Dam seepage intelligent monitoring and risk early warning method, system and device based on expert large model and storage medium

Abstract

The invention discloses a dam seepage intelligent monitoring and risk early warning method, system and device based on an expert large model and a storage medium, and belongs to the field of hydraulic engineering monitoring, and the method comprises the steps: collecting multi-source heterogeneous data from a dam seepage monitoring system, and marking structural domain labels according to data sources and structural attributes; executing corresponding feature fusion processing on different types of data according to the structural domain labels, and constructing a structured feature vector group in a unified format; inputting the feature vector group into a structure-driven expert large model, activating a bound expert path according to a structural domain label, and executing multi-task reasoning; the multi-task reasoning result is input into a dam seepage virtual twin system, a simulation module is driven to conduct seepage diffusion evolution simulation, and a space risk map is generated; and generating an early warning signal and an emergency plan text according to the risk map and the multi-task reasoning result.

Description

Technical Field

[0001] The present invention relates to the field of water conservancy project monitoring, and in particular to a dam seepage intelligent monitoring and risk early warning method, system, equipment and storage medium based on an expert large model. Background Art

[0002] Dams are crucial water conservancy infrastructure, and their safe and stable operation is directly linked to national and people's livelihoods, including flood control, water supply, and power generation. Seepage is one of the most common dam failures. Long-term or abnormal seepage can damage the dam structure and even lead to dam failure. Therefore, real-time, accurate monitoring of dam seepage and risk warnings are crucial.

[0003] Traditional dam seepage monitoring methods rely primarily on manual inspections and numerical simulations based on physical models. Manual inspections are inefficient, lack full coverage of the dam area, and are limited by the experience and subjective judgment of inspectors. While numerical simulations based on physical models can quantitatively analyze the seepage field, they are complex to calibrate model parameters and require extensive computation, making them inadequate for real-time early warning. Furthermore, model accuracy is limited by model assumptions and parameter accuracy.

[0004] In recent years, with the development of sensor technology, the Internet of Things (IoT), and big data technologies, dam safety monitoring systems have accumulated massive amounts of multi-source, heterogeneous monitoring data. This data contains rich information on seepage conditions and risks, but traditional methods have struggled to effectively utilize this massive data for intelligent analysis and risk warning.

[0005] Furthermore, existing dam seepage early warning methods are mostly based on threshold-based alarms, which provide limited warning information and are difficult to provide effective emergency decision-making support. When abnormal seepage occurs, managers often struggle to quickly and accurately determine the risk level and impact scope, and develop appropriate emergency response plans. Summary of the Invention

[0006] To solve the above technical problems, a dam seepage intelligent monitoring and risk warning method based on expert large model is proposed. The method includes collecting multi-source heterogeneous data from the dam seepage monitoring system and labeling the structural domain according to the data source and structural attributes.

[0007] Perform corresponding feature fusion processing on different types of data according to the structural domain labels to construct a structured feature vector group in a unified format;

[0008] The feature vector group is input into the structure-driven expert model, which activates the bound expert paths according to the structural domain labels and performs multi-task reasoning;

[0009] The multi-task reasoning results are input into the dam seepage virtual twin system, driving the simulation module to simulate the seepage diffusion evolution and generate a spatial risk map.

[0010] Generate early warning signals and emergency plan texts based on risk maps and multi-task reasoning results.

[0011] As a preferred solution of the dam seepage intelligent monitoring and risk warning method based on the expert large model described in the present invention, wherein: the multi-source heterogeneous data collected from the dam seepage monitoring system and the structural domain labels marked according to the data source and structural attributes include:

[0012] After collecting various monitoring data, a structural domain identification submodule is set up to build a mapping table between data source type and data format, judge the structural attributes of the input data, and embed structural domain labels in the data structure metadata;

[0013] The domain labels are the input data for feature fusion processing and expert large models.

[0014] As a preferred solution of the dam seepage intelligent monitoring and risk warning method based on the expert large model described in the present invention, wherein: the corresponding feature fusion processing is performed on different types of data according to the structural domain label to construct a structured feature vector group in a unified format, including:

[0015] According to the structural domain label value of each input sample, the corresponding data processing flow template is matched.

[0016] Each process template defines the structural transformation, time or space alignment strategy required for the corresponding category data to enter the unified feature space. The feature vector generated after processing has a unified data format, dimensional structure and position mark, and the data corresponding to different structural domain labels are dimensionally aligned and semantically mapped before being input into the expert model.

[0017] As a preferred solution of the dam seepage intelligent monitoring and risk warning method based on the expert macro model described in the present invention, wherein: the inputting of the feature vector group into the structure-driven expert macro model, activating the bound expert path according to the structural domain label and performing multi-task reasoning includes:

[0018] The expert model establishes a binding relationship structure between the domain labels and the expert paths during the initialization phase, indicating a set of expert paths corresponding to each type of domain label;

[0019] Each set of expert paths includes a sub-model module designed for specific structural attributes, with an internal feature modeling structure based on the attention mechanism to process the data features represented by the binding labels;

[0020] When the fused feature vector group is input into the expert model, the gated routing mechanism parses the structural domain label carried by the current vector and selectively activates the corresponding expert path set based on the binding relationship structure;

[0021] The activated expert path models the received structural vector and outputs the intermediate semantic expression result of the task;

[0022] The output results are sent to the multi-task fusion module inside the expert model, and vector weighting and target path selection are performed through task adaptation control logic to generate semantic input vectors for independent processing by each task channel.

[0023] As a preferred solution of the dam seepage intelligent monitoring and risk warning method based on the expert large model described in the present invention, the multi-task fusion module includes generating a plurality of semantic vector outputs according to the task adaptation control logic, each output vector is processed by a different task channel;

[0024] The percolation state prediction output channel generates state labels based on the classification structure and adopts a normalized classification strategy to construct the task boundary;

[0025] The anomaly detection output channel generates anomaly probability values ​​through a continuous scoring mechanism, and the output format is a single variable probability vector with a settable threshold;

[0026] The fault type determination output channel generates a multi-dimensional classification vector based on multi-class label logic to represent the response intensity of each type of fault;

[0027] The risk scoring channel performs risk level mapping processing on the structural semantic vector and generates continuous values ​​or interval level values ​​as strength references.

