processing and feature extraction parameters are shown in Table 4.

[0065] Table 4. Parameter settings for fluorescence spectroscopy signal processing and feature extraction

[0066] Table 4 shows the parameter settings for fluorescence spectral signal processing and feature extraction. The nonlinear drift features extracted through the above process can accurately reflect the dynamic changes of the characteristic peaks of toxin-producing algae, providing a basis for subsequent risk assessment.

[0067] In a preferred embodiment, reference Figure 5 The early warning server performs feature cross-fusion of hysteresis-driven features and nonlinear drift features in the latent space to generate a preliminary characterization value for toxin outbreak risk, including the following process: The early warning server performs orthogonal projection decomposition on the hysteresis-driven features and nonlinear drift features respectively, obtaining the principal component projection matrix and orthogonal residual matrix of the corresponding physical field and the principal component projection matrix and orthogonal residual matrix of the corresponding spectral field. Orthogonal projection decomposition can decompose the original feature space into mutually orthogonal subspaces. The principal component projection matrix contains the main information of the original features, and the orthogonal residual matrix contains the remaining information of the original features.

[0068] The early warning server generates a first cross-feature by performing a Hadamard product between the principal component projection matrix of the physical field and the orthogonal residual matrix of the spectral field, and a second cross-feature by performing the same Hadamard product between the principal component projection matrix of the spectral field and the orthogonal residual matrix of the physical field. The Hadamard product, which is the element-wise product of two matrices, enables element-wise interaction between features. Through the cross-interaction between the principal components and the residuals, the nonlinear coupling relationship between the physical and spectral fields can be captured.

[0069] The early warning server concatenates and fully connects the first and second cross-features to generate a pre-emergence risk characterization value for toxin outbreaks. The concatenation operation joins the two cross-feature vectors into a longer feature vector, and the fully connected mapping maps the concatenated feature vector to a scalar value, i.e., the pre-emergence risk characterization value for toxin outbreaks.

[0070] In this embodiment, the early warning server performs Hadamard product interactions between the principal component projection matrix of the physical field and the orthogonal residual matrix of the spectral field to generate the first cross feature, and performs Hadamard product interactions between the principal component projection matrix of the spectral field and the orthogonal residual matrix of the physical field to generate the second cross feature, including the following process: The early warning server performs singular value decomposition on the hysteresis-driven feature and the nonlinear drift feature respectively. For the hysteresis-driven feature matrix... Its singular value decomposition is:

[0071] in, It is a left singular vector matrix. It is a singular value diagonal matrix. It is a right singular vector matrix.

[0072] The early warning server extracts the left singular vectors corresponding to a preset number of preceding singular values ​​to construct the principal component projection matrix, and constructs an orthogonal residual matrix based on the left singular vectors corresponding to the remaining singular values. The preset number is determined based on the variance contribution rate of the features, typically selecting the features with a variance contribution rate of 95%. There are singular values. The principal component projection matrix of the physical field. From the beginning The orthogonal residual matrix of the physical field is composed of several left singular vectors. It consists of the remaining left singular vectors.

[0073] Similarly, for the nonlinear drift characteristic matrix Its singular value decomposition is:

[0074] Principal component projection matrix of the spectral field From the beginning The orthogonal residual matrix of the spectral field is composed of several left singular vectors. It consists of the remaining left singular vectors.

[0075] The early warning server performs element-wise multiplication of the row vectors of the principal component projection matrix of the physical field with the row vectors of the orthogonal residual matrix of the spectral field to generate the first cross feature, and performs element-wise multiplication of the row vectors of the principal component projection matrix of the spectral field with the row vectors of the orthogonal residual matrix of the physical field to generate the second cross feature. The first cross feature... With the second cross feature The calculation formula is:

[0076]

[0077] in, Represents the Hadama product. This indicates taking the first orthogonal residual matrix of the spectral field. List, This indicates taking the first orthogonal residual matrix of the physical field. List.

[0078] The early warning server applies polarization regularization constraints to the first and second cross features, maximizing the projection magnitude of the cross feature vectors in the orthogonal subspace, thus generating orthogonalized cross feature vectors. The loss function for the polarization regularization constraint is:

[0079] By minimizing the polarization regularization loss function, the discriminative power of cross features can be enhanced, thereby improving the accuracy of subsequent risk assessment.

[0080] In a preferred embodiment, reference Figure 6 The early warning server compares the pre-emergence risk profile of toxin outbreaks with a dynamic baseline threshold, including the following processes: The early warning server extracts historical pre-emergence risk profiles of toxin outbreaks based on a sliding time window, constructing a historical risk characteristic distribution sequence. The length of the sliding time window is determined based on the accumulation of historical data, for example, set to 30 days. The early warning server performs kernel density estimation on the historical risk characteristic distribution sequence to obtain a nonparametric probability density function. Kernel density estimation is a nonparametric statistical method that can estimate the probability density distribution of the population based on sample data without assuming that the data follows a specific distribution form.

