Hydraulic engineering safety monitoring method and system based on data processing

Abstract

The invention provides a water conservancy project safety monitoring method and system based on data processing, and relates to the technical field of monitoring, and the method comprises the steps: obtaining and carrying out the multi-dimensional preprocessing of water conservancy project multi-source heterogeneous monitoring data through the deployment of a sensor network, and extracting multi-scale space-time fusion features from the data; performing structural state modeling, anomaly prediction, risk assessment and early warning by using a long-short-term memory neural network model integrated with a multi-head attention mechanism; and intelligent suggestions oriented to maintenance decisions are generated, so that comprehensive, accurate and prospective evaluation and early warning of the structural state of the water conservancy project are finally realized, the exception identification and risk prediction capabilities are effectively improved, the false alarm rate is reduced, refined and initiative intelligent maintenance decisions are provided, resource allocation is optimized, and the service life of the project is prolonged.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of monitoring, and particularly relates to a water conservancy project safety monitoring method and system based on data processing. BACKGROUND

[0002] With the continuous construction and operation of water conservancy projects worldwide, they play a crucial role in flood control, power generation, irrigation, water supply and other fields of national economy and social development. However, water conservancy projects generally have the characteristics of large scale, long service period, complex structure and variable operating environment, and are easily affected by natural factors (such as earthquakes, floods, geological changes) and human factors (such as construction defects, operating loads), thus causing structural deformation, leakage, cracking and other safety hazards. Therefore, it is a crucial task to conduct normalized and refined safety monitoring of water conservancy projects to ensure their long-term stable operation and timely warning of potential risks.

[0003] In the prior art, water conservancy project safety monitoring usually relies on data collection by laying sensors, combined with periodic inspection and manual interpretation. However, with the increasing density of sensor networks and the diversification of monitoring parameters, the amount of data generated has shown explosive growth, and the data types are diverse and complex. Traditional offline, manual experience-based data processing methods cannot meet the real-time analysis needs of massive, heterogeneous data, resulting in that the monitoring data cannot be fully utilized, and potential safety hazards cannot be timely and accurately identified and predicted.

[0004] These limitations make the existing safety monitoring system have obvious deficiencies in dynamic response, fault diagnosis accuracy and intelligent level of risk assessment, which seriously affects the efficiency of water conservancy project safety management and the scientificity of decision-making. Therefore, how to build an efficient and intelligent data processing method and system to comprehensively improve the accuracy, real-time performance and early warning capability of water conservancy project safety monitoring is a key technical problem to be solved in the current field. SUMMARY

[0005] The present application is proposed in view of the many limitations of existing water conservancy project safety monitoring technologies in data acquisition, processing, analysis, and early warning decision-making. Existing technologies generally rely on decentralized and heterogeneous sensor networks, which collect a variety of data types with different sampling frequencies, resulting in serious data fragmentation and making it difficult to conduct comprehensive and real-time fusion analysis. Traditional data processing methods mainly focus on independent analysis of single data modalities or use simple statistical thresholds for judgment, ignoring the dynamic characteristics and deep correlations of water conservancy structures under complex environmental loads and multi-factor coupling. This makes the existing monitoring system insufficient in identifying structural abnormalities, predicting potential risks, and providing accurate early warnings, resulting in high false positive rates, high risk of missed reports, poor timeliness of early warnings, and difficulty in effectively identifying structural deterioration trends and potential damage patterns. Therefore, existing technologies cannot provide fine-grained and forward-looking safety state evaluation and maintenance decision support, leading to passive safety management and low resource allocation efficiency.

[0006] To solve the above technical problems, the present application provides a water conservancy project safety monitoring method and system based on data processing. The present application realizes comprehensive, accurate, and forward-looking evaluation and early warning of the health state of water conservancy structures by constructing a multi-scale, multi-modal, and deep fusion data processing and analysis framework. The present application integrates multi-source heterogeneous monitoring data through a unified data acquisition and preprocessing mechanism; extracts structured and physically meaningful spatio-temporal features from raw data through advanced feature engineering methods; performs multi-modal feature fusion and structure state evolution modeling through deep learning models to achieve accurate detection of abnormal states and effective prediction of future risks; finally, based on risk assessment and early warning level division, generates intelligent suggestions for maintenance decisions, thereby significantly improving the intelligence, refinement, and initiative level of water conservancy project safety monitoring.

[0007] According to one aspect of the present application, a water conservancy project safety monitoring method based on data processing is provided, which includes the following steps:

[0008] Acquiring multi-source heterogeneous monitoring data of water conservancy projects: Through the deployment of sensor networks in water conservancy project structures, surrounding environments, and operating conditions monitoring points, real-time collection and transmission of multiple types of monitoring data are achieved, including but not limited to structural displacement data, structural stress and strain data, seepage data, temperature data, humidity data, water level data, reservoir capacity data, rainfall data, and seismic load data. Structural displacement data is obtained through high-precision global positioning system displacement sensors, deflection meters, and inclinometers. Structural stress and strain data is obtained through fiber Bragg grating strain sensors and resistance strain gauges. Seepage data is obtained through piezometers, flow meters, and groundwater level gauges. Temperature and humidity data are obtained through temperature and humidity sensors. Water level and reservoir capacity data are obtained through ultrasonic water level meters or radar water level meters. Rainfall data is obtained through tipping bucket rain gauges. Seismic load data is obtained through acceleration sensors.

[0009] Multi-dimensional time series preprocessing of multi-source heterogeneous monitoring data: Preprocessing includes timestamp synchronization, missing value interpolation, outlier detection and correction, and data normalization. Timestamp synchronization is achieved through distributed high-precision time protocol to ensure strict alignment of different types of sensor data on the time axis. Missing value interpolation uses a long short-term memory neural network prediction model based on historical data patterns or a cubic spline interpolation method to accurately fill in missing data points. Outlier detection and correction uses an isolation forest or local outlier factor algorithm to identify and correct data points that deviate from the normal range. Data normalization uses the min-max normalization method to map data of different dimensions to the range of zero to one. The preprocessing operation aims to generate a set of unified, clean, and time-aligned structured time series data.

[0010] Extracting multi-scale spatio-temporal fusion features from structured time series data sets: Feature extraction includes the following sub-steps:

[0011] Extracting basic statistical features, including the mean, variance, skewness, kurtosis, and extreme value of each sensor data within a pre-set time window.

[0012] Extracting frequency domain features, including performing fast Fourier transform or wavelet transform on vibration sensor data to obtain its main frequency, amplitude, and energy distribution characteristics.

[0013] Extracting spatial correlation features, including calculating the Pearson correlation coefficient or mutual information entropy of displacement, stress, and seepage data between adjacent sensors or key structural points to reflect the coordinated deformation and force relationship of the structure as a whole or local area.

[0014] Temporal evolution features are extracted, including calculating trend items, seasonal items and residual items of data through sliding time windows to construct features reflecting long-term deformation trends and periodic changes of structures. Temporal evolution features use time convolution network model for multi-scale feature learning to capture sequence dependence and patterns at different time scales.

[0015] The multi-scale spatio-temporal fusion features are input into a multi-modal deep learning model for structure state modeling and anomaly prediction: the multi-modal deep learning model adopts a long short-term memory neural network architecture integrating a multi-head attention mechanism, and its input layer receives fusion features extracted from different sensor types and different time scales, and the multi-head attention mechanism is used to adaptively weight the importance of different feature channels. The long short-term memory neural network layer is used to capture the long-term time dependence and nonlinear evolution law of structure state data, and a dynamic baseline model of water conservancy projects under normal operation state is established. Anomaly prediction is performed by calculating the Mahalanobis distance or reconstruction error between the current real-time fusion features and the prediction output of the dynamic baseline model to determine whether there is a structure state deviation beyond the normal fluctuation range. The anomaly prediction model can distinguish between normal response and potential signs of deterioration, and identify early weak abnormal signals.

