Multi-sensor safety monitoring system for pumped storage power station

Abstract

The present invention relates to the technical field of spatio-temporal synchronous measurement of multi-sensor monitoring networks in pumped storage power stations, and particularly to a multi-sensor safety monitoring system for pumped storage power stations, including a sensor data acquisition module, a spatio-temporal synchronous network module, a data calibration module, a dynamic correlation model construction module, a multi-dimensional data fusion module, and a hierarchical early warning module. The dynamic correlation model construction module simulates the dynamic responses of seepage pressure and surrounding rock strain through a fluid-structure interaction equation and a discrete element digital twin model, updates the fracture permeability boundary conditions in real time, extracts cross-sensor spatio-temporal correlation features using a convolutional long short-term memory network improved by a multi-head attention mechanism, and suppresses the interference of low-confidence data with a dynamic weighted fusion coefficient. The present invention improves the real-time early warning accuracy of abnormal events and the reliability of structural health assessment in complex hydraulic and mechanical coupling environments.

Description

Technical Field

[0001] The present invention relates to the technical field of spatio-temporal synchronous measurement of multi-sensor monitoring networks in pumped storage power stations, and particularly to a multi-sensor safety monitoring system for pumped storage power stations. Background Art

[0002] In the existing multi-sensor safety monitoring system for pumped storage power stations, dynamic acquisition and analysis of key structural parameters are achieved by integrating a variety of high-precision sensors. This system uses devices such as strain sensors, displacement gauges, and piezometers to monitor key parts such as the surrounding rock mass of the underground powerhouse, penstock, dam, and water conveyance tunnel in real time, and combines temperature sensors and vibration sensors to obtain environmental variable and mechanical operation state data. Through the fusion processing of multiple data, a correlation model between structural response and load action is established, and pattern recognition algorithms are used to identify abnormal signals, thereby triggering a hierarchical early warning mechanism. The core of its technology lies in eliminating the uncertainty of single-sensor data through spatio-temporal correlation analysis, improving the reliability of monitoring results, providing a quantitative basis for structural health assessment and risk prevention and control during the entire life cycle of the power station, and effectively reducing potential safety hazards caused by material aging, hydraulic impact, or geological activities.

[0003] In the existing multi-sensor safety monitoring system for pumped storage power stations, the spatio-temporal synchronization and correlation of sensor data are insufficient, resulting in limited identification accuracy of the coupling mechanism of abnormal events. In a complex hydro-mechanical coupling environment, sensors of different physical quantities generate time series deviations and spatial calibration errors due to differences in acquisition frequencies or transmission delays, thereby weakening the synergistic effect of data fusion. For example, the asynchronous acquisition of the deformation of the surrounding rock mass of a pumped storage power station and the seepage flow data of the adjacent water conveyance tunnel masks the dynamic interaction characteristics between the two, and cannot accurately reflect the correlation between hydraulic fracturing and rock mass displacement, reducing the timeliness of early warning of structural instability precursors. Summary of the Invention

[0004] Aiming at the deficiencies of the existing technology, the present invention provides a multi-sensor safety monitoring system for pumped storage power stations, and the present invention solves the problem of deviation in the identification of the coupling mechanism of abnormal events caused by insufficient spatio-temporal synchronization accuracy of the multi-sensor monitoring network in a pumped storage power station under a complex hydro-mechanical coupling environment.

[0005] To solve the above technical problems, the specific technical solutions of the present invention are as follows:

[0006] The multi-sensor safety monitoring system for pumped storage power stations provided by the present invention includes: a sensor data acquisition module, a spatio-temporal synchronization network module, a data calibration module, a dynamic correlation model construction module, a multi-dimensional data fusion module, and a hierarchical early warning module;

[0007] The sensor data acquisition module is used to collect stress, seepage pressure, displacement and vibration signals of key monitoring points through redundant sensor nodes, and perform noise suppression processing on the collected signals to generate preprocessed data; the spatio-temporal synchronization network module is used to receive the preprocessed data output by the sensor data acquisition module, achieve microsecond-level time alignment of sensor nodes through a hierarchical clock synchronization architecture, and dynamically adjust the data transmission timing based on cluster-based communication optimization to output synchronized timing data;

[0008] The data calibration module is used to receive the synchronized timing data output by the spatio-temporal synchronization network module, align asynchronous data through a transmission delay compensation algorithm, and calibrate the global coordinate system of the sensor nodes in combination with a three-dimensional laser scanner to generate spatio-temporal aligned data;

[0009] The dynamic correlation model construction module is used to input the spatio-temporal aligned data output by the data calibration module into a fluid-structure interaction model, drive a digital twin model based on the discrete element method to simulate the dynamic responses of seepage pressure and surrounding rock strain, and output simulation verification data;

[0010] The multi-dimensional data fusion module is used to extract the spatio-temporal correlation features of the spatio-temporal aligned data and the simulation verification data, and generate fusion decision data through a weighted fusion algorithm;

[0011] The hierarchical early warning module is used to trigger multi-level early warning thresholds according to the fusion decision data, and iteratively optimize the clock synchronization parameters of the spatio-temporal synchronization network module and the fluid-structure interaction model coefficients of the dynamic correlation model construction module based on the fusion decision data.

[0012] Furthermore, for the multi-sensor safety monitoring system of the pumped storage power station of the present invention, the sensor data acquisition module includes:

[0013] A redundant sensor node group, an abnormal data rejection unit and an adaptive noise suppression unit, where:

[0014] The redundant sensor node group adopts a triple redundancy configuration and is deployed in the stress concentration area of the penstock. Three sensors of the same type are set at each monitoring point to form a data acquisition unit, and the three sensors synchronously collect signals of the same physical quantity;

[0015] The abnormal data rejection unit is connected to the redundant sensor node group and rejects abnormal data through a majority voting algorithm within a sliding time window. Specifically: normalize the data collected by the three sensors within the sliding time window, calculate the Euclidean distance between pairwise data. If the distance of a certain data from the other two data exceeds the dynamic threshold, it is determined as abnormal data and rejected, and the mean value of the remaining data is used as the effective output;

[0016] The adaptive noise suppression unit is connected to the abnormal data rejection unit and uses an adaptive wavelet threshold filtering algorithm to denoise the effective output signal, including: dynamically selecting a wavelet basis function based on the signal spectrum analysis result, separating the electromagnetic interference noise frequency band through an improved SUREShrink threshold estimation, and retaining the time-domain waveform characteristics and frequency-domain harmonic characteristics when reconstructing the signal, generating preprocessed data and outputting it to the spatio-temporal synchronization network module.

[0017] Further, in the multi-sensor safety monitoring system of the pumped storage power station according to the present invention, the spatio-temporal synchronization network module includes a clock synchronization sub-module and a cluster communication optimization sub-module, wherein:

[0018] The clock synchronization sub-module is used to receive the preprocessed data output by the sensor data acquisition module, including:

[0019] The master node obtains a reference clock signal from the Beidou satellite system through the satellite time service protocol and broadcasts the reference clock to all slave nodes;

[0020] Each slave node establishes a time synchronization link with the master node through the precision clock synchronization protocol, and at the same time uses a linear regression model to predict the cumulative drift of the crystal oscillator of each slave node in real time, and dynamically corrects the time stamp of the sensor data according to the drift amount for microsecond-level time alignment between sensor nodes;

[0021] The cluster communication optimization sub-module is connected to the clock synchronization sub-module and includes:

[0022] Dividing the monitoring area into multiple geological unit subnets according to the geological structure characteristics, and forming independent communication clusters for the sensor nodes within each subnet;

[0023] Within the subnet, a fixed data transmission time slot is allocated to each node through the time division multiple access mechanism. The master node monitors the channel load intensity in real time. When the data volume of the nodes within the subnet exceeds a preset threshold, a dynamic time slice adjustment algorithm is used to shorten the length of a single time slot and increase the time slot density to optimize the data transmission timing;

[0024] Output the data that has been time-aligned and timing-optimized as synchronous timing data and transmit it to the data calibration module.

[0025] Further, in the multi-sensor safety monitoring system of the pumped storage power station according to the present invention, the data calibration module includes a time delay compensation sub-module and a space calibration sub-module, wherein:

[0026] The time delay compensation sub-module is used to receive the synchronous timing data output by the spatio-temporal synchronization network module, including:

[0027] The two-way timestamp protocol is used to measure the round-trip transmission link delay between the master node and the slave node. The delay fluctuation curve is fitted by the least squares method to generate the benchmark transmission delay amount.

