An online identification and management system for planktonic species at nuclear power plant water intakes

Abstract

This invention relates to the field of nuclear power plant environmental monitoring technology and discloses an online identification and management system for planktonic species at nuclear power plant water intakes. The system acquires continuously collected multispectral underwater image sequences and aquatic environmental parameters, and generates multidimensional planktonic feature vectors through feature extraction. It constructs a spatiotemporal correlation graph structure using timestamps and spatial location information, employs incremental prototype network dynamic clustering training, and iteratively identifies and activates prototype nodes. The system determines fuzzy prototype nodes by analyzing the intersection of feature sets and calculates an adaptive learning rate to update the feature representation. Simultaneously, it identifies abnormal migration feature points and calculates the abnormal probability. This probability is then used to perform contextual semantic disambiguation on the feature vector to obtain an optimized feature representation. Finally, the optimized representation is input into a pre-trained multimodal classification network and outputs real-time planktonic species identification results and population abundance estimates.

Description

Technical Field

[0001] This invention relates to the field of nuclear power environmental monitoring technology, specifically to an online identification and management system for planktonic species at nuclear power plant water intakes. Background Technology

[0002] In the safe and stable operation of nuclear power plants, water environment monitoring at the intake plays an indispensable role, with the species and abundance of plankton being one of the key monitoring indicators. Nuclear power plant cooling water systems rely on the intake for continuous water intake, and the excessive proliferation or presence of specific plankton species can trigger a series of problems. For example, some plankton may adhere to and accumulate inside pipes, causing a decrease in cooling water flow rate and affecting heat exchange efficiency; the metabolic products of some plankton may also accelerate pipe corrosion and shorten equipment lifespan; even more seriously, the explosive growth of certain plankton may clog filters, directly threatening the normal operation of nuclear power equipment. Therefore, real-time and accurate identification and management of plankton at the intake is a crucial link in ensuring the safe and efficient operation of the nuclear power system.

[0003] Monitoring of planktonic organisms at nuclear power plant intakes primarily relies on a combination of traditional methods and partially automated technologies, but this approach has several limitations. Traditional manual sampling and laboratory analysis methods require monitoring personnel to collect water samples periodically, and then perform microscopic observation and morphological comparison to identify and count species. This method is not only resource-intensive but also has a long sampling cycle, making real-time monitoring difficult. It often fails to capture the dynamic changes in planktonic populations in a timely manner, resulting in a significant lag in monitoring results and hindering effective support for immediate decision-making.

[0004] Existing automated monitoring equipment mostly relies on single-spectral image acquisition and simple feature matching for identification, making it significantly affected by the complex underwater environment. Factors such as underwater light scattering and changes in water turbidity can lead to decreased image quality and inaccurate feature extraction. Furthermore, these devices typically ignore the temporal evolution and spatial distribution characteristics of plankton, relying solely on local features from a single frame for identification, which is insufficient to address the morphological variability and diverse species characteristics of plankton. In addition, existing technologies lack effective adaptive adjustment mechanisms when dealing with samples with ambiguous features, resulting in insufficient identification stability when facing anomalies such as population migration, thus limiting the reliability and practicality of monitoring results. Summary of the Invention

[0005] The purpose of this invention is to provide an online identification and management system for planktonic species at nuclear power plant water intakes, in order to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides an online identification and management system for planktonic species at nuclear power plant water intakes, the system comprising:

[0007] Acquire a sequence of continuously collected multispectral underwater images and corresponding aquatic environmental parameters from the water intake of a nuclear power plant;

[0008] Feature extraction is performed on the multispectral underwater image sequence to generate a multidimensional planktonic feature vector;

[0009] Based on the collected timestamps and spatial location information, the multidimensional planktonic feature vectors are mapped onto temporal spatial grid nodes to construct a spatiotemporal correlation graph structure;

[0010] An incremental prototype network is used to dynamically cluster the spatiotemporal association graph structure, and the currently active prototype node is identified in each iteration.

[0011] The fuzzy prototype node is determined based on the intersection relationship between the current feature set of the activated prototype node and the historical feature sets of other prototype nodes in a preset number of historical iteration stages.

[0012] Calculate the adaptive learning rate of the fuzzy prototype node;

[0013] The feature representations of the fuzzy prototype node and the activated prototype node are updated according to the adaptive learning rate.

[0014] Identify feature points with abnormal migration in the intersection relationship, and calculate their abnormal probability based on the neighborhood distribution density of the feature points;

[0015] The anomaly probability is combined with the contextual semantic disambiguation of the multidimensional planktonic feature vector to generate an optimized feature representation;

[0016] The optimized feature representation is input into a pre-trained plankton multimodal classification network, which outputs real-time plankton species identification results and population abundance estimates.

[0017] Preferably, the adaptive learning rate is determined based on a combination of the proportion of intersection features, the feature distance ratio between the activated prototype node and the fuzzy prototype node, the historical position drift variance ratio between the activated prototype node and the fuzzy prototype node, and the baseline learning rate of the activated prototype node.

[0018] Preferably, feature extraction of the multispectral underwater image sequence includes:

[0019] A depthwise separable convolutional network was used to extract morphological features from different spectral channels.

[0020] By integrating the morphological features with the corresponding aquatic environment parameters, a multidimensional planktonic feature vector containing the spectral-environment coupling relationship is generated through a feature cross-layer.

[0021] The constructed spatiotemporal correlation graph structure includes:

[0022] The water intake monitoring area at each data collection time is divided into a uniform spatial grid;

[0023] Set each grid center as the temporal spatial grid node;

[0024] The multidimensional planktonic feature vectors falling into the same grid are averaged and pooled to serve as the initial features for that node.

[0025] Based on the spatial distance between adjacent grid nodes and the interval between consecutive time nodes, the connection edges of the temporal spatial grid nodes with spatiotemporal weights are constructed.

[0026] Preferably, determining the fuzzy prototype node includes:

[0027] When the current feature set of the activated prototype node has a non-empty intersection with the historical feature set of other prototype nodes, the other prototype node is marked as a suspected fuzzy prototype node.

[0028] The outlier degree of the neighborhood density of each feature point in the intersection is calculated as the anomaly probability.

[0029] Feature points with an anomaly probability lower than a preset threshold are selected as stable migration feature points;

[0030] If the historical feature set of the suspected fuzzy prototype node contains the stable migration feature point, then it is determined to be the fuzzy prototype node.

[0031] Preferably, calculating the outlier degree of the neighborhood density of each feature point in the intersection includes:

[0032] Divide the target feature point into a multi-directional fan-shaped neighborhood region;

[0033] Calculate the average distance between the feature point and the target feature point within each of the said sector-shaped neighborhood regions;

[0034] The neighborhood density outlier is generated based on the ratio between the dispersion of the average distance of each sector region and the overall distribution mean.

[0035] Preferably, the contextual semantic disambiguation includes:

[0036] Extract the feature points whose anomaly probability is greater than a preset threshold and their corresponding spatiotemporal grid node information to form a set of features to be disambiguated.