[0016] Risk assessment and warning level classification are performed on the anomaly prediction results: risk assessment includes quantifying the risk of the identified abnormal state by combining the design standards, operation specifications, historical failure data and bearing capacity thresholds of water conservancy projects. Risk assessment is based on Bayesian network or fault tree analysis to calculate the probability of abnormal events and their impact on the overall safety of the project. According to the risk assessment results, the safety risk is divided into four levels: first-level warning, second-level warning, third-level warning and fourth-level warning, and the warning levels correspond to extremely low risk, low risk, medium risk and high risk, respectively. The warning level threshold is related to the safety margin of the engineering structure and the potential economic loss and personnel casualty risk, and is identified by color coding and priority sorting.

[0017] Intelligent suggestions for maintenance decisions are generated and output: intelligent suggestions are based on risk assessment results and warning levels, and combined with a pre-defined maintenance strategy knowledge base to provide specific and actionable maintenance measure suggestions for managers. Suggestions include but are not limited to immediate stop operation, limit operation load, strengthen on-site inspection, arrange expert consultation, start emergency rescue plan or conduct targeted repair. Intelligent suggestions also include prediction of the evolution trend of abnormal state and give specific intervention opportunities to achieve preventive maintenance and resource optimization. Intelligent suggestions are displayed to users through a visual interface and can be sent to the terminal devices of relevant responsible persons simultaneously.

[0018] According to a preferred embodiment of the present application, in the multi-dimensional time series preprocessing, the precision of timestamp synchronization reaches the microsecond level to ensure accurate capture of transient events. The error of missing value imputation is controlled within 0.1%. The outlier detection and correction can identify and correct data points with an amplitude exceeding three standard deviations.

[0019] According to a preferred embodiment of the present application, in the multi-scale spatio-temporal fusion feature extraction, the wavelet transform of frequency domain feature extraction uses Daubechies wavelet basis function, and the decomposition layer is set to five layers to capture the energy distribution of different frequency components. The calculation interval of spatial correlation features is set to every five meters of sensor array.

[0020] According to a preferred embodiment of the present application, the multi-modal deep learning model includes twenty-four layers of long short-term memory neural network, each layer containing one hundred and twenty-eight memory cells, and the multi-head attention mechanism is set to eight attention heads. The model is optimized using adaptive moment estimation optimization algorithm during training, with a learning rate of 0.0001, a batch size of 32, and a training period of 300 iterations. The training data set of the model comes from at least three years of normal operation monitoring data of water conservancy projects, which has been strictly cleaned and labeled.

[0021] According to a preferred embodiment of the present application, in the risk assessment and warning level division, the first level warning corresponds to a structural safety margin greater than 50%, the second level warning corresponds to a structural safety margin between 30% and 50%, the third level warning corresponds to a structural safety margin between 10% and 30%, and the fourth level warning corresponds to a structural safety margin less than 10%.

[0022] According to another aspect of the present application, a water conservancy project safety monitoring system based on data processing is provided, which comprises:

[0023] The data acquisition unit is used to acquire multi-source heterogeneous monitoring data of the water conservancy project. The data acquisition unit includes a plurality of high-precision sensors and a data transmission module. The sensors are arranged at monitoring points of the water conservancy project structure body, the surrounding environment and the operating condition. The sensors include global positioning system displacement sensors, deflection gauges, inclinometers, fiber Bragg strain sensors, resistance strain gauges, osmotic pressure gauges, flowmeters, groundwater level gauges, temperature and humidity sensors, ultrasonic water level gauges, radar water level gauges, tipping bucket rain gauges and acceleration sensors. The data transmission module transmits the collected monitoring data to the data processing server in real time through optical fiber communication or fifth generation mobile communication technology.

[0024] The data preprocessing module is connected with the data acquisition unit and is used for multi-dimensional time series preprocessing of multi-source heterogeneous monitoring data. The data preprocessing module includes a timestamp synchronizer, a missing value imputer, an outlier detector, and a data normalizer. The timestamp synchronizer realizes time alignment of data based on a high-precision time protocol. The missing value imputer performs missing data filling based on a long short-term memory neural network prediction model or a cubic spline interpolation algorithm. The outlier detector performs abnormal data identification based on an isolation forest or a local anomaly factor algorithm. The data normalizer performs minimum-maximum value normalization processing.

[0025] The feature engineering module is connected with the data preprocessing module and is used for extracting multi-scale spatio-temporal fusion features from a structured time series data set. The feature engineering module includes a statistical feature extractor, a frequency domain feature extractor, a spatial correlation feature extractor, and a time series evolution feature extractor. The statistical feature extractor calculates the mean, variance, skewness, kurtosis, and extreme value of data within a preset time window. The frequency domain feature extractor performs fast Fourier transform or wavelet transform. The spatial correlation feature extractor calculates the Pearson correlation coefficient or mutual information entropy between sensor data. The time series evolution feature extractor uses a time convolution network model for multi-scale feature learning.

[0026] The structure state modeling and anomaly prediction module is connected with the feature engineering module and is used for inputting multi-scale spatio-temporal fusion features into a multi-modal deep learning model for structure state modeling and anomaly prediction. The structure state modeling and anomaly prediction module includes a pre-trained multi-head attention mechanism long short-term memory neural network model, which is used to establish a dynamic baseline model of the water conservancy project structure state, and calculate the Mahalanobis distance or reconstruction error between the real-time fusion features and the predicted output of the dynamic baseline model to identify the structure abnormal state.

[0027] The risk assessment and early warning module is connected with the structure state modeling and anomaly prediction module and is used for risk assessment and early warning level division of the anomaly prediction result. The risk assessment and early warning module includes a risk assessment engine and a warning level classifier. The risk assessment engine quantifies risks based on the design standards, operation specifications, historical failure data, and bearing capacity threshold of the water conservancy project, using a Bayesian network or fault tree analysis method. The warning level classifier divides the safety risk into four levels of first-level warning, second-level warning, third-level warning, and fourth-level warning according to the risk assessment result.

[0028] The decision support module is connected with the risk assessment and early warning module and is used for generating and outputting intelligent suggestions for maintenance decisions. The decision support module includes a maintenance strategy knowledge base and a suggestion generator. The maintenance strategy knowledge base stores predefined maintenance measures and intervention schemes. The suggestion generator retrieves and generates specific maintenance measure suggestions from the maintenance strategy knowledge base according to the risk assessment result and the warning level, and predicts the evolution trend and intervention timing of the abnormal state.

[0029] The visualization and interaction module is connected with the data acquisition unit, the data preprocessing module, the feature engineering module, the structure state modeling and anomaly prediction module, the risk assessment and early warning module and the decision support module, and is used for providing a graphical display interface of data, analysis results, early warning information and decision suggestions.

[0030] Compared with the prior art, the application has the advantages and positive effects that:

[0031] The application realizes synchronous, continuous and high-resolution data acquisition of water conservancy engineering structures and environmental parameters by constructing a comprehensive monitoring network composed of high-precision and multiple types of sensors, solves the problem of incomplete data acquisition and time and space inconsistency in the prior art, and effectively improves the quality and availability of original data by using advanced timestamp synchronization, missing value interpolation, anomaly detection and correction and data normalization methods, thereby laying a solid foundation for subsequent in-depth analysis.

[0032] The application introduces a multi-scale spatiotemporal feature extraction mechanism, not only calculates basic statistical and frequency domain features, but also deeply mines the internal laws and complex dynamic behaviors of water conservancy engineering structures under the coupling action of multiple factors through spatial correlation features and time series evolution features. In particular, the use of time convolution network for multi-scale feature learning significantly enhances the ability to capture deep patterns in data, surpassing the simple feature engineering methods of the prior art.

[0033] The application uses a long short-term memory neural network integrated with a multi-head attention mechanism as a core multi-modal deep learning model, realizes deep fusion and adaptive weight distribution of heterogeneous and multi-dimensional features, and effectively overcomes the limitations of traditional methods in data fusion. The model can establish an accurate dynamic baseline of water conservancy engineering under normal operating conditions, and perform high-sensitivity anomaly detection through Mahalanobis distance or reconstruction error, significantly reducing the false alarm rate and improving the identification accuracy of initial anomaly signals.

