A method and system for early warning of hydroelectric generator set failure

Abstract

This invention discloses a fault early warning method for hydro-generator units, belonging to the field of fault diagnosis technology for hydro-generator units. It includes: collecting multi-source heterogeneous monitoring data and extracting multi-dimensional features; establishing a hydraulic-mechanical-electromagnetic multi-physics coupling mechanism model; embedding the mechanism model as a physical constraint into a deep learning network to construct a physical information neural network early warning model, whose loss function includes a data-driven loss term and a physical constraint loss term; using the trained model to calculate multi-dimensional health indicators online, and generating dynamic early warning thresholds based on adaptive operating conditions; after triggering an early warning, automatically locating the root cause of the fault through a fault causal graph and causal inference algorithm; and using a temporal convolutional network to predict the changing trend of health indicators and output hierarchical early warning information. This invention integrates physical mechanisms and data-driven approaches, achieving accurate early fault perception, adaptive early warning based on operating conditions, and automatic root cause location, significantly improving the accuracy and timeliness of early warnings.

Description

Technical Field

[0001] This invention relates to the field of hydro-generator condition monitoring and fault diagnosis technology, specifically to a method and system for early warning of faults in hydro-generators. Background Technology

[0002] As the core equipment of hydropower generation, the safe and stable operation of hydro-turbine generator units is directly related to the safety of the power grid and the reliability of the power system. With the rapid development of my country's hydropower industry, the installed capacity and single unit capacity of hydro-turbine generator units are becoming larger and larger, the unit structure is becoming more complex, and the operating conditions are more variable, which puts forward higher requirements for the fault early warning capabilities of the units.

[0003] Currently, the fault early warning system for hydro-generator units mainly suffers from the following technical deficiencies:

[0004] First, most existing fault early warning methods adopt a purely data-driven approach, using machine learning models such as neural networks and support vector machines to perform pattern recognition on collected vibration, temperature, and other signals. While these methods can identify faults to some extent, they lack in-depth modeling of the hydraulic-mechanical-electromagnetic multi-physics coupling mechanism of hydro-generator units, resulting in poor model generalization ability and a significant decrease in early warning accuracy under varying operating conditions and small sample scenarios. Especially when the unit's operating conditions change frequently (different heads, different loads), the false alarm rate and false negative rate of the model both increase significantly.

[0005] Second, most existing early warning methods are based on alarm strategies using fixed thresholds or simple operating condition zoning. Fixed thresholds cannot adapt to the dynamic changes in unit operating conditions, easily leading to frequent false alarms under non-fault conditions, while false alarms are missed in the early stages of actual faults due to excessively high threshold settings. Although some methods have proposed the idea of ​​operating condition zoning, the zoning method is crude and cannot achieve continuous and smooth adaptive adjustment of thresholds according to operating conditions.

[0006] Third, most existing methods can only determine "abnormal equipment status" but cannot clarify the causal relationship between the fault and the symptoms, nor can they automatically locate the root cause of the fault. In actual operation and maintenance, after being notified of "abnormal equipment," maintenance personnel still need to spend a lot of time on manual analysis and troubleshooting, making it difficult to quickly locate the source of the fault and take targeted measures.

[0007] Fourth, existing methods are insufficient in detecting early, subtle fault characteristics. Many faults in hydro-generator units (such as initiation of cavitation and micro-wear of guide bearings) have extremely weak characteristic signals in their nascent stages, which are drowned out by strong background noise. Existing methods struggle to capture these subtle signal changes during the fault's early stages, often triggering warnings only when the fault characteristics have become more apparent, thus missing the optimal maintenance window.

[0008] To address the aforementioned technical issues, there is an urgent need for a fault early warning method for hydro-generator units that can integrate physical mechanisms and data-driven approaches, achieve early detection of weak features, adapt to changes in operating conditions, and possess the ability to locate the root cause of faults. Summary of the Invention

[0009] To achieve accurate detection, adaptive early warning, and root cause localization of faults in hydro-generator units, and to improve the accuracy of fault early warning, this invention provides a fault early warning method and system for hydro-generator units, employing the following technical solution:

[0010] A method for early warning of faults in a hydro-generator unit includes:

[0011] Step S1: Collect multi-source heterogeneous monitoring data of the hydro-generator unit, preprocess the monitoring data, extract time-domain features and frequency-domain features, and construct a multi-dimensional feature vector;

[0012] Step S2: Based on the structural design parameters and operating environment parameters of the hydro-generator unit, establish a mechanism model of hydraulic-mechanical-electromagnetic multi-physics coupling, and generate theoretical reference values ​​for each monitored physical quantity under different operating conditions through the mechanism model;

[0013] Step S3: The mechanism model described in step S2 is used as a physical constraint and embedded into a deep learning network framework to construct a physical information neural network early warning model. The loss function of the physical information neural network early warning model includes a data-driven loss term and a physical constraint loss term. The physical information neural network early warning model is trained by minimizing the loss function.

[0014] Step S4: Using the trained physical information neural network early warning model, perform online inference on the real-time collected monitoring data, calculate multi-dimensional health indicators, determine dynamic early warning thresholds based on the current operating parameters of the unit, and trigger a primary early warning when any health indicator is lower than the corresponding dynamic early warning threshold.

