Intelligent prediction method for state of water turbine

Abstract

The invention discloses a water turbine state intelligent prediction method which comprises the following steps: collecting multi-source sensor data of a water turbine, including vibration, temperature, pressure, flow and electrical parameters; performing space-time alignment preprocessing on the multi-source sensor data to generate a space-time associated data set; extracting and fusing the features of the space-time associated data through a multi-modal space-time diagram network, and generating a joint feature vector; performing health state prediction on the joint feature vector based on a dynamic digital twinborn model, and outputting a health score, a fault probability and a confidence interval; analyzing a fault propagation path by using a causal reasoning module, positioning a fault root cause and generating an interpretable report; and triggering an early warning or maintenance decision according to the prediction result. According to the technical scheme, the technical problems that multi-source heterogeneous data fusion is difficult, fault coupling and propagation are uncertain, real-time performance and computing resources are contradictory, and interpretability and reliability are insufficient are solved.

Description

Technical Field

[0001] The present invention relates to the field of hydraulic turbines, and specifically, to an intelligent prediction method for the state of a hydraulic turbine. Background Art

[0002] As the core power equipment of a hydraulic power generation system, the operating state of a hydraulic turbine directly affects the power generation efficiency and the safety of the unit. The traditional maintenance method mainly focuses on regular maintenance and post-fault repair, which has problems such as high maintenance costs and large losses due to unplanned shutdowns. With the development of intelligent operation and maintenance technology, real-time monitoring of multi-source parameters such as vibration, temperature, and pressure, combined with data analysis to predict potential faults, has become the key means to achieve predictive maintenance of hydraulic turbines. This technology can effectively avoid sudden failures and optimize the maintenance cycle, which is of great significance for ensuring the stable operation of the power grid and reducing operation and maintenance costs.

[0003] Although certain progress has been made in the existing hydraulic turbine state prediction technology, there are still the following bottlenecks:

[0004] Difficulty in fusing multi-source heterogeneous data: The sampling frequencies and spatial distributions of sensor data such as vibration and temperature vary greatly, and it is difficult for traditional methods to achieve spatio-temporal alignment and effective correlation, resulting in insufficient feature extraction;

[0005] Uncertainty in fault coupling and propagation: Multiple types of faults such as mechanical and electrical faults often couple with each other, and existing models lack the ability to dynamically model the fault propagation path, resulting in high rates of missed detections and false alarms;

[0006] Contradiction between real-time performance and computing resources: Deep learning models rely on high-computing power platforms and are difficult to be deployed in real time on edge devices, unable to meet the online analysis requirements of high-frequency data (such as vibration signals);

[0007] Insufficient interpretability and reliability: The prediction results are mostly black-box outputs, lacking quantitative analysis of the root causes and confidence levels of faults, which restricts the credibility of operation and maintenance decisions.

[0008] The above problems severely limit the engineering application of hydraulic turbine state prediction technology, and there is an urgent need for an innovative method that takes into account accuracy, efficiency, and interpretability. Summary of the Invention

[0009] The purpose of the embodiments of this application is to propose an intelligent prediction method for the state of a hydraulic turbine to solve the technical problems of difficulty in fusing multi-source heterogeneous data, uncertainty in fault coupling and propagation, contradiction between real-time performance and computing resources, and insufficient interpretability and reliability.

[0010] The technical solution of the present invention is as follows:

[0011] An intelligent prediction method for the state of a hydraulic turbine, comprising the following steps:

[0012] S100. Collect multi-source sensor data of the water turbine, including vibration, temperature, pressure, flow rate, and electrical parameters;

[0013] S200. Perform spatio-temporal alignment preprocessing on the multi-source sensor data to generate a spatio-temporal correlation dataset;

[0014] S300. Extract and fuse the features of the spatio-temporal correlation data through a multi-modal spatio-temporal graph network to generate a joint feature vector;

[0015] S400. Based on the dynamic digital twin model, perform health status prediction on the joint feature vector, and output a health score, a fault probability, and a confidence interval;

[0016] S500. Use a causal inference unit to analyze the fault propagation path, locate the root cause of the fault, and generate an interpretability report;

[0017] S600. Trigger an early warning or a maintenance decision according to the prediction result.

[0018] In a possible implementation manner, in the step S200, the spatio-temporal alignment preprocessing includes:

[0019] Adopt dynamic time warping to perform time alignment on sensor data with different sampling frequencies;

[0020] Define a node topology structure based on the spatial adjacency relationship of the sensor installation positions.

[0021] In a possible implementation manner, in the step S300, the multi-modal spatio-temporal graph network includes:

[0022] A single-modal feature extraction unit: perform time-frequency domain decomposition on vibration signals, and perform statistical and frequency domain analysis on temperature and pressure signals;

[0023] A cross-modal fusion unit: learn the correlation weights between modalities through a multi-head attention mechanism;

[0024] A spatio-temporal graph convolution unit: model the spatio-temporal dependence relationship of multi-source data based on a dynamic graph structure, and the edge weights of the dynamic graph structure are dynamically updated by physical transmission paths and data correlations.

[0025] In a possible implementation manner, in the step S400, the dynamic digital twin model is implemented in the following manner:

[0026] Embed a simplified mechanism model of the water turbine to generate benchmark health features;

[0027] Adopt knowledge distillation technology to compress the multi-modal spatio-temporal graph network to generate a lightweight student model;

[0028] Deploy the student model on an edge device and calibrate the model parameters online in combination with Kalman filtering.

[0029] In a possible implementation manner, in the step S500, the causal reasoning unit includes:

[0030] Construct an initial causal graph based on domain knowledge and dynamically optimize the weights of the fault propagation path through reinforcement learning;

[0031] Use a Bayesian neural network to quantify the uncertainty of the prediction result and generate a confidence interval;

[0032] Through gradient backpropagation and SHAP value analysis, locate the root cause of the fault and trace back the maximum probability propagation chain.