[0037] Obtain the feature vectors of the neighboring nodes of the spatiotemporal grid node where the feature point to be disambiguated is located, and form a neighborhood feature vector;

[0038] The set of features to be disambiguated and the neighborhood feature vectors are input into the graph attention disambiguation module;

[0039] The graph attention disambiguation module generates semantic association weights between the feature point to be disambiguated and each neighboring feature.

[0040] The feature points to be disambiguated are reweighted based on the semantic association weights to generate the optimized feature representation.

[0041] Preferably, the system further includes:

[0042] Based on the characteristic distribution patterns of historical iteration stages, a probability diagram model of planktonic community succession is constructed.

[0043] The multidimensional planktonic feature vector of the current iteration stage is input into the planktonic community succession probability graph model;

[0044] Predict the probability of occurrence of key species and the trend of community structure changes in the next monitoring cycle;

[0045] When the predicted probability of the occurrence of the key species exceeds a preset warning threshold, a pre-start command for the water intake protection equipment is triggered.

[0046] Preferably, the update mechanism of the incremental prototype network includes:

[0047] After each clustering iteration, record the feature vector and corresponding timestamp of the current prototype node;

[0048] When new monitoring data is input, the feature similarity between the new feature vector and the historical prototype node is calculated.

[0049] If the feature similarity is lower than the dynamically adjusted similarity threshold, a new prototype node is created.

[0050] Otherwise, the newly added feature vector is merged into the feature set of the historical prototype node with the highest similarity.

[0051] The preferred workflow of the graph attention disambiguation module includes:

[0052] The feature points to be disambiguated are used as the query vector, and the neighborhood feature vector is used as the key vector.

[0053] Calculate the multi-scale semantic relevance between the query vector and each key vector;

[0054] The semantic association weights are generated by fusing the multi-scale semantic relevance through a gating mechanism.

[0055] The semantic association weights are used to weight and aggregate the neighborhood feature vectors to generate a context-enhanced vector.

[0056] The context enhancement vector is residually concatenated with the feature points to be disambiguated to obtain the optimized feature representation.

[0057] Preferably, the planktonic multimodal classification network includes:

[0058] Parallel processing of the morphological feature stream, spectral feature stream, and environmental feature stream in the optimized feature representation;

[0059] Abstract representations of different feature streams are fused through a cross-modal attention mechanism;

[0060] Multi-level classification tasks are performed based on the fused multimodal feature vectors, and species category identifiers and confidence estimates are output simultaneously.

[0061] Compared with the prior art, the beneficial effects of the present invention are:

[0062] This system, through multi-dimensional technology fusion and optimization, provides a more comprehensive and adaptable solution for online identification and management of plankton species at nuclear power plant water intakes. In the feature acquisition stage, by combining multispectral underwater image sequences with aquatic environment parameters, it overcomes the limitations of single image information, enabling the extracted multi-dimensional plankton feature vectors to more comprehensively reflect the biological characteristics of plankton and their environmental background, providing richer basic information for subsequent identification.

[0063] At the feature processing level, a spatiotemporal correlation graph structure is constructed based on timestamps and spatial location information, mapping plankton feature vectors to temporal spatial grid nodes. This allows previously scattered feature information to form correlations in both time and space. This correlation not only preserves the morphological change trajectories of plankton at different times but also reflects their aggregation or diffusion patterns in spatial distribution. It helps to capture potential patterns in population dynamics and avoids the one-sidedness that may result from isolated analysis of single-frame images or single-location data.

[0064] Employing an incremental prototype network for dynamic clustering training allows the system to continuously adapt to the evolution of planktonic populations during iterations. By identifying currently active prototype nodes and combining them with feature sets from historical iterations, fuzzy prototype nodes are determined, enabling the system to focus on samples with unclear features. Calculating the adaptive learning rate for fuzzy prototype nodes and updating their feature representations enhances the system's ability to handle samples with ambiguous features, avoiding overfitting or underfitting issues that may occur in traditional fixed-learning-rate models, and allowing the model to maintain better adaptability in complex and changing aquatic environments.

[0065] In terms of anomaly handling, abnormal migration feature points in the feature intersection are identified and their anomaly probabilities are calculated. Contextual semantic disambiguation is then performed based on these probabilities, effectively reducing the interference of abnormal data on feature representation. In underwater environments, factors such as sudden changes in light and water flow disturbances may cause anomalies in some feature points. These anomalies are identified and corrected through neighborhood distribution density analysis, making the generated optimized feature representation closer to the true characteristics of planktonic organisms and reducing the possibility of misidentification.

[0066] By inputting optimized feature representations into a pre-trained multimodal classification network, the advantages of the pre-trained model in feature extraction and pattern recognition are fully utilized. Combined with optimized features, the output real-time plankton species identification results are more closely aligned with actual conditions. Population abundance estimation is also more valuable due to the accuracy of the basic features and the integration of spatiotemporal correlations, providing a more reliable basis for water environment management at nuclear power plant intakes. This helps in the timely detection of potential ecological risks and meets the real-time, accurate, and stable requirements for plankton monitoring in nuclear power plant operations. Attached Figure Description

[0067] Figure 1 This is a timing diagram of the online identification and management system for planktonic species at nuclear power plant water intakes as described in this invention.

[0068] Figure 2 A flowchart for determining the adaptive learning rate;

[0069] Figure 3 A flowchart for feature extraction and spatiotemporal correlation graph construction;

[0070] Figure 4 A flowchart for contextual semantic disambiguation. Detailed Implementation

[0071] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0072] Please see Figure 1 This invention provides an online identification and management system for planktonic species at nuclear power plant water intakes, the system comprising:

[0073] Multispectral image sequences of the nuclear power plant's water intake area were continuously acquired using underwater cameras, simultaneously recording aquatic environmental parameters such as water temperature, salinity, and turbidity. The image sequences were processed using a depthwise separable convolutional network to extract morphological features from different spectral channels, which were then fused with environmental parameters to generate multidimensional planktonic feature vectors. Based on the acquisition timestamps and spatial location information, the feature vectors were mapped onto spatiotemporal grid nodes, constructing a graph structure with spatiotemporal weights. An incremental prototype network was used for dynamic clustering training of the graph structure, identifying activated prototype nodes and determining fuzzy prototype nodes. The adaptive learning rate of the fuzzy prototype nodes was calculated to update the node feature representations. Anomalies in feature point migration were detected, anomaly probabilities were calculated, and contextual semantic disambiguation was performed to generate optimized feature representations. Finally, the optimized features were input into a multimodal classification network, outputting planktonic species identification results and population abundance estimates.

[0074] Example 1: See Figure 2 This paper describes an adaptive learning rate calculation and update mechanism for an online plankton species identification and management system at a nuclear power plant intake. The mechanism aims to precisely adjust the feature representation of prototype nodes in a dynamically changing plankton monitoring environment to address changes in biological communities under different seasons and aquatic conditions. The adaptive learning rate calculation process integrates multiple dynamic indicators, comprehensively analyzing the feature distribution relationships, spatial location changes, and historical evolution patterns among nodes to form an environmentally adaptive learning strategy.

