A hierarchical limit linkage alarm water and electricity monitoring method and system

Abstract

The application belongs to the technical field of water and electricity monitoring, and particularly discloses a water and electricity monitoring method and system with hierarchical limit linkage alarm, wherein the method comprises the following steps: acquiring water and electricity unit multi-source sensing data collected synchronously, wherein the multi-source sensing data at least comprises vibration, swing, temperature, pressure and electrical parameters; pre-processing the multi-source sensing data and extracting time-frequency domain features; and constructing a dynamic benchmark model driven by physics and data, wherein the dynamic benchmark model integrates data-driven prediction of a deep residual shrinkage network and physical mechanism prediction based on a heat balance equation and a water hammer effect model; the application aims to solve the technical problems of high false alarm rate, high missed alarm rate and insufficient equipment health state sensing capability caused by the fact that the existing water and electricity monitoring technology adopts a static threshold or a pure data-driven model, which makes it difficult to realize dynamic collaborative early warning, fault root cause tracing and unknown working condition self-adaption in a complex operating environment with multiple working conditions and multiple physical field coupling.

Description

Technical Field

[0001] This invention relates to the technical field of hydropower monitoring, specifically to a hydropower monitoring method and system with graded limit linkage alarms. Background Technology

[0002] Hydropower stations, as core hubs for power grid peak shaving and frequency regulation, are of paramount importance for their safe and reliable operation. Modern hydropower monitoring systems utilize large-scale sensor arrays to achieve real-time monitoring of parameters such as unit vibration, sway, temperature, and pressure. Traditional alarm mechanisms primarily rely on preset static thresholds, triggering alarms when parameters exceed fixed limits. However, the operating conditions of hydropower units are complex and variable (changes in load, head, excitation current, etc.), and static thresholds cannot adapt to these fluctuations, resulting in high false alarm rates and a significant risk of missed alarms.

[0003] To address the aforementioned issues, existing technologies have proposed several improvement solutions. For example, patent CN112946469B discloses a monitoring method for adaptive dynamic alarm thresholds in hydro-generators. This method segments operating conditions by collecting active power, head, and excitation current data. It then uses historical data to calculate the quartiles, maximum values, and minimum values ​​of the monitored indicators under each operating condition, dynamically generating upper and lower threshold limits, and triggering alarms based on the exceedance ratio and duration. While this method achieves adaptive threshold generation under operating conditions, its threshold generation relies entirely on historical data statistics, lacking constraints on the unit's physical mechanisms. This may lead to unreasonable thresholds under extreme operating conditions or in data-sparse regions, and it cannot explain why the threshold falls within a specific range.

[0004] Patent CN121055497A proposes an intelligent AGC (Automatic Guided Control) method for hydropower stations based on multi-source data fusion. This method integrates power grid, generating unit, hydraulic engineering, and equipment status data, employs a spatiotemporal attention LSTM (STA-LSTM) for feature modeling, and establishes a multi-objective optimization function (frequency deviation, power tracking error, generating unit energy consumption, and equipment lifespan loss). By dynamically adjusting weights through fuzzy logic, intelligent AGC control is achieved. While this method is advanced in the control field, its focus is on load allocation and PID parameter optimization rather than monitoring and alarms. Furthermore, its parameter correlation analysis, based on an attention mechanism, can only capture correlations rather than causality, making it difficult to pinpoint the root cause of faults.

[0005] Patent CN120848222B discloses an intelligent start-up and shutdown system and method for hydropower units based on multi-dimensional dynamic optimization and high-precision collaborative control. Through modules such as dynamic priority management, vibration zone self-learning, collaborative interpolation control, V-curve recognition, intelligent diagnosis, and risk scoring, the system optimizes the start-up and shutdown process. While the system introduces vibration zone self-learning based on a sliding window and risk scoring based on FMEA, its fault diagnosis remains primarily classification-based, failing to dynamically correlate equipment health status with alarm thresholds and lacking assessment of model prediction uncertainty.

[0006] In summary, while existing technologies have made progress in adaptive thresholding, multi-source data fusion, and intelligent control, they still have the following shortcomings: 1. Threshold generation relies on pure data-driven approaches and lacks physical mechanism constraints; 2. Multi-parameter linkage is limited to correlation analysis and cannot achieve causal inference; 3. The characterization of abnormal states is limited to a single time scale, making it difficult to simultaneously capture transient shocks and slow degradation; 4. The model output lacks confidence assessment, making it prone to false alarms under unknown operating conditions. Therefore, there is an urgent need for a hydropower monitoring solution that can deeply integrate physical mechanisms and data-driven approaches, possess causal inference capabilities, multi-scale anomaly perception, and self-evolving capabilities. Summary of the Invention

[0007] This invention provides a graded limit-linked alarm method and system for hydropower monitoring. The purpose is to solve the technical problems of existing hydropower monitoring technologies, which use static thresholds or pure data-driven models. These technologies make it difficult to achieve dynamic collaborative early warning, fault root cause tracing, and adaptive response to unknown operating conditions in complex operating environments with multiple operating conditions and multiple physical fields, resulting in high false alarm rates, high missed alarm rates, and insufficient equipment health status perception capabilities.

