Hydropower plant deformation monitoring method and system

Abstract

The invention discloses a hydropower station workshop deformation monitoring method, which comprises the following steps of collecting multi-dimensional monitoring data and point cloud data of a hydropower station workshop, performing adaptive filtering and noise reduction on the multi-dimensional monitoring data, performing spatial registration on the multi-dimensional monitoring data and the point cloud data in a unified three-dimensional coordinate system, constructing a multi-channel time sequence deep learning model, and performing multi-channel time sequence deep learning on the multi-channel time sequence deep learning model; a dynamic early warning threshold model is constructed based on material characteristics, equipment operation parameters and historical monitoring data, risk assessment is carried out, a three-level response mechanism is triggered according to a risk assessment result, a sensor reliability evaluation index is established based on deviation analysis of monitoring data and an early warning result, and a data fusion weight is dynamically optimized. Adjusting the decision boundary of the classification model, and carrying out visualization and traceability analysis on the multi-dimensional monitoring data; the invention further discloses a hydropower station plant deformation monitoring system. According to the invention, a multi-dimensional and three-dimensional risk assessment system is established through data monitoring and real-time analysis, and the structural safety of the hydropower house is guaranteed to the maximum extent.

Description

Technical Field

[0001] The invention belongs to the technical field of water conservancy and hydropower engineering structure safety monitoring methods, relates to a hydropower station plant deformation monitoring method, and further relates to a hydropower station plant deformation monitoring system. Background Art

[0002] As the core hub of a hydropower generation system, the structural stability of a hydropower station building directly impacts its operational safety. Since powerhouses are often built on high slopes or near underground caverns, stress redistribution caused by excavation can easily lead to uneven foundation settlement. Furthermore, vibrations from turbine operation are transmitted to the structure, causing micron-level dynamic deformation accumulation and resulting in adjustments to its structural stability. Existing deformation monitoring technologies primarily rely on traditional monitoring methods, but face the following technical bottlenecks under complex operating conditions, limiting their effectiveness. Specifically, existing technologies often use a combination of point sensors (such as displacement gauges and inclinometers) and manual inspections. These methods can only acquire data at discrete points and struggle to capture overall deformation trends in the powerhouse. For example, the tensioning wire method is only suitable for linear displacement monitoring and cannot identify three-dimensional torsional deformation. Multi-point displacement gauges are limited in depth by their measurement points and can easily miss areas of deep deformation. Furthermore, horizontal displacement monitoring (such as GNSS observations) and vertical displacement monitoring (such as geometric leveling) operate independently, resulting in a lack of temporal and spatial correlation in the data, making comprehensive analysis difficult. Summary of the Invention

[0003] The purpose of the present invention is to provide a method for monitoring deformation of a hydropower plant building, which solves the problem that the monitoring methods in the prior art can only obtain discrete point data and are difficult to capture the overall deformation trend of the plant building.

[0004] A second object of the present invention is to provide a hydropower station building deformation monitoring system.

[0005] The first technical solution adopted by the present invention is a method for monitoring deformation of a hydropower plant building, comprising the following steps: Step 1: Collect multi-dimensional monitoring data and point cloud data of the hydropower station building; Step 2: After adaptive filtering and noise reduction, the multi-dimensional monitoring data is spatially registered with the point cloud data in a unified three-dimensional coordinate system; Step 3: Build a multi-channel time series deep learning model and construct a dynamic warning threshold model based on material properties, equipment operating parameters, and historical monitoring data to conduct risk assessment; Step 4: Trigger the three-level response mechanism based on the risk assessment results. Based on the deviation analysis between the monitoring data and the early warning results, establish the sensor reliability evaluation index and dynamically optimize the data fusion weight, adjust the classification model decision boundary, and visualize and trace the multi-dimensional monitoring data. The first technical solution of the present invention is also characterized in that: In step 1, strain sensors and inertial measurement units are deployed in key stress concentration areas of the hydropower plant building to collect multi-dimensional monitoring data. A drone equipped with a multi-lens camera is used to periodically perform 3D scans of the hydropower plant building's exterior facade to generate point cloud data. The method for determining key stress concentration areas is as follows: using a finite element simulation model to calculate the stress distribution cloud map of the powerhouse structure under different operating conditions, the nodes where the maximum principal stress exceeds a set ratio of the material yield strength are extracted as candidate monitoring areas. Vibration modal analysis of the candidate areas is then performed in combination with the vibration test data of the turbine unit. The nodes where vibration energy is concentrated and there is a mechanical transmission path with the equipment base are selected as the final layout points. The vibration modal analysis process is as follows: collect the vibration signal of the turbine unit during normal operation and extract the characteristic frequency components through Fourier transform:

[0006] in, is the frequency component of the signal in the frequency domain; is the time domain signal; f is the frequency, unit is Hz; e is a natural constant; t is time, unit is s; j is an imaginary unit; Calculate the resonant response amplitude of each node at the characteristic frequency:

[0007] in, For nodes i At the characteristic frequency f The resonance response amplitude under ; For nodes i External vibration force, unit: N; For nodes i The mass of, in kg; is the damping coefficient, unit is N·s / m; For nodes i The natural frequency, in Hz; Set a safety threshold and select nodes with amplitudes exceeding the safety threshold to increase the density of sensor deployment; The method for extracting characteristic frequency components in vibration modal analysis is as follows: perform wavelet packet decomposition on the vibration acceleration signal, extract the sub-band signal energy related to the turbine unit rotation frequency and blade passing frequency, and calculate the vibration transfer function of each node at different frequencies through the cross-power spectral density of the input unit base vibration signal and the output node vibration signal:

[0008] in, is the transfer function; is the cross power spectral density; is the power spectrum density of the input unit base vibration signal; Analyze the spectral relationship between the input signal and the output signal, set the gain threshold, and screen out the frequency components whose transfer function gain is higher than the gain threshold as dangerous resonance frequencies for continuous monitoring and tracking. Combined with the plant structure deformation monitoring data, determine whether there is a resonance risk.

[0009] The specific method of spatial registration in step 2 is as follows: pre-place a reference target on the surface of the hydropower station building. The target contains a geometric feature pattern with known three-dimensional coordinates. During the three-dimensional scanning process, the target pattern is preferentially identified, and the conversion relationship between the image coordinate system and the global coordinate system is established through the perspective transformation algorithm. For areas without target coverage, the feature point matching algorithm is used to extract the visual features of the strain sensor installation position, and the strain sensor spatial coordinates are embedded in the point cloud data; a spatiotemporal alignment model is established at the edge computing node to perform timestamp alignment and spatial position compensation on the multi-dimensional monitoring data and point cloud data. The timestamp alignment uses the GPS timing module to synchronize the clocks of each device, and motion speed compensation is performed on the drone image acquisition delay. The registration process is based on the plant BIM model to form a unified three-dimensional coordinate system. The spatial mapping relationship is established through the feature point matching algorithm, and the coordinate conversion error is corrected using an iterative optimization algorithm; The time delay compensation method in the spatiotemporal alignment model is as follows: the image acquisition position offset is calculated based on the UAV flight speed and the camera exposure interval. The position error is eliminated through a motion blur correction algorithm in the point cloud reconstruction stage. The clock deviation between the sensor data acquisition time and the image exposure time is compensated by polynomial fitting. The compensation amount is dynamically adjusted based on historical synchronization error data. An abnormal time difference detection mechanism is established. When the clock deviation exceeds the set range, forced clock synchronization based on the NTP protocol is initiated.

[0010] The method for constructing the dynamic warning threshold model in step 3 is as follows: establish a temperature-stress coupling relationship function for the concrete structure, correct the material elastic modulus parameters through real-time temperature data, and then calculate the dynamic adjustment range of the allowable deformation; construct a load-deformation transfer function based on the correlation between the turbine unit output parameters and the structural vibration intensity to determine the dynamic baseline value; use a sliding time window statistical method to analyze the distribution characteristics of historical monitoring data, monitor in real time whether the current monitoring data conforms to the distribution pattern of historical data, and trigger the threshold update mechanism when the real-time monitoring value deviates from the historical data distribution center by more than a preset probability; The load-deformation transfer function is established as follows: under different load conditions, the vibration acceleration of the concrete structure is measured using an accelerometer, while the output parameters of the turbine unit are recorded. The measured output parameters and the corresponding vibration acceleration data are then subjected to regression analysis to establish a functional relationship between output and vibration acceleration:

[0011] in, is the structural vibration acceleration, unit is m / s 2 ; To generate power for the turbine unit, is the regression coefficient, is the offset constant; Establish the load-deformation transfer function:

[0012] in, is the deformation of the concrete structure, unit is m; k is the stiffness of the concrete structure, in N / m; At the same time, the transfer coefficient of vibration energy to structural deformation is determined by regression analysis:

[0013]

[0014] in, is the transfer coefficient of vibration energy to structural deformation; is the vibration energy of the turbine unit or other equipment, unit is j; m is the mass of the plant structure, in kg; a is the vibration acceleration, unit is m / s 2 ; The multi-channel time series deep learning model includes an input layer, a hidden layer, and an output layer. The input layer receives the fused strain, displacement, acceleration, and environmental parameter data. The hidden layer uses a bidirectional recurrent neural network with an attention mechanism to extract spatiotemporal correlation features. The output layer synchronously outputs the structural deformation pattern classification results and risk probability prediction values. The multi-channel time series deep learning model extracts spatiotemporal correlation features in the data through a bidirectional recurrent neural network with an attention mechanism, outputs structural deformation patterns and risk prediction values, and updates the monitoring benchmark values ​​in real time through a dynamic early warning threshold model combined with material properties, equipment operating parameters, and historical data.

