A marine ecological threshold intelligent early warning method and system for a wind-solar-fishing complex

Abstract

This invention discloses an intelligent early warning method and system for ecological thresholds in the combined use of wind, solar, and fishing environments. It relates to the field of marine ecological environment monitoring and resource management data processing technology. Addressing the technical problems of existing static thresholds that ignore the nonlinear coupling and synergistic effects of multiple pressure sources and are difficult to predict and quantify instability gradients in advance, this method reconstructs the phase space of multidimensional environmental monitoring data to build an ecological phase space manifold benchmark; tracks the evolution trajectory of micro-disturbances, extracts autoregressive coefficients and state variances to generate a critical slowing characteristic index sequence to capture instability precursors; utilizes gradient backpropagation to calculate a dynamic composite ecological threshold surface when an early warning is triggered; performs sensitivity analysis along the manifold gradient direction to quantify the contribution of each pressure source; and automatically generates targeted intervention instructions for core environmental factors. This invention provides accurate early warning of marine ecological instability under multiple disturbances, improving the intelligent comprehensive management and control level of combined wind, solar, and fishing environments.

Description

Technical Field

[0001] This invention relates to the field of marine ecological environment monitoring and resource management data processing technology, and in particular to an intelligent early warning method and system for ecological thresholds in marine environments where wind, solar and fishing coexist. Background Technology

[0002] The integrated use of wind, solar, and fisheries in the sea refers to a three-dimensional marine resource development model that comprehensively deploys offshore wind power, offshore photovoltaic, and marine ranching facilities within the same sea area. This model involves the cross-operation of energy production and aquaculture, and the operation of various facilities and human activities will continuously have a cumulative impact on environmental factors such as sunlight transmittance, ocean current velocity, and water quality parameters in the surrounding sea area. In order to ensure the balance of the marine ecosystem and the safety of fisheries production, it is necessary to establish a supporting ecological threshold early warning mechanism. This mechanism should assess the current ecological carrying capacity of the sea area and formulate corresponding management and scheduling strategies through the processing and analysis of multi-dimensional environmental monitoring data.

[0003] Chinese patent publication number CN122288908A discloses a method and system for intelligent monitoring of coastal ecology based on multi-source data fusion. The method includes: Step S1, collecting multi-source monitoring data through a group of front-end sensing devices in a coordinated manner; Step S2, transmitting the multi-source monitoring data to a data resource platform, which cleans, stores, and manages the received multi-source monitoring data to obtain standard monitoring data; Step S3, performing multi-level data fusion processing on the standard monitoring data, including at least feature-level fusion and decision-level fusion, to obtain fused data; Step S4, inputting the fused data into a pre-trained AI model group to obtain risk identification results; Step S5, generating early warning information based on the risk identification results, thereby improving the real-time and accurate perception and early warning capabilities for complex coastal ecology and dynamic risks.

[0004] Existing monitoring and early warning systems for the marine ecological environment typically rely on networks of various water quality and meteorological sensors deployed in the sea. These sensors periodically collect environmental parameters such as dissolved oxygen, chlorophyll, and inorganic nitrogen, comparing them to pre-set, fixed safety standards. When a single environmental parameter exceeds the set limit, the system triggers an alarm and notifies management personnel to take manual intervention measures. Some management systems also incorporate time series analysis models to linearly extrapolate the future trends of individual environmental factors, using this as a reference for formulating aquaculture feeding or equipment operation and maintenance plans.

[0005] Existing technical solutions, when processing multi-source heterogeneous data in scenarios involving combined wind, solar, and aquaculture use of the ocean, treat each environmental factor as an independent variable for isolated evaluation, neglecting the nonlinear coupling effects of multiple stressors such as wind power disturbance, solar shading, and aquaculture on the marine ecosystem. Fixed, static thresholds are ill-suited to the dynamically evolving marine environment and fail to reflect the true state of the system approaching instability under the synergistic effects of multiple factors. Mechanisms relying on single-factor exceedances to trigger alarms lag behind the substantial decline in the ecosystem's resilience and lack quantitative analysis of the instability evolution gradient. This results in intervention strategies that are difficult to match with core stressors, limiting the level of intelligence in integrated marine use management. Summary of the Invention

[0006] To address the aforementioned issues, this invention provides an intelligent early warning method and system for marine ecological thresholds in the context of combined wind, solar, and fishery use. It employs phase space reconstruction and critical slowing feature extraction techniques to track the evolution trajectory of the ecological state, combines gradient backpropagation to solve the dynamic composite ecological threshold surface, and quantifies the contribution of pressure sources based on sensitivity analysis to automatically generate targeted intervention commands. This enables early warning and intelligent management and scheduling of precursors to marine ecological instability under multiple disturbances.

[0007] The above objectives can be achieved through the following approach: A method for intelligent early warning of ecological thresholds in marine areas used for both wind, solar, and fishing activities includes the following steps: acquiring multidimensional environmental time-series monitoring data of the marine area used for both wind, solar, and fishing activities, and performing time alignment and missing value imputation processing to generate a standardized multidimensional environmental state sequence; performing high-dimensional mapping transformation on the standardized multidimensional environmental state sequence based on a phase space reconstruction algorithm to construct an ecological phase space manifold benchmark; continuously collecting micro-disturbance events and tracking the evolution trajectory of the current ecological state point in the ecological phase space manifold benchmark in real time, calculating the time scale required for the current ecological state point to regress from the disturbance offset position to the equilibrium domain, extracting the autoregressive coefficient and state variance, and generating a critical slowing characteristic index sequence; performing time window sliding statistical analysis on the critical slowing characteristic index sequence, and determining when the autoregressive coefficient continuously rises and approaches one and the state variance significantly increases. Generate an ecological instability precursor signal; respond to the ecological instability precursor signal to initiate manifold space dimensionality reduction mapping calculation, perform nonlinear projection of the critical slowing characteristic index sequence in the low-dimensional feature space, identify the proximity of the current system state to historical instability samples, and trigger a dynamic threshold inversion calculation process when the distance metric is less than a preset safety threshold; use the gradient backpropagation algorithm to map the instability critical boundary in the low-dimensional feature space back to the original high-dimensional physical parameter space, and solve for the dynamic composite ecological threshold surface; perform sensitivity analysis along the gradient direction of the ecological phase space manifold benchmark to quantify the contribution of each pressure source to system instability, and generate a factor weight ranking list; automatically generate targeted intervention instructions based on the core environmental factor with the highest ranking in the factor weight ranking list, and send the intervention instructions to the corresponding execution control module.

[0008] Optionally, the step of acquiring multi-dimensional environmental time-series monitoring data of the sea area used for wind, solar and fishery co-operation and performing time alignment processing and missing value interpolation processing to generate a standardized multi-dimensional environmental state sequence includes: collecting raw environmental monitoring data streams reported by a distributed sensor network, the raw environmental monitoring data streams including time-series data of light transmittance, ocean current velocity, dissolved oxygen concentration, chlorophyll concentration, and dissolved inorganic nitrogen concentration; uniformly calibrating the timestamps of each sensor in the raw environmental monitoring data streams, interpolating and aligning data sequences with different sampling frequencies to a unified time grid to generate a time-synchronized environmental data matrix; detecting outliers and missing values ​​in the time-synchronized environmental data matrix, using the isolated forest algorithm to remove outliers, and using the cubic spline interpolation algorithm to complete the missing time periods to generate a cleaned environmental data matrix; and performing dimensional normalization processing on each environmental factor in the cleaned environmental data matrix to generate a standardized multi-dimensional environmental state sequence.

[0009] Optionally, the step of performing a high-dimensional mapping transformation on the standardized multidimensional environmental state sequence based on the phase space reconstruction algorithm to construct an ecological phase space manifold benchmark includes: calculating the autocorrelation function of the time series data of each environmental factor in the standardized multidimensional environmental state sequence to determine the time delay parameter; performing a phase space embedding transformation on the standardized multidimensional environmental state sequence according to the time delay parameter, expanding the one-dimensional time series of each environmental factor into a set of trajectory points in a high-dimensional phase space to generate a phase space trajectory coordinate matrix; performing manifold fitting calculation on the phase space trajectory coordinate matrix, and constructing the topological relationship between neighboring points through a local linear embedding algorithm to form an ecological phase space manifold benchmark.

[0010] Optionally, the continuous acquisition of micro-disturbance events and real-time tracking of the evolution trajectory of the current ecological state point in the ecological phase space manifold benchmark, calculating the time scale required for the current ecological state point to return from the disturbance offset position to the equilibrium domain, extracting autoregressive coefficients and state variance, and generating a critical slowing characteristic index sequence, includes: monitoring marine tidal change events, meteorological fluctuation events, and aquaculture operation events; when the abrupt change amplitude of each environmental factor exceeds the preset background fluctuation standard deviation, marking it as a micro-disturbance event and recording the time of occurrence of the disturbance; extracting the initial coordinates of the ecological state point corresponding to the time of occurrence of the micro-disturbance event in the ecological phase space manifold benchmark, continuously tracking the spatial position evolution of the state point in subsequent time windows, calculating the decay curve of the Euclidean distance between the state point and the center of the equilibrium domain over time; fitting the decay curve with a first-order autoregressive model, extracting the autoregressive coefficient as the first feature parameter, and simultaneously calculating the variance of the decay curve as the second feature parameter; organizing the first feature parameter and the second feature parameter in chronological order to generate a critical slowing characteristic index sequence.