[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A hydropower monitoring method with tiered limit linkage alarms includes: acquiring synchronously collected multi-source sensor data from hydropower units, wherein the multi-source sensor data includes at least vibration, sway, temperature, pressure, and electrical parameters; preprocessing the multi-source sensor data to extract time-frequency domain features; constructing a dynamic benchmark model driven by both physical and data aspects, wherein the dynamic benchmark model integrates data-driven prediction from a deep residual shrinkage network with physical mechanism prediction based on the thermal balance equation and water hammer effect model, and outputs the expected steady-state value and adaptive variance of each measuring point under the current operating condition; setting a three-level progressive alarm limit based on the expected steady-state value and adaptive variance of the operating condition, wherein the three-level progressive alarm limit includes a warning limit, an alarm limit, and an emergency protection limit; constructing a parameter causality graph based on Granger causality test, wherein when a parameter at a measuring point exceeds the warning limit, a set of associated parameters with a significant causal relationship with that measuring point is retrieved, and the associated parameters are calculated. The co-evolution trend of the data set within the same sliding time window is determined to be a true degradation if it exceeds a preset proportion, and an alarm of the corresponding level is triggered according to the degree of exceeding the limit; multi-scale abnormal energy integral is calculated, which includes short-cycle abnormal energy and long-cycle abnormal energy. When the short-cycle abnormal energy or long-cycle abnormal energy exceeds the corresponding threshold, a trend warning is triggered; Bayesian deep learning is used to output the confidence level of the expected steady-state value. When the confidence level is lower than the preset confidence threshold, the alarm limit is switched to a conservative mode; when concurrent alarms occur, based on the pre-constructed fault propagation map and the parameter causal graph, the source alarm is identified and the derived alarm is suppressed, and a root cause analysis report is pushed; reinforcement learning is used to dynamically optimize the sensitivity coefficients of the short-cycle abnormal energy and long-cycle abnormal energy and the significance level threshold of the Granger causality test based on the confirmation or rejection feedback of the operation and maintenance personnel.

[0009] In one aspect of the invention, in constructing a dynamic benchmark model driven by both physics and data, the dynamic benchmark model integrates data-driven prediction from a deep residual shrinkage network with physical mechanism prediction based on the thermal balance equation and water hammer effect model, outputting the expected steady-state value and adaptive variance of each measuring point under the current operating condition. The training of the dynamic benchmark model driven by both physics and data aims to minimize the error between data-driven predictions and measured values, as well as the deviation between data-driven predictions and physical mechanism predictions. The weights of the two are adjusted by a balancing coefficient.

[0010] In one aspect of the invention, when a parameter at a certain measuring point exceeds a warning limit, a set of associated parameters that have a significant causal relationship with that measuring point is retrieved. The co-evolution trend of the associated parameter set within the same sliding time window is calculated. If it exceeds a preset proportion, it is determined to be a true degradation, and an alarm of the corresponding level is triggered according to the degree of exceeding the limit. The significance level for the Granger causality test is set to be less than 0.05, and the set of associated parameters includes parameters that have a Granger causal relationship with the parameter that exceeds the limit and whose significance level is less than 0.05; The preset ratio is 60%, and the co-evolution trend refers to the proportion of parameters whose associated parameters exhibit a monotonically increasing or decreasing characteristic in sync with the main parameters.

[0011] In one aspect of the invention, the calculation of the multi-scale anomaly energy integral includes short-period anomaly energy and long-period anomaly energy. A trend warning is triggered when either the short-period or long-period anomaly energy exceeds a corresponding threshold. The short-period anomaly energy is obtained by integrating the squared residuals between the measured values ​​and the expected steady-state values ​​within a preset short-period observation period. The long-period anomaly energy is obtained by integrating the absolute value of the difference between the rate of change of the measured value and the rate of change of the expected steady-state value within a preset long-period observation period. The corresponding threshold is calculated and determined based on the adaptive variance of the operating condition, the observation duration, and the preset sensitivity coefficient.

[0012] In one aspect of the invention, the confidence level of the desired steady-state value is output using Bayesian deep learning. When the confidence level is lower than a preset confidence threshold, the alarm limit is switched to a conservative mode. The confidence level is calculated based on the upper and lower limits of the prediction interval output by Bayesian deep learning and the expected steady-state value. The preset confidence threshold is 90%. The conservative mode includes expanding the warning limit to the expected steady-state value plus or minus five times the adaptive variance of the operating condition, or switching the alarm limit to the static limit value.

[0013] In one aspect of the invention, the three-level progressive alarm limit is set based on the expected steady-state value and the adaptive variance of the operating condition. The three-level progressive alarm limit includes a warning limit, an alarm limit, and an emergency protection limit. The three-level progressive alarm limit setting method is as follows: the early warning limit is equal to the expected steady-state value plus twice the operating condition adaptive variance, the alarm limit is equal to the expected steady-state value plus three times the operating condition adaptive variance and weighted correction based on the equipment's historical operating extreme values, and the emergency protection limit is set with reference to the instantaneous limit strength allowed by the equipment's design life.

[0014] In another aspect, the present invention also relates to a water and electricity monitoring system with graded limit-linked alarms, comprising: The data sensing layer consists of a sensor array deployed at key parts of the hydropower unit, used to synchronously collect multi-source sensor data. The sensor array includes a swing sensor, a temperature sensor, a pressure pulsation sensor, an embedded thermal resistor sensor, and an acoustic emission sensor. All sensor channels are connected to a clock synchronization module based on the Global Positioning System timing protocol, with a synchronization error of less than 1 microsecond. The edge computing layer includes an embedded real-time processing terminal, which is equipped with an ARM processor and FPGA collaborative processing architecture. The FPGA is used for real-time acquisition, filtering and fast Fourier transform of high-frequency signals, and the ARM processor is used to run dynamic benchmark prediction models, multi-parameter linkage discrimination algorithms and multi-scale anomaly energy integral calculation. The central decision-making layer includes a high-performance computing server cluster, a distributed database, and a visualization interactive workstation. The server cluster is used to train deep residual shrinkage networks and physical and data-driven models. The distributed database is used to store raw waveform data and alarm event logs. The visualization interactive workstation is used for 3D digital twin display and root cause analysis push. The closed-loop feedback calibration module is used to monitor the handling results of alarms by maintenance personnel in real time, and to optimize the sensitivity coefficient and Granger causality test threshold using reinforcement learning algorithms.

[0015] In one aspect of the invention, the edge computing layer also integrates a miniaturized real-time database for caching high-frequency raw data from the past 24 hours, and automatically locking the data for that time period to form a black box data packet when an emergency protection alarm is triggered and uploading it to the central decision-making layer.

[0016] In one aspect of the invention, the communication network of the system adopts a ring fiber optic Ethernet topology, follows the IEC61850 standard, supports media redundancy protocol, and ensures that the network single-point failure recovery time is less than 50 milliseconds.