[0034] The application converts the anomaly prediction results into intuitive and operable safety risk levels based on quantitative risk assessment and multi-level early warning mechanism, and provides a clear early warning threshold and identification method. This enables management personnel to quickly respond according to the risk level, avoiding the drawbacks of unclear early warning information and difficult decision-making in the prior art.

[0035] The application further provides intelligent suggestions for maintenance decisions, which not only indicate specific maintenance measures, but also predict the evolution trend of abnormal states and the best intervention time, realizing the fundamental change from passive emergency rescue to active preventive maintenance. This greatly optimizes the allocation efficiency of maintenance resources, prolongs the service life of the water conservancy project, and effectively guarantees the long-term safe operation of the project. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 is the overall technical scheme architecture diagram of the application;

[0037] Figure 2 is the core principle framework diagram of the multi-modal deep fusion structure state modeling and abnormal prediction of the application. DETAILED DESCRIPTION

[0038] Please refer to Figure 1 and Figure 2 , the application proposes a water conservancy project safety monitoring method and system based on data processing, aiming to overcome the limitations of traditional water conservancy project safety monitoring technology in data processing, multi-source information fusion and abnormal behavior prediction. Existing monitoring schemes often rely on single modal data analysis or simple threshold alarm mechanism, which is difficult to fully reflect the complex mechanical behavior and environmental coupling effect of water conservancy project structure. These traditional methods are inefficient in processing massive, heterogeneous sensor data, and lack the ability to capture nonlinear and time-varying characteristics, resulting in frequent false positives and false negatives, affecting the timeliness and accuracy of safety warning. In addition, for potential structural degradation trends and hidden failure modes, traditional methods lack deep learning and multi-modal fusion capabilities, and cannot effectively identify and predict early, posing a major risk to the safe operation of water conservancy projects. The core idea of the application is to build a multi-modal deep fusion structure state modeling and abnormal prediction framework, use advanced data processing technology to realize accurate and intelligent monitoring and early warning of the safety state of water conservancy projects, and improve the intelligent level of water conservancy project safety management.

[0039] The application proposes a water conservancy project safety monitoring method based on data processing, which shows the overall technical scheme architecture of a process of systematically collecting, processing, analyzing water conservancy project monitoring data, and modeling and predicting abnormal states. The core principle framework of multi-modal deep fusion structure state modeling and abnormal prediction shows the specific technical details of the method in data fusion and model construction. The implementation of the method includes the following steps:

[0040] S101: Multi-source heterogeneous monitoring data collection;

[0041] S102: Monitoring data preprocessing and feature engineering;

[0042] S103: Multimodal data deep fusion representation learning;

[0043] S104: Structural state time series modeling and health assessment;

[0044] S105: Structural abnormal behavior pattern recognition and prediction;

[0045] S106: Safety risk level determination and early warning information generation.

[0046] As an embodiment of the present application, S101, multi-source heterogeneous monitoring data collection, specifically includes:

[0047] This step aims to obtain comprehensive, real-time, multimodal monitoring data from various monitoring subsystems of water conservancy projects, providing a data basis for subsequent safety assessment and early warning. Water conservancy projects usually cover various complex structures such as dams, embankments, sluices, and hydropower stations, and their safety state is influenced by multiple factors such as structural response, environmental load, and operating conditions. Traditional monitoring data collection systems may have problems such as single data type, mismatched sampling frequency, and uneven data quality, which may result in incomplete and comprehensive structural health portraits. Therefore, the present application deploys and integrates various advanced sensors and data collection equipment to achieve high-precision, synchronized collection of different physical quantities.

[0048] Specifically, the data collection process can be divided into the following sub-steps:

[0049] S1011: Structural response data acquisition: through sensors deployed in key positions of the main structure of the water conservancy project, such as the dam body, gate structure, powerhouse foundation, etc., real-time acquisition of physical quantity data such as deformation, stress, strain, vibration, and inclination of the structure.

[0050] Deformation data acquisition is achieved by deploying high-precision total station, global navigation satellite system receiver, laser range finder or fiber Bragg grating displacement sensor. Total station measures the coordinates of observation points periodically to calculate their three-dimensional displacement relative to the reference point. Global navigation satellite system receiver provides continuous and high-precision three-dimensional coordinate time series with sampling frequency up to hertz level or even higher, reflecting the slow deformation and transient displacement of the structure. Laser range finder measures the distance between two points by non-contact method, suitable for deformation monitoring in narrow space or dangerous area. Fiber Bragg grating displacement sensor uses the characteristic of fiber sensitivity to external strain to obtain local deformation of the structure by demodulating the change of grating center wavelength, with superior distributed monitoring capability and anti-electromagnetic interference performance. The raw data output by these sensors includes coordinate values, displacement difference values, center wavelength drift values, etc., usually stored in floating-point form with accurate time stamps. Data verification mechanisms include satellite signal quality check for global navigation satellite system raw observation values, multipath effect evaluation, and spectrum drift correction and temperature compensation for fiber Bragg grating sensors. Typical value ranges are determined according to engineering scale and monitoring requirements, for example, dam crest displacement may be in the range of several millimeters to several tens of centimeters, while local crack opening may be in the range of microns to millimeters.

[0051] Stress-strain data acquisition is achieved by embedding or pasting resistance strain gauges, vibrating wire strain gauges, fiber Bragg grating strain gauges, etc. Resistance strain gauges convert the deformation caused by structural stress into resistance change, and then calculate the stress value, suitable for accurate measurement of local stress concentration area. Vibrating wire strain gauges reflect strain by measuring frequency change, with good long-term stability and small temperature drift. Fiber Bragg grating strain gauges provide high-resolution and distributed strain monitoring capability. These sensors usually output current, voltage or frequency signals at a specific sampling frequency, such as once per minute to several tens of times per second, which are converted to digital signals by data acquisition modules. Data verification includes sensor zero drift correction, temperature compensation algorithm application, and cross-correlation analysis of multi-point synchronous acquisition data. Typical value range is within the allowable stress-strain range of the structure material, for example, strain of concrete structure is usually in the order of micro-strain, while strain of steel structure may be higher.

[0052] Vibration data acquisition is achieved through acceleration sensors such as piezoelectric accelerometers, servo-accelerometers, and velocity sensors such as electromagnetic tachometers. These sensors can capture the dynamic response of the structure under environmental loads such as wind loads, water flow impacts, and seismic actions, with a high sampling frequency, often up to hundreds or even thousands of hertz, to accurately record transient vibration signals. The output data is a time series of acceleration or velocity, with units corresponding to meters per second squared or meters per second, respectively. Data verification involves baseline drift correction, noise filtering, and spectral analysis to identify abnormal frequency components. Typical values range from zero-point-zero-zero meters per second squared for ambient micro-vibrations to several meters per second squared under seismic action.

[0053] Tilt data acquisition measures the attitude changes of the structure through tiltmeters such as inclinometers, pendulum tiltmeters, and other devices. Inclinometers measure the tilt angle in a vertical hole, reflecting deep displacement of the structure, and are often used to monitor landslides or foundation settlement. Pendulum tiltmeters directly measure the tilt angle of the structure surface relative to the direction of gravity. The data output is an angle value, expressed in degrees or radians, with a relatively low sampling frequency, usually once an hour or longer time intervals. Data verification includes zero-point calibration, temperature drift correction, and long-period stability analysis. Typical values range from micro-radians to several degrees, reflecting the overall tilt and rotation of the structure.

[0054] S1012: Environmental load data acquisition: Use environmental sensors to collect external load information that affects the safety of hydraulic engineering, such as temperature, humidity, wind speed, wind direction, rainfall, water level, reservoir capacity, seepage flow and uplift pressure, etc.