[0028] Based on the Kalman filter algorithm, the random delay deviation caused by network jitter is predicted. An interpolation function is constructed by combining the benchmark transmission delay amount to resample the sensor data arriving asynchronously on the time axis for microsecond-level time alignment.

[0029] The space calibration sub-module is connected to the time delay compensation sub-module and includes:

[0030] The three-dimensional point cloud data of the monitoring area is obtained by a three-dimensional laser scanner, and the actual point cloud coordinates of the installation positions of the sensor nodes are extracted.

[0031] The iterative closest point algorithm with normal vector constraint is used to register the actual point cloud coordinates with the theoretical coordinates in the design drawing, and the coordinate offset of the sensor nodes is calculated.

[0032] Based on the quaternion rotation matrix, the local coordinate systems of the sensor nodes are converted to the global engineering coordinate system, and the spatio-temporal alignment data is output to the dynamic correlation model construction module.

[0033] Furthermore, for the multi-sensor safety monitoring system of the pumped-storage power station of the present invention, the dynamic correlation model construction module includes a fluid-structure interaction solution sub-module and a digital twin driving sub-module, where:

[0034] The fluid-structure interaction solution sub-module is used to receive the spatio-temporal alignment data output by the data calibration module, including: based on the coupled partial differential equation of the seepage pressure gradient tensor and the surrounding rock strain tensor, the dynamic correlation relationship between the seepage field and the stress field is calculated by an implicit iterative solver to generate the coupled response data of seepage pressure and strain.

[0035] The digital twin driving sub-module is connected to the fluid-structure interaction solution sub-module and includes: a three-dimensional digital twin model of the surrounding rock fracture propagation is constructed by the discrete element method, the real-time seepage pressure data is mapped as the permeability boundary condition of the fracture surface, and the non-linear relationship between permeability and strain during the fracture propagation process is simulated by the dynamic relaxation algorithm.

[0036] The fracture network topology of the digital twin model is updated according to the coupled response data of seepage pressure and strain, driving the dynamic evolution of the digital twin model, and outputting the simulation verification data to the multi-dimensional data fusion module.

[0037] Furthermore, for the multi-sensor safety monitoring system of the pumped-storage power station of the present invention, the multi-dimensional data fusion module includes a spatio-temporal feature extraction sub-module and a dynamic weighted fusion sub-module, where:

[0038] The spatio-temporal feature extraction sub-module is used to receive the simulation verification data output by the dynamic association model construction module and the spatio-temporal alignment data output by the data calibration module, including:

[0039] A convolutional long short-term memory network improved by a multi-head attention mechanism, which inputs the time series data of seepage pressure and the spatial distribution data of displacement rate, extracts local deformation features through spatio-temporal convolutional kernels, and uses the multi-head attention mechanism to capture the correlation between pressure and displacement across sensor nodes, generating a spatio-temporal correlation feature matrix;

[0040] The dynamic weighted fusion sub-module is connected to the spatio-temporal feature extraction sub-module, including: calculating the model confidence weight according to the error rate between the simulation verification data and the measured data, and calculating the data reliability weight based on the historical stability index of the sensor node;

[0041] Performing a normalized product operation on the model confidence weight and the data reliability weight to generate the dynamic weighted fusion coefficient of each sensor node;

[0042] Performing multi-dimensional fusion on the spatio-temporal correlation feature matrix through a weighted decision function, and outputting the fusion decision data to the hierarchical early warning module.

[0043] Furthermore, for the multi-sensor safety monitoring system of the pumped storage power station of the present invention, the hierarchical early warning module includes an early warning trigger sub-module and a closed-loop parameter optimization sub-module, where:

[0044] The early warning trigger sub-module is used to receive the fusion decision data output by the multi-dimensional data fusion module, including:

[0045] Dividing the three-level early warning thresholds of seepage erosion, crack expansion and rock mass instability based on the structural safety margin model, where:

[0046] The seepage erosion threshold is determined by calculating the ratio of the seepage velocity to the critical permeability through Darcy's law;

[0047] The crack expansion threshold is based on calculating the stress intensity factor at the crack tip using the fracture mechanics J-integral model;

[0048] The rock mass instability threshold determines the safety factor of the surrounding rock shear slip surface according to the Mohr-Coulomb criterion;

[0049] Real-time comparison of the fusion decision data with the three-level thresholds, and when the continuous over-limit duration of the data exceeds the preset tolerance, triggering the corresponding level of audible and visual alarm signals;

[0050] The closed-loop parameter optimization sub-module is connected to the early warning trigger sub-module, including: using the clock synchronization error rate of the spatio-temporal synchronization network module and the root mean square of the prediction residuals of the dynamic association model construction module as the fitness function, and performing multi-objective optimization using the genetic algorithm;

[0051] Elite individuals are screened from the population through the tournament selection strategy. The crossover operator adopts arithmetic recombination, and the mutation operator introduces Gaussian perturbation to iteratively search for the optimal combination of clock synchronization parameters and fluid-structure interaction model coefficients.

[0052] The optimized parameters are fed back in real time to the clock synchronization sub-module of the spatio-temporal synchronization network module and the digital twin driving sub-module of the dynamic correlation model construction module.

[0053] Advantages of the present invention;

[0054] Through the triple-redundancy configuration of the redundant sensor node group combined with the sliding time window majority voting algorithm, the present invention effectively eliminates abnormal data caused by single sensor failure or environmental interference, and improves the reliability of multi-source heterogeneous sensor data acquisition; the spatio-temporal synchronization network module adopts a hierarchical clock synchronization architecture and a cluster-based communication optimization mechanism, predicts the crystal oscillator drift through the satellite timing protocol and the linear regression model, realizes microsecond-level time alignment and dynamically adjusts the data transmission timing, and solves the multi-node communication congestion problem in complex geological environments; the data calibration module compensates for the transmission delay based on the two-way timestamp protocol and the Kalman filtering algorithm, and completes the global coordinate system calibration by combining the normal vector constraint iterative closest point algorithm of the three-dimensional laser scanner, eliminating the influence of spatio-temporal reference deviation on data fusion; the dynamic correlation model construction module realizes the real-time update of the fracture permeability boundary conditions and simulates the non-linear correlation between seepage pressure and surrounding rock strain through the co-simulation of the fluid-structure interaction equation and the discrete element digital twin model, enhancing the analytical ability of the hydraulic and mechanical coupling mechanism; the multi-dimensional data fusion module uses a spatio-temporal convolutional network improved by the multi-head attention mechanism to extract cross-sensor correlation features, combines the dynamic weighted fusion algorithm to suppress the interference of low-confidence data, and optimizes the identification accuracy of abnormal event coupling features; the hierarchical early warning module establishes a three-level threshold judgment rule based on the structural safety margin model, and realizes the closed-loop optimization iteration of the clock synchronization parameters and model coefficients through the genetic algorithm, improving the adaptability and early warning timeliness of the system under complex working conditions. Description of the drawings

[0055] In order to more clearly illustrate the technical solutions of the present invention, the drawings required for use in the embodiments will be briefly introduced below. Obviously, for those of ordinary skill in the art, other drawings can be obtained according to the drawings without creative efforts.

[0056] Figure 1 It is the system architecture diagram of the multi-sensor safety monitoring system for pumped storage power stations provided by the embodiments of the present invention. Detailed implementation manners

[0057] To make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with specific embodiments of the present invention and the corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention. The technical solutions provided by each embodiment of the present invention are described in detail below in conjunction with the drawings. To better understand the objectives of the present invention, the present invention is further described in detail below.