[0017] Compared with the prior art, the present invention has the following beneficial effects: This invention constructs a physical and data-driven deep residual shrinkage network, using physical mechanisms such as the bearing thermal balance equation of the hydro-generator unit as loss function constraints. This ensures that the dynamic benchmark model strictly conforms to physical laws based on data-driven principles, effectively solving the problem of inaccurate predictions in pure data-driven models under extreme conditions or in data-sparse regions. It significantly improves the rationality and robustness of dynamic thresholds. Furthermore, it introduces a parametric causal graph based on Granger causality testing, moving from traditional correlation analysis to causal inference. In multi-parameter linkage discrimination, it can accurately trace the root cause of faults, significantly reducing the false alarm rate caused by instantaneous sensor interference or operating condition fluctuations. Experimental results show a substantial reduction in the average monthly number of false alarms. Finally, it defines a multi-scale abnormal energy integral, capturing second-level transient impacts through short-cycle abnormal energy and monitoring hour-level slow degradation through long-cycle abnormal energy, and weighted fusion. Subsequently, a comprehensive anomaly energy is formed, enabling unified early warning for both sudden and evolving faults. This overcomes the shortcomings of existing technologies that only focus on a single time scale, reducing the false negative rate. Bayesian deep learning is employed to quantify prediction uncertainty, dynamically adjusting the three-level limit width based on prediction variance. When the confidence level falls below a threshold, it automatically switches to a conservative mode, giving the alarm system self-awareness and avoiding blind alarms under unknown operating conditions, significantly enhancing system adaptability. A Q-learning-based reinforcement learning self-evolution mechanism is introduced, using alarm confirmation or rejection by maintenance personnel as a reward signal. This optimizes the sensitivity coefficient in multi-scale anomaly integration and the linkage threshold in causal linkage discrimination in real time, allowing alarm thresholds to adaptively adjust with equipment aging and changing operating conditions. This significantly reduces manual tuning workload and improves the accuracy of fault root cause localization. Through these multi-dimensional innovations, this invention constructs a full-chain monitoring system from physical perception to intelligent decision-making, completely overcoming the limitations of traditional static thresholds and existing adaptive methods. This significantly improves the intelligence level, operational reliability, and maintenance efficiency of hydropower station monitoring systems, providing solid technical support for unmanned operation and condition-based maintenance of hydropower stations. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 This is a flowchart of a water and electricity monitoring method with hierarchical limit linkage alarm according to the present invention.

[0020] Figure 2 This is a framework diagram of a hydropower monitoring system with hierarchical limit linkage alarm according to the present invention.

[0021] Figure 3This is a flowchart of the method in Embodiment 2 of the present invention. Detailed Implementation

[0022] The present invention will be further described below with reference to embodiments. These embodiments are merely some, not all, of the embodiments of the present invention. Other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are all within the protection scope of the present invention.

[0023] Please see Figure 1 as well as Figure 2 As shown in the figure, this embodiment discloses a water and electricity monitoring method and system with graded limit linkage alarm, wherein the method includes the following steps: S1, the original physical signals are synchronously acquired by a sensor array deployed in key parts of the hydropower station. The sensor array includes a swing sensor, a temperature sensor, a pressure pulsation sensor, an embedded thermal resistance sensor, and an acoustic emission sensor. All sampling channels are synchronized based on the GPS timing protocol, with a synchronization error of less than 1 microsecond.

[0024] S2. Preprocessing of the original signal: The signal is decomposed into intrinsic mode function components using adaptive noise complete set empirical mode decomposition (CEEMDAN), and high-frequency random noise is removed according to the Hearst exponent; time-frequency domain features are extracted from the reconstructed signal, including root mean square value, peak-to-peak value, kurtosis factor, centroid frequency, and dominant frequency amplitude.

[0025] S3 constructs a dynamic benchmark model driven by both physics and data: it acquires a multi-dimensional operating condition vector in real time, consisting of unit load, effective head, guide vane opening, speed, cooling water temperature, and sediment content; and predicts the expected steady-state values ​​of each measuring point under the current operating condition based on a deep residual shrinkage network (DRSN). and adaptive variance of operating conditions Simultaneously, the heat balance equation of the hydro-generator unit and the water hammer effect model are introduced to calculate the predicted values ​​of the physical mechanism. The data-driven loss and physical bias are weighted and fused together as the model training objective, and the loss function is:

[0026] in, These are measured values. For DRSN predicted values, Values ​​calculated for the physical model. is the balance coefficient, and N is the total number of samples.

[0027] S4, Set three-level progressive alarm limits: Warning Limit Alarm limit And a weighted correction is made based on the historical extreme values ​​of the equipment: Ensure that the limit does not exceed 95% of the historical extreme value, and implement emergency protection limits. Refer to the equipment's ultimate strength setting.

[0028] S5, Multi-parameter linkage discrimination based on causal inference: Construct a parameter causality graph based on Granger causality test, and calculate the causal relationship between parameters in real time. When a parameter at a certain measuring point exceeds the warning limit, the linkage verification program is initiated to retrieve the set of associated parameters (Granger causality test) that have a significant causal relationship with that measuring point. The co-evolution trend of the associated parameter set within the same sliding time window is calculated. If more than 60% of the parameters show a shift characteristic synchronized with the master parameter, it is determined to be a true degradation, and an alarm of the corresponding level is triggered according to the degree of exceeding the limit.

[0029] S6, Multi-scale anomaly energy integral: Defines short-period (second-level) anomaly energy and long-period (hour-level) abnormal energy To capture transient impacts and slow evolutionary trends respectively:

[0030]

[0031] when or When this occurs, a trend warning of the corresponding level is triggered; , For the observation period, , This is the sensitivity coefficient.

[0032] S7, Confidence Output and Limit Self-Evolution Based on Bayesian Deep Learning: A Monte Carlo dropout approximation of a Bayesian neural network is used for each predicted value. Output confidence interval Confidence level is defined as .when When the threshold is 90%, the system automatically switches the Level 3 limit to conservative mode: the warning limit is expanded to... The alarm limit is switched to a static limit value. Simultaneously, reinforcement learning (Q-learning) is used to dynamically optimize the sensitivity coefficient based on the confirmation / rejection feedback from operations and maintenance personnel regarding the alarm. , And Granger causality threshold, to achieve limit self-evolution.