[0015] Step S5: When the primary warning is triggered, construct the fault cause-effect graph of the hydro-generator unit, input the current abnormal monitoring data into the fault cause-effect graph, calculate the posterior probability of each candidate fault cause through the causal inference algorithm, and output the fault root cause sorting list according to the posterior probability from high to low.

[0016] Step S6: Based on the time series of health indicators within a continuous time window, a temporal convolutional network is used to predict the future trend of health indicators. The warning level is determined according to the current value and predicted value of the health indicators, and the warning information including fault type, root cause of fault, warning level and recommended maintenance measures is output.

[0017] Optionally, the multi-source heterogeneous monitoring data in step S1 includes vibration signals, sway signals, acoustic emission signals, temperature signals, pressure pulsation signals, electrical signals, and operating parameters; the preprocessing includes denoising the original signal using an adaptive noise complete set empirical mode decomposition method.

[0018] Optionally, the loss function of the physical information neural network early warning model in step S3 is:

[0019] ;

[0020] in, This is the set of trainable parameters for a neural network. This is the data-driven loss term, representing the deviation between the model's predicted values ​​and the actual monitored values; The physical constraint loss term represents the degree to which the model's predicted value deviates from the theoretical reference value of the mechanism model described in step S2; This is the physical constraint weighting coefficient, with a value range of 0.01-0.5;

[0021] The data-driven loss term The expression is:

[0022] ;

[0023] Where N is the number of training samples; Let be the input feature vector of the i-th training sample; For neural networks in input and parameters The predicted output is as follows; This represents the actual monitoring value of the i-th training sample. Represents the square of the Euclidean norm;

[0024] The physical constraint loss term The expression is:

[0025] ;

[0026] Where M is the number of physical constraint sampling points; This is the input feature vector of the j-th physical constraint sampling point; For neural networks in input and parameters The predicted output is as follows; For the mechanism model described in step S2, in the input The theoretical reference value is calculated below.

[0027] Optionally, the health indicators in step S4 include vibration health, temperature health, stability health, and comprehensive health, with each health indicator ranging from 0 to 100. The method for determining the dynamic early warning threshold is as follows: based on the current operating parameters of the unit, retrieve the historical distribution of health indicators under the same operating conditions from historical health data, calculate the probability distribution of health indicators under the operating conditions using the kernel density estimation method, and use the 5th percentile of the probability distribution as the dynamic early warning threshold.

[0028] Optionally, the fault cause-effect graph in step S5 uses fault type as the dependent variable node, each monitored physical quantity as the independent variable node, and the causal relationship driven by physical mechanism as the edge; the causal inference algorithm is a Bayesian network inference algorithm.

[0029] Optionally, the warning level in step S6 includes four levels: healthy, alert, warning, and danger; when the overall health level is greater than or equal to 80, it is judged as healthy; when the overall health level is greater than or equal to 60 and less than 80, it is judged as alert; when the overall health level is greater than or equal to 40 and less than 60, it is judged as warning; and when the overall health level is less than 40, it is judged as danger.

[0030] Optionally, the mechanism model in step S2 includes a turbine hydraulic model, a generator electromagnetic model, and a rotor-bearing system dynamic model; the turbine hydraulic model is established based on computational fluid dynamics, the generator electromagnetic model is established based on the finite element method, and the rotor-bearing system dynamic model is established based on the finite element method.

[0031] Optionally, the time-domain features in step S1 include the signal's peak value, root mean square value, peak-to-peak value, margin factor, and kurtosis; the frequency-domain features include the frequency shift amplitude, harmonic amplitude, half-frequency amplitude, spectral centroid frequency, and spectral root mean square frequency.

[0032] Secondly, the present invention provides a fault early warning system for a hydro-generator set, comprising:

[0033] The data acquisition and preprocessing module is configured to acquire multi-source heterogeneous monitoring data at a predetermined monitoring location of the hydro-generator unit, preprocess the monitoring data and extract time-domain and frequency-domain features to construct a multi-dimensional feature vector;

[0034] The mechanism model construction module is configured to establish a mechanism model of hydraulic-mechanical-electromagnetic multi-physics coupling based on the structural design parameters and operating environment parameters of the hydro-generator unit, and generate theoretical reference values ​​of each monitored physical quantity under different operating conditions through the mechanism model;

[0035] The physical information neural network early warning module is communicatively connected to the data acquisition and preprocessing module and the mechanism model construction module, respectively. It is configured to embed the mechanism model as a physical constraint into a deep learning network framework to construct a physical information neural network early warning model. The loss function of the physical information neural network early warning model includes a data-driven loss term and a physical constraint loss term. The physical information neural network early warning model is trained by minimizing the loss function, and the trained physical information neural network early warning model is used to perform online inference on the real-time collected monitoring data and output multi-dimensional health indicators.

[0036] The dynamic early warning threshold generation module is communicatively connected to the physical information neural network early warning module. It is configured to retrieve the historical distribution of health indicators under the same operating conditions from historical health data based on the current operating condition parameters of the unit, calculate the probability distribution of health indicators under the same operating conditions using the kernel density estimation method, and use the predetermined quantile of the probability distribution as the dynamic early warning threshold. When any health indicator is lower than the corresponding dynamic early warning threshold, a primary early warning is triggered.

