Hydraulic turbine unit state evaluation method and system

Abstract

The present application relates to the technical field of intelligent operation and maintenance of hydroelectric equipment, and discloses a method and system for evaluating the state of a hydraulic turbine unit. The method comprises: collecting multi-source operation data in real time through a sensor array and performing time alignment and preprocessing; extracting multi-dimensional characteristic parameters and constructing a dynamic reference curve in combination with historical data; constructing a state evaluation model based on a machine learning algorithm, outputting a state score and a warning signal according to real-time characteristics and the reference curve, and performing graded feedback control according to the score level; and continuously updating and optimizing the model using incremental data. The system correspondingly comprises data acquisition and preprocessing, characteristic extraction and reference construction, state evaluation and control, model optimization and update modules. The present application realizes real-time and accurate evaluation of the operating state of a hydraulic turbine unit, abnormality warning and adaptive control optimization.

Description

Technical Field

[0001] This application relates to the field of equipment condition monitoring and fault diagnosis technology, and in particular to a method and system for assessing the condition of a hydro turbine unit. Background Technology

[0002] Condition monitoring of hydroelectric turbine units is crucial for ensuring the safe and stable operation of hydropower plants. Currently, the widely used method relies on installing various sensors to collect signals such as vibration, temperature, and pressure, and then triggering alarms when limits are exceeded by setting fixed thresholds.

[0003] However, this method has systemic flaws. First, it is essentially a passive open-loop system of "monitoring-alarm." Data from each sensor is analyzed independently, and the evaluation model is usually trained based on limited offline historical data. Once deployed, it remains fixed and cannot reflect long-term degradation of equipment performance or newly emerging operating modes. This leads to a gradual disconnect between the evaluation benchmark and the actual state of the unit, resulting in insufficient early warning capabilities for slowly occurring or complex faults. Second, the results of the condition assessment are disconnected from the unit's operation and control. After the system issues a warning or alarm, manual intervention is usually required for analysis and decision-making. It cannot automatically and in real time execute differentiated control strategies based on the evaluation results, making it difficult to optimize operational efficiency while ensuring safety. Furthermore, traditional data preprocessing and feature extraction methods are relatively rigid and difficult to adapt to the complex and ever-changing operating conditions of hydroelectric units, affecting the accuracy of subsequent analysis.

[0004] Therefore, the lack of a closed-loop intelligent system in the current technology, which can continuously learn, accurately evaluate and immediately trigger corresponding control actions along with the operation of the unit, has become the main bottleneck to realizing intelligent operation and maintenance and forward-looking health management of hydro-turbine units. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this application aims to provide a method and system for assessing the condition of a hydro-turbine unit. By constructing a closed-loop intelligent system capable of self-evolution, real-time assessment, and proactive regulation, it overcomes the problems of static and fixed assessment models and the disconnect between assessment results and control execution in traditional hydro-turbine unit condition monitoring methods, thereby achieving adaptive perception of the unit's operating status and forward-looking safety protection.

[0006] To achieve the above objectives, this application adopts the following technical solution:

[0007] This application provides a method for assessing the condition of a hydro-turbine unit, the method comprising the following steps:

[0008] S101, real-time acquisition of multi-source operating data of the turbine unit through a distributed sensor array, including vibration signal, temperature signal, pressure signal, current signal, speed signal and flow signal; time synchronization protocol is used to time-align each signal to obtain time-aligned raw data;

[0009] S102, preprocessing the raw data includes: adaptive filtering and noise reduction based on dynamically adjusting the filtering parameters according to the signal frequency components; identifying and removing data points that deviate from the preset statistical threshold through a sliding window statistical model; and interpolating and aligning signals with inconsistent timing to obtain preprocessed data.

[0010] S103, extract multidimensional feature parameters from the preprocessed data. The multidimensional feature parameters include: the spectral characteristics of the vibration signal, the gradient rate of change of the temperature signal, the pulsation coefficient of the pressure signal, the harmonic distortion rate of the current signal, the standard deviation of the fluctuation of the speed signal, and the trend slope of the flow signal; combine the historical operating data of the turbine unit, and construct a dynamic benchmark curve through a sliding time window and a machine learning model.

[0011] S104, A state assessment model is constructed based on a machine learning algorithm. The state assessment model takes historical multidimensional feature parameters and dynamic benchmark curves as inputs and outputs unit state scores and abnormal early warning signals after training.

[0012] S105 inputs real-time multidimensional feature parameters and dynamic baseline curves into the state assessment model; if the output unit state score is lower than the preset threshold or an abnormal warning signal is triggered, hierarchical feedback control is executed based on the score level:

[0013] When the score falls within the first threshold range, maintenance recommendations are generated.

[0014] When the score falls within the second threshold range, the unit's operating parameters are dynamically adjusted.