[0033] S8, alarm convergence and root cause localization based on causal directed acyclic graph (DAG): When a complex fault causes a large number of concurrent alarms, the system traverses the pre-built fault propagation graph, combines it with the real-time causal graph, identifies the source event in the alarm stream, automatically suppresses derived alarms, and pushes a root cause analysis report.

[0034] Specifically, in this embodiment, a large hydropower station in Southwest China is used as an example. This station is equipped with four 300MW mixed-flow turbine generator units, which undertake the tasks of peak shaving and frequency regulation of the power grid. To verify the technical effect of the hydropower monitoring method and system with graded limit linkage alarm provided by the present invention, Unit #3 was selected as the implementation object for a 12-month field application and test. The following is a detailed description of the specific implementation process.

[0035] S1. System Hardware Deployment and Data Acquisition S1.1. Sensor Array Configuration Multiple types of sensors are deployed in key areas of Unit #3 to build a data sensing layer, with the specific configuration as follows: Table 1: Specific Configuration Data for Deploying Multiple Types of Sensors in Key Components of Unit #3

[0036] All sensor signals are transmitted to the field acquisition unit via shielded twisted-pair cables or optical fibers. To eliminate electromagnetic interference, the signal conditioning modules all employ isolation amplification technology, with input and output isolation voltage ≥3000V, and the signal cables are double-shielded and grounded at a single point.

[0037] S1.2. Clock Synchronization and Data Acquisition A GPS timing synchronization device (model: Beidou / GPS dual-mode timing server) was deployed to provide a unified clock reference for all acquisition cards via IRIG-B code. Testing showed that the synchronization error between each acquisition channel remained stable within 0.6μs, better than the design requirement of 1μs.

[0038] The data acquisition system is implemented using an edge computing layer, employing a NIPXIe-1073 chassis and equipped with an FPGA module (PXIe-7965R) and a real-time controller (PXIe-8840). The FPGA is responsible for parallel acquisition and preprocessing of high-speed channels, while the ARM controller runs a real-time operating system (PharlapETS) to execute dynamic benchmark model inference and linkage discrimination algorithms.

[0039] S2. Data Preprocessing and Feature Extraction S2.1. CEEMDAN noise reduction Taking the X-axis swing signal of a water-guided bearing as an example, the original signal is superimposed with 50Hz power frequency interference and water flow turbulence noise. The CEEMDAN algorithm is used for decomposition: Add Gaussian white noise with an amplitude of 0.2 times the standard deviation to the original signal, and set the number of integrations to 100.

[0040] Twelve intrinsic mode function components, IMF1~IMF12, were obtained through iterative decomposition.

[0041] Calculate the Hearst exponent for each component: For high-frequency noise components (such as IMF1 and IMF2), the Hearst exponent is <0.5, so they are discarded.

[0042] The remaining components were reconstructed to obtain the denoised signal. After denoising, the signal-to-noise ratio was improved from 15.3dB to 28.7dB.

[0043] S2.2. Time-Frequency Domain Feature Extraction The following features are extracted from the denoised signal: Time-domain characteristics: root mean square value (reflecting vibration energy), peak-to-peak value (reflecting maximum swing), kurtosis factor (reflecting impact characteristics).

[0044] Frequency domain characteristics: Perform a 2048-point FFT on the data per second to extract the centroid frequency (reflecting the centroid shift of the spectrum) and the main frequency amplitude (reflecting the frequency shift and harmonic amplitude).

[0045] Feature data is stored in the real-time database of the edge computing layer at 1-second intervals and simultaneously uploaded to the distributed database of the central decision-making layer.

[0046] S3. Construction of a Dynamic Benchmark Model Driven by Both Physics and Data S3.1. Definition of Working Condition Vector Taking into account the key factors affecting unit operation, a 7-dimensional operating condition vector is constructed: ,in: Active power (MW) Effective head (m) Guide vane opening (%) Rotational speed (rpm) Cooling water temperature (°C) Sediment content (kg / m³) Cooling water flow rate (m³ / s) S3.2 Deep Residual Shrinkage Network (DRSN) Structure DRSN is used to predict the expected steady-state values ​​of each measuring point from the load condition vector. The network structure is as follows: Input layer: 7 nodes, corresponding to working condition vectors.

[0047] Residual shrinkage blocks: A total of 6 residual shrinkage blocks are connected in series. Each block contains: Convolutional layer 1: 64 3×1 convolutional kernels, stride 1, activation function ReLU.

[0048] Batch normalization layer.

[0049] Convolutional layer 2: 64 3×1 convolutional kernels, stride 1, activation function ReLU.

[0050] Batch normalization layer.

[0051] The soft thresholding subnetwork performs global average pooling on the feature map, connects to two fully connected layers, and outputs the threshold τ for each channel, which is then passed through a soft thresholding function. Perform characteristic shrinkage.

[0052] Global average pooling layer: reduces the dimensionality of the feature map.

[0053] Fully connected output layer: The number of nodes equals the total number of measurement points (taking the thrust bearing temperature as an example, the output node is 1).

[0054] S3.3 Physical Mechanism Prediction Model To constrain the data-driven model, two physical models are introduced: Model 1: Thrust bearing thermal balance equation: Heat generated by bearing losses: ,in The coefficient of friction (calculated based on the thrust bearing area, oil film thickness, etc., taken as...) ).

[0055] Cooling water removes heat: ,in The specific heat capacity of water, For density, , This refers to the inlet and outlet temperatures of the cooling water.

[0056] steady state The bearing temperature prediction value can be derived:

[0057] Model 2: Water hammer pressure model: Pressure rise during guide vane closure:

[0058] in For the length of the pressure pipe, It is the acceleration due to gravity. This represents the rate of change in flow velocity (related to the rate of change in guide vane opening). This model is used to predict the baseline value for pressure pulsation measurement points.

[0059] S3.4 Joint Training DRSN prediction value Compared with the physical model predictions The fusion loss function is:

[0060] in (Determined through grid search).