[0058] Please refer to Figure 1 , the multi-sensor safety monitoring system for pumped storage power stations provided by the present invention includes: a sensor data acquisition module, a spatio-temporal synchronization network module, a data calibration module, a dynamic correlation model construction module, a multi-dimensional data fusion module, and a hierarchical early warning module;

[0059] The sensor data acquisition module is used to collect stress, seepage pressure, displacement, and vibration signals of key monitoring points through redundant sensor nodes, and perform noise suppression processing on the collected signals to generate preprocessed data;

[0060] The spatio-temporal synchronization network module is used to receive the preprocessed data, achieve microsecond-level time alignment of sensor nodes through a hierarchical clock synchronization architecture, and dynamically adjust the data transmission timing through cluster-based communication optimization to output synchronized timing data;

[0061] The data calibration module is used to receive the synchronized timing data, align asynchronous data through a transmission delay compensation algorithm, and calibrate the global coordinate system of sensor nodes through a three-dimensional laser scanner to generate spatio-temporal aligned data;

[0062] The dynamic correlation model construction module is used to input the spatio-temporal aligned data into a fluid-structure interaction model, drive a digital twin model based on the discrete element method to simulate the dynamic responses of seepage pressure and surrounding rock strain, and output simulation verification data;

[0063] The multi-dimensional data fusion module is used to extract the spatio-temporal correlation features of the spatio-temporal aligned data and the simulation verification data, and generate fusion decision data through a weighted fusion algorithm;

[0064] The hierarchical early warning module is used to trigger multi-level early warning thresholds according to the fusion decision data, and adjust the clock synchronization parameters of the spatio-temporal synchronization network module and the model coefficients of the dynamic correlation model construction module through a parameter optimization algorithm.

[0065] The sensor data acquisition module synchronously acquires stress, seepage pressure, displacement, and vibration signals at key monitoring points of the pumped-storage power station through redundant sensor nodes. The redundant sensor nodes adopt a triple-redundancy configuration and are deployed in the stress concentration area of the penstock and the sensitive area of surrounding rock deformation. Three sensors of the same type are set at each monitoring point to form a data acquisition unit, and abnormal data is eliminated through the majority voting algorithm within a sliding time window. After eliminating the abnormal data, the adaptive wavelet threshold filtering algorithm is used to denoise the effective signals. Based on the signal spectrum analysis results, the wavelet basis function is dynamically selected to separate the electromagnetic interference noise frequency band, retain the time-domain waveform characteristics and frequency-domain harmonic characteristics, and generate preprocessed data.

[0066] After receiving the preprocessed data, the spatio-temporal synchronization network module realizes the microsecond-level time alignment of sensor nodes through a hierarchical clock synchronization architecture. The master node obtains the reference clock signal through the satellite time synchronization protocol, and the slave nodes are synchronized to the master node through the precision clock synchronization protocol. The crystal oscillator drift is predicted using a linear regression model to correct the timestamps. The cluster-based communication optimization divides the monitoring subnet according to geological units. The time-division multiple access mechanism is used to allocate data transmission time slots within the subnet. The master node dynamically adjusts the time slice length according to the channel load to optimize the data transmission timing and outputs synchronized timing data.

[0067] After receiving the synchronized timing data, the data calibration module aligns the asynchronous data through the transmission delay compensation algorithm. The two-way timestamp protocol is used to measure the link delay, and the Kalman filter is used to predict the network jitter and generate an interpolation function for time-axis resampling. On this basis, the point cloud data of the monitoring area is obtained through a three-dimensional laser scanner. The iterative closest point algorithm with normal vector constraint is used to register the actual coordinates and design coordinates of the sensor nodes, and a quaternion rotation matrix is constructed to transform the local coordinate system to the global engineering coordinate system, generating spatio-temporal alignment data.

[0068] The dynamic correlation model construction module inputs the spatio-temporal alignment data into the fluid-structure interaction model and calculates the dynamic correlation equation of the seepage pressure gradient and the surrounding rock strain tensor through an implicit iterative solver. The digital twin model simulates the process of surrounding rock fracture propagation based on the discrete element method, maps the real-time seepage pressure data to the fracture surface permeability boundary condition, drives the dynamic evolution of the model, generates the coupled response data of seepage and strain, and outputs the simulation verification data.

[0069] The multi-dimensional data fusion module extracts the spatio-temporal correlation features of the spatio-temporal alignment data and the simulation verification data. The convolutional long short-term memory network improved by the multi-head attention mechanism is used to capture the cross-sensor correlation of the seepage pressure mutation and the displacement rate. The weighted fusion coefficient is generated according to the simulation error rate and the historical stability index of the sensor. Through the normalized product operation, the model confidence and the data reliability weight are combined to perform multi-dimensional fusion on the spatio-temporal feature matrix and generate the fusion decision data.

[0070] The hierarchical early warning module triggers multi-level early warning thresholds based on the fusion decision data. The seepage erosion threshold is determined by calculating the ratio of seepage velocity to critical permeability according to Darcy's law. The crack propagation threshold is based on the fracture mechanics J-integral model to determine the stress intensity factor at the crack tip. The rock mass instability threshold evaluates the safety factor of the surrounding rock shear slip surface according to the Mohr-Coulomb criterion. The parameter optimization algorithm uses the spatio-temporal synchronization error and the model prediction residual as the fitness function, and adopts a genetic algorithm combined with a tournament selection strategy to search for the optimal parameter combination, and feeds the optimized clock synchronization parameters and model coefficients back to the spatio-temporal synchronization network module and the dynamic correlation model construction module to form a closed-loop optimization link.

[0071] Specifically, for the multi-sensor safety monitoring system of the pumped-storage power station described in the present invention, the sensor data acquisition module includes:

[0072] A redundant sensor node group, an abnormal data rejection unit, and an adaptive noise suppression unit, where: the redundant sensor node group adopts a triple redundancy configuration and is deployed in the stress concentration area of the penstock. Three sensors of the same type are set at each monitoring point to form a data acquisition unit, and the three sensors synchronously collect the signals of the same physical quantity;

[0073] The abnormal data rejection unit is connected to the redundant sensor node group and rejects abnormal data through the majority voting algorithm within a sliding time window. Specifically: normalize the data collected by the three sensors within the sliding time window, calculate the Euclidean distance between two data, if the distance of a certain data from the other two data exceeds the dynamic threshold, it is determined as abnormal data and rejected, and the mean value of the remaining data is used as the effective output;

[0074] The adaptive noise suppression unit is connected to the abnormal data rejection unit and uses the adaptive wavelet threshold filtering algorithm to denoise the effective output signal, including: dynamically selecting the wavelet basis function based on the signal spectrum analysis result, separating the electromagnetic interference noise frequency band through the improved SUREShrink threshold estimation, and retaining the time-domain waveform characteristics and frequency-domain harmonic characteristics when reconstructing the signal, generating preprocessed data and outputting it to the spatio-temporal synchronization network module.

[0075] The redundant sensor node group of the sensor data acquisition module adopts a triple redundancy configuration and is deployed in the stress concentration area of the penstock. Three sensors of the same type, including strain sensors, piezometric sensors, and vibration sensors, are set at each monitoring point to form a data acquisition unit, and the synchronous acquisition of physical quantity signals is realized through a synchronous trigger circuit. The deployment of the redundant sensor node group in the stress concentration area covers the weld area and the high-curvature area of the penstock, and is densely arranged for the structurally deformed sensitive parts. The triple redundancy design improves data reliability through cross-validation.

[0076] The abnormal data rejection unit processes the original acquisition data of the redundant sensor node group through the majority voting algorithm within a sliding time window. The length of the sliding time window is dynamically adjusted according to the sensor sampling frequency and the change rate of the physical quantity. The three-way sensor data within the window is normalized to eliminate the dimension difference. After normalization, the Euclidean distance between pairwise data is calculated. The dynamic threshold is generated based on the statistical characteristics of historical data. If the Euclidean distance of a certain data continuously exceeds the threshold, it is determined as abnormal data and rejected. The remaining valid data is averaged as the output. This mechanism suppresses abnormal data caused by single sensor failure or environmental interference through multi-dimensional data comparison.

[0077] The adaptive noise suppression unit performs noise reduction processing on the effective signal after abnormal data rejection. Based on the signal spectrum analysis results, the wavelet basis function matching the main frequency of the signal is adaptively selected, such as the Daubechies wavelet or the Symlet wavelet. The electromagnetic interference noise frequency band is separated through an improved SUREShrink threshold estimation method. A frequency band energy weight factor is introduced during the threshold estimation process to preferentially suppress high-frequency noise components. At the same time, the mutation characteristics of the time-domain waveform and the frequency-domain harmonic components are retained during the signal reconstruction stage. The noise-reduced signal undergoes joint time-frequency domain verification to generate preprocessed data that meets the requirements of subsequent processing and is output to the spatio-temporal synchronization network module.

[0078] The data acquisition of the redundant sensor node group, the majority voting mechanism of the abnormal data rejection unit, and the frequency-domain noise reduction processing of the adaptive noise suppression unit form a series link. The synchronous acquisition of three redundant sensors provides a basis for data consistency. The sliding time window algorithm filters out abnormal values through a dynamic threshold, and the adaptive wavelet filtering further eliminates noise interference, finally generating preprocessed data with high confidence. The data transmission between units is realized through a standardized interface protocol to ensure the consistency of data format and processing timing.