[0015] When the score falls below the third threshold, a shutdown protection command is triggered.

[0016] S106 continuously updates the training set of the state assessment model through real-time incremental collection of multi-source operating data, and periodically performs incremental learning to optimize the parameters of the state assessment model.

[0017] As a preferred technical solution, the sensor array includes at least one self-calibrating sensor, and the method further includes:

[0018] During the data acquisition process, standard test signals are periodically injected into the self-calibrating sensor; the sensor's response data to the test signals is obtained; based on the deviation between the response data and the standard response model, the corresponding type of operating data is dynamically corrected or its reliability is marked; when the deviation continues to exceed the preset tolerance, a sensor calibration anomaly alarm is generated.

[0019] As a preferred technical solution, the sliding window statistical model in step S102 is a multi-model system based on working condition clustering, including:

[0020] Standard operating condition clusters are divided according to the real-time load and speed of the unit; an independent data statistical distribution model is established for each operating condition cluster to determine the preset statistical threshold of the signal; and anomalies in the real-time data are identified based on the model corresponding to the current operating condition cluster.

[0021] As a preferred technical solution, step S103 further includes:

[0022] Calculate the correlation matrix of multidimensional feature parameters within the sliding time window, extract the outer elements of the main diagonal elements to form the associated feature vector; use the associated feature vector and multidimensional feature parameters as input to construct a dynamic benchmark curve and a training state evaluation model.

[0023] As a preferred technical solution, the state evaluation model in step S104 is an ensemble model, and the training process includes:

[0024] Differentiated training subsets are constructed using a bootstrap sampling method; base learners are trained using each subset, including at least one cost-sensitive learner to handle class imbalanced data; the outputs are integrated using a weighted voting method, with weights dynamically adjusted based on the base learners' evaluation accuracy on recent data and outliers on the validation set.

[0025] As a preferred technical solution, in step S105:

[0026] When the score is in the second threshold range, the optimization strategy mapping table is called based on the dominant anomaly feature parameters to generate an adjustment instruction set; the predicted score of the virtual adjusted data by the state assessment model is used as feedback, and the optimal parameter combination is iteratively solved and executed through the particle swarm optimization algorithm.

[0027] As a preferred technical solution, the method further includes:

[0028] When an abnormal warning is triggered but the shutdown threshold is not reached, a simulated degradation verification procedure is initiated; the accelerated degradation process in the direction of the warning is simulated in the digital twin model; the simulated data is input into the status assessment model, and if the prediction shows continuous deterioration and serious failures can be predicted in advance, the warning is confirmed and a degradation report is output.

[0029] As a preferred technical solution, step S106, performing incremental learning includes:

[0030] An incremental dataset is constructed based on new data samples; the parameters of the state assessment model are updated using an elastic weight consolidation algorithm, while the importance of historical parameters is calculated; a penalty term for important parameters is added to the loss function to retain the ability to assess historical data and rare faults.

[0031] This application also provides a turbine generator condition assessment system, which includes a data acquisition and preprocessing module, a feature extraction and benchmark construction module, a condition assessment and control module, and a model optimization and update module.

[0032] The data acquisition and preprocessing module is used to acquire vibration, temperature, pressure, current, rotational speed and flow signals through the sensor array, and perform time alignment, adaptive filtering, outlier removal and interpolation preprocessing on them.

[0033] The feature extraction and benchmark construction module is used to extract multi-dimensional feature parameters from preprocessed data, including spectral features, gradient change rate, pulsation coefficient, harmonic distortion rate, fluctuation standard deviation and trend slope, and to construct a dynamic benchmark curve based on historical data and a sliding window.

[0034] The status assessment and control module includes a status assessment model and a control unit; the status assessment model is used to output status scores and early warning signals based on real-time characteristic parameters and dynamic benchmark curves; the control unit is used to perform graded feedback based on the score level: generating maintenance suggestions, adjusting operating parameters, or triggering shutdown protection;

[0035] The model optimization and update module is used to update the training set using incremental data and optimize the state evaluation model through periodic incremental learning.

[0036] Compared with the prior art, the beneficial effects of this application are as follows:

[0037] This application constructs a complete closed loop of "perception-assessment-decision-control-learning", which can automatically and instantly execute hierarchical control actions from early warning and adjustment to protection based on real-time assessment results. It realizes intelligent linkage between state perception and safe execution, and significantly improves the unit's active safety protection capability.

[0038] By introducing an incremental learning mechanism, this application makes the state assessment model no longer statically fixed. The system can continuously optimize itself using newly generated operational data in real time, so that the model parameters and dynamic benchmark curves can be adaptively updated as the equipment ages and operating conditions change, thus possessing self-evolution capabilities and maintaining the accuracy and foresight of the assessment over the long term.

