Hydraulic engineering seepage intelligent monitoring system and method

Abstract

The invention discloses a hydraulic engineering seepage intelligent monitoring system and method, and belongs to the technical field of hydraulic engineering safety monitoring. The system comprises the following modules: a multi-source data acquisition module used for acquiring multiple types of monitoring data in real time through an intelligent sensor network; the double-window time sequence analysis module is used for constructing a quick response window and a trend analysis window to realize double identification of sudden anomalies and long-term trends; the self-learning threshold optimization module is used for automatically extracting a key quantile threshold based on the distribution characteristics of the monitoring data and continuously optimizing weight configuration and early warning threshold setting of various statistical indexes; the multi-scale fusion early warning module is used for performing multi-source information fusion, generating a comprehensive change index and a multi-stage early warning state, and outputting a seepage abnormity early warning signal and a corresponding confidence coefficient; and the visual decision support module provides visual data display, emergency response guidance and intelligent decision support.

Description

Technical Field

[0001] This application relates to the field of water conservancy project safety monitoring technology, and more specifically, to a water conservancy project seepage intelligent monitoring system and method. Background Technology

[0002] Seepage is a key factor affecting the safe operation of water conservancy projects. Abnormal changes in seepage often indicate potential structural instability hazards. If these hazards are not detected and addressed in a timely manner, they may lead to landslides, collapses, or even project failure, causing severe economic losses and casualties. Therefore, accurate and real-time monitoring of seepage status in water conservancy projects is of great significance for ensuring project safety.

[0003] Existing seepage monitoring technologies in hydraulic engineering mainly rely on single-sensor data acquisition and often use static thresholds for anomaly warning, making them ill-suited to complex and ever-changing field environments and seepage characteristics. Traditional monitoring systems have significant shortcomings in data quality control, environmental impact compensation, and the sensitivity and accuracy of anomaly identification, and lack the ability to comprehensively analyze short-term emergencies and long-term trends. Furthermore, warning thresholds are typically set empirically, lacking dynamic adjustment mechanisms and failing to meet the needs of different engineering environments and seasonal variations. Moreover, existing systems primarily present two-dimensional data, lacking numerical reconstruction and visualization of the three-dimensional structure of the seepage field, limiting a deeper understanding of seepage evolution mechanisms and support for scientific decision-making.

[0004] Furthermore, with the development of IoT, big data, and AI technologies, how to fuse multi-source heterogeneous sensor data and combine it with intelligent algorithms to achieve efficient and accurate seepage anomaly identification and adaptive early warning has become a key direction for improving the safety management level of water conservancy projects. However, there is currently a lack of systematic solutions that integrate multi-level data acquisition, dual time-series analysis, self-learning threshold optimization, and multi-scale fusion early warning, which fails to fully realize the potential of intelligent technologies in seepage monitoring.

[0005] In summary, how to achieve high-precision, multi-dimensional real-time monitoring of seepage in water conservancy projects, combined with intelligent analysis and adaptive early warning mechanisms, to improve the accuracy and response speed of anomaly identification and ensure the safe operation of the project has become an urgent technical problem to be solved. Summary of the Invention

[0006] In order to overcome a series of defects in the existing technology, the purpose of this application is to provide an intelligent monitoring system for seepage in water conservancy projects, which includes the following modules:

[0007] The multi-source data acquisition module is used to collect various types of monitoring data in real time through an intelligent sensor network, and to perform quality control and environmental factor compensation to improve the accuracy and usability of the data.

[0008] The dual-window time series analysis module is configured to receive monitoring data output from the multi-source data acquisition module, construct a fast response window and a trend analysis window, and extract key statistical indicators to achieve dual identification of sudden anomalies and long-term trends.

[0009] The self-learning threshold optimization module responds to the analysis results output by the dual-window time series analysis module, automatically extracts key quantile thresholds based on the distribution characteristics of the monitoring data, and continuously optimizes the weight configuration and early warning threshold settings of various statistical indicators through reinforcement learning algorithms.

[0010] The multi-scale fusion early warning module is configured to integrate the output results of the self-learning threshold optimization module and the dual-window time series analysis module, perform multi-source information fusion, generate a comprehensive change index and multi-level early warning status, and output seepage anomaly early warning signals and their corresponding confidence levels.

[0011] The visualization decision support module responds to the warning signals from the multi-scale fusion warning module. Based on the real-time monitoring dashboard, three-dimensional seepage field reconstruction, and multi-channel warning push mechanism, it provides intuitive data display, emergency response guidance, and intelligent decision support.

[0012] Among them, the multi-source data acquisition module, the dual-window time series analysis module, the self-learning threshold optimization module, the multi-scale fusion early warning module, and the visualization decision support module work together to achieve efficient identification, accurate early warning, and intelligent auxiliary decision-making for seepage anomalies in water conservancy projects.

[0013] Furthermore, the multi-source data acquisition module includes a distributed sensor node array, a data quality assessment unit, and an environmental compensation and correction unit. The distributed sensor node array includes piezometers, flow sensors, water level sensors, soil pressure sensors, and tilt sensors. Each sensor node achieves synchronous data transmission via a LoRa+4G hybrid networking protocol. The data quality assessment unit is configured to perform outlier detection, missing value imputation, and noise filtering on the acquired data. Outlier detection employs a combined discrimination mechanism based on the improved 3σ criterion and box plot method, with the detection threshold dynamically adjusted according to the data distribution characteristics. The environmental compensation and correction unit performs temperature drift compensation and nonlinear correction on the raw sensor data based on environmental parameters, achieving a compensation accuracy of ±0.5% of the measured value, ensuring the consistency and accuracy of the monitoring data within an ambient temperature range of -20℃ to 60℃.

[0014] Furthermore, the basic window length of the rapid response window is set to 1 to 4 hours, with a sliding step of 15 minutes; the trend analysis window adopts a hierarchical time-weighted strategy, with a basic window length set to 24 to 30 days, extended to 45 days according to seasonal changes; the key statistical indicator extraction includes calculating the mean, standard deviation, coefficient of variation, skewness, kurtosis, first difference, and sliding correlation coefficient of the time series data. Through multi-dimensional joint analysis of statistical indicators, combined with wavelet transform for frequency domain feature extraction, accurate identification and quantitative description of seepage state change patterns are achieved.

[0015] Furthermore, the three-dimensional seepage field reconstruction is based on measured seepage pressure data and a simplified engineering geometric model to construct a three-dimensional seepage numerical model. The model adopts regular hexahedral mesh division, and the mesh size is adjusted between 0.2m and 1.0m according to the calculation accuracy requirements and computational resource constraints. The total number of meshes is controlled between 30,000 and 150,000. The permeability coefficient is determined through field test data and geological survey data. The boundary conditions include two types: constant head boundary and impermeable boundary. The numerical solution adopts the over-relaxation iteration method, with the relaxation factor set to 1.2, the convergence accuracy set to 10⁻⁴, and the upper limit of the number of iterations set to 1500. The reconstruction results are visualized in the form of three-dimensional constant head surface, seepage direction vector, and seepage head distribution cloud map.