[0015] The triggering logic of the three-level response mechanism in step 4 is as follows: the first-level response is activated when the deformation of any single monitoring point exceeds 120% of the dynamic threshold, triggering the monitoring frequency of the area to increase to 3 times the original rate, and retrieving the historical data of the last 24 hours for trend comparison to identify whether there is accelerated deformation or periodic fluctuations; the second-level response is activated when three or more adjacent monitoring points exceed the limit at the same time and the deformation direction shows mechanical correlation, generating a recommendation for equipment load reduction operation and starting the backup sensor for cross-validation to confirm the accuracy of the original monitoring data and eliminate single-point errors; the third-level response is activated when the probability of overall structural instability exceeds 90% and the consistency between the drone re-measurement data and the edge computing results is less than 80%, sending an encrypted shutdown command to the central control system and marking the coordinates of the dangerous area. At the same time, the central control system automatically marks the coordinates of the dangerous area corresponding to the current monitoring results; Increase monitoring frequency when local deformation exceeds the limit; generate equipment load adjustment instructions when multiple related areas exceed the limit; and execute automatic protection shutdown when the overall instability risk exceeds the critical value; The specific method for visualizing and tracing the multi-dimensional monitoring data is to superimpose the three-dimensional point cloud data and sensor monitoring values ​​on the surface of the factory BIM model, characterize the deformation distribution through color gradient mapping, and construct a multi-dimensional correlation map to show the temporal and spatial correlation between the structural deformation trend and the equipment vibration parameters and ambient temperature.

[0016] In step 1, while collecting multi-dimensional monitoring data, temperature and humidity sensors and vibration accelerometers are deployed to collect environmental interference parameters; in step 2, the noise reduction process dynamically adjusts the filter parameters according to the spectral characteristics of the environmental noise; in step 3, the dynamic warning threshold model updates the benchmark value in real time through the sliding window statistical method; in step 4, the sensor reliability evaluation index is obtained by comprehensively calculating the data continuity rate, noise variance and consistency error with adjacent sensors. The gradient descent algorithm is used to minimize the weighted sum of the mean square error of the prediction model and the warning false alarm rate. The weighted average algorithm is used to optimize the multi-source data fusion weight. The optimization goal is to minimize the weighted sum of the mean square error of the prediction model and the warning false alarm rate. The particle swarm optimization algorithm is used regularly to adjust the decision boundary of the classification model. The warning records of the last 30 days are used as the training set in the adjustment process, and the reduction of the false alarm rate and the stability of the missed alarm rate are used as convergence conditions.

[0017] In step 1, calculate the offset of the image acquisition position when the drone performs periodic 3D scanning:

[0018] in, v is the flight speed of the UAV, in m / s; is the image exposure time interval, unit s; is the offset, unit is m.

[0019] The temperature-stress coupling relationship function of concrete structure is:

[0020] in, Temperature T The stress under the pressure, unit is Pa; Temperature T The elastic modulus of the material under , in Pa; Temperature T strain under Through the stress-strain experimental data at different temperatures, the elastic modulus of the material at each temperature point is obtained and the corresponding strain , through regression analysis or fitting method, the relationship between temperature, stress and strain is obtained, and the elastic modulus of the material is dynamically corrected based on real-time temperature data , and then calculate the corresponding stress value to determine the deformation under temperature change; The dynamic adjustment range of allowable deformation is:

[0021] in, For the temperature T Maximum deformation under is the corresponding maximum stress, unit: Pa; It is the elastic modulus after temperature correction, in Pa.

[0022] The second technical solution adopted by the present invention is a hydropower station plant deformation monitoring system, which includes a multi-source data acquisition module, an edge data processing module, an analysis and decision-making module, and a system self-optimization module connected in sequence through industrial Ethernet. The analysis and decision-making module is connected to a hierarchical control execution module, and the system self-optimization module is connected to the edge data processing module and the analysis and decision-making module through a control bus.

[0023] The second technical solution of the present invention is also characterized in that: The multi-source data acquisition module includes a distributed sensor network and a drone equipped with a multi-lens camera. The distributed sensor network includes strain sensors, inertial measurement units, temperature and humidity sensors, and vibration accelerometers deployed in key stress concentration areas of the factory. The strain sensors, inertial measurement units, and drones are connected to the edge data processing module through wireless communication modules. The edge data processing module includes an adaptive filtering unit and a spatiotemporal registration unit. The adaptive filtering unit receives time series data transmitted by the sensor network and dynamically adjusts the filtering parameters according to the spectral characteristics of the ambient noise to eliminate interference signals. The spatiotemporal registration unit has a built-in factory building BIM model database and aligns the photogrammetric point cloud data with the sensor data in a unified three-dimensional coordinate system through a feature point matching algorithm. It also uses an iterative optimization algorithm to correct coordinate conversion errors. The output of the spatiotemporal registration unit is connected to the data analysis module. The analysis and decision-making module includes a multi-channel time-series deep learning model and a dynamic threshold generation unit. The input of the deep learning model receives fused strain, displacement, acceleration, and environmental parameter data. Its hidden layer deploys a bidirectional recurrent neural network with an attention mechanism to extract spatiotemporal correlation features. The output is connected to the risk classification unit. The dynamic threshold generation unit integrates a material property database and an equipment operating parameter interface, updates the warning baseline value in real time through a sliding window statistical method, and transmits the threshold data to the hierarchical control module. The hierarchical control execution module includes a three-level response controller, a command generator, and a central control system connected in sequence via an industrial bus. The input end of the three-level response controller receives the risk assessment results of the analysis and decision-making module. When local deformation exceeds the limit, the first-level response command is triggered, and the control data acquisition module increases the monitoring frequency of the corresponding area. When multiple related areas exceed the limit in coordination, the second-level response command is triggered, and the equipment load adjustment signal is generated and sent to the central control system. When the overall instability risk exceeds the limit, the third-level response command is triggered, and the encrypted shutdown control command is executed; The system self-optimization module includes a sensor reliability assessment unit and a model parameter adjustment unit. The reliability assessment unit calculates the data continuity rate, noise level and consistency error with adjacent sensors of each sensor in real time, generates a weight coefficient and feeds it back to the edge data processing module; the model parameter adjustment unit dynamically optimizes the data fusion weight and classification model decision boundary through a machine learning algorithm based on the early warning record database, and its output end is connected to the model update interface of the analysis and decision module.

[0024] The beneficial effects of the present invention are: The present invention establishes a multi-dimensional, three-dimensional risk assessment system through precise monitoring data and real-time analysis. The seamless connection between each step (such as data acquisition, spatial alignment, pattern recognition and response mechanism) is highly real-time and accurate. Through graded response, appropriate measures can be taken at different risk levels to maximize the structural safety of the hydropower station plant. At the same time, the system's self-optimization capability can dynamically adjust the monitoring model through continuous training and data analysis, improving the system's reliability and predictive capabilities. Finally, combined with the visualization and traceability analysis of three-dimensional point cloud data and sensor monitoring values, it can provide comprehensive diagnostic support when anomalies occur and effectively prevent potential catastrophic risks. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 is a flow chart of the present invention; Figure 2 is a flow chart of the method for determining the key stress concentration area in the present invention; Figure 3 It is an architectural diagram of the hydropower station plant deformation monitoring system of the present invention. DETAILED DESCRIPTION