[0039] This solution integrates multi-source data synchronization, adaptive preprocessing, dynamic feature extraction, and intelligent modeling. The dynamic baseline curve replaces the fixed threshold, more accurately depicting the normal operating range of the unit under current conditions. Furthermore, the fusion analysis of multi-dimensional features enhances the ability to identify early subtle anomalies and complex faults, significantly improving the sensitivity and reliability of condition assessment.

[0040] The hierarchical feedback control strategy achieves a balance between safety and economy. Based on the severity of the assessment score, the system adopts differentiated response measures, avoiding the loss of benefits caused by indiscriminate downtime. It adjusts parameters or provides maintenance suggestions in the early stages of potential faults, and decisively executes protection measures when a danger has already occurred, thus achieving intelligent operation and maintenance. Attached Figure Description

[0041] Figure 1 This is a flowchart illustrating the steps of the turbine unit condition assessment method in this application. Detailed Implementation

[0042] To enable those skilled in the art to better understand the present application, the technical solutions in specific embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0043] like Figure 1 As shown, this application provides a method for assessing the condition of a hydro-turbine unit, which includes the following steps.

[0044] S101 collects multi-source operating data of the turbine unit in real time through a distributed sensor array. The multi-source operating data includes vibration signals, temperature signals, pressure signals, current signals, speed signals, and flow signals. A time synchronization protocol is used to align the timing of each signal to obtain the time-aligned raw data.

[0045] In practice, the sensor array is optimized based on the structural mechanics and fluid dynamics characteristics of the turbine.

[0046] The vibration acceleration sensor adopts the IEPE type, with a frequency response range of 0.5Hz to 10kHz. It is arranged in three-dimensional space (X, Y, Z directions) in key parts such as the main shaft seal, top cover, upper and lower guide bearings and thrust bearing housing. The installation torque is strictly controlled within the specified value (e.g., 20N·m ± 0.5N·m).

[0047] The temperature sensor uses a Pt100 platinum resistance thermometer and is embedded in the bearing pad, winding coil, and cooler inlet and outlet, with a sampling rate of 1Hz.

[0048] Pressure pulsation sensors are located at the inlet of the volute, the inlet of the impeller, and the elbow of the tailrace pipe, with a range of -0.1 to 1.5 MPa.

[0049] The accuracy of the current and voltage transformers is 0.2S class.

[0050] All sensors are synchronized via a hardware clock module and use the IEEE 1588v2 (PTP) precision time protocol for network time synchronization, ensuring that the clock deviation of all network devices is less than 100 microseconds, thereby guaranteeing the accurate correspondence between key time-domain correlation features such as vibration peak and pressure pulsation phase.

[0051] S102, preprocessing the raw data. This includes adaptive filtering and noise reduction, which dynamically adjusts filtering parameters based on signal frequency components. A sliding window statistical model is used to identify and remove data points that deviate from a preset statistical threshold. Interpolation is performed on signals with inconsistent timing to obtain the preprocessed data.

[0052] In practice, adaptive filtering and noise reduction employs an improved variable-step LMS (LeastMeanSquare) algorithm, with a step size factor... With error signal Energy adaptive adjustment:

[0053] ,in , To achieve a balance between convergence speed and steady-state error, the filter order was set to 64, and the reference channel input was taken from adjacent normal operating condition data segments at the same measurement point.

[0054] For the sliding window statistical model, a working condition clustering based on a Gaussian mixture model (GMM) is adopted, and the feature vector is... .

[0055] The Expectation-Maximization (EM) algorithm is used to cluster historical data from one year, automatically determining the number of operating condition clusters (e.g., 5). For each operating condition cluster, a robust [mechanism / system] is established for each signal. (Three standard deviations) boundary model. When real-time data points arrive, calculate their Mahalanobis distance to the center of each operating condition cluster, assign them to the nearest cluster, and use the corresponding data of that cluster. Anomalies are detected and removed at the boundary.

[0056] For data with inconsistent sampling rates (e.g., vibration signal 10kHz, temperature signal 1Hz), synchronous resampling based on Lagrange interpolation is used to interpolate the low-frequency signal to a unified high-frequency timestamp. The formula is as follows:

[0057] For the temperature signal at time... interpolation ,in The interpolation order is 3.

[0058] This process effectively filters out high-frequency electromagnetic interference and low-frequency drift, and ensures strict alignment of multi-source data on the time axis.

[0059] S103 extracts multidimensional feature parameters from the preprocessed data. These parameters include the spectral characteristics of the vibration signal, the gradient rate of change of the temperature signal, the pulsation coefficient of the pressure signal, the harmonic distortion rate of the current signal, the standard deviation of the speed signal fluctuation, and the trend slope of the flow signal. Combined with historical operating data of the turbine unit, a dynamic baseline curve is constructed using a sliding time window and a machine learning model.

