Turbine runner fault early warning and self-learning identification method and device and storage medium

Abstract

The present application relates to the technical field of water turbine runner fault early warning, in particular to a water turbine runner fault early warning and self-learning identification method and device and computer storage medium. The present application collects multi-source sensing data, extracts time domain, frequency domain features and physical correlation features representing the correlation between water guide swing and active power, and constructs a multi-dimensional feature vector; a deep neural network early warning model is constructed based on normal and pre-fault state data, the multi-dimensional feature vector is input into the model to obtain the fault probability, and the multi-level early warning is triggered combined with the physical correlation features; when the early warning is triggered, the runner image is collected, and the image recognition model is used for defect identification; the early warning data confirming the defect is fed back to the early warning model for retraining, and the image recognition model is continuously optimized through artificial feedback and incremental learning, forming a closed-loop optimization. Through multi-source data fusion, physical correlation feature introduction and closed-loop self-learning, the present application improves the early warning accuracy and advance time, and reduces the false alarm and missed alarm rate.

Description

Technical Field

[0001] This invention relates to the field of turbine runner fault early warning technology, and in particular to a turbine runner fault early warning and self-learning identification method, device, equipment, and computer storage medium. Background Technology

[0002] As a key component of hydroelectric generator sets, the turbine runner endures alternating loads and hydraulic impacts under complex hydraulic environments for extended periods. Its operational status directly impacts the safety and economy of the unit. In related technologies, a preliminary diagnostic system for runner defects has been constructed using online vibration / sway monitoring, image visual inspection, and data-driven machine learning methods. Specifically, this system covers the entire process from vibration signal threshold alarms to neural network-based fault classification, and post-shutdown image recognition using industrial cameras, including key steps such as time-domain and frequency-domain feature extraction, single-data-source model training, and manual image interpretation.

[0003] However, existing turbine runner fault diagnosis methods directly use a single vibration index for threshold judgment without integrating multi-source sensor data and physical correlation features. This may lead to an inability to distinguish between different fault sources such as hydraulic imbalance and mechanical imbalance, thus affecting the accuracy of the diagnosis. Furthermore, image-based visual detection relies on shutdown and drainage conditions, making continuous monitoring during operation impossible. Traditional machine learning models also lack online self-learning capabilities, making it difficult to adapt to unit aging or changes in operating conditions. This results in insufficient early warning time and a high false alarm rate, seriously threatening the safe and stable operation of the hydro-generator unit. Summary of the Invention

[0004] Therefore, the technical problem to be solved by the present invention is to overcome the problems in the prior art that result in low accuracy of fault warning, insufficient advance time, high false alarm rate, and inability to achieve closed-loop diagnosis by combining online self-learning of the model with visual confirmation of defects due to reliance on a single data source and static model.

[0005] To address the aforementioned technical problems, this invention provides a method for early warning and self-learning identification of turbine runner faults, comprising:

[0006] Multi-source sensor data is collected during the operation of the water turbine. The multi-source sensor data is preprocessed and divided into operating conditions to extract multi-dimensional features. The multi-dimensional features include time-domain features, frequency-domain features, and physical correlation features. The physical correlation features are used to characterize the correlation between water conduction swing and active power. A deep neural network early warning model is constructed based on normal state data and pre-fault state data. The multi-dimensional features are input into the deep neural network early warning model to output the fault probability, and the physical correlation features are combined to trigger multi-level early warning. When the multi-level early warning at the preset level is triggered, the surface image of the turbine runner is collected after the unit is shut down. The image recognition model is used to identify defects in the surface image of the turbine runner to output the defect type and defect location. The early warning data confirming defects are fed back as new negative samples to the deep neural network early warning model for retraining. The image recognition model is continuously optimized through manual feedback and incremental learning to form a closed-loop optimization.

[0007] Preferably, the preprocessing and condition segmentation of the multi-source sensor data to extract multi-dimensional features includes: The raw data is cleaned to remove outliers; The operating data is divided into multiple detailed operating conditions based on the unit's active power, head, and guide vane opening. For each monitoring index under each working condition, the time domain characteristics, frequency domain characteristics, pressure pulsation characteristics, and temperature characteristics are extracted, and the correlation coefficient between the water conductance and active power is calculated as the physical correlation characteristics. All extracted features are combined into a multidimensional feature vector.

[0008] Preferably, the construction of the deep neural network early warning model based on normal state data and pre-fault state data includes: The feature vectors under each working condition are extracted from the normal state database as positive samples, and the feature vectors within a predetermined time window before the fault are extracted from the fault state database as negative samples to construct a training set. The deep neural network early warning model is trained using a binary cross-entropy loss function. The deep neural network early warning model includes an input layer, multiple hidden layers, and an output layer connected in sequence, wherein the output layer uses the Sigmoid activation function.

