Drainage hole clogging prediction and early warning method based on artificial intelligence

Abstract

The application discloses a drainage hole siltation prediction and early warning method based on artificial intelligence, which comprises feature quantization of a slope rock-soil body, data acquisition of a drainage hole, construction of a water force coupling topological graph, construction and training of a graph attention network model, drainage hole siltation prediction, and adaptive early warning threshold setting; in view of the problems of great siltation mechanism difference and strong inter-hole water force coupling effect caused by coexistence of limestone and loose accumulation body, flow, water pressure, and calcium ion concentration or turbidity data corresponding to geological features are collected, a water force coupling topological graph and a double-channel graph attention network model are constructed, a spatial propagation effect of siltation is simulated through dynamic updating of edge conduction coefficients, a siltation risk index, type, and propagation influence matrix are output, and then an adaptive three-level early warning threshold is set, and corresponding maintenance measures are matched; spatial propagation risk quantitative evaluation of different siltation mechanisms under complex geological conditions is realized, and the active operation and maintenance level and safety control ability of the slope drainage system are improved.

Description

Technical Field

[0001] This invention relates to the field of slope engineering drainage technology, and in particular to an artificial intelligence-based method for predicting and warning of drainage hole blockage. Background Technology

[0002] In large-scale slope engineering, dozens or even hundreds of drainage holes are often installed to form a drainage network in order to effectively lower the groundwater level and maintain slope stability. The geological conditions of such slopes are often extremely complex, typically involving the simultaneous presence of limestone and loose deposits within the same large slope site. However, the clogging patterns of drainage holes in limestone areas and loose deposit areas are drastically different: in limestone areas, groundwater is rich in calcium ions, leading to chemical crystallization clogging, primarily caused by calcium carbonate precipitation, at the drainage hole outlet and in nearby fissures; in loose deposit areas, clogging is mainly caused by mechanical deposition of fine-grained sediment, which migrates with seepage and gradually settles in the drainage channels. These two types of clogging differ fundamentally in their formation mechanisms, development rates, and response characteristics to water flow, significantly increasing the difficulty of predicting and identifying drainage hole clogging.

[0003] More importantly, the drainage network is not simply a combination of independently operating individual holes; each drainage hole forms a tightly coupled hydraulic network through the seepage field in the rock fissures of the slope. When a drainage hole at a critical location becomes clogged, its drainage capacity decreases, causing a local rise in water head. This redistributes the seepage field across the entire slope, leading to a surge in flow to upstream or adjacent drainage holes, accelerating siltation, or causing downstream drainage holes to gradually fail due to reduced water pressure. This spatial coupling effect means that the siltation of a single drainage hole can trigger a chain reaction, forming a siltation propagation chain along the fissure connection path, ultimately endangering the drainage capacity of the entire drainage system and the safety of the slope. However, in traditional point-based detection and analysis methods, this spatial coupling effect of the drainage network and the siltation propagation path are almost completely ignored.

[0004] The patent application CN202310278316.6, entitled "A Method for Predicting the State of Municipal Drainage Pipeline Networks Based on Graph and Deep Learning," analyzes the drainage pipeline network system as a hydraulically interconnected whole, using the established graph model and algorithm prediction results to troubleshoot faults and provide flood warnings. However, slope engineering drainage involves complex geological and hydrological conditions, far exceeding those of municipal drainage pipeline networks. Existing siltation assessment methods struggle to simultaneously consider the significant differences in the evolutionary patterns of chemically crystalline siltation and particulate mechanical depositional siltation, and are even less capable of dynamically predicting siltation types, occurrence probabilities, and their cascading effects on adjacent drainage outlets under conditions of coexistence of both. Faced with the complex conditions of coexistence of limestone and loose deposits, and the coupled spatial propagation of siltation mechanisms, traditional methods cannot provide effective early warnings. Often, reactive measures are only taken after localized siltation has developed into regional drainage failure, seriously threatening the long-term safety of large-scale slope engineering projects.

[0005] Therefore, given the different clogging patterns and spatial coupling propagation effects of drainage holes under the coexistence of limestone and loose deposits in the same large slope, there is an urgent need for an intelligent prediction and early warning method that can integrate multi-source monitoring information, reflect the hydraulic coupling topological relationship of the drainage hole network, and dynamically learn the mutual influence between nodes. This method can accurately predict, infer the propagation path, and provide graded early warning for chemical crystallization and particulate mechanical deposition clogging, supporting the proactive operation and maintenance of large slope drainage systems. Summary of the Invention

[0006] To achieve the above-mentioned objectives, this invention provides an artificial intelligence-based method for predicting and warning of drainage hole blockage, applicable to drainage hole networks in large-scale slope engineering projects where limestone and loose deposits coexist. The method includes the following steps: S1. Quantification of Slope Rock and Soil Characteristics: Based on geological survey and design data, the slope rock and soil are divided into limestone facies and loose sedimentary facies; the first... i Lithological composition tensor of each drainage hole R i =[ R ci , R si ],in R ci This represents the volume percentage of the limestone facies. R si The volume percentage of the loosely packed phase, and satisfying the following conditions: R ci + R si =1; S2. Data Acquisition of Drain Holes: For each drain hole, at least a flow meter and a pressure sensor should be installed to collect the drainage flow rate in real time. A 1 and water pressure A 2 Data; for R ci A drain hole with a diameter ≥0.8 mm is used to install an additional calcium ion selective electrode for real-time collection of Ca in the water. 2+ concentration A 3 Data; for R si For drainage holes with a diameter ≥0.8, an additional turbidity meter should be installed to collect the turbidity of the infiltrated water in real time. A 4 Data; for the remaining drainage holes, additional calcium ion selective electrodes and turbidimeters are installed; data from each sensor is recorded synchronously at preset time intervals, using precipitation events as units, to form a raw dataset covering the current and multiple historical precipitation events. A =(A 1 , A 2 , A 3 , A 4 ),like A 3 or A 4 If any data is missing, it is recorded as 0. S3. Construction of the hydraulic coupling topology: Construct the drainage network into a directed hydraulic coupling topology. G =( V , E ), where nodes v i ∈ V Corresponding drainage hole; edge ( i , j )∈ E Characterizing pores i and j The gaps between them and the hydraulic interference path; edge ( i , j The establishment of a ) requires the simultaneous fulfillment of the following dual criteria: Geometric accessibility: Based on geological survey profiles and fracture statistics, if the distance between two boreholes is greater than the site characteristic connectivity scale... L max If a continuous waterproof layer exists as an obstacle, no connection will be established, and the edge ( i , j ) does not exist; Hydraulic connectivity: Based on monitoring data from at least three independent precipitation events, calculate the Pearson correlation coefficient between the head response sequence of each drainage well and adjacent wells; if the Pearson correlation coefficient is greater than 0.6, or if the well is located after precipitation... j Peak head and orifice i If the time difference between the peak flow rates is less than 2 hours, hydraulic connectivity is confirmed. i , j )exist; Define edge ( i , j The initial conduction coefficient of ) C ij0 ∈(0,1], this coefficient comprehensively reflects the strength of the hydraulic connection between the holes:

[0007] in, Q ijr Hole under reference working conditions i Hole caused by unit disturbance j Flow response amplitude, Q maxrThis represents the maximum response amplitude across the entire network. d ij The hole spacing is... d 0 The characteristic attenuation distance; λ These are the weighting coefficients; S4. Construction and training of graph attention network models: including the following steps: S401, Extraction of temporal feature vectors: Constructing a temporal encoder based on gated recurrent units; for the first... i Each drainage hole collects data at preset time intervals during a single precipitation event. A The gating loop unit input to the timing encoder takes... A The hidden state at the last time step serves as the temporal feature vector of the well under that precipitation event. F i ; S402. Construction of Graph Attention Network: Using Hydraulically Coupled Topology Graph G =( V , E Using the time-series feature vector as the skeleton, a dual-channel graph attention network module is constructed; F i With lithological composition tensor R i To splice them together, form the first i Initial node feature vectors of each drainage hole H i =( F i , R ci , R si ), as input features for a dual-channel graph attention network; The dual-channel graph attention network module includes a chemical crystallization pathway and a particle deposition pathway; the chemical crystallization pathway is for... R ci Nodes with a value >0.5 enhance calcium ion concentration. A 3 The weights of features in attention calculation, and their contribution to chemical sludge output in the intermediate hidden layers of the network. S ci The particle deposition pathway is for R si Nodes with values ​​greater than 0.5 enhance turbidity. A 4 The weights of features in attention calculation contribute to siltation in the output of the intermediate hidden layers of the network. S si ; Calculate the first i Comprehensive blockage index of each drain hole Bi = R ci · S ci + R si · S si ,in, S ci For chemical crystallization blockage rate, S si For sediment deposition and clogging rate; calculate edge ( i , j Effective transmission coefficient C ijt = C ij0 ·(1- w 1 · B i ) ·exp(- w 2 · d ij );in, d ij The distance between the two holes; w 1 , w 2 All are influence coefficients; in each layer of the graph attention network, the edge weights are updated in real time; the dual-channel graph attention network module calculates the influence weights between nodes based on this, aggregates the neighbor information, and outputs the clogging risk index of the drainage hole through the fully connected layer mapping. S403. Model Training Based on Transfer Learning: A two-stage training strategy of simulation pre-training and on-site fine-tuning is adopted to overcome the problems of difficulty in obtaining real-world blockage labels and misjudgment of single indicators. Specifically, in the simulation pre-training stage, seepage analysis software is used to generate simulation datasets covering different rainfall intensities, lithological combinations, and fracture apertures to simulate the generated physical blockage rate. B sim Labels are used to train the model to grasp the basic physical laws of seepage field evolution and siltation development; in the fine-tuning stage, a model is constructed based on the original dataset. A Silting risk label Y i ; The mean squared error loss function is used as the main loss function to minimize the clogging risk index output by the model. Y i The error between them; the Adam optimizer is used to update the network parameters, so as to achieve adaptive fine-tuning of the model for the hydrogeological characteristics of specific work sites.

[0008] S5. Drainage hole blockage prediction: The original dataset... AWith lithological composition tensor R i Input the trained graph attention network model and execute the following prediction process: S501, Input Feature Construction: For the first... i Each drainage hole collects data at preset time intervals during a single precipitation event. A The gating loop unit input to the timing encoder takes... A The hidden state at the last time step serves as the temporal feature vector of the well under that precipitation event. F i ; Transform the time series feature vector F i With lithological composition tensor R i The features are then concatenated to form the input features of a dual-channel graph attention network. S502, Dynamic Graph Reasoning and Propagation: Input features are fed into the trained graph attention network model. In each layer of propagation calculation, a comprehensive congestion index is calculated. B i = w c · S ci + w s · S si And calculate the edges ( i , j Effective transmission coefficient C ijt This allows for real-time updates of edge weights during propagation at each layer of the graph attention network; attention aggregation is then performed, and the updated features are passed to the next layer. S503. Prediction Result Output: After forward computation by the network, the following three types of prediction results are output: siltation risk index P 1i ∈[0,1], output by the Sigmoid branch, indicating the th i The probability that a drainage hole will become significantly clogged after the current rainfall event; this value is the primary prediction target of the model and is directly supervised by the loss function during training. Contribution of chemical crystallization S ci Contribution of sediment deposition S si ; These are intermediate analytical parameters generated during the model's forward computation, used to provide auxiliary explanations of the dominant mechanism of clogging after the prediction results are output; when S ci > S si and Rci ≥0.5 indicates a chemically crystalline predominant type; otherwise, it indicates a particulate deposition predominant type. Propagation Influence Coefficient Matrix M Based on dynamic transmission coefficient C ijt The calculated derived index matrix; M elements m ij Indicates the hole i When blockage occurs, the hole j The degree of hydraulic disturbance; and m ij =1- C ijt / C ij0 ;in C ijt This represents the dynamic propagation coefficient under the current predicted state; this matrix is ​​the post-processing result of intermediate variables in the forward calculation process, which is updated in real time with each prediction and is used to assist in assessing the spatial propagation risk under the current working condition. S6. Adaptive Early Warning Threshold Setting: Collect historical records of drainage hole blockage and corresponding monitoring data for this slope project, and calculate the historical 90th percentile of the blockage risk index for each drainage hole. P 90 and the propagation impact coefficient M 90th percentile of all elements in history m 90 The initial warning threshold is set as follows: Level 1 Warning: P 1i ≥0.8 P 90 And there are at least two adjacent holes. j satisfy m ij ≥0.7 m 90 ; Level II Warning: 0.6 P 90 ≤ P 1i <0.8· P 90 , or 0.5 m 90 ≤ m ij <0.7· m 90 ; Level 3 warning: 0.4 P 90 ≤ P 1i<0.6· P 90 And all m ij <0.5· m 90 ; Recalculated every six months based on newly accumulated operation and maintenance data. P 90 and m 90 This enables adaptive updating of the threshold.