[0037] The causal inference root cause localization module is communicatively connected to the dynamic early warning threshold generation module. It is configured to construct a fault causal graph of the hydro-generator unit after the primary early warning is triggered, input the current abnormal monitoring data into the fault causal graph, calculate the posterior probability of each candidate fault cause through the causal inference algorithm, and output the fault root cause sorting list according to the posterior probability from high to low.

[0038] The trend prediction and graded early warning output module is communicatively connected to the causal inference and root cause localization module. It is configured to predict the future trend of health indicators based on the time series of health indicators within a continuous time window, and to determine the early warning level based on the current value and predicted value of the health indicators. It outputs early warning information including fault type, fault root cause, early warning level and recommended maintenance measures.

[0039] The central control module is communicatively connected to the data acquisition and preprocessing module, the mechanism model construction module, the physical information neural network early warning module, the dynamic early warning threshold generation module, the causal inference root cause location module, and the trend prediction and hierarchical early warning output module, and is configured to coordinate the data flow and timing control between the modules.

[0040] In summary, the present invention has at least one of the following beneficial technical effects:

[0041] First, by constructing the PINN early warning model, which integrates physical mechanisms and data-driven approaches, the coupling mechanism of hydraulic, mechanical, and electromagnetic multiphysics fields of hydro-generator units is embedded as a physical constraint into the deep learning framework. This solves the technical problem of poor generalization ability of purely data-driven models under varying operating conditions and small sample scenarios. The presence of physical constraints allows the model to maintain adherence to physical laws even with insufficient training data, significantly improving the early warning accuracy of the model under different operating conditions.

[0042] Second, by using a dynamic threshold generation method based on working condition adaptation, the probability distribution of health indicators under the current working condition is calculated using kernel density estimation and the quantile is used as the threshold. This achieves continuous and smooth adaptive adjustment of the warning threshold according to the working condition, overcoming the problems of high false alarm rate and false negative rate caused by fixed threshold and coarse working condition partitioning.

[0043] Third, by constructing a fault cause-effect graph and introducing a cause-effect inference algorithm, an end-to-end integrated early warning system from "anomaly detection" to "root cause localization" is realized. It can automatically output a fault root cause ranking list, which greatly shortens the fault investigation time and improves operation and maintenance efficiency.

[0044] Fourth, through comprehensive evaluation of multi-dimensional health indicators and trend prediction based on TCN, it is possible to perceive early-stage weak fault characteristics and predict fault development trends, and issue early warnings in the early stages of faults, thus providing maintenance personnel with sufficient maintenance window time. Attached Figure Description

[0045] Figure 1 This is a flowchart of the method in Embodiment 1 of this application;

[0046] Figure 2 This is a system diagram of Embodiment 2 of this application. Detailed Implementation

[0047] The following combination Figures 1-2 The present invention will be described in further detail below.

[0048] Example 1: Refer to Figure 1 This application provides a method for early warning of faults in a hydro-generator set, including:

[0049] Step S1: Acquisition and preprocessing of multi-source heterogeneous monitoring data:

[0050] Various types of sensors are installed at predetermined monitoring locations on the hydro-generator unit to collect multi-source heterogeneous monitoring data of the unit's operation in real time. The monitoring data includes vibration signals, sway signals, acoustic emission signals, temperature signals, pressure pulsation signals, electrical signals, and operating parameters.

[0051] The collected raw monitoring data is preprocessed, which includes: using the adaptive noise complete set empirical mode decomposition (CEEMDAN) method to denoise the raw signal; and extracting time-domain features and frequency-domain features from the denoised signal to construct a multi-dimensional feature vector.

[0052] Step S2: Construct a multiphysics coupling mechanism model for the hydro-generator unit:

[0053] Based on the structural design parameters and operating environment parameters of the hydro-generator unit, a mechanism model of hydraulic-mechanical-electromagnetic multiphysics coupling is established. The mechanism model includes a hydro-turbine hydraulic model, a generator electromagnetic model, and a rotor-bearing system dynamic model.

[0054] The hydraulic model of the turbine is used to describe the flow characteristics of water flowing through the turbine runner and the force exerted on the runner; the electromagnetic model of the generator is used to describe the electromagnetic force characteristics between the generator stator and rotor; and the dynamic model of the rotor-bearing system is used to describe the vibration characteristics of the rotor under bearing support.

[0055] Based on the aforementioned mechanism model, a benchmark for the normal operating status of the unit under different operating conditions is established, and theoretical reference values ​​for each monitored physical quantity are generated.

[0056] Step S3: Construct a physical information neural network early warning model;

[0057] The mechanistic model constructed in step S2 is used as a physical constraint and embedded into a deep learning network framework to construct a Physics-Informed Neural Network (PINN) early warning model. The loss function of the PINN early warning model includes a data-driven loss term and a physical constraint loss term.