[0026] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0027] The method for monitoring deformation of a hydropower station building comprises the following steps: Step 1: Strain sensors and inertial measurement units are deployed in key stress concentration areas of the hydropower station building to collect multi-dimensional monitoring data. Temperature and humidity sensors and vibration accelerometers are also deployed to collect environmental interference parameters to comprehensively monitor changes in the hydropower station building structure and environmental impacts. A drone equipped with a multi-lens camera is used to perform periodic 3D scans of the hydropower station building facade to generate point cloud data, which provides a basis for subsequent analysis. When the drone performs periodic 3D scanning, the image acquisition position will shift due to the drone's flight speed and the camera's exposure interval. To reduce this position error, the offset of the image acquisition position needs to be calculated:

[0028] in, v is the flight speed of the UAV, in m / s; is the image exposure time interval, unit s; is the offset, unit is m; The method for determining key stress concentration areas is as follows: by using a finite element simulation model to calculate the stress distribution cloud map of the powerhouse structure under different working conditions, stress concentration areas can be identified, especially those nodes where the maximum principal stress exceeds the set ratio of the material yield strength. These nodes are regarded as potential monitoring areas. Through simulation analysis, the areas that bear the greatest stress under different working conditions can be effectively identified, ensuring that these high-stress areas are monitored in a focused manner, thereby improving the pertinence and effectiveness of monitoring; nodes where the maximum principal stress exceeds the set ratio of the material yield strength are extracted as candidate monitoring areas, and vibration modal analysis is performed on the candidate areas in combination with the vibration test data of the turbine unit. Nodes where vibration energy is concentrated and there is a mechanical transmission path with the equipment base are selected as the final layout points to ensure that the key monitoring areas are covered, and that problems can be discovered in a timely manner and effective measures can be taken when potential risks occur, thereby ensuring the safe operation of the hydropower station powerhouse. These nodes are usually key areas closely related to the hydropower station powerhouse structure and equipment operation, and their deformation or failure may have a significant impact on the operational safety of the entire equipment; The vibration modal analysis process is as follows: collect the vibration signal of the turbine unit during normal operation and extract the characteristic frequency components through Fourier transform:

[0029] in, is the frequency component of the signal in the frequency domain; is the time domain signal; f is the frequency, unit is Hz; e is a natural constant; t is time, unit is s; j is an imaginary unit; Fourier transform can convert the vibration signal in the time domain into a frequency domain signal, thereby analyzing the vibration characteristics at different frequencies; Calculate the resonant response amplitude of each node at the characteristic frequency:

[0030] in, For nodes i At the characteristic frequency f The resonance response amplitude under ; For nodes i External vibration force, unit: N; For nodes i The mass of, in kg; is the damping coefficient, unit is N·s / m; For nodes i The natural frequency, in Hz; Set a safety threshold. The safety threshold is related to the durability and safety of the structure and the yield strength of the material. The specific value depends on the material properties (such as elastic modulus and yield strength), the structural design requirements, and the operating conditions of the hydropower station. 80%-90% of the structural yield strength is used as the safety threshold. Nodes with amplitudes exceeding the safety threshold are selected to increase the density of sensor deployment to more accurately monitor the deformation of these nodes. The method for extracting characteristic frequency components in vibration modal analysis is as follows: performing wavelet packet decomposition on the vibration acceleration signal to extract the sub-band signal energy related to the turbine unit rotation frequency and blade passing frequency. Wavelet packet decomposition is a time-frequency analysis method that can subdivide the vibration signal into multiple frequency bands and accurately extract the energy changes within the target frequency range. Through energy distribution analysis, the changes in vibration intensity within a specific frequency range can be identified, providing a data basis for subsequent analysis; the vibration transfer function of each node at different frequencies is calculated by inputting the cross-power spectral density of the turbine unit base vibration signal and the output node vibration signal:

[0031] in, is the transfer function; is the cross power spectral density; is the power spectrum density of the input unit base vibration signal; the vibration transfer function is used to characterize the dynamic response relationship between the input vibration signal (unit base vibration) and the output vibration signal (plant structure vibration). It can reflect the energy transfer between the two signals at different frequencies and help identify the vibration propagation path. By calculating the vibration transfer function, the impact of unit operation vibration on different nodes of the hydropower station plant structure can be quantified and the response characteristics of each node at different frequencies can be determined. By analyzing the spectral relationship between the input and output signals, setting a gain threshold, and screening out frequency components with transfer function gains above the gain threshold as dangerous resonance frequencies for continuous monitoring and tracking, the system combines this with plant structure deformation monitoring data to determine whether there is a resonance risk. By extracting key vibration signals through wavelet packet decomposition, calculating the vibration transfer function to analyze the signal propagation path, and screening out dangerous resonance frequencies for continuous monitoring, the system achieves precise monitoring and early warning of hydropower plant structure vibration. Ultimately, this monitoring method can effectively improve the operational safety of hydropower stations, reduce structural damage caused by vibration, and increase equipment service life and operational reliability. Step 2: After adaptive filtering and denoising the multi-dimensional monitoring data, spatial registration is performed with the point cloud data in a unified 3D coordinate system. The denoising process dynamically adjusts the filter parameters based on the spectral characteristics of the ambient noise to cope with the ambient noise and ensure data accuracy. The specific method of spatial registration is as follows: pre-set reference targets on the surface of the hydropower station building. The targets contain geometric feature patterns with known three-dimensional coordinates. The targets are used to provide accurate spatial reference points in the photogrammetry process. Since the geometric features and three-dimensional coordinates of the targets are known, they can be easily identified in the image, thus providing a basis for the conversion between the image coordinate system and the global coordinate system; during the three-dimensional scanning process, the target pattern is preferentially identified, and the conversion relationship between the image coordinate system and the global coordinate system is established through the perspective transformation algorithm. The perspective transformation algorithm can convert the two-dimensional coordinates in the image into actual three-dimensional spatial coordinates based on the known three-dimensional coordinates of the target, thereby realizing accurate registration of image data and spatial data; for areas without target coverage, the feature point matching algorithm is used to extract the visual features of the strain sensor installation position, and the strain sensor spatial coordinates are embedded in the point cloud data. The feature point matching algorithm can find specific visual feature points in the image, and these visual feature points have With strong spatial recognition, the position of the strain sensor in space can be further inferred by matching feature points, thereby embedding the spatial coordinates of the strain sensor into the point cloud data to ensure spatial consistency of different data sources. A spatiotemporal alignment model is established at the edge computing node to perform timestamp alignment and spatial position compensation on the multi-dimensional monitoring data and point cloud data. The timestamp alignment uses the GPS timing module to synchronize the clocks of each device to ensure that the data acquisition time of all devices is consistent, and motion speed compensation is performed on the drone image acquisition delay to minimize the impact of the image acquisition delay, thereby improving the synchronization accuracy of the image and other data sources. The alignment process is based on the factory building BIM model to form a unified three-dimensional coordinate system. The spatial mapping relationship is established through the feature point matching algorithm, and the coordinate conversion error is corrected by the iterative optimization algorithm to ensure the accuracy of the spatial alignment, thereby providing accurate spatiotemporal data support for subsequent analysis. The time delay compensation method in the spatiotemporal alignment model is as follows: the image acquisition position offset is calculated according to the UAV flight speed and the camera exposure interval, the position error is eliminated by the motion blur correction algorithm in the point cloud reconstruction stage, and the clock deviation between the sensor data acquisition time and the image exposure time is compensated by polynomial fitting. The compensation amount is dynamically adjusted according to the historical synchronization error data, and an abnormal time difference detection mechanism is established. When the clock deviation exceeds the set range, forced clock synchronization based on the NTP protocol is started; the position error can be eliminated by the motion blur correction algorithm, and the pixel position in the image is dynamically adjusted according to the flight trajectory of the UAV and the image exposure information to achieve the effect of eliminating motion blur, thereby improving the accuracy of point cloud reconstruction; when the sensor data is collected, there is a clock deviation between the sensor data acquisition time and the image exposure time. In order to accurately align the sensor data and image data, the clock deviation needs to be compensated. The specific process is: through the historical synchronization error data, the clock deviation is modeled using polynomial fitting, and a function of the clock deviation compensation amount is obtained by fitting. The compensation amount will be calculated according to the historical synchronization error data. The system dynamically adjusts to changes in historical data to ensure that the clock deviation is continuously optimized over time, thereby ensuring the timing accuracy of sensor data and image exposure data. In some cases, the clock deviation may change abnormally, causing the synchronization error to exceed the acceptable range. To avoid this, an abnormal time difference detection mechanism is established. This mechanism monitors the changes in clock deviation. Once the time difference exceeds the preset threshold, it will start the forced clock synchronization function based on the NTP protocol. The NTP protocol can accurately synchronize the clock of the computer system to ensure that the data acquisition time of the drone and the sensor is completely consistent, eliminating the error caused by clock deviation. The delay compensation method ensures the precise alignment of sensor data and image data in time and space by combining the drone flight speed, image acquisition interval, clock deviation correction and abnormal time difference detection mechanism. Through this series of steps and methods, not only the accuracy of image reconstruction is improved, but also the control capability of the hydropower station powerhouse deformation monitoring system for real-time data acquisition and timing synchronization is enhanced, which helps to timely discover and deal with potential deformation risks.