[0033] In a possible implementation manner, it further includes a transfer learning step:

[0034] Pre-train the multi-modal spatio-temporal graph network using the simulation data of multiple types of hydroturbines;

[0035] Fine-tune the pre-trained model with the measured data of the target hydroturbine.

[0036] In a possible implementation manner, in the step S600, the prediction result includes:

[0037] Health score, generated by linearly weighted fusion of the degradation degrees of multi-modal features;

[0038] Fault probability, the prediction probabilities of bearing wear, blade cavitation, and rotor imbalance output by the classification model;

[0039] Confidence interval, the upper and lower bounds of the 95% confidence level calculated based on the Monte Carlo Dropout method.

[0040] A hydroturbine state intelligent prediction system includes:

[0041] A data acquisition module that acquires multi-source sensor data of the hydroturbine, including vibration, temperature, pressure, flow rate, and electrical parameters;

[0042] A data preprocessing module that performs spatio-temporal alignment preprocessing on the multi-source sensor data to generate a spatio-temporal correlation data set;

[0043] A feature extraction module that extracts and fuses the features of the spatio-temporal correlation data through a multi-modal spatio-temporal graph network to generate a joint feature vector;

[0044] A state prediction module that performs health state prediction on the joint feature vector based on a dynamic digital twin model and outputs a health score, a fault probability, and a confidence interval;

[0045] A fault location module that uses a causal reasoning module to analyze the fault propagation path, locate the root cause of the fault, and generate an interpretability report;

[0046] An early warning module that triggers an early warning or maintenance decision based on the prediction result.

[0047] A computer device comprising

[0048] A memory in which computer-readable instructions are stored;

[0049] A processor that, when executing the computer-readable instructions, implements the steps of the hydroturbine state intelligent prediction method described in any one of the above.

[0050] A computer-readable storage medium having computer-readable instructions stored thereon, and when the computer-readable instructions are executed by a processor, implementing the steps of the hydroturbine state intelligent prediction method described in any one of the above.

[0051] Compared with the prior art, the embodiments of the present application mainly have the following beneficial effects:

[0052] The present invention realizes the time synchronization and spatial correlation of multi-source sensor data through spatio-temporal alignment preprocessing (dynamic time warping and three-dimensional topological mapping), and solves the difficulty of multi-source heterogeneous data fusion; through the dynamic graph convolution and multi-head attention mechanism of the multi-modal spatio-temporal graph network, it fuses spatio-temporal features, combines the reinforcement learning of the causal reasoning unit to optimize the fault propagation weight, and realizes the dynamic modeling and path tracing of the fault coupling relationship, and solves the uncertainty of fault propagation; through the knowledge distillation model compression of the dynamic digital twin and the edge-side FPGA deployment, it realizes high-frequency data low-latency real-time prediction and solves the contradiction between computing resources and real-time performance; through the confidence interval quantization of the Bayesian neural network and the root cause analysis of the SHAP value gradient backpropagation, it realizes the uncertainty evaluation of the prediction result and the output of the interpretability logic chain, and solves the problems of insufficient reliability and black-box decision-making. Each step forms a closed loop of "data fusion - dynamic modeling - lightweight reasoning - causal explanation", systematically overcoming the core technical bottleneck of hydroturbine state prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the solutions in the present application, the following will briefly introduce the drawings required for the description of the embodiments of the present application. Obviously, the drawings in the following description are some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0054] Figure 1 It is a flowchart of the hydroturbine state intelligent prediction method in this embodiment;

[0055] Figure 2It is the structural block diagram of the intelligent prediction system for the water turbine state in this embodiment;

[0056] Figure 3 It is the basic structural block diagram of the computer device in this embodiment. Specific implementation manners

[0057] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0058] As Figure 1 shown, this embodiment proposes a method for intelligent prediction of the water turbine state, including the following steps:

[0059] S100. Collect multi-source sensor data of the water turbine, including vibration, temperature, pressure, flow rate and electrical parameters.

[0060] Real-time collect the operation data of the water turbine through multi-source sensors (vibration, temperature, pressure, flow rate and electrical parameters), ensure that the monitoring covers multi-dimensional states such as machinery, electricity, and fluid, provide a comprehensive data basis for subsequent analysis, thereby improving the comprehensiveness and reliability of fault detection, and avoiding missed detection problems caused by the limitations of single-sensor data.

[0061] S200. Perform spatio-temporal alignment preprocessing on the multi-source sensor data to generate a spatio-temporal correlation data set.

[0062] Use the following dynamic time warping (DTW) to align sensor data with different sampling frequencies, and define the spatial adjacency relationship based on the sensor installation positions to solve the spatio-temporal misalignment problem of multi-source data, generate a spatio-temporal correlation data set, so that subsequent feature extraction can accurately reflect the true operating state of the device, and improve the adaptability of the model to non-stationary signals (such as vibration under variable working conditions).

[0063] S300. Extract and fuse the features of the spatio-temporal correlation data through a multi-modal spatio-temporal graph network to generate a joint feature vector.

[0064] Extract the spatio-temporal features of multi-modal data such as vibration and temperature through the spatio-temporal graph network (GCN-TCN), and use the following multi-head attention mechanism to fuse cross-modal correlations to generate a joint feature vector, effectively capturing the propagation law of fault features in the time and space dimensions, and significantly enhancing the sensitivity to early weak faults (such as slight wear of bearings).

[0065] S400. Perform a health state prediction on the joint feature vector based on the dynamic digital twin model, and output a health score, a fault probability and a confidence interval.

[0066] Perform health scoring and fault probability prediction on the joint feature vector based on the lightweight digital twin model, and output the confidence interval. Calibrate the model parameters through the following online Kalman filter to ensure that the prediction results are updated synchronously with the actual operating state, achieve high-precision and low-latency real-time state assessment, support edge device deployment, and reduce the dependence on cloud computing power.