[0058] ;

[0059] in, For the total loss function, This is the set of trainable parameters for a neural network, including the weight matrices and bias vectors of each layer. This is the data-driven loss term, representing the deviation between the model's predicted values ​​and the actual monitored values; The physical constraint loss term represents the degree to which the model's predicted value deviates from the theoretical reference value of the mechanism model described in step S2; This is the physical constraint weighting coefficient, used to adjust the proportion of physical constraints in the total loss, with a value range of 0.01-0.5;

[0060] The data-driven loss term The expression is:

[0061] ;

[0062] Where N is the number of training samples; The input feature vector for the i-th training sample includes operating parameters such as unit active power, working head, and guide vane opening, as well as multi-dimensional features of the pre-processed sensor monitoring data. For neural networks in input and parameters The predicted output below represents the predicted value for the unit's health status; The actual monitoring value (label) of the i-th training sample is derived from historical normal operation data and known fault data; Represents the square of the Euclidean norm;

[0063] The physical constraint loss term The expression is:

[0064] ;

[0065] Where M is the number of physical constraint sampling points; This is the input feature vector of the j-th physical constraint sampling point; For neural networks in input and parameters The predicted output is as follows; For the mechanism model described in step S2, in the input The theoretical reference value obtained from the calculation represents the theoretical expected value of this physical quantity under normal operating conditions of the unit.

[0066] By minimizing the total loss function The PINN early warning model is trained so that it can fit the actual monitoring data and follow the constraints of the physical mechanism, thus obtaining the trained PINN early warning model.

[0067] Step S4: Calculate multi-dimensional health indicators and generate dynamic early warning thresholds;

[0068] Using the trained PINN early warning model, online inference is performed on the real-time collected monitoring data to determine the degree of deviation between the current state of the computer group and the health baseline state, thereby obtaining multi-dimensional health indicators.

[0069] The health indicators include vibration health, temperature health, stability health, and overall health. Each health indicator ranges from 0 to 100, with higher values ​​indicating healthier equipment.

[0070] Based on the unit's current operating parameters (active power and head), historical health index distributions under the same or similar operating conditions are retrieved from historical health data. The probability distribution of the health index under this operating condition is calculated using kernel density estimation. The 5th percentile of this probability distribution is used as the dynamic early warning threshold. When any health index falls below the corresponding dynamic early warning threshold, a primary early warning is triggered.

[0071] Step S5: Root cause localization of the fault based on causal inference;

[0072] Upon triggering a primary warning, the fault root cause localization process is initiated. A fault cause-effect graph of the hydro-generator unit is constructed, wherein the fault type is the dependent variable node, each monitored physical quantity is the independent variable node, and the causal relationship driven by the physical mechanism is the edge, describing the causal chain between the fault and the monitored symptoms.

[0073] The current anomaly monitoring data is input into the fault cause-effect graph. The posterior probability of each candidate fault cause is calculated by the causal inference algorithm. The fault root cause sorting list is output according to the posterior probability from high to low, so as to realize the automatic location of the fault root cause.

[0074] Step S6: Fault Trend Prediction and Graded Early Warning Output:

[0075] Based on the time series of health indicators within the continuous time window in step S4, a Temporal Convolutional Network (TCN) is used to predict the future trend of health indicators and obtain the predicted value of health indicators within the predetermined time window in the future.

[0076] Based on the current and predicted values ​​of the health index, the warning level is divided into four levels: Healthy (overall health ≥ 80), Caution (60 ≤ overall health < 80), Warning (40 ≤ overall health < 60), and Danger (overall health < 40). When the warning level reaches Warning or above, a warning message including the fault type, root cause of the fault, warning level, and recommended maintenance measures is output.

[0077] Example 2: This example uses a 300MW mixed-flow turbine generator unit at a hydropower station as an example to illustrate the specific implementation process of the turbine generator unit fault early warning method of the present invention.

[0078] S1: Monitoring system deployment and data acquisition:

[0079] Sensors should be installed at the following key locations on the hydro-generator unit:

[0080] (1) Install three piezoelectric acceleration vibration sensors perpendicular to each other (X, Y and Z directions) on the upper frame, lower frame and stator base respectively, with the sampling frequency set to 2560Hz, to collect the vibration signal of the unit.

[0081] (2) Install X-axis and Y-axis eddy current swing sensors at the upper guide, lower guide and water guide positions of the generator shaft, respectively, and set the sampling frequency to 2560Hz to collect the radial swing signal of the shaft.

[0082] (3) Install acoustic emission sensors at the tailwater inlet gate and the volute inlet, respectively, with a sampling frequency of 1MHz, to collect acoustic emission signals during the initial cavitation stage.

[0083] (4) Install Pt100 platinum resistance temperature sensors at each guide bearing bush, and set the sampling frequency to 1Hz to collect the bush temperature signal.

[0084] (5) Install a pressure pulsation sensor on the inner wall of the tailwater pipe and set the sampling frequency to 1000Hz to collect the pressure fluctuation signal of the tailwater pipe.

[0085] (6) Install voltage transformers and current transformers at the generator outlet, and set the sampling frequency to 12800Hz to collect the voltage and current signals of the generator.