[0079] Specifically, the multi-sensor safety monitoring system of the pumped storage power station of the present invention includes a clock synchronization sub-module and a cluster communication optimization sub-module, where:

[0080] The clock synchronization sub-module is used to receive the preprocessed data of the sensor data acquisition module, including:

[0081] The master node obtains the reference clock signal from the Beidou satellite system through the satellite time service protocol and broadcasts the reference clock to all slave nodes;

[0082] Each slave node establishes a time synchronization link with the master node through the precision clock synchronization protocol. At the same time, a linear regression model is used to predict the cumulative drift of the crystal oscillator of each slave node in real time, and the timestamp of the sensor data is dynamically corrected according to the drift for microsecond-level time alignment between sensor nodes;

[0083] The cluster communication optimization sub-module is connected to the clock synchronization sub-module and includes:

[0084] Divide the monitoring area into multiple geological unit subnets according to geological structure characteristics, and sensor nodes within each subnet form an independent communication cluster;

[0085] Within the subnet, fixed data transmission time slots are allocated to each node through the time division multiple access mechanism. The master node monitors the channel load intensity in real time. When the data volume of the nodes within the subnet exceeds the preset threshold, the dynamic time slice adjustment algorithm is adopted to shorten the length of a single time slot and increase the time slot density, optimizing the data transmission timing sequence;

[0086] Output the data after time alignment and timing optimization as synchronous timing data and transmit it to the data calibration module.

[0087] The clock synchronization sub-module of the spatio-temporal synchronization network module establishes a global time reference through the satellite timing protocol. The master node accesses the navigation message signal of the Beidou satellite system, parses the coordinated universal time (UTC) as the reference clock source, and distributes the reference time information to all slave nodes through the low-latency broadcast protocol. The slave nodes establish a two-way communication link with the master node using the precision clock synchronization protocol, periodically exchange time synchronization messages, measure the transmission path delay, and compensate for the network asymmetry error. The slave nodes are built with high-precision temperature-compensated crystal oscillators, analyze the historical clock deviation data through a linear regression model, predict the crystal oscillator frequency drift trend, and dynamically correct the local clock counter value to achieve microsecond-level alignment of the sensor data timestamps. During the time synchronization process, the master node continuously monitors the clock offset of the slave nodes. When the offset exceeds the tolerance range, the resynchronization mechanism is triggered to maintain the time consistency of the entire network.

[0088] The cluster communication optimization sub-module divides the monitoring area subnets based on geological structure characteristics. According to the pressure steel pipe trend, surrounding rock fault distribution, and underground powerhouse structure form, the monitoring area is divided into multiple geological unit subnets, and each subnet corresponds to an independent communication cluster. Sensor nodes within the subnet elect a cluster head node through the self-organizing network protocol, and the cluster head node is responsible for coordinating the time slot allocation under the time division multiple access mechanism. The fixed time slot length is preset according to the sensor sampling rate and data packet size. The master node collects the channel load intensity indicators of each subnet in real time. When the concurrent data volume within the subnet exceeds the preset threshold, the dynamic time slice adjustment algorithm is started. This algorithm predicts the data transmission queue length based on the queuing theory model, compresses the duration of a single time slot proportionally, and increases the number of time slots within a unit period to improve the channel utilization rate. The optimized timing parameters are sent to the subnet cluster head node through control messages to synchronously update the time slot mapping table of each node, realizing the dynamic reconstruction of the data transmission timing sequence.

[0089] The clock synchronization sub-module and the cluster communication optimization sub-module work together, and the sensor data after timestamp correction is sent to the master node according to the optimized timing sequence. After the master node integrates the synchronous timing data of each subnet, it adds a time reference tag and a subnet identifier according to the standard data encapsulation format to generate a synchronous timing data stream with spatio-temporal reference information, which is transmitted to the data calibration module for subsequent processing. The time alignment mechanism eliminates the clock deviation of multi-source sensor data, and the cluster communication optimization reduces the probability of channel conflict. The two jointly ensure the real-time performance and integrity of high-density monitoring data.

[0090] Specifically, for the multi-sensor safety monitoring system of the pumped storage power station described in the present invention, the data calibration module includes:

[0091] A time delay compensation sub-module and a space calibration sub-module, where:

[0092] The time delay compensation sub-module is used to receive the synchronous timing data output by the spatio-temporal synchronization network module, including:

[0093] Measure the round-trip transmission link delay between the master node and the slave node using the two-way timestamp protocol, fit the delay fluctuation curve by the least squares method, and generate a reference transmission delay amount;

[0094] Predict the random delay deviation caused by network jitter based on the Kalman filter algorithm, construct an interpolation function in combination with the reference transmission delay amount, and perform time-axis resampling on the asynchronously arriving sensor data for microsecond-level time alignment;

[0095] The space calibration sub-module is connected to the time delay compensation sub-module and includes:

[0096] Obtain the three-dimensional point cloud data of the monitoring area with a resolution of 0.1 mm by a three-dimensional laser scanner, and extract the actual point cloud coordinates of the installation positions of the sensor nodes;

[0097] Adopt the iterative closest point algorithm with normal vector constraint to register the actual point cloud coordinates with the theoretical coordinates in the design drawing, and calculate the coordinate offset of the sensor nodes;

[0098] Based on the quaternion rotation matrix, convert the local coordinate systems of each sensor node to the global engineering coordinate system, and output the spatio-temporal alignment data to the dynamic association model construction module.

[0099] The time delay compensation sub-module of the data calibration module receives the synchronous timing data output by the spatio-temporal synchronization network module. The two-way transmission link delay is measured between the master node and the slave node through the two-way timestamp protocol. The protocol interaction process records the sending and receiving timestamps, calculates the round-trip time difference of the path, and eliminates the one-way transmission asymmetry error. Based on the continuous measurement data set, the least squares method is used to fit the delay fluctuation curve, extract the reference delay component of the transmission link, and eliminate the influence of the fixed propagation delay on time synchronization. For the random delay fluctuation caused by network jitter, a delay state prediction model is established through the Kalman filter algorithm, and a time axis interpolation function is generated in combination with the reference delay amount to resample the asynchronously arriving sensor data, aligning the multi-source data to a unified time reference to achieve a time synchronization accuracy of microseconds.

[0100] The space calibration sub-module performs a space coordinate system transformation on the time-calibrated data. The three-dimensional laser scanner performs a panoramic scan of the monitoring area along a predetermined scan path to generate high-precision three-dimensional point cloud data including the installation positions of the sensor nodes. In the point cloud data preprocessing stage, environmental noise points are removed, and the point cloud cluster where the sensor node is located is segmented through the region growing algorithm to extract the actual three-dimensional coordinates of the node installation base. The iterative closest point algorithm with normal vector constraint is used for point cloud registration. During the algorithm iteration process, the normal vector direction of the sensor installation surface is introduced as a rigid transformation constraint condition, and the least squares matching is performed between the actual point cloud coordinates and the theoretical coordinates in the design drawing to calculate the coordinate offset and rotation angle. A transformation model from the local coordinate system to the global coordinate system is constructed based on the quaternion rotation matrix to eliminate the measurement error caused by the deviation of the sensor installation position and output the calibrated data with a unified spatio-temporal reference.

[0101] The time delay compensation sub-module and the space calibration sub-module form a spatio-temporal joint calibration link. The data after time axis resampling carries accurate time tags, and the space coordinate transformation model maps the local measurement values to the global engineering coordinate system. The two work together to achieve the spatio-temporal consistency calibration of multi-source heterogeneous sensor data. The calibrated data integrates the triple information of the timestamp, space coordinates, and physical quantity measurement values, providing an input data source with spatio-temporal traceability for the dynamic association model. The time interpolation and space transformation parameters used in the calibration process are recorded through the calibration log to support subsequent data traceability and model parameter inversion verification.