[0016] The purpose of this application is also to provide an intelligent monitoring method for seepage in water conservancy projects, applied to the aforementioned seepage monitoring system for water conservancy projects, the method comprising the following steps:

[0017] A two-level monitoring architecture of rapid response window and trend analysis window is constructed to obtain real-time seepage monitoring data of the water conservancy project, and the analysis range of each time window is determined according to the time characteristics of the monitoring data.

[0018] Based on the time range of the fast response window and the trend analysis window, the dual-window time series analysis module is controlled to extract the statistical characteristic parameters of the seepage monitoring data in each time window;

[0019] Whenever the dual-window time series analysis module completes the extraction of statistical feature parameters, it marks the statistical feature parameters in the current time window as target analysis parameters, and configures the learning parameters of the self-learning threshold optimization module based on the distribution characteristics of the target analysis parameters.

[0020] Based on the learning parameters and using the self-learning threshold optimization module, a distribution histogram is constructed based on historical monitoring data, and key quantiles are extracted from the distribution histogram as threshold boundary reference values.

[0021] During the threshold boundary reference value extraction process, the self-learning threshold optimization module collects the change characteristics of the target analysis parameters in real time, uses the self-learning mechanism to collect the real-time weight data of each statistical indicator, and combines the threshold boundary reference value and the real-time weight data to identify whether the current weight configuration has reached the preset optimization standard.

[0022] If the current weight configuration does not meet the preset optimization standard, the self-learning mechanism will continue to iterate and optimize until the optimization standard is met.

[0023] If the current weight configuration has reached the preset optimization standard, the statistical feature parameters are weighted and fused using the dual-window time series analysis module, and adaptive threshold parameters are generated and the comprehensive change index is calculated through the self-learning threshold optimization module.

[0024] During the calculation of the comprehensive change index, the multi-source data acquisition module monitors the change trend of the seepage monitoring data in real time, and the dual-window time series analysis module analyzes whether the comprehensive change index exceeds the safe range of the adaptive threshold parameter.

[0025] If the comprehensive change index does not exceed the safety range, the current monitoring status is maintained and the multi-source data acquisition module is controlled to continue the data acquisition and analysis process;

[0026] If the comprehensive change index exceeds the safe range, the analysis results of the fast response window and the trend analysis window are integrated by the dual-window time series analysis module, and the multi-scale fusion early warning module is used to execute the multi-scale fusion decision-making mechanism to generate a seepage anomaly early warning signal. The visualization decision support module is then controlled to output the corresponding early warning level and emergency response suggestions.

[0027] Furthermore, the construction of a two-tiered monitoring architecture consisting of a rapid response window and a trend analysis window to acquire real-time seepage monitoring data of the water conservancy project, and the determination of the analysis range for each time window based on the temporal characteristics of the monitoring data, includes the following steps:

[0028] Time-frequency domain analysis is performed on historical seepage monitoring data to identify its periodicity and abrupt change characteristics. Then, the optimal duration and overlap ratio of the rapid response window and trend analysis window are calculated and determined to achieve accurate capture and continuous monitoring of the time characteristics of the data.

[0029] Based on the established window parameters, configure a fast response window and a trend analysis window to capture short-term anomalies and long-term trends respectively, while establishing a data synchronization and time alignment mechanism between the two.

[0030] By using a dynamic adjustment mechanism based on data variability and signal-to-noise ratio, the boundaries of the fast response window and trend analysis window are adaptively optimized, thereby intelligently adjusting the window length under different data fluctuations and operating conditions.

[0031] A data stream distributor is established to synchronously distribute the real-time collected seepage monitoring data to the rapid response window and the trend analysis window;

[0032] By tracking the data processing load, analysis accuracy, and response time metrics of each window in real time, the system assesses the statistical stability of the data within the window and the consistency of the analysis results. It also automatically triggers parameter optimization processes when performance degradation or misconfiguration is detected, ensuring that the monitoring architecture always maintains optimal working condition.

[0033] Furthermore, based on the time range of the rapid response window and the trend analysis window, the dual-window time series analysis module is controlled to extract the statistical characteristic parameters of the seepage monitoring data within each time window, including the following steps:

[0034] For the rapid response window, the central trend, dispersion and instantaneous changes of the seepage data are quantified by calculating the basic statistical characteristics and rate of change of the seepage data within the window, providing quantitative indicators for short-term anomaly detection;

[0035] For the trend analysis window, by performing fitting analysis and correlation calculation on time series data, its long-term trend, periodicity and seasonality characteristics are extracted, thereby constructing a quantitative description of trend evolution;

[0036] Frequency domain analysis is performed on each time window to extract power spectrum, main frequency and spectral features. Multi-scale time-frequency decomposition is performed through wavelet transform and empirical mode decomposition to construct a comprehensive multi-scale feature vector.

[0037] Based on the statistical analysis of historical abnormal events, information gain, chi-square test and correlation analysis methods are used to evaluate the importance ranking of each characteristic parameter, establish a dynamic weight allocation mechanism, and select the feature subset with the greatest discriminative power for seepage anomalies.

[0038] Quality control is performed on the extracted statistical feature parameters, including stability testing and rationality verification, and uncertainty assessment and confidence intervals are established to ensure the reliability and standardization of the output feature data.

[0039] Furthermore, whenever the dual-window time series analysis module completes the extraction of statistical feature parameters, it marks the statistical feature parameters within the current time window as target analysis parameters, and configures the learning parameters of the self-learning threshold optimization module based on the distribution characteristics of the target analysis parameters, including the following steps:

[0040] Establish a completion status monitoring mechanism for the dual-window time series analysis module to detect the feature parameter extraction progress of the fast response window and trend analysis window in real time, and automatically add a unique timestamp and window identifier to all statistical feature parameters in the current time window upon completion;

[0041] Based on the numerical stability, physical meaning, and historical performance of the characteristic parameters, the key parameters most sensitive to changes in seepage state are selected as target analysis parameters.

[0042] The selected target analysis parameters are subjected to probability distribution fitting and statistical characteristic analysis to determine their optimal distribution model and distribution characteristics.

[0043] Based on the distribution characteristics and numerical range of the target analysis parameters, the key learning parameters of the self-learning threshold optimization module are initialized and dynamically adjusted through an adaptive configuration strategy.

[0044] Based on the time correlation and seasonality characteristics of the target analysis parameters, the optimal length of the historical data window and the data weight decay strategy are determined.

[0045] The effectiveness and stability of the learning parameter configuration are verified through rationality checks and small sample tests. An automatic tuning mechanism is established to prevent learning instability. Finally, the configuration is completed and the self-learning threshold optimization module is activated to start the threshold learning and optimization process.

[0046] Furthermore, the construction of the distribution histogram adopts an adaptive binning strategy. In the initial stage (sample size less than 500), a fixed 15 bins are used. As the sample size increases, the number of bins is dynamically adjusted: 20-35 bins are used when the sample size is 500-2000, 35-60 bins are used when the sample size is 2000-10000, and 60-100 bins are used when the sample size is greater than 10000. At the same time, the bin boundaries are finely adjusted in combination with the skewness and kurtosis characteristics of the data distribution.