[0060] This step also includes calculating the correlation matrix of the multidimensional feature parameters within the sliding time window, and extracting the outer elements of the main diagonal elements to form the associated feature vector. The associated feature vector and the multidimensional feature parameters are then used as input to construct a dynamic benchmark curve and train the state evaluation model.

[0061] In practice, spectral feature extraction employs a 1024-point FFT with a Hanning window to calculate the amplitudes of the first harmonic (1X), second harmonic (2X), ... up to the tenth harmonic (10X), as well as the total energy within the 0.4-0.5 harmonic (subsynchronous) and 2.5-3.5 harmonic (supersynchronous) frequency bands. The rate of change of the temperature gradient is calculated as follows: ,in Take 5 minutes. Pressure pulsation coefficient. Calculated over one frequency cycle. Total harmonic distortion of current. ,in , This represents the RMS value of the fundamental current. The standard deviation of the rotational speed fluctuation is also shown. Calculated based on 600 sampling points per second. The flow trend slope is obtained by fitting the flow data of the past 10 minutes using the least squares method. .

[0062] The sliding time window is set to 72 hours. At each moment... The calculation window contains 25 characteristic parameters across 6 categories (e.g., 10 vibration frequency bands, 1 temperature gradient, 1 pressure pulsation coefficient, 1 THD, 1 speed standard deviation, 1 flow rate slope, etc.). Pearson correlation coefficient matrix .

[0063] extract The 300 coefficients in the upper triangular portion (excluding the main diagonal) constitute the associated feature vector. The dynamic baseline curve is constructed using a seasonal autoregressive integral moving average (SARIMA) model, with each feature parameter's time series... The model is in the form of ,in (Daily periodicity).

[0064] The optimal parameters (p, d, q, P, D, Q) are determined through a grid search, and their range is [missing information]. , .

[0065] After model training, the 95% confidence interval of the predicted feature values ​​for the next hour is used as a dynamic benchmark. The introduction of associated feature vectors enables the model to detect latent fault symptoms such as "a sudden weakening of the correlation between the 1X component of the upper bearing vibration and the amplitude of the 5th harmonic of the excitation current".

[0066] S104, a state assessment model is constructed based on machine learning algorithms. The model takes historical multidimensional feature parameters and a dynamic baseline curve as input, and outputs a unit state score and anomaly warning signals after training. The state assessment model is an ensemble model, and the training process includes constructing a differentiated training subset using a bootstrap sampling method.

[0067] Base learners are trained using subsets of data, including at least one cost-sensitive learner to handle imbalanced data. The outputs are then integrated using a weighted voting method, with weights dynamically adjusted based on the base learners' accuracy in evaluating recent data and outliers on the validation set.

[0068] In practice, the state assessment model adopts a three-level cascaded integrated architecture.

[0069] The first stage is feature extraction and converter, which includes a one-dimensional convolutional neural network (1D-CNN, with a kernel size of 3 and a number of filters of 16, used to capture local temporal patterns) and a gated recurrent unit network (GRU, with 32 hidden layer units, used to capture long-term dependencies).

[0070] The second level is the base learner layer, which contains three heterogeneous models:

[0071] 1) Support Vector Data Description (SVDD), its Gaussian kernel parameters This is used to describe the hypersphere boundary of health data, assigning a high anomaly probability to points that deviate from the boundary.

[0072] 2) Gradient Boosting Decision Tree (LightGBM), parameters are: , , ;

[0073] 3) A deep autoencoder (DAE) with an encoder structure of [input layer]-

[64] -

[32] -

[16] (number of neurons), a symmetrical decoder, and reconstruction error used to measure anomalies.

[0074] The third level is the meta-learner, which uses logistic regression. The training subset of the base learner is used to handle the class imbalance problem through the SMOTEENN method (a combination of oversampling and undersampling).

[0075] Dynamic weights of weighted voting Calculate using the following formula: ,in It is the first The AUC (Area Under the Receiver Operating Characteristic) value of each base learner on the validation set of the most recent 30 days. It is its F1 score across the entire historical fault sample. The final crew condition score. ,in Output by the meta-learner and passed through Function normalization.

[0076] S105 inputs real-time multi-dimensional feature parameters and dynamic baseline curves into the status assessment model. If the output unit status score is lower than a preset threshold or an abnormal warning signal is triggered, hierarchical feedback control is executed based on the score level. When the score is in the first threshold range, maintenance suggestions are generated. When the score is in the second threshold range, unit operating parameters are dynamically adjusted. When the score is lower than the third threshold, a shutdown protection command is triggered.

[0077] When the score falls within the second threshold range, an adjustment instruction set is generated by calling the optimization strategy mapping table based on the dominant anomaly feature parameters. Using the predicted score of the virtual adjusted data from the state assessment model as feedback, the optimal parameter combination is iteratively solved using the particle swarm optimization algorithm and then executed.