[0102] Specifically, for the multi-sensor safety monitoring system of the pumped storage power station described in the present invention, the dynamic association model construction module includes:

[0103] A fluid-structure interaction solution sub-module and a digital twin drive sub-module, where:

[0104] The fluid-structure interaction solution sub-module is used to receive the spatio-temporal alignment data output by the data calibration module, including:

[0105] Based on the coupled partial differential equations of the seepage pressure gradient tensor and the surrounding rock strain tensor, the dynamic correlation between the seepage field and the stress field is calculated by an implicit iterative solver to generate seepage pressure and strain coupling response data;

[0106] The digital twin driving sub-module is connected to the fluid-structure interaction solving sub-module, including:

[0107] A three-dimensional digital twin model of surrounding rock fracture propagation is constructed using the discrete element method. The real-time seepage pressure data is mapped as the permeability boundary condition of the fracture surface, and the non-linear relationship between permeability and strain during fracture propagation is simulated by the dynamic relaxation algorithm;

[0108] According to the seepage pressure and strain coupling response data, the fracture network topology of the digital twin model is updated, driving the digital twin model to dynamically evolve with a time resolution of seconds, and outputting simulation verification data to the multi-dimensional data fusion module.

[0109] The fluid-structure interaction solving sub-module of the dynamic correlation model construction module receives the spatio-temporal alignment data output by the data calibration module. The spatio-temporal alignment data includes the seepage pressure field, the surrounding rock strain field, and the corresponding spatio-temporal coordinate information. The fluid-structure interaction model is based on the theories of porous media seepage mechanics and elastoplastic mechanics to establish the coupled control equations of the seepage pressure gradient tensor and the surrounding rock strain tensor. The implicit iterative solver uses the Newton-Raphson method to handle the non-linear terms of the equations, updates the iteration step size through the Jacobian matrix, and calculates the dynamic interaction relationship between the seepage field and the stress field. An adaptive time step control strategy is introduced during the solving process, and the calculation step size is dynamically adjusted according to the seepage pressure change rate, and a seepage and strain coupling response data set is output, including the evolution law of fracture permeability and the distribution characteristics of surrounding rock displacement.

[0110] The digital twin driving sub-module constructs a three-dimensional numerical model of the surrounding rock fracture network based on the discrete element method. In the model initialization stage, geological exploration data and rock mechanics parameters are imported, and the initial contact stiffness and friction coefficient of the fracture surface are defined. The real-time seepage pressure data is mapped as the permeability boundary condition of the fracture surface through the spatial interpolation algorithm, and the dynamic relaxation algorithm is used to solve the contact force and motion equations of the fracture blocks to simulate the fracture propagation behavior caused by the seepage pressure fluctuation. During the fracture propagation process, the non-linear relationship between permeability and strain tensor is characterized by a state-dependent stiffness decay model. The model updates the fracture network topology connection relationship and mechanical property parameters in real time according to the coupling response data output by the fluid-structure interaction solving sub-module.

[0111] The fluid-structure interaction solution sub-module and the digital twin-driven sub-module achieve collaborative computing through a data interaction interface. The coupled seepage and strain response data triggers the update of the boundary conditions of the digital twin model, and the result of the fracture network evolution inversely corrects the permeability tensor term in the fluid-structure interaction equation, forming a two-way coupling iteration mechanism. The digital twin model outputs dynamic simulation data of fracture aperture, permeability distribution, and surrounding rock strain field at a second-level time resolution. After the simulation data and the measured data are aligned in space and time, they are transmitted to the multi-dimensional data fusion module for consistency verification. The model parameter update log records key variables such as the permeability attenuation coefficient and the contact stiffness correction amount, providing a basis for reverse verification for subsequent parameter optimization.

[0112] Specifically, for the multi-sensor safety monitoring system of the pumped-storage power station described in the present invention, the multi-dimensional data fusion module includes:

[0113] A spatio-temporal feature extraction sub-module and a dynamic weighted fusion sub-module, where:

[0114] The spatio-temporal feature extraction sub-module is used to receive the simulation verification data output by the dynamic correlation model construction module and the spatio-temporal alignment data output by the data calibration module, including:

[0115] A convolutional long short-term memory network improved by a multi-head attention mechanism, which inputs the seepage pressure time-series data and the displacement rate spatial distribution data, extracts local deformation features through spatio-temporal convolutional kernels, and uses the multi-head attention mechanism to capture the correlation between pressure and displacement across sensor nodes, generating a spatio-temporal correlation feature matrix;

[0116] The dynamic weighted fusion sub-module is connected to the spatio-temporal feature extraction sub-module and includes:

[0117] Calculate the model confidence weight according to the error rate between the simulation verification data and the measured data, and at the same time calculate the data reliability weight based on the historical stability index of the sensor node;

[0118] Perform a normalized product operation on the model confidence weight and the data reliability weight to generate the dynamic weighted fusion coefficient of each sensor node;

[0119] Perform multi-dimensional fusion on the spatio-temporal correlation feature matrix through a weighted decision function, and output the fusion decision data to the hierarchical warning module.

[0120] The spatio-temporal feature extraction sub-module of the multi-dimensional data fusion module receives the spatio-temporally aligned data output by the data calibration module and the simulation verification data output by the dynamic correlation model construction module. The spatio-temporally aligned data contains the temporal sequence of seepage pressure and the spatial distribution information of displacement rate, and the simulation verification data provides the numerical simulation results of fracture permeability evolution and surrounding rock strain field. The spatio-temporal feature extraction sub-module adopts a convolutional long short-term memory network improved by the multi-head attention mechanism. The input layer of the network normalizes the temporal sequence data of seepage pressure, and at the same time converts the displacement rate data into a spatial grid distribution matrix. The spatio-temporal convolution kernel slides along the time axis and the spatial axis to extract the change gradient of seepage pressure and the spatial correlation features of displacement rate in the local area. The multi-head attention mechanism calculates the correlation weights of pressure and displacement between different sensor nodes in parallel to generate a spatio-temporal correlation feature matrix across dimensions.

[0121] The dynamic weighted fusion sub-module calculates the model confidence weight according to the residual analysis of the simulation verification data and the measured data. The residual analysis adopts a sliding window statistical method to calculate the mean square error rate of the simulation data and the measured values at the corresponding spatio-temporal positions. The error rate is mapped to the model confidence weight coefficient after exponential smoothing processing. The historical stability index of the sensor node is generated by statistically calculating the signal packet loss rate, noise variance and long-term drift amount, and the data reliability weight is quantified based on the index entropy method. The model confidence weight and the data reliability weight are fused through normalized product operation to generate a dynamic weighted fusion coefficient matrix, and the weights of low-confidence or low-reliability nodes in the coefficient matrix are adaptively suppressed.

[0122] The weighted decision function performs an element-wise product operation on the spatio-temporal correlation feature matrix and the dynamic weighted fusion coefficient matrix. The fused feature vector is reduced in dimension through a fully connected layer to generate fusion decision data representing the system safety state. The fusion decision data integrates multi-dimensional information such as abnormal fluctuations of seepage pressure, sudden changes in displacement rate, and fracture expansion trend, and is output to the hierarchical early warning module to trigger threshold determination. During the feature extraction and weighted fusion process, the dimension of the spatio-temporal correlation feature matrix is consistent with the spatial distribution of the sensor nodes, and the dynamic update frequency of the weight coefficient matrix is synchronized with the data sampling period to ensure the timeliness and spatial resolution of the fusion result.

[0123] Specifically, for the multi-sensor safety monitoring system of the pumped storage power station described in the present invention, the hierarchical early warning module includes:

[0124] An early warning trigger sub-module and a closed-loop parameter optimization sub-module, where:

[0125] The early warning trigger sub-module is used to receive the fusion decision data output by the multi-dimensional data fusion module, including:

[0126] Dividing the three-level early warning thresholds of seepage erosion, crack expansion and rock mass instability based on the structural safety margin model, where:

[0127] The seepage erosion threshold is determined by calculating the ratio of the seepage velocity to the critical permeability through Darcy's law;

[0128] The crack propagation threshold is based on calculating the stress intensity factor at the crack tip using the fracture mechanics J-integral model;

[0129] The rock mass instability threshold is used to determine the safety factor of the shear slip surface of the surrounding rock according to the Mohr-Coulomb criterion;

[0130] The fusion decision data is compared with the three-level thresholds in real time. When the continuous overrun duration of the data exceeds the preset tolerance, an audible and visual alarm signal at the corresponding level is triggered;

[0131] The closed-loop parameter optimization sub-module is connected to the early warning trigger sub-module and includes:

[0132] Taking the clock synchronization error rate of the spatio-temporal synchronization network module and the root mean square of the prediction residuals of the dynamic correlation model construction module as the fitness function, a genetic algorithm is used for multi-objective optimization;

[0133] Elite individuals are selected from the population through the tournament selection strategy. The crossover operator uses arithmetic recombination, and the mutation operator introduces Gaussian perturbation to iteratively search for the optimal combination of clock synchronization parameters and model coefficients;

[0134] The optimized parameters are fed back to the clock synchronization sub-module of the spatio-temporal synchronization network module and the digital twin driving sub-module of the dynamic correlation model construction module in real time.

