A coastal nuclear power biodiversity monitoring method

Abstract

The present application provides a kind of coastal nuclear power biodiversity monitoring method, belong to the technical field of coastal nuclear power ecological influence, the present application is by in the nuclear power plant surrounding coastal area layout multi-sensor array acquisition site synchronous acquisition water quality parameter, acoustic signal, image data and plankton density such as multidimensional ecological information, constructs multidimensional ecological characteristic data set, utilizes the dense connection network architecture of marine ecological correlation analysis model to learn the nonlinear mapping relationship between multi-source data and outputs biodiversity evaluation index and community structure characteristic parameter vector, combines ecological dynamic early warning model to carry out future ecological change trend prediction, when detecting ecological abnormal signal, start ecological risk comprehensive regulation mode and carry out adaptive optimization adjustment of monitoring parameter, solve the technical problem that coastal nuclear power station ecological monitoring data is multi-source heterogeneous and insufficient in space-time correlation analysis capability.

Description

Technical Field

[0001] This invention belongs to the field of coastal nuclear power ecological impact technology, and specifically relates to a method for monitoring biodiversity in coastal nuclear power plants. Background Technology

[0002] Ecological environment monitoring at coastal nuclear power plants is a crucial technological area for ensuring the coordinated development of nuclear power generation and marine ecosystems. Traditional biodiversity monitoring primarily employs methods such as single-sensor fixed-point sampling, regular manual inspections, and laboratory sample analysis to acquire water quality parameters and biological population data, followed by ecological status assessment using statistical methods. However, traditional monitoring technologies suffer from limitations such as single data acquisition dimensions, limited spatiotemporal coverage, and weak multi-source data fusion processing capabilities. These limitations hinder the in-depth correlation analysis between water quality parameters and biological community structure, and prevent accurate identification of potential impact patterns of nuclear power operation on marine ecosystems. Given the increasing number of coastal nuclear power plants, the complexity and dynamism of marine ecosystems make it difficult for traditional monitoring methods to establish nonlinear mapping relationships between multidimensional ecological characteristic data, thus hindering the prediction, early warning, and adaptive regulation and control of ecological risks. In other words, existing technologies suffer from the technical challenges of multi-source heterogeneous ecological monitoring data from coastal nuclear power plants and insufficient spatiotemporal correlation analysis capabilities. Summary of the Invention

[0003] In view of this, the present invention provides a method for monitoring biodiversity in coastal nuclear power plants, which can solve the technical problems of multi-source heterogeneity and insufficient spatiotemporal correlation analysis capabilities of ecological monitoring data in coastal nuclear power plants in the prior art.

[0004] This invention is implemented as follows: It provides a method for monitoring biodiversity in coastal nuclear power plants, comprising deploying multi-sensor array data collection stations in the coastal area surrounding the nuclear power plant. Each multi-sensor array data collection station is equipped with water quality parameter sensors, acoustic biological monitoring equipment, underwater camera devices, and plankton samplers, covering the surrounding sea area. The multi-sensor array data collection stations collect real-time data on seawater temperature, salinity, dissolved oxygen concentration, turbidity, and chlorophyll a concentration; collect underwater acoustic signal spectrum data; collect seabed biological image data; collect plankton density data; and simultaneously record time-series data on tidal cycle parameters, wind speed values, and ocean current velocity parameters. The collected data are then processed using a multi-scale adaptive filtering algorithm for noise suppression. The process involves: extracting statistical features to obtain water quality statistical feature vectors; performing frequency domain analysis to obtain acoustic spectrum feature parameters; performing image recognition processing to obtain biological species identification results; and performing species classification statistics to obtain plankton community structure parameters. These parameters are then combined to construct a multidimensional ecological feature dataset. This dataset is then input into a marine ecological correlation analysis model for nonlinear relationship modeling, outputting a biodiversity assessment index and a community structure feature parameter vector. The time-series data of the biodiversity assessment index and community structure feature parameter vector are input into an ecological dynamic early warning model for trend prediction analysis. When the biodiversity assessment index, community structure deviation index, and water temperature anomaly index meet preset conditions, a comprehensive ecological risk regulation mode is activated.

[0005] The distance between the multi-sensor array data collection stations ranges from 500m to 2000m, covering a sea area with a radius of 5000m around the nuclear power plant, with a data collection frequency of once per minute.

[0006] Specifically, the multi-scale adaptive filtering algorithm decomposes the seawater temperature signal, salinity signal, and dissolved oxygen concentration signal into different frequency band components through wavelet decomposition, and identifies and filters out low-frequency noise components and high-frequency noise components caused by tidal period parameters, ocean current velocity parameters, and wind speed values.

[0007] Specifically, the water quality statistical feature vector is a 20-dimensional feature vector obtained by performing statistical analysis on the mean, variance, skewness, and kurtosis of seawater temperature, salinity, dissolved oxygen concentration, turbidity, and chlorophyll a concentration data.

[0008] Specifically, the acoustic spectrum characteristic parameters are eight-dimensional characteristic parameters extracted from the underwater acoustic signal spectrum data through frequency domain analysis, including the main frequency, bandwidth, signal strength, and spectral entropy.

[0009] Specifically, the biological species identification results are obtained by classifying marine biological image data using deep learning image recognition algorithms, resulting in species category labels and individual quantity statistics.

[0010] Specifically, the planktonic community structure parameters are a six-dimensional parameter vector obtained by statistically analyzing planktonic density data according to species classification, including species richness, evenness index, dominance index, and diversity index.

[0011] Specifically, the multidimensional ecological feature dataset is a 34-dimensional feature data set formed by combining water quality statistical feature vectors, acoustic spectrum feature parameters, biological species identification results, and plankton community structure parameters in a time series manner.

[0012] Specifically, the structure of the marine ecological association analysis model is as follows: the input layer receives a 34-dimensional multidimensional ecological feature dataset, and performs feature extraction and feature fusion through three dense connection blocks. Each dense connection block contains 64 neurons and residual connection structures. The output layer generates a 1-dimensional biodiversity assessment index and a 6-dimensional community structure feature parameter vector.

[0013] Specifically, the ecological dynamic early warning model adopts a long short-term memory network architecture combined with a multi-head attention mechanism. It predicts future trends by analyzing historical time-series data of biodiversity assessment index and community structure characteristic parameter vectors, with a prediction period of 7 to 30 days.

[0014] Specifically, the training of the marine ecological association analysis model is carried out using the Adam optimizer to update the model parameters, with a learning rate of 0.001, a batch size of 128, and 200 training rounds.

[0015] The community structure deviation index is specifically obtained by normalizing the Euclidean distance between the current community structure feature parameter vector and the historical baseline vector. The calculation formula is as follows: .

[0016] The water temperature anomaly index is specifically calculated using the standardized difference between the current seawater temperature and the historical average for the same period. The calculation formula is as follows: .

[0017] Specifically, the preset conditions are that the biodiversity assessment index is less than 0.6, the community structure deviation index is greater than 15%, and the water temperature anomaly index is greater than 2.0.

[0018] Specifically, the ecological risk comprehensive regulation mode adjusts the data acquisition frequency of the multi-sensor array acquisition stations to once every 10 seconds through a multi-parameter coordinated control algorithm, increases the sampling density parameter by 50%, reduces the early warning threshold parameter by 20%, and initiates manual inspection and verification.

[0019] Before constructing the multidimensional ecological feature dataset, the process also includes collecting monitoring data from the multi-sensor array acquisition sites of the coastal nuclear power plant over a period of three years as the basic dataset. Through data cleaning and standardization, a training set containing 150,000 samples and a validation set containing 30,000 samples are constructed.

[0020] This invention achieves simultaneous, multi-dimensional monitoring of water quality parameters, acoustic signals, image data, and plankton density by constructing a multi-sensor array acquisition station. It establishes a marine ecological correlation analysis model to learn the nonlinear mapping relationships between different ecological characteristic parameters, outputting a comprehensive biodiversity assessment index and community structure characteristic parameter vectors. A multi-scale adaptive filtering algorithm effectively removes environmental noise interference, and a densely connected network structure enables deep feature fusion of multi-source heterogeneous data. Combined with an ecological dynamic early warning model, it accurately predicts future ecological change trends. When an ecological anomaly signal is detected, an ecological risk comprehensive regulation mode is automatically activated to optimize and adjust parameters. This invention establishes a quantitative correlation between water quality environment and biological community, achieving effective integration and spatiotemporal correlation analysis of multi-source ecological monitoring data, solving the problems of data isolation and insufficient correlation analysis capabilities in traditional technologies. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method of the present invention.