[0081] The formula for calculating kernel density estimation is:

[0082] in, For the sample size, For bandwidth, For kernel function, For the first Each sample value.

[0083] The early warning server selects the eigenvalue corresponding to the cumulative probability reaching a preset quantile on the nonparametric probability density function as the initial baseline reference point. The preset quantile is determined according to the sensitivity requirements of the risk warning, for example, set to the 95th quantile. The early warning server calculates the environmental fluctuation correction coefficient based on the velocity scalar of the current ocean current field vector data and the pressure gradient change rate of the meteorological disturbance data. The environmental fluctuation correction coefficient reflects the degree of fluctuation in the current marine environment; the greater the environmental fluctuation, the larger the correction coefficient.

[0084] The formula for calculating the environmental fluctuation correction factor is:

[0085] in, Let the current flow velocity be a scalar. The historical average flow velocity scalar. This represents the absolute value of the current rate of change of the pressure gradient. This represents the absolute value of the historical average rate of change of the pressure gradient. and These are the weighting coefficients.

[0086] The early warning server multiplies the initial baseline reference point by an environmental fluctuation correction factor to obtain the dynamic baseline threshold. The formula for calculating the dynamic baseline threshold is:

[0087] in, This serves as the initial baseline reference point.

[0088] When the pre-outbreak risk indicator value exceeds the dynamic baseline threshold, the early warning server generates an alarm command corresponding to the risk level. The risk level is categorized based on the degree to which the pre-outbreak risk indicator value exceeds the dynamic baseline threshold, for example, into three levels: low risk, medium risk, and high risk. Different risk levels correspond to different alarm methods and responses. The early warning server sends the alarm command to the display terminal in the aquaculture area, which then displays the corresponding risk level and response based on the alarm command, guiding aquaculture personnel to take appropriate prevention and control measures.

[0089] In a preferred embodiment, after generating an alarm command, the early warning server stores the alarm information and corresponding multimodal data in a database for subsequent model updates and performance evaluation. The early warning server periodically fine-tunes the multimodal signal spatiotemporal alignment model and feature extraction network using newly accumulated historical data to adapt to long-term changes in the marine environment. Model fine-tuning employs transfer learning, training the last few layers of the model with new data while maintaining the model's low-level feature extraction capabilities and updating its high-level classification and regression capabilities.

[0090] In a preferred embodiment, the environmental parameter acquisition node also collects water temperature and salinity data, and the underwater fluorescence spectroscopy acquisition node also collects water turbidity data. The early warning server incorporates water temperature, salinity, and turbidity data into the processing flow of the multimodal signal spatiotemporal alignment model and feature extraction network, further improving the accuracy of toxin outbreak risk early warning. Water temperature and salinity are important environmental factors affecting algal growth, and turbidity affects the acquisition quality of fluorescence spectral signals. Incorporating these data into the model can more comprehensively reflect the environmental conditions of the aquaculture water and the physiological state of the algae.

[0091] In a preferred embodiment, the early warning server is also linked with the automatic control system of the aquaculture area. When a high-risk alarm command is generated, the early warning server sends a control command to the automatic control system, which then initiates corresponding prevention and control measures based on the command, such as closing the inlet gate of the aquaculture area, activating water purification equipment, and increasing dissolved oxygen levels in the water. Through this linkage with the automatic control system, automated prevention and control of shellfish toxin risks can be achieved, reducing delays in human intervention and minimizing losses caused by toxin outbreaks.

Claims

1. A high-risk early warning system for shellfish toxins in marine aquaculture, characterized in that, This includes environmental parameter acquisition nodes, underwater fluorescence spectroscopy acquisition nodes, and an early warning server; The environmental parameter acquisition node acquires ocean current field vector data and meteorological disturbance data, and the underwater fluorescence spectroscopy acquisition node acquires algal fluorescence spectral signals excited by specific wavelengths underwater, and transmits the ocean current field vector data, the meteorological disturbance data and the algal fluorescence spectral signals to the early warning server; The early warning server constructs a multimodal signal spatiotemporal alignment model, and uses the multimodal signal spatiotemporal alignment model to synchronize the physical field disturbance period with the spectral feature change period with timestamps and spatial coordinate mapping to generate spatiotemporal aligned data; The early warning server uses a feature extraction network based on long short-term memory network and temporal convolution coupling to process the spatiotemporal aligned data, extract the hysteretic driving features of physical field perturbation on algal proliferation and the nonlinear drift features of the characteristic peaks of toxin-producing algae in the fluorescence spectrum, and perform feature cross-fusion of the hysteretic driving features and the nonlinear drift features in the latent space to generate a pre-characteristic value of toxin outbreak risk. The early warning server compares the pre-risk characterization value of the toxin outbreak with the dynamic baseline threshold. When the pre-risk characterization value of the toxin outbreak exceeds the dynamic baseline threshold, it generates an alarm command corresponding to the risk level and sends it to the display terminal in the breeding area.