[0028] As a preferred solution of the dam seepage intelligent monitoring and risk warning method based on the expert large model described in the present invention, wherein: the multi-task reasoning results are input into the dam seepage virtual twin system, the simulation module is driven to simulate the seepage diffusion evolution, and the spatial risk map is generated, including:

[0029] The seepage state prediction labels and fault type determination results output by the expert large model are used as simulation condition inputs. The corresponding dam structural parameters, material permeability characteristics and monitoring layout information are loaded, and a spatial model is constructed in combination with GIS geographic data. A three-dimensional model including the coupling relationship of dam body-foundation-seepage field is constructed. The seepage pressure disturbance boundary is set according to the abnormal area indicated by the prediction results, and the diffusion evolution calculation based on the finite volume method or the dual seepage model is performed. The three-dimensional dynamic risk map including risk heat, propagation path and time window is output to characterize the evolution process of seepage risk in spatial and temporal dimensions.

[0030] As a preferred solution of the dam seepage intelligent monitoring and risk warning method based on the expert large model described in the present invention, the method of generating warning signals and emergency plan texts based on the risk map and multi-task reasoning results includes:

[0031] The diffusion rate, heat gradient and affected area extracted from the three-dimensional dynamic risk map of the virtual twin system are used to

[0032] The anomaly detection probability and risk score results output by the expert large model are integrated to construct a multi-dimensional risk factor input vector;

[0033] The risk factors are input into the hierarchical fuzzy inference model based on expert rules, and the fitness of each risk level is calculated through the membership function;

[0034] The current risk level is determined according to the maximum membership principle, and an information package containing the risk level, key impact points, propagation scope and disposal suggestions is output. It is then provided as input to the emergency plan text generation module to generate a structured response strategy text.

[0035] Another object of the present invention is to provide a dam seepage intelligent monitoring and risk warning system based on an expert large model. The present invention can integrate multi-source heterogeneous monitoring data, construct a structure-driven expert reasoning mechanism, and combine virtual twin simulation and hierarchical fuzzy reasoning technology to realize multi-task intelligent identification and risk level judgment of the dam seepage state, thereby solving the problems of insufficient data fusion capability, lack of structural adaptability of model response, low accuracy of risk evolution simulation, and insufficient support for early warning information decision-making in the existing technology.

[0036] As a preferred solution of the dam seepage intelligent monitoring and risk early warning system based on the expert large model described in the present invention, it is characterized by including a structural domain label recognition and feature fusion module, a structural domain design submodule, a multi-task fusion module, a driving simulation module, and an emergency plan text generation module;

[0037] The structural domain label recognition and feature fusion module collects multi-source heterogeneous data from the dam seepage monitoring system, identifies the structural domain labels according to the data source and structural attributes, and constructs a structured feature vector group in a unified format according to the label matching and fusion processing flow;

[0038] The structural domain design submodule is a component of the expert model. It receives a feature vector group with structural domain labels and activates the submodel path bound to it according to the label to complete the preliminary reasoning of feature modeling and semantic expression.

[0039] The multi-task fusion module aggregates the semantic vectors output by each sub-model module, generates corresponding task input vectors according to the preset task adaptation mechanism, and outputs multi-task reasoning results including seepage state prediction, anomaly scoring, fault classification and risk scoring;

[0040] The driving simulation module receives prediction labels related to structural anomalies from the multi-task reasoning results, builds a spatial model by combining dam structural parameters with GIS geographic information, and performs seepage pressure disturbance settings and diffusion evolution calculations based on the finite volume method or dual seepage model.

[0041] The emergency plan text generation module integrates the spatial indicators in the three-dimensional dynamic risk map with the reasoning results to construct a multidimensional risk factor vector, performs fuzzy reasoning and risk level judgment based on expert rules, and automatically generates structured warning information and emergency response text.

[0042] A computer device includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the method for intelligent monitoring and risk warning of dam seepage based on an expert large model are implemented.

[0043] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a method for intelligent monitoring and risk warning of dam seepage based on an expert large model.

[0044] The present invention has the following beneficial effects: It can structuredly fuse multi-source heterogeneous monitoring data and perform multi-task identification by binding expert paths, improving the model's accuracy in determining seepage status and identifying faults. Combined with the risk map and fuzzy reasoning module generated by the virtual twin system, the system can dynamically simulate seepage diffusion trends and provide graded early warnings. This effectively addresses the shortcomings of traditional methods, such as weak structural adaptability, insufficient risk evolution expression, and delayed early warning response processes. It improves the intelligence level of dam seepage monitoring and the timeliness of risk management. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0046] Figure 1 The present invention provides an overall flow chart of a method for intelligent monitoring and risk warning of dam seepage based on an expert large model according to an embodiment of the present invention.

[0047] Figure 2 A schematic diagram of a structure-driven hybrid expert model for a dam seepage intelligent monitoring and risk early warning method based on an expert large model provided by one embodiment of the present invention.

[0048] Figure 3This is a performance display diagram of four expert models of a dam seepage intelligent monitoring and risk early warning method based on an expert large model provided by an embodiment of the present invention.

[0049] Figure 4 This is a rendering of the seepage diffusion simulation and risk visualization process of a dam seepage intelligent monitoring and risk early warning system based on an expert large model provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0050] To make the above-mentioned objects, features, and advantages of the present invention more clearly understood, the following detailed description of the specific embodiments of the present invention is given in conjunction with the accompanying drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in this field without creative work should fall within the scope of protection of the present invention.