[0022] Figure 2 The image shows the processing effect of the multi-scale adaptive filtering algorithm in Example 2.

[0023] Figure 3 This is a time-series graph showing the changes in acoustic, salinity, and biodiversity assessment indices in Example 2.

[0024] Figure 4 This is a time-series graph showing the changes in plankton enrichment, chlorophyll a, and water temperature in Example 2. Detailed Implementation

[0025] 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.

[0026] like Figure 1 The diagram shown is a flowchart of a biodiversity monitoring method for coastal nuclear power plants provided by this invention. The method includes the following steps:

[0027] S01. Deploy multi-sensor array data collection stations in the coastal area surrounding the nuclear power plant. Each multi-sensor array data collection station is equipped with water quality parameter sensors, acoustic biological monitoring equipment, underwater camera devices, and plankton samplers. The distance between the multi-sensor array data collection stations is within the range of [500m, 2000m], covering a sea area with a radius of 5000m around the nuclear power plant.

[0028] S02. Real-time data collection of seawater temperature, salinity, dissolved oxygen concentration, turbidity, and chlorophyll a concentration is collected using a multi-sensor array at the collection station. Underwater acoustic signal spectrum data is collected using an acoustic biological monitoring device. Underwater biological image data is collected using an underwater camera device. Plankton density data is collected using a plankton sampler. The data collection frequency is once per minute. At the same time, time series data of tidal cycle parameters, wind speed values, and ocean current speed parameters are recorded.

[0029] S03. A multi-scale adaptive filtering algorithm is used to suppress noise in seawater temperature, salinity, dissolved oxygen concentration, turbidity, chlorophyll a concentration data, underwater acoustic signal spectrum data and seabed biological image data. The seawater temperature signal, salinity signal and dissolved oxygen concentration signal are decomposed into different frequency band components through wavelet decomposition, and low-frequency noise components and high-frequency noise components caused by tidal period parameters, ocean current velocity parameters and wind speed values ​​are identified and filtered out.

[0030] S04. Statistical feature extraction is performed on the filtered seawater temperature, salinity, dissolved oxygen concentration, turbidity, and chlorophyll a concentration data to obtain water quality statistical feature vectors. Frequency domain analysis is performed on the underwater acoustic signal spectrum data to obtain acoustic spectrum feature parameters. Image recognition processing is performed on the seabed biological image data to obtain biological species identification results. Species classification statistics are performed on the planktonic density data to obtain planktonic community structure parameters. The water quality statistical feature vectors, acoustic spectrum feature parameters, biological species identification results, and planktonic community structure parameters are combined to construct a multidimensional ecological feature dataset.

[0031] S05. Input the multidimensional ecological feature dataset into the marine ecological association analysis model to perform nonlinear relationship modeling. The marine ecological association analysis model learns the mapping relationship between water quality statistical feature vector and planktonic community structure parameters through dense connection layers, learns the association relationship between acoustic spectrum feature parameters and biological species identification results, and outputs biodiversity assessment index and community structure feature parameter vector.

[0032] S06. Input the time series data of biodiversity assessment index and community structure characteristic parameter vector into the ecological dynamic early warning model for trend prediction analysis. The ecological dynamic early warning model calculates the predicted value of biodiversity assessment index and community structure characteristic parameter vector in the future time period in combination with ecological dynamic constraints. The prediction time period is [7 days, 30 days].

[0033] S07. When the biodiversity assessment index is <0.6, the community structure deviation index is >15%, and the water temperature anomaly index is >2.0, the ecological risk comprehensive regulation mode is activated. The data acquisition frequency, sampling density parameter, and early warning threshold parameter of the multi-sensor array acquisition station are adjusted through a multi-parameter coordinated control algorithm to achieve adaptive optimization control.

[0034] The multi-sensor array acquisition station refers to a combination of integrated monitoring equipment arranged at regular intervals in the sea area. Each multi-sensor array acquisition station includes the synchronous data acquisition function of water quality parameter sensors, acoustic biological monitoring equipment, underwater camera devices, and plankton samplers.

[0035] The water quality parameter sensors refer to instruments and equipment used to measure the physicochemical properties of seawater, including temperature sensors, salinity sensors, dissolved oxygen sensors, turbidity sensors, and chlorophyll a sensors.

[0036] The acoustic biological monitoring equipment refers to an underwater acoustic receiving device that collects acoustic signals emitted by marine organisms through passive acoustic listening technology, and is used to identify the bioacoustic characteristics of fish, marine mammals and crustaceans.

[0037] The underwater camera device refers to a high-resolution camera device installed on the seabed, used to continuously capture images of seabed biological activities and habitat conditions.

[0038] The planktonic sampler refers to an automated sampling device that obtains phytoplankton and zooplankton samples by filtering seawater, and is equipped with a microscope imaging system for species identification and counting.

[0039] The multi-scale adaptive filtering algorithm is based on wavelet transform and multi-level decomposition filtering technology. It removes environmental noise by analyzing the characteristics of the signal at different time and frequency scales.

[0040] The water quality statistical feature vector is a 20-dimensional feature vector obtained by performing statistical analysis of mean, variance, skewness, and kurtosis on seawater temperature, salinity, dissolved oxygen concentration, turbidity, and chlorophyll a concentration data.

[0041] The acoustic spectrum characteristic parameters are eight-dimensional characteristic parameters extracted from the underwater acoustic signal spectrum data through frequency domain analysis, including the main frequency, bandwidth, signal strength, and spectral entropy.

[0042] The species identification results are obtained by classifying marine biological image data using deep learning image recognition algorithms, resulting in species category labels and individual counts.

[0043] The phytoplankton community structure parameters are a six-dimensional parameter vector consisting of species richness, evenness index, dominance index, and diversity index, obtained by statistically analyzing phytoplankton density data according to species classification.

[0044] The multidimensional ecological feature dataset is a 34-dimensional feature data set formed by combining water quality statistical feature vectors, acoustic spectrum feature parameters, biological species identification results, and plankton community structure parameters in a time series manner.

[0045] The specific structure of the marine ecological association analysis model is as follows: the input layer receives a 34-dimensional multidimensional ecological feature dataset, and performs feature extraction and feature fusion through three dense connection blocks. Each dense connection block contains 64 neurons and residual connection structures. The output layer generates a 1-dimensional biodiversity assessment index and a 6-dimensional community structure feature parameter vector.

[0046] The steps for establishing the training dataset for the marine ecological correlation analysis model specifically include collecting monitoring data from multi-sensor array acquisition sites of coastal nuclear power plants over a period of 3 years as the basic dataset. The basic dataset includes water quality statistical feature vectors, acoustic spectrum feature parameters, biological species identification results, and plankton community structure parameters. Through data cleaning and standardization, a training set containing 150,000 samples and a validation set containing 30,000 samples are constructed.

[0047] The training steps for the marine ecological association analysis model specifically include updating model parameters using the Adam optimizer, with a learning rate of 0.001, a batch size of 128, and 200 training epochs. The training process employs a data augmentation-based sample diversity expansion mechanism, which increases sample diversity and improves robustness in the training data preprocessing step by adding 5% Gaussian noise to the water quality statistical feature vector and random pruning of time-series data; and utilizes a gradient-based training stability enhancement mechanism, by adding... The regularization term stabilizes the training dynamics and prevents pattern collapse during the adversarial training step.

[0048] The dense connection block refers to a network structure module in a neural network in which each layer is directly connected to all the preceding layers. Through dense connections between forward layers, features are reused and parameter redundancy is reduced during network propagation.

[0049] The biodiversity assessment index is a dimensionless index that comprehensively reflects the richness, evenness, and stability of marine biological communities. Its value ranges from [0, 1], and the closer the value is to 1, the better the biodiversity status.

[0050] The community structure characteristic parameter vector is a 6-dimensional parameter vector containing species richness, evenness index, dominance index, diversity index, community stability index, and biomass index, used to quantitatively describe the structural characteristics of biological communities.

[0051] The ecological dynamic early warning model adopts a long short-term memory network architecture combined with a multi-head attention mechanism, and predicts future trends by analyzing historical time-series data of biodiversity assessment index and community structure characteristic parameter vectors.

[0052] The community structure deviation index is obtained by normalizing the Euclidean distance between the current community structure feature parameter vector and the historical baseline vector. It reflects the degree of deviation of the community structure from the normal state. The calculation formula is as follows: ,in This is the current community structure feature parameter vector. This is the historical baseline vector.