[0078] In practice, the scoring threshold is set as follows: First threshold interval Second threshold interval The third threshold .

[0079] When the score falls into the second range At that time, the control unit initiates the parameter optimization process.

[0080] First, identify the three characteristic parameters that deviate most from their dynamic baseline (such as "upper guide X-axis vibration 1X amplitude", "upper guide tile temperature gradient", and "tailrace pipe pressure pulsation coefficient"), and query the "symptom-countermeasure mapping table" in the expert knowledge base. For example, for "high vibration 1X amplitude", the mapped adjustable parameter is... .

[0081] Then, construct the optimization objective function: The predicted unit condition score is subject to the following constraints: , , φ And ensure that the power output fluctuation is less than .

[0082] A particle swarm optimization algorithm with a shrinkage factor is used, and the particle position vector... The speed update formula is: Among them, contraction factor Inertial weight Learning factor Set the number of particles. Number of iterations .

[0083] In each iteration, for the position of each particle (i.e., a set of running parameters), the fast prediction interface of the state evaluation model (using a lightweight model) is called to calculate its prediction score. .

[0084] After optimization, the optimal parameter combination is sent to the unit speed governor and excitation controller for execution. This process takes place during... Completed within minutes, enabling online adaptive adjustment.

[0085] S106 continuously updates the training set of the state assessment model through real-time incremental collection of multi-source operating data, and periodically performs incremental learning to optimize the parameters of the state assessment model.

[0086] Incremental learning involves constructing an incremental dataset based on new data samples. An elastic weight consolidation algorithm is used to update the parameters of the state evaluation model, while simultaneously calculating the importance of historical parameters. A penalty term for important parameters is added to the loss function to preserve the evaluation capability for historical data and rare faults.

[0087] In practice, the incremental learning process is triggered on a weekly cycle. Each week, newly added data is cleaned and labeled to form an incremental dataset. .

[0088] An improved Online Elastic Weight Consolidation (Online-EWC) algorithm is used to update the model. For the parameters in the model... Its importance is approximated by the diagonal Fisher information matrix. ,in For the amount of data from historical tasks, This represents the likelihood of the data given the parameters.

[0089] In incremental learning, the total loss function is defined as:

[0090] ,in For the cross-entropy loss on the new data, To set the regularization strength, set to . These are the parameter values ​​from the previous training round.

[0091] Importance matrix Updated after each incremental learning step: The attenuation factor This is used to balance the importance of history with that of the present.

[0092] This method allows the model to adapt to new operating modes (such as high-temperature, high-load summer modes) while firmly "remembering" rare failure patterns that occurred during winter or low-load conditions, effectively mitigating catastrophic forgetting. Model versions are managed iteratively, and regression testing on the historical validation set is performed after each update to ensure performance does not degrade by more than [a certain percentage]. Only then can it be deployed.

[0093] As a preferred technical solution, the sensor array includes at least one self-calibrating sensor. The method further includes periodically injecting a standard test signal into the self-calibrating sensor during data acquisition. The sensor's response data to the test signal is acquired, and based on the deviation between the response data and the standard response model, the corresponding type of operating data is dynamically corrected or its reliability is marked. When the deviation continuously exceeds a preset tolerance, a sensor calibration anomaly alarm is generated.

[0094] In practice, approximately [number] sensors are deployed in each type of sensor network. A self-calibrating smart sensor node. This node integrates a high-precision standard source and a microprocessor. For example, the self-calibrating vibration sensor incorporates a piezoelectric ceramic exciter, capable of generating frequencies of [frequency range missing]. Amplitude The standard sinusoidal acceleration signal.

[0095] The self-calibrating temperature sensor incorporates a high-precision resistor network, which can simulate... , Two temperature points. The verification program starts automatically during the low-load period in the early morning each day.

[0096] The sensor collects its own response to a standard stimulus. Compared with the standard calibration curve stored in the node ( Expected output (for factory calibration) Compare the results. Calculate the relative error. ( (Full scale). If If the error is not specified, the sensor status is "Excellent," and this error is used to perform linear compensation on the data of sensors of the same batch (same model, same batch installed) on the same day:

[0097] .like The status is "Good," and the data is marked "Needs Attention." If... continued When the status is "fault", an alarm is triggered and the data of that node is removed from the data stream, and the data of the adjacent node is used to fill the gap through spatial interpolation.

[0098] Sensor health is also used as a feature input state evaluation model.

[0099] As a preferred technical solution, the sliding window statistical model in step S102 is a multi-model system based on operating condition clustering, which includes dividing standard operating condition clusters according to the unit's real-time load and speed. An independent data statistical distribution model is established for each operating condition cluster to determine the preset statistical threshold for the signal. Anomalies in the real-time data are identified based on the model corresponding to the current operating condition cluster.