[0009] Preferably, the step of triggering multi-level early warnings by combining the physical association features includes: Obtain the normal feature distribution based on the current operating conditions to calculate the comprehensive anomaly degree of the current real-time feature vector; The real-time feature vector is input into the deep neural network early warning model to obtain the fault probability; Based on the numerical range of the fault probability and the comprehensive anomaly degree, and combined with the threshold judgment of the physical correlation characteristics, different severity levels of warning are triggered.

[0010] Preferably, the defect identification of the wheel surface image using an image recognition model includes: The acquired images of the wheel surface are input into an improved target detection model, which includes a backbone network, a neck network, and a head network. The backbone network includes an attention module for enhancing the focus on key features, the neck network adopts a multi-scale feature fusion structure and introduces an attention module, and the head network includes multiple detection heads. The model is optimized using a metric function based on the vertex distance and / or bounding box size between the predicted box and the ground truth box as the loss function.

[0011] Preferably, the continuous optimization of the image recognition model through manual feedback and incremental learning includes: Provides a human feedback interface for receiving verification and annotation of the recognition results; When the number of new samples reaches a preset threshold, incremental training is initiated to update the model parameters while maintaining the original recognition ability. Images whose recognition confidence level is within a preset range or that have been manually verified as incorrect are identified as difficult examples and given higher sampling weights in subsequent training.

[0012] Preferably, it is characterized by further comprising: The contribution of each feature to the model output is calculated using the SHAP value method. The contribution is used to quantify the degree of influence of the feature on the fault probability prediction. Based on the contribution level, fault symptom patterns are mined to generate an interpretable list of abnormal indicators and fault type inferences.

[0013] The present invention also provides a turbine runner fault early warning and self-learning identification device, comprising: The data acquisition and feature extraction module is used to acquire multi-source sensor data during the operation of the water turbine, preprocess the multi-source sensor data and classify the operating conditions to extract multi-dimensional features. The multi-dimensional features include time domain features, frequency domain features and physical correlation features. The physical correlation features are used to characterize the correlation between water conductance and active power. The early warning module is used to construct a deep neural network early warning model based on normal state data and pre-fault state data, input the multi-dimensional features into the deep neural network early warning model to output the fault probability, and trigger multi-level early warnings in combination with the physical correlation features. The image recognition module is used to collect images of the turbine surface after the unit is shut down when the multi-level early warning of the preset level is triggered, and to use the image recognition model to identify defects in the turbine surface images in order to output the defect type and defect location. The closed-loop optimization module is used to feed back the early warning data of confirmed defects as new negative samples to the deep neural network early warning model for retraining, and to continuously optimize the image recognition model through manual feedback and incremental learning to form a closed-loop optimization.

[0014] This invention also provides a turbine runner fault early warning and self-learning identification device, comprising: Memory, used to store computer programs; The processor is used to implement the steps of the above-mentioned method for early warning and self-learning identification of turbine runner faults when executing the computer program.

[0015] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described method for early warning and self-learning identification of turbine runner faults.

[0016] The technical solution of the present invention has the following advantages compared with the prior art: The turbine runner fault early warning and self-learning identification method described in this invention collects multi-source sensor data and extracts multi-dimensional features including physical correlation characteristics (such as features characterizing the correlation between water conduction swivel and active power). It then combines this with a deep neural network early warning model to predict fault probability and, based on this, integrates physical correlation features to trigger multi-level early warnings. This effectively overcomes the shortcomings of existing technologies where single-indicator judgment is easily interfered with and has a high false alarm rate, significantly improving the accuracy and lead time of fault early warnings. Simultaneously, after triggering an early warning, the method uses an image recognition model to visually confirm the fault, and feeds the confirmed fault data back to the early warning model as negative samples. Combined with human feedback and incremental learning, the image recognition model is continuously optimized, forming a closed-loop optimization system where data and images, early warning and identification mutually verify and continuously evolve. This significantly reduces the false alarm rate and the missed alarm rate, enhances the model's adaptability to changes in operating conditions and new fault types, and ensures the long-term reliability and intelligent level of turbine runner fault diagnosis. Attached Figure Description

[0017] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein: Figure 1 This is a flowchart illustrating the implementation of a turbine runner fault early warning and self-learning identification method provided by the present invention. Figure 2 This is a structural block diagram of a turbine runner fault early warning and self-learning identification device provided in an embodiment of the present invention. Detailed Implementation

[0018] The core of this invention is to provide a method, device, equipment, and computer storage medium for early warning and self-learning identification of turbine runner faults, so as to effectively solve the problems of low accuracy, insufficient advance time, high false alarm rate, and inability to achieve closed-loop diagnosis caused by single data source, lack of physical correlation features, and inability of model to learn on its own in the prior art.