[0011] Optionally, the step of performing time-window sliding statistical analysis on the critical slowing characteristic index sequence, generating an ecological instability precursor signal when the autoregressive coefficient continuously rises and approaches one and the state variance significantly increases, includes: performing sliding window traversal on the critical slowing characteristic index sequence according to a preset sliding time window length, calculating the trend slope of the autoregressive coefficient and the coefficient of variation of the state variance for each window position, and generating a dynamic statistical characteristic sequence; establishing an autoregressive coefficient trend criterion, marking a resilience decay state when the trend slope of the autoregressive coefficient in the dynamic statistical characteristic sequence is greater than zero and the mean of the autoregressive coefficient enters a preset threshold range; establishing a state variance fluctuation criterion, marking a system instability enhancement state when the coefficient of variation of the state variance in the dynamic statistical characteristic sequence exceeds a preset multiple threshold of the historical baseline level; and generating an ecological instability precursor signal when the resilience decay state and the system instability enhancement state simultaneously satisfy a duration exceeding a preset minimum warning period.

[0012] Optionally, in response to the ecological instability precursor signal, a manifold space dimensionality reduction mapping calculation is initiated, which nonlinearly projects the critical slowing characteristic index sequence into a low-dimensional feature space to identify the proximity of the current system state to historical instability samples. When the distance metric is less than a preset safety threshold, a dynamic threshold inversion calculation process is triggered, including: loading a historical ecological instability event sample library, which contains labeled critical slowing characteristic index sequence samples before instability; performing low-dimensional manifold learning on the samples in the historical ecological instability event sample library to construct a historical instability feature analysis. The system generates a low-dimensional feature vector of the current system state by inputting the currently collected critical slowdown feature index sequence into a pre-trained nonlinear dimensionality reduction mapping model. It then calculates the Mahalanobis distance between the low-dimensional feature vector and each historical instability sample point in the historical instability feature distribution map, identifies the nearest neighbor historical instability sample, and uses the minimum Mahalanobis distance as the instability proximity metric of the current system state. Finally, it compares the instability proximity metric with a preset safety threshold. If the instability proximity metric is less than the preset safety threshold, it triggers a dynamic threshold inversion calculation process; otherwise, it maintains the current monitoring state.

[0013] Optionally, the step of using the gradient backpropagation algorithm to map the instability critical boundary in the low-dimensional feature space back to the original high-dimensional physical parameter space and solve for the dynamic composite ecological threshold surface includes: extracting the instability critical boundary that separates the safe area from the instability danger area in the low-dimensional feature space; parameterizing the instability critical boundary to generate a low-dimensional critical boundary function; constructing an iterative inversion optimizer from the low-dimensional feature space to the original high-dimensional physical parameter space; inputting the low-dimensional critical boundary function into the iterative inversion optimizer through backpropagation of gradient information to generate a critical boundary point cloud in the high-dimensional physical parameter space; and performing surface fitting on the critical boundary point cloud to establish a multivariate functional relationship jointly defined by the environmental factors to generate the dynamic composite ecological threshold surface.

[0014] Optionally, the sensitivity analysis along the gradient direction of the ecological phase space manifold baseline, quantifying the contribution of each pressure source to system instability, and generating a factor weight ranking list includes: calculating the normal and tangent vectors of the manifold surface at the current ecological state point of the ecological phase space manifold baseline to determine the gradient direction of the system state evolution towards instability; performing differential perturbation on each environmental factor along the gradient direction, applying a unit perturbation amount to each environmental factor, measuring the displacement distance of the system state point in the manifold space caused by each perturbation, and generating a factor perturbation response vector; calculating the cosine value of the angle between each component of the factor perturbation response vector and the system instability gradient direction, normalizing the cosine value of each component to obtain the contribution weight of the corresponding environmental factor, and arranging them in descending order of weight value to generate a factor weight ranking list.

[0015] Optionally, the step of automatically generating targeted intervention instructions based on the highest-ranked core environmental factor in the factor weight ranking list and sending the intervention instructions to the corresponding execution control module includes: selecting the core environmental factor with the highest weight value from the factor weight ranking list; querying the stress source type corresponding to the core environmental factor, where the stress source type includes wind power disturbance sources, photovoltaic shading sources, or aquaculture sources; and generating a rule base by calling the corresponding intervention strategy based on the stress source type. When the core environmental factor is dissolved oxygen concentration and the stress source type is aquaculture source, an instruction to reduce the feeding frequency in a specific cage area is generated. The following commands are generated: When the core environmental factor is light transmittance and the pressure source type is photovoltaic shading source, an instruction to adjust the tilt angle of the photovoltaic panel is generated; when the core environmental factor is ocean current velocity and the pressure source type is wind power disturbance source, an instruction to optimize wind turbine operating parameters is generated; when other environmental factors and pressure source types are combined, corresponding conventional control instructions are generated according to the rule base generated by the intervention strategy; the generated intervention instructions are encapsulated into control instruction data packets and sent to the corresponding execution control module through an industrial communication protocol. The execution control module includes an aquaculture automation control unit, a photovoltaic array attitude adjustment unit, or a wind turbine operation management unit.

[0016] This invention also provides an intelligent early warning system for ecological thresholds in marine environments where wind, solar and fishery co-use are conducted, comprising: The data acquisition module is used to acquire multi-dimensional environmental time-series monitoring data of the sea area where wind, solar and fishery are used together, and to perform time alignment processing and missing value interpolation processing to generate a standardized multi-dimensional environmental state sequence. The phase space reconstruction module is used to perform high-dimensional mapping transformation on the standardized multidimensional environmental state sequence based on the phase space reconstruction algorithm to construct an ecological phase space manifold benchmark. The critical slowing identification module is used to continuously collect micro-disturbance events and track the evolution trajectory of the current ecological state point in the ecological phase space manifold benchmark in real time. It calculates the time scale required for the current ecological state point to return from the disturbance offset position to the equilibrium domain, extracts the autoregressive coefficient and state variance, and generates a critical slowing characteristic index sequence. The precursor signal generation module is used to perform time window sliding statistical analysis on the critical slowing characteristic index sequence. When the autoregressive coefficient continues to rise and approaches one and the state variance increases significantly, an ecological instability precursor signal is generated. The dimension reduction mapping module is used to initiate manifold space dimension reduction mapping calculation in response to the ecological instability precursor signal, and to perform nonlinear projection of the critical slowing characteristic index sequence in the low-dimensional feature space to identify the proximity between the current system state and historical instability samples. When the distance metric is less than a preset safety threshold, the dynamic threshold inversion calculation process is triggered. The threshold inversion module is used to map the unstable critical boundary in the low-dimensional feature space back to the original high-dimensional physical parameter space using the gradient backpropagation algorithm, and solve for the dynamic composite ecological threshold surface. The sensitivity analysis module is used to perform sensitivity analysis along the gradient direction of the ecological phase space manifold baseline, quantify the contribution of each pressure source to system instability, and generate a factor weight ranking list. The intervention instruction generation module is used to automatically generate targeted intervention instructions based on the core environmental factor with the highest ranking in the factor weight ranking list, and send the intervention instructions to the corresponding execution control module.

[0017] Compared with the prior art, the present invention has the following advantages: This invention achieves the tracking of the evolutionary trajectory of marine ecosystem disturbances and the capture of instability precursor signals by constructing an ecological phase space manifold benchmark and extracting critical slowing characteristic index sequences. Furthermore, it utilizes the gradient backpropagation algorithm to map the low-dimensional instability critical boundary back to the high-dimensional physical parameter space, and calculates the dynamic composite ecological threshold surface, thus overcoming the limitation that static thresholds cannot adapt to the synergistic effects of multiple environmental factors. At the same time, it performs sensitivity analysis along the gradient direction of the ecological phase space manifold benchmark to quantify the contribution of various pressure sources to system instability, thereby automatically generating targeted intervention instructions and distributing them to the corresponding execution control modules, improving the level of automation in ecological management in the scenario of wind, solar and fishery co-use of the sea. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram of the ecological phase space manifold reconstruction and critical slowing recovery mechanism of micro-disturbance trajectory according to the present invention; Figure 3 This is a schematic diagram of the evolution curves of the critical slowing autoregressive coefficient and the state variance sliding window dual indexes of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Reference Figure 1 As shown, a smart early warning method for marine ecological thresholds in the context of simultaneous wind, solar, and fishery use is provided. The method includes: S1. Acquire multi-dimensional environmental time-series monitoring data of the sea area where wind, solar and fishery are used together, and perform time alignment and missing value imputation to generate a standardized multi-dimensional environmental state sequence. In one embodiment of the present invention, step S1 includes the following steps: collecting raw environmental monitoring data streams reported by a distributed sensor network, the raw environmental monitoring data streams including time-series data of light transmittance, ocean current velocity, dissolved oxygen concentration, chlorophyll concentration, and dissolved inorganic nitrogen concentration; uniformly calibrating the timestamps of each sensor in the raw environmental monitoring data streams, interpolating and aligning data sequences with different sampling frequencies to a unified time grid to generate a time-synchronized environmental data matrix; detecting outliers and missing values ​​in the time-synchronized environmental data matrix, using an isolated forest algorithm to remove outliers, and using a cubic spline interpolation algorithm to complete missing time periods to generate a cleaned environmental data matrix; and performing dimensional normalization processing on each environmental factor in the cleaned environmental data matrix to generate a standardized multidimensional environmental state sequence.