[0067] S500. Use the causal inference unit to analyze the fault propagation path, locate the root cause of the fault, and generate an interpretability report.

[0068] Combine the following causal graph optimized by reinforcement learning and the Bayesian neural network to analyze the fault propagation path and quantify the uncertainty, locate the root cause (e.g., abnormal vibration is due to bearing wear rather than rotor imbalance), and generate an interpretability report, enabling maintenance personnel to quickly understand the fault logic and formulate targeted maintenance strategies.

[0069] S600. Trigger an early warning or maintenance decision according to the prediction results.

[0070] Automatically trigger an early warning or maintenance work order according to the following health score, fault probability, and confidence interval, realize the closed-loop management from prediction to execution, avoid manual misjudgment or delayed response, and at the same time extend the service life of the equipment by dynamically optimizing the maintenance cycle.

[0071] In this embodiment, in step S200, the spatio-temporal alignment preprocessing includes:

[0072] Perform time alignment on sensor data with different sampling frequencies using dynamic time warping.

[0073] The implementation steps of dynamic time warping are as follows:

[0074] S211. Input data: Assume there are two sensor data sequences with different sampling frequencies, for example:

[0075] Low-frequency signal (such as temperature, sampling frequency 1Hz): T = [t1, t2,..., t

[0077] , i ,

[0076] , N ,

[0079] ,

[0078] , , , i , M , , , , a total of M sampling points;

[0076] High-frequency signal (such as vibration, sampling frequency 100Hz): V = [v1, v2,..., v N , a total of N sampling points.

[0077] S212. Construct the cumulative distance matrix:

[0078] Calculate the local distance d(i, j) = |t i - v i | for each pair of time points (i, j);

[0079] Initialize the cumulative distance matrix where D(1, 1) = d(1, 1);

[0080] Recursive filling matrix:

[0081] D(i, j) = d(i, j) + min(D(i - 1, j), D(i, j - 1), D(i - 1, j - 1)),

[0082] where D(i, j) represents the minimum cumulative cost from the starting point (1, 1) to the point (i, j).

[0083] S213. Backtracking the optimal alignment path:

[0084] Search backward from D(M, N) for the minimum cumulative path to generate the alignment mapping relationship {(i1, j1), (i2, j2), …, (i K , j K ))} (K is the total length of the alignment path), where each i k corresponds to the low-frequency time point and j k corresponds to the best matching position of the high-frequency time point.

[0085] S214. Data alignment:

[0086] Segment and aggregate the high-frequency signal V according to the alignment path (such as taking the mean value) to generate a sequence synchronized with the low-frequency signal

[0087] Dynamic Time Warping (DTW) solves the alignment distortion problem of traditional linear interpolation methods under non-stationary signals (such as sudden increase in turbine vibration caused by load mutation) by adaptively bending the time sequence path, and can accurately match the time sequence characteristics of data with different sampling frequencies (such as phase alignment of fault impact waveforms), avoiding feature fusion errors caused by time misalignment, thereby improving the sensitivity of subsequent fault detection models to early abnormal signals.

[0088] Define the node topology structure based on the spatial adjacency relationship of sensor installation positions.

[0089] The specific steps to define the node topology structure are as follows:

[0090] S221. Obtain the sensor position coordinates:

[0091] According to the turbine design drawings or on-site measurements, record the three-dimensional installation coordinates of each sensor (such as the bearing sensor coordinates (x1, y1, z1), the runner sensor coordinates (x2, y2, z2), etc.)

[0092] S222. Calculate the Euclidean distance:

[0093] For any two sensor nodes i and j, calculate their spatial distance:

[0094]

[0095] Among them, (x i , y i , z i ) are the three-dimensional installation coordinates of sensor i; (x j , y j , z j ) are the three-dimensional installation coordinates of sensor j; d ij is the spatial distance between sensors i and j.

[0096] S223. Define the adjacency relationship:

[0097] Set the distance threshold d th . If d ij ≤ d th , then mark it as an adjacent node (adjacency matrix A ij = 1); otherwise A ij = 0 (A ij is the connection status of nodes i and j in the adjacency matrix, 1 means connected, and 0 means disconnected).

[0098] S224. Construct the topological structure:

[0099] Generate an adjacency matrix A ∈ {0, 1} N×N , as the spatial relationship input of the spatio-temporal graph network.

[0100] Based on the definition of the spatial adjacency relationship of the physical installation positions of the sensors, directly using the actual layout information of the equipment to construct the topological structure not only reduces the implementation cost but also enhances the model's ability to capture actual physical associations (such as the fault propagation of adjacent components being more easily associated), thereby improving the fault location accuracy; at the same time, this method is also applicable to the equipment of old units, expanding the universality of the technical solution.

[0101] In this embodiment, in step S300, the multi-modal spatio-temporal graph network includes:

[0102] Single-modal feature extraction unit: Perform time-frequency domain decomposition on the vibration signal, and perform statistical quantity and frequency domain analysis on the temperature and pressure signals.

[0103] Specifically, perform wavelet packet decomposition on the vibration signal, and calculate the energy entropy of the k-th sub-frequency band. The calculation formula is as follows:

[0104]

[0105] Among them, W k.i is the coefficient of the i-th node in the k-th sub-frequency band after wavelet packet decomposition; N is the number of signal segments.

[0106] Physical meaning of wavelet packet decomposition of vibration signals: According to the Nyquist sampling theorem, a 1 kHz vibration signal is decomposed into 4 layers of wavelet packets, generating 16 sub-bands (0 - 62.5 Hz, 62.5 - 125 Hz, …, 937.5 - 1000 Hz), covering key frequency bands such as blade passing frequency and bearing fault characteristic frequency. In the normal state, the energy entropy is evenly distributed in each frequency band; when there is an inner ring fault in the bearing, the energy entropy in the high-frequency band (such as 500 - 1000 Hz) increases significantly, thus judging the occurrence of the fault.