[0053] The sea temperature anomaly index is calculated using the standardized difference between the current seawater temperature and the historical average for the same period. It is used to quantify the degree to which the seawater temperature deviates from the normal range. The calculation formula is as follows: ,in Given the current seawater temperature, This is the historical average seawater temperature for the same period. This represents the standard deviation of seawater temperature for the same historical period.

[0054] The multi-parameter coordinated control algorithm utilizes a structured prediction framework based on a probabilistic graphical model. It models the constraint relationships between the biodiversity assessment index, community structure deviation index, and water temperature anomaly index as edge connections in a graph. Probabilistic information is propagated between graph nodes through a message passing algorithm. Variational inference techniques are used to approximate the posterior distribution, and Gibbs sampling is used to generate structured prediction results.

[0055] The aforementioned comprehensive ecological risk regulation mode refers to an emergency monitoring state that is automatically activated when an abnormal ecological signal is detected. This includes comprehensive measures such as adjusting the data collection frequency to once every 10 seconds, increasing the sampling density parameter by 50%, reducing the warning threshold parameter by 20%, and initiating manual inspection and verification.

[0056] The sampling density parameter refers to the number of times a planktonic sampler is used per unit sea area. Under normal conditions, it is twice per hour, and under the comprehensive ecological risk regulation mode, it is adjusted to three times per hour.

[0057] The warning threshold parameters refer to the critical values ​​of key indicators that trigger ecological risk warnings, including a biodiversity assessment index threshold of 0.6, a community structure deviation index threshold of 15%, and a water temperature anomaly index threshold of 2.0.

[0058] This invention is also implemented using a computer to form a biodiversity monitoring system for coastal nuclear power plants. The computer is equipped with a readable storage medium that stores program instructions, which are used to execute the above-described methods when the computer is run.

[0059] The specific implementation methods of the above steps are described in detail below.

[0060] The specific implementation of step S01 involves establishing a marine spatial layout plan. First, a topographic map of the sea area within a 5000m radius surrounding the nuclear power plant is drawn using a marine geographic information system, identifying key parameters such as water depth distribution, seabed slope, and sediment type. Next, a grid-based monitoring method is used to divide the monitoring area into several monitoring units, with each monitoring unit's area controlled to be less than 0.25 square kilometers. Up to 4 The monitoring stations are evenly distributed within a distance of 500m to 2000m. Priority monitoring areas are then determined based on ocean current direction, tidal influence, and the discharge path of cooling water from the nuclear power plant. Within these areas, the density of monitoring stations is increased to 1.5 times the original density. Finally, a comprehensive monitoring system, including water quality parameter sensors, acoustic biological monitoring equipment, underwater cameras, and plankton samplers, is installed at each selected location. All equipment is connected to an onshore data center via submarine cable or wireless transmission. The purpose of this step is to construct a comprehensive and rationally laid-out multi-parameter synchronous monitoring network to provide sufficient spatial data support for subsequent biodiversity assessments.

[0061] The specific implementation of step S02 involves establishing a multi-parameter synchronous data acquisition system. First, a water quality parameter sensor group collects key water quality indicators such as seawater temperature, salinity, dissolved oxygen concentration, turbidity, and chlorophyll a concentration every minute, with acquisition accuracies reaching ±0.1℃, ±0.1‰, ±0.1mg / L, ±1NTU, and ±0.1μg / L, respectively. Simultaneously, an acoustic biological monitoring device continuously records underwater acoustic signals in the frequency range of 20Hz to 20kHz, with a sampling frequency set to 48kHz to ensure signal integrity. Digital filters are used to remove interference from ship noise and mechanical noise. An underwater camera continuously records seabed biological activity at a frequency of 30 high-resolution images per minute, with an image resolution of no less than 1920×1080 pixels, and is equipped with infrared supplemental lighting to adapt to different lighting conditions. A plankton sampler extracts 1L of seawater sample per minute using a constant-volume sampling method. After multi-stage filtration, a microscope imaging system is used to identify and count individual plankton of different particle sizes. Simultaneously record environmental parameters such as tidal height changes, wind speed and direction, and ocean current speed and direction, maintaining the recording frequency consistent with biological monitoring data. The purpose of this step is to obtain multidimensional basic data reflecting the state of the marine ecosystem, laying the data foundation for subsequent feature extraction and correlation analysis.

[0062] The specific implementation of step S03 involves multi-level noise suppression processing. First, the original monitoring data is preprocessed. An outlier detection algorithm identifies and marks outlier data points exceeding three standard deviations of the normal range, and linear interpolation is used to supplement missing data. Next, a multi-scale adaptive filtering algorithm is applied to process seawater temperature, salinity, dissolved oxygen concentration, turbidity, and chlorophyll a concentration data. This algorithm, based on wavelet transform theory, selects the Daubechies wavelet as the mother wavelet function, decomposing the signal into five sub-signal components with different frequency bands. Then, frequency domain analysis identifies periodic noise components generated by tidal cycles, ocean current variations, and wind speed fluctuations. Low-frequency noise (0.001Hz to 0.01Hz) and high-frequency noise (above 2Hz) are filtered out by setting a frequency threshold range of 0.001Hz to 0.01Hz. For the underwater acoustic signal spectrum data, a noise reduction strategy combining spectral subtraction and Wiener filtering is employed. The background noise power spectral density is first estimated, and then the signal-to-noise ratio is enhanced through adaptive weight adjustment. Wavelet domain-based image denoising was performed on the seabed biological image data. A soft thresholding function was used to suppress image noise while preserving the integrity of the biological contour features. The aim of this step was to improve the signal-to-noise ratio and reliability of the monitoring data, ensuring the accuracy of subsequent feature extraction.

[0063] The specific implementation of step S04 involves constructing a multidimensional ecological feature dataset. First, statistical features are extracted from the filtered water quality parameter data, calculating the mean, variance, skewness, and kurtosis of each parameter to form a water quality statistical feature vector containing 20 feature components. Frequency domain analysis is performed on the underwater acoustic signal spectrum data, converting the time-domain signal to a frequency-domain representation using a fast Fourier transform. Eight acoustic spectrum feature parameters are extracted, including dominant frequency, bandwidth, signal strength, and spectral entropy. The dominant frequency reflects the vocalization characteristics of dominant species, while the bandwidth characterizes the degree of species diversity. A deep convolutional neural network is used for species identification from the seabed biological image data. The network structure includes five convolutional layers and three fully connected layers. End-to-end training enables automatic feature learning and classification, outputting biological species identification results including species category labels and confidence scores. Statistical analysis is performed on the plankton sampling data according to species classification, calculating six community structure parameters, including the Shannon diversity index, Simpson evenness index, and Berger-Parker dominance index, to reflect the compositional characteristics and stability of the plankton community. Finally, the four types of feature data mentioned above are aligned and fused according to time series to construct a multidimensional ecological feature dataset containing 34 feature dimensions. The purpose of this step is to transform the raw monitoring data into a standardized feature representation that can be used for ecological analysis, providing high-quality input data for subsequent correlation modeling.

[0064] The specific implementation of step S05 involves establishing a marine ecological correlation analysis model. First, the model's network architecture is constructed. The input layer receives a 34-dimensional multidimensional ecological feature dataset, which is then processed through three densely connected blocks for feature extraction and fusion. Each densely connected block employs the design concept of dense connections between layers in forward propagation. The input of the current layer includes the output features of all previous layers, reducing parameter redundancy and enhancing gradient propagation through a feature reuse mechanism. Each densely connected block contains 64 neurons. The activation function is a modified linear unit function to enhance nonlinear expression, and a residual connection structure is introduced to alleviate the gradient vanishing problem in deep networks. The network learns the complex mapping relationship between water quality statistical feature vectors and plankton community structure parameters to identify the influence of environmental factors such as water temperature and salinity on plankton community composition. Simultaneously, it learns the correlation between acoustic spectral feature parameters and biological species identification results, establishing a correspondence model between acoustic signal features and species distribution. The output layer generates a 1-dimensional biodiversity assessment index and a 6-dimensional community structure feature parameter vector. The biodiversity assessment index ranges from 0 to 1, with higher values ​​indicating better ecosystem health. The purpose of this step is to use deep learning methods to uncover the inherent correlations between multi-source ecological data, thereby enabling a quantitative assessment of the status of marine biodiversity.