[0009] Preferably, a siltation risk label Y i The construction uses single precipitation events or sliding time windows as sample units to analyze the first precipitation event. i Individual drainage hole construction risk label Y i ∈{0, 0.5, 1}, defined as follows: When the presence of physical blockages in the borehole is confirmed based on the borehole cleaning, endoscopy, or jet dredging records within 7 days following the rainfall event, it is classified as a positive case of sludge blockage. Y i =1; If there is no physical confirmation record, but any of the following hydraulic anomalies, chemical anomalies, or turbidity anomalies are met, the sample is judged as a suspected siltation sample. Y i =0.5; where the hydraulic anomaly condition is the flow rate. A 1 The water pressure decreased for three consecutive time steps. A 2 The rate of increase has continued, with the magnitude of change exceeding 1.5 standard deviations of the historical average for the same period; the chemical anomaly conditions are for... R ci Holes >0.5 A 3 A single increase exceeding twice the background value; the turbidity anomaly condition is for... R si Holes >0.5 A 4 A single elevation exceeds three times the background value; If the above-mentioned hydraulic anomaly conditions, chemical anomaly conditions, and turbidity anomaly conditions are not met, or if the well cleaning record confirms that the well is clean, the sample is judged to be a normal sample. Y i =0; In the loss function, for Y i Positive samples with a value of 1 are assigned the highest weight for blockages. Y i Suspected clogged samples with a value of 0.5 are assigned a moderate weight.Y i Normal samples with a value of 0 are assigned the lowest weight.

[0010] Preferably, the graph attention network model adopts an end-to-end deep neural network architecture, which is composed of a temporal feature extraction module, a dual-channel graph attention network module, and a fully connected mapping and output module connected in series. The temporal feature extraction module is equipped with a gated loop unit, used to extract the first time series feature. i Each drainage well collects data at preset time intervals during a single precipitation event. A Encoding is performed; this module outputs a time-series feature vector characterizing the hydraulic dynamic evolution of the orifice during this period. F i ; The dual-channel graph attention network module uses a hydraulically coupled topology graph. G To compute the skeleton, receive the temporal feature vectors. F i With lithological composition tensor R i The initial node feature vector formed by concatenation H i As input; the dual-channel graph attention network module internally sets up parallel chemical crystallization and grain deposition pathways, based on the lithological composition tensor. R i Assign attention weights; for R ci Nodes with a value >0.5 exhibit enhanced chemical crystallization pathways. A 3 Weights in message passing are used to calculate and output the contribution of chemical sludge. S ci ;for R si Nodes with a density >0.5 exhibit enhanced particle deposition pathways. A 4 Weights in message passing are used to calculate and output the contribution of siltation. S si In each layer of graph attention convolution operation, the module calculates based on the comprehensive congestion index. B i Real-time edge updates ( i , j Effective transmission coefficient C ijt The spatial propagation effect of blockage in porous media is simulated by using dynamic weights to control the information aggregation intensity between neighboring nodes. The fully connected mapping and output module receives the high-order node features updated by the graph attention module, maps them to the hidden layer space, and then outputs the first node feature via a Sigmoid activation function branch. iRisk index of clogging of each drain hole P 1i The sludge type is then determined via the Softmax classification branch.

[0011] Preferably, maintenance measures are recommended based on the warning level: Level 1 warning: Immediately arrange high-pressure water jet or chemical cleaning, and compare monitoring data before and after cleaning to assess whether additional drainage holes are needed; Level 2 warning: Perform endoscopic examination or local patency testing within one week, and increase monitoring frequency to once a day; Level 3 warning: Included in the monthly dredging and observation plan, maintaining regular monitoring frequency.