[0086] (7) Read the active power, reactive power, working head, guide vane opening, excitation current and other operating parameters from the local control unit (LCU) of the unit, and set the sampling frequency to 1Hz.

[0087] All sensor data is transmitted to a data acquisition server via industrial Ethernet for time synchronization and data aggregation.

[0088] S2: Data Preprocessing

[0089] The collected raw monitoring data undergoes preprocessing, specifically including:

[0090] (1) Signal denoising: The original vibration signal, sway signal and acoustic emission signal are decomposed by the adaptive noise complete set empirical mode decomposition (CEEMDAN) method to obtain several intrinsic mode function (IMF) components. The correlation coefficient between each IMF component and the original signal is calculated, and noise components with correlation coefficients lower than 0.1 are removed. The remaining IMF components are then reconstructed to obtain the denoised signal.

[0091] (2) Temporal Feature Extraction: For the denoised vibration and sway signals, the following temporal feature parameters are calculated: peak value (difference between the maximum and minimum signal values), root mean square (RMS) value (square root of the mean of the squared signal values), peak-to-peak value (difference between the maximum and minimum signal values), margin factor (ratio of peak value to RMS value), and kurtosis (ratio of the fourth central moment to the squared variance). The formula for calculating the RMS value is as follows:

[0092] Where T is the number of signal sampling points; s(t) is the signal amplitude at time t.

[0093] 3) Frequency domain feature extraction: Perform Fast Fourier Transform (FFT) on the noise-reduced vibration signal and swing signal to extract the amplitude at characteristic frequencies such as rotational frequency (50Hz, corresponding to the rated speed of the unit 3000r / min), harmonic frequency (2×, 3×, ..., 10× rotational frequency), half frequency (0.5× rotational frequency), as well as the centroid frequency and root mean square frequency of the spectrum.

[0094] (4) Acoustic emission signal feature extraction: Calculate the entropy rate of low-frequency cavitation acoustic signal and the instantaneous energy value of high-frequency cavitation acoustic signal for acoustic emission signal.

[0095] S3: Construction of a multiphysics coupling mechanism model;

[0096] Based on the structural design parameters of the 300MW mixed-flow turbine generator unit, a multiphysics coupling mechanism model is constructed:

[0097] (1) Turbine hydraulic model: Based on computational fluid dynamics (CFD) method, a three-dimensional model of the entire turbine flow channel was established. The SST k-ω turbulence model was used for numerical simulation to calculate the velocity field and pressure field at the inlet and outlet of the runner under different combinations of head and guide vane opening, and then the radial force and axial thrust of the water flow on the runner were calculated. The boundary conditions of the CFD simulation include: given flow rate or velocity at the inlet, given pressure at the outlet, and no-slip boundary conditions on the wall.

[0098] (2) Electromagnetic model of generator: Based on the finite element method, a two-dimensional transient electromagnetic field model of generator is established, and the electromagnetic force distribution under different excitation current and load conditions is calculated to obtain the radial electromagnetic tension between the stator and rotor.

[0099] (3) Dynamic model of rotor-bearing system: The dynamic model of rotor-bearing system is established based on the finite element method. The hydraulic excitation force calculated by the hydraulic model of the turbine and the electromagnetic force calculated by the electromagnetic model of the generator are used as external excitation inputs. The theoretical vibration response of the rotor under each working condition is calculated, and the theoretical vibration amplitude and frequency components at each monitoring position are obtained.

[0100] Based on the above mechanism model, the theoretical reference values ​​of various monitored physical quantities of the computer unit under various operating conditions within the range of 30% to 100% rated load and water head within the range of 80m to 120m are obtained. And construct a mapping database between operating parameters and theoretical reference values.

[0101] S4: Training the PINN early warning model;

[0102] The number of features in the input layer are 256, 128, 64, and 32, respectively, and the ReLU activation function is used. The input feature vector of the input layer... It includes 86 features, such as: active power (normalized to [0,1]), operating head (normalized to [0,1]), guide vane opening (normalized to [0,1]), root mean square value of vibration signal at each monitoring point (normalized), peak-to-peak value of swing signal at each monitoring point (normalized), temperature value of each bearing (normalized), entropy rate of acoustic emission signal (normalized), and pressure pulsation amplitude (normalized). The output layer outputs an 8-dimensional vector, corresponding to vibration health, swing health, temperature health, acoustic emission health, pressure health, electrical health, stability health, and overall health.

[0103] The theoretical reference value calculated by the mechanism model in step three. As a physical constraint, it is embedded in the loss function. The weight coefficient of the physical constraint is selected as λ=0.1. The historical operating data of the unit over the past 3 years (including normal operation data and data from 12 known failure events) is used as the training set, totaling approximately 2.6 million samples. The Adam optimization algorithm is used to train the network parameters θ, with an initial learning rate of 0.001, a batch size of 256, and 500 training epochs. During training, the model performance is evaluated on the validation set every 50 epochs. Training is terminated early when the validation set loss no longer decreases for 10 consecutive epochs, resulting in the trained PINN early warning model.