[0075] During system operation, each prototype node maintains its own feature set and historical state records. When new monitoring data is input into the system, the incremental prototype network identifies the currently active set of nodes. These nodes are compared with other nodes from previous iterations to detect any feature intersections. The statistical analysis of intersection features considers not only the quantity but also the distribution density of feature points in multidimensional space. The system establishes a spatiotemporal index for feature points, recording the location and timestamp of each point in historical iterations.

[0076] The calculation process first processes the feature overlap ratio. This ratio reflects the similarity between two types of nodes. It is calculated by counting the number of feature points shared by the currently active node and the target blurred node, and then dividing each by the total feature set of both nodes. The ratio is calculated using a sliding window averaging method, taking into account the results of the last five iterations to smooth out short-term fluctuations. The system also monitors the dynamic trend of this ratio, and adjusts its weight on the final learning rate accordingly when a continuous upward or downward pattern is detected.

[0077] The calculation of the feature distance ratio involves a measurement of a multi-dimensional feature space. The system employs a modified Euclidean distance formula, introducing feature importance weights when calculating point-to-point distances. These weights are obtained through offline training and reflect the contribution of different feature dimensions to species differentiation. The ratio of the current distance to the historical average distance undergoes a logarithmic transformation to reduce the impact of extreme values. The distance ratio is updated every three iterations to maintain the stability of the indicator while avoiding excessive lag.

[0078] The position drift variance ratio reflects the regularity differences in node motion. The system establishes a position trajectory database for each prototype node, recording its coordinate changes in the feature space. Variance calculation is based on position data from the most recent ten iterations, using an unbiased estimation method. The variance ratio of active nodes to fuzzy nodes is normalized and mapped to a fixed range. This index pays particular attention to changes in node motion patterns under specific seasonal or environmental conditions, such as aquatic environmental changes caused by sudden temperature fluctuations.

[0079] The baseline learning rate is set according to the lifecycle principle. Newly created nodes have a higher initial learning rate, which is beneficial for rapid adaptation to environmental characteristics. As the node's lifespan increases, the learning rate decays exponentially, and the decay rate is related to the frequency of biological activity in the node's region. The system periodically evaluates the learning performance of the nodes, and when a decrease in feature update efficiency is detected, a local adjustment mechanism for the learning rate is triggered to avoid premature convergence or continuous oscillation.

[0080] The four core indicators mentioned above are weighted and fused to form the final adaptive learning rate. The weight coefficients are not fixed values ​​but are dynamically adjusted based on the stability of the monitoring environment. When water environment parameters fluctuate significantly, the weights of the distance ratio and drift variance are increased; when the biological community structure is relatively stable, the influence of the feature overlap ratio is emphasized. The weight adjustment strategy is implemented through a reinforcement learning framework, and the system continuously optimizes the weight allocation scheme based on feedback from the effects of historical decisions.

[0081] The learning rate update process employs a phased strategy. In the initial calculation phase, the system generates a base learning rate value, which is applied to the coarse adjustment of fuzzy nodes. Subsequently, a fine-tuning phase is initiated, analyzing the local effects of feature updates and fine-tuning the learning rate. This two-phase method balances convergence speed and accuracy requirements. The node feature update operation uses incremental gradient descent, adjusting only a subset of dimensions in each iteration to avoid drastic changes that could disrupt the existing knowledge structure.

[0082] The system also establishes a safe boundary mechanism for the learning rate. Upper and lower thresholds are set to prevent excessively large learning rates from causing oscillations or excessively small learning rates from causing stagnation. The boundary values ​​adaptively change according to the overall network state; the upper limit is automatically widened when a large number of new nodes are detected, while the lower limit is tightened during network convergence. This dynamic boundary management effectively maintains the stability of the training process.

[0083] To cope with sudden environmental changes, the system has an emergency learning rate adjustment channel. This channel can be triggered when the water environment sensor detects abnormal parameter fluctuations or when image analysis reveals drastic changes in the biological community. In this case, the learning rate calculation will incorporate additional environmental factors, shortening the time window for indicator statistics and improving the response speed to emergencies. The emergency adjustment is time-sensitive; it automatically switches back to the standard calculation mode after the environment returns to normal.

[0084] The adaptive learning rate mechanism is deeply integrated with other modules of the system. The quality assessment data provided during the feature extraction phase influences the initial learning rate setting. The node topology information maintained by the spatiotemporal graph structure provides contextual reference for positional drift analysis. The output of the anomaly detection module marks nodes requiring special handling, and their learning rate updates employ a conservative strategy. This cross-module collaboration ensures that learning rate adjustments are based on global information.

[0085] The mechanism demonstrates good adaptability to environmental changes in practice. In response to changes in biological activity patterns caused by day-night cycles, the learning rate automatically matches the feature update needs of different time periods. Faced with changes in species composition due to seasonal changes, the mechanism maintains stable learning performance by adjusting indicator weights. In abnormal situations such as sudden water pollution, the activation of the emergency channel effectively prevents a sharp decline in model performance.

[0086] The system displays the dynamic changes in the learning rate through a visual interface. Operations personnel can observe the real-time values ​​and curves of the learning rate at different nodes to understand the network's training status. The interface also provides historical data analysis tools, allowing users to filter and view learning rate evolution trends by time range, geographical location, or species type. These functions assist in manual supervision and system optimization.

[0087] The learning rate calculation module employs a distributed architecture, supporting parallel processing of large-scale node sets. Computational tasks are grouped and distributed across different processor units according to the number of nodes, and intermediate results are aggregated via a high-speed data bus. This design meets real-time requirements, maintaining iteration frequency even in scenarios with tens of thousands of nodes. The module implements efficient memory management, and the feature comparison process uses an approximate nearest neighbor search algorithm, significantly reducing computational complexity.

[0088] To ensure long-term operational reliability, the system incorporates a built-in learning rate health monitoring subsystem. This subsystem periodically checks the learning rate distribution, update frequency, and status of related metrics for each node, initiating a self-repair process upon detecting abnormal patterns. The subsystem records complete diagnostic logs, supporting post-event analysis and algorithm improvement. This self-monitoring mechanism ensures the adaptive learning rate function continues to operate stably.

[0089] Example 2: See Figure 3This paper covers the feature extraction and spatiotemporal modeling mechanism in the planktonic identification system of nuclear power plant water intakes, as well as the determination process of fuzzy prototype nodes. This part handles the conversion process from multispectral images to spatiotemporal feature structures and establishes a node fuzziness identification standard in dynamic environments.

[0090] The processing of multispectral underwater image sequences employs a hierarchical feature extraction architecture. Raw images captured by underwater cameras are grouped by spectral band and input into processing channels, each configured with an independent depthwise separable convolutional branch. The first convolutional layer uses a 3×3 kernel size to extract microtexture features while preserving positional information. The second layer employs dilated convolution to expand the receptive field and capture the contour features of cell populations. The third layer implements a channel attention mechanism to enhance the response to important features. The outputs of each spectral channel are fused element-wise to form a comprehensive morphological feature representation. Aquatic environmental parameters, including temperature, salinity, turbidity, dissolved oxygen, and pH, are normalized and input into the feature cross-layer. The cross-layer uses a bilinear interaction model to perform an outer product operation between the morphological feature vector and the environmental parameter vector, generating a high-dimensional interaction matrix. This matrix is ​​then dimensionality-reduced through low-rank decomposition, preserving the main coupling relationships, and finally outputting a 128-dimensional planktonic feature vector.