[0012] In summary, compared with the prior art, the present invention has the following beneficial effects: 1) Achieving quantitative characterization and unified modeling of different siltation mechanisms under complex geological conditions: This invention, for the first time, proposes a lithological composition tensor to quantitatively describe the geological environment of each drainage hole in complex slope engineering where limestone and loose deposits coexist. It also constructs a dual-channel graph attention network module composed of chemical crystallization pathways and particle deposition pathways. This module can adaptively enhance the weight of calcium ion concentration or turbidity features in attention calculation, thereby capturing the evolution of chemical crystallization siltation and particle mechanical deposition siltation within the same model framework. Compared with the shortcomings of traditional methods that cannot distinguish between the two siltation mechanisms or can only be processed separately, this invention achieves dynamic discrimination of different siltation types, contribution separation, and integrated calculation of comprehensive siltation indicators, significantly improving the characterization ability of siltation behavior under complex hydrogeological conditions. 2) This invention is the first to introduce the hydraulic coupling topology of drainage hole networks into siltation prediction, enabling a mathematical characterization of spatial propagation effects: Breaking away from the limitations of traditional point-based detection which neglects the hydraulic connections between holes, this invention proposes a directed hydraulic coupling topology based on dual criteria of geometric reachability and hydraulic connectivity. By introducing an effective transmission coefficient, when a drainage hole becomes silted up, its hydraulic driving capacity to adjacent drainage holes is proportionally weakened. This simulates the increased flow resistance, flow redistribution, and chain propagation effects of siltation along fracture connectivity paths caused by local blockage. This mechanism allows the model to dynamically update edge weights and output a propagation influence coefficient matrix, quantitatively assessing the interference of a single hole blockage on the entire network. It provides a computable mathematical tool for identifying key nodes and predicting siltation propagation chains, filling the gap in modeling the spatial coupling effects of slope drainage hole networks. 3) Integrating multi-dimensional time-series monitoring data with graph attention mechanisms to improve the accuracy and interpretability of siltation prediction: Dynamic evolution characteristics of time-series data from multiple sensors, such as drainage flow, water pressure, calcium ion concentration, and turbidity, are extracted through gated circulation units. Initial node features are constructed by combining lithological composition tensors. Then, a graph attention network is used to aggregate neighborhood information on the hydraulic coupling topology graph, thereby enabling the simultaneous learning of the siltation evolution law of each drainage hole and its influence from hydraulic disturbances of adjacent holes. The output results not only include the siltation risk index but also the contribution of chemical crystallization, the contribution of sediment deposition, and the propagation influence coefficient matrix, making the prediction results clearly physically interpretable. That is, maintenance personnel can clearly understand the dominant mechanism of siltation, the risk probability, and the range of adjacent holes that may be affected, thereby supporting accurate proactive maintenance decisions. 4) Achieving an adaptive hierarchical early warning and proactive operation and maintenance closed loop significantly improves the safety management level of slope drainage systems: Based on historical data of slope engineering and siltation, the maximum value of the siltation risk index and the maximum value of the propagation influence coefficient of each well are dynamically calculated, and a three-level progressive early warning threshold is set accordingly; the threshold is automatically updated every six months based on the new accumulated operation and maintenance data, so that the early warning standard can be adaptively adjusted with the service life of the slope and the evolution of the geological environment, avoiding false alarms or missed alarms caused by fixed thresholds; at the same time, clear maintenance measures are given for different early warning levels, realizing the transformation from passive treatment to proactive prevention, and providing quantifiable technical support for the long-term safe operation of drainage systems for large-scale slope engineering projects. Attached Figure Description

[0013] Figure 1 This is a flowchart of an artificial intelligence-based method for predicting and warning of drainage hole blockage, as shown in an embodiment of the present invention. Figure 2 This is a schematic diagram of the hydraulic coupling topology shown in an embodiment of the present invention. Detailed Implementation

[0014] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments given herein are for illustration and explanation only and are not intended to limit the present invention.

[0015] This application discloses, as follows: Figure 1-2 The AI-based method for predicting and warning of drainage hole blockage, as shown, is applied to drainage hole networks in large-scale slope engineering projects where limestone and loose deposits coexist. The method includes the following steps: S1. Quantification of characteristics of slope rock and soil: Based on geological survey and design data, the slope rock and soil are divided into limestone facies and loose sedimentary facies. Specifically, based on borehole core identification, stratigraphic columnar section or geophysical logging, if the carbonate rock layer is identified as mainly gray or dark gray cryptocrystalline or microcrystalline, with well-developed fractures and rich in calcium, it is determined to be limestone facies. Based on the loose, uncemented or weakly cemented Quaternary deposits revealed by the boreholes, with particle composition mainly of sand, gravel and clay, loose structure and high permeability, it is determined to be loose sedimentary facies.

[0016] Definition of the first i Lithological composition tensor of each drainage hole R i =[ R ci , R si ],in R ci This represents the volume percentage of the limestone facies. R si The volume percentage of the loosely packed phase, and satisfying the following conditions: R ci + R si =1; In specific implementation, for the first i For each drainage hole, the lithological composition tensor is calculated based on the thickness of each rock layer penetrated by the screen section in the borehole columnar section. R i =[ R ci , R si ],in, R ci This is the ratio of the total thickness of the limestone layer to the total length of the screen tube section. R si It is the ratio of the total thickness of the loose deposit layer to the total length of the screen pipe section; within the length of the borehole screen pipe section, it is assumed that the cross-sectional area of ​​each rock layer is the same, and the thickness ratio is equivalent to the volume ratio.

[0017] S2. Data Acquisition of Drain Holes: For each drain hole, at least a flow meter and a pressure sensor should be installed to collect the drainage flow rate in real time. A 1 and water pressure A 2 Data; for R ci A drain hole with a diameter ≥0.8 mm is used to install an additional calcium ion selective electrode for real-time collection of Ca in the water. 2+ concentration A 3 Data; for R siFor drainage holes with a diameter ≥0.8, an additional turbidity meter should be installed to collect the turbidity of the infiltrated water in real time. A 4 Data; for the remaining drainage holes, additional calcium ion selective electrodes and turbidimeters are installed; data from each sensor is recorded synchronously at preset time intervals, using precipitation events as units, to form a raw dataset covering the current and multiple historical precipitation events. A =( A 1 , A 2 , A 3 , A 4 ),like A 3 or A 4 Missing data is recorded as 0.