[0051] Example 1, with reference to Figures 1 and 2 , which is the first embodiment of the present invention, provides a dam seepage intelligent monitoring and risk early warning method based on an expert large model, comprising:

[0052] S1. Collect multi-source heterogeneous data from the dam seepage monitoring system and annotate the structural domain labels according to the data source and structural attributes.

[0053] After collecting various monitoring data, a structural domain identification submodule is set up to build a mapping table between data source type and data format, judge the structural attributes of the input data, and embed structural domain labels in the data structure metadata;

[0054] The domain labels are the input data for feature fusion processing and expert large models.

[0055] Collect multi-source heterogeneous data from the dam seepage monitoring system, including time series sensor data, spatial structural response data, environmental interference variables and historical fault knowledge data in the form of graphs.

[0056] S2. Perform corresponding feature fusion processing on different types of data according to the structural domain labels to construct a structured feature vector group in a unified format.

[0057] According to the structural domain label value of each input sample, the corresponding data processing flow template is matched.

[0058] Each process template defines the structural transformation, time or space alignment strategy required for the corresponding category data to enter the unified feature space. The feature vector generated after processing has a unified data format, dimensional structure and position mark, and the data corresponding to different structural domain labels are dimensionally aligned and semantically mapped before being input into the expert model.

[0059] The feature fusion processing in the preferred embodiment of the present invention is as follows: the collected data are identified according to their source structural features, and structural domain labels are generated, which are marked as temporal class (T), spatial class (S), environmental class (E) and atlas class (G);

[0060] Table 1. Samples of collected data and structural domain classification

[0061]

[0062]

[0063] Description of the structural domain label classification rules.

[0064] [T-type - Time Series Data]: If the data contains "timestamp + sensor number + single variable or multivariate value", it is marked as T (such as seepage pressure, pore pressure, flow rate).

[0065] [Class S - Spatial structure data]: If the data contains "spatial coordinates (x, y, z) + node number + displacement or stress tensor", it is marked as S.

[0066] [Category E - Environmental variables]: If the data source is "external natural environment", such as meteorology, hydrology, etc., and contains time stamps and interference factor values ​​(temperature, rainfall, water level, etc.), it is marked as E.

[0067] [G-Graph Knowledge Data]: If the data is described in triples or graph relational structures, or references historical failure cases and expert knowledge, it is marked as G.

[0068] According to the structural domain labels, various types of data are sent to the corresponding preset fusion path, and time series normalization, spatial interpolation, conditional standardization and graph embedding are performed respectively to obtain a feature vector group in a unified format, which constitutes a feature fusion data set of structural grouping.

[0069] In an optional embodiment of the feature fusion process, rather than performing differentiated strategies such as normalization, interpolation, and standardization based on the structural domain labels, a unified compression fusion path is employed. In this approach, after labeling the structural domain labels, the system does not classify the data with different labels and feed them into a pre-defined fusion process. Instead, the data corresponding to all structural domain labels is uniformly mapped into vector form and then directly concatenated to form a multi-source merged hybrid input feature set.

[0070] To extract semantic features, this solution deploys a shared fusion encoding network to process a unified input feature set and generate a fixed-length structured representation. This fusion network can be based on a multi-layer perceptron architecture (such as an MLP) or a shallow self-attention mechanism to extract compressed semantic vectors for subsequent model input. The entire processing process disregards the original structural properties of the data; all structural domain features are uniformly encoded and fed into subsequent modules. Expert path routing driven by structural labels is no longer performed in the system.

[0071] Although the use of a unified compression fusion strategy in the optional implementation of the present invention can formally simplify the processing path and reduce the number of fusion modules, it still has obvious shortcomings in the actual seepage monitoring scenario. Since the data of all structural domains are uniformly spliced ​​and encoded in the fusion stage and handed over to a common shared network for feature extraction, this method does not differentiate the structural attributes of data from different sources, resulting in a serious weakening of the original structural semantics. There are significant differences in physical meaning and information density between time series data and spatial data. If isomorphic compression is performed indiscriminately, it is easy to cause an imbalance in the weight of feature expression. In particular, in seepage tasks, key indicators may be weakened by the compression process, resulting in a decrease in subsequent reasoning accuracy.

[0072] More critically, this alternative bypasses the domain label-driven expert path selection mechanism, failing to achieve a one-to-one match between model structure and data structure. This results in a lack of targeted expert path activation, rendering the core structure-driven mechanism of the present invention ineffective and weakening the interpretability and adaptability of the expert model. In experimental testing, the model using this fusion strategy significantly underperformed the domain-based path selection approach in multi-task collaborative performance. In particular, the accuracy of the two tasks, anomaly detection and fault classification, declined significantly, and the task consistency scores fluctuated significantly, making it difficult to meet the dual requirements of stability and accuracy in practical engineering.

[0073] In contrast, the preferred solution described in the present invention makes full use of structural domain labels to perform path-by-path fusion processing on data. Different structural domain types of data adopt customized normalization, interpolation, standardization or graph embedding strategies to ensure that each type of feature has a unified dimensional structure and clear semantic boundaries before entering the expert model. This mechanism not only retains the structural information of the data, but also achieves precise binding with the expert path, improving the model's ability in structural adaptability and task expression consistency. Through a structure-driven routing mechanism and a multi-task decoupled output method, the preferred solution of the present invention demonstrates higher accuracy and stability in key tasks such as seepage state identification and risk level judgment.

[0074] Therefore, from the perspectives of structural matching accuracy, feature expression capability, and engineering applicability, the present invention recommends adopting the structural domain-driven fusion path as the preferred implementation strategy, while unified compression fusion is only an optional but non-priority processing method.

[0075] S3. Input the feature vector group into the structure-driven expert model, activate the bound expert path according to the structural domain label and perform multi-task reasoning.