[0022] Specifically, the raw environmental monitoring data stream reported by the distributed sensor network is collected. This raw environmental monitoring data stream includes time-series data on light transmittance, ocean current velocity, dissolved oxygen concentration, chlorophyll concentration, and dissolved inorganic nitrogen concentration. The raw environmental monitoring data stream represents a set of unprocessed environmental parameters transmitted in real time by various monitoring devices deployed in the sea area. The light transmittance time-series data reflects the change in the ability of light to penetrate water over time; the ocean current velocity time-series data records the evolution of seawater flow speed over time; the dissolved oxygen concentration time-series data characterizes the fluctuation of dissolved oxygen content in the water over time; the chlorophyll concentration time-series data indicates the change in phytoplankton biomass in the water over time; and the dissolved inorganic nitrogen concentration time-series data reflects the dynamics of nutrient levels in the water over time.

[0023] The timestamps of each sensor in the raw environmental monitoring data stream are uniformly calibrated. A timestamp represents the precise moment when data acquisition occurred. Data sequences with different sampling frequencies are interpolated and aligned to a unified time grid. This unified time grid represents a reference time axis with fixed time intervals, generating a time-synchronized environmental data matrix. This matrix represents a two-dimensional data structure after all environmental parameters are aligned at the same time node. The time alignment process is achieved through a linear interpolation formula, as detailed below: , in, Represents the target time on a unified time grid Synchronized data values, and These represent the times immediately preceding the target in the original environmental monitoring data stream. The actual sampling time before and after and The data value.

[0024] Next, outliers and missing values ​​are detected in the time synchronization environment data matrix. Outliers refer to erroneous records that deviate from the normal data distribution range, while missing values ​​refer to blank data points not recorded due to equipment failure or communication interruption. Outliers are removed by constructing a random tree to isolate them. Outliers represent data nodes identified as outliers. The method calculates the outlier score for each data point using the following formula: , in, Representative data points Abnormal scores, Representative data points The average path length in multiple random trees. Represents the average path length constant used for normalization, when When the value exceeds a set threshold, the threshold can be set based on data to determine the data point. Outliers were identified and removed. A piecewise polynomial smoothing fitting method was used to fill in missing time periods. This method constructs the following cubic polynomial between adjacent data points: , in, Representing the The missing time period is at time Complete the data values. Representing the The start time of the segment , , , This represents the polynomial coefficients obtained by solving the boundary conditions.

[0025] Finally, a cleaned environmental data matrix is ​​generated. This matrix represents the complete dataset after outlier removal and missing time periods are filled in. Dimensional normalization is performed on each environmental factor in the cleaned matrix. Environmental factors represent the specific environmental parameter types that constitute the cleaned matrix. Dimensional normalization refers to converting data from different units to a dimensionless, unified interval. A standardized multidimensional environmental state sequence is then generated. This sequence represents multidimensional time series data after eliminating dimensional differences. The normalization formula is as follows: , in, Represents the normalized values ​​in a standardized multidimensional environmental state sequence. This represents the original values ​​in the cleaned environmental data matrix. and These represent the minimum and maximum values ​​of the corresponding environmental factors in historical records, respectively.

[0026] For example, raw environmental monitoring data streams reported by a distributed sensor network are collected. Light transmittance time-series data is available at time... The record is 80% at time [time]. Recorded as 70%, ocean current velocity time series data at time [time value missing]. The record was 1.5 meters per second, at time... The record is 1.7 meters per second. A target time is set using a unified time grid. t lie in and At the very center, the light transmittance synchronization data value in the time synchronization environment data matrix is ​​calculated according to the time alignment formula. Percentage, ocean current velocity synchronization data value Meters per second, detecting outliers and missing values ​​in the time-synchronized environmental data matrix. This is applied to specific data points within the time-series data of light transmittance. Calculate its average path length across multiple random trees. Set the normalized average path length constant to 5.0. The value is 10.0, calculated according to the outlier score formula. The anomaly score threshold was set to 0.6. Since 0.707 is greater than 0.6, this data point was determined to be an anomaly. Outliers are identified and removed.

[0027] For missing time periods in ocean current velocity time series data, a piecewise polynomial smoothing fitting method is used to fill in the gaps, given the polynomial coefficients. It is 1.6. It is 0.1. It is 0.05. The value is 0.01, and the calculation time is... t Distance from start time Time difference The time limit is 2 seconds to calculate the completed data value using a polynomial formula. Meters per second, generating a cleaned environmental data matrix, and performing dimensional normalization on each environmental factor in the cleaned environmental data matrix, setting the historical minimum value of the light transmittance time series data. 10%, the highest historical value. The percentage is 90%. For the light transmittance value X in the post-cleaning environmental data matrix, which is 75%, the normalized value in the standardized multidimensional environmental state sequence is calculated using the normalization formula. This completes the transformation from raw environmental monitoring data streams to standardized multidimensional environmental state sequences.

[0028] S2. Based on the phase space reconstruction algorithm, perform high-dimensional mapping transformation on the standardized multidimensional environmental state sequence to construct an ecological phase space manifold benchmark. In one embodiment of the present invention, step S2 includes the following steps: calculating the autocorrelation function of the time series data of each environmental factor in the standardized multidimensional environmental state sequence, and determining the time delay parameter; performing a phase space embedding transformation on the standardized multidimensional environmental state sequence according to the time delay parameter, expanding the one-dimensional time series of each environmental factor into a set of trajectory points in a high-dimensional phase space, and generating a phase space trajectory coordinate matrix; performing manifold fitting calculation on the phase space trajectory coordinate matrix, and constructing the topological relationship between neighboring points through a local linear embedding algorithm to form an ecological phase space manifold benchmark.

[0029] Specifically, the autocorrelation function of each environmental factor time series data in the standardized multidimensional environmental state sequence is calculated. The standardized multidimensional environmental state sequence represents multidimensional time series data after eliminating dimensional differences. Each environmental factor time series data refers to the set of values ​​of independent environmental parameters such as light transmittance as a function of time. The autocorrelation function is used to measure the degree of numerical similarity of the same time series at different time intervals. The calculation formula is as follows: , in, The time delay is represented as The autocorrelation function value at time, Represents the total length of time series data for each environmental factor. Represents the time step index. The standardized multidimensional environmental state sequence is in the first... The number of steps, Represents a standardized multidimensional environmental state sequence in lag The time step after the first time step The value of the step is used to determine the time delay parameter. The time delay parameter represents the number of delay steps required for the autocorrelation function to decrease to a specific proportion of its initial value. It is set based on empirical criteria in dynamical system theory that ensure the delay coordinates are independent. Typically, it is selected when the autocorrelation function value decreases to a specific proportion of its initial value. The value is used to perform a phase space embedding transformation on the standardized multidimensional environmental state sequence based on the time delay parameter. Phase space embedding transformation refers to the mathematical operation of reconstructing the original sequence into a multidimensional spatial vector using time delay coordinates. This expands the one-dimensional time series of each environmental factor into a set of trajectory points in a high-dimensional phase space. The one-dimensional time series represents the historical data of a single environmental factor, while the high-dimensional phase space represents a mathematical space containing all dynamic characteristics of the system. The trajectory point set represents the set of multidimensional coordinates showing the evolution of the system state over time. The expanded formula is as follows: , in, Representing the The phase space vector at each time step. Represents the embedding dimension. The setting is determined based on the pseudo nearest neighbor method, representing the minimum dimension required to reconstruct the phase space. The phase space trajectory coordinate matrix is ​​generated, which is a matrix structure composed of all phase space vectors arranged in chronological order. Its formula is expressed as: , in, Represents the phase space trajectory coordinate matrix. This represents the matrix transpose operation.

[0030] The phase space trajectory coordinate matrix is ​​then subjected to manifold fitting calculation. Manifold fitting calculation refers to the process of finding the inherent low-dimensional geometric structure of data in a high-dimensional space. Topological relationships between neighboring points are constructed using the Locally Linear Embedding (LAN) algorithm. LAN is a technique for dimensionality reduction or manifold learning that preserves local neighborhood geometric features. Neighboring points represent the data points closest to the target trajectory point in the high-dimensional phase space, and topological relationships represent the linear reconstruction weights between a data point and its neighbors. The calculation process is achieved by minimizing the reconstruction error cost function as follows: , in, This represents the overall reconstruction error. Represents the target trajectory point. The neighborhood points representing the target trajectory point This represents the reconstruction weight of the target trajectory point by the neighboring points.

[0031] The solution process must satisfy the constraint that the sum of the weights is a constant, forming an ecological phase space manifold benchmark. The ecological phase space manifold benchmark represents the benchmark geometric manifold that reflects the normal evolutionary law of the marine ecosystem, and is composed of the phase space trajectory coordinate matrix and topological relationships.

[0032] For example, consider time-series data of light transmittance in a standardized multidimensional environmental state sequence. Set the total length of the time-series data for each environmental factor. Given a time delay of 1000, calculate the autocorrelation function. When the value is 0, calculate the initial autocorrelation function value. It is 1.0. When the time delay... When increased to 5, calculate according to the formula. The value is 0.36. A specific percentage threshold is set at 0.368. Since 0.36 is the first time it falls below 0.368, the time delay parameter is determined. Set the embedding dimension to 5. The value is 3. A phase space embedding transformation is performed on the standardized multidimensional environmental state sequence based on the time delay parameter, for the time step index... Set the value to 1, and extract the numerical values ​​from the standardized multidimensional environmental state sequence. It is 0.8125. It is 0.7500. The value is 0.6800. The one-dimensional time series of each environmental factor is extended into a set of trajectory points in a high-dimensional phase space, generating a phase space vector. Similarly, the phase space vectors for subsequent time steps are generated, and these vectors are arranged in chronological order to generate the phase space trajectory coordinate matrix. .