[0055] Temperature and humidity data acquisition is achieved through temperature and humidity sensors such as resistance temperature and humidity sensors, capacitance temperature and humidity sensors, and other devices deployed inside the engineering area such as inside the concrete, on the surface of the steel reinforcement, and in the external environment. These sensors provide accurate environmental temperature and relative humidity data, which are crucial for evaluating the thermal expansion and contraction of structural materials, the hydration heat effect of concrete, and the operating status of equipment. The data output is in degrees Celsius or Fahrenheit and relative humidity percentage. Data verification includes regular calibration, comparison with other weather station data, and detection of abnormal mutations. Typical values cover the normal fluctuation range of local climate, such as minus forty degrees Celsius to plus sixty degrees Celsius, and ten percent to ninety-five percent relative humidity.

[0056] Wind speed and direction data acquisition is achieved by installing wind speed and direction instruments such as ultrasonic wind speed and direction instruments, mechanical wind speed and direction instruments, etc. on high places or windward surfaces of the project to capture the influence of wind load on structures such as high towers and cable bridges. Ultrasonic wind speed and direction instruments measure wind speed and direction using ultrasonic propagation time difference, with advantages such as no mechanical wear and tear, fast response, etc. The data output is wind speed in meters per second or kilometers per hour and wind direction in angular value. Data verification includes wind cup speed calibration, wind vane calibration, and data robustness evaluation under extreme weather conditions. The typical value range is from zero to twenty meters per second or higher.

[0057] Rainfall data acquisition is achieved by deploying rain gauges such as tipping bucket rain gauges, siphon rain gauges, etc. in the upstream basin and monitoring area of the project. Rainfall data is an important basis for evaluating hydrological loads such as floods and debris flows. Tipping bucket rain gauges record rainfall by measuring the number of times the water bucket is turned over, with advantages such as simple structure and easy maintenance. The data output is in millimeters or inches. Data verification includes funnel cleaning, winter heating to prevent freezing, and comparison with data from other hydrological stations. The typical value range is from zero to several hundred millimeters per day.

[0058] Water level and storage capacity data acquisition is achieved by using water level meters such as float type water level meters, piezoresistive water level meters, radar water level meters, etc. and combining with topographic storage capacity curves for calculation. Water level data directly reflects the pressure of water on the structure, and storage capacity data reflects the water resource reserve situation. Radar water level meters measure non-contact, with advantages such as high precision and no influence from water quality. The data output is water level elevation in meters or feet and storage capacity in cubic meters. Data verification includes zero point calibration, temperature compensation, and comparison with other observation methods such as manual measurement results. The typical value range is determined according to the design water level of the reservoir.

[0059] Seepage flow and uplift pressure data acquisition is achieved by using seepage pressure meters, pressure measuring tubes, etc. to monitor the seepage and uplift pressure distribution of dam bodies or foundations. Seepage and uplift pressure are key factors affecting the stability of dams, and abnormal changes may indicate the penetration failure of the project. The data output is seepage flow in liters per second or cubic meters per day and uplift pressure in pascals or meters of water head. Data verification includes sensor sealing inspection, temperature correction, and long-term trend analysis. The typical value range is determined according to the design seepage standard of the project.

[0060] S1013: Operation condition data acquisition: Collects the operation parameters of the water conservancy project, such as the running state of the unit, the opening degree of the gate, the power generation, the discharge, etc. These data reflect the actual operation load and operation behavior of the project.

[0061] The unit operation state data is obtained through the operation monitoring system of the generator set, including key parameters such as unit start-stop time, output size, speed, vibration, bearing temperature, etc. These data reflect the dynamic influence of unit load on the structure of the plant. The data output is the Boolean start-stop state, megawatt output, revolutions per minute speed, and Celsius temperature. Data verification includes synchronization check with the monitoring and data acquisition system data, and sensor fault detection.

[0062] The gate opening data is obtained through gate opening sensors such as pull-wire displacement sensors, encoders, etc. The gate opening directly affects the water discharge, and in turn changes the downstream scouring and the impact of water flow on the structure. The data output is in meters or percentage opening. Data verification includes periodic calibration and comparison with hydraulic system feedback values.

[0063] The power generation and discharge data is obtained through flow meters such as ultrasonic flow meters, electromagnetic flow meters, etc., or based on a calculation model of water head and gate opening. These data reflect the utilization and scheduling of water resources, and are closely related to water pressure loads. The data output is in megawatt-hours or cubic meters per second. Data verification includes meter calibration and coordinated analysis with other hydrological observation data.

[0064] All the above types of data are transmitted to the data aggregation layer through a unified data acquisition interface and communication protocol such as industrial Ethernet, transmission control protocol / internet protocol, message queue telemetry transmission, etc., for preliminary storage and time stamp synchronization. To ensure data integrity and consistency, the acquisition system implements a multi-level redundant backup mechanism, and encrypts the data transmission process to prevent data loss or tampering. In addition, the system sets up an intelligent filtering module to perform preliminary quality checks on the raw data collected, such as removing obvious transmission errors or sensor offline data, to ensure data availability.

[0065] As an embodiment of the present application, S102, the monitoring data preprocessing and feature engineering, specifically includes:

[0066] This step aims to convert the raw multi-source heterogeneous monitoring data obtained in S101 into high-quality, uniformly formatted feature vectors that can be used for deep learning model training. Raw data often has problems such as noise, missing, abnormal, inconsistent sampling frequency, and redundancy, and direct use in model training will seriously affect the accuracy and generalization ability of the model. In addition, different types of data such as time series, categorical data, and numerical data have different statistical characteristics and expression forms, and need to be properly extracted and converted to better reveal the structural safety information contained in the data. Traditional preprocessing methods may not effectively handle large-scale, multi-modal data, and feature engineering often relies on the experience of domain experts, which is inefficient and difficult to adapt to complex and changing water conservancy engineering scenarios.

[0067] Specifically, the data preprocessing and feature engineering process can be subdivided into the following sub-steps:

[0068] S1021: Data cleaning and missing value handling:

[0069] Data cleaning involves identifying and correcting or removing obvious errors and noise in the raw data. For example, for structural response data, there can be spike noise or baseline drift due to sensor transient failure, which can be smoothed by median filtering, Kalman filtering, or wavelet denoising. For environmental load data, meteorological mutations can cause short-time data anomalies, which need to be verified in combination with historical data and physical models. Data validation includes outlier detection such as three-sigma criteria, isolation forest algorithm, and data consistency checks such as correlation analysis between cross-sensor data.

[0070] Missing value handling aims to fill in data gaps caused by sensor failure, communication interruption, or manual recording omission. Processing strategies include interpolation methods such as linear interpolation, polynomial interpolation, spline interpolation, or machine learning-based interpolation such as K-Nearest Neighbor imputation; for time series data, missing values can be filled forward, backward, or predicted based on historical trends. For longer periods of continuous missing data, more complex models such as Long Short-Term Memory networks or Gated Recurrent Units may be needed for data reconstruction. All processing procedures need to be recorded to trace and evaluate their impact on data authenticity.

[0071] S1022: Data synchronization and alignment:

[0072] Given that different types of sensors have different sampling frequencies, such as displacement sensors that may sample once an hour, while acceleration sensors may sample hundreds of times per second, and there can be problems with asynchronous sampling times, this sub-step will unify all monitoring data to a pre-set sampling frequency and time reference.

[0073] Data synchronization ensures that the timestamps of all data are calibrated to a unified standard through the Precision Time Protocol or Global Positioning System time synchronization module.

[0074] Data alignment usually uses resampling techniques, such as averaging or median for high-frequency data downsampling, and interpolation or piecewise constant for low-frequency data upsampling. For example, all data are unified to a sampling frequency of once per minute, where high-frequency acceleration data is averaged down, and low-frequency displacement data is linearly interpolated up. This process ensures the comparability and feasibility of fusion of different modal data in the time dimension.

[0075] S1023: Data normalization and standardization:

[0076] Since different physical quantities have different dimensions and value ranges, for example, strain values are small, while water level values can be large, directly using them for model training can cause some features to contribute too much to the model output, while others are drowned.

[0077] Data normalization maps data to a specific range, for example, zero to one, achieved by max-min normalization:

[0078]

[0079] where X is the original data value X min and X max are the minimum and maximum values of the feature X norm is the normalized data.