[0135] The early warning trigger sub-module of the hierarchical early warning module receives the fusion decision data output by the multi-dimensional data fusion module. The fusion decision data includes multi-dimensional characteristic quantities such as the abnormal index of seepage pressure, the displacement rate gradient, and the crack expansion rate. The early warning trigger sub-module sets three-level early warning thresholds based on the structural safety margin model. The seepage erosion threshold is calculated by the ratio of the seepage velocity in the monitoring area to the critical permeability of the rock mass through Darcy's law. When the ratio continuously exceeds the material durability threshold, it is determined as the seepage erosion risk; the crack expansion threshold is based on the fracture mechanics J-integral model to calculate the relative relationship between the stress intensity factor at the crack tip and the rock fracture toughness, and dynamically tracks the crack expansion trend; the rock mass instability threshold uses the Mohr-Coulomb criterion to analyze the ratio of the shear strength to the shear stress of the shear slip surface of the surrounding rock, and combines the safety factor threshold to determine the instability risk. The real-time monitoring data is compared with the three-level thresholds using a sliding window. When the continuous overrun duration of the characteristic quantity exceeds the preset tolerance window, an audible and visual alarm signal at the corresponding level is triggered and an early warning log is generated.

[0136] The closed-loop parameter optimization sub-module takes the spatio-temporal synchronization error rate and the model prediction residual as the optimization objectives. The spatio-temporal synchronization error rate is calculated through the timestamp deviation statistic between the master node and the slave nodes, and the model prediction residual is the root mean square error between the digital twin simulation data and the measured data. In the initialization stage of the genetic algorithm, population individuals containing clock synchronization parameters and fracture permeability attenuation coefficients are randomly generated. The fitness function is defined as the weighted sum of the spatio-temporal synchronization error rate and the model residual. The tournament selection strategy selects the top 10% of the elite individuals in terms of fitness from the population. The arithmetic recombination operator performs linear interpolation on the parameters of the elite individuals to generate offspring, and the Gaussian perturbation operator applies random perturbations to the offspring parameters to maintain population diversity. During the iterative optimization process, the search is terminated when the fitness value converges to a preset threshold, and the optimal parameter combination is output.

[0137] The optimized clock synchronization parameters are sent to the clock synchronization sub-module of the spatio-temporal synchronization network module through control messages to update the crystal oscillator drift correction coefficient; the fracture permeability attenuation coefficient is fed back to the digital twin drive sub-module of the dynamic correlation model construction module to correct the fracture surface permeability boundary condition. The parameter update triggers the model recalibration and data synchronization mechanisms, forming a closed-loop optimization link from early warning determination to parameter adaptation. The early warning logs and parameter update records are stored in the system database to support subsequent fault tracing and model iterative training.

[0138] The explanations of the technical features in the technical solution of the present invention are as follows:

[0139] The redundant sensor node group adopts a triple redundancy configuration, which means that three sensors of the same type (such as strain, seepage pressure, vibration sensors) are deployed at each monitoring point, and the signals of the same physical quantity are collected through a synchronous trigger circuit. The stress concentration area of the penstock is a sensitive area for structural deformation. The triple redundancy design cross-verifies the data through a majority voting mechanism to improve the acquisition reliability. The sliding time window majority voting algorithm dynamically adjusts the window length to adapt to the sampling frequency. After normalization to eliminate the dimension difference, the Euclidean distance between two-by-two data is calculated. The dynamic threshold is generated based on historical data statistics. The continuously exceeding-limit data is determined as abnormal and excluded, and the remaining data is averaged and output. The adaptive wavelet threshold filtering algorithm selects the Daubechies or Symlet wavelet basis based on spectrum analysis. The improved SUREShrink threshold estimation introduces a frequency band energy weight factor to preferentially suppress high-frequency electromagnetic noise, and the reconstructed signal retains the time-domain waveform mutation and frequency-domain harmonic components.

[0140] The satellite time synchronization protocol (PTP / NTP) enables the master node to obtain the UTC reference clock from the Beidou system and distribute it to the slave nodes through a low-latency broadcast protocol. The Precision Clock Synchronization Protocol (IEEE 1588v2) establishes a two-way synchronization link between the master and slave nodes to compensate for the path delay asymmetry. The linear regression model analyzes the historical deviation data of the slave node crystal oscillator, predicts the frequency drift trend, and dynamically corrects the value of the local clock counter to achieve microsecond-level timestamp alignment. The cluster communication optimization sub-module divides the subnet according to geological units (such as the direction of the penstock and the surrounding rock fault). The cluster head node is elected through the self-organization protocol within the subnet. Time Division Multiple Access (TDMA) pre-allocates time slots, and the dynamic time slice adjustment algorithm compresses the time slot length and increases the density according to the channel load intensity to optimize the data transmission timing.

[0141] The Two-Way Timestamp Protocol (TWTT) measures the round-trip delay between the master and slave nodes, and the least squares method is used to fit the delay fluctuation curve to separate the fixed delay component. The Kalman filter algorithm predicts the random delay deviation caused by network jitter, and constructs an interpolation function to resample and align the asynchronous data on the time axis. The 3D laser scanner acquires point cloud data with a resolution of 0.1 mm. The region growing algorithm is used to segment the point cloud clusters of the sensor nodes. The Iterative Closest Point (ICP) algorithm with normal vector constraint introduces the normal vector direction constraint of the installation surface, registers the actual coordinates with the theoretical coordinates, and calculates the offset and then converts it to the global coordinate system through the quaternion rotation matrix.

[0142] The fluid-structure interaction model is based on the partial differential equations of the seepage pressure gradient and the surrounding rock strain tensor. The implicit iterative solver (such as the Newton-Raphson method) updates the iteration step through the Jacobian matrix to calculate the dynamic correlation between the seepage field and the stress field. The Discrete Element Method (DEM) digital twin model initializes the fracture contact stiffness and friction coefficient, maps the real-time seepage pressure data as the permeability boundary condition of the fracture surface, and the dynamic relaxation algorithm solves the block motion equation to simulate the fracture propagation. The non-linear relationship between permeability and strain is characterized by the stiffness decay model, and the fracture topology structure is updated with a second-level resolution.

[0143] The Conv-LSTM network improved by the multi-head attention mechanism normalizes the seepage pressure time series data. The displacement rate data is converted into a spatial grid matrix. The spatio-temporal convolution kernel extracts local deformation features. The multi-head attention calculates the cross-node pressure-displacement correlation weights in parallel to generate a spatio-temporal correlation feature matrix. The dynamic weighted fusion coefficient generates the model confidence weight based on the mean square error rate of the sliding window statistical simulation data and the measured data, combines the sensor historical packet loss rate and the noise variance entropy value method to calculate the data reliability weight, normalizes the product to generate the fusion coefficient matrix, and the weighted decision function fuses the feature matrix to output the decision data.

[0144] The seepage erosion threshold is calculated by the ratio of seepage velocity to critical permeability through Darcy's law. The crack propagation threshold is based on the J-integral model to calculate the ratio of the stress intensity factor at the crack tip to the rock fracture toughness. The rock mass instability threshold analyzes the safety factor of the shear slip surface of the surrounding rock according to the Mohr-Coulomb criterion. The genetic algorithm uses the clock synchronization error rate and the model residual as the fitness function, selects elite individuals by the tournament strategy, arithmetically recombines the linear interpolation parameters, and uses Gaussian perturbation to maintain the population diversity. It outputs the optimal clock synchronization parameters (crystal oscillator drift correction coefficient) and the crack permeability attenuation coefficient, which are fed back to the clock synchronization sub-module and the digital twin model to form a closed-loop optimization link.