[0076] The expert model establishes a binding relationship structure between the domain labels and the expert paths during the initialization phase, indicating a set of expert paths corresponding to each type of domain label;

[0077] Each set of expert paths includes a sub-model module designed for specific structural attributes, with an internal feature modeling structure based on the attention mechanism to process the data features represented by the binding labels;

[0078] When the fused feature vector group is input into the expert model, the gated routing mechanism parses the structural domain label carried by the current vector and selectively activates the corresponding expert path set based on the binding relationship structure;

[0079] The activated expert path models the received structural vector and outputs the intermediate semantic expression result of the task.

[0080] The output results are sent to the multi-task fusion module inside the expert model, and vector weighting and target path selection are performed through task adaptation control logic to generate semantic input vectors for independent processing by each task channel.

[0081] The structure-driven expert big model in the preferred embodiment of the present invention is: the expert big model uses the SD-MoE architecture of the trinity of structure-domain-oriented expert routing mechanism + multi-task decoupling output strategy + graph enhancement path.

[0082] Its underlying logic is to route input data to bound expert clusters based on their structural domain labels (such as T: time series, S: space, G: graph). Each type of expert network embeds a set of independent sub-modules in the Transformer backbone to extract data features of their respective expertise. The MoE module determines the activation path through the prior structure label + distribution-aware gating mechanism, while the multi-task head (Mix-of-Head) structure corresponds to output tasks such as seepage state prediction, anomaly detection, fault classification, and risk scoring. The model also introduces a graph expert path and uses a graph neural network to introduce graph structure latent variables into the output representation, thereby achieving cross-modal fusion + multi-task collaboration + structural alignment optimization.

[0083] Finally, through multiple rounds of joint fine-tuning, the model is endowed with domain generalization capabilities and precise expression capabilities.

[0084] like Figure 2First, the system sends multi-source heterogeneous data into the corresponding preset fusion path according to the structural domain labels generated by the structure recognition module (T: time series, S: space, E: environmental disturbance, G: graph knowledge), and performs customized feature preprocessing: T-type data is time-series normalized, S-type data is spatially interpolated, E-type data is standardized, and G-type data is graph-embedded through a graph neural network (GAT) to uniformly generate a standardized group of structured feature vectors.

[0085] These feature vectors are then fed into a structure-driven MoE module built on a Transformer backbone. The MoE first parses and matches existing structural labels through a gating mechanism, accurately activating the corresponding expert clusters (such as T-domain experts, S-domain experts, and G-domain experts). Each expert cluster contains an independent Transformer submodule that performs self-attention calculations, feed-forward neural layer activations (FFN), feature compression, and cross-channel calculations on the input vector, outputting its own unique semantic vector.

[0086] After the output of multiple expert paths, they enter the multi-task fusion module, which will construct a "task fitness matrix" to perform weighted aggregation on the expert outputs: that is, dynamically weighting the results of different experts according to the current task goals, so that the prediction results can fully combine spatiotemporal characteristics, historical knowledge and environmental disturbances, thereby achieving synergistic enhancement of cross-domain information.

[0087] Based on the unified feature vector output by the multi-task fusion module, each of the four task heads employs a dedicated neural architecture to perform feature decoupling and target alignment on the input, achieving targeted output. The percolation state prediction head employs a cross-entropy loss for classification modeling, outputting a softmax vector representing the probability distribution of "normal-abnormal" categories. The category corresponding to the maximum value is the predicted label. The anomaly detection head continuously scores the degree of anomaly using sigmoid activation and outputs a threshold-determinable anomaly probability. The fault type determination head, based on a multi-layer perceptron network with multiple class labels, learns the boundaries between different faults in a structural semantic space and outputs a multidimensional fault probability vector. The category corresponding to the maximum component is the determination result. The preliminary risk scoring head estimates the risk intensity of the fused features using a regression or segmented classification model, outputting the corresponding level or risk score. The four heads share the underlying code but maintain decoupling of the task outputs, enabling the models to maintain their respective predictive accuracy while collaborating on the tasks.

[0088] It should be further explained that:

[0089] The structure-driven hybrid expert model (SD-MoE) proposed in the present invention is constructed with Transformer as the main backbone architecture, integrating the expert cluster activation mechanism controlled by structural domain labels and the multi-task decoupling module, and introducing a GNN enhancement path specifically for graph data. First, during construction, by analyzing the structural properties of multi-source data (such as whether it is time series, space, graph or environmental perturbation), the system will label each type of input with a structural domain label (T / S / G / E). These labels are input into the gating network as prior variables to control the expert activation path inside the MoE module. Each type of structural data will only activate the expert cluster bound to it, thereby ensuring the consistency of data-expert-task. During operation, the input data is first formed into a structured vector after feature fusion, and enters the bound expert cluster according to the structural label; the Transformer submodule inside each expert cluster performs self-attention modeling, feature extraction and transformation, and outputs the intermediate results. Next, the outputs of multiple activated experts are integrated into a multi-task fusion module and then fed in parallel into decoupled task head modules (corresponding to percolation prediction, anomaly detection, fault identification, and risk scoring, respectively). This ultimately completes the refined reasoning and joint output of each task. The entire model's operational logic embodies a layered, progressive architecture: structural perception → expert activation → multi-task decoupling → output fusion.

[0090] In the expert model of this invention, the multi-task fusion module is a critical link in the overall structure, connecting the upstream expert outputs with the downstream predictions for specific tasks. Because the four tasks of seepage state prediction, anomaly detection, fault type determination, and risk scoring each focus on different perspectives, it's not possible to simply distribute all expert outputs directly to each task. Instead, an intelligent judgment mechanism is required to determine which experts are best suited for each task and how much information should be allocated. This is precisely what the multi-task fusion module accomplishes.

[0091] When multiple expert paths are activated, each expert provides a representation of their understanding of the input, which we call an "expert output vector." These vectors essentially represent information from different perspectives—some specialize in temporal changes, some focus on spatial deformations, and some are rooted in graph knowledge. At this point, the multi-task fusion module acts as a "dispatching control center." Rather than blindly averaging the expert advice, it weights and selects the information provided by each expert based on the specific requirements of the task.