[0047] Furthermore, the comprehensive change index is calculated using a time series decomposition and reconstruction method, which decomposes the original monitoring data into three parts: trend component, seasonal component, and residual component. The weight coefficient of the trend component is set to 0.5, the weight coefficient of the seasonal component is set to 0.3, and the weight coefficient of the residual component is set to 0.2. The comprehensive change index is calculated by weighted summation, and the index value is standardized to a range of 0 to 1. A time decay factor is introduced in the calculation process. The weight coefficient of recent data is 1.0. Within the fast response window, the weight decays by 0.02 to 0.05 for every 15 minutes forward, and within the trend analysis window, the weight decays by 0.05 to 0.1 for every 24 hours forward.

[0048] Compared with the prior art, this application has the following beneficial effects:

[0049] This application constructs a multi-module collaborative intelligent monitoring and early warning architecture. It achieves high-precision data acquisition through a multi-source sensor network, and combines dual-window time-series analysis to identify both short-term anomalies and long-term trends. A self-learning threshold optimization module dynamically adjusts early warning thresholds and statistical indicator weights based on data distribution and reinforcement learning. Furthermore, a multi-scale fusion module integrates multi-source information to generate a comprehensive change index and multi-level early warning signals. In addition, it is equipped with visualization-based decision support based on three-dimensional seepage field reconstruction, providing intuitive monitoring and intelligent emergency guidance. Overall, it improves the efficiency of seepage anomaly identification, the accuracy of early warning, and the intelligent auxiliary decision-making capability of water conservancy projects, realizing intelligent management of the entire process from data acquisition and analysis to early warning and decision-making. Attached Figure Description

[0050] Figure 1 This is a communication timing diagram, which describes the interaction process between various modules of a water conservancy engineering seepage intelligent monitoring system disclosed in an embodiment of this application.

[0051] Figure 2 This is a flowchart illustrating an intelligent monitoring method for seepage in water conservancy projects disclosed in an embodiment of this application. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be described in more detail below with reference to the accompanying drawings. In the drawings, the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The described embodiments are some embodiments of this invention, but not all embodiments.

[0053] Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] The embodiments and directional terms described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0055] Figure 1 This is a communication timing diagram, which describes the interaction process between various modules of a water conservancy engineering seepage intelligent monitoring system disclosed in an embodiment of this application.

[0056] like Figure 1 As shown, a smart seepage monitoring system for water conservancy projects includes the following modules:

[0057] The multi-source data acquisition module is used to collect various types of monitoring data in real time through an intelligent sensor network, and to perform quality control and environmental factor compensation to improve the accuracy and usability of the data.

[0058] The dual-window time series analysis module is configured to receive monitoring data output from the multi-source data acquisition module, construct a fast response window and a trend analysis window, and extract key statistical indicators to achieve dual identification of sudden anomalies and long-term trends.

[0059] The self-learning threshold optimization module responds to the analysis results output by the dual-window time series analysis module, automatically extracts key quantile thresholds based on the distribution characteristics of the monitoring data, and continuously optimizes the weight configuration and early warning threshold settings of various statistical indicators through reinforcement learning algorithms.

[0060] The multi-scale fusion early warning module is configured to integrate the output results of the self-learning threshold optimization module and the dual-window time series analysis module, perform multi-source information fusion, generate a comprehensive change index and multi-level early warning status, and output seepage anomaly early warning signals and their corresponding confidence levels.

[0061] The visualization decision support module responds to the warning signals from the multi-scale fusion warning module. Based on the real-time monitoring dashboard, three-dimensional seepage field reconstruction, and multi-channel warning push mechanism, it provides intuitive data display, emergency response guidance, and intelligent decision support.

[0062] Among them, the multi-source data acquisition module, the dual-window time series analysis module, the self-learning threshold optimization module, the multi-scale fusion early warning module, and the visualization decision support module work together to achieve efficient identification, accurate early warning, and intelligent auxiliary decision-making for seepage anomalies in water conservancy projects.

[0063] This intelligent seepage monitoring system for a water conservancy project achieves efficient monitoring and intelligent response to abnormal seepage conditions through modular design. First, a multi-source data acquisition module enables real-time and continuous acquisition of various monitoring data, and the introduction of data quality control and environmental compensation mechanisms effectively improves the reliability and stability of the raw data. Second, a dual-window time-series analysis module, by constructing a dual time-window mechanism for rapid response and trend tracking, can simultaneously capture sudden seepage anomalies and potential trend risks, improving the sensitivity and coverage of the response without increasing the burden on sensors. A self-learning threshold optimization module further combines the time-varying characteristics of monitoring data and dynamically optimizes the early warning threshold settings through reinforcement learning, enhancing the system's adaptability under different operating conditions. A multi-scale fusion early warning module deeply integrates multi-source information, outputting graded early warning signals and their confidence evaluations, making the early warning results more interpretable and valuable for decision-making. Finally, through a visualization decision support module, not only is dynamic visual reconstruction of the seepage field achieved, but also, combined with early warning information push and operational suggestions, significantly improves the on-site response efficiency and scientific decision-making capabilities of maintenance personnel. Overall, the system has established a closed-loop linkage mechanism in data acquisition, intelligent analysis, early warning generation and decision support, which can significantly improve the accuracy of water conservancy projects in perceiving seepage risks and their management level.

[0064] Furthermore, the multi-source data acquisition module includes a distributed sensor node array, a data quality assessment unit, and an environmental compensation and correction unit. The distributed sensor node array includes piezometers, flow sensors, water level sensors, soil pressure sensors, and tilt sensors. Each sensor node achieves synchronous data transmission via a LoRa+4G hybrid networking protocol. The data quality assessment unit is configured to perform outlier detection, missing value imputation, and noise filtering on the acquired data. Outlier detection employs a combined discrimination mechanism based on the improved 3σ criterion and box plot method, with the detection threshold dynamically adjusted according to the data distribution characteristics. The environmental compensation and correction unit performs temperature drift compensation and nonlinear correction on the raw sensor data based on environmental parameters, achieving a compensation accuracy of ±0.5% of the measured value, ensuring the consistency and accuracy of the monitoring data within an ambient temperature range of -20℃ to 60℃.

[0065] In this embodiment, the distributed sensor node array covers key monitoring indicators such as seepage pressure, water level, flow rate, soil pressure, and structural tilt, forming a complete seepage state sensing system. Various sensor nodes work collaboratively through a hybrid LoRa and 4G networking protocol, balancing the needs of low-power long-distance communication and high-bandwidth data backhaul. This effectively solves the problem of uneven communication network distribution and improves the stability and real-time performance of data transmission. In terms of data processing, the data quality assessment unit employs a discrimination mechanism combining an improved 3σ rule and box plots, dynamically adjusting thresholds to adapt to data fluctuation characteristics under different monitoring scenarios. This improves the accuracy of abnormal data identification. Simultaneously, interpolation and filtering significantly reduce the impact of data noise and missing data, enhancing the basic reliability of subsequent analysis. Furthermore, the environmental compensation and correction unit can correct temperature drift and adjust nonlinear response based on changes in the ambient temperature of the sensors, with compensation accuracy controlled within ±0.5%, effectively ensuring the consistency and comparability of monitoring data across a wide temperature range.