[0091] The spatiotemporal correlation graph structure employs a hierarchical grid mapping strategy. The water intake monitoring area is divided into square grids with sides of 20 cm, with the grid center point designated as a spatial node. Each node is associated with three-dimensional coordinate information, including planar location and water depth data. The time dimension is divided with minute-level precision, forming an independent time layer for each complete monitoring cycle. Feature vectors are mapped to corresponding spatial grid nodes based on the GPS coordinates and underwater depth data of the acquisition equipment. When multiple feature vectors fall into the same grid, mean pooling is performed to generate initial features for the node. A time decay factor is added during the pooling process, with more recent data receiving higher weights. The establishment of node connections follows the principle of spatiotemporal proximity. In the spatial dimension, adjacent grid nodes are connected, with the weight decreasing by 20% for every 20 cm increase in distance. In the time dimension, nodes from consecutive monitoring cycles are connected, with the weight decreasing by 15% for every 5 minutes increase in time interval. The edge weight is ultimately expressed as the product of the spatial decay coefficient and the time decay coefficient.

[0092] The process for identifying fuzzy prototype nodes involves multi-stage analysis. After the incremental prototype network identifies an active node, the system retrieves the intersection of the node's current feature set with the features of historical prototype nodes. Nodes with non-empty intersections are marked as suspected fuzzy nodes. For each feature point in the intersection, an eight-directional fan-shaped neighborhood analysis is performed: the feature space is divided into eight fan-shaped regions at 45-degree intervals, centered on the feature point. The average distribution density of neighboring feature points within each region is calculated, and the standard deviation of the density across the eight regions is calculated. When the standard deviation is below a set threshold, the feature point is considered to have a stable neighborhood structure. The system scans all intersection feature points and filters out a set of stable feature points. For each suspected fuzzy node, the proportion of stable feature points in its historical feature set is calculated. When this proportion exceeds a 60% threshold, and stable feature points exhibit a continuous occurrence pattern in the node's historical record, the system formally identifies it as a fuzzy prototype node. The determination result is recorded in the node attributes, triggering the subsequent adaptive learning rate calculation process.

[0093] The feature extraction process is subject to online quality monitoring. A feature confidence evaluation module is added to the output layer of the convolutional network to analyze the uniformity of the activation distribution of the feature vectors. When feature degradation is detected, the weight constraints of the convolutional kernels are automatically adjusted. A dynamic optimization mechanism is implemented for spatiotemporal grid partitioning, automatically adjusting the grid size based on the feature point distribution density. A fine 10-cm grid is used in densely populated plankton areas, while a coarse 30-cm grid is used in sparse areas. The grid size adjustment follows a gradual principle, with a maximum change of no more than 5% per iteration to avoid drastic reconstruction affecting system stability.

[0094] The fuzzy node determination process incorporates time consistency verification. The system traces the historical state changes of suspected nodes and analyzes their characteristic interaction patterns with active nodes. If periodic interaction patterns are detected, such as feature migrations related to day-night cycles or tidal cycles, the fuzzy determination weight for that node is increased. For newly created nodes, an initial observation period is set, during which fuzzy determination is temporarily suspended to avoid misjudgments due to insufficient early data.

[0095] During operation, the feature extraction thread and the spatiotemporal modeling thread adopt a producer-consumer model. The feature vectors output by the image processing pipeline are stored in a circular buffer, and the spatiotemporal modeling module reads the data according to a fixed time window. The two threads are synchronized through semaphores to ensure the timeliness of data processing. The fuzzy node determination runs as an independent background task, utilizing idle computing periods to perform depth analysis, thus avoiding blocking the real-time recognition process.

[0096] The system maintains a closed-loop optimization mechanism for feature extraction and spatiotemporal modeling. Node distribution information fed back from the spatiotemporal graph structure guides the focus of feature extraction on key regions. When a specific grid node is detected to exhibit consistently high variability, the image processing resolution of the corresponding region is automatically enhanced. This cross-module collaborative optimization enables the system to dynamically adapt to changes in phytoplankton distribution.

[0097] A comprehensive anomaly handling mechanism was established during implementation. When consecutive failed frames occur during feature extraction, an image quality enhancement subprocess is initiated, employing multi-frame fusion technology to compensate for single-frame defects. When the spatiotemporal grid mapping encounters boundary feature points, bilinear interpolation is performed to assign them to adjacent nodes. In cases of disputed decisions regarding ambiguous node determination, a multi-evidence voting mechanism is activated, comprehensively considering factors such as feature similarity, historical trajectory, and environmental parameters to make a decision.

[0098] Example 3: See Figure 4 This research focuses on neighborhood density analysis and semantic disambiguation mechanisms in a plankton identification system for nuclear power plant water intakes, specifically addressing outlier identification and feature optimization in the feature space. The system module achieves accurate characterization of plankton features through multi-dimensional distribution analysis combined with graph attention mechanisms.

[0099] The neighborhood density outlier calculation employs a joint spatial-temporal analysis method. For each point to be evaluated in the feature space, a polar coordinate system centered on that point is established, dividing the surrounding area into eight 45-degree sector analysis regions. Within each sector, the K nearest feature points to the center point are counted, and the average standardized distance between these points and the center point is calculated. The standardized distance calculation considers the feature dimension weights, with different feature dimensions assigned different importance based on their discriminative ability in species classification. The average distances of the eight sector regions constitute the distance vector D = [ , ,..., The outlier index is calculated using this vector. :

[0100]

[0101] in: The standard deviation of the distance vector is represented by... This represents the mean of the distance vector. This indicates the frequency of occurrence of the feature point in the most recent N iterations. and This is a parameter that adjusts the rate of time decay. The formula takes into account both the spatial dispersion and the temporal regularity of the distribution; the larger the outlier value, the more likely the point is to be an anomalous feature.

[0102] The contextual semantic disambiguation process constructs a three-stage processing pipeline. The first stage collects the feature points to be disambiguated and their spatiotemporal context information, including the coordinates of their respective grid nodes, the collection timestamp, and features of neighboring nodes. The second stage implements a graph attention mechanism, using the point to be disambiguated as the query vector Q, and neighboring feature points as the key vector K and value vector V. Attention weights are calculated using multi-scale similarity metrics: cell structure similarity is calculated at the local morphological scale, group spatial relationships are evaluated at the regional distribution scale, and feature change trajectories are analyzed at the temporal evolution scale. These three-scale attention weights are fused through a gating network to form the final semantic association matrix. The third stage performs feature reconstruction, weighting and aggregating neighboring features based on the association matrix to generate a context-enhanced vector. This vector is merged with the original features through residual connections, and then processed by layer normalization to output an optimized feature representation.