[0033] Manifold fitting calculations are performed on the phase space trajectory coordinate matrix for the target trajectory points. Select its nearest neighbor point and Known The coordinates are , The coordinates are [0.7900, 0.7300, 0.6600]. A topological relationship between neighboring points is constructed using the local linear embedding algorithm, and the reconstruction weights are calculated. and After minimizing the reconstruction error cost function, we obtain... It is 0.6. The value is 0.4, and the sum of the verification weights is... The reconstructed vector is calculated as follows: Calculate the reconstruction error for This forms the basis for the ecological phase space manifold.

[0034] S3. Continuously collect micro-disturbance events and track the evolution trajectory of the current ecological state point in the ecological phase space manifold benchmark in real time. Calculate the time scale required for the current ecological state point to return from the disturbance offset position to the equilibrium domain. Extract the autoregressive coefficient and state variance to generate a critical slowing characteristic index sequence. In one embodiment of the present invention, step S3 includes the following steps: monitoring tidal change events, meteorological fluctuation events, and aquaculture operation events in the marine area; when the abrupt change amplitude of each environmental factor exceeds the preset background fluctuation standard deviation, it is marked as a micro-disturbance event and the time of occurrence of the disturbance is recorded; extracting the initial coordinates of the ecological state point corresponding to the time of occurrence of the micro-disturbance event in the ecological phase space manifold reference, continuously tracking the spatial position evolution of the state point in subsequent time windows, calculating the decay curve of the Euclidean distance between the state point and the center of the equilibrium domain over time; fitting the decay curve with a first-order autoregressive model, extracting the autoregressive coefficient as the first feature parameter, and simultaneously calculating the variance of the decay curve as the second feature parameter; organizing the first feature parameter and the second feature parameter in chronological order to generate a critical slowing characteristic index sequence.

[0035] Specifically, refer to Figure 2 As shown, the monitoring system tracks marine tidal changes, meteorological fluctuations, and aquaculture operation events. Tidal changes represent environmental parameter fluctuations caused by the periodic rise and fall of seawater. Meteorological fluctuations refer to water disturbances caused by weather changes such as storms or rainfall. Aquaculture operation events represent the short-term impact of human activities such as feeding or fishing on water quality. When the abrupt change of any environmental factor exceeds the preset background fluctuation standard deviation, it is marked as a micro-disturbance event, and the time of occurrence is recorded. The abrupt change amplitude represents the absolute value of the difference between the current normalized value of the environmental factor and its historical mean. The background fluctuation standard deviation represents the statistical dispersion of the environmental factor during historical stable periods. The criteria for this setting are calculated based on long-term historical undisturbed monitoring data, and the specific judgment formula is as follows: , in, Represents the magnitude of the mutation. Represents the normalized environmental factor value at the current moment. Representing the historical average, when Greater than the standard deviation of background fluctuation When a micro-perturbation event is triggered, the current moment is recorded as the moment the perturbation occurred. .

[0036] Extract the initial coordinates of the ecological state points corresponding to the occurrence of micro-perturbation events in the ecological phase space manifold reference. The ecological state points represent the high-dimensional phase space position of the system at a specific moment, and their initial coordinates are defined as follows: The spatial evolution of state points within subsequent time windows is continuously tracked. Each subsequent time window represents a fixed observation period after a disturbance occurs, and is set based on empirical statistical values ​​of the recovery time of marine ecosystem characteristics. Spatial evolution represents the continuous sequence of changes in state point coordinates over time steps. Calculate the decay curve of the Euclidean distance between the state point and the center of the equilibrium domain over time. The center of the equilibrium domain represents the geometric center of the phase space of the system in a undisturbed steady state, denoted as . The decay curve represents the distance sequence of the state points regressing to the center of the equilibrium domain. The distance is calculated using the following formula: , in, Representative moment Euclidean distance, This represents the embedding dimension determined by the preceding steps. The state point is at the 1st dimensional coordinate components, The center of the equilibrium domain is in the The coordinate components of the dimension.

[0037] A first-order autoregressive model is applied to fit the decay curve. First-order autoregressive model fitting refers to the process of linearly predicting the distance value at the current time step using the distance value from the previous time step. The fitting formula is: , in, Represents the autoregressive coefficient. Represents the fitting residual ， Represents the previous moment The Euclidean distance between the state point and the center of the equilibrium domain is used to extract the autoregressive coefficient as the first characteristic parameter. This first characteristic parameter reflects the rate at which the system recovers to equilibrium after being disturbed. Simultaneously, the variance of the decay curve is calculated as the second characteristic parameter. The variance of the decay curve represents the degree of fluctuation in the distance sequence within the time window. The variance calculation formula is as follows: , in, Represents the second characteristic parameter. This represents the total number of time steps contained in the subsequent time window. The first and second feature parameters are organized in chronological order to generate a critical slowing feature index sequence, which represents a dynamic feature set characterizing the trend of system stability degradation.

[0038] For example, for monitoring marine tidal change events, meteorological fluctuation events, and aquaculture operation events using time-series data on light transmittance, the historical average is known. The standard deviation of background fluctuation is 0.5000. It is 0.0500 at a specific time. The normalized value of light transmittance was detected. The value is 0.5800. The magnitude of the mutation is calculated using the formula. Since 0.0800 is greater than 0.0500, it is marked as a micro-perturbation event and the time of occurrence of the perturbation is recorded. For the 100th time step, extract the initial coordinates of the ecological state point corresponding to the occurrence of the micro-disturbance event in the ecological phase space manifold reference. Given the embedding dimension The initial coordinates are 3, according to the phase space embedding formula. The three components correspond to , and The measured values ​​obtained were 0.5800, 0.5600, and 0.5400, respectively. The center of the equilibrium domain is known. The three components are 0.5000, 0.5000, and 0.5000, respectively, continuously tracking the spatial evolution of the state point within subsequent time windows. The total number of time steps included in each subsequent time window is set. The value is 3. At the 100th time step, the Euclidean distance between the state point and the center of the equilibrium domain is calculated. At the 101st time step, the state point coordinates evolved to 0.5600, 0.5400, and 0.5200. Calculate the Euclidean distance. At the 102nd time step, the state point coordinates evolved to 0.5400, 0.5200, and 0.5100. The Euclidean distance was then calculated. The resulting Euclidean distance decay curves over time are 0.1077, 0.0748, and 0.0458.

[0039] A first-order autoregressive model was fitted to the decay curve, and the autoregressive coefficients were solved using the least squares method. The autoregressive coefficient is extracted as the first feature parameter, and its value is... Simultaneously, the variance of the decay curve is calculated as the second characteristic parameter, and the distance mean is calculated. Calculate the variance The first feature parameter 0.6678 and the second feature parameter 0.000639 are organized in chronological order to generate a critical slowdown feature index sequence.

[0040] S4. Perform time window sliding statistical analysis on the critical slowdown characteristic index sequence. When the autoregressive coefficient continues to rise and approaches unit 1 and the state variance increases significantly, generate an ecological instability precursor signal. In one embodiment of the present invention, step S4 includes the following steps: performing a sliding window traversal on the critical slowing characteristic index sequence according to a preset sliding time window length; calculating the trend slope of the autoregressive coefficient and the coefficient of variation of the state variance for each window position to generate a dynamic statistical feature sequence; establishing an autoregressive coefficient trend criterion: when the trend slope of the autoregressive coefficient in the dynamic statistical feature sequence is greater than zero and the mean of the autoregressive coefficient enters a preset threshold range, it is marked as a resilience decay state; establishing a state variance fluctuation criterion: when the coefficient of variation of the state variance in the dynamic statistical feature sequence exceeds a preset multiple threshold of the historical baseline level, it is marked as a system instability enhancement state; when the resilience decay state and the system instability enhancement state simultaneously satisfy a duration exceeding a preset minimum warning period, an ecological instability precursor signal is generated.

[0041] Specifically, refer to Figure 3 As shown, a sliding window traversal is performed on the critical slowing characteristic index sequence according to the preset sliding time window length. The critical slowing characteristic index sequence contains autoregressive coefficients and state variance arranged in chronological order. The preset sliding time window length represents the number of data points included in each statistical analysis, and is set based on the empirical time span for capturing the dynamic change trend of the system. For each window position, the trend slope of the autoregressive coefficient and the coefficient of variation of the state variance are calculated. The trend slope reflects the rate of change of the autoregressive coefficient over time and is obtained through linear regression. The coefficient of variation measures the relative fluctuation of the state variance. A dynamic statistical characteristic sequence is generated, which is composed of the trend slope and coefficient of variation calculated at each window position in chronological order. The formula for calculating the trend slope is as follows: , in, The trend slope representing the autoregressive coefficient. This represents the preset sliding time window length. Represents the first in the window Each time step This represents the average value of the time steps within the window. Represents the first in the window The autoregressive coefficients at each time step The mean of the autoregressive coefficients within the window is represented by the coefficient of variation, which is calculated using the following formula: , in, The coefficient of variation represents the variance of the state. The standard deviation represents the variance of the states within the window. This represents the mean of the variance of the states within the window.