[0018] S3. Construction of the hydraulic coupling topology: Construct the drainage network into a directed hydraulic coupling topology. G =( V , E ), where nodes v i ∈ V Corresponding drainage hole; edge ( i , j )∈ E Characterizing pores i and j The gaps between them and the hydraulic interference path; edge ( i , j The establishment of a ) requires the simultaneous fulfillment of the following dual criteria: Geometric accessibility: Based on geological survey profiles and fracture statistics, if the distance between two boreholes is greater than the site characteristic connectivity scale... L max If a continuous waterproof layer exists as an obstacle, no connection will be established, and the edge ( i , j ) does not exist; Hydraulic connectivity: Based on monitoring data from at least three independent precipitation events, calculate the Pearson correlation coefficient between the head response sequence of each drainage well and adjacent wells; if the Pearson correlation coefficient is greater than 0.6, or if the well is located after precipitation... j Peak head and orifice i If the time difference between the peak flow rates is less than 2 hours, hydraulic connectivity is confirmed. i , j )exist; Define edge ( i , j The initial conduction coefficient of ) C ij0∈(0,1], this coefficient comprehensively reflects the strength of the hydraulic connection between the holes: (1) in, Q ijr To establish a seepage numerical model based on geological surveys and hydrogeological parameters at the work site, the borehole was tested under baseline conditions. i The orifice obtained by numerical solution after applying a unit head disturbance j Flow response amplitude; Q maxr This represents the maximum response amplitude across the entire network. d ij The hole spacing is... d 0 The characteristic attenuation distance represents the distance at which the hydraulic interference intensity between two boreholes decreases to 1 / e (approximately 36.8%) of its initial value as the distance increases in a homogeneous, isotropic seepage medium without fractured dominant channels. It comprehensively reflects the average hydraulic diffusion capacity and the connectivity of the fracture network of the site. d 0 A larger value indicates that the hydraulic disturbance propagates further in the seepage field, while a smaller value indicates that the hydraulic coupling effect is limited to boreholes with close proximity. d 0 Typical values ​​range from 3m to 30m; λ The weighting coefficient is used in specific implementation. λ = 0.6 to 0.8, with priority given to the actual measured flow response.

[0019] S4. Construction and training of graph attention network models: including the following steps: S401, Extraction of temporal feature vectors: Constructing a temporal encoder based on gated recurrent units; for the first... i Each drainage hole collects data at preset time intervals during a single precipitation event. A The gating loop unit input to the timing encoder takes... A The hidden state at the last time step serves as the temporal feature vector of the well under that precipitation event. F i .

[0020] S402. Construction of Graph Attention Network: Using Hydraulically Coupled Topology Graph G =( V , E Using a framework, a dual-channel graph attention network module is constructed to address the different clogging mechanisms of limestone and loose deposits; the temporal feature vectors are then used as the backbone. F i With lithological composition tensor R i To splice them together, form the first iInitial node feature vectors of each drainage hole H i =( F i , R ci , R si ), as input features for a dual-channel graph attention network; The dual-channel graph attention network module includes a chemical crystallization pathway and a particle deposition pathway; the chemical crystallization pathway is for... R ci Nodes with a value >0.5 enhance calcium ion concentration. A 3 The weights of features in attention calculation, and their contribution to chemical sludge output in the intermediate hidden layers of the network. S ci The particle deposition pathway is for R si Nodes with values ​​greater than 0.5 enhance turbidity. A 4 The weights of features in attention calculation contribute to siltation in the output of the intermediate hidden layers of the network. S si ; S ci and S si These are characteristic response values ​​of the intermediate layer of the network. Their values ​​reflect the degree of influence of chemical crystallization or grain deposition factors on the node state under corresponding lithological conditions, and are used as a subsequent comprehensive blockage index. B i The calculation input provides an intermediate reference for determining the clogging mechanism.

[0021] Calculate the first i Comprehensive blockage index of each drain hole B i = R ci · S ci + R si · S si ,in, S ci For chemical crystallization blockage rate, S si The siltation blockage rate, and S ci and S si These are all feature values ​​output from the intermediate layer of the dual-channel graph attention network module, reflecting the relative influence of the corresponding clogging factor on the current node's clogging state; the edges ( i , jEffective transmission coefficient C ijt = C ij0 ·(1- w 1 · B i ) ·exp(- w 2 · d ij );in, d ij The distance between the two holes; w 1 , w 2 These are all influence coefficients; in actual implementation, w 1 The value range is [0.8, 1.0]. w 2 The typical value is 0.05m. -1 Through this mechanism, when upstream nodes experience blockage, their comprehensive blockage index... B i The increase in volume proportionally weakens the hydraulic drive capacity of the drain hole to downstream or adjacent holes, thus simulating the increased flow resistance and flow redistribution caused by local blockage. Through this dynamic update mechanism, the model can capture the weakening effect of a single drain hole blockage on the hydraulic coupling relationship between adjacent nodes and further propagate it to more distant nodes, achieving a mathematical characterization of the spatial propagation effect of blockage. In each layer of the graph attention network, edge weights are updated in real time; the dual-channel graph attention network module calculates the influence weights between nodes accordingly, aggregates neighbor information, and outputs the blockage risk index of the drain hole via a fully connected layer.

[0022] S403. Model Training Based on Transfer Learning: A two-stage training strategy of simulation pre-training and on-site fine-tuning is adopted to overcome the problems of difficulty in obtaining real-world blockage labels and misjudgment of single indicators. Specifically, in the simulation pre-training stage, seepage analysis software is used to generate simulation datasets covering different rainfall intensities, lithological combinations, and fracture apertures to simulate the generated physical blockage rate. B sim Labels are used to train the model to grasp the basic physical laws of seepage field evolution and siltation development; in the fine-tuning stage, a model is constructed based on the original dataset. A Silting risk label Y i ; In specific implementation, a single precipitation event or a sliding time window is used as the sample unit to analyze the data. i Individual drainage hole construction risk label Y i∈{0, 0.5, 1} represents three risk levels: "normal," "suspected blockage," and "confirmed blockage," respectively. This label is used to train the model to identify the blockage risk category after the current precipitation event and serves only as a supervisory signal for risk grading. Y i The definition is as follows: When the presence of physical blockages in the borehole is confirmed based on the borehole cleaning, endoscopy, or jet dredging records within 7 days following the rainfall event, it is classified as a positive case of sludge blockage. Y i =1; If there is no physical confirmation record, but any of the following hydraulic anomalies, chemical anomalies, or turbidity anomalies are met, the sample is judged as a suspected siltation sample. Y i =0.5; where the hydraulic anomaly condition is the flow rate. A 1 The water pressure decreased for three consecutive time steps. A 2 The rate of increase has continued, with the magnitude of change exceeding 1.5 standard deviations of the historical average for the same period; the chemical anomaly conditions are for... R ci Holes >0.5 A 3 A single increase exceeding twice the background value; the turbidity anomaly condition is for... R si Holes >0.5 A 4 A single elevation exceeds three times the background value; the above thresholds are default initial values ​​set based on engineering experience in this field, and can be optimized and adjusted according to the statistical distribution of historical monitoring data of the work site in actual application.