[0032] Step 3: Build a multi-channel time-series deep learning model for deformation pattern recognition. A dynamic warning threshold model is constructed based on material properties, equipment operating parameters, and historical monitoring data. When the real-time monitoring value deviates from the historical data distribution center by more than a preset probability threshold, the threshold update mechanism is triggered to update the dynamic warning threshold to avoid false alarms or missed alarms due to environmental changes or special operating conditions. The dynamic warning threshold model uses a sliding window statistical method to update the baseline value in real time for risk assessment. The construction method of the dynamic warning threshold model is to establish the temperature-stress coupling relationship function of the concrete structure:

[0033] in, Temperature T The stress under the pressure, unit is Pa; Temperature T The elastic modulus of the material under , in Pa; Temperature T It can reflect the stress response of concrete under different temperature changes and accurately monitor and warn the deformation of hydropower station buildings under different working conditions. Through the stress-strain experimental data at different temperatures, the elastic modulus of the material at each temperature point is obtained and the corresponding strain , through regression analysis or fitting method, the relationship between temperature, stress and strain is obtained, and the elastic modulus of the material is dynamically corrected based on real-time temperature data , and then calculate the corresponding stress value to determine the deformation under temperature changes. The elastic modulus parameters of the material are corrected through the real-time collected temperature data, and then the dynamic adjustment range of the allowable deformation is calculated:

[0034] in, For the temperature T Maximum deformation under is the corresponding maximum stress, unit: Pa; is the elastic modulus after temperature correction, unit: Pa; Based on the correlation between turbine unit output parameters and structural vibration intensity, a load-deformation transfer function is constructed to determine the dynamic baseline value. A sliding time window statistical method is used to analyze the distribution characteristics of historical monitoring data, and the current monitoring data is monitored in real time to ensure that it conforms to the distribution pattern of historical data. When the real-time monitoring value deviates from the historical data distribution center by more than a preset probability, a threshold update mechanism is triggered. The load-deformation transfer function is established as follows: under different load conditions, the vibration acceleration of the concrete structure is measured using an accelerometer, while the output parameters of the turbine unit are recorded. The measured output parameters and the corresponding vibration acceleration data are then subjected to regression analysis to establish a functional relationship between output and vibration acceleration:

[0035] in, is the structural vibration acceleration, unit is m / s 2 ; To generate power for the turbine unit, is the regression coefficient, is the offset constant; Establish the load-deformation transfer function:

[0036] in, is the deformation of the concrete structure, unit is m; k is the stiffness of the concrete structure, in N / m; it is used to describe the transfer relationship between the turbine unit output (i.e., load) and the deformation of the hydropower station building structure; it can dynamically determine the baseline value of the structural deformation according to the load changes of the turbine unit. The construction of the load-deformation transfer function relies on the correspondence between the turbine unit output and the structural vibration acceleration under different load conditions; At the same time, the transfer coefficient of vibration energy to structural deformation is determined by regression analysis:

[0037]

[0038] in, is the transfer coefficient of vibration energy to structural deformation; is the vibration energy of the turbine unit or other equipment, unit is j; m is the mass of the plant structure, in kg; a is the vibration acceleration, unit is m / s 2 The corresponding relationship between turbine unit output and structural vibration acceleration under different load conditions is recorded, and regression analysis is used to calculate how vibration energy is converted into structural deformation. The core of this process is to infer structural deformation from vibration data, thereby providing a mathematical model for the relationship between load and deformation. The multi-channel time series deep learning model includes an input layer, a hidden layer and an output layer. The input layer receives the fused strain, displacement, acceleration and environmental parameter data. The hidden layer uses a bidirectional recurrent neural network with an attention mechanism to extract spatiotemporal correlation features. The output layer synchronously outputs the structural deformation pattern classification results and risk probability prediction values. The multi-channel time series deep learning model extracts spatiotemporal correlation features in the data through a bidirectional recurrent neural network with an attention mechanism, outputs structural deformation patterns and risk prediction values, and updates the monitoring baseline value in real time through a dynamic early warning threshold model combined with material properties, equipment operating parameters and historical data. The bidirectional recurrent neural network with an attention mechanism extracts spatiotemporal correlation features in the data, outputs structural deformation patterns and risk prediction values, ensuring that potential risks can be discovered and early warnings can be issued in a timely manner.

[0039] Step 4: Trigger a three-level response mechanism based on the risk assessment results. The three-level response mechanism constitutes a complete early warning and emergency control process. Its core is to set a hierarchical response strategy according to different risk levels, and achieve accurate monitoring, rapid response and effective disposal through data-driven trigger logic. Based on the deviation analysis between monitoring data and early warning results, establish sensor reliability evaluation indicators and dynamically optimize data fusion weights. Adjust the classification model decision boundary through machine learning algorithms, and visualize and trace the multi-dimensional monitoring data to reduce the false alarm rate. The sensor reliability evaluation indicator is obtained by comprehensively calculating the data continuity rate, noise variance and consistency error with adjacent sensors. The data continuity rate is used to measure the consistency of sensor data collection within a certain period of time. Sensors with many missing values ​​or discontinuous data are considered unreliable. The noise variance can evaluate the noise level of the sensor output. Sensors with large noise will reduce the accuracy and stability of the system data. The consistency error is compared with the data of adjacent sensors to evaluate the error between each sensor. If the data consistency between adjacent sensors is poor, it indicates that there is a sensor failure or external interference. These evaluation indicators together constitute the sensor reliability score, which can be used to dynamically The performance of each sensor is dynamically evaluated, providing a basis for adjusting data fusion weights during the optimization process. The gradient descent algorithm is used to minimize the weighted sum of the mean squared error of the prediction model and the false alarm rate of the warning. The mean squared error represents the accuracy of the model's prediction results, while the false alarm rate reflects the misjudgment of the warning. The gradient descent algorithm can continuously adjust the weight of each sensor data based on error feedback, allowing the system to minimize the error while reducing the false alarm rate, thereby improving the system's stability and accuracy. The weighted average algorithm is used to optimize the multi-source data fusion weights, with the optimization goal of minimizing the weighted sum of the mean squared error of the prediction model and the false alarm rate of the warning. The particle swarm optimization algorithm (PSO) is used to regularly adjust the decision boundary of the classification model to adapt to changes in the monitoring system over different time periods. Through optimization, the particle swarm optimization algorithm continuously reduces the false alarm rate and stabilizes the false alarm rate until the predetermined performance standard is met. The particle swarm optimization algorithm helps the system adjust the classification model based on historical data, enabling it to cope with long-term changes in the monitoring environment and ensure the reliability of the warning mechanism. The adjustment process uses the warning records of the last 30 days as the training set, and the reduction of the false alarm rate and the stabilization of the false alarm rate as the convergence conditions. The triggering logic of the three-level response mechanism is as follows: the first-level response is activated when the deformation of any single monitoring point exceeds 120% of the dynamic threshold, triggering the monitoring frequency of the area to increase to 3 times the original rate, and retrieving the historical data of the last 24 hours for trend comparison to identify whether there is accelerated deformation or periodic fluctuations; the second-level response is activated when three or more adjacent monitoring points exceed the limit at the same time and the deformation direction shows mechanical correlation, generating equipment load reduction operation suggestions and starting backup sensors for cross-validation to confirm the accuracy of the original monitoring data and eliminate single-point errors; the third-level response is activated when the probability of overall structural instability exceeds 90% and the drone re-measurement data is consistent with the edge computing results When the risk of local deformation exceeds the limit, the monitoring frequency is increased and manual review is initiated. When multiple related areas exceed the limit, equipment load adjustment instructions are generated. When the overall instability risk exceeds the critical value, an automatic protection shutdown is executed. The specific method of visualizing and tracing the multi-dimensional monitoring data is as follows: superimposing the three-dimensional point cloud data and sensor monitoring values ​​on the surface of the plant BIM model. The plant BIM model provides the spatial structure data of the hydropower plant, and the three-dimensional point cloud data obtains the actual deformation information of the surface of the hydropower plant through laser scanning; the deformation distribution is represented by color gradient mapping, and a multi-dimensional correlation map is constructed to show the temporal and spatial correlation between the structural deformation trend and the equipment vibration parameters and ambient temperature; the sensor monitoring values ​​provide real-time data on the deformation, vibration and temperature of various parts of the plant. The superposition of these data enables the hydropower plant model to truly reflect the deformation state, and the color gradient mapping is used to express different deformation amounts in different colors, thereby intuitively showing the deformation distribution and facilitating the operator to quickly identify the problem area; by constructing a multi-dimensional correlation map, the temporal and spatial relationship between the structural deformation of the hydropower plant and the factors of equipment vibration and ambient temperature can be further analyzed. This multi-dimensional correlation map includes time axis, spatial position, vibration frequency, and other parameters. The system can associate data from multiple dimensions, such as rate and temperature changes, to reveal whether the deformation of the hydropower station building is related to the operating status of the equipment or ambient temperature fluctuations. This not only helps to identify potential structural problems in advance, but also helps to determine whether the external environment has an impact on the hydropower station building structure. The design of the traceability query function provides support for the intelligentization of the hydropower station building deformation monitoring system. When the hydropower station building deformation monitoring system detects an anomaly, it automatically generates a list of associated sensors, indicating areas with large deformation or vibration. At this time, through the traceability query function, users can view the historical deformation curve of the area and the timeline markers of key events related to the area, thereby helping engineers quickly locate the time and cause of the problem. This function significantly improves the early warning and response capabilities of the hydropower station building deformation monitoring system. Combined with BIM models and real-time sensor data, it not only improves the accuracy of deformation monitoring, but also strengthens the early warning capabilities of potential risks, allowing engineers to take timely measures before problems occur to avoid greater losses.