[0145] In the specific implementation manner of the present invention, aiming at the complex hydraulic and mechanical coupling environment of the multi-sensor safety monitoring system of the pumped storage power station, high-precision spatio-temporal synchronization and the identification of the coupling mechanism of abnormal events are realized through multi-level collaborative technology. The sensor data acquisition module deploys three-redundancy sensor nodes in the stress concentration area of the penstock and the surrounding rock deformation sensitive area. Each monitoring point is equipped with three sensors of the same type to form a data acquisition unit, and the synchronous acquisition of stress, seepage pressure and vibration signals is realized through a synchronous trigger circuit. The sliding time window majority voting algorithm dynamically adjusts the window length according to the sensor sampling frequency, normalizes the three-way data and calculates the Euclidean distance. The dynamic threshold is generated based on the statistical characteristics of historical data. Continuously exceeding the limit data is determined as an outlier and removed, and the mean value of the remaining data is used as the effective output. The adaptive wavelet threshold filtering algorithm selects the Daubechies or Symlet wavelet basis function based on the signal spectrum analysis result, separates the high-frequency electromagnetic interference noise through the improved SUREShrink threshold estimation, and reconstructs the signal to retain the time-domain waveform mutation and frequency-domain harmonic characteristics, generating preprocessed data.

[0146] The spatio-temporal synchronization network module obtains the reference clock signal through the Beidou satellite timing protocol. The master node analyzes the UTC time and broadcasts it to the slave nodes. The slave nodes establish a two-way synchronization link using the precision clock synchronization protocol. The linear regression model analyzes the historical deviation data of the slave node crystal oscillator in real time, predicts the cumulative drift amount and dynamically corrects the timestamp to achieve microsecond-level time alignment. The cluster communication optimization sub-module divides the monitoring subnet according to geological units. The cluster head node is elected within the subnet through the self-organizing network protocol, and fixed time slots are allocated based on the time division multiple access mechanism. The master node monitors the channel load intensity of the subnet in real time. When the data volume exceeds the preset threshold, the dynamic time slice adjustment algorithm compresses the single time slot length and increases the time slot density to optimize the data transmission timing. The data after timestamp correction is encapsulated into a standard format stream and transmitted to the data calibration module.

[0147] The data calibration module measures the transmission delay between the master node and the slave node using the two-way timestamp protocol, generates the reference delay amount by fitting the delay fluctuation curve with the least squares method, predicts the random deviation caused by network jitter with the Kalman filter, and constructs a time-axis interpolation function to align asynchronous data. The 3D laser scanner scans the monitoring area at a resolution of 0.1 mm to generate point cloud data. The region growing algorithm is used to segment the point cloud clusters of the sensor nodes. The iterative closest point algorithm with normal vector constraint is used to register the actual coordinates and the theoretical coordinates. After calculating the coordinate offset, it is converted to the global engineering coordinate system based on the quaternion rotation matrix. The dynamic correlation model construction module inputs the spatio-temporally aligned data into the fluid-structure interaction model, and the implicit iterative solver solves the coupled equations of the seepage pressure gradient and the surrounding rock strain tensor, and outputs the seepage pressure and strain response data. The discrete element method is used to construct a digital twin model of the surrounding rock fractures. The real-time seepage pressure data is mapped as the permeability boundary condition of the fracture surface. The dynamic relaxation algorithm simulates the non-linear relationship between permeability and strain, updates the fracture network topology, and outputs the simulation verification data at the second-level resolution.

[0148] The multi-dimensional data fusion module uses a convolutional long short-term memory network improved by the multi-head attention mechanism. The spatio-temporal convolutional kernel extracts the temporal gradient of seepage pressure and the spatial distribution characteristics of displacement rate. The multi-head attention mechanism calculates the cross-node pressure-displacement correlation weight to generate a spatio-temporal correlation feature matrix. The dynamic weighted fusion sub-module generates the model confidence weight based on the mean square error rate of the sliding window of the simulation data and the measured data, calculates the data reliability weight in combination with the historical packet loss rate and noise variance index of the sensor, generates a dynamic fusion coefficient matrix through normalized product operation, and the weighted decision function fuses the feature matrix to output the decision data. The hierarchical early warning module sets the seepage erosion threshold by calculating the ratio of the seepage velocity to the critical permeability according to Darcy's law, sets the crack propagation threshold by calculating the stress intensity factor at the crack tip with the J integral model, sets the rock mass instability threshold by judging the safety factor of the shear slip surface of the surrounding rock with the Mohr-Coulomb criterion, and continuously triggers the acoustic and optical alarm when exceeding the limit. The genetic algorithm uses the clock synchronization error rate and the model residual as the fitness function, selects elite individuals with the tournament strategy, generates the optimal parameter combination through arithmetic recombination and Gaussian perturbation, and feeds it back to the clock synchronization sub-module to correct the crystal oscillator drift coefficient and update the fracture permeability boundary condition of the digital twin model, forming a closed-loop optimization link.

[0149] The present invention solves the problem of mechanism identification deviation of multi-sensor monitoring networks in pumped storage power stations under complex coupling environments through a multi-level spatio-temporal synchronization architecture and a dynamic coupling model collaboration mechanism. The sensor data acquisition module adopts a three-redundant node configuration and a sliding time window majority voting algorithm to synchronously collect stress, seepage pressure, and displacement signals, eliminates abnormal data through normalization processing and Euclidean distance calculation, suppresses electromagnetic noise by combining adaptive wavelet threshold filtering, and generates high-confidence preprocessed data. The spatio-temporal synchronization network module is based on the Beidou satellite timing protocol and a hierarchical clock synchronization architecture, predicts the oscillator drift amount through a linear regression model and dynamically corrects the time stamp to achieve microsecond-level time alignment; the cluster communication optimization sub-module divides the monitoring subnet by geological unit, adopts a dynamic time slice adjustment algorithm to optimize the data transmission timing, reduces the probability of channel conflict, and ensures the real-time and integrity of data.

[0150] The data calibration module predicts the network transmission delay through a two-way time stamp protocol and a Kalman filter, constructs an interpolation function to resample asynchronous data on the time axis, and eliminates the timing deviation. The three-dimensional laser scanner obtains the point cloud data of the monitoring area, registers the actual coordinates and theoretical coordinates of the sensor nodes by using the iterative closest point algorithm with normal vector constraints, and converts them to the global engineering coordinate system based on the quaternion rotation matrix to achieve spatial calibration. The dynamic correlation model construction module inputs the spatio-temporally aligned data into the fluid-structure coupling model, calculates the dynamic correlation equation of seepage pressure gradient and surrounding rock strain through an implicit iterative solver, drives the discrete element method digital twin model to simulate the crack propagation process, updates the permeability boundary conditions and crack network topology structure in real time, generates seepage-strain coupling response data, and improves the simulation accuracy of the hydraulic and mechanical interaction mechanism.

[0151] The multi-dimensional data fusion module adopts a convolutional long short-term memory network improved by a multi-head attention mechanism to extract the cross-sensor spatio-temporal correlation features of seepage pressure and displacement rate, combines the simulation error rate and sensor stability index to generate a dynamic weighted fusion coefficient, and suppresses the interference of low-confidence data. The hierarchical early warning module sets three-level early warning thresholds based on Darcy's law, the J integral model, and the Mohr-Coulomb criterion, uses the spatio-temporal synchronization error rate and model residual as fitness functions through a genetic algorithm to optimize the clock synchronization parameters and crack permeability attenuation coefficient, and forms a closed-loop feedback mechanism. The parameter optimization results adjust the oscillator drift correction coefficient of the clock synchronization sub-module and the boundary conditions of the digital twin model in real time, and dynamically improve the system's identification ability of the coupling relationship of abnormal events.

Claims

1. A multi-sensor safety monitoring system for a pumped storage power station, characterized in that, Including: A sensor data acquisition module, a spatio-temporal synchronization network module, a data calibration module, a dynamic correlation model construction module, a multi-dimensional data fusion module, and a hierarchical warning module; The sensor data acquisition module is used to collect stress, seepage pressure, displacement, and vibration signals of key monitoring points through redundant sensor nodes, and perform noise suppression processing on the collected signals to generate preprocessed data; The spatio-temporal synchronization network module is used to receive the preprocessed data output by the sensor data acquisition module, achieve microsecond-level time alignment of sensor nodes through a hierarchical clock synchronization architecture, and dynamically adjust the data transmission timing based on cluster-based communication optimization to output synchronized timing data; The data calibration module is used to receive the synchronized timing data output by the spatio-temporal synchronization network module, align asynchronous data through a transmission delay compensation algorithm, and calibrate the global coordinate system of sensor nodes in combination with a three-dimensional laser scanner to generate spatio-temporally aligned data; The dynamic correlation model construction module is used to input the spatio-temporally aligned data output by the data calibration module into a fluid-structure interaction model, drive a digital twin model based on the discrete element method to simulate the dynamic responses of seepage pressure and surrounding rock strain, and output simulation verification data; The multi-dimensional data fusion module is used to extract the spatio-temporal correlation features of the spatio-temporally aligned data and the simulation verification data, and generate fusion decision data through a weighted fusion algorithm; The hierarchical warning module is used to trigger multi-level warning thresholds according to the fusion decision data, and iteratively optimize the clock synchronization parameters of the spatio-temporal synchronization network module and the fluid-structure interaction model coefficients of the dynamic correlation model construction module based on the fusion decision data.