2. The high-risk early warning system for shellfish toxins in marine aquaculture according to claim 1, characterized in that, The early warning server uses the multimodal signal spatiotemporal alignment model to synchronize the physical field disturbance period with the spectral feature change period using timestamps and spatial coordinate mapping to generate spatiotemporal aligned data, including: the early warning server extracts the flow field topology constraint boundary from the ocean current field vector data, and calculates the fluid transport lag time function between the environmental parameter acquisition node and the underwater fluorescence spectroscopy acquisition node based on the flow field topology constraint boundary; The early warning server performs reverse compensation and stretching operations on the meteorological disturbance data and the algal fluorescence spectral signal in the time dimension based on the fluid transport lag time function, aligning the meteorological data response sequence and the spectral data response sequence caused by the same physical disturbance event to the same time axis reference point. The early warning server constructs a spatial coordinate system mapping matrix based on the Lagrange particle trajectories in the ocean current field vector data, and projects the time-aligned meteorological disturbance data and the algal fluorescence spectral signal onto a unified spatial grid to generate the spatiotemporally aligned data.

3. The high-risk early warning system for shellfish toxins in marine aquaculture according to claim 1, characterized in that, The early warning server employs a feature extraction network based on long short-term memory network coupled with temporal convolution to extract the lag driving features of physical field disturbances on algal proliferation. The early warning server constructs a spatiotemporal graph structure from the ocean current field vector data and the meteorological disturbance data in the spatiotemporal aligned data, where nodes are spatial grid positions and edges are fluid connectivity relationships. The early warning server applies a spatiotemporal graph convolution operation to the spatiotemporal graph structure to extract the initial propagation features of meteorological disturbances under spatial topological correlation. The early warning server inputs the initial propagation characteristics of the meteorological disturbance into a long short-term memory network configured with an adaptive delay matrix. The adaptive delay matrix dynamically adjusts the delay receptive field scale according to the Reynolds number of the flow field, calculates the state transition correlation matrix with time lag effect, and outputs the lag driving characteristics of the physical field disturbance on algal proliferation.

4. The high-risk early warning system for shellfish toxins in marine aquaculture according to claim 1, characterized in that, The early warning server employs a feature extraction network based on long short-term memory network coupled with temporal convolution to extract the nonlinear drift features of the characteristic peaks of toxic algae in the fluorescence spectrum. This includes: the early warning server performing multi-scale wavelet packet decomposition on the algal fluorescence spectrum signal in the spatiotemporally aligned data to obtain spectral sub-band signals at different frequency band levels. The early warning server calculates the wavelet packet energy entropy of each of the spectral sub-band signals, and constructs a frequency band energy distribution matrix that evolves over time based on the wavelet packet energy entropy; The early warning server inputs the frequency band energy distribution matrix into an extended causal convolutional network, and uses the receptive field corresponding to different expansion rates to extract the nonlinear shift trajectory vectors of the feature peak center frequency at different time scales. The nonlinear shift trajectory vectors at each scale are combined to form the nonlinear drift characteristics of the toxic algae feature peaks in the fluorescence spectrum.

5. The high-risk early warning system for shellfish toxins in marine aquaculture according to claim 1, characterized in that, The early warning server performs feature cross-fusion of the hysteresis driving feature and the nonlinear drift feature in the latent space to generate a pre-characterization value of toxin outbreak risk, including: the early warning server performs orthogonal projection decomposition on the hysteresis driving feature and the nonlinear drift feature respectively to obtain the principal component projection matrix and orthogonal residual matrix of the corresponding physical field and the principal component projection matrix and orthogonal residual matrix of the corresponding spectral field. The early warning server performs a Hadamard product interaction between the principal component projection matrix of the physical field and the orthogonal residual matrix of the spectral field to generate a first cross feature, and performs a Hadamard product interaction between the principal component projection matrix of the spectral field and the orthogonal residual matrix of the physical field to generate a second cross feature. The early warning server concatenates and fully connects the first cross feature and the second cross feature to generate the pre-characteristic value of the toxin outbreak risk.