[0107] By performing frequency domain decomposition on the vibration signal, the energy entropy quantifies the complexity of the frequency band, thereby effectively distinguishing normal vibration and fault impact signals.

[0108] Cross-modal fusion unit: Learn the correlation weights between modalities through the multi-head attention mechanism.

[0109] The cross-modal fusion module uses the multi-head attention mechanism to map the feature vectors of different modalities (such as vibration, temperature, pressure) to query (Query), key (Key), and value (Value) matrices respectively, and calculates the attention weights between modalities. The specific steps are as follows:

[0110] S321: Perform a linear transformation on the features of each modality to generate independent Q, K, and V matrices.

[0111] S322: Split the matrix by attention heads, calculate the attention scores of each head respectively and fuse them with weights.

[0112] S323: Concatenate the multi-head outputs and generate a joint feature vector through a fully connected layer.

[0113] The calculation formula for the attention weights between modalities is as follows:

[0114]

[0115] Among them, Q is the query matrix (Query); K is the key matrix (Key); V is the value matrix (Value); d k is the single-head attention dimension; QK T is the similarity matrix between the query and the key, and softmax is the normalization function.

[0116] The cross-modal fusion unit dynamically captures the non-linear correlation relationships among modalities such as vibration, temperature, and pressure through the multi-head attention mechanism, solving the problem of insufficient feature representation caused by manually designed weights or simple concatenation in traditional methods. Its advantages are as follows: it can adapt to changes in the correlation strength among modalities under different working conditions (for example, the correlation between temperature and vibration increases under high load), and at the same time extracts multi-dimensional interaction features through parallel multi-head calculations, significantly enhancing the representation ability of the joint feature vector for complex fault modes, thereby enhancing the detection sensitivity and classification accuracy of the model for coupled faults, and finally achieving high-precision prediction of the health status of the water turbine.

[0117] Spatio-temporal graph convolution unit: Based on the dynamic graph structure, it models the spatio-temporal dependence relationships of multi-source data, and the edge weights of the dynamic graph structure are dynamically updated by the physical transmission path and data correlation.

[0118] The spatio-temporal graph convolution unit models the spatio-temporal dependence relationships of multi-source data through the dynamic graph structure, dynamically updates the edge weights using the physical transmission path and data correlation, and combines the multi-head attention mechanism to achieve cross-modal feature fusion, effectively solving the problem of insufficient spatio-temporal correlation modeling of heterogeneous data in traditional methods; its beneficial effects are as follows: through dynamic time warping, it precisely aligns the sensor data with different sampling frequencies in time, and constructs a node topology structure based on the spatial adjacency relationship of the sensor installation positions, enabling the wavelet packet energy entropy features of vibration signals and the statistical features of temperature and pressure signals to be correlated and analyzed in a unified spatio-temporal framework, not only improving the sensitivity of early fault feature extraction, but also enhancing the generalization ability of the model under variable working conditions through the weight balance of physical constraints and data-driven (α coefficient optimization), and at the same time, the spatio-temporal correlation features output by this module provide high-precision input for the subsequent digital twin model, forming a closed-loop optimization chain from data alignment to feature extraction.

[0119] In this embodiment, in step S400, the dynamic digital twin model is implemented in the following manner:

[0120] Embed the simplified mechanism model of the water turbine to generate benchmark health features.

[0121] Specifically, based on the physical characteristics of the water turbine (such as blade geometric parameters, material properties) and operating mechanisms (fluid mechanics, structural dynamics), a simulation model coupling the simplified Navier-Stokes equation and rotor dynamics equation is established;

[0122] Run the simulation model under normal working conditions to generate benchmark health features such as vibration spectrum, temperature distribution, and pressure fluctuation;

[0123] Align the simulation results with the actual sensor data in the feature space, and optimize the simulation model parameters (such as friction coefficient, cavitation rate) by minimizing the mean square error.

[0124] By embedding a simplified mechanism model and combining physical laws to generate benchmark health features, it provides prior knowledge support for data-driven models, effectively making up for the deficiency of actual fault data, enhancing the generalization ability of the model in variable working conditions or small-sample scenarios, and avoiding the misjudgment risk caused by training data deviation in pure data-driven methods.

[0125] Adopt knowledge distillation technology to compress the multi-modal spatio-temporal graph network and generate a lightweight student model.

[0126] Specifically, train the original multi-modal spatio-temporal graph network (teacher model) to enable it to fully learn the spatio-temporal correlation and fault patterns of multi-source data;

[0127] Construct a student model with a streamlined structure (such as reducing the number of graph convolution layers and the number of attention heads), and design a weighted combination of a feature matching loss function (L2 distance) and a prediction error loss function;

[0128] Use the intermediate layer features (such as spatio-temporal graph convolution output) and prediction results of the teacher model as supervision signals to guide the training of the student model until the student model reaches a prediction accuracy similar to that of the teacher model.

[0129] Through knowledge distillation technology, transfer the knowledge of the complex teacher model to the lightweight student model. On the premise of retaining the multi-modal spatio-temporal correlation modeling ability, significantly reduce the model parameter quantity and computational complexity, enable the model to adapt to the limited computing power resources of edge devices, and solve the problem that traditional deep learning models are difficult to deploy in real time due to excessive computational resource requirements.

[0130] Deploy the student model on edge devices and combine Kalman filtering to calibrate model parameters online.

[0131] Specifically, convert the student model into a hardware description language executable by FPGA (such as Verilog), and optimize the parallelism of computing units and memory occupancy;

[0132] Deploy the model on edge devices (such as industrial PLCs), receive sensor data in real time and output prediction results;

[0133] Design a Kalman filter observer, use the actual sensor data as the observation value and the model prediction value as the state quantity, and dynamically update model parameters (such as fault thresholds, weight coefficients) to reduce the model deviation caused by environmental interference and equipment aging.