[0092] To achieve this "demand-based allocation," the system maintains an internal "expert-task association table," which can be thought of as a database of accumulated experience. This table isn't statically configured; instead, it continuously learns and adjusts during training. For example, if spatial experts perform better at anomaly detection across multiple historical tasks, the system will gradually learn to prioritize their input when handling such tasks. Conversely, graph experts often provide a reliable reference for risk scoring, and the system will increase their influence on risk scoring tasks.

[0093] The entire fusion process is essentially like a multi-round consultation, with different experts expressing their opinions. The fusion module then assigns weights based on the experts' previous credibility in a particular task type, ensuring that each task head receives the most appropriate comprehensive advice. More importantly, this fusion is not a one-time process, but rather "adaptive": it automatically adjusts as the task scenario changes. For example, if environmental variables suddenly become dominant in a monitoring task, the fusion module will "detect" this change and shift more attention to the environmental expert's output. Similarly, in another fault identification task, if the atlas expert performs exceptionally well, the system will automatically increase its absorption of their output.

[0094] After fusion is complete, each task head receives not a jumble of chaotically overlapping data but a customized task input. This input, deeply integrated with the opinions of multiple experts, carries a clear semantic objective, ultimately enabling the task head to make more accurate judgments, such as predicting the presence of an anomaly, determining the type of fault, and assigning a risk level score.

[0095] This design allows the entire large model to maintain a unified architecture while demonstrating adaptability and judgment for each subtask, improving the practicality and interpretability of the model in complex, multi-dimensional task scenarios.

[0096] The core difference between this method and general expert models lies in the introduction of structural domain labels as prior information for expert path selection during the model design phase, eliminating the reliance on single feature similarity for soft routing decisions. Each label type is explicitly associated with a set of expert clusters, ensuring semantic consistency in the processing flow and interpretability of path selection. This approach is particularly suitable for environments with distinct data types, such as engineering structural scenarios, and improves both the specificity and generalization capabilities of the model.

[0097] An optional embodiment of the structure-driven expert model is a lower-level embodiment based on the Soft-MoE and task adapter structure. Its design logic is as follows:

[0098] Construct multiple general Transformer expert submodules that are not bound to specific domain labels;

[0099] Using the softmax gating mechanism for expert activation, each input data will simultaneously activate multiple expert paths and weight their outputs by probability;

[0100] The output features are uniformly fed into the task Adapter module. Different tasks (percolation state prediction, anomaly detection, fault identification, and risk scoring) are configured with their own lightweight Adapter structure to achieve target-specific mapping.

[0101] The optional embodiment does not introduce structural tags or build binding paths. Its design simplifies the model structure and has certain application value in scenarios with limited computing resources or low task complexity. However, this solution has the following problems:

[0102] The structural attributes of the input data are not perceived, resulting in inconsistency between the expert path and the semantics, which easily causes semantic information interference; the general expert path performs unevenly on different tasks, and the prediction accuracy fluctuates greatly, especially in graph knowledge tasks. The effect is significantly reduced; the adapter has a fixed output dimension and weak channel decoupling ability, making it difficult to adapt to the output heterogeneity between complex tasks, especially in multi-task cross-coupling scenarios. It is prone to failure.

[0103] Therefore, although Soft-MoE+Adapter has the characteristics of easy deployment and lightweight reasoning, it has significant deficiencies in the ability to perceive multi-source heterogeneous data structures, task expression accuracy, and result path independence, making it difficult to meet the needs of intelligent monitoring of dam seepage under complex working conditions.

[0104] S4. Input the multi-task reasoning results into the dam seepage virtual twin system, drive the simulation module to simulate the seepage diffusion evolution, and generate a spatial risk map.

[0105] The percolation state prediction output channel generates state labels based on the classification structure and adopts a normalized classification strategy to construct the task boundary;

[0106] The anomaly detection output channel generates anomaly probability values ​​through a continuous scoring mechanism, and the output format is a single variable probability vector with a settable threshold;

[0107] The fault type determination output channel generates a multi-dimensional classification vector based on multi-class label logic to represent the response intensity of each type of fault;

[0108] The risk scoring channel performs risk level mapping processing on the structural semantic vector and generates continuous values ​​or interval level values ​​as strength references.

[0109] After receiving the state prediction labels and fault type determination results from the expert large-scale model, the system inputs them as trigger conditions into the dam seepage virtual twin system, initiating a 3D physical simulation engine. The simulation model first retrieves basic information such as the dam's structural geometry, material permeability, dam body hierarchical structure, and monitoring deployment locations. This information is then spatially mapped and integrated with GIS geographic information to construct a complete coupled model of the dam body, foundation, and seepage field. Subsequently, based on the abnormal areas or fault nodes identified in the prediction results, the system sets the initial seepage pressure anomaly source and boundary perturbation conditions. The seepage path is gradually evolved using the finite volume method or a dual seepage model (e.g., saturated-unsaturated flow). During the calculation, the propagation direction of hydraulic head potential energy, pressure gradient, and seepage stress are tracked in real time, dynamically recording the diffusion trajectory and impact area of ​​high-risk pathways. Finally, the system outputs a 3D dynamic risk map with time as the horizontal axis, spatial location as the vertical axis, and seepage pressure heat as the color channel. This map clearly depicts the evolutionary trends and potential spread areas of seepage risk at different time scales, providing refined spatial support for subsequent risk response and emergency dispatch.

[0110] S5. Generate warning signals and emergency plan texts based on the risk map and multi-task reasoning results.