[0080] Data standardization converts data to a distribution with mean zero and variance one, achieved by Z-score standardization:

[0081]

[0082] where μ is the mean of the feature σ is the standard deviation of the feature, and X std is the standardized data.

[0083] According to the distribution characteristics of the data and the requirements of the model, choose the appropriate processing method to eliminate the influence of dimension, speed up the convergence of the model, and improve the performance of the model.

[0084] S1024: Feature extraction and construction:

[0085] This sub-step aims to extract or construct deep features that can effectively represent the structural state and abnormal behavior from the original or pre-processed data. Traditional feature extraction methods based on statistics or signal processing have limitations and are difficult to capture complex nonlinear and high-dimensional correlations.

[0086] For time series data such as vibration, strain: extract time domain features such as mean, variance, kurtosis, skewness, root mean square, margin factor, waveform factor, and frequency domain features such as power spectral density, spectral entropy, main frequency and its amplitude. These features can describe the intensity, distribution and frequency characteristics of the structural dynamic response from different angles.

[0087] For multivariate data: construct correlation features between multiple variables, such as covariance between different measurement point displacements, and cross-correlation coefficient between temperature and strain, to reveal the coordination of the overall response of the structure.

[0088] For categorical data such as operating modes: perform one-hot encoding to convert to numerical vectors.

[0089] Feature engineering can also include advanced feature construction based on physical models or expert knowledge, such as damage-sensitive features, structural stiffness change indicators, fatigue accumulation indicators, etc. Finally, all extracted features are integrated into a unified feature vector, ready to be input into the deep fusion model.

[0090] As an embodiment of the present application, S103, multi-modal data deep fusion representation learning, specifically includes:

[0091] This step is the core part of the present application, aiming to deeply fuse the multi-source heterogeneous feature vectors processed by S102, to generate a multi-dimensional fusion representation that can comprehensively and robustly represent the current health state of the water conservancy project structure. The safety monitoring data of the water conservancy project contains structural responses such as displacement, stress, vibration, environmental loads such as temperature, water level, and operating conditions such as gate opening, unit load, etc. There are complex nonlinear correlations and complementary information between these modalities. Simple data concatenation or shallow fusion is difficult to capture the inherent coupling mechanism. Traditional multi-modal fusion methods often face the problems of semantic gap between modalities, feature redundancy, and poor fusion effect, which cannot effectively support subsequent accurate modeling and anomaly prediction. Therefore, the present application adopts a deep learning-based fusion strategy to realize collaborative learning of heterogeneous features and extraction of high-level semantic features by constructing a dedicated fusion network.

[0092] Specifically, the multi-modal data deep fusion representation learning process can be divided into the following sub-steps:

[0093] S1031: Intra-modal feature encoding:

[0094] For each type of modal data, first, an independent deep neural network encoder is used for feature learning to extract deep, abstract features within the modality. For example:

[0095] Structural response modality encoder: For time series data such as structural displacement, strain, and vibration, a recurrent neural network such as a time convolution network or a long short-term memory network is used to model the time dependence and capture the dynamic evolution features. The time convolution network realizes large receptive field through dilated convolution, effectively capturing long-range temporal dependencies while avoiding the computational bottleneck caused by the recurrent structure. For example, the input is a historical N time step structural vibration feature sequence, and the output is a low-dimensional vibration encoding vector that can represent the current vibration pattern.

[0096] Environmental load modal encoder: For environmental characteristics such as temperature, water level, rainfall, etc., convolutional neural networks or fully connected networks are used for feature extraction. Convolutional neural networks can extract the spatial correlation or local patterns of environmental features through multiple convolution kernels, and fully connected networks can capture the nonlinear mapping relationship of environmental factors. For example, the input is the current value of multiple environmental parameters, and the output is the environmental influence encoding vector.

[0097] Operating condition modal encoder: For operating parameters such as gate opening, unit load, etc., fully connected networks or embedding layers are used for encoding. Embedding layers can map discrete operating modes to continuous vector space, and fully connected networks can perform nonlinear transformation on continuous parameters to extract their influence characteristics on structural state. For example, the input is the current operating condition parameter, and the output is the operating state encoding vector.

[0098] The output of each encoder is a low-dimensional, semantically rich feature vector of the respective modal.

[0099] S1032: Cross-modal feature interaction and fusion:

[0100] The modal encoding vectors obtained in S1031 are input into the cross-modal fusion module, and through the design of a specific interaction mechanism, deep information sharing and complementation between modalities are realized.

[0101] One implementation uses an attention mechanism for fusion. For example, a multi-head self-attention network is constructed, allowing features from different modalities to "focus" on each other, dynamically allocating weights to enhance the contribution of important modalities and suppress redundant information.

[0102] The attention fusion calculation can be represented as:

[0103]

[0104] where Q, K, and V are the query, key, and value matrices generated by linear transformation of the modal encoding vectors, d k is the dimension of the key vector. This formula describes how to determine the weighted sum of value vectors by calculating the similarity between queries and keys, thereby realizing the interaction and aggregation between different modal information.

[0105] For example, the structural response modal encoding vector can be used as the query Q, and the environmental load modal encoding vector and the operating condition modal encoding vector can be used as the key K and the value V. In this way, the features of the structural response can be dynamically weighted according to the current environmental load and operating condition, thereby better reflecting the true state of the structure under specific external conditions.

[0106] Another implementation employs a gated fusion network that learns a gating mechanism to control the flow and fusion degree of different modal information. For example, a gated recurrent unit or a variant of long short-term memory network can be designed to dynamically adjust the input, forget, and output of each modal information through its internal gating units, such as update gate and reset gate, to achieve more fine-grained fusion control.

[0107] S1033: Fusion representation generation and optimization:

[0108] After the cross-modal interaction in S1032, the fused features are further transformed through one or more fully connected layers for non-linear transformation, and finally a high-dimensional, unified structural state multi-dimensional fusion representation vector is generated.

[0109] The fusion representation vector has the following characteristics:

[0110] Multi-dimensional semantics: It not only contains unique information of each modality, but also contains complex coupling relationships between modalities, such as the influence of temperature on structural stress and the contribution of water level change to dam deformation.

[0111] Robustness: In the face of partial sensor failure or data loss, due to the complementarity of multi-modal information, the model can be compensated by the data of other modalities to improve the robustness to abnormal situations.

[0112] Interpretability: During the training process, the contribution of different modalities to the final structural state representation can be revealed to some extent by analyzing the attention weights or gating coefficients, enhancing the interpretability of the model.

[0113] The fusion process is trained in an end-to-end manner, and the optimization goal is to make the generated fusion representation best support the subsequent structural state modeling and anomaly prediction tasks. The training data includes historical monitoring data and corresponding structural health state labels such as normal, slight abnormal, and severe abnormal.

[0114] As an embodiment of the present application, S104, structural state time series modeling and health degree evaluation, specifically includes:

[0115] This step aims to construct a time series model that can capture the dynamic evolution law of the water conservancy structure using the multi-dimensional fusion representation generated in S103, and based on this model, the current health status of the engineering structure is quantitatively evaluated. The state of the water conservancy structure has significant time-dependent and nonlinear characteristics, and its health degree is not static but continuously changes with time, environment and operating conditions. Traditional methods usually use empirical thresholds or simple statistical models for health assessment, which lack a deep understanding of the intrinsic complex dynamics and degradation process of the structure, resulting in rough and lagging evaluation results. The present invention uses a deep time series model to correlate the historical state and current state of the structure, achieving more refined and forward-looking health assessment.

[0116] Specifically, the structure state time series modeling and health degree evaluation process can be divided into the following sub-steps:

[0117] S1041: Time series model construction:

[0118] Select and construct a deep learning model suitable for processing time series data and having long-term dependence capturing capability, such as long short-term memory network, gated recurrent unit or time convolution network.