[0065] The specific implementation of step S06 involves constructing an ecological dynamic early warning and prediction system. First, historical time-series data of the biodiversity assessment index and community structure characteristic parameter vector are used as input, with a data time window length set to 90 days to capture seasonal variation patterns. A Long Short-Term Memory (LSTM) network is adopted as the basic architecture of the prediction model, achieving effective learning of long-term dependencies through coordinated control of memory gates, forget gates, and output gates. The network structure contains two LSM layers, each with 128 hidden units, and a random inactivation rate of 0.2 is used between layers to prevent overfitting. A multi-head attention mechanism is introduced to enhance the model's ability to focus on key time steps; the number of attention heads is set to 8, and the diversity of feature expression is improved by parallel computing of attention weights in different subspaces. The prediction results are corrected by incorporating ecological dynamic constraints, including the assumption of continuity of the biodiversity index and the principle of gradual change in community structure. The model outputs predicted values ​​of the biodiversity assessment index and community structure characteristic parameter vector for the next 7 to 30 days, with a prediction confidence interval set at 95%. The purpose of this step is to predict future trends in ecosystem changes through time-series analysis, providing a scientific basis for ecological risk early warning.

[0066] The specific implementation of step S07 involves implementing comprehensive ecological risk regulation and control. First, a multi-indicator joint judgment mechanism is established. When the biodiversity assessment index is below the 0.6 threshold, the community structure deviation index exceeds the 15% threshold, and the water temperature anomaly index is greater than the 2.0 threshold, the comprehensive ecological risk regulation mode is triggered when all three conditions are met simultaneously. The community structure deviation index is obtained by normalizing the Euclidean distance between the current community structure characteristic parameter vector and the historical baseline vector, reflecting the degree of deviation of the community structure from its normal state. The water temperature anomaly index is calculated by the standardized difference between the current seawater temperature value and the historical average for the same period, quantifying the degree of seawater temperature deviation from the normal range. After the regulation mode is activated, the multi-parameter coordinated control algorithm automatically adjusts the monitoring parameter configuration, increasing the data acquisition frequency from once per minute to once every 10 seconds, increasing the plankton sampling density from twice per hour to three times per hour, and reducing various warning threshold parameters by 20% to improve sensitivity. The algorithm employs a probabilistic graphical model framework to establish structured constraints among various indicators. Probabilistic information is propagated between graph nodes via a message passing algorithm. Variational inference techniques are used to approximate the posterior distribution, and Gibbs sampling is employed to generate structured prediction results. Simultaneously, a manual inspection and verification procedure is initiated, dispatching marine biologists to verify anomalies on-site and collect supplementary samples for laboratory analysis. The purpose of this step is to respond rapidly and implement targeted regulatory measures upon detecting ecological anomalies, minimizing the impact of ecological risks on the safe operation of the nuclear power plant.

[0067] Further explanation is needed regarding the detailed structure of the marine ecological association analysis model, which employs a deep residual dense connection architecture, comprising four main components: an input layer, a feature extraction layer, a feature fusion layer, and an output layer. The input layer receives a 34-dimensional multidimensional ecological feature dataset. First, the input features are standardized through a batch normalization layer. Then, an initial feature transformation is performed through a fully connected layer containing 128 neurons. This layer uses a modified linear unit activation function and applies a random deactivation rate of 0.1 to prevent overfitting. The feature extraction layer consists of three cascaded dense connection blocks, each containing four sub-layers. The first sub-layer of the first dense connection block receives the initial feature transformation result and contains 64 neurons. The second sub-layer receives the connection features between the output of the first sub-layer and the original input, increasing the number of neurons to 128. The third sub-layer receives the connection features between the output of the first two layers and the original input, with 192 neurons. The fourth sub-layer receives all the connection features between the output of the first three layers and the original input, reaching 256 neurons. Each sub-layer uses a standard configuration of batch normalization, modified linear unit activation, and a random deactivation rate of 0.2. The second densely connected block uses the same internal structure, but expands the input feature dimension to include all output features from the first block and the original features. A transformation layer compresses the feature dimension to 50% of its original size to control parameter growth. The third densely connected block maintains the same sub-layer structure, ultimately outputting a 1024-dimensional high-level semantic feature representation. The feature fusion layer first converts spatial features into fixed-length vectors using global average pooling, then performs feature dimensionality reduction through a fully connected layer with 512 neurons. This layer uses a modified linear unit activation function and a random inactivation rate of 0.3. The output layer contains two parallel branches: a biodiversity branch connected to one output neuron via a hidden layer with 128 neurons, using a sigmoid activation function to ensure output values ​​are within the range of 0 to 1; and a community structure branch connected to six output neurons via a hidden layer with 256 neurons, using a linear activation function to maintain the continuous numerical characteristics of the parameters. The loss function employs a multi-task learning framework, combining mean squared error loss and binary cross-entropy loss. The total loss function is a weighted sum of the losses from the two branches, with weight coefficients of 0.4 and 0.6, respectively. The training dataset was established through six detailed stages: data collection, quality control, feature engineering, expert annotation, data augmentation, and ensemble partitioning. The data collection stage involved selecting valid samples from 36 months of continuous monitoring data of the sea area surrounding the Binhai Nuclear Power Plant. This included complete monitoring records covering all four seasons (spring, summer, autumn, and winter), various weather conditions (sunny, rainy, snowy, foggy), tidal changes (spring and neap tides), and sea conditions (calm and stormy). The original dataset comprised 5.2 million time-series records.In the quality control phase, strict data screening standards were established to eliminate invalid data from sensor calibration periods, equipment maintenance periods, and periods under abnormal weather conditions. Missing data segments were identified through time series continuity testing. Data with missing lengths less than 30 minutes were filled using cubic spline interpolation, while time periods with missing lengths exceeding 30 minutes were entirely discarded. After quality control, 3.8 million sets of valid data were retained. In the feature engineering phase, statistical features were extracted from the time-series data of each monitoring station according to a 24-hour time window. Eight statistical indicators were calculated, including mean, median, standard deviation, skewness, kurtosis, maximum value, minimum value, and trend. Simultaneously, Fourier coefficients of periodic features, including daily and tidal cycles, were extracted to form a comprehensive feature vector for each station daily. The expert annotation phase involved a team of five professionals with over 10 years of experience in marine ecological research. They annotated feature vectors using real ecological data obtained through various methods, including on-site diving surveys, benthic organism sampling, plankton microscopic analysis, and fish acoustic identification. The annotations included six community structure parameters (Shannon diversity index, Simpson evenness index, species richness, community stability index, etc.) and a comprehensive biodiversity assessment index. Annotation consistency was measured using a Pearson correlation coefficient based on multi-expert scoring, achieving a correlation coefficient above 0.92. The data augmentation phase employed multiple strategies to expand the training samples. These included adding 5% random noise (following a normal distribution) to the water quality feature vectors to simulate sensor measurement errors, randomly pruning time-series data to generate samples at different time scales, and using synthetic minority oversampling techniques to balance the data distribution for rare ecological state samples. Ultimately, the number of effective samples was expanded to 180,000. In the dataset partitioning phase, a stratified random sampling method was used to ensure that different ecological state categories maintained the same distribution ratio in the training and validation sets. The datasets were divided into 150,000 training samples and 30,000 validation samples at an 8:2 ratio, with an additional 10,000 samples reserved as the final test set to evaluate the model's generalization performance. The model training process employed the Adam adaptive optimization algorithm, with an initial learning rate set to 0.001, using a cosine annealing learning rate scheduling strategy, and a minimum learning rate of [missing value]. The batch size was set to 128, and the total training epochs were 200. An early stopping strategy was applied during training to monitor the validation set loss; training was stopped when the validation loss no longer decreased after 15 consecutive epochs to prevent overfitting. To improve model robustness, an adversarial example generation mechanism was introduced during training. Adversarial examples were generated using the fast gradient sign method and added to the training batches, with the adversarial example ratio set to 20%. Simultaneously, label smoothing technology was used to convert hard labels to a soft label distribution, with a smoothing parameter set to 0.1 to reduce the model's sensitivity to mislabeling. Training stability was ensured through gradient clipping and weight decay techniques, with the gradient clipping threshold set to 1.0 and the weight decay coefficient set to... .