[0111] In the intelligent early warning phase, the system first semantically aligns and fuses the anomaly detection probabilities and preliminary risk scores from the expert large-scale model with the spatial risk map generated by the virtual twin simulation to construct a multidimensional risk factor input vector. This fusion process uses a regional indexing mechanism to weightedly combine spatial risk indicators (such as diffusion rate, heat gradient, and affected area) for high-heat regions in the map with the anomaly probability and score values ​​at the corresponding locations, forming a composite risk representation with spatial awareness. The system then introduces a hierarchical fuzzy inference model (existing technology), using multiple expert rule bases as the inference core. This model quantifies the uncertainty and fuzzy boundaries in the fused vector, calculates the suitability of each risk level (G1-G4) in the current scenario through membership functions, and implements multi-rule superposition and conflict resolution within the inference network. Finally, the system outputs the most suitable risk level based on the maximum membership decision principle and automatically generates a multi-level warning information package containing the warning level, coordinates of key impact points, predicted propagation range, and disposal recommendations (such as enhanced monitoring and preliminary drainage) to support intelligent scheduling and response deployment. The entire process maintains the unified integration of spatial modeling, model output and expert experience, improving the accuracy and operability of risk warnings.

[0112] On the morning of July 12, 2024, the system detected an abnormal rise at the seepage pressure sensor P123 in the right dam abutment area. The expert large model output an abnormal probability of 0.87, and the fault type was determined to be "possible development of seepage ditch", with a risk score of 3.2 / 5.

[0113] The virtual twin system then conducted a spatial diffusion simulation of the right dam abutment area. The output map showed the existence of a high-heat seepage path in the area with a maximum diffusion length of 22 meters, which is expected to be transmitted to the downstream monitoring well within the next three hours.

[0114] The system integrates risk characteristics from the following three aspects:

[0115] Anomaly probability (0.87), risk score (3.2) from the expert model;

[0116] Spatial indicators extracted from the atlas: influence radius, diffusion heat, and osmotic pressure gradient;

[0117] Similarity matching degree of historical fault graphs of the same type.

[0118] The resulting eigenvector is: [abnormal probability = 0.87, risk score = 3.2, heat mean = 0.74, spatial diffusion length = 22m, historical similarity = 0.68]

[0119] S5 step 2 of the present invention: fuzzy reasoning and maximum membership decision

[0120] The above vector is input into the hierarchical fuzzy inference system and reasoning is performed through the preset rule base, such as:

[0121] If the probability of abnormality is high and the spread is large → it belongs to G3 level;

[0122] If the heat is concentrated but the probability of abnormality is medium → it belongs to G2 level;

[0123] If the anomaly probability is extremely high + the historical pattern match is strong → it belongs to G4 level.

[0124] The system calculates the membership of each level as follows (illustrative):

[0125] G1: 0.05

[0126] G2: 0.19

[0127] G3: 0.64

[0128] G4: 0.47

[0129] According to the principle of maximum membership, the final output is a G3 level warning.

[0130] Based on the G3 level rule template and combined with the twin graph output, the system automatically constructs the following warning information:

[0131] {

[0132] "Warning Level":"G3",

[0133] "Key Influencing Point": "Right Dam Abutment - Coordinates (34.82, 121.17)",

[0134] "Predicted spread range": "Currently 22m, expected to spread to within 34m, time window: next 3 hours",

[0135] "Disposal suggestion":[

[0136] "Schedule and increase the number of monitoring points (shorten the spacing to 5m)",

[0137] "Dredge the lateral drainage ditches of the dam to release pressure",

[0138] "Arrange a structural diagnosis team to conduct on-site inspections to detect the development path of the seepage ditch." ]

[0140] }

[0141] The information package is pushed to the dispatch system and mobile terminals to support automated plan linkage and dispatch response.

[0142] Based on the generated early warning signal, the natural language generation module is triggered to generate a structured multi-scenario emergency plan text containing response strategies, disposal processes and scheduling nodes according to the current risk level, regional impact range and preset emergency rule template;

[0143] During the natural language emergency plan generation phase, the system first outputs a warning signal as the primary control input. This signal contains key information such as the risk level (e.g., G3), spatial impact range, key node locations, and time prediction window. Based on this information, the natural language generation module retrieves a library of emergency rule templates that matches the current risk level. This library predefines response strategies, task responsibilities, and action node structures for different risk levels. The system then uses the Prompt orchestration module to embed the warning signal into the template's placeholder fields. For example, it maps "right abutment, G3 level, diffusion radius 22m" to fields such as "response area," "action priority," and "dispatch path." The system then selects an appropriate language mode (e.g., imperative, descriptive, or tabular) based on the scenario context, driving a language modeler (large-scale models such as ChatGPT and DeepSeek) to generate a complete, structured plan text. During the generation process, the system simultaneously calls the scheduling mapping table to embed the nodes of the involved scheduling units, such as the monitoring group, drainage group, and structural diagnosis group, into the process. The output text not only describes the emergency strategy and steps but also clearly defines the responsible parties and execution timelines for each link. Ultimately, the multi-scenario emergency plan text is encoded into a structured format (such as JSON or a table) and can be sent to the dispatch platform, duty system, and mobile terminals to achieve readable, callable, and interactive linkage response support.

[0144] The platform records users' responses to generated plans and feedback on actual event developments, converts response deviations, handling efficiency, and feedback label data into enhanced samples, and uses them for fine-tuning expert large models and filling in virtual twin system parameters to complete closed-loop optimization of the intelligent system.

[0145] Example 2, reference Figures 3 and 4 , which is the second embodiment of the present invention, provides a dam seepage intelligent monitoring and risk warning method based on an expert large model. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through experiments.

[0146] This experiment aims to verify the advantages of the structure-driven hybrid expert model (SD-MoE) proposed in this invention in processing multi-source heterogeneous monitoring data of dam seepage and performing multiple tasks (seepage state prediction, anomaly detection, fault identification, and risk scoring) in key indicators such as accuracy, collaborative efficiency, interpretability, and operational efficiency, and to objectively compare it with current mainstream solutions.

[0147] This experiment constructed a simulated dam multi-source seepage monitoring dataset, which includes the following data:

[0148] Seepage pressure time series data: 4500 (simulated working condition data recorded at a frequency of minutes);

[0149] Spatial displacement data: 2300 items (3D structure node displacement information);

[0150] Environmental disturbance data: 3100 items (rainfall, water level and other environmental variables);

[0151] Graph knowledge data: 1,200 items (historical accident rules, causal graphs);

[0152] Label information: Each sample contains four types of labels: seepage status, abnormal label, fault type and risk level (G1-G4).