[0042] Simultaneously, an autoregressive coefficient trend criterion is established. When the trend slope of the autoregressive coefficient in the dynamic statistical characteristic sequence is greater than zero and the mean of the autoregressive coefficient enters the preset threshold range, it is marked as a state of weakened resilience. The preset threshold range represents the typical range of values ​​for the autoregressive coefficient when the system is close to the critical point, and is usually set to an interval close to one unit. A state variance fluctuation criterion is also established. When the coefficient of variation of the state variance in the dynamic statistical characteristic sequence exceeds a preset multiple threshold of the historical baseline level, it is marked as a state of increased system instability. The historical baseline level represents the average value of the coefficient of variation of the state variance of the system during the stable period. The preset multiple threshold is used to determine whether the current fluctuation is abnormal. When the state of weakened resilience and the state of increased system instability simultaneously meet the condition that the duration exceeds the preset minimum warning period, an ecological instability precursor signal is generated. The preset minimum warning period represents the continuous time length required to confirm the change in system state, and is used to filter out short-term random fluctuations.

[0043] For example, a preset sliding time window length is set. The value is 5, representing the time steps contained within a specific window position of the critical slowing characteristic index sequence. The values ​​are 1, 2, 3, 4, and 5, respectively. These correspond to the autoregressive coefficients. The values ​​are 0.80, 0.82, 0.85, 0.88, and 0.90, respectively. Calculate the average value per time step. Calculate the mean of the autoregressive coefficients Calculate the trend slope according to the formula. Variance of the state within this window Calculate the mean of the state variance for values ​​of 0.010, 0.012, 0.015, 0.020, and 0.025. Standard deviation of state variance Calculate the coefficient of variation of the state variance according to the formula. Simultaneously, an autoregressive coefficient trend criterion is established. Given that the trend slope is 0.026 and greater than zero, a preset threshold range is set to [0.8, 1.0). When the mean autoregressive coefficient of 0.85 enters this range, it is marked as a state of weakened resilience. A state variance fluctuation criterion is established. Given that the historical baseline level is 0.10 and the preset multiple threshold is 3, when the previous coefficient of variation of 0.335 exceeds 0.30, it is marked as a state of increased system instability. A preset minimum warning period is set to 3 consecutive windows. If both of the above states are marked within 3 consecutive windows, an ecological instability precursor signal is generated.

[0044] S5. In response to the ecological instability precursor signal, start the manifold space dimensionality reduction mapping calculation, perform nonlinear projection of the critical slowing characteristic index sequence in the low-dimensional feature space, identify the proximity of the current system state and historical instability samples, and trigger the dynamic threshold inversion calculation process when the distance metric is less than the preset safety threshold. In one embodiment of the present invention, step S5 includes the following steps: loading a historical ecological instability event sample library, the historical ecological instability event sample library containing labeled critical slowing characteristic index sequence samples before instability; performing low-dimensional manifold learning on the samples in the historical ecological instability event sample library to construct a historical instability characteristic distribution map; inputting the currently collected critical slowing characteristic index sequence into a pre-trained nonlinear dimensionality reduction mapping model to generate a low-dimensional feature vector of the current system state; calculating the Mahalanobis distance between the low-dimensional feature vector and each historical instability sample point in the historical instability characteristic distribution map, identifying the nearest neighbor historical instability sample, and using the minimum Mahalanobis distance as the instability proximity metric of the current system state; comparing the instability proximity metric with a preset safety threshold; when the instability proximity metric is less than the preset safety threshold, triggering a dynamic threshold inversion calculation process; otherwise, maintaining the current monitoring state.

[0045] Specifically, a historical ecological instability event sample library is loaded. This sample library contains a set of characteristic data recorded before past ecosystem collapses. The sample library contains labeled pre-instability critical slowing characteristic indicator sequence samples. The labeled pre-instability critical slowing characteristic indicator sequence samples refer to historical autoregressive coefficient and state variance sequences with clear instability labels. Low-dimensional manifold learning is performed on the samples in the historical ecological instability event sample library. Low-dimensional manifold learning refers to the method of mapping high-dimensional data to low-dimensional space by preserving the local neighborhood structure of the data. A historical instability characteristic distribution map is constructed, which represents the set of coordinates of historical samples in low-dimensional space.

[0046] The currently collected critical slowing feature index sequence is then input into a pre-trained nonlinear dimensionality reduction mapping model. This model refers to a mathematical algorithm that optimizes parameters using historical data and can convert high-dimensional sequences into low-dimensional coordinates, generating a low-dimensional feature vector of the current system state. This low-dimensional feature vector represents the coordinates of the current data in a low-dimensional space. The formula for the mapping process is: ， in, A low-dimensional feature vector representing the current system state. The mapping function represents the pre-trained nonlinear dimensionality reduction mapping model. Representing the currently collected critical slowing feature index sequence, this nonlinear dimensionality reduction mapping model is specifically a differentiable neural network built based on a deep autoencoder. Its pre-training steps are as follows: an offline training set is constructed by collecting critical slowing feature index sequences from the entire historical steady-state and unstable periods of the sea area; the mean squared error (MSE) between the input sequence and the network's decoded and reconstructed sequence is used as the loss function; the error gradient is calculated using the gradient descent algorithm, and the connection weights and biases of the hidden layers of the network are iteratively optimized in reverse until the loss function on the validation set converges to a set threshold. The mapping function... This refers to the nonlinear matrix calculation performed by calling the feedforward network parameters of the encoder part in the autoencoder.

[0047] Calculate the Mahalanobis distance between the low-dimensional eigenvector and each historical instability sample point in the historical instability feature distribution map. Mahalanobis distance represents a distance metric that considers the covariance between different dimensions of the data. Historical instability sample points represent specific coordinate records in the historical instability feature distribution map. The Mahalanobis distance calculation formula is: , in, The low-dimensional feature vector representing the current system state and its number are... The Mahalanobis distance between historical unstable sample points. Representative number is Historical instability sample points Represents the matrix transpose operation. The inverse of the covariance matrix representing the distribution spectrum of historical instability features is used to identify the nearest neighbor historical instability samples. These samples represent the historical records closest to the current state. The minimum Mahalanobis distance is used as the instability proximity metric for the current system state. This metric indicates the degree of danger of the current state approaching a historical instability state. The extraction formula is as follows: , in, Represents the measure of proximity to instability. This represents the operation of taking the minimum value.

[0048] Finally, the instability proximity metric is compared with the preset safety threshold. The preset safety threshold represents the critical distance limit for determining whether the system is about to become unstable. It is set based on the statistical lower boundary of the Mahalanobis distance between historical normal state samples and unstable samples. When the instability proximity metric is less than the preset safety threshold, the dynamic threshold inversion calculation process is triggered. The dynamic threshold inversion calculation process refers to the subsequent steps of reverse derivation of the safety environment parameter boundary. Otherwise, the current monitoring state is maintained, which means that the routine data collection and analysis continue.

[0049] For example, a historical ecological instability event sample library is loaded, and a historical instability feature distribution map is constructed. It is known that the inverse matrix of the covariance matrix of the historical instability feature distribution map in two-dimensional space has elements of 2.0 for all diagonal elements and 0.5 for all off-diagonal elements. The currently collected critical slowdown feature index sequence is input into a pre-trained nonlinear dimensionality reduction mapping model to generate a low-dimensional feature vector of the current system state. The coordinates of the low-dimensional feature vector of the current system state are set to [1.0, 1.5]. The low-dimensional... The Mahalanobis distance between the eigenvector and each historical instability sample point in the historical instability feature distribution map is calculated. Taking the historical instability sample point numbered 1 in the historical instability feature distribution map, with coordinates [1.2, 1.3], the coordinate difference vector is calculated as [-0.2, 0.2]. According to the Mahalanobis distance calculation formula, the product of the transpose vector and the inverse matrix is ​​calculated as [-0.3, 0.3]. The inner product of this result and the difference vector is calculated as 0.12. The Mahalanobis distance corresponding to the historical instability sample point numbered 1 is then calculated. Similarly, the Mahalanobis distance between the historical unstable sample point numbered two and the sample point numbered two is calculated to be 0.512. The nearest neighbor historical unstable sample point is identified as the historical unstable sample point numbered one. The minimum Mahalanobis distance of 0.346 is used as the unstable proximity metric of the current system state.

[0050] The unstable proximity metric is then compared with a preset safety threshold, which is set to 0.500. Since the unstable proximity metric of 0.346 is less than the preset safety threshold of 0.500, the dynamic threshold inversion calculation process is triggered.

[0051] S6. Using the gradient backpropagation algorithm, the unstable critical boundary in the low-dimensional feature space is mapped back to the original high-dimensional physical parameter space to solve the dynamic composite ecological threshold surface. In one embodiment of the present invention, step S6 includes the following steps: extracting the instability critical boundary that separates the safe region from the instability danger region in the low-dimensional feature space; parameterizing the instability critical boundary to generate a low-dimensional critical boundary function; constructing an iterative inversion optimizer from the low-dimensional feature space to the original high-dimensional physical parameter space; inputting the low-dimensional critical boundary function into the iterative inversion optimizer by backpropagating gradient information to generate a critical boundary point cloud in the high-dimensional physical parameter space; performing surface fitting on the critical boundary point cloud to establish a multivariate function relationship jointly defined by the environmental factors, and generating a dynamic composite ecological threshold surface.

[0052] Specifically, the instability critical boundary that separates the safe region from the instability danger region is extracted in the low-dimensional feature space. The low-dimensional feature space represents the data representation space formed after the dimensionality reduction process in the previous steps. The safe region represents the data distribution range of the system in a stable state, and the instability danger region represents the data distribution range of the system tending to collapse. The instability critical boundary refers to the geometric boundary line that divides these two states. The instability critical boundary is parameterized.