[0100] The specific implementation method has been detailed in S102. Its core is to use GMM for soft clustering instead of hard partitioning. Each data point belongs to multiple operating condition clusters with a certain probability. Its anomaly score is the weighted average of the anomaly probabilities under each cluster model. This is more in line with the continuity of operating condition transition in actual operation and improves the flexibility of threshold setting and the accuracy of anomaly detection.

[0101] As a preferred technical solution, step S103 further includes calculating the correlation matrix of the multidimensional feature parameters within the sliding time window, and extracting the outer elements of the main diagonal elements to form the associated feature vector. The associated feature vector and the multidimensional feature parameters are used together as input to construct a dynamic benchmark curve and a training state evaluation model.

[0102] The specific implementation method has been detailed in S103. This method encodes time-domain and frequency-domain features and the time-varying correlations between features (a total of 300 dimensions), greatly enriching the state characterization information. When constructing the dynamic reference curve, not only is a SARIMA model established for each individual feature, but also an ARIMA model is established for the correlation coefficients of key feature pairs (such as vibration 1X and the 5th harmonic of current), thereby forming a multi-dimensional dynamic reference surface, which can more sensitively capture the drift of system coupling characteristics.

[0103] As a preferred technical solution, the state evaluation model in step S104 is an ensemble model. The training process includes constructing differentiated training subsets using a bootstrap sampling method. Base learners are trained using each subset, including at least one cost-sensitive learner to handle class-imbalanced data. The outputs are then integrated using a weighted voting method, with the weights dynamically adjusted based on the base learners' evaluation accuracy on the validation set for recent data and anomalous states.

[0104] The specific implementation method is detailed in S104. The designed three-level cascaded ensemble model consists of a first-level CNN-GRU deep network that automatically learns high-level feature representations, a second-level heterogeneous base learner (SVDD, LightGBM, DAE) that makes decisions based on different mathematical principles (boundary, tree ensemble, reconstruction), and a third-level meta-learner that performs fusion. This structure combines the representation learning capabilities of deep learning with the interpretability and robustness of traditional machine learning, and through dynamic weighting and focusing on the latest data and key anomalies, its performance is significantly better than that of a single model.

[0105] As a preferred technical solution, in step S105, when the score is within the second threshold range, an adjustment instruction set is generated by calling the optimization strategy mapping table based on the dominant anomaly feature parameters. Using the predicted score of the virtual adjusted data from the state assessment model as feedback, the optimal parameter combination is iteratively solved using the particle swarm optimization algorithm and then executed.

[0106] The specific implementation method has been detailed in S105. This scheme realizes closed-loop control of "perception-assessment-decision-optimization". Its core innovation lies in embedding the state assessment model as a "digital agent" into the optimization loop, and using the PSO algorithm to search for operating parameters that maximize the output of the "digital agent" (i.e., the predicted health score). This is essentially "health optimization" control of the unit in the digital space, enabling the unit to automatically adjust to the operating point most beneficial to equipment health when there are minor abnormalities, thereby extending the maintenance cycle.

[0107] As a preferred technical solution, the method also includes initiating a simulated degradation verification procedure when an abnormal warning is triggered but the shutdown threshold is not reached. The accelerated degradation process in the direction of the warning is simulated in a digital twin model. The simulated data is input into the state assessment model; if the prediction shows continuous deterioration and severe failures can be predicted in advance, the warning is confirmed and a degradation report is output.

[0108] In practice, the digital twin model is formed by coupling multiple domain models:

[0109] 1) Mechanical module: Based on the Lagrange equation, a multibody dynamics model of "bearing-shaft-wheel" is established, including mass, stiffness, and damping matrix.

[0110] 2) Hydraulic module: A pressure pipe-volute-runner-tailrace pipe system model is established based on the one-dimensional unsteady flow characteristic line method.

[0111] 3) Electrical module: Establish the Park equation model for the synchronous generator.

[0112] 4) Degradation model: For common faults (such as imbalance, misalignment, cavitation), establish the mapping relationship between fault parameters and model parameters (such as stiffness change, excitation force change).

[0113] When an early warning is triggered (e.g., the score drops to 92 for the first time), the system, in its twin model, increases the excitation force amplitude of the "misalignment" fault by ΔF=F0*(1+k*t) based on the early warning characteristics (e.g., a second harmonic increase in the vibration spectrum), where k is an acceleration factor (e.g., 10) and t is the simulation time. In a 10x speedup simulation, the degradation process over the next 24 hours (simulation time) is run, and simulated sensor data is generated.