[0023] If the above-mentioned hydraulic anomaly conditions, chemical anomaly conditions, and turbidity anomaly conditions are not met, or if the well cleaning record confirms that the well is clean, the sample is judged to be a normal sample. Y i =0; In the loss function, for Y i Positive samples with a value of 1 are assigned the highest weight for blockages. Y i Suspected clogged samples with a value of 0.5 are assigned a moderate weight. Y i Normal samples with a value of 0 are assigned the lowest weight.

[0024] Using mean squared error as the main loss function, the model output clogging risk index is minimized. Y i The error between them; during network training, the comprehensive congestion index B iApply a smoothness constraint to make B i The value variation conforms to the physical law of gradual siltation development, and is used as an auxiliary regularization term to guide the learning direction of intermediate layer features, avoiding intermediate parameters from producing values ​​that violate physical common sense. The Adam optimizer is used to update network parameters, realizing adaptive fine-tuning of the model for specific hydrogeological characteristics of the work site.

[0025] In specific implementation, the graph attention network model adopts an end-to-end deep neural network architecture, which is composed of a temporal feature extraction module, a dual-channel graph attention network module, and a fully connected mapping and output module connected in series. The temporal feature extraction module is equipped with a gated loop unit, used to extract the first time series feature. i Each drainage well collects data at preset time intervals during a single precipitation event. A Encoding is performed; this module outputs a time-series feature vector characterizing the hydraulic dynamic evolution of the orifice during this period. F i ; The dual-channel graph attention network module uses a hydraulically coupled topology graph. G To compute the skeleton, receive the temporal feature vectors. F i With lithological composition tensor R i The initial node feature vector formed by concatenation H i As input; the dual-channel graph attention network module internally sets up parallel chemical crystallization and grain deposition pathways, based on the lithological composition tensor. R i Assign attention weights; for R ci Nodes with a value >0.5 exhibit enhanced chemical crystallization pathways. A 3 Weights in message passing are used to calculate and output the contribution of chemical sludge. S ci ;for R si Nodes with a density >0.5 exhibit enhanced particle deposition pathways. A 4 Weights in message passing are used to calculate and output the contribution of siltation. S si In each layer of graph attention convolution operation, the module calculates based on the comprehensive congestion index. B i Real-time edge updates ( i , j Effective transmission coefficient C ijtThe spatial propagation effect of blockage in porous media is simulated by using dynamic weights to control the information aggregation intensity between neighboring nodes. The fully connected mapping and output module receives the high-order node features updated by the graph attention module, maps them to the hidden layer space, and then outputs the first node feature via a Sigmoid activation function branch. i Risk index of clogging of each drain hole P 1i The sludge type is then determined via the Softmax classification branch.

[0026] S5. Drainage hole blockage prediction: The original dataset... A With lithological composition tensor R i Input the trained graph attention network model and execute the following prediction process: S501, Input Feature Construction: For the first... i Each drainage hole collects data at preset time intervals during a single precipitation event. A The gating loop unit input to the timing encoder takes... A The hidden state at the last time step serves as the temporal feature vector of the well under that precipitation event. F i ; Transform the time series feature vector F i With lithological composition tensor R i The features are then concatenated to form the input features of a dual-channel graph attention network. S502, Dynamic Graph Reasoning and Propagation: Input features are fed into the trained graph attention network model. In each layer of propagation calculation, a comprehensive congestion index is calculated. B i = w c · S ci + w s · S si And calculate the edges ( i , j Effective transmission coefficient C ijt This allows for real-time updates of edge weights during propagation at each layer of the graph attention network; attention aggregation is then performed, and the updated features are passed to the next layer. S503. Prediction Result Output: After forward computation by the network, the following three types of prediction results are output: siltation risk index P 1i ∈[0,1], output by the Sigmoid branch, indicating the th iThe probability of a drainage hole becoming significantly clogged after the current precipitation event; Contribution of chemical crystallization S ci Contribution of sediment deposition S si ;when S ci > S si and R ci ≥0.5 indicates a chemically crystalline predominant type; otherwise, it indicates a particulate deposition predominant type. Propagation Influence Coefficient Matrix M ; M elements m ij Indicates the hole i When blockage occurs, the hole j The degree of hydraulic disturbance; and m ij =1- C ijt / C ij0 ;in C ijt This represents the dynamic transmission coefficient under the current predicted state. The more severe the upstream blockage, the more... C ijt The smaller, m ij The closer the value is to 1, the stronger the downstream interference. This matrix is ​​updated in real time with each prediction, reflecting the spatial propagation risk under the current operating conditions.

[0027] S6. Adaptive Early Warning Threshold Setting: Collect historical records of drainage hole siltation and corresponding monitoring data for the past 3-5 years for this slope project, and calculate the historical 90th percentile of the siltation risk index for each drainage hole. P 90 and the propagation impact coefficient M 90th percentile of all elements in history m 90 The purpose of selecting the historical 90th percentile is to eliminate the interference of extreme outliers on the threshold setting. The initial warning threshold is set as follows: Level 1 Warning: P 1i ≥0.8 P 90 And there are at least two adjacent holes. j satisfy m ij ≥0.7 m 90 ; Level II Warning: 0.6 P 90≤ P 1i <0.8· P 90 , or 0.5 m 90 ≤ m ij <0.7· m 90 ; Level 3 warning: 0.4 P 90 ≤ P 1i <0.6· P 90 And all m ij <0.5· m 90 ; Recalculated every six months based on newly accumulated operation and maintenance data. P 90 and m 90 This enables adaptive updating of thresholds; the proportional coefficients of 0.4, 0.5, 0.6, 0.7, and 0.8 are default initial values ​​set based on engineering experience. In actual applications, they can be optimized and adjusted according to the statistical distribution of historical monitoring data of the work site and the requirements of operation and maintenance management, while maintaining a hierarchical progression.