[0040] The hydropower station plant deformation monitoring system includes a multi-source data acquisition module, an edge data processing module, an analysis and decision module, and a system self-optimization module, which are connected in sequence through industrial Ethernet. The analysis and decision module is connected to the hierarchical control execution module, and the system self-optimization module is connected to the edge data processing module and the analysis and decision module through a control bus. Among them: the output end of the multi-source data acquisition module provides a standardized data stream to the analysis and decision module after preprocessing by the edge data processing module. The preprocessing includes data fusion and time-space alignment based on edge computing; the output end of the analysis and decision module transmits a risk assessment level signal to the hierarchical control execution module, and at the same time sends the early warning record back to the system self-optimization module; the parameter adjustment instructions generated by the system self-optimization module are synchronously sent to the edge data processing module and the analysis and decision module through the control bus.

[0041] The multi-source data acquisition module includes a distributed sensor network and a drone equipped with a multi-lens camera. The drone is used to perform periodic 3D scanning of the hydropower station's powerhouse facade and generate high-precision point cloud data. The distributed sensor network includes strain sensors, inertial measurement units, temperature and humidity sensors, and vibration accelerometers deployed in key stress concentration areas of the powerhouse. The strain sensors, inertial measurement units, and drones are connected to the edge data processing module via wireless communication modules. The edge data processing module includes an adaptive filtering unit and a spatiotemporal registration unit. The adaptive filtering unit receives time series data transmitted by the sensor network and dynamically adjusts the filtering parameters according to the spectral characteristics of the ambient noise to eliminate interference signals. The spatiotemporal registration unit has a built-in factory building BIM model database and aligns the photogrammetric point cloud data with the sensor data in a unified three-dimensional coordinate system through a feature point matching algorithm. It also uses an iterative optimization algorithm to correct coordinate conversion errors. The output of the spatiotemporal registration unit is connected to the data analysis module. The analysis and decision-making module includes a multi-channel time-series deep learning model and a dynamic threshold generation unit. The input of the deep learning model receives fused strain, displacement, acceleration, and environmental parameter data. Its hidden layer deploys a bidirectional recurrent neural network with an attention mechanism to extract spatiotemporal correlation features. The output is connected to the risk classification unit. The dynamic threshold generation unit integrates a material property database and an equipment operating parameter interface, updates the warning baseline value in real time through a sliding window statistical method, and transmits the threshold data to the hierarchical control module. The hierarchical control execution module includes a three-level response controller, a command generator, and a central control system connected in sequence via an industrial bus. The input end of the three-level response controller receives the risk assessment results of the analysis and decision-making module. When local deformation exceeds the limit, the first-level response command is triggered, and the control data acquisition module increases the monitoring frequency of the corresponding area. When multiple related areas exceed the limit in coordination, the second-level response command is triggered, and the equipment load adjustment signal is generated and sent to the central control system. When the overall instability risk exceeds the limit, the third-level response command is triggered, and the encrypted shutdown control command is executed; The system self-optimization module includes a sensor reliability assessment unit and a model parameter adjustment unit. The reliability assessment unit calculates the data continuity rate, noise level and consistency error with adjacent sensors of each sensor in real time, generates a weight coefficient and feeds it back to the edge data processing module; the model parameter adjustment unit dynamically optimizes the data fusion weight and classification model decision boundary through a machine learning algorithm based on the early warning record database, and its output end is connected to the model update interface of the analysis and decision module.

[0042] Temperature and humidity sensors are used to collect temperature data from the hydropower plant's surrounding environment. Temperature directly affects the expansion or contraction of structural materials, which in turn affects the plant's structural deformation. For example, temperature changes can cause thermal expansion and contraction of building materials such as concrete and steel bars, thereby altering the strain sensor readings. Changes in air humidity, especially relative humidity, affect the hygroscopicity of building materials, further affecting their strength and stiffness. These humidity changes can impact the durability of equipment and structures within the hydropower plant, as well as the stability of the strain data detected by the sensors. Vibration accelerometers are used to monitor vibration disturbances inside and outside the hydropower plant. These vibrations can originate from a variety of factors, such as turbine operation, equipment operating noise, and vibrations from the surrounding environment (such as traffic and construction). Vibration acceleration directly affects the response of the strain sensor. Excessive vibration can cause unnecessary strain fluctuations, interfering with the actual structural deformation data, necessitating compensation. By deploying these environmental disturbance sensors, the hydropower plant's deformation monitoring system can effectively capture the impact of the external environment, thereby improving the accuracy of the overall data.

[0043] Example 1: The method for monitoring deformation of a hydropower station building comprises the following steps: Step 1: Strain sensors and inertial measurement units are deployed in key stress concentration areas of the hydropower station building to collect multi-dimensional monitoring data. Temperature and humidity sensors and vibration accelerometers are also deployed to collect environmental interference parameters. A drone equipped with a multi-lens camera is used to perform periodic 3D scans of the hydropower station building facade to generate point cloud data. The offset of the image acquisition position is calculated during the periodic 3D scanning by the drone:

[0044] in, v is the flight speed of the UAV, in m / s; is the image exposure time interval, unit s; is the offset, unit is m; The method for determining key stress concentration areas is as follows: using a finite element simulation model to calculate the stress distribution cloud map of the powerhouse structure under different operating conditions, the nodes where the maximum principal stress exceeds a set ratio of the material yield strength are extracted as candidate monitoring areas. Vibration modal analysis of the candidate areas is then performed in combination with the vibration test data of the turbine unit. The nodes where vibration energy is concentrated and there is a mechanical transmission path with the equipment base are selected as the final layout points. The vibration modal analysis process is as follows: collect the vibration signal of the turbine unit during normal operation and extract the characteristic frequency components through Fourier transform:

[0045] in, is the frequency component of the signal in the frequency domain; is the time domain signal; f is the frequency, unit is Hz; e is a natural constant; t is time, unit is s; j is an imaginary unit; Calculate the resonant response amplitude of each node at the characteristic frequency:

[0046] in, For nodes i At the characteristic frequency f The resonance response amplitude under ; For nodes i External vibration force, unit: N; For nodes i The mass of, in kg; is the damping coefficient, unit is N·s / m; For nodes i The natural frequency, in Hz; Set a safety threshold and select nodes with amplitudes exceeding the safety threshold to increase the density of sensor deployment; The method for extracting characteristic frequency components in vibration modal analysis is as follows: perform wavelet packet decomposition on the vibration acceleration signal, extract the sub-band signal energy related to the turbine unit rotation frequency and blade passing frequency, and calculate the vibration transfer function of each node at different frequencies through the cross-power spectral density of the input unit base vibration signal and the output node vibration signal:

[0047] in, is the transfer function; is the cross power spectral density; is the power spectrum density of the input unit base vibration signal; Analyze the spectral relationship between the input and output signals, set a gain threshold, and screen out frequency components with transfer function gains higher than the gain threshold as dangerous resonance frequencies for continuous monitoring and tracking. Combined with plant structure deformation monitoring data, determine whether there is a resonance risk. Step 2: After adaptive filtering and noise reduction, the multi-dimensional monitoring data is spatially registered with the point cloud data in a unified three-dimensional coordinate system; Step 3: Build a multi-channel time series deep learning model and construct a dynamic warning threshold model based on material properties, equipment operating parameters, and historical monitoring data to conduct risk assessment; Step 4: Trigger the three-level response mechanism based on the risk assessment results. Based on the deviation analysis between the monitoring data and the early warning results, establish the sensor reliability evaluation index and dynamically optimize the data fusion weight, adjust the classification model decision boundary, and visualize and trace the multi-dimensional monitoring data.