[0103] The polar coordinate sector analysis region is divided using a dynamic adjustment strategy. When a significant cluster of feature points is detected in a certain direction, the sector in that direction is automatically subdivided into smaller angular intervals. The subdivision threshold is adaptively determined based on the overall distribution density; high-density regions are subdivided at 22.5 degrees, while low-density regions retain their original 45-degree subdivision. This dynamic subdivision ensures more accurate outlier assessment in regions with varying high density.

[0104] The calculation of the time decay factor traces the occurrence history of feature points. The system maintains a lifecycle record for each feature point, statistically analyzing its frequency and interval patterns within the most recent 24-hour monitoring period. For periodically occurring feature points, even with high spatial outlier rates, the final outlier score is appropriately reduced. The length of the time window is set based on the activity patterns of plankton, with shorter time windows for diurnal organisms and longer time windows for nocturnal organisms.

[0105] The graph attention disambiguation module enables multi-granularity feature interaction. At the local morphological scale, it focuses on microscopic features such as cell wall texture and pigment distribution, using 3×3 convolutional kernels to extract local similarity. At the regional distribution scale, it calculates the spatial relationships between feature points in the grid, introducing relative position encoding to enhance spatial awareness. At the temporal evolution scale, it analyzes the changing patterns of feature points across consecutive time frames, using one-dimensional temporal convolution to capture dynamic features. Attention heads at the three scales are computed in parallel, with output dimensions of 64, 32, and 32 respectively, and are then fused after being uniformly mapped to the same dimension through a fully connected layer.

[0106] The feature reconstruction process employs a dual verification mechanism. The initial reconstruction result is compared with the original features using cosine similarity; if the similarity falls below a threshold, a second reconstruction is triggered. The second reconstruction expands the neighborhood, incorporating more context nodes to recalculate attention weights. Before the final output, outlier pruning is performed, limiting the maximum contribution weight of a single neighborhood feature to a preset upper limit to prevent interference from extreme values.

[0107] This module employs a heterogeneous computing architecture. The density analysis component is deployed on a GPU array, leveraging parallel computing to accelerate sector processing. The attention mechanism runs on a TPU unit, optimizing matrix operation efficiency. The feature reconstruction stage is controlled by the CPU, coordinating memory data transfer and intermediate result storage. A dedicated cache is provided to store recent feature optimization records, supporting real-time backtracking analysis.

[0108] A closed-loop quality monitoring system is established during operation. The distribution changes of outlier scores are monitored in real time, and evaluation parameters are automatically adjusted when an overall distribution shift is detected. The changes in each feature point before and after disambiguation are recorded, and the optimization effect is analyzed periodically. A dedicated optimization strategy cache is established for frequently occurring feature point categories to accelerate the processing of recurring patterns.

[0109] The system maintains a knowledge base of disambiguation rules, storing typical feature patterns under different seasons and water conditions. When a new recurring pattern is detected, the feature rules are automatically extracted and stored in the knowledge base. Subsequent processing prioritizes matching rules from the knowledge base to improve the disambiguation efficiency of common patterns. The knowledge base implements version management, supporting rule rollback and effect comparison.

[0110] A dynamic parameter adjustment mechanism continuously optimizes processing performance. It automatically adjusts the parallel computing granularity based on hardware load, balancing latency and throughput. It monitors latency at each stage of the feature processing pipeline and dynamically allocates computing resources. A feedback loop for processing performance is established, backpropagating the final classification result to disambiguation parameter adjustments. An anomaly handling mechanism ensures system robustness. When consecutive feature point loss is detected, a data recovery process is initiated, reconstructing missing features using spatiotemporal correlation. In the face of sudden changes in feature distribution caused by sudden water quality variations, it automatically switches to a robust processing mode, relaxing some judgment thresholds. The system retains a manual intervention interface, supporting expert experience injection and special case annotation. This module is deeply integrated with other parts of the system. The quality score provided during feature extraction guides the initial parameter settings for outlier calculation. The topological relationships maintained by the spatiotemporal graph structure provide extended neighborhood information for attention calculation. Feedback signals from the classification network are used to optimize the attention weight fusion strategy. This cross-module collaboration achieves an end-to-end feature optimization process. A multi-level caching system is established during implementation. High-frequency access feature patterns are cached in a fast memory lookup area, mid-frequency patterns are stored in a solid-state drive cache, and low-frequency patterns remain on disk. The cache replacement strategy considers the temporal locality and spatial correlation of feature points, prioritizing the retention of patterns that are likely to be reused. Cache consistency is maintained through a version stamp mechanism to ensure data synchronization in a distributed environment.

[0111] Example 4: A community succession prediction and protection linkage mechanism in a nuclear power plant water intake plankton identification system. This module analyzes historical characteristic distribution patterns, constructs a probabilistic graphical model to predict the trend of biological community changes, and forms a closed-loop control with the water intake protection equipment.

[0112] The system collects monitoring data from the past three years to construct a training dataset, divided into four subsets according to season: spring, summer, autumn, and winter. Each seasonal dataset contains approximately 90 days of continuous monitoring records, encompassing feature vectors, species labels, and corresponding aquatic environmental parameters. The data processing phase begins with feature clustering to identify typical community structure patterns for each season. Summer datasets may exhibit a pattern dominated by diatoms and dinoflagellates, while winter datasets may shift to a pattern dominated by copepods and protozoa. Core feature vectors for each pattern are extracted as state nodes, recording their temporal distribution and duration.

[0113] The construction of the probabilistic graphical model comprises two core components: a state transition matrix and an environmental response matrix. The state transition matrix describes the evolutionary relationships between different community patterns, calculating transition probabilities by statistically analyzing the frequency of pattern transitions in historical data. For example, analysis revealed that the summer diatom pattern has a higher probability of transitioning to an autumn diatom-dinoflagellate mixed pattern as temperatures decrease. The environmental response matrix establishes the association rules between aquatic environmental parameters and state transitions, stored in the form of conditional probability tables. Water temperature parameters are discretized into five intervals (e.g., <10℃, 10-15℃, 15-20℃, 20-25℃, >25℃), each temperature interval corresponding to a different state transition probability distribution. Salinity, turbidity, and other parameters are also discretized, collectively forming a multidimensional conditional probability model.

[0114] The prediction mechanism operates as follows: When the feature vector of a new monitoring period is input into the system, it is first matched against the historical pattern library for similarity. The matching process uses a weighted Euclidean distance metric, assigning higher weights to seasonally sensitive features. Successfully matched patterns are used as the current state node and input into the probabilistic graphical model for inference calculations. The model considers current water environment parameters and queries the environmental response matrix to obtain transition probability correction coefficients. A message passing algorithm is used to calculate the probability of occurrence of each state node within the next 24 hours, with particular attention paid to the probability of biological categories that may block the water intake. To more clearly demonstrate the correspondence between prediction results and protection responses, a summer monitoring day is used as an example, recording in detail the key species, occurrence probabilities, and corresponding protection response levels at different time points. This intuitively reflects the practical application effect of community succession prediction and protection linkage, as shown in Table 1 below.