[0153] 1. Comparison Model:

[0154] SD-MoE (this invention): structured label controlled routing + expert binding + multi-task fusion + graph injection.

[0155] Universal MoE: traditional softmax gating + shared expert path.

[0156] MoE+Adapter: Superimposes the task decoupling module on the general MoE.

[0157] Single-task Baseline: A deep model trained separately for each task.

[0158] 2. Task definition:

[0159] Task 1: Seepage state prediction (classification); Task 2: Anomaly detection (F1 evaluation); Task 3: Fault identification (multi-class classification); Task 4: Risk scoring (regression, G1-G4)

[0160] 3. Evaluation indicators:

[0161] Accuracy, F1 value, RMSE: measure task performance.

[0162] TCS (Task Consistency Score): Assess the degree of coordination between tasks.

[0163] Inference time: average inference time per batch of data (seconds).

[0164] E-Score: Expert path interpretability score.

[0165] Table 2 Experimental results and comparison table

[0166]

[0167] Reference Figure 3 The performance of the structure-driven expert model (SD-MoE) of the present invention in four core tasks is compared with three mainstream models (general MoE, MoE+Adapter, and single-task Baseline). The four sub-graphs correspond to: seepage state prediction accuracy, anomaly detection F1 score, fault identification accuracy, and risk score stability. It can be clearly observed from the figure that SD-MoE maintains performance advantages in all tasks, especially in fault identification and anomaly detection with higher task complexity, its accuracy and robustness far exceed other models. This shows that the structural domain labeling mechanism and expert routing strategy of SD-MoE play a key role in improving the consistency of task expression and model generalization ability, and can more effectively adapt to the distribution characteristics of multi-source monitoring data of dams and realize accurate collaborative reasoning of multiple tasks.

[0168] Reference Figure 4The path heat map shows that the expert path selection of SD-MoE has a structure-expert-task chain consistency, while the general MoE path activation distribution is scattered and lacks logic. In the present invention, the virtual twin module combines the expert large model to perform the overall process of seepage diffusion simulation and risk visualization. On the left is a cross-sectional simulation diagram, which simulates the changes in different seepage paths (streamlines) and total head equipotential lines, and marks the response of high, medium and low risk areas after the water level rises; on the right is a three-dimensional risk heat map based on the simulation output, showing the changes in risk level with the diffusion evolution process in different time (Time) and space (Distance) dimensions. Colors from blue to red represent low to extremely high risk levels. It can be observed that the risk hotspot area gradually moves downstream of the dam over time. This result verifies the accuracy and timeliness of the present invention in risk path modeling and dynamic early warning visualization, and supports the closed-loop risk prevention and control goal of "from prediction to deduction to operational early warning."

[0169] Example 3 is the third embodiment of the present invention, which differs from the first two embodiments in that:

[0170] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0171] The logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing the logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (e.g., a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device). For purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0172] More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection with one or more wires (electronic devices), a portable computer disk cartridge (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), a fiber optic device, and a portable compact disc read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering, or processing in another suitable manner as necessary, and then stored in a computer memory.

[0173] It should be understood that various parts of the present invention can be implemented using hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one or combination of the following technologies known in the art can be used: a discrete logic circuit having logic gate circuits for implementing logic functions on data signals, an application-specific integrated circuit having suitable combinational logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0174] Example 4, reference Figure 4 , which is the fourth embodiment of the present invention, provides a dam seepage intelligent monitoring and risk warning system based on an expert large model, including a structural domain label recognition and feature fusion module, a structural domain design submodule, a multi-task fusion module, a driving simulation module, and an emergency plan text generation module;

[0175] The structural domain label recognition and feature fusion module collects multi-source heterogeneous data from the dam seepage monitoring system, identifies the structural domain labels according to the data source and structural attributes, and constructs a structured feature vector group in a unified format according to the label matching and fusion processing flow;

[0176] The structural domain design submodule is a component of the expert model. It receives a feature vector group with structural domain labels and activates the submodel path bound to it according to the label to complete the preliminary reasoning of feature modeling and semantic expression.

[0177] The multi-task fusion module aggregates the semantic vectors output by each sub-model module, generates corresponding task input vectors according to the preset task adaptation mechanism, and outputs multi-task reasoning results including seepage state prediction, anomaly scoring, fault classification and risk scoring;

[0178] The driving simulation module receives prediction labels related to structural anomalies from the multi-task reasoning results, builds a spatial model by combining dam structural parameters with GIS geographic information, and performs seepage pressure disturbance settings and diffusion evolution calculations based on the finite volume method or dual seepage model.

[0179] The emergency plan text generation module integrates the spatial indicators in the three-dimensional dynamic risk map with the reasoning results to construct a multidimensional risk factor vector, performs fuzzy reasoning and risk level judgment based on expert rules, and automatically generates structured warning information and emergency response text.

[0180] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A dam seepage intelligent monitoring and risk early warning method based on an expert large model, characterized by: include, Collect multi-source heterogeneous data from the dam seepage monitoring system and annotate the structural domain labels according to the data source and structural attributes; Perform corresponding feature fusion processing on different types of data according to the structural domain labels to construct a structured feature vector group in a unified format; The feature vector group is input into the structure-driven expert model, which activates the bound expert paths according to the structural domain labels and performs multi-task reasoning; The multi-task reasoning results are input into the dam seepage virtual twin system, driving the simulation module to simulate the seepage diffusion evolution and generate a spatial risk map. Generate early warning signals and emergency plan texts based on risk maps and multi-task reasoning results.