[0119] The long short-term memory network or gated recurrent unit model can effectively solve the gradient vanishing or gradient explosion problem encountered by traditional recurrent neural networks when processing long sequences through its internal gating mechanism, thus better learning the long-term evolution law and short-term fluctuation characteristics of the structure state. The input is the fusion representation sequence output by S103, and the output is the state prediction of the structure at a future time step or the hidden representation of the current state of the structure.

[0120] The time convolution network model, by stacking causal convolution layers and dilated convolution layers, effectively expands the receptive field to capture long-range dependencies while maintaining causal relationships, and has parallel computing advantages. It has strong ability to capture the periodic or trend changes of water conservancy structures affected by environmental loads.

[0121] The specific structure and parameters of the model, such as the number of layers, the number of units, the dilution rate, etc., are trained and adjusted on the historical data set through cross-validation and hyperparameter optimization techniques to achieve optimal prediction performance.

[0122] S1042: Structure state evolution prediction:

[0123] Using the trained time series model, input the fusion representation sequence of the current time and the past period of time to predict the state of the water conservancy structure at one or more future time steps.

[0124] The prediction can be the predicted value of a key response parameter of the structure, such as displacement, strain, vibration amplitude, or the predicted value of an abstract structure health indicator. For example, the model can predict the maximum displacement of the dam crest in the next twenty-four hours, or predict the strain change trend of a key measurement point in the next week.

[0125] The prediction model infers the future behavior of the structure by learning patterns in historical data, such as periodic changes, trend changes, and recovery patterns after sudden events. This prediction capability is of great significance for identifying potential risks in advance and planning maintenance and repair.

[0126] S1043: Health assessment index system construction and evaluation:

[0127] Based on the structure state evolution prediction result of S1042, combined with the preset structure safety specification and engineering experience, a comprehensive health assessment index system is constructed, and the health of the water conservancy project structure is quantitatively evaluated.

[0128] The health assessment index can include:

[0129] Deviation index: the deviation between the predicted value of the structure response and the normal operation baseline value, such as the historical average value, the design value, such as displacement deviation, strain deviation.

[0130] Change rate index: the rate of change of the structure response with time, such as displacement change rate, strain change rate, and rapid change rate may indicate structural deterioration.

[0131] Cumulative damage index: cumulative effects such as fatigue damage and plastic deformation calculated based on historical data, which reflect the long-term health loss of the structure.

[0132] Stability index: by analyzing the frequency spectrum characteristics of the vibration signal, the modal parameters such as natural frequency and damping ratio of the structure are evaluated, and the change of which may indicate the change of the stiffness or connection performance of the structure.

[0133] These indicators are integrated by weighted summation, fuzzy comprehensive evaluation or expert system, etc. to obtain a comprehensive health score. For example, a percentage score, where a high score indicates good health, and a low score indicates a health risk. The score result provides a quantitative basis for subsequent abnormal behavior recognition and safety risk judgment.

[0134] The evaluation result can be generated periodically, such as daily, weekly, and compared with the historical health curve to find the long-term evolution trend of the health.

[0135] As an embodiment of the present application, S105, structure abnormal behavior pattern recognition and prediction, specifically includes:

[0136] This step aims to further explore the data on the basis of the structural health assessment results of S104, identify abnormal behavior patterns that deviate significantly from the normal operation mode, and predict potential abnormal events. In the operation process of water conservancy projects, due to factors such as material aging, environmental erosion, load mutation or construction defects, various forms of abnormalities may occur, such as local cracking, leakage aggravation, deformation acceleration, vibration mutation, etc. Traditional anomaly detection methods often rely on fixed thresholds or simple statistical models, which are difficult to effectively identify hidden abnormalities and multi-modal coupling abnormalities in complex nonlinear modes, and lack prediction ability. The present application uses the powerful pattern recognition ability of deep learning models, combined with multi-modal fusion representation, to realize accurate identification and early prediction of abnormal behavior.

[0137] Specifically, the structural abnormal behavior pattern recognition and prediction process can be divided into the following sub-steps:

[0138] S1051: Abnormal pattern learning and discrimination:

[0139] Using the multi-dimensional fusion representation generated by S103 and the time series model output of S104, an anomaly detection model is trained to learn the data pattern under normal operation state.

[0140] One implementation adopts anomaly detection based on reconstruction error. For example, use autoencoder or variational autoencoder to train normal data. The model learns how to efficiently encode and decode normal data, i.e. capture the internal regularity of normal data in the hidden space. When inputting abnormal data, the reconstruction error of the model will increase significantly, because the abnormal data does not conform to the normal pattern it has learned.

[0141] The reconstruction error calculation can be represented as:

[0142]

[0143] where x i is the original input data, is the data reconstructed by the autoencoder, and n is the number of samples. When L recon exceeds the preset threshold, it is determined to be abnormal.

[0144] Another implementation adopts classification-based anomaly detection. For example, the normal, slight abnormal, severe abnormal, etc. in the structural health assessment are used as class labels, and a classifier such as support vector machine, random forest or deep neural network is trained to classify the current structure state. The classifier can directly output which health level the current state belongs to, and identify the abnormal pattern.

[0145] The training data usually includes a large amount of normal operation data and a small amount of expert-labeled abnormal data. For the case of scarce abnormal data, semi-supervised learning or unsupervised learning methods can be used.

[0146] S1052: Abnormal trend analysis and early warning:

[0147] On the basis of identifying abnormal patterns, further analysis of the evolution trend of abnormal behavior is carried out, and early warning is carried out.

[0148] Abnormal trend analysis includes analysis of the curve of abnormal indicators such as reconstruction error, health score drop rate, abnormal frequency, etc. with time. For example, the abnormal indicators are smoothed by methods such as moving average, exponential smoothing, etc., and the future trend of the abnormal indicators is predicted by regression analysis such as linear regression, polynomial regression or Kalman filter.

[0149] The early warning mechanism is realized by setting multiple levels of warning thresholds. For example, when the reconstruction error or the health score first reaches the "slight abnormality" threshold, the system triggers a first-level warning and notifies the relevant personnel to pay attention; when the abnormality continues to worsen and reaches the "serious abnormality" threshold, a second-level warning is triggered, requiring immediate action. The setting of the warning threshold takes into account engineering design standards, historical abnormal data statistics, and expert experience.

[0150] The prediction of the occurrence time or severity of the abnormality can be achieved by using a prediction model such as a long short-term memory network to predict the future trend of the abnormal indicators, and when the predicted value will reach the warning threshold at a certain time in the future, an early warning is given. This prediction capability gives the manager sufficient time to intervene and avoid small abnormalities from evolving into major accidents.

[0151] As an embodiment of the present application, S106, safety risk level determination and early warning information generation, specifically includes:

[0152] This step is the final output of the entire safety monitoring method, which aims to quantitatively determine the safety risk of the water conservancy project structure based on the results of abnormal behavior identification and prediction in S105, combined with engineering actual situation, risk management strategy and relevant specification standards, and generate an early warning report containing detailed information. Traditional safety warning systems may only provide a simple "abnormal / normal" binary decision, lacking detailed differentiation of risk level and richness of early warning information, making it difficult for managers to make quick and accurate decisions. The present application improves the practicality of early warning and decision support capability through multi-dimensional risk assessment and structured early warning information generation.

[0153] Specifically, the safety risk level determination and early warning information generation process can be divided into the following sub-steps:

[0154] S1061: Safety risk assessment model construction:

[0155] Based on the abnormal patterns and prediction trends identified in S105, a comprehensive safety risk assessment model is constructed. This model not only considers the severity of the anomaly, but also considers its impact on the overall stability and operational function of the project.

[0156] The risk assessment model can use Bayesian networks, fuzzy comprehensive evaluation, analytic hierarchy process, or expert systems, etc. For example, the abnormal indicators output by S105, such as reconstruction error, health score, abnormal duration, and abnormal change rate, are taken as input, combined with the safety factor of engineering design, material degradation model, and the probability of possible failure modes such as leakage, sliding, and fracture, to calculate a comprehensive safety risk index.