[0115] The protection command triggering mechanism sets two levels of response thresholds. When the probability of jellyfish appearance exceeds 30%, a primary response is triggered: the low-frequency vibration mode of the water intake grille is activated to prevent organism attachment. When the probability of specific algae (such as Sargassum) exceeds 45%, a higher-level response is triggered: the ultrasonic dispersing device is activated and the filter cleaning frequency is increased. The command generation module encapsulates the control parameters into standard protocol messages and sends them to the field PLC controller via the industrial bus. Simultaneously, an early warning notification is generated, including a list of predicted species and a probability distribution chart, and pushed to the mobile terminals of maintenance personnel.

[0116] The model update mechanism implements rolling optimization. Daily comparisons and analyses are performed between actual monitoring results and predicted values ​​to calculate the prediction deviation for each state node. When the average deviation for a specific node exceeds 15% for three consecutive days, model parameter recalibration is initiated. The recalibration process employs incremental learning, adjusting only relevant transition probability parameters to maintain overall model stability. Comprehensive validation is performed during seasonal transitions, and cross-validation is conducted before loading the new seasonal sub-model.

[0117] Table 1: Example data of a forecast for a certain summer monitoring day

[0118] Predicted time point Key Species probability of occurrence Protection Response Level 06:00 Noctiluca scintillans 28% Monitoring status 10:00 Skeletalella costatum 52% ultrasonic start 12:00 Moon jellyfish 38% Grille cleaning 15:00 Scripplia conica 67% Dual device linkage 18:00 Noctiluca scintillans 45% ultrasonic start 22:00 Chinese Philosopher Water Flea 28% Monitoring status

[0119] The prediction results visualization system provides spatiotemporal distribution maps. A planar map displays a heatmap of species probability in different areas of the water intake, while a depth profile shows the predicted vertical distribution of organisms. The timeline view presents a 24-hour probability change curve, allowing users to click to view detailed prediction data for any given time point. The interface includes a prediction-actual comparison mode, displaying the difference between predicted and measured values ​​at each time point using a two-color bar chart.

[0120] The equipment linkage module implements a safety interlock mechanism. The ultrasonic dispersion equipment is linked with the water intake flow sensor, automatically reducing the dispersion intensity when an abnormal drop in flow is detected. The bar screen cleaning operation is coordinated with the underwater robot inspection plan to avoid interference from simultaneous operations. All protective operations are logged in detail, including triggering conditions, execution parameters, and actual effect evaluation data.

[0121] The model training system implements version control and rollback functionality. A model snapshot is created before each major parameter update, storing training data and parameter configurations. A rollback mechanism is automatically triggered when the performance of the new model version on the test set drops by more than 5%. Model health reports are generated periodically to analyze the prediction accuracy drift at each state node and highlight weaknesses requiring close monitoring. This module is deployed as an independent microservice architecture, receiving feature vector inputs via a message queue and outputting protection commands to the equipment control system. Service operation status is monitored in real time, automatically switching to simplified mode when prediction calculation latency exceeds a set threshold, calculating only the probability of key species. System resource allocation implements dynamic priority scheduling, automatically increasing computing resource quotas during water quality change warnings. The historical data management system adopts a hierarchical storage strategy. Data from the most recent three months is stored in a high-speed storage array, supporting real-time queries; intermediate-term data is compressed and stored in warm storage; long-term historical data is archived in a cold storage system. The data retrieval interface supports multi-condition combined queries, such as filtering training samples by species type, seasonal characteristics, and water temperature range. The model interpretation module provides prediction basis analysis functions. Operations and maintenance personnel can query the derivation path of any prediction result and view key historical events and environmental factors that affect the probability calculation. For important prediction conclusions, the system automatically generates attribution reports, listing the top five historical similar cases with the highest contribution to the transition probability, enhancing the credibility of the prediction results.

[0122] Example 5: A dynamic update and management mechanism for an incremental prototype network in a plankton identification system at a nuclear power plant intake. This mechanism is designed to adaptively maintain and optimize the prototype node set during continuous monitoring to adapt to the dynamic changes in the plankton community.

[0123] The system configures a circular data buffer for each active prototype node, storing the feature vectors and related timestamps from the most recent 100 iterations. The buffer employs a first-in, first-out (FIFO) management strategy, overwriting the oldest record when a new feature vector arrives, maintaining a constant storage capacity. Each feature vector is accompanied by metadata such as acquisition time, spatial location, and feature confidence level, forming a complete feature profile. The node buffer has status flags indicating its current state, such as active, awaiting evaluation, or dormant.

[0124] When new monitoring data is input into the network, a multi-stage similarity calculation process is executed. First, the input features are standardized and preprocessed to eliminate systematic biases between different monitoring devices. Then, the cosine similarity between the feature and all historical nodes is calculated in parallel. The similarity calculation employs a dimensional weighting strategy, assigning higher weights to feature dimensions with strong species differentiation. The initial similarity threshold is set at 0.85 and dynamically adjusted based on network complexity: when the total number of system nodes exceeds 500, the threshold decreases linearly by 0.01 for every 50 new nodes, with a minimum of 0.7. A hysteresis protection is implemented during the dynamic adjustment process, limiting the threshold change to no more than 0.05 every ten minutes.

[0125] Node operation decisions are based on similarity comparison results. When the highest similarity is below the current threshold, a new node creation process is triggered. New node initialization includes: assigning a unique identifier, creating an initial buffer, setting a baseline learning rate, and registering to the node management directory. New nodes have a two-week observation period, during which a special update strategy is used, increasing the learning rate by 30% to accelerate feature stabilization. When a historical node with a similarity higher than the threshold exists, a feature merging operation is performed. Cross-validation is performed before merging: the average similarity between the new feature and the target node's five most recent feature vectors is checked; if the fluctuation exceeds 15%, merging is postponed. After successful validation, the new feature is stored in the target node's buffer, and the node's representative vector is updated.

[0126] The representative vector calculation employs a time-weighted strategy. An exponential decay function is used to allocate weights, with the most recent feature vector receiving the highest weight. The decay coefficient is related to the stability of the node's features. New nodes use a larger decay coefficient to accelerate convergence, while stable nodes use a smaller coefficient to maintain continuity. After the representative vector is updated, the system automatically calculates the offset between the new vector and historical vectors. When an abnormal offset is detected, a buffer integrity check is triggered.

[0127] A node dormancy mechanism periodically assesses node activity. All nodes are scanned every 24 hours, and the number of feature updates in the most recent 48 hours is counted. Nodes without updates for two consecutive weeks are marked as dormancy candidates. Before dormancy determination, a feature value assessment is performed: the contribution of the node's feature patterns to historical recognition tasks is analyzed, and nodes with high contributions are retained in a pending evaluation state. Officially dormant nodes retain feature data but exit real-time computation, releasing computing resources. A wake-up channel is set for dormant nodes: when the similarity between a new input feature and a dormant node exceeds its pre-dormant threshold of 5%, the node is reactivated and its computational state resumes.