[0048] Example 2: The method for monitoring deformation of a hydropower station building comprises the following steps: Step 1: Collect multi-dimensional monitoring data and point cloud data of the hydropower station building; Step 2: After adaptive filtering and denoising the multi-dimensional monitoring data, spatial registration is performed with the point cloud data in a unified 3D coordinate system. The denoising process dynamically adjusts the filter parameters based on the spectral characteristics of the ambient noise. The specific method of spatial registration is as follows: pre-place benchmark targets on the surface of the hydropower station building. The targets contain geometric feature patterns with known three-dimensional coordinates. During the three-dimensional scanning process, the target patterns are preferentially identified, and the conversion relationship between the image coordinate system and the global coordinate system is established through the perspective transformation algorithm. For areas without target coverage, the feature point matching algorithm is used to extract the visual features of the strain sensor installation position, and the strain sensor spatial coordinates are embedded in the point cloud data; a spatiotemporal alignment model is established at the edge computing node to perform timestamp alignment and spatial position compensation on the multi-dimensional monitoring data and point cloud data. The timestamp alignment uses the GPS timing module to synchronize the clocks of each device, and motion speed compensation is performed on the drone image acquisition delay. The registration process uses the plant BIM model as the benchmark to form a unified three-dimensional coordinate system. The spatial mapping relationship is established through the feature point matching algorithm, and the coordinate conversion error is corrected using an iterative optimization algorithm. The time delay compensation method in the spatiotemporal alignment model is as follows: the image acquisition position offset is calculated based on the drone's flight speed and the camera exposure interval. During the point cloud reconstruction phase, a motion blur correction algorithm is used to eliminate position errors. The clock deviation between the sensor data acquisition time and the image exposure time is compensated by polynomial fitting. The compensation amount is dynamically adjusted based on historical synchronization error data. An abnormal time difference detection mechanism is established. When the clock deviation exceeds the set range, forced clock synchronization based on the NTP protocol is initiated. Step 3: Build a multi-channel time series deep learning model and construct a dynamic warning threshold model based on material properties, equipment operating parameters, and historical monitoring data to conduct risk assessment; Step 4: Trigger the three-level response mechanism based on the risk assessment results. Based on the deviation analysis between the monitoring data and the early warning results, establish the sensor reliability evaluation index and dynamically optimize the data fusion weight, adjust the classification model decision boundary, and visualize and trace the multi-dimensional monitoring data.

[0049] Example 3: The method for monitoring deformation of a hydropower station building comprises the following steps: Step 1: Collect multi-dimensional monitoring data and point cloud data of the hydropower station building; Step 2: After adaptive filtering and noise reduction, the multi-dimensional monitoring data is spatially registered with the point cloud data in a unified three-dimensional coordinate system; Step 3: Build a multi-channel time series deep learning model and construct a dynamic warning threshold model based on material properties, equipment operating parameters, and historical monitoring data. The dynamic warning threshold model uses a sliding window statistical method to update the baseline value in real time to perform risk assessment. The construction method of the dynamic warning threshold model is to establish the temperature-stress coupling relationship function of the concrete structure:

[0050] in, Temperature T The stress under the pressure, unit is Pa; Temperature T The elastic modulus of the material under , in Pa; Temperature T strain under Through the stress-strain experimental data at different temperatures, the elastic modulus of the material at each temperature point is obtained and the corresponding strain , through regression analysis or fitting method, the relationship between temperature, stress and strain is obtained, and the elastic modulus of the material is dynamically corrected based on real-time temperature data , and then calculate the corresponding stress value to determine the deformation under temperature changes. The elastic modulus parameters of the material are corrected through the real-time collected temperature data, and then the dynamic adjustment range of the allowable deformation is calculated:

[0051] in, For the temperature T Maximum deformation under is the corresponding maximum stress, unit: Pa; is the elastic modulus after temperature correction, unit: Pa; Based on the correlation between turbine unit output parameters and structural vibration intensity, a load-deformation transfer function is constructed to determine the dynamic baseline value. A sliding time window statistical method is used to analyze the distribution characteristics of historical monitoring data, and the current monitoring data is monitored in real time to ensure that it conforms to the distribution pattern of historical data. When the real-time monitoring value deviates from the historical data distribution center by more than a preset probability, a threshold update mechanism is triggered. The load-deformation transfer function is established as follows: under different load conditions, the vibration acceleration of the concrete structure is measured using an accelerometer, while the output parameters of the turbine unit are recorded. The measured output parameters and the corresponding vibration acceleration data are then subjected to regression analysis to establish a functional relationship between output and vibration acceleration:

[0052] in, is the structural vibration acceleration, unit is m / s 2 ; To generate power for the turbine unit, is the regression coefficient, is the offset constant; Establish the load-deformation transfer function:

[0053] in, is the deformation of the concrete structure, unit is m; k is the stiffness of the concrete structure, in N / m; At the same time, the transfer coefficient of vibration energy to structural deformation is determined by regression analysis:

[0054]

[0055] in, is the transfer coefficient of vibration energy to structural deformation; is the vibration energy of the turbine unit or other equipment, unit is j; m is the mass of the plant structure, in kg; a is the vibration acceleration, unit is m / s 2 ; The multi-channel time series deep learning model includes an input layer, a hidden layer, and an output layer. The input layer receives fused strain, displacement, acceleration, and environmental parameter data. The hidden layer uses a bidirectional recurrent neural network with an attention mechanism to extract spatiotemporal correlation features. The output layer simultaneously outputs the structural deformation pattern classification results and risk probability prediction values. The multi-channel time series deep learning model extracts spatiotemporal correlation features from the data through a bidirectional recurrent neural network with an attention mechanism, outputs the structural deformation pattern and risk prediction value, and updates the monitoring baseline value in real time through a dynamic warning threshold model combined with material properties, equipment operating parameters, and historical data. Step 4: Trigger the three-level response mechanism based on the risk assessment results. Based on the deviation analysis between the monitoring data and the early warning results, establish the sensor reliability evaluation index and dynamically optimize the data fusion weight, adjust the classification model decision boundary, and visualize and trace the multi-dimensional monitoring data.

[0056] Example 4: The method for monitoring deformation of a hydropower station building comprises the following steps: Step 1: Collect multi-dimensional monitoring data and point cloud data of the hydropower station building; Step 2: After adaptive filtering and noise reduction, the multi-dimensional monitoring data is spatially registered with the point cloud data in a unified three-dimensional coordinate system; Step 3: Build a multi-channel time series deep learning model and construct a dynamic warning threshold model based on material properties, equipment operating parameters, and historical monitoring data to conduct risk assessment; Step 4: Trigger a three-level response mechanism based on the risk assessment results. Based on the deviation analysis between monitoring data and warning results, establish sensor reliability evaluation indicators and dynamically optimize data fusion weights. Use machine learning algorithms to adjust the classification model decision boundary, and perform visualization and traceability analysis on multi-dimensional monitoring data. The sensor reliability evaluation indicator is calculated by comprehensively calculating the data continuity rate, noise variance, and consistency error with adjacent sensors. The gradient descent algorithm is used to minimize the weighted sum of the mean square error of the prediction model and the warning false alarm rate. The weighted average algorithm is used to optimize the multi-source data fusion weights. The optimization goal is to minimize the weighted sum of the mean square error of the prediction model and the warning false alarm rate. The particle swarm optimization algorithm is used regularly to adjust the decision boundary of the classification model. The warning records of the last 30 days are used as the training set during the adjustment process, and the convergence conditions are a decrease in the false alarm rate and a stable omission rate. The triggering logic of the three-level response mechanism is as follows: the first-level response is activated when the deformation of any single monitoring point exceeds 120% of the dynamic threshold, triggering the monitoring frequency of the area to increase to 3 times the original rate, and retrieving the historical data of the last 24 hours for trend comparison to identify whether there is accelerated deformation or periodic fluctuations; the second-level response is activated when three or more adjacent monitoring points exceed the limit at the same time and the deformation direction shows mechanical correlation, generating a recommendation for equipment load reduction operation and starting the backup sensor for cross-validation to confirm the accuracy of the original monitoring data and eliminate single-point errors; the third-level response is activated when the probability of overall structural instability exceeds 90% and the consistency between the drone re-measurement data and the edge computing results is less than 80%, sending an encrypted shutdown command to the central control system and marking the coordinates of the dangerous area. At the same time, the central control system automatically marks the coordinates of the dangerous area corresponding to the current monitoring results; Increase monitoring frequency when local deformation exceeds the limit; generate equipment load adjustment instructions when multiple related areas exceed the limit; and execute automatic protection shutdown when the overall instability risk exceeds the critical value; The specific method for visualizing and tracing the multi-dimensional monitoring data is to superimpose the three-dimensional point cloud data and sensor monitoring values ​​on the surface of the factory BIM model, characterize the deformation distribution through color gradient mapping, and construct a multi-dimensional correlation map to show the temporal and spatial correlation between the structural deformation trend and the equipment vibration parameters and ambient temperature.