[0066] Furthermore, the basic window length of the rapid response window is set to 1 to 4 hours, with a sliding step of 15 minutes; the trend analysis window adopts a hierarchical time-weighted strategy, with a basic window length set to 24 to 30 days, extended to 45 days according to seasonal changes; the key statistical indicator extraction includes calculating the mean, standard deviation, coefficient of variation, skewness, kurtosis, first difference, and sliding correlation coefficient of the time series data. Through multi-dimensional joint analysis of statistical indicators, combined with wavelet transform for frequency domain feature extraction, accurate identification and quantitative description of seepage state change patterns are achieved.

[0067] In this embodiment, a dual-window structure and a multi-dimensional statistical index fusion mechanism are introduced in the time-series data analysis stage, significantly improving the accuracy and response efficiency in identifying changes in seepage status in water conservancy projects. The rapid response window, by setting a basic analysis cycle of 1 to 4 hours and a sliding step of 15 minutes, enables high-frequency monitoring and dynamic updates of short-term seepage anomalies, suitable for identifying sudden seepage disturbances, and possessing good real-time performance and sensitivity. The trend analysis window, with a basic cycle of 24 to 30 days, can be extended to 45 days according to the seepage variation patterns of different seasons, employing a layered time-weighted strategy. This allows the system to more effectively capture potential trend changes and periodic fluctuations in long-term monitoring, balancing the timeliness and seasonal adaptability of the seepage process. In terms of data processing, a relatively complete time-series behavior description model is established by extracting various statistical indicators such as mean, standard deviation, coefficient of variation, skewness, kurtosis, first-order difference, and sliding correlation coefficient. By further combining wavelet transform to mine frequency domain features, the hidden periodicity and abrupt change points in seepage signals can be revealed at different time scales, improving the depth and accuracy of identifying abnormal evolution processes.

[0068] Furthermore, the self-learning threshold optimization module employs an improved deep Q-network algorithm to construct a reinforcement learning framework, using LeakyReLU as the activation function to avoid gradient vanishing. The state space includes the normalized value of the current statistical indicator, the moving average trend, and historical threshold performance scores. The action space is defined as a discretized selection of the threshold adjustment range, including four adjustment levels: ±1%, ±3%, ±5%, and ±10%. The reward function is designed as a weighted combination of R = 0.6 × accuracy - 0.3 × false alarm rate - 0.1 × false negative rate. In determining the threshold boundary, a probability density function is first constructed using an improved kernel density estimation method. Then, based on engineering safety margin requirements and historical accident statistics, the optimal quantile combination is intelligently selected as the four-level early warning threshold boundary, and dynamic corrections are made according to actual operating conditions to achieve engineering adaptation of threshold settings and dynamic optimization of seepage risk assessment standards.

[0069] In this embodiment, the self-learning threshold optimization module is built on the improved Deep Q-Network (DQN) framework and uses the LeakyReLU activation function to effectively alleviate the gradient vanishing problem in traditional neural networks during training, thereby ensuring the stability and convergence speed of the learning process. Regarding state space construction, by integrating the normalized values ​​of current statistical indicators, moving average trends, and historical threshold performance scores, multi-angle perception of the monitoring status is achieved. The action space uses discretized adjustment levels to set threshold variation strategies, balancing fine control and computational efficiency, facilitating rapid model response in complex scenarios. The reward function considers accuracy, false alarm rate, and false negative rate through weight combinations, effectively guiding the learning process towards improving early warning quality. For threshold boundary setting, an improved kernel density estimation method is used to establish the probability density distribution of monitoring data, combined with engineering safety margins and historical accident data to select representative quantiles, forming multi-level early warning boundaries. Simultaneously, through real-time feedback during operation, the threshold boundaries can be automatically fine-tuned based on actual performance, thereby achieving continuous optimization of the threshold setting mechanism.

[0070] Furthermore, the multi-source information fusion includes four steps: evidence extraction, credibility allocation, conflict detection, and fusion decision. Specifically: the evidence extraction step obtains anomaly indication evidence from the fast response window and trend analysis window respectively; the credibility allocation step assigns basic probability values ​​based on the historical accuracy of each evidence source, allocating high credibility (0.8-0.9) to sources with accuracy above 90%, medium credibility (0.5-0.7) to sources with accuracy between 70% and 90%, and low credibility (0.2-0.4) to sources with accuracy below 70%; the conflict detection step calculates the conflict coefficient between evidence sources, and initiates a credibility reassessment procedure when the conflict coefficient exceeds 0.6; the fusion decision step uses the standard Dempster synthesis rule to synthesize evidence, ultimately outputting a warning status.

[0071] In this embodiment, a systematic evidence processing procedure is introduced to improve the stability and accuracy of seepage anomaly identification. By extracting anomaly evidence from different time windows and assigning credibility based on the historical performance of each evidence source, the dominant role of high-quality data in early warning judgment is ensured. Simultaneously, a conflict detection mechanism is established to effectively identify and handle discrepancies in judgments from different sources, avoiding the accumulation and amplification of erroneous information. Finally, the Dempster synthesis rule is used to achieve the organic fusion of multi-source evidence, making the early warning results more consistent and engineering-usable.

[0072] Furthermore, the real-time monitoring dashboard adopts a modular interface design, supporting adaptive display of multiple devices. The main interface of the dashboard is divided into a status overview area, a data trend area, an early warning information area, and a device monitoring area. The status overview area displays the current values, changes, and status levels of key monitoring parameters in real time using a dashboard and digital display. The data trend area uses multi-axis line charts and bar charts to display the data change trends of the past 24 hours and 7 days. The early warning information area intuitively displays the current early warning level and the scope of impact through color coding and icon flashing. The device monitoring area displays the online status, battery level, and signal strength of each sensor node in real time.

[0073] In this embodiment, the real-time monitoring dashboard achieves multi-terminal adaptive display through modular design, improving user convenience and information acquisition efficiency in different scenarios. The interface structure is clear, with each function clearly defined, and can reflect the real-time operating status of the monitoring system. The status overview area facilitates quick understanding of core indicators, the data trend area supports trend analysis, the early warning information area enhances response efficiency through visual enhancement, and the equipment monitoring area ensures the visibility of the sensor network's operation, enhancing the system's maintainability and on-site control capabilities.

[0074] Furthermore, the three-dimensional seepage field reconstruction is based on measured seepage pressure data and a simplified engineering geometric model to construct a three-dimensional seepage numerical model. The model adopts regular hexahedral mesh division, and the mesh size is adjusted between 0.2m and 1.0m according to the calculation accuracy requirements and computational resource constraints. The total number of meshes is controlled between 30,000 and 150,000. The permeability coefficient is determined through field test data and geological survey data. The boundary conditions include two types: constant head boundary and impermeable boundary. The numerical solution adopts the over-relaxation iteration method, with the relaxation factor set to 1.2, the convergence accuracy set to 10⁻⁴, and the upper limit of the number of iterations set to 1500. The reconstruction results are visualized in the form of three-dimensional constant head surface, seepage direction vector, and seepage head distribution cloud map.