[0128] The system implements a node relationship maintenance subsystem. It records the feature migration history between nodes, and establishes an association index when a continuous bidirectional feature flow is detected between two nodes. Associated node groups share some computing resources and are prioritized for matching during similarity calculations. For strongly correlated node pairs, the system supports a semi-automatic merging operation, which requires manual confirmation before node fusion can be performed.

[0129] The technology employs a distributed in-memory database to store node data. Each node is allocated an independent storage partition, which is further divided into a hot data area and a cold data area. The hot data area stores the most recently accessed 20% of data, using in-memory computing for acceleration. The cold data area stores historical data, which is compressed and then stored on a solid-state drive. Data access implements a read-write separation mechanism; computing threads read data through cached copies, and update operations are asynchronously committed to a log queue.

[0130] The monitoring system tracks key metrics in real time. It records the time distribution of node creation / dormancy events, calculates similarity calculation latency, and monitors buffer usage. When the node growth rate exceeds a warning threshold, the similarity threshold is automatically tightened; when frequent node creation-dormancy oscillations are detected, the threshold is relaxed and a configuration check alert is issued. All operations are logged in an audit log, including detailed parameter snapshots that inform decision-making.

[0131] This mechanism establishes a feedback channel with the feature extraction module. When new nodes continuously emerge in a specific area, the image acquisition system is notified to adjust the shooting parameters for that area. The node distribution heatmap guides the pan-tilt-zoom (PTZ) of the underwater camera, optimizing the allocation of monitoring resources. The system periodically generates node evolution reports, visually displaying the dynamic trends of the feature space and assisting biologists in analyzing community succession patterns.

[0132] The anomaly handling process includes a multi-level recovery mechanism. When node data is corrupted, the most recent backup is restored from persistent storage; when similarity calculation fails, a simplified algorithm is switched to ensure service continuity; when a buffer overflow occurs, an emergency compression program is initiated to preserve core characteristics. All anomaly events trigger diagnostic data collection, forming a fault analysis case library.

[0133] The system provides a management interface to support manual intervention. Operators can manually adjust the similarity threshold of specific nodes, freeze the status of key nodes, or forcibly activate dormant nodes. The interface displays a node relationship graph and supports replaying the node evolution history along a timeline. Expert annotation functionality allows for the addition of semantic annotations to special feature patterns, enhancing the interpretability of node information.

[0134] The incremental update mechanism is designed with long-term operational stability in mind. A daily maintenance window is set to perform data consistency checks, and weekly defragmentation optimizes the storage layout. Memory management employs an elastic allocation strategy, dynamically expanding the resource pool during periods of rapid node growth. Network communication implements data verification and retransmission mechanisms to ensure complete synchronization of data across distributed nodes.

[0135] Example 6: This example describes a multimodal classification and output mechanism within a nuclear power plant water intake planktonic identification system. This module processes the optimized feature representations and generates the final species identification results and abundance estimates. The system employs a parallel multi-stream architecture to fuse heterogeneous features and achieves comprehensive judgment through cross-modal interaction.

[0136] After optimizing the feature representation input to the classification network, feature decoupling and splitting are first performed. The morphological feature stream focuses on the structural information of organisms and contains five cascaded residual convolutional blocks. The first block extracts basic cell contour features and uses 3×3 convolutional kernels to capture local edge responses. The second block introduces dilated convolutions to expand the receptive field and identify group arrangement patterns. The third block implements channel reweighting to enhance the activation intensity of key morphological features. The fourth block fuses multi-scale contextual information and integrates regional features through spatial pyramid pooling. The fifth block outputs a 128-dimensional abstract morphological representation, preserving high-order structural information. The spectral feature stream processes pigment absorption characteristics, and the input features are separated into different band responses by a spectral demultiplexing layer. Each band channel is configured with an independent one-dimensional convolutional layer to extract spectral curve features. The spectral attention module analyzes the discriminative weights of each band and suppresses redundant band responses. The feature fusion layer aggregates all band outputs to generate a 64-dimensional spectral feature vector. The environmental feature stream processes aquatic environmental parameters and learns nonlinear relationships through a three-layer fully connected network. The first layer maps the original parameters to a high-dimensional space, the second layer performs feature selection to retain the influence of key environmental factors, and the third layer outputs a 32-dimensional environmental embedding vector.

[0137] The cross-modal fusion mechanism implements spatial alignment and feature interaction in stages. The spatial alignment module receives morphological and spectral feature maps and performs adaptive registration using deformable convolution. The registration process considers the refractive distortion characteristics of underwater imaging and compensates for spatial deformation through offset learning. The aligned feature map is input to a cross-modal attention unit, which contains dual interaction channels. The morphological-to-spectral channel calculates the attention weights of each position of the morphological feature to the spectral feature, while the spectral-to-morphological channel performs reverse attention. The weight matrix is ​​normalized using softmax and weighted aggregation to generate interaction features. The environmental feature vector is expanded to the feature map size via spatial broadcasting and fused with the interaction features element-wise using gating. The gating signal is generated by the affine transformation of the environmental and interaction features, controlling the injection intensity of environmental information. The fused multimodal feature map is compressed into a 256-dimensional feature vector using global pooling.

[0138] The multi-level classification task simultaneously performs species identification and abundance estimation. The species classification branch adopts a hierarchical structure. The first-level coarse classification network outputs the probability distribution of phylum and class categories, containing 16 basic categories. This network consists of two fully connected layers, with batch normalization and Dropout regularization added in between. The second-level fine classification network receives the coarse classification results and the concatenated vector of multimodal features, outputting genus and species-level identification results, covering 120 species commonly found in local waters. The fine classification network contains three hidden layers and uses label smoothing loss to mitigate class imbalance. The abundance estimation branch implements ordinal regression, dividing population density into six levels. The regression network contains four fully connected layers, with the last layer using softmax to output the membership probability of each density level. The loss function is designed as a three-term weighted sum: coarse classification cross-entropy loss, fine classification focus loss, and abundance estimation ordinal regression loss. The weight coefficients are dynamically adjusted with each training epoch, initially focusing on coarse classification learning, and gradually increasing the weights of fine classification and abundance estimation in later stages.

[0139] The output module integrates all branch results to generate a monitoring report. Species identification results use the species identifier corresponding to the highest probability in the fine-grained classification network. When the highest probability falls below a set threshold, it reverts to the coarse-grained classification result and marks it as insufficient confidence. Abundance estimation uses the expectation method, multiplying the probability of each density level by the median of the corresponding level and then summing the results. Anomaly detection markers are added to the report. Anomaly markers are triggered when the following conditions occur: confidence differences in morphological and spectral features exceed 40%; environmental and biological features significantly deviate from historical patterns; and the distribution of abundance estimates exhibits a bimodal characteristic. The report data is encapsulated as a structured message, including timestamps, spatial grid coordinates, species identifiers, confidence scores, abundance estimates, and anomaly marker fields.