[0068] Further explanation is needed regarding the ecological dynamic early warning model, which employs a hybrid temporal prediction architecture based on long short-term memory (LSTM) networks and multi-head attention mechanisms. The network input consists of 90 days of historical time-series data, including time series of biodiversity assessment indices and 6-dimensional community structure feature parameter vectors. The input data undergoes feature transformation first through a temporal embedding layer, which includes a positional encoding mechanism to preserve temporal information; the embedding dimension is set to 256. The temporal encoder comprises two LTM layers. The first layer contains 128 hidden units and employs a bidirectional structure to simultaneously learn forward and backward temporal dependencies, with an output dimension of 256. The forgetting gate of the LTM units controls the degree of historical information retention through a sigmoid activation function, with the forgetting bias initialized to 1.0 to mitigate the gradient vanishing problem. The input gate and candidate value gate control the storage of new information and the generation of candidate states, respectively. The output gate determines the information content to be output to the next layer at the current moment. The second layer, Long Short-Term Memory (LSTM), also contains 128 hidden units. It receives the output of the first layer as input and further extracts high-level temporal feature representations. A random inactivation rate of 0.3 is used between layers to prevent overfitting, ultimately outputting a 256-dimensional temporal feature vector. The multi-head attention mechanism comprises eight parallel attention heads, each with a 32-dimensional query, key-value, and numerical matrix. It learns dependencies between different positions in the temporal sequence through a scaled dot-product attention mechanism. Attention weights are calculated using softmax normalization to ensure a sum of weights equal to 1. The attention mechanism determines the degree of attention given to different historical moments by learning the similarity between the query vector and the key-value vector; moments with higher similarity receive greater attention weights. The multi-head mechanism allows the model to learn different types of dependencies in different representation subspaces. The outputs of the eight attention heads are fused through a linear transformation to form a 256-dimensional attention feature representation. The temporal feature vector and attention feature representation are fused through residual connections and layer normalization to avoid gradient problems in deep networks and accelerate the convergence process. The prediction layer comprises three fully connected layers. The first fully connected layer contains 512 neurons and uses a modified linear unit activation function (MRU). The second fully connected layer contains 256 neurons and also uses the MRU. The third fully connected layer directly outputs a 7-dimensional prediction result, including a 1-dimensional biodiversity assessment index prediction and a 6-dimensional community structure feature parameter vector prediction. The output layer does not use an activation function to preserve the numerical characteristics of the prediction values. The loss function uses a weighted combination of mean squared error loss and mean absolute error loss with a weight ratio of 7:3. A temporal consistency constraint loss term is also introduced. This constraint term maintains the temporal smoothness of the prediction results by calculating the mean squared error between the first difference of the predicted sequence and the first difference of the true sequence. The constraint weight coefficient is set to 0.1.The training dataset was constructed using a sliding window method, generating time-series sample pairs from two years of historical monitoring data. The input window size was set to 90 days, the output prediction window size to 30 days, and the sliding step size to 1 day to ensure full utilization of historical data. The original monitoring data contained 730 days of continuous records. After sliding window processing, 610 sets of basic training samples were generated. Each set of samples contained an input sequence of 90 time steps and a target sequence of 30 time steps. During data preprocessing, missing values ​​were handled in the time-series data. Forward imputation was used to handle missing segments shorter than 4 hours, and linear interpolation was used to handle missing segments between 4 and 12 hours. Samples corresponding to missing segments longer than 12 hours were directly discarded. Time-series standardization employed a sliding window standardization method, applying zero-mean, unit variance standardization to the data within each 90-day input window to avoid the impact of data distribution differences across different periods on model training. Data augmentation strategies included time-series jittering, noise injection, and time warping. Temporal jitter is achieved by randomly selecting 10% of the time steps in the input sequence and making small offsets, with the offset amplitude controlled within 2% of the original value, to simulate time synchronization errors in real monitoring. Noise injection adds Gaussian white noise to the input features, with the noise standard deviation set to 5% of the feature standard deviation, improving the model's robustness to measurement noise. Time warping simulates different temporal dynamics by performing nonlinear time axis transformations on the input sequence, with the warping intensity controlled within 10% of the original time axis. After data augmentation, the number of training samples is expanded to 3050 groups, divided into a training set of 2440 groups, a validation set of 305 groups, and a test set of 305 groups in an 8:1:1 ratio. Model training uses the AdamW optimization algorithm, which combines the adaptive learning rate characteristics of the Adam optimizer with the weight decay regularization effect. The initial learning rate is set to 0.0005, and the weight decay coefficient is set to 0.01. A warm-up strategy is used for learning rate scheduling, with the learning rate increasing linearly for the first 20 training rounds. The learning rate was increased to 0.0005, and then a cosine annealing strategy was used to gradually reduce the learning rate to... The batch size was set to 32 to balance training efficiency and memory usage, and the total number of training epochs was set to 300. An early stopping strategy was employed during training to monitor the mean squared error loss on the validation set; training was stopped early when the validation loss no longer decreased after 20 consecutive epochs. To improve the ability to quantify prediction uncertainty, Bayesian neural network technology was introduced during training. Variational inference was used to learn the probability distribution of network weights rather than deterministic values, enabling the model to output confidence intervals for prediction results. Training stability was ensured through gradient pruning, with the gradient norm threshold set to 2.0 to prevent gradient explosion. An exponential moving average technique was used to smoothly update model parameters, with a smoothing coefficient set to 0.999.

[0069] It should be noted that the key technical ideas of this invention include four aspects: multi-sensor array collaborative monitoring, multi-scale adaptive filtering, deep learning correlation modeling, and ecological dynamic early warning and prediction. Multi-sensor array collaborative monitoring technology integrates various monitoring methods such as water quality sensors, acoustic monitoring equipment, underwater cameras, and plankton samplers to achieve comprehensive and three-dimensional observation of the marine ecosystem. Compared with traditional single-sensor monitoring methods, it can obtain richer and more comprehensive ecological information, effectively compensating for the limitations of single monitoring methods and significantly improving the accuracy and reliability of ecological status assessment. Multi-scale adaptive filtering technology, based on wavelet transform theory, decomposes the monitoring signal into different frequency band components through multi-level decomposition, specifically filtering out noise interference caused by environmental factors such as tides, currents, and wind speed. Compared with traditional fixed-parameter filtering methods, it has stronger adaptive capabilities, dynamically adjusting filtering parameters according to signal characteristics, suppressing noise to the greatest extent while maintaining the integrity of useful signals, and significantly improving data quality and feature extraction accuracy. Deep learning-based correlation modeling technology employs a densely connected network architecture to learn the complex nonlinear mapping relationships between multi-source ecological data. Through feature reuse and gradient enhancement mechanisms, it improves the model's expressive power and training stability. Compared to traditional statistical analysis methods, it can uncover deeper ecological correlation patterns, accurately quantify the impact mechanisms of environmental factors on biological community structure, and achieve precise assessment of biodiversity status. Ecological dynamic early warning and prediction technology combines long short-term memory networks and multi-head attention mechanisms. By learning long-term dependencies and key time patterns in historical time-series data, it accurately predicts future ecosystem change trends. Compared to traditional experience-based judgment methods, it has stronger scientific rigor and foresight, providing effective early warnings before ecological anomalies occur, thus gaining valuable time for the timely implementation of ecological protection measures. The synergistic effect of these four technological approaches forms a complete ecological monitoring and early warning system. It achieves technological breakthroughs in all aspects from data acquisition, processing, and analysis to prediction. Compared to existing technologies, it has comprehensive advantages in terms of wide monitoring coverage, high data processing accuracy, strong correlation analysis capabilities, and timely early warning response, providing comprehensive and reliable technical support for the ecological environmental protection of coastal nuclear power plants.

[0070] It should be noted that existing marine ecological monitoring technologies mainly rely on single-type sensors for data collection, lacking the comprehensive perception capability of multi-dimensional ecological factors such as water quality chemical parameters, bioacoustic characteristics, seabed image information, and planktonic community structure. This results in an inability to accurately capture the complex response mechanisms of environmental changes to biological communities. This invention, by constructing an integrated multi-sensor array acquisition station, achieves synchronous high-frequency monitoring of multi-level information of the marine ecosystem. It utilizes the densely connected network structure of a marine ecological correlation analysis model to establish a nonlinear mapping relationship between environmental parameters and biological community structure, enabling real-time quantitative analysis of the impact of changes in key environmental factors such as temperature, salinity, and dissolved oxygen on biological indicators such as species richness and community stability. By outputting standardized biodiversity assessment indices and community structure characteristic parameter vectors, it achieves quantitative characterization of multi-dimensional environment-biological response relationships, providing a scientific basis for assessing the health status of marine ecosystems. Furthermore, traditional ecological monitoring systems typically employ static threshold judgments and manual intervention for risk identification and response, failing to predict future ecological states based on historical data trends and lacking intelligent control mechanisms for automatically adjusting monitoring strategies based on prediction results. This invention introduces an ecological dynamic early warning model, employing a long short-term memory network architecture combined with a multi-head attention mechanism to analyze historical temporal variation patterns of biodiversity assessment indices and community structure characteristic parameters. This enables accurate prediction of ecological change trends over 7 to 30 days, achieving a technological shift from passive monitoring to proactive early warning. When the system detects a biodiversity assessment index less than 0.6, a community structure deviation index greater than 15%, and a water temperature anomaly index greater than 2.0, it automatically activates an ecological risk comprehensive regulation mode. Through a multi-parameter coordinated control algorithm, it dynamically adjusts key parameters such as data acquisition frequency, sampling density, and early warning thresholds, achieving adaptive optimization control of the monitoring system and significantly improving the accuracy and response speed of ecological risk identification.