[0053] Parametric representation refers to the quantitative description of geometric boundaries using mathematical equations, generating low-dimensional critical boundary functions. These low-dimensional critical boundary functions represent the mathematical expressions describing the instability critical boundaries, as follows: , in, This represents the value of the low-dimensional critical boundary function. Represents a coordinate vector in a low-dimensional feature space. The normal vector representing the boundary. The intercept constant representing the boundary. This represents the matrix transpose operation.

[0054] An iterative inversion optimizer is constructed to transform a low-dimensional feature space into a primitive high-dimensional physical parameter space. The primitive high-dimensional physical parameter space represents a multi-dimensional data space containing the original dimensions of all environmental factors. The iterative inversion optimizer refers to a computational model capable of reverting low-dimensional features to high-dimensional physical parameters. The iterative inversion optimizer backpropagates gradient information, which involves using the chain rule to calculate the partial derivatives of the objective function with respect to the input variables, thereby guiding the iterative update of the high-dimensional parameters. The low-dimensional critical boundary function is input into the iterative inversion optimizer to generate a critical boundary point cloud in the high-dimensional physical parameter space. The critical boundary point cloud represents a set of numerous high-dimensional discrete data points satisfying the critical conditions. The iterative update formula is: , in: This represents a high-dimensional physical parameter vector for the next iteration. This represents the high-dimensional physical parameter vector of the current iteration. Represents the learning rate parameter. This represents the operation of calculating the gradient of high-dimensional physical parameters. This represents the target low-dimensional eigenvector point that satisfies the low-dimensional critical boundary function.

[0055] Surface fitting is performed on the critical boundary point cloud. Surface fitting refers to the process of approximating a discrete point set using a continuous mathematical surface. A multivariate functional relationship, jointly defined by various environmental factors, is established. This multivariate functional relationship represents the mathematical equation describing the high-dimensional threshold surface, generating a dynamic composite ecological threshold surface. This dynamic composite ecological threshold surface represents a high-dimensional geometric surface reflecting the system's instability and critical state under the synergistic effect of multiple environmental factors. The specific fitting formula is as follows: , in, This represents a dynamic composite ecological threshold surface function. This represents the total number of basis functions. Representing the The fitting coefficient of the term, Representing the The polynomial basis functions of the terms, It represents a coordinate vector in a high-dimensional physical parameter space.

[0056] For example, the instability critical boundary separating the safe region from the instability hazard region is extracted in a low-dimensional feature space. The low-dimensional feature space is known to be two-dimensional. The instability critical boundary is parameterized to generate a low-dimensional critical boundary function. A normal vector is defined. for intercept constant for The low-dimensional critical boundary function is specifically as follows: Select low-dimensional coordinate points that satisfy the function. for .

[0057] Construct an iterative inversion optimizer to transform the low-dimensional feature space into the original high-dimensional physical parameter space. Input the low-dimensional critical boundary function into the iterative inversion optimizer. The iterative inversion optimizer generates the critical boundary point cloud in the high-dimensional physical parameter space by backpropagating gradient information. Set the initial high-dimensional parameter vector. for Set the learning rate parameter The value is 0.1, after passing through a nonlinear dimensionality reduction mapping function. The current low-dimensional mapping point is calculated as follows: Calculate the distance from the target point The error vector is Given the gradient vector of the error function with respect to the high-dimensional parameters in the current state. for Calculate the high-dimensional parameter vector for the next iteration based on the iterative update formula. After repeated iterations, when the error approaches zero, the final high-dimensional parameter vector is added to the critical boundary point cloud.

[0058] Assuming that the final result is a specific data point in the critical boundary point cloud. Surface fitting is performed on the critical boundary point cloud to establish a multivariate functional relationship jointly defined by various environmental factors. The total number of basis functions is set. The polynomial basis function is 4. for , for , for , Given a constant of 1, the fitting coefficients are solved using the least squares method. Assuming the calculation yields... It is 2.0. It is 1.5. for , for Generate a dynamic composite ecological threshold surface. The specific function of the dynamic composite ecological threshold surface is as follows: Verify specific data Substituting into the function, the result is It conforms to the relationship of multivariate functions.

[0059] S7. Perform sensitivity analysis along the gradient direction of the ecological phase space manifold baseline, quantify the contribution of each pressure source to system instability, and generate a factor weight ranking list. In one embodiment of the present invention, step S7 includes the following steps: calculating the normal vector and tangent vector of the manifold surface at the current ecological state point of the ecological phase space manifold reference, and determining the gradient direction of the system state evolution towards instability; performing differential perturbation on each environmental factor along the gradient direction, applying a unit perturbation amount to each environmental factor, measuring the displacement distance of the system state point in the manifold space caused by each perturbation, and generating a factor perturbation response vector; calculating the cosine value of the angle between each component of the factor perturbation response vector and the system instability gradient direction, normalizing the cosine value of each component to obtain the contribution weight of the corresponding environmental factor, and arranging them in descending order of weight value to generate a factor weight sorting list.

[0060] Specifically, the normal and tangent vectors of the manifold surface are calculated at the current ecological state point of the ecological phase space manifold benchmark. The ecological phase space manifold benchmark represents the benchmark geometric manifold reflecting the normal evolutionary law of the marine ecosystem. The current ecological state point represents the high-dimensional phase space position of the system at the current moment. The normal vector of the manifold surface represents a multidimensional vector perpendicular to the tangent plane of the manifold surface, and the tangent vector represents a multidimensional vector tangent to the manifold surface. The gradient direction of the system state evolving towards instability is determined. The gradient direction represents the fastest evolutionary path of the system state towards the dynamic composite ecological threshold surface. The calculation formula is as follows: , in, Represents the gradient direction. The normal vector represents the manifold surface. The normal vector is obtained by taking the partial derivative of the dynamic composite ecological threshold surface function generated in the previous step.

[0061] Then, differential perturbations are applied to each environmental factor along the gradient direction. Differential perturbation refers to the process of applying a small numerical change to a single environmental parameter to observe the system response. A unit perturbation amount is applied to each environmental factor, where the unit perturbation amount represents a set small change value, based on the minimum observable change according to the measurement resolution of various environmental sensors. The displacement distance of the system state point in the manifold space caused by each perturbation is measured. The displacement distance represents the position change vector of the state point in multidimensional space due to the perturbation. A factor perturbation response vector is generated, representing the vector set composed of the displacement distances of all environmental factors. The formula for calculating the displacement distance is: , in, Representing the The displacement distance vector corresponding to each environmental factor. Represents the manifold surface with respect to the first The tangent vector of the environmental factors, The unit of measurement is the disturbance amount.

[0062] The cosine of the angle between each component of the disturbance response vector and the system instability gradient direction is calculated. Each component represents the displacement distance vector corresponding to each environmental factor. The cosine of the angle is used to measure the degree of consistency between the disturbance direction and the instability direction. The calculation formula is as follows: , in, Representing the The cosine of the angle corresponding to each environmental factor Represents the vector dot product operation. and These represent the magnitudes of the corresponding vectors.

[0063] Finally, the cosine values ​​of each component are normalized to obtain the contribution weights of the corresponding environmental factors. The contribution weights represent the relative influence of each environmental factor on system instability, and their normalization formula is as follows: , in, Representing the The contribution weight of each environmental factor Represents the total number of environmental factors. The absolute value of the cosine of the angle represents the factor weight sorting list, which is generated by sorting the factors in descending order of their weight values. The factor weight sorting list represents the environmental factor sequence arranged from largest to smallest according to their contribution weight.

[0064] For example, the normal and tangent vectors of the manifold surface are calculated at the current ecological state point on the ecological phase space manifold baseline, given that the current ecological state point contains three environmental factor dimensions. The normal vector of the manifold surface is then calculated. for Determine the gradient direction of the system state evolving toward instability. for Differential perturbations are applied to each environmental factor along the gradient direction, and a unit perturbation amount is applied to each environmental factor. The unit perturbation amount is set. The value is 0.01. The displacement distance of the system state point in the manifold space induced by each disturbance is measured. The tangent vector of the manifold surface with respect to the first environmental factor is known. for Calculate the displacement distance vector corresponding to the first environmental factor according to the formula. Similarly, the displacement distance vector corresponding to the second environmental factor can be obtained. for The displacement distance vector corresponding to the third environmental factor for This generates a factor perturbation response vector, which contains , and Three components.

[0065] Next, the cosine values ​​of the angles between each component of the factor disturbance response vector and the direction of the system instability gradient are calculated. For the first environmental factor, the vector inner product is calculated. Calculate the modulus Calculate the modulus. Calculate the cosine value of the included angle. Similarly, the cosine value of the angle between the second environmental factor and the second environmental factor can be calculated. The value is 0.745, which is the cosine of the angle between the third environmental factor and the surrounding area. The cosine value of each component is normalized to obtain the contribution weight of the corresponding environmental factor. The sum of the absolute values ​​of the cosine values ​​is 1.784. The contribution weight of the first environmental factor is calculated as W1=0.472, the contribution weight of the second environmental factor is calculated as W2=0.418, and the contribution weight of the third environmental factor is calculated as W3=0.110. The factors are sorted in descending order of their weight values ​​to generate a sorted list of factor weights. The sorted results are the first environmental factor, the second environmental factor, and the third environmental factor.