[0157] This index is usually mapped to discrete risk levels, such as "normal", "attention", "general warning", "serious warning", "emergency rescue", etc. Each level corresponds to a different risk level and the level of action required.

[0158] The model needs to fully absorb expert knowledge in the field of hydraulic engineering during the establishment process, and combine historical accident cases for calibration and verification to ensure the reliability and authority of the evaluation results.

[0159] S1062: Early warning information generation and push:

[0160] According to the safety risk level determined by S1061, the system automatically generates a structured early warning report containing detailed information, and timely pushes it to the relevant responsible persons through various channels.

[0161] The early warning report should at least include the following content:

[0162] Risk level: Clearly indicate the safety risk level determined at present, such as "serious warning".

[0163] Abnormal type and location: Describe the identified abnormal pattern in detail, such as "abnormal increase in seepage flow of dam body left bank bedrock", and point out the specific measuring point or area where the anomaly occurs.

[0164] Data support: Provide key monitoring data charts and characteristic values that support the determination of anomalies, such as time series curve of seepage flow at abnormal measuring point, historical comparison data, and deviation value from normal baseline.

[0165] Impact assessment: Preliminary assessment of the impact of the anomaly on the safety of the project structure, operational function, or economic benefits.

[0166] Suggestion: Provide preliminary response suggestions based on risk level and anomaly type, such as "recommend immediate activation of emergency plan, organize expert on-site investigation, and encrypt seepage flow monitoring frequency."

[0167] Early warning information push channels include but are not limited to: SMS, email, mobile application push notifications, desktop client pop-up windows, and interface docking with water conservancy department emergency command system. The system should have flexible early warning push configuration functions, allowing users to receive different levels of early warning information according to their roles and permissions. For example, a first-level warning can be pushed to on-duty personnel, and a second-level warning can be pushed to on-duty personnel, project management responsible person and emergency command center. To ensure the timeliness and reliability of information, the system can set up multiple notification mechanisms, such as automatically switching to phone notification when SMS sending fails.

[0168] The water conservancy project safety monitoring system based on data processing aims to provide hardware and software support for the above methods. The system is a comprehensive platform that integrates data acquisition, data processing, model analysis, health assessment and early warning.

[0169] The system includes: data acquisition module, data preprocessing module, multi-modal fusion module, time series modeling module, anomaly prediction module, risk assessment module, early warning management module, data storage module and user interaction module.

[0170] The data acquisition module is responsible for performing various data acquisition tasks in S101. This module is composed of various sensors such as displacement sensors, strain sensors, vibration sensors, temperature and humidity sensors, water level meters, osmotic pressure meters, etc., as well as data acquisition units or data recorders. The data acquisition unit is responsible for analog-to-digital conversion, preliminary filtering and timestamp labeling of sensor signals, and transmits data to the central processing unit through wired communication networks such as optical fiber, Ethernet or wireless communication networks such as 5th generation mobile communication technology, long-distance wireless communication technology. The module also includes data redundancy and fault self-diagnosis functions to ensure data continuity and reliability.

[0171] The data preprocessing module is responsible for performing various data preprocessing and feature engineering tasks in S102. This module is a software function unit deployed on the central processing unit or distributed computing cluster. It integrates data cleaning algorithms such as noise filtering, outlier detection, missing value imputation, data synchronization and alignment algorithms such as resampling, timestamp calibration, and data normalization standardization algorithms. In addition, the module also contains a rich feature engineering toolbox that can automatically extract time domain, frequency domain, statistical features, and generate high-level physical meaning features according to the configuration. The output of this module is a unified formatted, high-quality multi-modal feature vector.

[0172] The multi-modal fusion module is responsible for performing the multi-modal data deep fusion representation learning task in S103. The module is based on a deep learning framework such as Tensorflow or Pytorch, and includes multiple intra-modal feature encoders such as temporal convolutional networks and fully connected networks, as well as a cross-modal feature interaction fusion network such as attention mechanisms and gated fusion networks. The module receives feature vectors from the data preprocessing module, learns the complex relationships between different modalities, and outputs a high-dimensional fusion representation vector that can comprehensively represent the structural health state. The training and inference processes of the module can be completed on a high-performance computing server.

[0173] The time series modeling module is responsible for performing the structural state time series modeling and health degree evaluation task in S104. The module includes one or more deep time series prediction models such as long short-term memory networks, gated recurrent units, and temporal convolutional networks, which are used to learn the time series evolution of the fusion representation. The module can predict the future changes of key parameters of the water conservancy structure, and through a pre-defined health degree evaluation index system, it can quantify and score the health degree of the engineering structure based on the current and predicted state.

[0174] The anomaly prediction module is responsible for performing the structural anomaly behavior pattern recognition and prediction task in S105. Based on the output of the time series modeling module, the module deploys autoencoders, classifiers, or anomaly detection algorithms such as Isolation Forest and Local Outlier Factor. The module can monitor whether the structure state deviates from the normal mode in real time, identify specific abnormal patterns, and predict the occurrence time or severity of the anomaly through trend analysis, providing technical support for early warning.

[0175] The risk assessment module is responsible for performing the safety risk level determination task in S106. The module integrates risk assessment models such as Bayesian networks and fuzzy comprehensive evaluation models. It receives the identification results from the anomaly prediction module, combines with the engineering characteristics database, historical accident database and expert knowledge base, and comprehensively calculates the safety risk index of the current water conservancy project, and maps it to the pre-defined risk level.

[0176] The early warning management module is responsible for generating and pushing the early warning information in S106. The module automatically generates structured early warning reports based on the risk level output by the risk assessment module, and manages the sending of early warning information. It supports multiple notification channels such as SMS, email, and application push, and can be flexibly configured according to the warning level and recipient role. The module also has the functions of early warning log recording, early warning cancellation and feedback management.

[0177] A data storage module is used to store all data generated during system operation, including original monitoring data, pre-processed feature data, fusion representation, model parameters, historical evaluation results, early warning logs, etc. This module can use a distributed file system such as a distributed file system, object storage, or a combination of a relational database and a non-relational database to support the storage, efficient query, and analysis of massive data. The data storage has high availability, scalability, and data backup and recovery mechanisms.

[0178] A user interaction module provides a user-friendly interface for engineering managers and technical experts to configure the system, visualize data, query health scores, view abnormal reports, and respond to early warnings. This module provides services through a graphical user interface, web browser, or mobile application, supporting multi-user access and permission management. Users can view real-time curves, historical trends, health score changes, and abnormal event details for various types of monitoring data through this module, and confirm and handle early warning information.

Claims

1. A method for safety monitoring of water conservancy projects based on data processing, characterized in that, include: Acquire multi-source heterogeneous monitoring data of the water conservancy project structure, surrounding environment and operating conditions. The multi-source heterogeneous monitoring data includes structural displacement data, structural stress and strain data, seepage data, temperature data, humidity data, water level data, reservoir capacity data, rainfall data and seismic load data. The multi-source heterogeneous monitoring data is subjected to multi-dimensional time series preprocessing, including timestamp synchronization, missing value imputation, outlier detection and correction, and data normalization, to generate a unified, clean, and time-aligned structured time series dataset. Extract basic statistical features, frequency domain features, spatial correlation features, and temporal evolution features from the structured time-series dataset; multi-scale spatiotemporal fusion features. The multi-scale spatiotemporal fusion features are input into a multimodal deep learning model with a long short-term memory neural network architecture that integrates a multi-head attention mechanism for deep fusion representation learning, so as to obtain a multi-dimensional fusion representation vector of the structural state of the hydraulic engineering. Based on the multidimensional fusion representation vector, a dynamic baseline model of the water conservancy project under normal operation is established. By calculating the Mahalanobis distance or reconstruction error between the current real-time fusion features and the predicted output of the dynamic baseline model, structural anomalies that deviate significantly from the normal operation mode are identified. The identified structural anomalies are quantitatively assessed for risk and classified into multiple levels of early warning. Combined with a predefined maintenance strategy knowledge base, intelligent suggestions for maintenance decision-making are generated and output.