[0071] Specifically, the principle of this invention is as follows: The technical solution of this invention can solve the technical problem of multi-source heterogeneity and insufficient spatiotemporal correlation analysis capability of ecological monitoring data from coastal nuclear power plants, mainly based on the following principles and mechanisms. First, by deploying a multi-sensor array acquisition station network covering a radius of 5000 meters in the sea area surrounding the nuclear power plant, each station integrates water quality parameter sensors, acoustic biological monitoring equipment, underwater camera devices, and plankton samplers, synchronous high-frequency acquisition of multi-dimensional information of the marine ecosystem is achieved, providing a rich data foundation for subsequent correlation analysis. Second, a multi-scale adaptive filtering algorithm based on wavelet transform technology is used to suppress noise in the acquired multi-source data. Interference signals caused by environmental factors such as tides, ocean currents, and wind speed are identified and filtered out through frequency domain decomposition, significantly improving data quality and reliability. Then, through statistical feature extraction, frequency domain analysis, image recognition, and species classification statistics, the original monitoring data is converted into water quality statistical feature vectors, acoustic spectrum feature parameters, biological species identification results, and plankton community structure parameters, constructing a 34-dimensional multi-dimensional ecological feature dataset, laying a standardized data structure foundation for correlation analysis. The core marine ecological correlation analysis model employs a densely connected network architecture. Through feature extraction and fusion processing of three densely connected blocks, it learns the complex nonlinear mapping relationship between water quality parameters and biological community structure, outputting a quantified biodiversity assessment index and community structure characteristic parameter vector, thus achieving deep correlation analysis of multi-source heterogeneous data. Finally, combined with the long short-term memory network architecture and multi-head attention mechanism of the ecological dynamic early warning model, it predicts future ecological change trends through historical time-series data analysis. When the biodiversity assessment index, community structure deviation index, and water temperature anomaly index exceed preset thresholds, the model automatically initiates an ecological risk comprehensive regulation mode to adaptively optimize and adjust monitoring parameters, achieving intelligent ecological risk management.

[0072] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0073] In this embodiment, the specific implementation methods of steps S01-S02 and S05-S06 are the same as those described above, and will not be repeated in detail here.

[0074] The specific implementation of step S03 involves multi-level noise suppression processing, using a multi-scale adaptive filtering algorithm to process the seawater temperature, salinity, dissolved oxygen concentration, turbidity, and chlorophyll a concentration data. The mathematical expression for the wavelet decomposition process is:

[0075] ;

[0076] In the formula, Wavelet coefficients, dimensions relative to the input signal same; For input signals; The mother wavelet function is dimensionless. The scale parameter is dimensionless and its empirical value is an integer power of 2. ; For time shift parameters, the unit is the time variable. same; Indicates complex conjugation; Let be the time variable, in seconds (s). The filtering expression for signal reconstruction is:

[0077] ;

[0078] In the formula, This is the filtered signal; For the first Layer Wavelet coefficients, For decomposition layer index, , This is the index of the coefficient position in this layer; For the first Layer Each scaling factor; For the first Layer Wavelet functions; For the first Layer One scaling function; The decomposition level is typically set to 5. The frequency threshold for noise component identification is expressed as follows:

[0079] ;

[0080] In the formula, This is the frequency domain filter response, which is dimensionless. Frequency, in Hz; This is the low-frequency cutoff frequency, with a default value of 0.001Hz; This is the high-frequency cutoff frequency, which is 2Hz by default.

[0081] The specific implementation of step S04 involves constructing a multidimensional ecological feature dataset and extracting statistical features from the filtered water quality parameter data. The expression for calculating the water quality statistical feature vector is as follows:

[0082] ;

[0083] ;

[0084] ;

[0085] ;

[0086] In the formula, For the first The average of each water quality parameter, For water quality parameter index, These correspond to seawater temperature, salinity, dissolved oxygen concentration, turbidity, and chlorophyll a concentration, respectively. For the first The variance of each water quality parameter; It is a skewness, dimensionless; Kuroshi is dimensionless; For the first The parameter of the first Each sample value, For sampling sequence number, ; This represents the total number of samples. The expression for calculating the spectral entropy, a key characteristic parameter of the acoustic spectrum, is:

[0087] ;

[0088] In the formula, The spectral entropy is dimensionless. For the first Normalized power spectral density of each frequency component For frequency component index, , ; The total number of frequency components. The formula for calculating the Shannon diversity index, a parameter of planktonic community structure, is:

[0089] ;

[0090] In the formula, The Shannon diversity index is dimensionless. For the first The relative abundance of each species, For species indexing, , ,in For the first The number of individuals of a species The total number of individuals of all species. ; This represents the total number of species. The Simpson evenness index is calculated as follows:

[0091] ;

[0092] In the formula, is the Simpson uniformity index, which is dimensionless; The meaning is the same as in the Shannon index; The meaning is the same as in the Shannon index.

[0093] The specific implementation method of step S07 is to implement comprehensive ecological risk regulation and control, and to activate the regulation mode when the detection indicators meet the preset conditions. The calculation expression for the community structure deviation index is:

[0094] ;

[0095] In the formula, is a dimensionless index representing the deviation of community structure. This is the current community structure feature parameter vector; For historical baseline vectors; It is the Euclidean norm. The formula for calculating the water temperature anomaly index is:

[0096] ;

[0097] In the formula, This is a dimensionless index representing an abnormal water temperature. The current seawater temperature is expressed in °C. The average seawater temperature for the same period in history, in °C; represents the standard deviation of seawater temperature for the same historical period, in °C. The node probability update expression for the probabilistic graphical model in the multi-parameter coordinated control algorithm is:

[0098] ;

[0099] In the formula, For nodes Given parent node The conditional probability is dimensionless. For nodes and Potential function between; For nodes The set of parent nodes; For nodes The possible values ​​of ; For node indexing.

[0100] It should be explained that the specific calculation of the community structure deviation index requires obtaining a historical baseline vector. This vector is obtained as follows: Community structure monitoring data for 12 consecutive months during the normal operation of the nuclear power plant is selected, data from periods of abnormal weather and equipment failure are excluded, the monthly average value of each community structure characteristic parameter is calculated, forming a 12-dimensional monthly baseline feature vector, which is then smoothed over time to obtain a 6-dimensional historical baseline vector. The current community structure characteristic parameter vector is obtained from real-time monitoring data using the same statistical method, including six components: species richness, evenness index, dominance index, diversity index, community stability index, and biomass index.

[0101] The calculation of the water temperature anomaly index requires the establishment of a historical temperature database for the same period. This database contains seawater temperature observation records for the same period over the past three years, grouped and statistically analyzed by month and tidal cycle. The historical average seawater temperature for the same period is also required. The standard deviation of seawater temperature for the corresponding historical period was obtained by averaging all temperature observations for that period. The current seawater temperature is calculated using the sample standard deviation of temperature observations for the corresponding period. The instantaneous temperature value is obtained for real-time monitoring.

[0102] The potential function in the multi-parameter coordinated control algorithm Using the Gaussian radial basis function form, the specific expression is as follows:

[0103] ;

[0104] In the formula, The connection weights are dimensionless and have empirical values ​​ranging from 0.5 to 2.0. This is the width parameter of the radial basis function, with the same units as the characteristic dimension, and is typically taken as 0.1 to 0.5 times the characteristic standard deviation; is the Euclidean distance between the node feature vectors.

[0105] The mathematical expression for wavelet decomposition is based on the theory of continuous wavelet transform, using the scale parameter. and time shift parameters The adjustment enables signal analysis in different time and frequency domains, where the calculation of wavelet coefficients follows... The integral transform form.

[0106] The mother wavelet function, the Daubechies wavelet, ensures perfect signal reconstruction. This formula effectively separates periodic and random noise, exhibiting stronger adaptability and higher signal-to-noise ratio improvement compared to traditional fixed-parameter filtering methods. The filtering expression for signal reconstruction achieves noise suppression through selective preservation and reconstruction of wavelet coefficients. The reconstruction process follows... The linear combination form.