[0075] In this embodiment, the three-dimensional seepage field reconstruction method establishes a numerical model that balances accuracy and computational efficiency by combining measured data with engineering geometry. Regular mesh generation and control of the total number of meshes effectively balance computational load and simulation accuracy. The reliable source of the permeability coefficient and reasonable boundary conditions ensure the accuracy of the model's physical meaning. The use of an over-relaxation iterative method improves computational convergence efficiency. Finally, the results are visually presented using isostatic head surfaces, flow direction vectors, and head distribution diagrams, enhancing the visualization and engineering interpretation capabilities of the seepage process.

[0076] Figure 2 This is a flowchart illustrating an intelligent monitoring method for seepage in water conservancy projects disclosed in an embodiment of this application.

[0077] like Figure 2 As shown, a method for intelligent monitoring of seepage in water conservancy projects includes the following steps:

[0078] A two-level monitoring architecture of rapid response window and trend analysis window is constructed to obtain real-time seepage monitoring data of the water conservancy project, and the analysis range of each time window is determined according to the time characteristics of the monitoring data.

[0079] Based on the time range of the fast response window and the trend analysis window, the dual-window time series analysis module is controlled to extract the statistical characteristic parameters of the seepage monitoring data in each time window;

[0080] Whenever the dual-window time series analysis module completes the extraction of statistical feature parameters, it marks the statistical feature parameters in the current time window as target analysis parameters, and configures the learning parameters of the self-learning threshold optimization module based on the distribution characteristics of the target analysis parameters.

[0081] Based on the learning parameters and using the self-learning threshold optimization module, a distribution histogram is constructed based on historical monitoring data, and key quantiles are extracted from the distribution histogram as threshold boundary reference values.

[0082] During the threshold boundary reference value extraction process, the self-learning threshold optimization module collects the change characteristics of the target analysis parameters in real time, uses the self-learning mechanism to collect the real-time weight data of each statistical indicator, and combines the threshold boundary reference value and the real-time weight data to identify whether the current weight configuration has reached the preset optimization standard.

[0083] If the current weight configuration does not meet the preset optimization standard, the self-learning mechanism will continue to iterate and optimize until the optimization standard is met.

[0084] If the current weight configuration has reached the preset optimization standard, the dual-window time series analysis module is used to perform weighted fusion of each statistical feature parameter, and the self-learning threshold optimization module generates adaptive threshold parameters and calculates the comprehensive change index.

[0085] During the calculation of the comprehensive change index, the multi-source data acquisition module monitors the change trend of the seepage monitoring data in real time, and the dual-window time series analysis module analyzes whether the comprehensive change index exceeds the safe range of the adaptive threshold parameter.

[0086] If the comprehensive change index does not exceed the safety range, the current monitoring status is maintained and the multi-source data acquisition module is controlled to continue the data acquisition and analysis process;

[0087] If the comprehensive change index exceeds the safe range, the analysis results of the fast response window and the trend analysis window are integrated by the dual-window time series analysis module, and the multi-scale fusion early warning module is used to execute the multi-scale fusion decision-making mechanism to generate a seepage anomaly early warning signal. The visualization decision support module is then controlled to output the corresponding early warning level and emergency response suggestions.

[0088] This intelligent seepage monitoring method for water conservancy projects achieves multi-timescale dynamic monitoring of seepage status by constructing a two-level monitoring architecture of a rapid response window and a trend analysis window. The rapid response window focuses on capturing short-term abnormal fluctuations, ensuring timely response to emergencies; the trend analysis window conducts in-depth analysis of long-term changes, improving the ability to proactively identify potential risks. The dual-window time-series analysis module extracts multi-dimensional statistical feature parameters, providing rich data support for subsequent threshold optimization and risk assessment. The self-learning threshold optimization module constructs a distribution histogram based on historical data and, combined with a real-time weight adjustment mechanism, achieves dynamic adaptation of threshold settings, improving the flexibility and accuracy of the early warning model. During the calculation and analysis of the comprehensive change index, it can determine in real time whether the current seepage status is within a safe range, ensuring the continuity and effectiveness of monitoring. When an abnormal state is identified, combined with the decision-making mechanism of the multi-scale fusion early warning module, it can accurately generate multi-level early warning signals and output clear early warning levels and emergency suggestions through the visualization decision support module, realizing closed-loop management from data collection and analysis to early warning decision-making.

[0089] Furthermore, the construction of a two-tiered monitoring architecture consisting of a rapid response window and a trend analysis window to acquire real-time seepage monitoring data of the water conservancy project, and the determination of the analysis range for each time window based on the temporal characteristics of the monitoring data, includes the following steps:

[0090] Time-frequency domain analysis is performed on historical seepage monitoring data to identify its periodicity and abrupt change characteristics. Then, the optimal duration and overlap ratio of the rapid response window and trend analysis window are calculated and determined to achieve accurate capture and continuous monitoring of the time characteristics of the data.

[0091] Based on the established window parameters, configure a fast response window and a trend analysis window to capture short-term anomalies and long-term trends respectively, while establishing a data synchronization and time alignment mechanism between the two.

[0092] By using a dynamic adjustment mechanism based on data variability and signal-to-noise ratio, the boundaries of the fast response window and trend analysis window are adaptively optimized, thereby intelligently adjusting the window length under different data fluctuations and operating conditions.

[0093] A data stream distributor is established to synchronously distribute the real-time collected seepage monitoring data to the rapid response window and the trend analysis window;

[0094] By tracking the data processing load, analysis accuracy, and response time metrics of each window in real time, the system assesses the statistical stability of the data within the window and the consistency of the analysis results. It also automatically triggers parameter optimization processes when performance degradation or misconfiguration is detected, ensuring that the monitoring architecture always maintains optimal working condition.

[0095] In summary, the two-tiered monitoring architecture, through in-depth time-frequency domain analysis of historical seepage data, determined the duration and overlap ratio of the rapid response window and trend analysis window, achieving precise capture of the periodic and abrupt changes in seepage data. This not only ensures agile response to short-term anomalies but also strengthens robust analysis of long-term trends, improving the comprehensiveness and accuracy of monitoring. By establishing data synchronization and time alignment mechanisms, coordinated processing of data from both windows is guaranteed, avoiding the risk of misjudgment caused by information silos and time misalignments. The dynamic adjustment mechanism, based on data variability and signal-to-noise ratio, endows the monitoring window with adaptive adjustment capabilities, flexibly adjusting the analysis range according to actual data fluctuations and operational status. The introduction of a data stream distributor effectively realizes the synchronous distribution and processing of real-time data, ensuring the efficient operation of the monitoring process. Simultaneously, real-time performance evaluation and automatic parameter optimization functions ensure that the monitoring architecture remains in optimal operating condition, avoiding decreased monitoring accuracy or response delays due to improper configuration.