[0066] S8. Automatically generate targeted intervention instructions based on the core environmental factor with the highest ranking in the factor weight ranking list, and send the intervention instructions to the corresponding execution control module; In one embodiment of the present invention, step S8 includes the following steps: selecting the core environmental factor ranked first in weight value from the factor weight ranking list; querying the pressure source type corresponding to the core environmental factor, wherein the pressure source type includes wind power disturbance source, photovoltaic shading source, or aquaculture source; calling the corresponding intervention strategy to generate a rule base according to the pressure source type; when the core environmental factor is dissolved oxygen concentration and the pressure source type is aquaculture source, generating an instruction to reduce the feeding frequency of a specific cage area; when the core environmental factor is light transmittance and the pressure source type is photovoltaic shading source, generating an instruction to adjust the tilt angle of the photovoltaic panel; when the core environmental factor is ocean current velocity and the pressure source type is wind power disturbance source, generating an instruction to optimize wind turbine operating parameters; when it is a combination of other environmental factors and pressure source types, generating corresponding conventional control instructions according to the intervention strategy rule base; encapsulating the generated intervention instructions into a control instruction data packet, and sending it to the corresponding execution control module through an industrial communication protocol, wherein the execution control module includes an aquaculture automation control unit, a photovoltaic array attitude adjustment unit, or a wind turbine operation management unit.

[0067] Specifically, the core environmental factor with the highest weight is selected from the factor weight ranking list. The factor weight ranking list represents the sequence of environmental factors arranged from largest to smallest according to their contribution weight. The core environmental factor represents the environmental parameter that contributes the most to system instability. The extraction formula is as follows: , in, Represents core environmental factors, The environmental factor that ranks first in the weighted list of representative factors.

[0068] The system also queries the stress source types corresponding to the core environmental factors. The stress source type represents the external disturbance source that causes abnormal fluctuations in the environmental factors. The stress source types include wind power disturbance sources, photovoltaic shading sources, or aquaculture sources. Wind power disturbance sources represent water disturbances caused by the operation of offshore wind turbines. Photovoltaic shading sources represent the shading effect of offshore photovoltaic arrays on sunlight. Aquaculture sources represent feeding and fish metabolic activities during cage aquaculture.

[0069] Based on the type of pressure source, the corresponding intervention strategy is invoked to generate a rule base. This rule base represents a pre-defined mapping table of control actions for different environmental anomalies and pressure sources. When the core environmental factor is dissolved oxygen concentration and the pressure source is aquaculture, a command to reduce the feeding frequency in a specific cage area is generated. This command represents a control command to reduce the number of times feed is given in the aquaculture area. When the core environmental factor is light transmittance and the pressure source is photovoltaic shading, a command to adjust the tilt angle of the photovoltaic panels is generated. This command represents a control command to change the angle between the photovoltaic modules and the horizontal plane. When the core environmental factor is ocean current velocity and the pressure source is wind power disturbance, a command to optimize wind turbine operating parameters is generated. This command represents a control command to adjust the wind turbine's speed or pitch angle. When there are other combinations of environmental factors and pressure source types, corresponding routine control commands are generated based on the intervention strategy rule base. These routine control commands represent standard response commands for atypical environmental factor anomalies. The command generation logic formula is as follows: , in, This represents the generated intervention instructions. The mapping function that represents the rule base for generating intervention strategies. This represents the type of pressure source.

[0070] The generated intervention commands are encapsulated into control command data packets. Each control command data packet represents a set of command data formatted according to a specific communication standard. The encapsulation formula is as follows: , in, This represents a control command data packet. Represents the protocol header identifier. This represents a data concatenation operation. Represents the network address of the target device. The checksum used for error detection is sent to the corresponding execution control module via an industrial communication protocol. The industrial communication protocol represents the standard communication rules for data exchange between industrial control devices. The execution control module represents the hardware device that receives and executes control commands. The execution control module includes an aquaculture automation control unit, a photovoltaic array attitude adjustment unit, or a wind turbine operation management unit. The aquaculture automation control unit represents a system that controls aquaculture equipment such as feeders. The photovoltaic array attitude adjustment unit represents a system that controls the photovoltaic panel support motor. The wind turbine operation management unit represents a system that controls the operating status of wind turbine generators.

[0071] For example, the core environmental factor with the highest weight value is selected from the factor weight ranking list. It is known that the environmental factor ranked first in the factor weight ranking list generated in the previous steps is light transmittance. Based on the extraction formula, the core environmental factor is determined to be light transmittance. The corresponding stress source type is then queried. By querying the system configuration table, the stress source type corresponding to light transmittance is determined to be photovoltaic shading source. Based on the stress source type, the corresponding intervention strategy is called to generate the rule base. Since the core environmental factor is light transmittance and the stress source type is photovoltaic shading source, a logical mapping function is generated based on the instruction to generate an instruction to adjust the tilt angle of the photovoltaic panel. The instruction code for this intervention instruction is set to 1011.

[0072] The generated intervention command is encapsulated into a control command data packet, with the protocol header identifier set to 170. The network address of the target device is set to 192. The checksum is calculated as the remainder of the arithmetic sum of the protocol header identifier, command code, and network address divided by 256. The arithmetic sum is 170 + 1011 + 192, which equals 1373. The remainder of 1373 divided by 256 is 93, resulting in a checksum of 93. Based on the data concatenation operation, a control command data packet containing the data sequence 170, 1011, 192, and 93 is generated. This control command data packet is sent to the corresponding execution control module via the industrial communication protocol. Since the target is to adjust the photovoltaic panel, the corresponding execution control module is the photovoltaic array attitude adjustment unit. After receiving the control command data packet, the photovoltaic array attitude adjustment unit parses the command to adjust the tilt angle of the photovoltaic panel and executes the corresponding action.

[0073] It should be noted that the electrical connections between the various units described above do not necessarily represent direct or indirect connections. Any indirect connection method can be applied to the embodiments of the present invention as long as it achieves the purpose of the present invention. The above are merely exemplary embodiments of the present invention and should not be construed as limiting the scope of the present invention.

[0074] All equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of other embodiments of this invention upon considering the specification and the disclosure of practical truth. This application is intended to cover any variations, uses, or adaptations of this invention that follow the general principles of this invention and include common knowledge or conventional techniques in the art not described herein.

Claims

1. A method for intelligent early warning of ecological thresholds for marine use involving both wind, solar, and fisheries, characterized in that, Includes the following steps: Multidimensional environmental time-series monitoring data of the sea area where wind, solar and fishery are used are acquired and time alignment and missing value imputation are performed to generate a standardized multidimensional environmental state sequence. Based on the phase space reconstruction algorithm, a high-dimensional mapping transformation is performed on the standardized multidimensional environmental state sequence to construct an ecological phase space manifold benchmark. Continuously collect micro-disturbance events and track the evolution trajectory of the current ecological state point in the ecological phase space manifold benchmark in real time. Calculate the time scale required for the current ecological state point to return from the disturbance offset position to the equilibrium domain, extract the autoregressive coefficient and state variance, and generate a critical slowing characteristic index sequence. A time window sliding statistical analysis is performed on the critical slowdown characteristic index sequence. When the autoregressive coefficient continues to rise and approaches one and the state variance increases significantly, an ecological instability precursor signal is generated. In response to the ecological instability precursor signal, the manifold space dimensionality reduction mapping calculation is initiated, and the critical slowing characteristic index sequence is nonlinearly projected in the low-dimensional feature space to identify the proximity between the current system state and historical instability samples. When the distance metric is less than a preset safety threshold, the dynamic threshold inversion calculation process is triggered. The gradient backpropagation algorithm is used to map the unstable critical boundary in the low-dimensional feature space back to the original high-dimensional physical parameter space, and the dynamic composite ecological threshold surface is calculated. Sensitivity analysis is performed along the gradient direction of the ecological phase space manifold baseline to quantify the contribution of each pressure source to system instability and generate a factor weight ranking list. Based on the core environmental factor ranked highest in the factor weight sorting list, a targeted intervention instruction is automatically generated and sent to the corresponding execution control module.

2. The intelligent early warning method for ecological thresholds of marine use in conjunction with wind, solar and fishery activities according to claim 1, characterized in that, The process of acquiring multidimensional environmental time-series monitoring data of the sea area used for wind, solar and fishery co-operation, and performing time alignment and missing value interpolation to generate a standardized multidimensional environmental state sequence includes: collecting raw environmental monitoring data streams reported by a distributed sensor network, which include time-series data on light transmittance, ocean current velocity, dissolved oxygen concentration, chlorophyll concentration, and dissolved inorganic nitrogen concentration; uniformly calibrating the timestamps of each sensor in the raw environmental monitoring data stream, interpolating and aligning data sequences with different sampling frequencies to a unified time grid to generate a time-synchronized environmental data matrix; detecting outliers and missing values ​​in the time-synchronized environmental data matrix, using the isolated forest algorithm to remove outliers, and using the cubic spline interpolation algorithm to complete the missing time periods to generate a cleaned environmental data matrix; and performing dimensional normalization processing on each environmental factor in the cleaned environmental data matrix to generate a standardized multidimensional environmental state sequence.