[0134] Achieve local real-time inference through edge deployment, avoid the response lag caused by cloud transmission, and at the same time combine Kalman filtering to calibrate model parameters online, continuously adapt to the dynamic operating state of the water turbine (such as load fluctuations, component wear), improve the long-term stability and environmental robustness of the prediction model, and ensure the timeliness of early warning under complex working conditions.

[0135] The dynamic digital twin model achieves lightweighting through the mechanism model from the benchmark to knowledge distillation, and from knowledge distillation lightweighting to edge deployment and online calibration to ensure real-time performance and adaptability, forming a closed loop, layer by layer solving the conflicts of data fusion, computing resources and dynamic working conditions, and systematically improving the engineering practicability of the prediction method.

[0136] In this embodiment, in step S500, the causal reasoning unit includes:

[0137] Construct an initial causal graph based on domain knowledge and dynamically optimize the weights of the fault propagation paths through reinforcement learning.

[0138] Specifically, define the initial causal graph nodes (fault types) and directed edges (propagation directions) according to the water turbine fault mechanism (such as bearing wear causing shaft eccentricity → rotor imbalance → increased vibration);

[0139] Convert the causal graph into a Markov decision process (MDP), where the state is the current sensor features, the action is to adjust the edge weights of the causal graph, and the reward function is the negative value of the prediction error;

[0140] Use a deep Q-network (DQN) to train the agent, and dynamically optimize the edge weights through interaction with the environment (measured data) until the prediction accuracy of the causal graph converges.

[0141] By combining domain knowledge to define the initial causal graph structure and using reinforcement learning to adaptively optimize the weights of the fault propagation paths, it not only retains the reliability of expert experience but also overcomes the limitations of traditional methods that rely on fixed causal relationships, significantly improving the recognition accuracy of complex coupled faults (such as mechanical-electrical combined faults). At the same time, through dynamic weight update, it adapts to the differences in different water turbine models and operating conditions, enhancing the generalization ability of the model.

[0142] Use a Bayesian neural network to quantify the uncertainty of the prediction results and generate a confidence interval.

[0143] Specifically, construct a Bayesian neural network (BNN), replace the weights of the fully connected layers with probability distributions (such as Gaussian distributions), and use variational inference to approximate the posterior distribution;

[0144] Introduce the KL divergence loss in the training stage to constrain the deviation between the weight distribution and the prior distribution;

[0145] In the inference stage, calculate the mean and variance of the prediction results through Monte Carlo sampling (such as 100 random forward propagations) to generate a 95% confidence interval.

[0146] Quantify the uncertainty of the prediction results through a Bayesian neural network, providing confidence interval references for operation and maintenance personnel (such as "the bearing wear probability is 80% ± 5%"), avoiding the misjudgment risks caused by traditional point estimation models ignoring data noise and model bias. At the same time, support dynamic risk assessment (such as high-confidence faults triggering emergency shutdowns), improving decision-making reliability and reducing economic losses caused by sudden failures.

[0147] Locate the root cause of the fault and trace back the maximum probability propagation chain through gradient backpropagation and SHAP value analysis.

[0148] Specifically, calculate the gradient of the prediction result (such as the bearing wear probability) with respect to the input features (such as vibration spectrum energy), and select the top 10% of the key features ranked by the absolute value of the gradient.

[0149] Analyze the contribution degree of each feature to the prediction result based on SHAP (Shapley Additive Explanations) values, and combine the causal graph node weights for ranking.

[0150] Traverse the node path with the highest contribution degree in reverse along the causal graph (such as vibration anomaly → shaft eccentricity → bearing wear), and generate a visual fault propagation chain report.

[0151] Jointly locate the root cause of the fault through gradient backpropagation and SHAP value analysis, trace back the propagation path in combination with the causal graph, and transform the output of the traditional black box model into an interpretable fault logic chain (such as "blade cavitation causes abnormal flow, which in turn leads to rotor vibration"), helping operation and maintenance personnel quickly understand the fault mechanism and formulate targeted maintenance strategies. At the same time, enhance the transparency and user trust of the technical solution through a visual report.

[0152] The causal reasoning unit forms a progressive analysis chain from dynamic optimization of the causal graph to uncertainty quantification to root cause visualization tracing. It delves deeper layer by layer from fault propagation modeling, risk quantification to decision support, systematically solving the problems of "high fault coupling uncertainty" and "lack of interpretability" in the background technology, achieving a leap from "whether there is a fault" to "why there is a fault" in the prediction results, and promoting the transformation of hydro turbine operation and maintenance from experience-driven to data-knowledge dual-driven mode.

[0153] In this embodiment, it also includes a transfer learning step:

[0154] Pre-train a multi-modal spatio-temporal graph network using simulation data of multiple types of hydro turbines.

[0155] Specifically, establish a parametric simulation model covering different types of hydro turbines such as Francis turbines, Kaplan turbines, and tubular turbines, and set different operating conditions (such as head height, load rate, and rotational speed).

[0156] Generate multi-source simulation data (vibration, temperature, pressure) through numerical simulation, inject typical fault modes (such as bearing wear, blade cavitation), and construct a labeled training set;

[0157] Initialize the multi-modal spatio-temporal graph network (MTSG), input the simulation data into the spatio-temporal alignment module, and train the network to learn the common spatio-temporal features and fault modes across models;

[0158] Freeze the underlying parameters of the network (such as the spatio-temporal graph convolution layer), and save the pre-trained model weights.

[0159] Through pre-training with multi-model turbine simulation data, use virtual data to cover different structures, operating conditions, and fault scenarios, break through the limitation of scarce measured data of a single turbine, enable the model to learn a universal spatio-temporal feature expression, significantly improve the initial performance of the model on the target turbine, reduce the dependence on labeled data of the target device at the same time, and shorten the engineering deployment cycle.