[0096] Furthermore, based on the time range of the rapid response window and the trend analysis window, the dual-window time series analysis module is controlled to extract the statistical characteristic parameters of the seepage monitoring data within each time window, including the following steps:

[0097] For the rapid response window, the central trend, dispersion and instantaneous changes of the seepage data are quantified by calculating the basic statistical characteristics and rate of change of the seepage data within the window, providing quantitative indicators for short-term anomaly detection;

[0098] For the trend analysis window, by performing fitting analysis and correlation calculation on time series data, its long-term trend, periodicity and seasonality characteristics are extracted, thereby constructing a quantitative description of trend evolution;

[0099] Frequency domain analysis is performed on each time window to extract power spectrum, main frequency and spectral features. Multi-scale time-frequency decomposition is performed through wavelet transform and empirical mode decomposition to construct a comprehensive multi-scale feature vector.

[0100] Based on the statistical analysis of historical abnormal events, information gain, chi-square test and correlation analysis methods are used to evaluate the importance ranking of each characteristic parameter, establish a dynamic weight allocation mechanism, and select the feature subset with the greatest discriminative power for seepage anomalies.

[0101] Quality control is performed on the extracted statistical feature parameters, including stability testing and rationality verification, and uncertainty assessment and confidence intervals are established to ensure the reliability and standardization of the output feature data.

[0102] In summary, the dual-window time-series analysis module extracts statistical features from seepage monitoring data in a stratified manner, enabling precise characterization of short-term anomalies and long-term trends. The rapid response window focuses on calculating basic statistics and rate of change indicators, effectively reflecting instantaneous fluctuations and abnormal signals in seepage data, providing a reliable basis for timely early warning. The trend analysis window reveals the periodicity and seasonality of seepage changes through fitting analysis and correlation calculation. The introduction of frequency domain analysis and multi-scale time-frequency decomposition extracts richer spectral features, enhancing the ability to identify complex signals. Combined with historical anomalies, information gain and chi-square tests are used to dynamically evaluate the importance of features, ensuring the selected feature set has high discriminative power. Strict quality control is implemented on the extracted feature parameters, establishing a stability and rationality verification mechanism, supplemented by uncertainty assessment and confidence interval setting, ensuring the accuracy and usability of the feature data.

[0103] Furthermore, whenever the dual-window time series analysis module completes the extraction of statistical feature parameters, it marks the statistical feature parameters within the current time window as target analysis parameters, and configures the learning parameters of the self-learning threshold optimization module based on the distribution characteristics of the target analysis parameters, including the following steps:

[0104] Establish a completion status monitoring mechanism for the dual-window time series analysis module to detect the feature parameter extraction progress of the fast response window and trend analysis window in real time, and automatically add a unique timestamp and window identifier to all statistical feature parameters in the current time window upon completion;

[0105] Based on the numerical stability, physical meaning, and historical performance of the characteristic parameters, the key parameters most sensitive to changes in seepage state are selected as target analysis parameters.

[0106] The selected target analysis parameters are subjected to probability distribution fitting and statistical characteristic analysis to determine their optimal distribution model and distribution characteristics.

[0107] Based on the distribution characteristics and numerical range of the target analysis parameters, the key learning parameters of the self-learning threshold optimization module are initialized and dynamically adjusted through an adaptive configuration strategy.

[0108] Based on the time correlation and seasonality characteristics of the target analysis parameters, the optimal length of the historical data window and the data weight decay strategy are determined.

[0109] The effectiveness and stability of the learning parameter configuration are verified through rationality checks and small sample tests. An automatic tuning mechanism is established to prevent learning instability. Finally, the configuration is completed and the self-learning threshold optimization module is activated to start the threshold learning and optimization process.

[0110] In summary, by automatically identifying and analyzing key statistical parameters after dual-window feature extraction, the learning process of the self-learning threshold optimization module is precisely driven. First, a completion status monitoring mechanism helps ensure effective integration between the feature parameter extraction process and subsequent learning modules. Second, the selected target analysis parameters have high sensitivity and representativeness, accurately reflecting changes in seepage state and providing stable input for threshold optimization. Distribution fitting and statistical characteristic analysis ensure that the learning parameters have a sound theoretical basis and engineering adaptability. Combining the temporal characteristics of the target parameters with historical data windows and weight decay rules ensures that the learning results take into account both long-term trends and short-term fluctuations, thereby improving the rationality of threshold setting and the accuracy of early warning response. Simultaneously, small-sample testing and rationality verification enhance the robustness and generalization ability of the learning parameters in different scenarios, avoiding bias or overfitting during the learning process.

[0111] Furthermore, the construction of the distribution histogram adopts an adaptive binning strategy. In the initial stage (sample size less than 500), a fixed 15 bins are used. As the sample size increases, the number of bins is dynamically adjusted: 20-35 bins are used when the sample size is 500-2000, 35-60 bins are used when the sample size is 2000-10000, and 60-100 bins are used when the sample size is greater than 10000. At the same time, the bin boundaries are finely adjusted in combination with the skewness and kurtosis characteristics of the data distribution.

[0112] In this embodiment, an adaptive binning strategy is introduced to improve the adaptability and accuracy of the distribution histogram for sample data of different sizes. Using fixed binning when the sample size is small ensures that basic distribution characteristics are captured. As the sample size increases, dynamically increasing the number of bins helps to reveal detailed changes in the data structure. Simultaneously, fine-tuning the bin boundaries based on skewness and kurtosis further enhances the accuracy and stability of distribution modeling, providing a more reliable probabilistic basis for subsequent threshold analysis and anomaly detection.

[0113] Furthermore, the comprehensive change index is calculated using a time series decomposition and reconstruction method, which decomposes the original monitoring data into three parts: trend component, seasonal component, and residual component. The weight coefficient of the trend component is set to 0.5, the weight coefficient of the seasonal component is set to 0.3, and the weight coefficient of the residual component is set to 0.2. The comprehensive change index is calculated by weighted summation, and the index value is standardized to a range of 0 to 1. A time decay factor is introduced in the calculation process. The weight coefficient of recent data is 1.0. Within the fast response window, the weight decays by 0.02 to 0.05 for every 15 minutes forward, and within the trend analysis window, the weight decays by 0.05 to 0.1 for every 24 hours forward.

[0114] In this embodiment, seepage monitoring data is subdivided into three parts—trend, seasonal, and residual—through time series decomposition, and weights are rationally allocated to make the comprehensive change index more accurately reflect the change characteristics at different time scales. The introduction of a time decay factor effectively improves the sensitivity to recent data changes, ensuring a rapid response to sudden anomalies. Simultaneously, the weight decay setting balances short-term fluctuations and long-term trends, achieving a good coordination between stability and sensitivity in the comprehensive index. The standardized index facilitates unified threshold setting and multi-module data fusion, enhancing the dynamic monitoring and accurate judgment of seepage status.