6. The high-risk early warning system for shellfish toxins in marine aquaculture according to claim 1, characterized in that, The comparison between the pre-emergence risk characterization value of the toxin outbreak and the dynamic baseline threshold by the early warning server includes: the early warning server extracting the historical pre-emergence risk characterization value of the toxin outbreak based on a sliding time window, and constructing a historical risk feature distribution sequence; The early warning server performs kernel density estimation on the historical risk characteristic distribution sequence to obtain a nonparametric probability density function; The early warning server selects the feature value whose cumulative probability reaches a preset quantile on the nonparametric probability density function as the initial baseline reference point. The early warning server calculates the environmental fluctuation correction coefficient based on the current ocean current field vector data velocity scalar and the air pressure gradient change rate of the meteorological disturbance data, and multiplies the initial baseline reference point with the environmental fluctuation correction coefficient to obtain the dynamic baseline threshold.

7. The high-risk early warning system for shellfish toxins in marine aquaculture according to claim 2, characterized in that, The early warning server extracts the flow field topological constraint boundary from the ocean current field vector data, and calculates the fluid transport lag time function between the environmental parameter acquisition node and the underwater fluorescence spectroscopy acquisition node based on the flow field topological constraint boundary, including: the early warning server calculates a finite-time Lyapunov exponential field based on the ocean current field vector data, and extracts ridges with exponential values ​​greater than a preset threshold in the finite-time Lyapunov exponential field as Lagrange coherent structures; The early warning server divides the Lagrange coherent structure into topological boundaries that separate water transport, and calculates the shortest fluid trajectory path of fluid particles from the environmental parameter acquisition node to the underwater fluorescence spectrum acquisition node along the topological boundaries. The early warning server integrates the flow velocity vector along the shortest fluid trajectory path to generate the fluid transport lag time function, which varies with spatial position and time.

8. The high-risk early warning system for shellfish toxins in marine aquaculture according to claim 3, characterized in that, The adaptive delay matrix dynamically adjusts the delay sensing field scale based on the flow field Reynolds number, including: the early warning server calculates the local flow field Reynolds number based on the ocean current field vector data, and maps the local flow field Reynolds number to a delay control factor; The early warning server constructs a delay memory unit containing multiple parallel delay branches, each of which is configured with a different fixed delay step size; The early warning server uses the delay control factor to weight and fuse the output states of each of the parallel delay branches. When the local flow field Reynolds number increases, the weight of the parallel delay branch with a small delay step size is increased to shrink the delay receptive field scale to extract recent sudden disturbance correlation features. When the local flow field Reynolds number decreases, the weight of the parallel delay branch with a large delay step size is increased to expand the delay receptive field scale to extract long-term lag correlation features.

9. The high-risk early warning system for shellfish toxins in marine aquaculture according to claim 4, characterized in that, The early warning server calculates the wavelet packet energy entropy of each of the spectral sub-band signals and constructs a frequency band energy distribution matrix that evolves over time based on the wavelet packet energy entropy, including: for each of the spectral sub-band signals, the early warning server calculates the proportion of the signal energy to the total energy of the entire frequency band as a relative energy entropy value; The early warning server moves an analysis window of a fixed length along the time axis, extracts the relative energy entropy values ​​of all the spectral sub-band signals within each analysis window, and constructs a column vector characterizing the frequency band energy distribution. The early warning server concatenates the column vectors obtained in continuous time steps in chronological order to form the frequency band energy distribution matrix, in which the row direction represents time evolution and the column direction represents the frequency band level. The early warning server performs a difference operation on the frequency band energy distribution matrix to extract the transient migration vector of the characteristic peak energy between different frequency bands.

10. The high-risk early warning system for shellfish toxins in marine aquaculture according to claim 5, characterized in that, The early warning server performs Hadamard product interaction between the principal component projection matrix of the physical field and the orthogonal residual matrix of the spectral field to generate a first cross feature, and performs Hadamard product interaction between the principal component projection matrix of the spectral field and the orthogonal residual matrix of the physical field to generate a second cross feature. This includes: the early warning server performs singular value decomposition on the hysteresis driving feature and the nonlinear drift feature respectively, extracts the left singular vectors corresponding to the preceding preset number of singular values ​​to construct the principal component projection matrix, and constructs the orthogonal residual matrix based on the left singular vectors corresponding to the remaining singular values. The early warning server performs element-wise multiplication of the row vectors of the principal component projection matrix of the physical field with the row vectors of the orthogonal residual matrix of the spectral field to generate the first cross feature, and performs element-wise multiplication of the row vectors of the principal component projection matrix of the spectral field with the row vectors of the orthogonal residual matrix of the physical field to generate the second cross feature. The early warning server applies polarization regularization constraints to the first cross feature and the second cross feature, thereby maximizing the projection modulus of the cross feature vector in the orthogonal subspace and generating an orthogonalized cross feature vector.

Images