[0061] The training data consisted of approximately 31 million samples from the full year of normal operation of Unit #3 in 2019 (excluding start-up and shutdown processes). The Adam optimizer was used with a learning rate of 0.001, a batch size of 256, and 50 iterations. After training, the model's mean absolute error in predicting thrust bearing temperature was 0.32℃, which is superior to both the pure DRSN (0.45℃) and the pure physics model (0.78℃).

[0062] S4. Three-level progressive alarm limit setting Taking the thrust bearing temperature as an example, the dynamic benchmark model outputs the expected steady-state value under the current operating conditions. Adaptive variance of operating conditions .set up: Warning limits:

[0063] Alarm limits: And introduce historical extreme value weighted correction: ,in This is the historical maximum value under this operating condition.

[0064] Emergency protection limits: (Refer to the allowable limit temperature of thrust-type Babbitt alloy).

[0065] In actual operation, , It changes dynamically with operating conditions. For example, under rated operating conditions (P=300MW, H=150m), , ,but , Under low load conditions (P=150MW), , , , .

[0066] S5. Multi-parameter linkage discriminant based on Granger causality test S5.1 Constructing a parametric cause-effect graph Twenty measurement points related to the thrust bearing were selected (thrust bearing temperature, thrust oil sump temperature, lubricating oil pressure, thrust bearing housing vibration, unit speed, active power, etc.), and Granger causality tests were performed on each pair of parameters. A VAR model was used, with the lag order selected according to the AIC criterion (usually 5-10), and the significance level set to [value missing]. .

[0067] Example of test results: Lubricating oil pressure indicates thrust bearing temperature ( The ':' indicates that the change in lubricating oil pressure is a Granger cause of the change in thrust bearing temperature.

[0068] The vibration of the thrust bearing housing points to the temperature of the thrust bearing pad. Vibration changes are the cause of temperature changes.

[0069] The thrust bearing temperature indicates the thrust oil tank temperature. ): Temperature changes cause changes in oil temperature.

[0070] This constructs a causal directed graph, which is then stored in the edge computing layer.

[0071] S5.2 Linkage Verification Process At 14:23:05 on March 15, 20XX, the measured temperature of the thrust bearing reached 61.5℃, exceeding the warning limit for the current operating condition. The system immediately initiates a linked verification: Retrieve the causal nodes pointing to "thrust bearing temperature" from the causal graph, namely "lubricating oil pressure" and "thrust bearing housing vibration".

[0072] Extract the sequence data of three parameters within 1 minute before and after this moment (14:22:35~14:24:05).

[0073] The Pearson correlation coefficients (used for auxiliary verification, mainly based on the cause-effect plot) and trend consistency of the three parameters with the thrust bearing temperature were calculated. The results show: The lubricating oil pressure decreased from 0.32 MPa to 0.28 MPa during the same period, showing a monotonically decreasing trend (inversely correlated with the temperature rise, but in accordance with physical laws: the pressure drop leads to a thinner oil film and a rise in temperature).

[0074] The vibration amplitude of the thrust bearing housing increased monotonically from 1.2 mm / s to 1.5 mm / s.

[0075] Both correlated parameters exhibited offset characteristics synchronized with the main parameter (temperature rise, pressure fall, vibration rise), with a coordination ratio of 100% > 60%.

[0076] The issue was determined to be a genuine degradation trend, triggering a Level 2 alarm (alarm limit). It has not yet exceeded the limit, but a level-two warning has been triggered based on the linkage logic.

[0077] After receiving the alarm, maintenance personnel inspected the system and found that the lubricating oil pump outlet filter was clogged, causing a drop in oil pressure. After replacing the filter, the temperature returned to normal. If a traditional static threshold were used, the temperature would not have exceeded the fixed alarm value (usually set at 75°C), and the fault would have been ignored, potentially developing into a bearing failure.

[0078] S6. Multiscale anomaly energy integral S6.1 Parameter Settings short cycle Seconds are used to capture transient impacts (such as blade impacts or foreign object intrusion).

[0079] Long cycle Hours are used to capture slow evolution (such as bearing wear, decreased cooling efficiency).

[0080] Sensitivity coefficient: , (Initial value, which will be optimized through reinforcement learning).

[0081] S6.2 Calculation Example On July 12, 20XX, intermittent spikes appeared in the pressure pulsation monitoring data of the tailrace pipe. Calculate the short-cycle energy. (Points every 10 seconds):

[0082] Between 03:00 and 04:00 on July 12th, Exceeding the threshold multiple times (Unit: MPa²·s), reaching a maximum of 0.52, triggering a level two alarm. Upon inspection, a loose component was found inside the tailwater pipe; the issue was resolved after the machine was shut down.

[0083] In September 20XX, the thrust bearing temperature slowly rose from 57.5℃ to 59.2℃ within 6 hours, with the instantaneous value consistently below the warning limit for this operating condition. (The expected value output by the dynamic benchmark model under the current operating conditions) Adaptive standard deviation of operating conditions Warning limit Although the instantaneous value did not exceed the limit, the system calculated the long-period abnormal energy:

[0084] By integrating the temperature sampling points over 6 hours (21600 seconds), the temperature rises linearly, and the absolute value of the deviation gradually decreases from an initial 1.0℃ to 0.7℃, with an average absolute deviation of approximately 0.85℃. The threshold value is .because The system triggered a long-term trend warning, indicating "abnormal thrust bearing temperature trend, possibly due to poor lubrication." Maintenance personnel checked and found a slight drop in lubricating oil pressure. After adjustment, the temperature returned to normal, preventing a potential malfunction.

[0085] Using another case: In October 20XX, the X-axis runout of the upper guide bearing slowly increased from 120μm to 160μm over 3 days, with stable operating conditions, and the expected value... Approximately 125μm Calculate the 3-day (72-hour) mean absolute deviation:

[0086] S6.3 Threshold Adaptive Sensitivity coefficient , The initial value is determined through historical data statistics and subsequently optimized by reinforcement learning.