2. The multi-sensor safety monitoring system for a pumped storage power station according to claim 1, characterized in that The sensor data acquisition module includes: A redundant sensor node group, an abnormal data elimination unit, and an adaptive noise suppression unit, where: The redundant sensor node group adopts a triple redundancy configuration and is deployed in the stress concentration area of the penstock. Three sensors of the same type are set at each monitoring point to form a data acquisition unit, and the three sensors synchronously collect signals of the same physical quantity; The abnormal data elimination unit is connected to the redundant sensor node group and eliminates abnormal data through a majority voting algorithm within a sliding time window. Specifically: normalize the data collected by the three sensors within the sliding time window, calculate the Euclidean distance between pairwise data. If the distance of a certain data from the other two data exceeds the dynamic threshold, it is determined as abnormal data and eliminated, and the mean value of the remaining data is used as the effective output; The adaptive noise suppression unit is connected to the abnormal data elimination unit and uses an adaptive wavelet threshold filtering algorithm to denoise the effective output signal, including: dynamically selecting a wavelet basis function based on the signal spectrum analysis result, separating the electromagnetic interference noise frequency band through an improved SUREShrink threshold estimation, and retaining the time-domain waveform characteristics and frequency-domain harmonic characteristics during signal reconstruction to generate preprocessed data and output it to the spatio-temporal synchronization network module.

3. The multi-sensor safety monitoring system for a pumped storage power station according to claim 1, characterized in that The spatio-temporal synchronization network module includes a clock synchronization sub-module and a cluster communication optimization sub-module, where: The clock synchronization sub-module is used to receive the preprocessed data output by the sensor data acquisition module, including: The master node obtains the reference clock signal from the Beidou satellite system through the satellite time synchronization protocol and broadcasts the reference clock to all slave nodes; Each slave node establishes a time synchronization link with the master node through the precision clock synchronization protocol. Meanwhile, a linear regression model is adopted to predict the cumulative drift of the crystal oscillator of each slave node in real time, and the timestamp of the sensor data is dynamically corrected according to the drift amount for microsecond-level time alignment between sensor nodes; The cluster communication optimization sub-module is connected to the clock synchronization sub-module and includes: The monitoring area is divided into multiple geological unit subnets according to the geological structure characteristics, and the sensor nodes within each subnet form an independent communication cluster; In the subnet, the time division multiple access mechanism is used to allocate fixed data transmission time slots for each node. The master node monitors the channel load intensity in real time. When the data volume of the nodes in the subnet exceeds the preset threshold, the dynamic time slice adjustment algorithm is adopted to shorten the length of a single time slot and increase the time slot density to optimize the data transmission timing; The data after time alignment and timing optimization is output as synchronous timing data and transmitted to the data calibration module.

4. The multi-sensor safety monitoring system for a pumped storage power station according to claim 1, characterized in that, The data calibration module includes a time delay compensation sub-module and a space calibration sub-module, where: The time delay compensation sub-module is used to receive the synchronous timing data output by the space-time synchronization network module and includes: The two-way transmission link delay between the master node and the slave node is measured by the two-way timestamp protocol, and the delay fluctuation curve is fitted by the least square method to generate the reference transmission delay amount; Based on the Kalman filter algorithm, the random delay deviation caused by network jitter is predicted, and an interpolation function is constructed in combination with the reference transmission delay amount to resample the sensor data arriving asynchronously on the time axis for microsecond-level time alignment; The space calibration sub-module is connected to the time delay compensation sub-module and includes: The three-dimensional point cloud data of the monitoring area is obtained by a three-dimensional laser scanner, and the actual point cloud coordinates of the installation positions of the sensor nodes are extracted; The iterative closest point algorithm with normal vector constraint is used to register the actual point cloud coordinates with the theoretical coordinates in the design drawing, and the coordinate offset of the sensor node is calculated; Based on the quaternion rotation matrix, the local coordinate system of each sensor node is converted to the global engineering coordinate system, and the space-time alignment data is output to the dynamic association model construction module.

5. The multi-sensor safety monitoring system for a pumped storage power station according to claim 1, characterized in that, The dynamic association model construction module includes a fluid-structure interaction solution sub-module and a digital twin drive sub-module, where: The fluid-structure interaction solution sub-module is used to receive the space-time alignment data output by the data calibration module and includes: Based on the coupled partial differential equation of the seepage pressure gradient tensor and the surrounding rock strain tensor, the dynamic association relationship between the seepage field and the stress field is calculated by an implicit iterative solver to generate the coupled response data of seepage pressure and strain; The digital twin drive sub-module is connected to the fluid-structure interaction solution sub-module and includes: A three-dimensional digital twin model of the surrounding rock fracture propagation is constructed by the discrete element method, the real-time seepage pressure data is mapped as the permeability boundary condition of the fracture surface, and the nonlinear relationship between permeability and strain during the fracture propagation process is simulated by the dynamic relaxation algorithm; Update the fracture network topology of the digital twin model according to the coupled response data of seepage pressure and strain, drive the dynamic evolution of the digital twin model, and output simulation verification data to the multi-dimensional data fusion module.

6. The multi-sensor safety monitoring system for a pumped storage power station according to claim 1, characterized in that, The multi-dimensional data fusion module includes a spatio-temporal feature extraction sub-module and a dynamic weighted fusion sub-module, where: The spatio-temporal feature extraction sub-module is used to receive the simulation verification data output by the dynamic correlation model construction module and the spatio-temporal alignment data output by the data calibration module, including: A convolutional long short-term memory network improved by a multi-head attention mechanism, inputting the time series data of seepage pressure and the spatial distribution data of displacement rate, extracting local deformation features through spatio-temporal convolutional kernels, and using the multi-head attention mechanism to capture the correlation between pressure and displacement across sensor nodes, generating a spatio-temporal correlation feature matrix; The dynamic weighted fusion sub-module is connected to the spatio-temporal feature extraction sub-module, including: calculating the model confidence weight according to the error rate between the simulation verification data and the measured data, and calculating the data reliability weight based on the historical stability index of the sensor node; Perform a normalized product operation on the model confidence weight and the data reliability weight to generate the dynamic weighted fusion coefficient of each sensor node; Perform multi-dimensional fusion on the spatio-temporal correlation feature matrix through a weighted decision function, and output the fusion decision data to the hierarchical warning module.

7. The multi-sensor safety monitoring system for a pumped storage power station according to claim 1, characterized in that The hierarchical warning module includes a warning trigger sub-module and a closed-loop parameter optimization sub-module, where: The warning trigger sub-module is used to receive the fusion decision data output by the multi-dimensional data fusion module, including: Based on the structural safety margin model, divide the three-level warning thresholds for seepage erosion, crack propagation, and rock mass instability, where: The seepage erosion threshold is determined by calculating the ratio of the seepage velocity to the critical permeability through Darcy's law; The crack propagation threshold is based on calculating the stress intensity factor at the crack tip using the fracture mechanics J integral model; The rock mass instability threshold determines the safety factor of the surrounding rock shear slip surface according to the Mohr-Coulomb criterion; Compare the fusion decision data with the three-level thresholds in real time. When the continuous over-limit duration of the data exceeds the preset tolerance, trigger the corresponding level of audible and visual alarm signal; The closed-loop parameter optimization sub-module is connected to the warning trigger sub-module, including: using the clock synchronization error rate of the spatio-temporal synchronization network module and the root mean square of the prediction residuals of the dynamic correlation model construction module as the fitness function, and using a genetic algorithm for multi-objective optimization; Select elite individuals from the population through the tournament selection strategy, use arithmetic recombination for the crossover operator, introduce Gaussian perturbation for the mutation operator, and iteratively search for the optimal combination of clock synchronization parameters and fluid-structure coupling model coefficients; Real-time feedback the optimized parameters to the clock synchronization sub-module of the spatio-temporal synchronization network module and the digital twin drive sub-module of the dynamic correlation model construction module.

Images