[0160] Fine-tune the pre-trained model with the measured data of the target turbine.

[0161] Specifically, collect multi-source sensor data of the target turbine in the real operating environment (including at least 1 week of normal operating condition data and a small amount of fault segments);

[0162] Perform spatio-temporal alignment, noise filtering, and feature standardization processing on the measured data that is consistent with the simulation data;

[0163] Load the pre-trained model weights, unfreeze and retrain the top layer of the network (such as the classifier layer), and use a small learning rate (such as 1e-5) to optimize the model to adapt to the characteristics of the target device;

[0164] Evaluate the performance of the fine-tuned model through cross-validation. If the accuracy is insufficient, gradually unfreeze more underlying network layers for iterative training.

[0165] Based on the transfer learning fine-tuning of the pre-trained model, use a small amount of measured data of the target turbine to quickly adapt to the individual differences of the device and the on-site environmental noise, solve the domain drift problem between the simulation data and the real scenario (such as the sediment wear effect not considered in the simulation), improve the prediction accuracy in the target scenario while ensuring the generalization ability, achieve the balance of "general model - special optimization", and reduce the data collection and annotation costs.

[0166] Transfer learning forms a "simulation - measurement" closed-loop through the synergistic effect of pre-training and fine-tuning. The combination of the two significantly reduces the requirement for the historical fault data volume of the target turbine, supports rapid cross-device deployment, and is especially suitable for new power stations with few fault cases or the renovation scenarios of old units.

[0167] In this embodiment, in step S600, the prediction results include:

[0168] Health score, generated by linearly weighted fusion of the degradation degrees of multi-modal features.

[0169] Specifically, extract the features of each modality (such as vibration energy entropy, temperature trend slope, pressure harmonic amplitude) from the multi-modal spatio-temporal graph network and normalize them to the interval [0, 1];

[0170] Based on expert experience or genetic algorithm to optimize the weight coefficients (such as vibration weight 0.5, temperature 0.3, pressure 0.2), calculate the weighted sum: where HI is the quantitative value of the comprehensive health state of the water turbine, and the range is [0, 1]; w i is the contribution weight of the i-th modality feature (such as vibration, temperature, pressure) to the overall health score; f i is the current feature value; f i,healthy and f i,faulty are the health and failure reference values.

[0171] Set a threshold (such as HI < 0.7 to trigger an alarm), and output the health score in real time.

[0172] By linearly weighted fusion of the degradation degrees of multi-modal features, complex multi-dimensional data is transformed into a single quantifiable health score, which intuitively reflects the overall degradation state of the water turbine, solves the subjectivity defect of the traditional method relying on manual experience for comprehensive evaluation, and at the same time supports horizontal comparison of the health states of different devices (such as the priority ranking of multiple units), provides a unified decision-making basis for preventive maintenance, and reduces the over-maintenance cost caused by misjudgment of local faults.

[0173] Failure probability, the predicted probabilities of bearing wear, blade cavitation, and rotor imbalance output by the classification model.

[0174] Specifically, construct a multi-label classification model, with the input being the joint feature vector of the multi-modal spatio-temporal graph network, and the output layer using the Sigmoid activation function;

[0175] Use the cross-entropy loss function to train the model, and the optimization goal is to minimize the prediction errors of fault categories such as bearing wear, blade cavitation, and rotor imbalance;

[0176] In the inference stage, classify the input features and output the occurrence probabilities of each fault type.

[0177] By synchronously outputting the probabilities of multiple fault types through the multi-label classification model, the occurrence possibilities of different faults are accurately quantified, overcoming the limitation of the traditional binary classification model that cannot handle multi-fault coupling scenarios, helping the operation and maintenance personnel quickly lock in the high-probability fault types (such as giving priority to dealing with bearing wear rather than low-probability rotor imbalance), shortening the fault diagnosis time and improving the maintenance efficiency.

[0178] Confidence interval, the upper and lower bounds of 95% confidence calculated based on the Monte Carlo Dropout method.

[0179] Specifically, during the inference stage of the Bayesian neural network, keep the Dropout layer and set the random dropout rate;

[0180] Perform T forward propagations on the same input data and record the failure probability of each output;

[0181] Calculate the mean of the T results as the predicted value, and the standard deviation as the uncertainty measure, and generate a 95% confidence interval: Confidence interval = [μ - 1.96σ, μ + 1.96σ], where μ is the arithmetic mean of all output results after performing T Monte Carlo Dropout forward propagations on the same input data; σ is the standard deviation of the T output results, which is used to quantify the uncertainty of the model prediction.

[0182] Quantify the uncertainty of the model prediction through Monte Carlo Dropout, attach a confidence interval to the failure probability (such as "bearing wear probability 0.82 ± 0.06"), reveal the credibility of the prediction result, avoid misleading decisions with a single probability value (such as rechecking low-confidence results), and at the same time support dynamic adjustment of maintenance strategies (such as immediate shutdown for high-confidence failures and enhanced monitoring for low-confidence ones), significantly improving the robustness of operation and maintenance decisions and reducing the risk of unplanned shutdowns caused by misjudgment.

[0183] The health score, failure probability, and confidence interval work together to construct a decision support system of "condition assessment - fault location - risk quantification". The combination of the three breaks through the limitations of traditional single-index prediction, forming a hierarchical and interpretable intelligent operation and maintenance closed-loop.

[0184] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Among them, Artificial Intelligence (AI) is a theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use knowledge to obtain the best results.

[0185] Artificial intelligence basic technologies generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction systems, and mechatronics. Artificial intelligence software technologies mainly include several major directions such as computer vision technology, robotics, biometric technology, speech processing technology, natural language processing technology, and machine learning / deep learning.