[0087] S7. Bayesian Deep Learning Confidence and Conservative Mode Switching S7.1 Confidence Calculation In DRSN prediction, Monte Carlo Dropout was used: Dropout was enabled during the testing phase (dropout rate 0.1%), and 50 forward propagations were performed to obtain 50 predicted values. Calculate the mean of these 50 predicted values. and standard deviation Then the 95% confidence interval is Confidence level is defined as:

[0088] in, This is the average of multiple predictions. This represents the standard deviation of multiple predictions. This value reflects the uncertainty of the model's predictions. The lower the value, the higher the uncertainty.

[0089] when At that time, it was considered that the model had high prediction uncertainty under the current working conditions, and thus entered a conservative mode.

[0090] S7.2 Conservative Mode Switching Logic In November 20XX, the unit operated at extremely low load (P=80MW), a condition rarely seen in historical data. The model's prediction of thrust bearing temperature... , ,calculate Conservative mode is triggered: Warning limit from Expand to (The actual situation is still relatively lenient).

[0091] Alarm limit value switched to static limit value (The original dynamic alarm limit was) However, the model is unreliable at this point, so a conservative static value is adopted.

[0092] In subsequent actual operation, the temperature remained stable at around 50℃ without any abnormalities, thus avoiding false alarms caused by model uncertainties.

[0093] S7.3 Reinforcement Learning Parameter Optimization Define reinforcement learning state Current operating condition category (ultra-low load, low load, rated load, overload), false alarm count in the last hour, and average confidence level.

[0094] action :Adjustment , Granger causality test The thresholds have step sizes of ±0.1, ±0.1, and ±0.01, respectively.

[0095] award : +1 for confirmation of alarm by maintenance personnel, -1 for false alarm, -2 for missed alarm discovered afterward.

[0096] Q-learning was used with a learning rate of 0.1 and a discount factor of 0.9. The Q-table was updated every 24 hours based on the day's logs. After 3 months of learning, It converged from 1.5 to 1.6. It converged from 2.0 to 1.9. The threshold converged from 0.05 to 0.045, further reducing the false alarm rate by 15%.

[0097] S8. Alarm Convergence and Root Cause Analysis S8.1 Fault Propagation Map Construction Based on the schematic diagram of the hydro-generator system, a directed acyclic graph (DAG) of fault propagation is pre-constructed. For example, a fault in the governor's hydraulic system may lead to uncontrolled guide vanes, which in turn causes a series of secondary faults such as speed fluctuations, power fluctuations, increased vibration, and voltage fluctuations. In the graph, nodes represent fault types, and edges represent propagation relationships.

[0098] S8.2 Concurrent Alarm Handling On January 5, 20XX, a power grid failure caused the generating unit to shed load, instantly triggering 187 alarms. The system initiated alarm convergence. Sort all real-time alerts by timestamp.

[0099] Traversing the fault propagation graph, the source node is identified: a fault in the governor hydraulic system (the actual load shedding is caused by the power grid, but the unit itself is not faulty; this is only a simulation). In a real scenario, power grid faults are not internal unit faults, so there may be no corresponding node in the graph. At this time, the system detects that all alarms occur almost simultaneously, and there is no obvious time-series propagation chain, which is judged as an external disturbance. Only the aggregated display "Power grid disturbance caused unit load shedding" is shown, suppressing derived alarms (such as increased speed, voltage fluctuations, etc.).

[0100] If the fault is internal, such as a series of alarms caused by overheating of the thrust bearing, the root cause is traced back to the decrease in lubricating oil pressure according to the cause-effect diagram (decreased lubricating oil pressure - increased temperature - increased oil temperature - increased vibration), and a message is sent saying "Lubricating oil system fault, it is recommended to check the filter".

[0101] S8.3 Visualization The central decision-making visualization workstation overlays alarm information onto a 3D digital twin model, highlighting the root cause device (such as the lubricating oil pump) in red and the affected devices in yellow, and dynamically displaying the fault propagation path with arrows. Maintenance personnel can view detailed diagnostic reports with a single click.

[0102] S9. System Performance Comparison Test To quantitatively evaluate the effectiveness of this invention, a traditional static threshold system and the adaptive threshold system described in the prior art were simultaneously deployed on Unit #3 and operated in parallel with this invention for 6 months (July 20XX - December 20XX), recording alarm data. The results are as follows: Table 2: Alarm Log Data Table

[0103] Note: Extreme operating conditions refer to data sparse operating conditions (such as ultra-low load, high head anomaly zone).

[0104] Therefore, this embodiment verifies the effectiveness and advancement of the graded limit-linked alarm hydropower monitoring method and system provided by this invention through actual deployment and testing on Unit #3 of a large hydropower station. The physical-data dual-drive model improves the robustness of the dynamic benchmark; Granger causality testing realizes the leap from correlation to causality for multiple parameters; multi-scale anomaly energy integration takes into account both transient and evolutionary faults; Bayesian confidence and conservative mode avoid false alarms caused by model uncertainty; and alarm convergence and root cause analysis significantly improve operation and maintenance efficiency. All indicators are significantly superior to existing technologies, providing solid technical support for intelligent monitoring of hydropower stations.

[0105] Example 2 Please see Figure 3As shown, this embodiment is a further optimization based on Embodiment 1. In this embodiment, the method further includes: the step of solving the problem using an optimization algorithm based on vibration risk constraints, which further includes a rolling optimization process of Model Predictive Control (MPC). A state-space model of unit vibration risk is established, with the state vector xi(k) describing the dynamics of vibration risk of unit i. Its evolution relationship and output equation are as follows:

[0106]

[0107] in For discrete time steps, For the model parameter matrix, For the rate of change of output, For process noise, To observe noise; In each control cycle, based on the current time Based on the actual measured conditions, predict the future. The vibration risk changes within a time step are used to construct a multi-objective optimization problem:

[0108] ST power balance constraints, output upper and lower limit constraints, ramp rate constraints, and risk threshold constraints; in To predict the time domain, To control the time domain, This is the water consumption rate coefficient. As a risk penalty weight, To smooth out changes in output, a penalty weight is applied. Solving the above optimization problem yields the optimal control sequence, taking only the control input at the first time step. Issue and execute; In the next control cycle, update the measured status and start rolling optimization again.