[0115] Furthermore, the judgment criteria for the safe range are determined based on a dynamic threshold mechanism. The upper limit of the safe range is defined as the 95th percentile of historical data multiplied by a seasonal adjustment factor, and the lower limit is defined as the 5th percentile of historical data multiplied by a seasonal adjustment factor. The dynamic threshold is adjusted according to the seasonal variation pattern of historical monitoring data. The threshold adjustment factor is 1.2 to 1.5 in winter, 1.0 to 1.2 in spring, 0.9 to 1.1 in summer, and 0.8 to 1.0 in autumn. To ensure the reliability of the judgment, a dual verification mechanism is adopted: when the comprehensive change index exceeds the safe range determined based on the quantile, it is also verified whether it exceeds the auxiliary judgment range of the historical mean ± 2 times the standard deviation. When the comprehensive change index of 3 consecutive sampling points exceeds the safe range or a single sampling point exceeds the safe range by 1.5 times, it is judged to be outside the safe range.

[0116] In this embodiment, by introducing a seasonally adjusted dynamic quantile threshold, the adaptability of the safety range determination to seasonal changes is effectively improved. The upper and lower limits are set based on the 5% and 95th quantiles of historical data, and dynamically corrected using adjustment coefficients for different seasons, making the determination more consistent with actual operational characteristics. The introduction of a dual verification mechanism enhances sensitivity while suppressing the risk of misjudgment. When anomalies occur continuously or their intensity significantly exceeds the limit, early warnings can be triggered in a timely manner, helping to enhance the identification and stability of abnormal seepage states and ensuring the reliability and practicality of monitoring results.

[0117] Furthermore, the emergency response recommendation generation mechanism is based on a hybrid reasoning system combining an expert knowledge base and machine learning algorithms. The expert knowledge base contains no fewer than 200 rules for handling seepage anomalies, covering response strategies for different anomaly types, different engineering types, and different environmental conditions. The machine learning algorithm employs an ensemble of decision trees and random forests, with the decision tree depth limited to 5 to 15 layers and the random forest containing 50 to 200 decision trees. The emergency response recommendations include four levels: immediate response measures, short-term control plans, long-term remediation plans, and risk assessment reports. Immediate response measures are required to be generated within 5 minutes of receiving the warning signal, short-term control plans within 30 minutes, and long-term remediation plans within 2 hours. The recommendations include specific operational steps, required personnel configuration, estimated response time, and expected effect assessment, ensuring the scientific nature and operability of the emergency response.

[0118] In this embodiment, the emergency response recommendation mechanism combines extensive expert experience with advanced machine learning technology, significantly improving the accuracy and timeliness of seepage anomaly responses. By integrating decision tree and random forest algorithms, it can efficiently handle complex and ever-changing anomaly situations and automatically generate scientifically sound and reasonable response plans. The multi-layered recommendation system covers everything from immediate response to long-term repair, ensuring targeted measures at each stage. Strict time requirements guarantee rapid response and effective control, while detailed operational guidelines and personnel allocation plans enhance the operability and effectiveness of actual implementation, comprehensively strengthening the management capabilities and safety assurance level of seepage anomalies in water conservancy projects.

[0119] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A smart monitoring system for seepage in water conservancy projects, characterized in that, Includes the following modules: The multi-source data acquisition module is used to collect various types of monitoring data in real time through an intelligent sensor network, and to perform quality control and environmental factor compensation. The dual-window time series analysis module constructs a rapid response window and a trend analysis window, respectively extracting key statistical indicators to achieve dual identification of sudden anomalies and long-term trends; The self-learning threshold optimization module automatically extracts key quantile thresholds based on the distribution characteristics of monitoring data and continuously optimizes the weight configuration and early warning threshold settings of various statistical indicators. The multi-scale fusion early warning module generates a comprehensive change index and multi-level early warning status, and outputs seepage anomaly early warning signals and their corresponding confidence levels; The visualization decision support module, based on a real-time monitoring dashboard, three-dimensional seepage field reconstruction, and multi-channel early warning push mechanism, provides intuitive data display, emergency response guidance, and intelligent decision support. Among them, the multi-source data acquisition module, the dual-window time series analysis module, the self-learning threshold optimization module, the multi-scale fusion early warning module, and the visualization decision support module work together to achieve efficient identification, accurate early warning, and intelligent auxiliary decision-making for seepage anomalies in water conservancy projects.

2. The intelligent monitoring system for seepage in water conservancy projects according to claim 1, characterized in that, The multi-source data acquisition module includes a distributed sensor node array, a data quality assessment unit, and an environmental compensation and correction unit. The distributed sensor node array includes piezometers, flow sensors, water level sensors, soil pressure sensors, and tilt sensors. Each sensor node transmits data synchronously via a LoRa+4G hybrid networking protocol. The data quality assessment unit is configured to perform outlier detection, missing value imputation, and noise filtering on the acquired data. Outlier detection employs a combined discrimination mechanism based on the improved 3σ criterion and box plot method, with the detection threshold dynamically adjusted according to the data distribution characteristics. The environmental compensation and correction unit performs temperature drift compensation and nonlinear correction on the raw sensor data based on environmental parameters, achieving a compensation accuracy of ±0.5% of the measured value, ensuring the consistency and accuracy of the monitoring data within an ambient temperature range of -20℃ to 60℃.

3. The intelligent monitoring system for seepage in water conservancy projects according to claim 1, characterized in that, The basic window length of the rapid response window is set to 1 to 4 hours, with a sliding step of 15 minutes. The trend analysis window adopts a stratified time-weighted strategy, with a basic window length set to 24 to 30 days, which is extended to 45 days according to seasonal changes. The key statistical indicator extraction includes calculating the mean, standard deviation, coefficient of variation, skewness, kurtosis, first difference, and sliding correlation coefficient of the time series data. Through multi-dimensional joint analysis of statistical indicators, combined with wavelet transform for frequency domain feature extraction, the accurate identification and quantitative description of seepage state change patterns can be achieved.

4. The intelligent monitoring system for seepage in water conservancy projects according to claim 1, characterized in that, The three-dimensional seepage field reconstruction is based on measured seepage pressure data and a simplified engineering geometric model to construct a three-dimensional seepage numerical model. The model adopts regular hexahedral mesh, and the mesh size is adjusted between 0.2m and 1.0m according to the calculation accuracy requirements and computational resource constraints. The total number of meshes is controlled between 30,000 and 150,000. The permeability coefficient is determined through field test data and geological survey data. The boundary conditions include two types: constant head boundary and impermeable boundary. The numerical solution adopts the over-relaxation iteration method, with the relaxation factor set to 1.2, the convergence accuracy set to 10⁻⁴, and the upper limit of the number of iterations set to 1500. The reconstruction results are visualized in the form of three-dimensional constant head surface, seepage direction vector, and seepage head distribution cloud map.