[0186] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through computer-readable instructions. These computer-readable instructions can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above methods. Among them, the aforementioned storage medium can be a non-volatile storage medium such as a magnetic disk, an optical disk, a read-only memory (ROM), etc., or a random access memory (RAM), etc.

[0187] It should be understood that although the steps in the flowchart of the accompanying drawings are shown in sequence according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless there is a clear indication in this article, the execution of these steps has no strict order restriction, and they can be executed in other orders. Moreover, at least some of the steps in the flowchart of the accompanying drawings may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. Their execution order is not necessarily sequential, but can be executed alternately or alternately with at least a part of other steps or sub-steps or stages of other steps.

[0188] Further referring to Figure 2 , as an implementation of the method shown above Figure 1 , the present application provides an intelligent prediction system for the state of a water turbine. The system embodiment corresponds to the method embodiment shown in Figure 1 , and this system can be specifically applied to various electronic devices.

[0189] As shown in Figure 2 , an intelligent prediction system for the state of a water turbine described in this embodiment includes:

[0190] A data acquisition module that acquires multi-source sensor data of the water turbine, including vibration, temperature, pressure, flow rate, and electrical parameters.

[0191] The data acquisition module real-time obtains multi-source sensor data such as vibration and temperature, provides comprehensive input for subsequent analysis, and solves the problem of incomplete feature coverage caused by a single data source in traditional methods.

[0192] A data preprocessing module that performs spatio-temporal alignment preprocessing on the multi-source sensor data to generate a spatio-temporal correlation data set.

[0193] The data preprocessing module uses spatio-temporal alignment technology to unify the spatio-temporal benchmarks of multi-modal data, eliminates information fragmentation caused by differences in sampling frequency and spatial distribution, and improves the accuracy of subsequent modeling.

[0194] The feature extraction module extracts and fuses the features of the spatio-temporal correlation data through a multi-modal spatio-temporal graph network to generate a joint feature vector.

[0195] The feature extraction module relies on a multi-modal spatio-temporal graph network (MTSG) to fuse cross-modal features such as vibration spectra and temperature trends, and captures the spatio-temporal correlation of fault propagation by combining a dynamic graph structure, breaking through the limitation of simple splicing of heterogeneous data by traditional methods and enhancing the representation ability of complex faults.

[0196] The state prediction module performs a health state prediction on the joint feature vector based on a dynamic digital twin model, and outputs a health score, a fault probability, and a confidence interval.

[0197] The state prediction module realizes real-time inference at the edge based on lightweight dynamic digital twin (DDT), and outputs a health score, a fault probability, and a confidence interval through the fusion of a mechanism model and data-driven, taking into account both prediction speed and accuracy, and solving the contradiction between high computing power dependence and real-time performance.

[0198] The fault location module analyzes the fault propagation path by using a causal reasoning module, locates the root cause of the fault, and generates an interpretable report.

[0199] The fault location module traces the fault propagation chain with the help of causal reasoning technology, quantifies the feature contribution degree by combining SHAP values, generates an interpretable report, solves the problem of opaque decision logic of the black box model, and improves the trust of maintenance personnel.

[0200] The warning module triggers a warning or a maintenance decision according to the prediction result.

[0201] The warning module dynamically triggers a hierarchical response strategy (such as warning, shutdown) according to the prediction result, evaluates the risk level by combining the confidence interval, and optimizes the allocation of maintenance resources.

[0202] Each module collaborates to form a closed loop of "data perception - feature fusion - state prediction - root cause tracing - decision response", systematically solving core problems such as difficult multi-source data fusion, high uncertainty of fault coupling, and insufficient real-time performance, realizing the transformation from passive maintenance to predictive maintenance, and significantly reducing the unplanned shutdown rate and operation and maintenance costs.

[0203] The hydro-turbine state intelligent prediction system provided by the embodiments of the present invention can implement all the processes of the hydro-turbine state intelligent prediction method in the above embodiments. The functions of each module in the device and the achieved technical effects are respectively the same as the functions and the achieved technical effects of the hydro-turbine state intelligent prediction method in the above embodiments, and will not be elaborated here.

[0204] To solve the above technical problems, the embodiments of the present application also provide a computer device. Specifically, please refer to Figure 3 , Figure 3This is the basic structural block diagram of the computer device in this embodiment.

[0205] The computer device 3 includes a memory 31, a processor 32, and a network interface 33 that are communicatively connected to each other through a system bus. It should be noted that only the computer device 3 with components 31-33 is shown in the figure, but it should be understood that it is not required to implement all the shown components, and more or fewer components can be alternatively implemented. Among them, those skilled in the art of this technology can understand that the computer device here is a device that can automatically perform numerical calculations and / or information processing according to pre-set or stored instructions, and its hardware includes but is not limited to microprocessors, application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), embedded devices, etc.

[0206] The computer device 3 can be a computing device such as a desktop computer, a notebook, a palm computer, and a cloud server. The computer device 3 can perform human-computer interaction with the user through a keyboard, a mouse, a remote control, a touchpad, or a voice control device, etc.

[0207] The memory 31 includes at least one type of readable storage medium, and the readable storage medium includes flash memory, a hard disk, a multimedia card, a card-type memory (such as an SD or DX memory, etc.), a random access memory (RAM), a static random access memory (SRAM), a read-only memory (ROM), an electrically erasable programmable read-only memory (EEPROM), a programmable read-only memory (PROM), a magnetic memory, a magnetic disk, an optical disk, etc. In some embodiments, the memory 31 can be an internal storage unit of the computer device 3, such as the hard disk or memory of the computer device 3. In other embodiments, the memory 31 can also be an external storage device of the computer device 3, such as a plug-in hard disk equipped on the computer device 3, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. Of course, the memory 31 can also include both the internal storage unit and the external storage device of the computer device 3. In this embodiment, the memory 31 is generally used to store the operating system and various application software installed on the computer device 3, such as computer-readable instructions of the intelligent prediction method for the water turbine state. In addition, the memory 31 can also be used to temporarily store various data that have been output or will be output.