[0109] The state space model parameters Obtained through system identification methods, specifically including: Collect historical operating data of the generating units, including output sequences. and the corresponding comprehensive vibration risk coefficient ; The data is preprocessed to remove outliers and then normalized. The subspace identification algorithm (such as N4SID) is used to estimate the state-space model parameters, and the model order is determined by the AIC criterion. The identified model is used to predict future risks, and the state estimation is corrected online by Kalman filtering to improve the prediction accuracy.

[0110] Specifically, this embodiment takes four generating units (numbered G1~G4) of a hydropower station as an example, with a total installed capacity of 1000MW, participating in AGC frequency regulation. Model predictive control (MPC) is used to achieve dynamic load allocation considering vibration risk, with a control cycle of... minute.

[0111] Step 1: Identification of the state-space model for vibration risk Collect historical operating data from the past three months, including the output per minute for each unit. and the corresponding comprehensive vibration risk coefficient (Calculated from the Direction 1 model). Continuous operating data under normal conditions, totaling 20,000 samples, were selected. The N4SID subspace identification algorithm was used, with the model order set to 2, to obtain the state-space model parameters for each unit. Taking G1 as an example, the identification results are as follows:

[0112] The model output equation is ,in This represents the state vector. Model validation: Using the remaining 20% ​​of the data for testing, the root mean square error of the prediction is 0.042, which meets the control requirements.

[0113] Step 2: MPC Optimization Problem Construction Set the prediction time domain Minutes, controlling the time domain Minutes. Current moment. Total load command of the power grid MW. The current output and risk factor of each unit are as follows: Table 3: Current Output and Risk Factors of Units G1-G4

[0114] Water consumption rate coefficient (m³ / kWh) determined based on the current water head: G1: 3.2, G2: 3.1, G3: 3.3, G4: 3.0. Risk penalty weights. Smoothing weights for output changes .

[0115] MPC optimization issues:

[0116] Constraints: Power balance: Future load instructions are provided by dispatch forecasts, assuming the load remains unchanged (850MW) for the next 15 minutes.

[0117] Upper and lower limits of output:

[0118] Slope rate:

[0119] Risk threshold: (Mandatory constraints on G3) Step 3: Rolling optimization solution The above problem is solved using a quadratic programming (QP) algorithm, with an upper limit of 100 iterations and a convergence accuracy of [missing information]. The optimal control sequence is obtained by solving the problem (only the output command for the first time step is shown): Table 4: Output Data of Units G1-G4

[0120] The total output is 225 + 215 + 195 + 215 = 850 MW, which meets the requirements. The high-risk unit G3 was appropriately reduced in load, and the other units shared the load, thus controlling the overall risk.

[0121] Step 4: State Update and Scrolling Issue and execute the above instructions. One minute later, collect new tilt angle data and update the actual risk coefficients for each unit. Assuming the actual risk of G3 drops to 0.60 and G1 rises to 0.38, reconstruct the MPC optimization problem and repeat steps 2-3.

[0122] Step 5: Online calibration of Kalman filter To improve prediction accuracy, a Kalman filter is introduced into the state-space model. Let the state equation and observation equation be:

[0123]

[0124] in , , Noise covariance estimated using offline data. , In each control cycle, the state estimate is first updated using Kalman filtering, and then substituted into MPC optimization to improve the accuracy of the state estimate.

[0125] The MPC method in this embodiment is compared with the single static optimization method in Embodiment 1. Under the same load fluctuation scenario (continuous 24-hour operation, including 3 load jumps), the statistical results are as follows: Table 5: Statistical data under the same load fluctuation scenario

[0126] MPC, through rolling optimization, anticipates risk changes in advance, preventing high-risk units from bearing excessive load in subsequent periods. It also smooths out output changes, reducing transient impacts and achieving a synergistic improvement in safety, economy, and stability. Especially after load surges, MPC can adjust allocation schemes in advance, reducing the exposure time of high-risk units by more than 70%, significantly enhancing the proactive risk control capabilities of the AGC system.

[0127] This embodiment enhances the physical mechanism of the vibration risk model, providing more accurate risk prediction input for MPC in direction two. For example, the mixed risk coefficient output in direction one... It can be directly used for state-space model identification in Direction 2 to improve model accuracy, while the MPC framework utilizes the real-time risk provided in Direction 1, combined with a dynamic evolution model, to achieve forward-looking control. The two are tightly coupled through the "model-prediction-control" chain, forming a complete closed loop from perception to decision-making.

[0128] The two approaches together address the two major pain points in Example 1: the lack of physical basis in the risk model and the lack of predictability in optimization decisions. This upgrades AGC load allocation from passive response to proactive prevention, significantly improving the flexibility and safety of hydropower stations in scenarios with a high proportion of renewable energy access.

[0129] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A hydropower monitoring method with graded limit-linked alarm, characterized in that, include: Acquire synchronously collected multi-source sensor data from the hydropower unit, wherein the multi-source sensor data includes at least vibration, sway, temperature, pressure, and electrical parameters; The multi-source sensor data is preprocessed to extract time-frequency domain features; A dynamic benchmark model driven by both physics and data is constructed. The dynamic benchmark model integrates data-driven prediction from a deep residual shrinkage network with physical mechanism prediction based on the thermal balance equation and water hammer effect model, and outputs the expected steady-state value and adaptive variance of each measuring point under the current working condition. Based on the expected steady-state value and the adaptive variance of the operating condition, a three-level progressive alarm limit is set, which includes a warning limit, an alarm limit, and an emergency protection limit. Based on the Granger causality test, a parameter causality graph is constructed. When a parameter at a certain measuring point exceeds the warning limit, the set of associated parameters that have a significant causal relationship with the measuring point is retrieved, and the co-evolution trend of the associated parameter set within the same sliding time window is calculated. If it exceeds a preset ratio, it is determined to be a real deterioration, and an alarm of the corresponding level is triggered according to the degree of exceeding the limit. Calculate the multi-scale anomaly energy integral, which includes short-period anomaly energy and long-period anomaly energy. When the short-period anomaly energy or the long-period anomaly energy exceeds the corresponding threshold, a trend warning is triggered. The confidence level of the expected steady-state value is output using Bayesian deep learning. When the confidence level is lower than a preset confidence threshold, the alarm limit is switched to conservative mode. When concurrent alarms occur, based on the pre-built fault propagation graph and the parameter cause-effect graph, the source alarm is identified and the derived alarm is suppressed, and a root cause analysis report is pushed. By using reinforcement learning, the sensitivity coefficients of the short-cycle and long-cycle anomaly energies, as well as the significance level threshold of the Granger causality test, are dynamically optimized based on the confirmation or rejection feedback from maintenance personnel regarding the alarms.