2. The method for intelligent monitoring and risk early warning of dam seepage based on an expert large model according to claim 1, characterized in that: The method of collecting multi-source heterogeneous data from the dam seepage monitoring system and labeling the structural domain labels according to the data source and structural attributes includes: After collecting various monitoring data, a structural domain identification submodule is set up to build a mapping table between data source type and data format, judge the structural attributes of the input data, and embed structural domain labels in the data structure metadata; The domain labels are the input data for feature fusion processing and expert large models.

3. The method for intelligent monitoring and risk early warning of dam seepage based on an expert large model according to claim 2, characterized in that: The method of performing corresponding feature fusion processing on different types of data according to the structural domain labels and constructing a structured feature vector group in a unified format includes: According to the structural domain label value of each input sample, the corresponding data processing flow template is matched. Each process template defines the structural transformation, time or space alignment strategy required for the corresponding category data to enter the unified feature space. The feature vector generated after processing has a unified data format, dimensional structure and position mark, and the data corresponding to different structural domain labels are dimensionally aligned and semantically mapped before being input into the expert model.

4. The method for intelligent monitoring and risk early warning of dam seepage based on an expert large model according to claim 3, characterized in that: The step of inputting the feature vector group into the structure-driven expert model, activating the bound expert path according to the structural domain label and performing multi-task reasoning includes: The expert model establishes a binding relationship structure between the domain labels and the expert paths during the initialization phase, indicating a set of expert paths corresponding to each type of domain label; Each set of expert paths includes a sub-model module designed for specific structural attributes, with an internal feature modeling structure based on the attention mechanism to process the data features represented by the binding labels; When the fused feature vector group is input into the expert model, the gated routing mechanism parses the structural domain label carried by the current vector and selectively activates the corresponding expert path set based on the binding relationship structure; The activated expert path models the received structural vector and outputs the intermediate semantic expression result of the task; The output results are sent to the multi-task fusion module inside the expert model, and vector weighting and target path selection are performed through task adaptation control logic to generate semantic input vectors for independent processing by each task channel.

5. The method for intelligent monitoring and risk early warning of dam seepage based on an expert large model according to claim 4, characterized in that: The multi-task fusion module includes generating several semantic vector outputs according to the task adaptation control logic, each output vector is processed by a different task channel; The percolation state prediction output channel generates state labels based on the classification structure and adopts a normalized classification strategy to construct the task boundary; The anomaly detection output channel generates anomaly probability values ​​through a continuous scoring mechanism, and the output format is a single variable probability vector with a settable threshold; The fault type determination output channel generates a multi-dimensional classification vector based on multi-class label logic to represent the response intensity of each type of fault; The risk scoring channel performs risk level mapping processing on the structural semantic vector and generates continuous values ​​or interval level values ​​as strength references.

6. The method for intelligent monitoring and risk early warning of dam seepage based on an expert large model according to claim 5, characterized in that: The multi-task reasoning results are input into the dam seepage virtual twin system to drive the simulation module to simulate the seepage diffusion evolution and generate a spatial risk map, including: The seepage state prediction labels and fault type determination results output by the expert large model are used as simulation condition inputs. The corresponding dam structural parameters, material permeability characteristics and monitoring layout information are loaded, and a spatial model is constructed in combination with GIS geographic data. A three-dimensional model including the coupling relationship of dam body-foundation-seepage field is constructed. The seepage pressure disturbance boundary is set according to the abnormal area indicated by the prediction results, and the diffusion evolution calculation based on the finite volume method or the dual seepage model is performed. The three-dimensional dynamic risk map including risk heat, propagation path and time window is output to characterize the evolution process of seepage risk in spatial and temporal dimensions.

7. The method for intelligent monitoring and risk early warning of dam seepage based on an expert large model according to claim 6, characterized in that: The generation of warning signals and emergency plan texts based on risk maps and multi-task reasoning results includes: The diffusion rate, heat gradient and affected area extracted from the three-dimensional dynamic risk map of the virtual twin system are used to The anomaly detection probability and risk score results output by the expert large model are integrated to construct a multi-dimensional risk factor input vector; The risk factors are input into the hierarchical fuzzy inference model based on expert rules, and the fitness of each risk level is calculated through the membership function; The current risk level is determined according to the maximum membership principle, and an information package containing the risk level, key impact points, propagation scope and disposal suggestions is output. It is then provided as input to the emergency plan text generation module to generate a structured response strategy text.

8. A dam seepage intelligent monitoring and risk early warning system based on an expert large model, applying the dam seepage intelligent monitoring and risk early warning method based on an expert large model as claimed in any one of claims 1 to 7, characterized in that: include: Structural domain label recognition and feature fusion module, structural domain design submodule, multi-task fusion module, drive simulation module, and emergency plan text generation module; The structural domain label recognition and feature fusion module collects multi-source heterogeneous data from the dam seepage monitoring system, identifies the structural domain labels according to the data source and structural attributes, and constructs a structured feature vector group in a unified format according to the label matching and fusion processing flow; The structural domain design submodule is a component of the expert model. It receives a feature vector group with structural domain labels and activates the submodel path bound to it according to the label to complete the preliminary reasoning of feature modeling and semantic expression. The multi-task fusion module aggregates the semantic vectors output by each sub-model module, generates corresponding task input vectors according to the preset task adaptation mechanism, and outputs multi-task reasoning results including seepage state prediction, anomaly scoring, fault classification and risk scoring; The driving simulation module receives prediction labels related to structural anomalies from the multi-task reasoning results, builds a spatial model by combining dam structural parameters with GIS geographic information, and performs seepage pressure disturbance settings and diffusion evolution calculations based on the finite volume method or dual seepage model. The emergency plan text generation module integrates the spatial indicators in the three-dimensional dynamic risk map with the reasoning results to construct a multidimensional risk factor vector, performs fuzzy reasoning and risk level judgment based on expert rules, and automatically generates structured warning information and emergency response text.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method for intelligent monitoring and risk warning of dam seepage based on an expert large model according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of a dam seepage intelligent monitoring and risk early warning method based on an expert large model according to any one of claims 1 to 7 are implemented.

Images