[0107] Different decomposition levels correspond to different frequency resolutions; lower levels correspond to high-frequency details, while higher levels correspond to low-frequency trends. Thresholding removes noise components while preserving useful signal features, significantly improving the accuracy of subsequent feature extraction. The calculation expression for the water quality statistical feature vector is based on probability and statistics theory, where the mean is calculated using... The arithmetic mean form.

[0108] The mean reflects the central tendency of the parameter, and the variance is expressed as a function of the mean. The degree of dispersion of the quantified data is skewness. Kurtosis describes the asymmetry of the distribution. Characterizes the sharpness of the distribution.

[0109] These four statistics can comprehensively describe the distribution characteristics of water quality parameters, and compared with the simple mean statistical method, they can capture more of the inherent patterns in the data. The Shannon diversity index calculation formula is based on information theory and follows... The form of information entropy calculation.

[0110] The community diversity index quantifies the degree of community diversity by taking the logarithmic weighted sum of relative species abundance. This index considers both species quantity and evenness, and can more accurately reflect the health of the ecosystem compared to simple species counting methods. The calculation expression of the community structure deviation index is based on vector space theory and employs... The normalized Euclidean distance form.

[0111] The deviation of the current state from the normal state is quantified by normalizing the Euclidean distance. This index has dimensionless properties, facilitating comprehensive comparison of different parameters, and can achieve multi-dimensional comprehensive evaluation compared to single-parameter anomaly detection methods. The calculation expression of the water temperature anomaly index is based on standardized statistical theory and employs... The Z-score standardized form.

[0112] Standardization is achieved by dividing the difference between the current value and the historical mean by the historical standard deviation. This index eliminates the influence of seasonal variations and can more accurately identify true temperature anomalies compared to absolute temperature comparison methods. The node probability update expression of the probabilistic graphical model is based on Markov random field theory and follows... The Gibbs distribution.

[0113] By modeling the probability distribution through the exponential form of the potential function and the normalization of the partition function, this model can capture the interdependencies among multiple monitoring indicators and achieves more coordinated system optimization compared to independent parameter adjustment methods. The Gaussian radial basis function form of the potential function is based on kernel function theory and employs... The Gaussian kernel form is used. The similarity between nodes is quantified by the exponential decay of distance. The local properties of the radial basis function make similar states more correlated, and this functional form can better describe nonlinear parameter dependencies compared to the linear correlation model.

[0114] To better understand and implement this invention, the following is a specific application scenario example 2: A technical team conducted biodiversity monitoring work in the waters surrounding a coastal nuclear power plant, employing multi-sensor array integrated monitoring technology to comprehensively monitor the marine ecosystem. The team deployed 16 multi-sensor array acquisition stations at 1200-meter intervals within a 5000-meter radius around the nuclear power plant, forming a three-dimensional monitoring network covering the entire monitoring area. Each multi-sensor array acquisition station was equipped with a CTD-type water quality parameter sensor, broadband acoustic biological monitoring equipment, a 4K high-resolution underwater camera, and an automatic plankton sampler, ensuring continuous real-time monitoring of the marine environment and biological conditions.

[0115] The technical team collected key water quality data, including seawater temperature, salinity, dissolved oxygen concentration, turbidity, and chlorophyll a concentration, using water quality parameter sensors. During implementation, the seawater temperature monitoring range was 8.2–28.6℃, the salinity monitoring range was 32.1–35.8℃, and the dissolved oxygen concentration monitoring range was 5.8–9.4℃. The turbidity monitoring range was 0.5–12.3 NTU, and the chlorophyll a concentration monitoring range was 1.2–15.8 NTU. The acoustic biological monitoring equipment collects data in the frequency range of 10Hz to 50Hz. The underwater acoustic signal spectrum data can identify bioacoustic features such as the swimming sounds of fish, the breathing sounds of marine mammals, and the feeding sounds of crustaceans. The underwater camera continuously captures images of seabed biological activity, recording the habitat behavior and community distribution of benthic organisms. The plankton sampler collects phytoplankton and zooplankton samples of different sizes through an automatic filtration system, and its equipped microscope imaging system enables automatic species identification and counting.

[0116] like Figure 2 As shown, the technical team preprocessed the collected multi-source heterogeneous data using a multi-scale adaptive filtering algorithm. This algorithm, based on wavelet transform multi-level decomposition technology, decomposes the seawater temperature signal into eight frequency band components. By analyzing the energy distribution characteristics of each frequency band component, it identifies and filters out the 0.08 ohm coefficient caused by tidal periodic variations. Low-frequency noise components and 2.5% caused by wind and wave disturbances. The above high-frequency noise components. After similar decomposition and filtering processing, the signal-to-noise ratios of salinity and dissolved oxygen concentration signals were improved to 23.6 dB and 27.2 dB, respectively. Turbidity and chlorophyll a concentration data were processed using adaptive threshold denoising technology to eliminate the effects of sensor drift and environmental interference.

[0117] After filtering and preprocessing, the data enters the feature extraction stage. The technical team performs statistical analysis on the water quality parameter data, calculating statistical quantities such as mean, variance, skewness, and kurtosis for each parameter, and constructs a 20-dimensional water quality statistical feature vector. Frequency domain analysis is performed on the underwater acoustic signal spectrum data, extracting 8-dimensional acoustic spectral feature parameters such as dominant frequency, bandwidth, signal strength, and spectral entropy. Seabed biological image data collected by underwater cameras is processed using deep learning image recognition algorithms to identify 18 major species categories, including flatfish, shrimp and crabs, echinoderms, and mollusks, and the population distribution of each species is statistically analyzed. After species classification and statistical analysis, 6-dimensional plankton community structure parameters, including species richness, evenness index, dominance index, and diversity index, are calculated from the plankton sampling data.

[0118] The technical team combined extracted water quality statistical feature vectors, acoustic spectral feature parameters, biological species identification results, and plankton community structure parameters according to time series to construct a 34-dimensional multidimensional ecological feature dataset. This dataset covers comprehensive information from multiple ecological dimensions, including the physicochemical properties of the marine environment, bioacoustic activity characteristics, benthic community structure, and plankton community dynamics. Figure 3 and Figure 4 As shown, the multidimensional ecological feature dataset, after standardization, is input into the marine ecological association analysis model for nonlinear relationship modeling.

[0119] The marine ecological correlation analysis model employs a deep neural network architecture. The input layer receives 34-dimensional feature data, which is extracted and fused through three densely connected blocks. Each densely connected block contains 64 neurons and residual connection structures, enabling it to learn complex nonlinear mapping relationships between different ecological elements. The model learns the ecological response relationship between water quality statistical feature vectors and plankton community structure parameters through the densely connected layers, establishing a correlation model between acoustic spectral feature parameters and species identification results. The output layer generates a 1-dimensional biodiversity assessment index and a 6-dimensional community structure feature parameter vector, quantitatively assessing the health status of the marine ecosystem.

[0120] As shown in Table 1, the technical team collected three years of historical monitoring data as the basic dataset during the model training process.

[0121] Table 1. Composition of the training dataset for the marine ecological association analysis model

[0122] After data cleaning and standardization, the training dataset was updated using the Adam optimizer with a learning rate of 0.001, a batch size of 128, and 200 training epochs. The training process employed a data augmentation-based sample diversity enhancement mechanism, increasing the diversity of training samples by adding 5% Gaussian noise to the water quality statistical feature vector and randomly pruning time-series data. Simultaneously, a gradient-based training stability enhancement mechanism was used, adding [a certain parameter] to the loss function. Regularization term stabilizes training dynamics.

[0123] The trained marine ecological association analysis model achieved a 92.8% accuracy in predicting the biodiversity assessment index of the test data, with a mean absolute error of 0.063 for predicting the community structure characteristic parameter vector. The biodiversity assessment index output by the model ranges from 0 to 1, with values ​​closer to 1 indicating better biodiversity. In actual monitoring, the typical value of this index is 0.72–0.89, reflecting a good level of ecological health in the monitored sea area.

[0124] The technical team further inputs time-series data of biodiversity assessment indices and community structure characteristic parameter vectors into an ecological dynamic early warning model for trend prediction analysis. This early warning model employs a long short-term memory network architecture combined with a multi-head attention mechanism, enabling it to capture both long-term dependencies and short-term fluctuations in ecosystem changes. The model analyzes the changing patterns of historical time-series data to predict the trends in biodiversity assessment indices and the direction of community structure succession over the next 7 to 30 days.