[0104] S5: Health index calculation and dynamic early warning threshold generation;

[0105] The trained PINN early warning model is used to perform online inference on real-time collected monitoring data. Taking a certain moment as an example, the model is input with the current operating parameters (active power 240MW, head 95m, guide vane opening 78%) and multi-dimensional features of real-time monitoring data. The model outputs the following health indicators: vibration health 72, sway health 68, temperature health 85, acoustic emission health 55, pressure health 70, electrical health 90, stability health 65, and overall health 68.

[0106] Taking vibration health as an example, the method for generating the dynamic early warning threshold is illustrated as follows: Historical health data samples (approximately 12,000 sets) with active power ranging from 230MW to 250MW and water head ranging from 90m to 100m are retrieved from the historical health database. The vibration health values ​​of these samples are extracted. The probability density distribution of vibration health is estimated using the Gaussian kernel density estimation method, with the kernel bandwidth determined using the Silverman rule. The 5th percentile of this probability distribution is calculated, yielding a dynamic early warning threshold of 58 for vibration health. The current vibration health value of 72 is higher than the threshold of 58, therefore no early warning is triggered. Similarly, the dynamic early warning threshold corresponding to an acoustic emission health value of 55 is 52; the current value of 55 is higher than the threshold of 52, therefore no early warning is triggered.

[0107] S6: Root cause analysis;

[0108] During a certain operation, when the unit was running at an active power of 260MW and a head of 88m, its overall health dropped to 52, which was lower than the dynamic early warning threshold of 56 for that condition, triggering a primary early warning.

[0109] Initiate the root cause analysis process. The constructed cause-effect graph includes the following dependent variable nodes (fault types): runner cavitation, guide bearing wear, rotor imbalance, stator core loosening, and tailrace vortex band; independent variable nodes (monitored symptoms) include: vibration characteristic frequency amplitude, swing characteristic frequency amplitude, acoustic emission entropy rate, pressure pulsation dominant frequency amplitude, bearing temperature change rate, and more than 30 other nodes; the causal relationships are determined based on physical mechanisms. For example, runner cavitation leads to an increase in the acoustic emission signal entropy rate and the instantaneous energy value of the high-frequency acoustic emission signal, accompanied by an increase in the low-frequency component of the tailrace pressure pulsation amplitude.

[0110] The current anomaly monitoring data is input into the fault cause-effect graph, and a Bayesian network inference algorithm is used to calculate the posterior probability of each candidate fault cause. The calculation results are as follows: posterior probability of runner cavitation 78%, posterior probability of tailrace vortex band 12%, posterior probability of guide bearing wear 5%, posterior probability of rotor imbalance 3%, and posterior probability of stator core loosening 2%. Based on the posterior probability ranking, the root cause of the fault is determined to be "runner cavitation", and a warning message is output: the fault type is initial cavitation, the warning level is warning, and the recommended measure is "It is recommended to check whether there is cavitation erosion at the runner blade outlet edge during the next shutdown maintenance, and to review the coordination relationship between the unit's operating head and the guide vane opening."

[0111] S7: Fault Trend Prediction;

[0112] Based on the comprehensive health status time series calculated every 10 minutes over the past 72 hours (a total of 432 data points), a Temporal Convolutional Network (TCN) was used to predict the trend of comprehensive health status changes over the next 24 hours. The TCN model used 5 layers of dilated convolutions with dilation factors of 1, 2, 4, 8, and 16, and a kernel size of 3. The prediction results showed that if the current operating conditions remained unchanged, the comprehensive health status would drop to 38 after 16 hours, entering the "dangerous" level. Based on this prediction, the maintenance personnel arranged a load adjustment plan in advance, gradually reducing the unit's active power from 260MW to 200MW. After the power adjustment, the comprehensive health status rebounded to 65, successfully avoiding a turbine runner damage accident caused by increased cavitation.

[0113] Example 3: This example uses a 150MW reversible hydro-turbine generator unit in a pumped storage power station as an example to illustrate the application effect of the method of the present invention in a scenario with frequent changes in operating conditions.

[0114] This pumped-storage unit frequently switches between power generation and pumping operations daily, resulting in drastic changes in operating conditions. Traditional fixed-threshold early warning methods have a false alarm rate as high as 23%. After adopting the method of this invention:

[0115] During the transition from power generation to pumping operation, the unit's active power changes from +150MW to -130MW, and the water head fluctuates between 50m and 70m. The physical constraint loss term in the PINN early warning model... This ensures that the model adheres to the constraints of physical mechanisms during transient transitions between operating conditions, and the health indicators output by the model accurately reflect the equipment status rather than spurious fluctuations caused by changes in operating conditions. The dynamic warning threshold is smoothly adjusted as operating conditions continuously change, reducing the false alarm rate from 23% to 4.2%.

[0116] During a pumping operation, the model detected that the acoustic emission health level continuously decreased from 82 to 61. Although this was still above the dynamic warning threshold of 58, TCN trend prediction indicated that the acoustic emission health level would drop to 52 after 8 hours. The model issued an early "caution" warning, indicating "abnormal acoustic emission signal, suspected initial cavitation; it is recommended to check the runner after this pumping operation." Maintenance personnel conducted an inspection after the planned shutdown and found slight cavitation erosion spots on the runner blades at the water outlet edge. These were promptly repaired, preventing further cavitation expansion.