3. The intelligent early warning method for ecological thresholds of marine use in the same area as wind, solar and fishery as described in claim 1, characterized in that, The step of performing a high-dimensional mapping transformation on the standardized multidimensional environmental state sequence based on the phase space reconstruction algorithm to construct an ecological phase space manifold benchmark includes: calculating the autocorrelation function of the time series data of each environmental factor in the standardized multidimensional environmental state sequence to determine the time delay parameter; performing a phase space embedding transformation on the standardized multidimensional environmental state sequence according to the time delay parameter, expanding the one-dimensional time series of each environmental factor into a set of trajectory points in a high-dimensional phase space to generate a phase space trajectory coordinate matrix; and performing manifold fitting calculation on the phase space trajectory coordinate matrix, constructing the topological relationship between neighboring points through a local linear embedding algorithm to form an ecological phase space manifold benchmark.

4. The intelligent early warning method for ecological thresholds of marine use in conjunction with wind, solar and fishery activities according to claim 1, characterized in that, The process involves continuously collecting micro-disturbance events and tracking the evolution trajectory of the current ecological state point in the ecological phase space manifold benchmark in real time. It calculates the time scale required for the current ecological state point to return from the disturbance offset position to the equilibrium domain, extracts autoregressive coefficients and state variance, and generates a critical slowing characteristic index sequence. This includes monitoring tidal change events, meteorological fluctuation events, and aquaculture operation events in the marine area. When the abrupt change amplitude of each environmental factor exceeds the preset background fluctuation standard deviation, it is marked as a micro-disturbance event, and the time of occurrence of the disturbance is recorded. The initial coordinates of the ecological state point corresponding to the time of occurrence of the micro-disturbance event in the ecological phase space manifold benchmark are extracted. The spatial position evolution of the state point within subsequent time windows is continuously tracked, and the Euclidean distance between the state point and the center of the equilibrium domain is calculated over time as a decay curve. A first-order autoregressive model is fitted to the decay curve, and the autoregressive coefficient is extracted as the first feature parameter. Simultaneously, the variance of the decay curve is calculated as the second feature parameter. The first and second feature parameters are organized in chronological order to generate a critical slowing characteristic index sequence.

5. The intelligent early warning method for ecological thresholds of marine use in conjunction with wind, solar and fishery activities according to claim 1, characterized in that, The step of performing time-window sliding statistical analysis on the critical slowing characteristic index sequence, generating an ecological instability precursor signal when the autoregressive coefficient continuously rises and approaches one and the state variance significantly increases, includes: performing sliding window traversal on the critical slowing characteristic index sequence according to a preset sliding time window length, calculating the trend slope of the autoregressive coefficient and the coefficient of variation of the state variance for each window position, and generating a dynamic statistical characteristic sequence; establishing an autoregressive coefficient trend criterion, marking a resilience decay state when the trend slope of the autoregressive coefficient in the dynamic statistical characteristic sequence is greater than zero and the mean of the autoregressive coefficient enters a preset threshold range; establishing a state variance fluctuation criterion, marking a system instability enhancement state when the coefficient of variation of the state variance in the dynamic statistical characteristic sequence exceeds a preset multiple threshold of the historical baseline level; and generating an ecological instability precursor signal when the resilience decay state and the system instability enhancement state simultaneously satisfy a duration exceeding a preset minimum warning period.

6. The intelligent early warning method for ecological thresholds of marine use in conjunction with wind, solar and fishery activities according to claim 1, characterized in that, The process involves initiating a manifold space dimensionality reduction mapping calculation in response to the ecological instability precursor signal. This calculation performs a nonlinear projection of the critical slowing characteristic index sequence onto a low-dimensional feature space to identify the proximity of the current system state to historical instability samples. When the distance metric is less than a preset safety threshold, a dynamic threshold inversion calculation process is triggered, including: loading a historical ecological instability event sample library containing labeled critical slowing characteristic index sequence samples before instability; performing low-dimensional manifold learning on the samples in the historical ecological instability event sample library to construct a historical instability feature distribution map. The system analyzes the following steps: First, it inputs the currently collected critical slowdown characteristic index sequence into a pre-trained nonlinear dimensionality reduction mapping model to generate a low-dimensional feature vector of the current system state. Then, it calculates the Mahalanobis distance between the low-dimensional feature vector and each historical instability sample point in the historical instability characteristic distribution map, identifies the nearest neighbor historical instability sample, and uses the minimum Mahalanobis distance as the instability proximity metric of the current system state. Next, it compares the instability proximity metric with a preset safety threshold. If the instability proximity metric is less than the preset safety threshold, it triggers a dynamic threshold inversion calculation process; otherwise, it maintains the current monitoring state.

7. The intelligent early warning method for ecological thresholds of marine use in conjunction with wind, solar and fishery activities according to claim 1, characterized in that, The process of mapping the instability critical boundary in the low-dimensional feature space back to the original high-dimensional physical parameter space using the gradient backpropagation algorithm to calculate the dynamic composite ecological threshold surface includes: extracting the instability critical boundary separating the safe region from the instability danger region in the low-dimensional feature space; parameterizing the instability critical boundary to generate a low-dimensional critical boundary function; constructing an iterative inversion optimizer from the low-dimensional feature space to the original high-dimensional physical parameter space; inputting the low-dimensional critical boundary function into the iterative inversion optimizer through backpropagation of gradient information to generate a critical boundary point cloud in the high-dimensional physical parameter space; and performing surface fitting on the critical boundary point cloud to establish a multivariate functional relationship jointly defined by the environmental factors to generate the dynamic composite ecological threshold surface.

8. The intelligent early warning method for ecological thresholds of marine use in conjunction with wind, solar and fishery activities according to claim 1, characterized in that, The sensitivity analysis along the gradient direction of the ecological phase space manifold baseline, quantifying the contribution of each pressure source to system instability, and generating a factor weight ranking list includes: calculating the normal and tangent vectors of the manifold surface at the current ecological state point of the ecological phase space manifold baseline to determine the gradient direction of the system state evolution towards instability; performing differential perturbation on each environmental factor along the gradient direction, applying a unit perturbation amount to each environmental factor, measuring the displacement distance of the system state point in the manifold space caused by each perturbation, and generating a factor perturbation response vector; calculating the cosine value of the angle between each component of the factor perturbation response vector and the system instability gradient direction, normalizing the cosine value of each component to obtain the contribution weight of the corresponding environmental factor, and arranging them in descending order of weight value to generate a factor weight ranking list.

9. The intelligent early warning method for ecological thresholds of marine use in the co-use of wind, solar and fisheries as described in claim 1, characterized in that, The process of automatically generating targeted intervention instructions based on the highest-ranked core environmental factor in the factor weight ranking list and sending the intervention instructions to the corresponding execution control module includes: selecting the core environmental factor with the highest weight value from the factor weight ranking list; querying the pressure source type corresponding to the core environmental factor, where the pressure source type includes wind power disturbance source, photovoltaic shading source, or aquaculture source; calling the corresponding intervention strategy to generate a rule base based on the pressure source type; generating an instruction to reduce the feeding frequency in a specific cage area when the core environmental factor is dissolved oxygen concentration and the pressure source type is aquaculture source; generating an instruction to adjust the tilt angle of photovoltaic panels when the core environmental factor is light transmittance and the pressure source type is photovoltaic shading source; generating an instruction to optimize wind turbine operating parameters when the core environmental factor is ocean current velocity and the pressure source type is wind power disturbance source; and generating corresponding conventional control instructions based on the intervention strategy rule base when other environmental factors and pressure source types are combined; and encapsulating the generated intervention instructions into a control instruction data packet and sending it to the corresponding execution control module via an industrial communication protocol, where the execution control module includes an aquaculture automation control unit, a photovoltaic array attitude adjustment unit, or a wind turbine operation management unit.

10. A system employing the intelligent early warning method for marine ecological thresholds in the context of combined wind, solar, and fishery use as described in any one of claims 1-9, characterized in that, include: The data acquisition module is used to acquire multi-dimensional environmental time-series monitoring data of the sea area where wind, solar and fishery are used together, and to perform time alignment processing and missing value interpolation processing to generate a standardized multi-dimensional environmental state sequence. The phase space reconstruction module is used to perform high-dimensional mapping transformation on the standardized multidimensional environmental state sequence based on the phase space reconstruction algorithm to construct an ecological phase space manifold benchmark. The critical slowing identification module is used to continuously collect micro-disturbance events and track the evolution trajectory of the current ecological state point in the ecological phase space manifold benchmark in real time. It calculates the time scale required for the current ecological state point to return from the disturbance offset position to the equilibrium domain, extracts the autoregressive coefficient and state variance, and generates a critical slowing characteristic index sequence. The precursor signal generation module is used to perform time window sliding statistical analysis on the critical slowing characteristic index sequence. When the autoregressive coefficient continues to rise and approaches one and the state variance increases significantly, an ecological instability precursor signal is generated. The dimension reduction mapping module is used to initiate manifold space dimension reduction mapping calculation in response to the ecological instability precursor signal, and to perform nonlinear projection of the critical slowing characteristic index sequence in the low-dimensional feature space to identify the proximity between the current system state and historical instability samples. When the distance metric is less than a preset safety threshold, the dynamic threshold inversion calculation process is triggered. The threshold inversion module is used to map the unstable critical boundary in the low-dimensional feature space back to the original high-dimensional physical parameter space using the gradient backpropagation algorithm, and solve for the dynamic composite ecological threshold surface. The sensitivity analysis module is used to perform sensitivity analysis along the gradient direction of the ecological phase space manifold baseline, quantify the contribution of each pressure source to system instability, and generate a factor weight ranking list. The intervention instruction generation module is used to automatically generate targeted intervention instructions based on the core environmental factor with the highest ranking in the factor weight ranking list, and send the intervention instructions to the corresponding execution control module.

Images