[0057] Example 5: The hydropower station plant deformation monitoring system includes a multi-source data acquisition module, an edge data processing module, an analysis and decision module, and a system self-optimization module, which are connected in sequence through industrial Ethernet. The analysis and decision module is connected to a hierarchical control execution module, and the system self-optimization module is connected to the edge data processing module and the analysis and decision module through a control bus.

[0058] Example 6: The hydropower station powerhouse deformation monitoring system includes a multi-source data acquisition module, an edge data processing module, an analysis and decision-making module, a hierarchical control execution module, and a system self-optimization module, which form a closed-loop control loop through industrial Ethernet. The multi-source data acquisition module includes a distributed sensor network and a drone equipped with a multi-lens camera. The distributed sensor network includes strain sensors, inertial measurement units, temperature and humidity sensors, and vibration accelerometers deployed in key stress concentration areas of the factory. The strain sensors, inertial measurement units, and drones are connected to the edge data processing module through wireless communication modules. The edge data processing module includes an adaptive filtering unit and a spatiotemporal registration unit. The adaptive filtering unit receives time series data transmitted by the sensor network and dynamically adjusts the filtering parameters according to the spectral characteristics of the ambient noise to eliminate interference signals. The spatiotemporal registration unit has a built-in factory building BIM model database and aligns the photogrammetric point cloud data with the sensor data in a unified three-dimensional coordinate system through a feature point matching algorithm. It also uses an iterative optimization algorithm to correct coordinate conversion errors. The output of the spatiotemporal registration unit is connected to the data analysis module. The analysis and decision-making module includes a multi-channel time-series deep learning model and a dynamic threshold generation unit. The input of the deep learning model receives fused strain, displacement, acceleration, and environmental parameter data. Its hidden layer deploys a bidirectional recurrent neural network with an attention mechanism to extract spatiotemporal correlation features. The output is connected to the risk classification unit. The dynamic threshold generation unit integrates a material property database and an equipment operating parameter interface, updates the warning baseline value in real time through a sliding window statistical method, and transmits the threshold data to the hierarchical control module. The hierarchical control execution module includes a three-level response controller, a command generator, and a central control system connected in sequence via an industrial bus. The input end of the three-level response controller receives the risk assessment results of the analysis and decision-making module. When local deformation exceeds the limit, the first-level response command is triggered, and the control data acquisition module increases the monitoring frequency of the corresponding area. When multiple related areas exceed the limit in coordination, the second-level response command is triggered, and the equipment load adjustment signal is generated and sent to the central control system. When the overall instability risk exceeds the limit, the third-level response command is triggered, and the encrypted shutdown control command is executed; The system self-optimization module includes a sensor reliability assessment unit and a model parameter adjustment unit. The reliability assessment unit calculates the data continuity rate, noise level and consistency error with adjacent sensors of each sensor in real time, generates a weight coefficient and feeds it back to the edge data processing module; the model parameter adjustment unit dynamically optimizes the data fusion weight and classification model decision boundary through a machine learning algorithm based on the early warning record database, and its output end is connected to the model update interface of the analysis and decision module.

Claims

1. A method for monitoring deformation of a hydropower plant building, characterized in that: The following steps are involved: Step 1: Collect multi-dimensional monitoring data and point cloud data of the hydropower station building; Step 2: After adaptive filtering and noise reduction, the multi-dimensional monitoring data is spatially registered with the point cloud data in a unified three-dimensional coordinate system; Step 3: Build a multi-channel time series deep learning model and construct a dynamic warning threshold model based on material properties, equipment operating parameters, and historical monitoring data to conduct risk assessment; Step 4: Trigger the three-level response mechanism based on the risk assessment results. Based on the deviation analysis between the monitoring data and the early warning results, establish the sensor reliability evaluation index and dynamically optimize the data fusion weight, adjust the classification model decision boundary, and visualize and trace the multi-dimensional monitoring data.

2. The method for monitoring deformation of a hydropower plant building according to claim 1, characterized in that: In step 1, strain sensors and inertial measurement units are deployed in key stress concentration areas of the hydropower station building to collect multi-dimensional monitoring data. A drone equipped with a multi-lens camera is used to periodically perform three-dimensional scanning of the hydropower station building facade to generate point cloud data. The method for determining key stress concentration areas is as follows: using a finite element simulation model to calculate the stress distribution cloud map of the powerhouse structure under different operating conditions, the nodes where the maximum principal stress exceeds a set ratio of the material yield strength are extracted as candidate monitoring areas. Vibration modal analysis of the candidate areas is then performed in combination with the vibration test data of the turbine unit. The nodes where vibration energy is concentrated and there is a mechanical transmission path with the equipment base are selected as the final layout points. The vibration modal analysis process is as follows: collect the vibration signal of the turbine unit during normal operation and extract the characteristic frequency components through Fourier transform: in, is the frequency component of the signal in the frequency domain; is the time domain signal; f is the frequency, unit is Hz; e is a natural constant; t is time, unit is s; j is an imaginary unit; Calculate the resonant response amplitude of each node at the characteristic frequency: in, For nodes i At the characteristic frequency f The resonance response amplitude under ; For nodes i External vibration force, unit: N; For nodes i The mass of, in kg; is the damping coefficient, unit is N·s / m; For nodes i The natural frequency, in Hz; Set a safety threshold and select nodes with amplitudes exceeding the safety threshold to increase the density of sensor deployment; The method for extracting characteristic frequency components in vibration modal analysis is as follows: perform wavelet packet decomposition on the vibration acceleration signal, extract the sub-band signal energy related to the turbine unit rotation frequency and blade passing frequency, and calculate the vibration transfer function of each node at different frequencies through the cross-power spectral density of the input unit base vibration signal and the output node vibration signal: in, is the transfer function; is the cross power spectral density; is the power spectrum density of the input unit base vibration signal; Analyze the spectral relationship between the input signal and the output signal, set the gain threshold, and screen out the frequency components whose transfer function gain is higher than the gain threshold as dangerous resonance frequencies for continuous monitoring and tracking. Combined with the plant structure deformation monitoring data, determine whether there is a resonance risk.

3. The method for monitoring deformation of a hydropower plant building according to claim 2, characterized in that: The specific method of spatial registration in step 2 is as follows: pre-setting a reference target on the surface of the hydropower station building, the target containing a geometric feature pattern with known three-dimensional coordinates, prioritizing the recognition of the target pattern during the three-dimensional scanning process, and establishing a conversion relationship between the image coordinate system and the global coordinate system through a perspective transformation algorithm. For areas not covered by the target, a feature point matching algorithm is used to extract the visual features of the strain sensor installation position, and the strain sensor spatial coordinates are embedded in the point cloud data; A spatiotemporal alignment model is established at the edge computing node to perform timestamp alignment and spatial position compensation on the multi-dimensional monitoring data and point cloud data. The timestamp alignment uses the GPS timing module to synchronize the clocks of each device, and motion speed compensation is performed on the drone image acquisition delay. The alignment process uses the factory building BIM model as the benchmark to form a unified three-dimensional coordinate system. The spatial mapping relationship is established through a feature point matching algorithm, and an iterative optimization algorithm is used to correct coordinate conversion errors. The time delay compensation method in the spatiotemporal alignment model is as follows: the image acquisition position offset is calculated based on the UAV flight speed and the camera exposure interval. The position error is eliminated through a motion blur correction algorithm in the point cloud reconstruction stage. The clock deviation between the sensor data acquisition time and the image exposure time is compensated by polynomial fitting. The compensation amount is dynamically adjusted based on historical synchronization error data. An abnormal time difference detection mechanism is established. When the clock deviation exceeds the set range, forced clock synchronization based on the NTP protocol is initiated.