[0125] In practical applications, the technical team set trigger conditions for the integrated ecological risk regulation mode. When the monitoring system detects that the biodiversity assessment index drops to 0.58, the community structure deviation index rises to 18.5%, and the water temperature anomaly index reaches 2.3, the system automatically activates the integrated ecological risk regulation mode. At this time, the data acquisition frequency of the multi-sensor array acquisition stations is adjusted from once per minute to once every 10 seconds, the sampling density parameter is increased from twice per hour to three times per hour, and the warning threshold parameter is correspondingly reduced by 20%.

[0126] As shown in Table 2, the adjustment of various monitoring parameters after the activation of the integrated ecological risk regulation mode reflects the system's adaptive response capability.

[0127] Table 2 Comparison of Parameter Adjustments for Integrated Ecological Risk Regulation Model

[0128] The technical team employed a multi-parameter coordinated control algorithm to optimize the monitoring strategy under the risk adjustment mode. This algorithm utilizes a structured prediction framework based on a probabilistic graphical model, modeling the constraints between biodiversity assessment indices, community structure deviation indices, and water temperature anomaly indices as edge connections in a graph. Probabilistic information is propagated between graph nodes via a message-passing algorithm, variational inference techniques are used to approximate the posterior distribution, and Gibbs sampling is used to generate structured prediction results. This coordinated control mechanism ensures logical consistency and response coordination among different monitoring parameters.

[0129] In practical monitoring, the community structure deviation index is obtained by normalizing the Euclidean distance between the current community structure characteristic parameter vector and the historical baseline vector. Current monitoring data shows that diatoms are the dominant group in the phytoplankton community in spring, dinoflagellates become dominant in summer, and a diversified distribution pattern emerges in autumn. In the benthic community, the abundance index of mollusks is 0.42, that of crustaceans is 0.36, and that of echinoderms is 0.22. The water temperature anomaly index is calculated by standardizing the difference between the current seawater temperature value and the historical average for the same period. The average water temperature anomaly index is 0.8 in spring, rises to 1.4 in summer, and falls back to 0.6 in autumn.

[0130] After six months of continuous monitoring, this technical approach successfully captured three significant ecosystem change events. The first event occurred during the spring algal bloom, with chlorophyll a concentration peaking at 24.6%. The biodiversity assessment index dropped to 0.55 in a short period, prompting the system to promptly activate its early warning mechanism and adjust its monitoring strategy. The second event occurred during the summer heatwave, with the water temperature anomaly index exceeding 2.0 for 15 consecutive days, leading to significant changes in the benthic community structure. The system's predictive model accurately predicted the succession trend of the community structure. The third event occurred during an autumn typhoon, with turbidity surging to 45.8 NTU and significant changes in the acoustic signal spectrum characteristics. The multi-parameter coordinated control algorithm effectively coordinated the response strategies of different sensors.

[0131] This invention represents a significant technological advancement over traditional marine biodiversity monitoring methods. Traditional methods rely primarily on periodic sampling and manual analysis, resulting in low monitoring frequency, limited coverage, and delayed response, making it difficult to capture rapid changes in ecosystems. This invention achieves real-time, continuous observation of the marine environment and biological communities through collaborative monitoring using a multi-sensor array, overcoming the limitations of traditional methods in terms of spatiotemporal resolution. A multi-scale adaptive filtering algorithm effectively eliminates complex noise interference in the marine environment, improving data quality and monitoring accuracy. A marine ecological correlation analysis model, employing deep learning technology, establishes a nonlinear mapping relationship between water quality and biological communities, revealing ecological response mechanisms that are difficult to detect using traditional statistical methods. An ecological dynamic early warning model, combining long short-term memory networks and attention mechanisms, enables forward-looking predictions of ecosystem change trends, providing a scientific basis for early warning and timely intervention of ecological risks. An integrated ecological risk regulation model, through a multi-parameter coordinated control algorithm, achieves adaptive optimization of monitoring strategies, dynamically adjusting monitoring parameters according to ecological risk levels, improving the response efficiency of the monitoring system and the rationality of resource allocation.

[0132] It should be noted that the variables involved in this invention are explained in detail in Table 3 below.

[0133] Table 3. Variable Explanation Table

[0134] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for monitoring biodiversity in coastal nuclear power plants, characterized in that, This includes deploying multi-sensor array data collection stations in the coastal area surrounding the nuclear power plant. Each station is equipped with water quality parameter sensors, acoustic biological monitoring equipment, underwater cameras, and plankton samplers, covering the surrounding sea area. The stations will collect real-time data on seawater temperature, salinity, dissolved oxygen concentration, turbidity, and chlorophyll a concentration; underwater acoustic signal spectrum data; seabed biological image data; and plankton density data. Simultaneously, time-series data on tidal cycle parameters, wind speed, and ocean current velocity parameters will be recorded. The collected data will be processed using a multi-scale adaptive filtering algorithm for noise suppression. Statistical feature extraction will yield water quality statistical feature vectors; frequency domain analysis will yield acoustic spectrum feature parameters; image recognition processing will yield biological species identification results; and species classification statistics will yield plankton community structure parameters. These parameters will be combined to construct a multi-dimensional ecological feature dataset. The multidimensional ecological characteristic dataset is input into the marine ecological correlation analysis model to model nonlinear relationships, and outputs the biodiversity assessment index and community structure characteristic parameter vector. The time series data of the biodiversity assessment index and community structure characteristic parameter vector are input into the ecological dynamic early warning model for trend prediction analysis. When the biodiversity assessment index, community structure deviation index and water temperature anomaly index meet the preset conditions, the ecological risk comprehensive regulation mode is activated.

2. The method for monitoring biodiversity in coastal nuclear power plants according to claim 1, characterized in that, The distance between the multi-sensor array data collection stations ranges from 500m to 2000m, covering a sea area with a radius of 5000m around the nuclear power plant, with a data collection frequency of once per minute.

3. The method for monitoring biodiversity in coastal nuclear power plants according to claim 2, characterized in that, The steps of the multi-scale adaptive filtering algorithm are as follows: the seawater temperature signal, salinity signal, and dissolved oxygen concentration signal are decomposed into different frequency band components by wavelet decomposition, and low-frequency noise components and high-frequency noise components caused by tidal period parameters, ocean current velocity parameters, and wind speed values ​​are identified and filtered out.

4. The method for monitoring biodiversity in coastal nuclear power plants according to claim 3, characterized in that, The water quality statistical feature vector is specifically a 20-dimensional feature vector obtained by performing statistical analysis of the mean, variance, skewness, and kurtosis of seawater temperature, salinity, dissolved oxygen concentration, turbidity, and chlorophyll a concentration data.

5. The method for monitoring biodiversity in coastal nuclear power plants according to claim 4, characterized in that, The acoustic spectrum characteristic parameters are specifically eight-dimensional characteristic parameters extracted from the underwater acoustic signal spectrum data through frequency domain analysis, including the main frequency, bandwidth, signal strength, and spectral entropy.

6. The method for monitoring biodiversity in coastal nuclear power plants according to claim 5, characterized in that, The biological species identification results are specifically obtained by classifying marine biological image data using deep learning image recognition algorithms, resulting in species category labels and individual quantity statistics.

7. The method for monitoring biodiversity in coastal nuclear power plants according to claim 6, characterized in that, The so-called planktonic community structure parameters are specifically a six-dimensional parameter vector of species richness, evenness index, dominance index, and diversity index obtained by statistically analyzing planktonic density data according to species classification.

8. The method for monitoring biodiversity in coastal nuclear power plants according to claim 7, characterized in that, The multidimensional ecological feature dataset is specifically a 34-dimensional feature data set formed by combining water quality statistical feature vectors, acoustic spectrum feature parameters, biological species identification results, and plankton community structure parameters in a time series manner.

9. The method for monitoring biodiversity in coastal nuclear power plants according to claim 8, characterized in that, The structure of the marine ecological association analysis model is as follows: the input layer receives a 34-dimensional multidimensional ecological feature dataset, and performs feature extraction and feature fusion through three dense connection blocks. Each dense connection block contains 64 neurons and residual connection structures. The output layer generates a 1-dimensional biodiversity assessment index and a 6-dimensional community structure feature parameter vector.

10. The method for monitoring biodiversity in coastal nuclear power plants according to claim 9, characterized in that, The ecological dynamic early warning model specifically adopts a long short-term memory network architecture combined with a multi-head attention mechanism. It predicts future trends by analyzing historical time-series data of biodiversity assessment index and community structure characteristic parameter vectors, with a prediction period of 7 to 30 days.

Images