[0140] The system implements a distributed computing architecture. Morphological feature streams are deployed on GPU accelerator cards, utilizing CUDA cores for parallel convolution operations. Spectral feature streams run on TPU processors, optimizing matrix multiplication and addition efficiency. Environmental feature streams and the fusion module are CPU-controlled, coordinating memory data exchange. Classification tasks are distributed to multi-core processors for parallel execution, exchanging intermediate results via shared memory. The output module connects to a real-time database, writing recognition results while triggering alarm event subscription services. A quality monitoring system implements full-process tracking. It records the output distribution statistics of each feature stream, monitoring feature drift. The classification decision process records the top-3 candidate species and their probability differences to assist in subsequent result verification. The abundance estimation module saves probability distribution histograms for each density level, supporting error source analysis. The system periodically generates quality assessment reports, analyzing recognition stability under different lighting conditions and water quality states. The visualization interface provides multi-dimensional result displays. The main view presents a heat map of the water intake's planar distribution, with color coding representing dominant species types and brightness indicating population density. The depth profile view displays the vertical distribution characteristics of plankton. The species evolution view presents the community structure change trend in a timeline format. Clicking on a single grid node displays a detailed identification report, including a thumbnail of the original image, a feature response heatmap, and an explanation of the classification decision path. This module is designed with operational and maintenance needs in mind. The classification network supports online model updates, and new species samples can be integrated into the existing model through incremental learning. A sliding control for setting the confidence threshold allows operators to adjust the identification sensitivity according to task requirements. The results export function supports standard ecological monitoring data formats and is compatible with third-party analysis tools. The system maintains a complete operation log, recording the parameter configuration and execution details of each identification task.

[0141] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0142] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A nuclear power water intake plankton species online identification management system, characterized in that, The method comprises the following steps: acquiring a multi-spectral underwater image sequence continuously collected by a nuclear power water intake and corresponding water environment parameters; extracting features from the multi-spectral underwater image sequence to generate a multi-dimensional plankton feature vector; mapping the multi-dimensional plankton feature vector to a time-space grid node based on a collection timestamp and spatial position information to construct a time-space correlation graph structure; using an incremental prototype network to perform dynamic clustering training on the time-space correlation graph structure, and identifying a current active prototype node in each iteration; determining a fuzzy prototype node according to an intersection relationship between a current feature set of the active prototype node and a historical feature set of other prototype nodes in a preset number of historical iteration stages; calculating an adaptive learning rate of the fuzzy prototype node; updating feature representations of the fuzzy prototype node and the active prototype node according to the adaptive learning rate; identifying feature points with abnormal migration in the intersection relationship, and calculating an abnormal probability of the feature points based on a neighborhood distribution density of the feature points; performing context semantic disambiguation on the multi-dimensional plankton feature vector based on the abnormal probability to generate an optimized feature representation; inputting the optimized feature representation into a pre-trained plankton multi-modal classification network to output a real-time plankton species recognition result and a population abundance estimate; the context semantic disambiguation comprises: extracting feature points with an abnormal probability greater than a preset threshold and spatial grid node information belonging to the feature points to form a disambiguation feature set; acquiring adjacent node feature vectors of the time-space grid node where the feature points are located to form a neighborhood feature vector; inputting the disambiguation feature set and the neighborhood feature vector into a graph attention disambiguation module; generating semantic correlation weights of the disambiguation feature points and each neighborhood feature through the graph attention disambiguation module; performing feature re-weighting on the disambiguation feature points according to the semantic correlation weights to generate the optimized feature representation; the working process of the graph attention disambiguation module comprises: taking the disambiguation feature points as query vectors and the neighborhood feature vectors as key vectors; calculating multi-scale semantic correlation degrees of the query vectors and each key vector; generating the semantic correlation weights by fusing the multi-scale semantic correlation degrees through a gating mechanism; performing weighted aggregation on the neighborhood feature vectors by using the semantic correlation weights to generate a context-enhanced vector; performing residual connection between the context-enhanced vector and the disambiguation feature points to obtain the optimized feature representation.

2. The online identification management system for phytoplankton species of a nuclear power water intake according to claim 1, characterized in that, The adaptive learning rate is determined based on an intersection feature quantity proportion, a feature distance ratio of the active prototype node to the fuzzy prototype node, a historical position drift variance ratio of the active prototype node to the fuzzy prototype node, and a baseline learning rate of the active prototype node.

3. The online identification management system for phytoplankton species of a nuclear power water intake according to claim 1, characterized in that, The feature extraction on the multi-spectral underwater image sequence comprises: extracting morphological features in different spectral channels by using a depth separable convolutional network; fusing the morphological features and corresponding water environment parameters to generate the multi-dimensional plankton feature vector containing spectral-environment coupling relationships through a feature cross layer; the construction of the time-space correlation graph structure comprises: dividing a monitoring area of the water intake at each collection time into uniform spatial grids; Each grid center is set as the time-space grid node; The multi-dimensional plankton feature vectors falling into the same grid are mean-pooled as the initial features of the node; According to the spatial distance of adjacent grid nodes and the interval length of continuous time nodes, the connection edges of the time-space grid node with space-time weight are constructed.

4. The online identification management system for phytoplankton species of a nuclear power water intake according to claim 1, characterized in that, The determination of the fuzzy prototype node includes: When the current feature set of the activated prototype node and the historical feature set of other prototype nodes have a non-empty intersection, mark the other prototype node as a suspected fuzzy prototype node; Calculate the neighborhood density outlying degree of each feature point in the intersection as the anomaly probability; Screen the feature points with anomaly probability lower than the preset threshold as stable migration feature points; If the historical feature set of the suspected fuzzy prototype node contains the stable migration feature points, it is determined as the fuzzy prototype node.

5. The online identification management system for phytoplankton species of a nuclear power water intake according to claim 4, characterized in that, The calculation of the neighborhood density outlying degree of each feature point in the intersection includes: Divide multi-direction sector neighborhood regions with the target feature point as the center; Calculate the average distance between the feature points in each sector neighborhood region and the target feature point; According to the proportional relationship between the discrete degree of the average distance of each sector region and the overall distribution mean, the neighborhood density outlying degree is generated.

6. The online identification management system for phytoplankton species of a nuclear power water intake according to claim 1, characterized in that, Further includes: Based on the feature distribution law of the historical iteration stage, a plankton community succession probability graph model is constructed; The multi-dimensional plankton feature vector of the current iteration stage is input into the plankton community succession probability graph model; The appearance probability of key species and the community structure change trend in the next monitoring period are predicted; When the predicted appearance probability of the key species exceeds the preset warning threshold, the pre-start instruction of the water intake protection equipment is triggered.

7. The online identification management system for phytoplankton species of a nuclear power water intake according to claim 1, characterized in that, The update mechanism of the incremental prototype network includes: After each clustering iteration, the feature vector and corresponding timestamp of the current prototype node are recorded; When new monitoring data is input, the feature similarity between the new feature vector and the historical prototype nodes is calculated; If the feature similarity is lower than the dynamically adjusted similarity threshold, a new prototype node is created; Otherwise, the new feature vector is merged into the feature set of the historical prototype node with the highest similarity.

8. The online identification management system for phytoplankton species of a nuclear power water intake according to claim 1, characterized in that, The plankton multi-modal classification network includes: The morphological feature flow, spectral feature flow and environmental feature flow in the optimized feature representation are processed in parallel.

Images