2. The water conservancy project safety monitoring method based on data processing according to claim 1, characterized in that, The acquisition of multi-source heterogeneous monitoring data of water conservancy projects includes: High-precision sensors deployed on the main body of the hydraulic engineering structure are used to collect structural displacement data, structural stress and strain data, structural vibration data and structural tilt data in real time. By deploying sensors in the surrounding environment, real-time data such as temperature, humidity, wind speed and direction, rainfall, water level, reservoir capacity, seepage flow, and uplift pressure are collected. The monitoring system for acquiring water conservancy project operation parameters collects real-time data on unit operating status, gate opening, power generation, and discharge flow. The collected monitoring data are transmitted to the data aggregation layer through a unified data acquisition interface and communication protocol for initial storage and timestamp synchronization.

3. The water conservancy project safety monitoring method based on data processing according to claim 2, characterized in that, Multi-dimensional time-series preprocessing is performed on the multi-source heterogeneous monitoring data, including: Data cleaning, including median filtering, Kalman filtering, or wavelet denoising, is performed on the multi-source heterogeneous monitoring data to correct or remove obvious errors and noise in the original data; The missing values ​​in the monitoring data are imputed using a long short-term memory neural network prediction model based on historical data patterns or a cubic spline interpolation method. An algorithm based on isolated forest or local anomaly factors is used to detect and correct outliers in the monitoring data; The monitoring data are normalized using the minimum-maximum normalization method or the Z-score normalization method.

4. The water conservancy project safety monitoring method based on data processing according to claim 3, characterized in that, Extracting multi-scale spatiotemporal fusion features from the structured time-series dataset includes: The mean, variance, skewness, kurtosis, and extreme values ​​of each sensor data are calculated within a preset time window to extract basic statistical characteristics. Perform fast Fourier transform or wavelet transform on vibration sensor data to obtain the frequency domain characteristics of its dominant frequency, amplitude, and energy distribution; Calculate the Pearson correlation coefficient or mutual information entropy of displacement, stress and seepage data between adjacent sensors or key structural points to extract spatial correlation features; The trend, seasonal, and residual terms of the data are calculated by using a sliding time window, and a temporal convolutional network model is used for multi-scale feature learning to construct temporal evolution features that reflect the long-term deformation trend and periodic changes of the structure.

5. The water conservancy project safety monitoring method based on data processing according to claim 4, characterized in that, The multi-scale spatiotemporal fusion features are input into a multimodal deep learning model with a long short-term memory neural network architecture integrating a multi-head attention mechanism for deep fusion representation learning, so as to obtain a multi-dimensional fusion representation vector of the hydraulic engineering structural state, including: The structural displacement data, structural stress-strain data, structural vibration data, and time-series evolution characteristic structural response modal data are encoded using an independent deep neural network encoder to extract the structural response encoding vector. The environmental load modal data, such as temperature data, water level data, and rainfall data, are encoded using an independent deep neural network encoder to extract the environmental impact coding vector. The gate opening data, unit load data, and operating condition modal data are encoded using an independent deep neural network encoder to extract the operating status encoding vector. The structural response encoding vector, environmental influence encoding vector, and operating state encoding vector are input into a multi-head self-attention network. Cross-modal feature interaction and fusion are achieved through dynamic weight allocation to generate the multi-dimensional fusion representation vector.

6. The water conservancy project safety monitoring method based on data processing according to claim 5, characterized in that, A dynamic baseline model of a water conservancy project under normal operating conditions is established based on the multidimensional fusion representation vector. By calculating the Mahalanobis distance or reconstruction error between the current real-time fusion features and the predicted output of the dynamic baseline model, structural anomalies that significantly deviate from the normal operating mode are identified, including: Long Short-Term Memory (LSTM) networks, gated recurrent units (GRUs), or temporal convolutional networks (TCRs) are selected as temporal models to train the multidimensional fused representation vector sequence in order to learn the dynamic evolution law of hydraulic engineering structures. Using the trained time series model, and inputting the multidimensional fusion representation vector sequence of the current time and a past period, the state of the water conservancy project structure in one or more future time steps is predicted, so as to establish a dynamic baseline model of the water conservancy project under normal operation. Based on the prediction results, structural anomalies are identified by calculating the Mahalanobis distance or reconstruction error between the current real-time multidimensional fusion representation vector and the predicted output of the dynamic baseline model.

7. The water conservancy project safety monitoring method based on data processing according to claim 6, characterized in that, The identified structural anomalies that deviate significantly from normal operating modes include: The trained autoencoder or variational autoencoder is used to learn and identify abnormal patterns in the multidimensional fusion representation vector under normal operating conditions. When abnormal data is input, the reconstruction error increases significantly, and it is judged as abnormal. Anomaly trend analysis is performed on the decline in the reconstruction error or health score, and the future trend of the abnormal indicators is predicted by regression analysis or Kalman filtering to achieve early warning.

8. The water conservancy project safety monitoring method based on data processing according to claim 7, characterized in that, The process involves quantifying the risk assessment and classifying the identified structural anomalies into multi-level early warning systems. Combined with a predefined maintenance strategy knowledge base, intelligent suggestions for maintenance decision-making are generated and output, including: Based on the identified abnormal patterns and predicted trends, a comprehensive security risk assessment model is constructed using Bayesian networks, fuzzy comprehensive evaluation, or hierarchical analysis to quantify the risk index of the structural abnormal state. Based on the aforementioned risk index, safety risks are divided into four levels: Level 1 warning, Level 2 warning, Level 3 warning, and Level 4 warning. By combining a predefined maintenance strategy knowledge base, specific maintenance measures are generated based on the risk assessment results and early warning levels. The evolution trend of abnormal states and the timing of intervention are also predicted to output intelligent suggestions for maintenance decision-making.

9. A monitoring system for the data processing-based safety monitoring method for water conservancy projects according to any one of claims 1-8, characterized in that, include: The data acquisition unit is used to acquire multi-source heterogeneous monitoring data of the water conservancy project structure, surrounding environment and operating conditions. The monitoring data includes structural displacement data, structural stress and strain data, seepage data, temperature data, humidity data, water level data, reservoir capacity data, rainfall data and seismic load data. The data preprocessing module is used to perform multi-dimensional time-series preprocessing on the multi-source heterogeneous monitoring data, including timestamp synchronization, missing value imputation, outlier detection and correction, and data normalization, in order to generate a unified, clean, and time-aligned structured time-series dataset. The feature engineering module is used to extract basic statistical features, frequency domain features, spatial correlation features, and multi-scale spatiotemporal fusion features of temporal evolution features from the structured time series dataset. The multimodal fusion module is used to input the multi-scale spatiotemporal fusion features into a multimodal deep learning model with a long short-term memory neural network architecture that integrates a multi-head attention mechanism for deep fusion representation learning, so as to obtain a multi-dimensional fusion representation vector of the structural state of the hydraulic engineering. The structural state modeling and anomaly identification module is used to establish a dynamic baseline model of the water conservancy project under normal operation based on the multi-dimensional fusion representation vector, and to identify structural anomalies that deviate significantly from the normal operation mode by calculating the Mahalanobis distance or reconstruction error between the current real-time fusion features and the predicted output of the dynamic baseline model. The risk assessment and decision support module is used to quantitatively assess the risks of the identified structural anomalies and classify them into multi-level warnings. It also combines a predefined maintenance strategy knowledge base to generate and output intelligent suggestions for maintenance decisions.

10. The water conservancy project safety monitoring system based on data processing according to claim 9, characterized in that, The data preprocessing module is used for: Data cleaning algorithms such as median filtering, Kalman filtering, or wavelet denoising are applied to correct or remove obvious errors and noise from the original monitoring data. The missing values ​​in the monitoring data are imputed by executing a long short-term memory neural network prediction model or a cubic spline interpolation algorithm based on historical data patterns. The algorithm based on isolated forest or local anomaly factors is used to detect and correct outliers in the monitoring data; The monitoring data is normalized by performing a minimum-maximum normalization method or a Z-score normalization method.

Images