[0117] Example 4: This example uses a super-large mixed-flow turbine generator unit with a rated power of 700MW in a hydropower station as the object to illustrate the applicability of the method of the present invention in a small sample scenario.

[0118] This unit is a newly commissioned unit, with less than 6 months of operation time. There is sufficient normal operation data, but fault data is extremely scarce (only one abnormal temperature event in the guide bearing). Traditional pure data-driven models cannot be effectively trained due to the lack of fault samples.

[0119] After adopting the method of this invention, the PINN early warning model includes a physical constraint loss term. Even in the absence of fault samples, the model can establish a reasonable mapping from normal operation to fault state through the constraints of physical mechanisms. In the eighth month after the unit's commissioning, the model detected a slow downward trend in the guide bearing swing health index, gradually decreasing from 92 to 78. Although still within the normal range, TCN trend prediction indicated that this index would drop below 65 within the next 48 hours. The model issued an early "attention" level warning. Based on this, maintenance personnel arranged a special inspection, discovering minor wear on the guide bearing, which was promptly addressed. Throughout the entire process from commissioning to fault handling, the model utilized only a very small number of fault samples (1 time), achieving a 100% warning accuracy rate, verifying the effectiveness of the method in small sample scenarios.

[0120] Example 5, refer to Figure 2 This embodiment provides a fault early warning system for hydro-generator sets, including:

[0121] The data acquisition and preprocessing module is configured to acquire multi-source heterogeneous monitoring data at a predetermined monitoring location of the hydro-generator unit, preprocess the monitoring data and extract time-domain and frequency-domain features to construct a multi-dimensional feature vector;

[0122] The mechanism model construction module is configured to establish a mechanism model of hydraulic-mechanical-electromagnetic multi-physics coupling based on the structural design parameters and operating environment parameters of the hydro-generator unit, and generate theoretical reference values ​​of each monitored physical quantity under different operating conditions through the mechanism model;

[0123] The physical information neural network early warning module is communicatively connected to the data acquisition and preprocessing module and the mechanism model construction module, respectively. It is configured to embed the mechanism model as a physical constraint into a deep learning network framework to construct a physical information neural network early warning model. The loss function of the physical information neural network early warning model includes a data-driven loss term and a physical constraint loss term. The physical information neural network early warning model is trained by minimizing the loss function, and the trained physical information neural network early warning model is used to perform online inference on the real-time collected monitoring data and output multi-dimensional health indicators.

[0124] The dynamic early warning threshold generation module is communicatively connected to the physical information neural network early warning module. It is configured to retrieve the historical distribution of health indicators under the same operating conditions from historical health data based on the current operating condition parameters of the unit, calculate the probability distribution of health indicators under the same operating conditions using the kernel density estimation method, and use the predetermined quantile of the probability distribution as the dynamic early warning threshold. When any health indicator is lower than the corresponding dynamic early warning threshold, a primary early warning is triggered.

[0125] The causal inference root cause localization module is communicatively connected to the dynamic early warning threshold generation module. It is configured to construct a fault causal graph of the hydro-generator unit after the primary early warning is triggered, input the current abnormal monitoring data into the fault causal graph, calculate the posterior probability of each candidate fault cause through the causal inference algorithm, and output the fault root cause sorting list according to the posterior probability from high to low.

[0126] The trend prediction and graded early warning output module is communicatively connected to the causal inference and root cause localization module. It is configured to predict the future trend of health indicators based on the time series of health indicators within a continuous time window, and to determine the early warning level based on the current value and predicted value of the health indicators. It outputs early warning information including fault type, fault root cause, early warning level and recommended maintenance measures.

[0127] The central control module is communicatively connected to the data acquisition and preprocessing module, the mechanism model construction module, the physical information neural network early warning module, the dynamic early warning threshold generation module, the causal inference root cause location module, and the trend prediction and hierarchical early warning output module, and is configured to coordinate the data flow and timing control between the modules.

[0128] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for early warning of faults in a hydro-generator unit, characterized in that, include: Step S1: Collect multi-source heterogeneous monitoring data of the hydro-generator unit, preprocess the monitoring data, extract time-domain features and frequency-domain features, and construct a multi-dimensional feature vector; Step S2: Based on the structural design parameters and operating environment parameters of the hydro-generator unit, establish a mechanism model of hydraulic-mechanical-electromagnetic multi-physics coupling, and generate theoretical reference values ​​for each monitored physical quantity under different operating conditions through the mechanism model; Step S3: The mechanism model described in step S2 is used as a physical constraint and embedded into a deep learning network framework to construct a physical information neural network early warning model. The loss function of the physical information neural network early warning model includes a data-driven loss term and a physical constraint loss term. The physical information neural network early warning model is trained by minimizing the loss function. Step S4: Using the trained physical information neural network early warning model, perform online inference on the real-time collected monitoring data, calculate multi-dimensional health indicators, determine dynamic early warning thresholds based on the current operating parameters of the unit, and trigger a primary early warning when any health indicator is lower than the corresponding dynamic early warning threshold. Step S5: When the primary warning is triggered, construct the fault cause-effect graph of the hydro-generator unit, input the current abnormal monitoring data into the fault cause-effect graph, calculate the posterior probability of each candidate fault cause through the causal inference algorithm, and output the fault cause sorting list according to the posterior probability from high to low. Step S6: Based on the time series of health indicators within a continuous time window, a temporal convolutional network is used to predict the future trend of health indicators. The warning level is determined according to the current value and predicted value of the health indicators, and the warning information including fault type, root cause of fault, warning level and recommended maintenance measures is output.