[0114] Input this data into the status assessment model to obtain a simulated score curve. If the curve falls below the downtime threshold (80) within the simulated 12 hours, it is judged as a "high-risk warning," and the estimated remaining available time (RUL) is derived in reverse, generating an in-depth report that includes fault development prediction and maintenance window. Otherwise, it is judged as an "observable warning."

[0115] As a preferred technical solution, step S106, performing incremental learning includes constructing an incremental dataset based on new data samples. The parameters of the state evaluation model are updated using an elastic weight consolidation algorithm, while simultaneously calculating the importance of historical parameters. A penalty term for important parameters is added to the loss function to preserve the evaluation capability for historical data and rare faults.

[0116] The specific implementation method is detailed in S106. The Online-EWC algorithm adopted, by introducing an importance decay factor γ, achieves dynamic and gradual updates to the importance of parameters. This allows the model to adapt to normal performance changes caused by equipment aging during continuous learning, while also retaining critical failure modes learned in the early stages of operation (which may not recur for many years). This ensures the reliability of the model's evaluation throughout its entire lifecycle and is crucial for achieving lifelong health management of equipment.

[0117] This application also provides a turbine generator condition assessment system, which includes a data acquisition and preprocessing module, a feature extraction and benchmark construction module, a condition assessment and control module, and a model optimization and update module.

[0118] The data acquisition and preprocessing module is used to acquire vibration, temperature, pressure, current, speed and flow signals through the sensor array, and perform time alignment, adaptive filtering, outlier removal and interpolation preprocessing on them.

[0119] The feature extraction and benchmark construction module is used to extract multi-dimensional feature parameters from preprocessed data, including spectral features, gradient change rate, pulsation coefficient, harmonic distortion rate, fluctuation standard deviation and trend slope, and to construct a dynamic benchmark curve based on historical data and a sliding window.

[0120] The condition assessment and control module includes a condition assessment model and a control unit. The condition assessment model outputs condition scores and early warning signals based on real-time characteristic parameters and dynamic baseline curves. The control unit performs tiered feedback based on the score level, generating maintenance recommendations, adjusting operating parameters, or triggering shutdown protection.

[0121] The model optimization and update module is used to update the training set with incremental data and optimize the state evaluation model through periodic incremental learning.

[0122] In practice, the system adopts a "cloud-edge-device" collaborative architecture. The "device" side consists of intelligent sensor nodes with preliminary filtering and self-verification functions. The "edge" side comprises data acquisition and preprocessing modules and feature extraction modules deployed on the power plant's industrial control computer, achieving millisecond-level real-time processing. The "cloud" side consists of state assessment and control modules, model optimization and update modules, and a digital twin simulation engine deployed on a private cloud server.

[0123] The modules interact securely and semantically via the OPCUA protocol. Maintenance suggestions generated by the control unit are pushed to maintenance personnel through a human-machine interface (HMI), parameter adjustment commands are sent to the unit PLC via the Modbus TCP protocol, and shutdown protection commands are directly connected to the protection system via hardwiring. The model optimization and update module is integrated with the version control system to ensure that model iterations are traceable and rollbackable. This system forms a complete solution from physical sensing, digital modeling, intelligent analysis to closed-loop control.

[0124] It should be noted that the terms "first," "second," and similar terms used in this application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, "an" or "a" and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. "A plurality" or "several" indicates at least two. Unless otherwise stated, terms such as "front," "back," "left," "right," "lower," and / or "upper" are for illustrative purposes only and are not limited to a location or spatial orientation. Terms such as "comprising" or "including" indicate that the elements or objects preceding "comprising" encompass the elements or objects listed following "comprising" or "including" and their equivalents, and do not exclude other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect.

[0125] The singular forms “a,” “the,” and “the” used in this application specification and appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0126] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A method for assessing the condition of a hydro-turbine unit, characterized in that, Includes the following steps: S101, real-time acquisition of multi-source operating data of the turbine unit through a distributed sensor array, including vibration signal, temperature signal, pressure signal, current signal, speed signal and flow signal; time synchronization protocol is used to time-align each signal to obtain time-aligned raw data; S102, preprocessing the raw data includes: adaptive filtering and noise reduction based on dynamically adjusting the filtering parameters according to the signal frequency components; identifying and removing data points that deviate from the preset statistical threshold through a sliding window statistical model; and interpolating and aligning signals with inconsistent timing to obtain preprocessed data. S103, extract multidimensional feature parameters from the preprocessed data. The multidimensional feature parameters include: the spectral characteristics of the vibration signal, the gradient rate of change of the temperature signal, the pulsation coefficient of the pressure signal, the harmonic distortion rate of the current signal, the standard deviation of the fluctuation of the speed signal, and the trend slope of the flow signal; combine the historical operating data of the turbine unit, and construct a dynamic benchmark curve through a sliding time window and a machine learning model. S104, A state assessment model is constructed based on a machine learning algorithm. The state assessment model takes historical multidimensional feature parameters and dynamic benchmark curves as inputs and outputs unit state scores and abnormal early warning signals after training. S105 inputs real-time multidimensional feature parameters and dynamic baseline curves into the state assessment model; if the output unit state score is lower than the preset threshold or an abnormal warning signal is triggered, hierarchical feedback control is executed based on the score level: When the score falls within the first threshold range, maintenance recommendations are generated. When the score falls within the second threshold range, the unit's operating parameters are dynamically adjusted. When the score falls below the third threshold, a shutdown protection command is triggered. S106 continuously updates the training set of the state assessment model through real-time incremental collection of multi-source operating data, and periodically performs incremental learning to optimize the parameters of the state assessment model.