[0028] In practice, maintenance measures are recommended based on the warning level: Level 1 warning: Immediately arrange high-pressure water jet or chemical cleaning, and compare monitoring data before and after cleaning to assess whether additional drainage holes are needed; Level 2 warning: Perform endoscopic examination or local patency testing within one week, and increase monitoring frequency to once a day; Level 3 warning: Included in the monthly dredging and observation plan, maintaining regular monitoring frequency.

[0029] The above describes one or more embodiments of the present invention in a relatively specific and detailed manner, but it should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this patent should be determined by the appended claims.

Claims

1. A method for predicting and warning of a drain clogging based on artificial intelligence, characterized by, Comprising the following steps: S1, Quantification of the characteristics of the slope rock-soil mass: According to the geological survey and design data, the slope rock-soil mass is divided into limestone phase and loose accumulation phase; Definition of the first i Lithological composition tensor of each drainage hole R i =[ R ci , R si ],in R ci This represents the volume percentage of the limestone facies. R si The volume percentage of the loosely packed phase, and satisfying the following conditions: R ci + R si =1; S2, Data collection of drainage holes: for each drainage hole, at least install flow meter and pressure sensor to collect real-time drainage flow A 1 and water pressure A 2 data; for R ci drainage holes with >0.8, additionally install calcium ion selective electrode to collect real-time Ca 2+ concentration in water A 3 data; for R si drainage holes with >0.8, additionally install turbidimeter to collect real-time turbidity of infiltrated water A 4 data; for the rest of the drainage holes, additionally install both calcium ion selective electrode and turbidimeter; record the values of each sensor simultaneously at preset time intervals in units of precipitation events to form the original data set A A 1 A 2 A 3 A 4 , if the data of A 3 or A 4 is missing, it is recorded as 0;​​​​ S3, Construction of hydraulic coupling topological graph: Construct the drainage hole network as a directed hydraulic coupling topological graph G =( V , E ), wherein the nodes v i ∈ V correspond to drainage holes; edges i , j )∈ E represent the fracture communication and hydraulic interference path between hole i and j ; The establishment of the edge ( i , j ) must satisfy the following criteria simultaneously:​ Geometric accessibility: Based on geological investigation profile and fracture statistics, if the distance between two holes is greater than the site characteristic connectivity scale L max or there is a continuous aquiclude barrier, then no connection is established; Hydraulic connectivity: Based on the monitoring data of not less than three independent rainfall events, the Pearson correlation coefficient between the water head response sequence of each drainage hole and the adjacent hole is calculated; If the Pearson correlation coefficient is greater than 0.6, or the time difference between the peak water head of the hole j and the peak flow of the hole i is less than 2 hours after the precipitation, it is confirmed that there is hydraulic connection between the holes. i , j ​​ Define edge ( i , j The initial conduction coefficient of ) C ij0 ∈(0,1], this coefficient comprehensively reflects the strength of the hydraulic connection between the holes: （1） wherein, Q ijr is the reference case hole i is the hole caused by a unit perturbation j is the flow response amplitude, Q maxr is the maximum response amplitude across the network, d ij is the hole spacing, d 0 is the characteristic decay distance; Lambda is the weight coefficient; S4, Construction and training of the graph attention network model: comprising the following steps: S401, extraction of the time sequence feature vector: a time sequence encoder based on a gated recurrent unit is constructed; for the first drainage hole, the input collected at a preset time interval in a single rainfall event is input to the gated recurrent unit of the time sequence encoder, and the hidden state of the last time step is taken as the time sequence feature vector of the hole in the rainfall event i A A F i ;​​​ S402, construction of a graph attention network: a hydraulic coupling topological graph G =( V , E ) is the skeleton, and a double-channel graph attention network module is constructed; the time sequence feature vector F i and the lithology composition tensor R i are spliced to form the initial node feature vector of the i drainage hole H i =( F i , R ci , R si ), as the input feature of the double-channel graph attention network; The dual-channel graph attention network module includes a chemical crystallization channel and a particle deposition channel; the chemical crystallization channel is for R ci Nodes with a feature weight greater than 0.5, reinforced calcium ion concentration A 3 The weight of the feature in the attention calculation, output chemical congestion contribution S ci ; the particle deposition channel is for R si Nodes with a feature weight greater than 0.5, reinforced turbidity A 4 The weight of the feature in the attention calculation, output sediment congestion contribution S si ; The comprehensive blocking index of the first drainage hole is calculated i B i R ci S ci R si S si S ci is a chemical crystallization blocking rate, S si is a sediment deposition blocking rate; the effective conduction coefficient of the edge (e i , j ) is calculated as C ijt C ij0 w 1 B i w 2 d ij ; wherein, d ij is a two-hole spacing; w 1 , w 2 are all influence coefficients; in each layer propagation of the graph attention network, the edge weight is updated in real time; the double-channel graph attention network module calculates the influence weight between nodes according to this, aggregates neighbor information, maps through a fully connected layer, and outputs the siltation risk index of the drainage hole;​​​​​​​​​​​ S403、based on the model training of migration learning: adopt two-stage training strategy of simulation pre-training and field fine-tuning; wherein, in the simulation pre-training stage, the simulation dataset covering different rainfall intensities, lithological combinations and fracture openings is generated by using seepage analysis software to simulate the generated physical plugging rate B sim For the label, the trained model masters the basic physical law of seepage field evolution and silt development; in the fine-tuning stage, the silt risk label based on the original dataset A Y i is constructed; the mean square error loss function is adopted to minimize the error between the silt risk index output by the model and the Y i ; the Adam optimizer is used to update the network parameters to realize the adaptive fine-tuning of the model to the hydrogeological characteristics of the specific work site.