2. The method for early warning of faults in a hydro-generator unit according to claim 1, characterized in that, The multi-source heterogeneous monitoring data mentioned in step S1 includes vibration signals, sway signals, acoustic emission signals, temperature signals, pressure pulsation signals, electrical signals, and operating parameters; the preprocessing includes denoising the original signals using an adaptive noise complete set empirical mode decomposition method.

3. The method for early warning of faults in a hydro-generator unit according to claim 1, characterized in that, The loss function of the physical information neural network early warning model described in step S3 is: ； in, This is the set of trainable parameters for a neural network. This is the data-driven loss term, representing the deviation between the model's predicted values ​​and the actual monitored values; The physical constraint loss term represents the degree to which the model's predicted value deviates from the theoretical reference value of the mechanism model described in step S2; This is the physical constraint weighting coefficient, with a value range of 0.01-0.5; The data-driven loss term The expression is: ； Where N is the number of training samples; Let be the input feature vector of the i-th training sample; For neural networks in input and parameters The predicted output is as follows; This represents the actual monitoring value of the i-th training sample. Represents the square of the Euclidean norm; The physical constraint loss term The expression is: ； Where M is the number of physical constraint sampling points; This is the input feature vector of the j-th physical constraint sampling point; For neural networks in input and parameters The predicted output is as follows; For the mechanism model described in step S2, in the input The theoretical reference value is calculated below.

4. The method for early warning of faults in a hydro-generator unit according to claim 3, characterized in that, The health indicators mentioned in step S4 include vibration health, temperature health, stability health and comprehensive health, and the value range of each health indicator is 0-100; The method for determining the dynamic early warning threshold is as follows: based on the operating parameters of the unit's current operation, retrieve the historical distribution of health indicators under the same operating conditions from historical health data, calculate the probability distribution of health indicators under the operating conditions using the kernel density estimation method, and use the 5th percentile of the probability distribution as the dynamic early warning threshold.

5. The method for early warning of faults in a hydro-generator unit according to claim 1, characterized in that, The fault cause-effect graph described in step S5 uses fault type as the dependent variable node, each monitored physical quantity as the independent variable node, and the causal relationship driven by physical mechanism as the edge; the causal inference algorithm is a Bayesian network inference algorithm.

6. The method for early warning of faults in a hydro-generator unit according to claim 1, characterized in that, The warning levels mentioned in step S6 include four levels: healthy, alert, warning, and danger. When the overall health level is greater than or equal to 80, it is judged as healthy; when the overall health level is greater than or equal to 60 and less than 80, it is judged as alert; when the overall health level is greater than or equal to 40 and less than 60, it is judged as warning; and when the overall health level is less than 40, it is judged as danger.

7. The method for early warning of faults in a hydro-generator unit according to claim 4, characterized in that, The mechanism model mentioned in step S2 includes a turbine hydraulic model, a generator electromagnetic model, and a rotor-bearing system dynamic model; the turbine hydraulic model is established based on computational fluid dynamics, the generator electromagnetic model is established based on the finite element method, and the rotor-bearing system dynamic model is established based on the finite element method.

8. The method for early warning of faults in a hydro-generator unit according to claim 1, characterized in that, The time-domain features mentioned in step S1 include the signal's peak value, root mean square value, peak-to-peak value, margin factor, and kurtosis; the frequency-domain features include the frequency shift amplitude, harmonic amplitude, half-frequency amplitude, spectral centroid frequency, and spectral root mean square frequency.

9. A fault early warning system for a hydro-generator set, employing the fault early warning method for a hydro-generator set as described in any one of claims 1-8, characterized in that, include: The data acquisition module is used to obtain the store operation dataset of the target store within a preset statistical period; The data preprocessing module is used to clean, align, and remove outliers from the store operation dataset to generate standardized store operation data. The indicator system construction module is used to construct a multi-dimensional value assessment indicator system for store operation data assets based on the standardized store operation data. The weight calculation module is used to determine the objective weights of each dimension in the multi-dimensional value assessment index system using the entropy weight method. The comprehensive scoring calculation module is used to calculate the comprehensive quality score of store operation data assets based on the indicator values ​​and objective weights of each dimension. The revenue sharing calculation module is used to calculate the actual revenue sharing rate of store operating data assets based on the comprehensive quality score and the preset basic revenue sharing rate, through a dynamic revenue sharing rate adjustment model. The revenue sharing calculation module is used to obtain the total operating revenue of the target store and calculate the value sharing amount of the store's operating data assets based on the total operating revenue and the actual revenue sharing rate. The allocation module is used to distribute the value share among the participants based on their contribution factors. The model update module is used to update the parameters in the dynamic division rate adjustment model online using data from the new period.

Images