[0208] In some embodiments, the processor 32 may be a Central Processing Unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chips. The processor 32 is generally used to control the overall operation of the computer device 3. In this embodiment, the processor 32 is used to run the computer-readable instructions stored in the memory 31 or process data, such as running the computer-readable instructions of the intelligent prediction method for the water turbine state.

[0209] The network interface 33 may include a wireless network interface or a wired network interface, and the network interface 33 is generally used to establish a communication connection between the computer device 3 and other electronic devices.

[0210] The computer device provided by the present application has the beneficial effects of the independent claim.

[0211] The present application also provides another implementation manner, that is, to provide a computer-readable storage medium storing computer-readable instructions, and the computer-readable instructions can be executed by at least one processor to enable the at least one processor to execute the steps of the intelligent prediction method for the water turbine state as described above.

[0212] The computer-readable storage medium provided by the present application has the beneficial effects of the independent claim.

[0213] Through the description of the above embodiments, those skilled in the art can clearly understand that the above embodiment methods can be implemented by means of software plus a necessary general hardware platform. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation manner. Based on such an understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art can be embodied in the form of a software product. The computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to enable a terminal device (which may be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the methods described in the various embodiments of the present application.

[0214] The above are only the preferred embodiments of the present application and are not intended to limit the present application. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. An intelligent prediction method for the state of a water turbine, characterized in that, It includes the following steps: S100. Collect multi-source sensor data of the water turbine, including vibration, temperature, pressure, flow rate, and electrical parameters; S200. Conduct spatio-temporal alignment preprocessing on the multi-source sensor data to generate a spatio-temporal correlation dataset; S300. Extract and fuse the features of the spatio-temporal correlation data through a multi-modal spatio-temporal graph network to generate a joint feature vector; S400. Based on a dynamic digital twin model, perform health status prediction on the joint feature vector, and output a health score, a fault probability, and a confidence interval; S500. Use a causal reasoning unit to analyze the fault propagation path, locate the root cause of the fault, and generate an interpretability report; S600. Trigger an early warning or a maintenance decision according to the prediction result.

2. The intelligent prediction method for the state of a hydraulic turbine according to claim 1, wherein In the step S200, the spatio-temporal alignment preprocessing includes: Adopt dynamic time warping to perform time alignment on sensor data with different sampling frequencies; Define a node topology structure based on the spatial adjacency relationship of the sensor installation positions.

3. The intelligent prediction method for the state of a water turbine according to claim 1, wherein In the step S300, the multi-modal spatio-temporal graph network includes: A single-modal feature extraction unit: perform time-frequency domain decomposition on the vibration signal, and perform statistical and frequency domain analysis on the temperature and pressure signals; A cross-modal fusion unit: learn the correlation weights between modalities through a multi-head attention mechanism; A spatio-temporal graph convolution unit: model the spatio-temporal dependence relationship of multi-source data based on a dynamic graph structure, and the edge weights of the dynamic graph structure are dynamically updated by the physical transmission path and data correlation.

4. The intelligent prediction method for the state of a hydraulic turbine according to claim 1, characterized in that, In the step S400, the dynamic digital twin model is implemented in the following way: Embed a simplified mechanism model of the water turbine to generate benchmark health features; Adopt knowledge distillation technology to compress the multi-modal spatio-temporal graph network to generate a lightweight student model; Deploy the student model on an edge device and combine it with Kalman filtering to calibrate the model parameters online.

5. The intelligent prediction method for the state of a hydraulic turbine according to claim 1, wherein In the step S500, the causal reasoning unit includes: Construct an initial causal graph based on domain knowledge and dynamically optimize the fault propagation path weights through reinforcement learning; Adopt a Bayesian neural network to quantify the uncertainty of the prediction result and generate a confidence interval; Through gradient backpropagation and SHAP value analysis, locate the root cause of the fault and trace back the maximum probability propagation chain.

6. The intelligent prediction method for the state of a water turbine according to claim 1, characterized in that It also includes a transfer learning step: Pre-train the multi-modal spatio-temporal graph network using simulation data of multiple models of water turbines; Fine-tune the pre-trained model with the measured data of the target water turbine.

7. The intelligent prediction method for the state of a water turbine according to claim 1, characterized in that In the step S600, the prediction results include: A health score, which is generated by linearly weighted fusion of the degradation degrees of multi-modal features; A fault probability, which is the prediction probability of bearing wear, blade cavitation, and rotor imbalance output by the classification model; A confidence interval, which is the upper and lower bounds of the 95% confidence level calculated based on the Monte Carlo Dropout method.

8. An intelligent prediction system for the state of a water turbine, characterized in that, It includes: A data acquisition module, which collects multi-source sensor data of the water turbine, including vibration, temperature, pressure, flow rate, and electrical parameters; A data preprocessing module, which conducts spatio-temporal alignment preprocessing on the multi-source sensor data to generate a spatio-temporal correlation dataset; A feature extraction module, which extracts and fuses the features of the spatio-temporal correlation data through a multi-modal spatio-temporal graph network to generate a joint feature vector; A state prediction module that predicts the health state of the combined feature vector based on a dynamic digital twin model and outputs a health score, a failure probability, and a confidence interval; A fault location module that analyzes the fault propagation path using a causal inference module, locates the root cause of the fault, and generates an interpretability report; An early warning module that triggers an early warning or a maintenance decision based on the prediction result.

9. A computer device, characterized in that, Comprising A memory that stores computer-readable instructions; A processor that, when executing the computer-readable instructions, implements the steps of the water turbine state intelligent prediction method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, Computer-readable instructions are stored on the computer-readable storage medium, and when the computer-readable instructions are executed by a processor, the steps of the water turbine state intelligent prediction method according to any one of claims 1 to 7 are implemented.

Images