​ S5, Drainage hole clogging prediction: the original data set A with the lithology composition tensor R i Input the trained graph attention network model, and execute the following prediction process: S501、Input feature construction: for the first drainage hole, the time series features of the drainage hole in the single rainfall event are constructed by inputting the collected data at preset time intervals into the gating recurrent unit of the time series encoder, taking the hidden state of the last time step as the time series feature vector of the hole in the rainfall event i A A F i F i R i ​​​​​​ S502, dynamic graph reasoning and propagation: input the input features into the trained graph attention network model, and calculate the comprehensive congestion index in each layer propagation calculation B i = w c · S ci + w s · S si , and calculate the effective conduction coefficient of the edge i , j ) C ijt , so as to update the edge weight in real time in each layer propagation of the graph attention network; perform attention aggregation, and pass the updated features to the next layer; S503, Prediction result output: After forward calculation of the network, the following three types of prediction results are output: Clogging risk index P 1i ∈ [0,1] output by the Sigmoid branch, representing the probability of a significant clogging of the nth drain after the current precipitation event; i Clogging risk index Contribution of chemical crystallization S ci Contribution of sediment deposition S si ;when S ci > S si and R ci ≥0.5 indicates a chemically crystalline predominant type; otherwise, it indicates a particulate deposition predominant type. Propagation Influence Coefficient Matrix M ; M elements m ij Indicates the hole i When blockage occurs, the hole j The degree of hydraulic disturbance; and m ij =1- C ijt / C ij0 ;in C ijt This represents the dynamic propagation coefficient under the current predicted state; this matrix is ​​updated in real time with each prediction, reflecting the spatial propagation risk under the current operating conditions. S6, adaptive early warning threshold setting: collect the history records of the drainage hole blockage of the slope engineering and the corresponding monitoring data, respectively, and count the historical 90% quantile of the drainage hole blockage risk index of each drainage hole P 90 , and the propagation influence coefficient M , and the historical 90% quantile of all elements m 90 ; the initial early warning threshold is set as follows: First level warning: P 1i ≥ 0.8 · P 90 and there are at least two adjacent holes j satisfies m ij ≥ 0.7 · m 90 ; Second level warning: 0.6 P 90 ≤ P 1i < 0.8 P 90 or 0.5 m 90 ≤ m ij < 0.7 m 90 ; Level 3 warning: 0.4 P 90 ≤ P 1i < 0.6 P 90 , and all m ij < 0.5 m 90 ; re-computed periodically based on newly accumulated O&M data P 90 and m 90 enabling adaptive update of the threshold. 2.The artificial intelligence-based scupper clogging prediction and early warning method of claim 1, wherein Risk of clogging label Y i of the construction, to a single rainfall event or sliding time window as a sample unit, the first i Risk of clogging label for the first drainage hole Y i ∈{0, 0.5, 1}, defined as follows: A clogged positive sample was determined when there was a physical blockage in the hole as confirmed by a clear hole, endoscopic, or jetting record within 7 days after the precipitation event and Y i = 1; No physical confirmation record, but when any of the following hydraulic anomaly conditions, chemical anomaly conditions, turbidity anomaly conditions are met, it is determined as a suspected blockage sample and Y i =0.5; wherein the hydraulic anomaly condition is that the flow rate A 1 drops for 3 consecutive time steps and the water pressure A 2 continuously rises and the change amplitude exceeds 1.5 times the standard deviation of the historical same period mean; the chemical anomaly condition is that the hole R ci >0.5, A 3 single rise exceeds 2 times the background value; the turbidity anomaly condition is that the hole R si >0.5, A 4 single rise exceeds 3 times the background value; At the same time, if the above hydraulic anomaly condition, chemical anomaly condition, turbidity anomaly condition are not met, or the hole is confirmed to be clean by hole cleaning record, it is determined as a normal sample and Y i =0; In the loss function, the highest weight is given to clogging positive samples for which Y i = 1, a medium weight is given to suspected clogging samples for which Y i = 0.5, and the lowest weight is given to normal samples for which Y i = 0. 3.The artificial intelligence-based scupper clogging prediction and early warning method of claim 1, wherein, The graph attention network model adopts an end-to-end deep neural network architecture, which is composed of a time series feature extraction module, a double-channel graph attention network module, and a full connection mapping and output module in series; The time sequence feature extraction module is configured with a gating cycle unit, which is used for encoding the first i drainage holes collected at preset time intervals in a single rainfall event. A ​ The module outputs a time sequence feature vector representing the hydraulic dynamic evolution law of the hole in this time period F i ; Dual-channel graph attention network module with hydraulic coupling topological graph G To compute the skeleton, receive a time-series feature vector F i With lithology composition tensor R i Spliced initial node feature vector H i As input; dual-channel graph attention network module internally parallel sets chemical crystallization channel and particle deposition channel, according to lithology composition tensor R i Attention weight is allocated; For nodes with R ci > 0.5, chemical crystallization channel strengthening A 3 Weight in message passing, calculate and output chemical jam contribution S ci ; for nodes with R si > 0.5, particle deposition channel strengthening A 4 Weight in message passing, calculate and output sediment jam contribution S si ; in each layer of graph attention convolution operation, the module calculates the comprehensive jam index according to B i Real-time update of the effective conduction coefficient of edge i , j ) C ijt ; the information aggregation strength between adjacent nodes is controlled by using dynamic weight, so as to simulate the spatial propagation effect of jam in pore medium; The full connection mapping receives the high-order node features updated by the graph attention module, maps the high-order node features to a hidden layer space, and outputs the first i risk index of the first drain hole P 1i ; and outputs the clogging type discrimination result through a Softmax classification branch. 4.The artificial intelligence-based scupper clogging prediction and early warning method of claim 1, wherein, According to the early warning level, recommend maintenance measures: First-level warning: Immediately arrange high-pressure water jet or chemical cleaning, and compare monitoring data before and after cleaning to evaluate whether additional drainage holes need to be added; Second-level warning: Endoscopy detection or local dredging within a week, and increase the monitoring frequency to once a day; Third-level warning: Included in the monthly dredging observation plan, and keep the regular monitoring frequency.

Images