5. A method for intelligent monitoring of seepage in water conservancy projects, applied to the seepage monitoring system of water conservancy projects as described in any one of claims 1-4, characterized in that, The method includes the following steps: A two-level monitoring architecture of rapid response window and trend analysis window is constructed to obtain real-time seepage monitoring data of the water conservancy project, and the analysis range of each time window is determined according to the time characteristics of the monitoring data. Based on the time range of the fast response window and the trend analysis window, the dual-window time series analysis module is controlled to extract the statistical characteristic parameters of the seepage monitoring data in each time window; Whenever the dual-window time series analysis module completes the extraction of statistical feature parameters, it marks the statistical feature parameters in the current time window as target analysis parameters, and configures the learning parameters of the self-learning threshold optimization module based on the distribution characteristics of the target analysis parameters. Based on the learning parameters and using the self-learning threshold optimization module, a distribution histogram is constructed based on historical monitoring data, and key quantiles are extracted from the distribution histogram as threshold boundary reference values. During the threshold boundary reference value extraction process, the self-learning threshold optimization module collects the change characteristics of the target analysis parameters in real time, uses the self-learning mechanism to collect the real-time weight data of each statistical indicator, and combines the threshold boundary reference value and the real-time weight data to identify whether the current weight configuration has reached the preset optimization standard. If the current weight configuration does not meet the preset optimization standard, the self-learning mechanism will continue to iterate and optimize until the optimization standard is met. If the current weight configuration has reached the preset optimization standard, the statistical feature parameters are weighted and fused using the dual-window time series analysis module, and adaptive threshold parameters are generated and the comprehensive change index is calculated through the self-learning threshold optimization module. During the calculation of the comprehensive change index, the multi-source data acquisition module monitors the change trend of the seepage monitoring data in real time, and the dual-window time series analysis module analyzes whether the comprehensive change index exceeds the safe range of the adaptive threshold parameter. If the comprehensive change index does not exceed the safety range, the current monitoring status is maintained and the multi-source data acquisition module is controlled to continue the data acquisition and analysis process; If the comprehensive change index exceeds the safe range, the analysis results of the fast response window and the trend analysis window are integrated by the dual-window time series analysis module, and the multi-scale fusion early warning module is used to execute the multi-scale fusion decision-making mechanism to generate a seepage anomaly early warning signal. The visualization decision support module is then controlled to output the corresponding early warning level and emergency response suggestions.

6. The intelligent monitoring method for seepage in water conservancy projects according to claim 5, characterized in that, The two-tiered monitoring architecture, consisting of a rapid response window and a trend analysis window, is used to acquire real-time seepage monitoring data of the water conservancy project. The analysis range for each time window is determined based on the temporal characteristics of the monitoring data. This process includes the following steps: Time-frequency domain analysis is performed on historical seepage monitoring data to identify its periodicity and abrupt change characteristics. Then, the optimal duration and overlap ratio of the rapid response window and trend analysis window are calculated and determined to achieve accurate capture and continuous monitoring of the time characteristics of the data. Based on the established window parameters, configure a fast response window and a trend analysis window to capture short-term anomalies and long-term trends respectively, while establishing a data synchronization and time alignment mechanism between the two. By using a dynamic adjustment mechanism based on data variability and signal-to-noise ratio, the boundaries of the fast response window and trend analysis window are adaptively optimized, thereby intelligently adjusting the window length under different data fluctuations and operating conditions. A data stream distributor is established to synchronously distribute the real-time collected seepage monitoring data to the rapid response window and the trend analysis window; By tracking the data processing load, analysis accuracy, and response time metrics of each window in real time, the system assesses the statistical stability of the data within the window and the consistency of the analysis results. It also automatically triggers parameter optimization processes when performance degradation or misconfiguration is detected, ensuring that the monitoring architecture always maintains optimal working condition.

7. The intelligent monitoring method for seepage in water conservancy projects according to claim 5, characterized in that, Based on the time range of the rapid response window and the trend analysis window, the dual-window time series analysis module is controlled to extract the statistical characteristic parameters of the seepage monitoring data within each time window, including the following steps: For the rapid response window, the central trend, dispersion and instantaneous changes of the seepage data are quantified by calculating the basic statistical characteristics and rate of change of the seepage data within the window, providing quantitative indicators for short-term anomaly detection; For the trend analysis window, by performing fitting analysis and correlation calculation on time series data, its long-term trend, periodicity and seasonality characteristics are extracted, thereby constructing a quantitative description of trend evolution; Frequency domain analysis is performed on each time window to extract power spectrum, main frequency and spectral features. Multi-scale time-frequency decomposition is performed through wavelet transform and empirical mode decomposition to construct a comprehensive multi-scale feature vector. Based on the statistical analysis of historical abnormal events, information gain, chi-square test and correlation analysis methods are used to evaluate the importance ranking of each characteristic parameter, establish a dynamic weight allocation mechanism, and select the feature subset with the greatest discriminative power for seepage anomalies. Quality control is performed on the extracted statistical feature parameters, including stability testing and rationality verification, and uncertainty assessment and confidence intervals are established to ensure the reliability and standardization of the output feature data.

8. The intelligent monitoring method for seepage in water conservancy projects according to claim 5, characterized in that, Whenever the dual-window time series analysis module completes the extraction of statistical feature parameters, the statistical feature parameters within the current time window are marked as target analysis parameters. Based on the distribution characteristics of the target analysis parameters, the learning parameters of the self-learning threshold optimization module are configured, including the following steps: Establish a completion status monitoring mechanism for the dual-window time series analysis module to detect the feature parameter extraction progress of the fast response window and trend analysis window in real time, and automatically add a unique timestamp and window identifier to all statistical feature parameters in the current time window upon completion; Based on the numerical stability, physical meaning, and historical performance of the characteristic parameters, the key parameters most sensitive to changes in seepage state are selected as target analysis parameters. The selected target analysis parameters are subjected to probability distribution fitting and statistical characteristic analysis to determine their optimal distribution model and distribution characteristics. Based on the distribution characteristics and numerical range of the target analysis parameters, the key learning parameters of the self-learning threshold optimization module are initialized and dynamically adjusted through an adaptive configuration strategy. Based on the time correlation and seasonality characteristics of the target analysis parameters, the optimal length of the historical data window and the data weight decay strategy are determined. The effectiveness and stability of the learning parameter configuration are verified through rationality checks and small sample tests. An automatic tuning mechanism is established to prevent learning instability. Finally, the configuration is completed and the self-learning threshold optimization module is activated to start the threshold learning and optimization process.

9. The intelligent monitoring method for seepage in water conservancy projects according to claim 5, characterized in that, The distribution histogram is constructed using an adaptive binning strategy. In the initial stage (sample size less than 500), a fixed 15 bins are used. As the sample size increases, the number of bins is dynamically adjusted: 20-35 bins are used when the sample size is 500-2000, 35-60 bins are used when the sample size is 2000-10000, and 60-100 bins are used when the sample size is greater than 10000. At the same time, the bin boundaries are finely adjusted based on the skewness and kurtosis characteristics of the data distribution.

10. The intelligent monitoring method for seepage in water conservancy projects according to claim 5, characterized in that, The comprehensive change index is calculated using a time series decomposition and reconstruction method, which decomposes the original monitoring data into three parts: trend component, seasonal component, and residual component. The weight coefficient of the trend component is set to 0.5, the weight coefficient of the seasonal component is set to 0.3, and the weight coefficient of the residual component is set to 0.

2. The comprehensive change index is calculated by weighted summation, and the index value is standardized to a range of 0 to 1. The calculation process incorporates a time decay factor. The weight coefficient for recent data is 1.

0. Within the fast response window, the weight decays by 0.02 to 0.05 for every 15 minutes forward. Within the trend analysis window, the weight decays by 0.05 to 0.1 for every 24 hours forward.

Images