2. The hydropower monitoring method with graded limit linkage alarm as described in claim 1, characterized in that, In constructing the dynamic benchmark model driven by both physics and data, the dynamic benchmark model integrates data-driven prediction from a deep residual shrinkage network with physical mechanism prediction based on the thermal balance equation and water hammer effect model, outputting the expected steady-state value and adaptive variance of each measuring point under the current operating condition. The training of the dynamic benchmark model driven by both physics and data aims to minimize the error between data-driven predictions and measured values, as well as the deviation between data-driven predictions and physical mechanism predictions. The weights of the two are adjusted by a balancing coefficient.

3. The hydropower monitoring method with graded limit linkage alarm as described in claim 1, characterized in that, In the parameter causality graph constructed based on Granger causality test, when a parameter at a certain measuring point exceeds the warning limit, the set of associated parameters that have a significant causal relationship with that measuring point is retrieved, and the co-evolution trend of the associated parameter set within the same sliding time window is calculated. If it exceeds a preset proportion, it is determined to be a true degradation, and an alarm of the corresponding level is triggered according to the degree of exceeding the limit. The significance level for the Granger causality test is set to be less than 0.05, and the set of associated parameters includes parameters that have a Granger causal relationship with the parameter that exceeds the limit and whose significance level is less than 0.05; The preset ratio is 60%, and the co-evolution trend refers to the proportion of parameters whose associated parameters exhibit a monotonically increasing or decreasing characteristic in sync with the main parameters.

4. The hydropower monitoring method with graded limit linkage alarm as described in claim 1, characterized in that, The calculation of multi-scale anomaly energy integrals includes short-period anomaly energy and long-period anomaly energy. When either the short-period or long-period anomaly energy exceeds a corresponding threshold, a trend warning is triggered. The short-period anomaly energy is obtained by integrating the squared residuals between the measured values ​​and the expected steady-state values ​​within a preset short-period observation period. The long-period anomaly energy is obtained by integrating the absolute value of the difference between the rate of change of the measured value and the rate of change of the expected steady-state value within a preset long-period observation period. The corresponding threshold is calculated and determined based on the adaptive variance of the operating condition, the observation duration, and the preset sensitivity coefficient.

5. A hydropower monitoring method with graded limit linkage alarm as described in claim 1, characterized in that, The system uses Bayesian deep learning to output the confidence level of the expected steady-state value. When the confidence level is lower than a preset confidence threshold, the alarm limit is switched to a conservative mode. The confidence level is calculated based on the upper and lower limits of the prediction interval output by Bayesian deep learning and the expected steady-state value. The preset confidence threshold is 90%. The conservative mode includes expanding the warning limit to the expected steady-state value plus or minus five times the adaptive variance of the operating condition, or switching the alarm limit to the static limit value.

6. A hydropower monitoring method with graded limit linkage alarm as described in claim 1, characterized in that, The three-level progressive alarm limit is set based on the expected steady-state value and the adaptive variance of the operating condition. The three-level progressive alarm limit includes a warning limit, an alarm limit, and an emergency protection limit. The three-level progressive alarm limit setting method is as follows: the early warning limit is equal to the expected steady-state value plus twice the operating condition adaptive variance, the alarm limit is equal to the expected steady-state value plus three times the operating condition adaptive variance and weighted correction based on the equipment's historical operating extreme values, and the emergency protection limit is set with reference to the instantaneous limit strength allowed by the equipment's design life.

7. A graded limit-linked alarm system for hydropower monitoring, characterized in that, To implement the method according to any one of claims 1 to 7, comprising: The data sensing layer consists of a sensor array deployed at key parts of the hydropower unit, used to synchronously collect multi-source sensor data. The sensor array includes a swing sensor, a temperature sensor, a pressure pulsation sensor, an embedded thermal resistor sensor, and an acoustic emission sensor. All sensor channels are connected to a clock synchronization module based on the Global Positioning System timing protocol, with a synchronization error of less than 1 microsecond. The edge computing layer includes an embedded real-time processing terminal, which is equipped with an ARM processor and FPGA collaborative processing architecture. The FPGA is used for real-time acquisition, filtering and fast Fourier transform of high-frequency signals, and the ARM processor is used to run dynamic benchmark prediction models, multi-parameter linkage discrimination algorithms and multi-scale anomaly energy integral calculation. The central decision-making layer includes a high-performance computing server cluster, a distributed database, and a visualization interactive workstation. The server cluster is used to train deep residual shrinkage networks and physical and data-driven models. The distributed database is used to store raw waveform data and alarm event logs. The visualization interactive workstation is used for 3D digital twin display and root cause analysis push. The closed-loop feedback calibration module is used to monitor the handling results of alarms by maintenance personnel in real time, and to optimize the sensitivity coefficient and Granger causality test threshold using reinforcement learning algorithms.

8. A graded limit-linked alarm hydropower monitoring system according to claim 7, characterized in that, The edge computing layer also integrates a miniaturized real-time database to cache high-frequency raw data from the past 24 hours, and automatically locks the data for that time period to form a black box data packet when an emergency protection alarm is triggered and uploads it to the central decision-making layer.

9. A graded limit-linked alarm hydropower monitoring system according to claim 7, characterized in that, The system's communication network adopts a ring fiber optic Ethernet topology, conforms to the IEC61850 standard, supports media redundancy protocol, and ensures that the network single-point failure recovery time is less than 50 milliseconds.

Images