2. The method for assessing the condition of a hydro-turbine unit according to claim 1, characterized in that, The sensor array includes at least one self-calibrating sensor, and the method further includes: During the data acquisition process, standard test signals are periodically injected into the self-calibrating sensor; the sensor's response data to the test signals is obtained; based on the deviation between the response data and the standard response model, the corresponding type of operating data is dynamically corrected or its reliability is marked; when the deviation continues to exceed the preset tolerance, a sensor calibration anomaly alarm is generated.

3. The method for assessing the condition of a hydro-turbine unit according to claim 1 or 2, characterized in that, The sliding window statistical model in step S102 is a multi-model system based on working condition clustering, including: Standard operating condition clusters are divided according to the real-time load and speed of the unit; an independent data statistical distribution model is established for each operating condition cluster to determine the preset statistical threshold of the signal; and anomalies in the real-time data are identified based on the model corresponding to the current operating condition cluster.

4. The method for assessing the condition of a hydro-turbine unit according to claim 3, characterized in that, Step S103 further includes: Calculate the correlation matrix of multidimensional feature parameters within the sliding time window, extract the outer elements of the main diagonal elements to form the associated feature vector; use the associated feature vector and multidimensional feature parameters as input to construct a dynamic benchmark curve and a training state evaluation model.

5. The method for assessing the condition of a hydro-turbine unit according to claim 1, characterized in that, The state evaluation model in step S104 is an ensemble model, and the training process includes: Differentiated training subsets are constructed using a bootstrap sampling method; base learners are trained using each subset, including at least one cost-sensitive learner to handle class imbalanced data; the outputs are integrated using a weighted voting method, with weights dynamically adjusted based on the base learners' evaluation accuracy on recent data and outliers on the validation set.

6. The method for assessing the condition of a hydro-turbine unit according to claim 1, characterized in that, In step S105: When the score is in the second threshold range, the optimization strategy mapping table is called based on the dominant anomaly feature parameters to generate an adjustment instruction set; the predicted score of the virtual adjusted data by the state assessment model is used as feedback, and the optimal parameter combination is iteratively solved and executed through the particle swarm optimization algorithm.

7. The method for assessing the condition of a hydro-turbine unit according to claim 6, characterized in that, The method further includes: When an abnormal warning is triggered but the shutdown threshold is not reached, a simulated degradation verification procedure is initiated; the accelerated degradation process in the direction of the warning is simulated in the digital twin model; the simulated data is input into the status assessment model, and if the prediction shows continuous deterioration and serious failures can be predicted in advance, the warning is confirmed and a degradation report is output.

8. The method for assessing the condition of a hydro-turbine unit according to claim 1, characterized in that, In step S106, incremental learning includes: An incremental dataset is constructed based on new data samples; the parameters of the state assessment model are updated using an elastic weight consolidation algorithm, while the importance of historical parameters is calculated; a penalty term for important parameters is added to the loss function to retain the ability to assess historical data and rare faults.

9. A turbine generator condition assessment system, characterized in that, It includes a data acquisition and preprocessing module, a feature extraction and benchmark construction module, a state assessment and control module, and a model optimization and update module; The data acquisition and preprocessing module is used to acquire vibration, temperature, pressure, current, rotational speed and flow signals through the sensor array, and perform time alignment, adaptive filtering, outlier removal and interpolation preprocessing on them. The feature extraction and benchmark construction module is used to extract multi-dimensional feature parameters from preprocessed data, including spectral features, gradient change rate, pulsation coefficient, harmonic distortion rate, fluctuation standard deviation and trend slope, and to construct a dynamic benchmark curve based on historical data and a sliding window. The status assessment and control module includes a status assessment model and a control unit; the status assessment model is used to output status scores and early warning signals based on real-time characteristic parameters and dynamic benchmark curves; the control unit is used to perform graded feedback based on the score level: generating maintenance suggestions, adjusting operating parameters, or triggering shutdown protection; The model optimization and update module is used to update the training set using incremental data and optimize the state evaluation model through periodic incremental learning.

Images