4. The method for monitoring deformation of a hydropower plant building according to claim 3, characterized in that: The method for constructing the dynamic warning threshold model in step 3 is as follows: establishing a temperature-stress coupling relationship function for the concrete structure, correcting the material elastic modulus parameter through real-time collected temperature data, and then calculating the dynamic adjustment range of the allowable deformation; constructing a load-deformation transfer function based on the correlation between the turbine unit output parameter and the structural vibration intensity to determine the dynamic reference value; using a sliding time window statistical method to analyze the distribution characteristics of historical monitoring data, monitoring in real time whether the current monitoring data conforms to the distribution pattern of historical data, and triggering a threshold update mechanism when the real-time monitoring value deviates from the historical data distribution center by more than a preset probability; The load-deformation transfer function is established as follows: under different load conditions, the vibration acceleration of the concrete structure is measured using an accelerometer, while the output parameters of the turbine unit are recorded. The measured output parameters and the corresponding vibration acceleration data are then subjected to regression analysis to establish a functional relationship between output and vibration acceleration: in, is the structural vibration acceleration, unit is m / s 2 ; To generate power for the turbine unit, is the regression coefficient, is the offset constant; Establish the load-deformation transfer function: in, is the deformation of the concrete structure, unit is m; k is the stiffness of the concrete structure, in N / m; At the same time, the transfer coefficient of vibration energy to structural deformation is determined by regression analysis: in, is the transfer coefficient of vibration energy to structural deformation; is the vibration energy of the turbine unit or other equipment, unit is j; m is the mass of the plant structure, in kg; a is the vibration acceleration, unit is m / s 2 ; The multi-channel time series deep learning model includes an input layer, a hidden layer, and an output layer. The input layer receives the fused strain, displacement, acceleration, and environmental parameter data. The hidden layer uses a bidirectional recurrent neural network with an attention mechanism to extract spatiotemporal correlation features. The output layer synchronously outputs the structural deformation pattern classification results and risk probability prediction values. The multi-channel time series deep learning model extracts spatiotemporal correlation features in the data through a bidirectional recurrent neural network with an attention mechanism, outputs structural deformation patterns and risk prediction values, and updates the monitoring benchmark values ​​in real time through a dynamic early warning threshold model combined with material properties, equipment operating parameters, and historical data.

5. The method for monitoring deformation of a hydropower plant building according to claim 4, characterized in that: The triggering logic of the three-level response mechanism in step 4 is as follows: the first-level response is activated when the deformation of any single monitoring point exceeds 120% of the dynamic threshold, triggering the monitoring frequency of the area to be increased to 3 times the original rate, and retrieving the historical data of the last 24 hours for trend comparison to identify whether there is accelerated deformation or periodic fluctuation; the second-level response is activated when three or more adjacent monitoring points exceed the limit at the same time and the deformation direction shows mechanical correlation, generating a recommendation for equipment load reduction operation and starting the backup sensor for cross-validation to confirm the accuracy of the original monitoring data and eliminate single-point errors; the third-level response is activated when the probability of overall structural instability exceeds 90% and the consistency between the drone re-measurement data and the edge computing results is less than 80%, sending an encrypted shutdown command to the central control system and marking the coordinates of the dangerous area. At the same time, the central control system automatically marks the coordinates of the dangerous area corresponding to the current monitoring result; Increase monitoring frequency when local deformation exceeds the limit; generate equipment load adjustment instructions when multiple related areas exceed the limit; and execute automatic protection shutdown when the overall instability risk exceeds the critical value; The specific method for visualizing and tracing the multi-dimensional monitoring data is to superimpose the three-dimensional point cloud data and sensor monitoring values ​​on the surface of the factory BIM model, characterize the deformation distribution through color gradient mapping, and construct a multi-dimensional correlation map to show the temporal and spatial correlation between the structural deformation trend and the equipment vibration parameters and ambient temperature.

6. The method for monitoring deformation of a hydropower plant building according to claim 5, characterized in that: In step 1, while collecting multi-dimensional monitoring data, temperature and humidity sensors and vibration accelerometers are deployed to collect environmental interference parameters; in step 2, the noise reduction process dynamically adjusts the filter parameters according to the spectral characteristics of the environmental noise; in step 3, the dynamic warning threshold model updates the benchmark value in real time through the sliding window statistical method; in step 4, the sensor reliability evaluation index is obtained by comprehensively calculating the data continuity rate, noise variance and consistency error with adjacent sensors, and the weighted sum of the mean square error of the prediction model and the warning false alarm rate is minimized by the gradient descent algorithm. The weighted average algorithm is used to optimize the multi-source data fusion weight. The optimization goal is to minimize the weighted sum of the mean square error of the prediction model and the warning false alarm rate. The particle swarm optimization algorithm is used regularly to adjust the decision boundary of the classification model. The warning records of the last 30 days are used as the training set in the adjustment process, and the reduction of the false alarm rate and the stability of the missed alarm rate are used as convergence conditions.

7. The method for monitoring deformation of a hydropower plant building according to claim 2, characterized in that: In step 1, the offset of the image acquisition position is calculated when the drone performs periodic three-dimensional scanning: in, v is the flight speed of the UAV, in m / s; is the image exposure time interval, unit s; is the offset, unit is m.

8. The method for monitoring deformation of a hydropower plant building according to claim 4, characterized in that: The temperature-stress coupling relationship function of the concrete structure is: in, Temperature T The stress under the pressure, unit is Pa; Temperature T The elastic modulus of the material under , in Pa; Temperature T strain under Through the stress-strain experimental data at different temperatures, the elastic modulus of the material at each temperature point is obtained and the corresponding strain , through regression analysis or fitting method, the relationship between temperature, stress and strain is obtained, and the elastic modulus of the material is dynamically corrected based on real-time temperature data , and then calculate the corresponding stress value to determine the deformation under temperature change; The dynamic adjustment range of the allowable deformation is: in, For the temperature T Maximum deformation under is the corresponding maximum stress, unit: Pa; It is the elastic modulus after temperature correction, in Pa.

9. A hydropower plant building deformation monitoring system, used to implement the hydropower plant building deformation monitoring method according to any one of claims 1 to 8, characterized in that: It includes a multi-source data acquisition module, an edge data processing module, an analysis and decision module, and a system self-optimization module, which are connected in sequence through industrial Ethernet. The analysis and decision module is connected to a hierarchical control execution module, and the system self-optimization module is connected to the edge data processing module and the analysis and decision module through a control bus.

10. The hydropower plant building deformation monitoring system according to claim 9, characterized in that: The multi-source data acquisition module includes a distributed sensor network and a drone equipped with a multi-lens camera. The distributed sensor network includes strain sensors, inertial measurement units, temperature and humidity sensors, and vibration accelerometers deployed in key stress concentration areas of the factory building. The strain sensors, inertial measurement units, and drones are respectively connected to the edge data processing module through wireless communication modules. The edge data processing module includes an adaptive filtering unit and a spatiotemporal registration unit. The adaptive filtering unit receives time series data transmitted by the sensor network and dynamically adjusts the filtering parameters according to the spectral characteristics of the ambient noise to eliminate interference signals. The spatiotemporal registration unit has a built-in factory building BIM model database, aligns the photogrammetric point cloud data with the sensor data in a unified three-dimensional coordinate system through a feature point matching algorithm, and uses an iterative optimization algorithm to correct coordinate conversion errors. The output end of the spatiotemporal registration unit is connected to the data analysis module. The analysis and decision-making module includes a multi-channel time-series deep learning model and a dynamic threshold generation unit. The input end of the deep learning model receives the fused strain, displacement, acceleration and environmental parameter data. Its hidden layer deploys a bidirectional recurrent neural network with an attention mechanism to extract spatiotemporal correlation features. The output end is connected to the risk classification unit. The dynamic threshold generation unit integrates a material property database and an equipment operation parameter interface, updates the warning baseline value in real time through a sliding window statistical method, and transmits the threshold data to the hierarchical control module. The hierarchical control execution module includes a three-level response controller, a command generator, and a central control system connected in sequence via an industrial bus. The input end of the three-level response controller receives the risk assessment result of the analysis and decision-making module. When local deformation exceeds the limit, the first-level response command is triggered, and the data acquisition module is controlled to increase the monitoring frequency of the corresponding area. When multiple related areas exceed the limit in a coordinated manner, the second-level response command is triggered, and the equipment load adjustment signal is generated and sent to the central control system. When the overall instability risk exceeds the limit, the third-level response command is triggered, and the encrypted shutdown control command is executed; The system self-optimization module includes a sensor reliability assessment unit and a model parameter adjustment unit. The reliability assessment unit calculates the data continuity rate, noise level and consistency error with adjacent sensors of each sensor in real time, generates a weight coefficient and feeds it back to the edge data processing module; the model parameter adjustment unit dynamically optimizes the data fusion weight and classification model decision boundary based on the early warning record database through a machine learning algorithm, and its output end is connected to the model update interface of the analysis and decision module.

Images