[0019] To enable those skilled in the art to better understand the present invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please refer to Figure 1. Figure 1 The flowchart illustrates the implementation of a turbine runner fault early warning and self-learning identification method provided by this invention; the specific operation steps are as follows: S101: Collect multi-source sensor data during the operation of the water turbine, preprocess the multi-source sensor data and classify the operating conditions to extract multi-dimensional features. The multi-dimensional features include time domain features, frequency domain features and physical correlation features. The physical correlation features are used to characterize the correlation between water conduction swing and active power. The method first constructs a multi-dimensional sensor data acquisition system covering key components of the turbine to obtain various physical quantity signals reflecting the unit's operating status, including vibration, runout, pressure, temperature, and operating parameters such as active power, head, and guide vane opening. The acquired raw data is preprocessed to remove invalid data segments caused by sensor malfunctions or communication anomalies. Based on operating parameters such as active power, head, and guide vane opening, the operating data is divided into multiple fine-grained operating condition intervals, providing a unified benchmark for subsequent analysis. On this basis, for each monitoring signal under each operating condition, feature parameters characterizing the signal's statistical properties and energy distribution are extracted from both the time and frequency domains. Features reflecting hydraulic pulsation characteristics are extracted from the pressure signal, and features reflecting thermal state changes are extracted from the temperature signal. Furthermore, to reveal the physical mechanism of hydraulic imbalance faults, the correlation coefficient between the runner runout and active power is calculated. This coefficient quantifies the linear correlation between runout and load changes. When the absolute value of this coefficient exceeds a preset threshold, it can be used as an important indicator of abnormal flow field in the turbine blades. All the above features are combined into a high-dimensional feature vector, which comprehensively covers the unit's operating status information under specific operating conditions, providing physically interpretable input for subsequent fault early warning models. As one implementation method, the time-domain features may include mean, variance, standard deviation, peak value, peak-to-peak value, root mean square value, kurtosis, and skewness; the frequency-domain features may include spectral centroid, spectral dispersion, dominant frequency amplitude, and harmonic amplitude; the pressure pulsation features may include pulsation amplitude, pulsation frequency, and pulsation energy; the temperature features may include temperature rise rate, temperature difference, and temperature fluctuation amplitude; the correlation coefficient is calculated using the Pearson correlation coefficient formula, with a value range of [-1, 1], and is considered significantly correlated when the absolute value is not less than 0.6.

[0021] Through the above steps, a high-dimensional feature space integrating multi-source sensor information and physical correlation features was constructed. This feature space not only comprehensively reflects the multi-dimensional signal changes of the unit under normal and abnormal conditions, but also enhances the sensitivity and interpretability of the features to hydraulic imbalance faults by introducing the correlation coefficient between water conductance and load. This provides high-quality and high-discrimination input for subsequent deep learning early warning models, thus laying the data foundation for achieving early and accurate fault warning.

[0022] S102: Construct a deep neural network early warning model based on normal state data and pre-fault state data, input the multi-dimensional features into the deep neural network early warning model to output the fault probability, and trigger multi-level early warnings in combination with the physical correlation features; A deep neural network early warning model is constructed based on normal state data and pre-fault state data. This model takes multi-dimensional feature vectors as input and automatically learns the complex nonlinear mapping relationship between the normal state and the pre-fault state through a deep learning network structure, outputting a fault probability that characterizes the degree to which the current operating state deviates from normal. This fault probability serves as the core early warning indicator, comprehensively reflecting the abnormal information contained in multi-source sensor data. Furthermore, the correlation coefficient between water conductance and active power is combined. This correlation coefficient characterizes the significance of the specific physical fault symptom of hydraulic imbalance. By jointly judging the fault probability and the correlation coefficient, a multi-level early warning mechanism is triggered. Specifically, based on the different numerical ranges of the fault probability and whether the correlation coefficient reaches a preset threshold, the system can distinguish different severity levels of early warning, thereby achieving a tiered response from early attention to emergency alarm. As one implementation method, the deep neural network early warning model can contain multiple hidden layers. Each hidden layer uses the ReLU activation function and introduces the Dropout mechanism to prevent overfitting. The output layer uses the Sigmoid activation function to map the output to a fault probability value between 0 and 1. The model training can use the binary cross-entropy loss function and the Adam optimizer. The multi-level early warning can be set to the attention level, the suspicious level, and the dangerous level. The triggering of the suspicious level and the dangerous level must simultaneously meet the fault probability threshold and the correlation coefficient threshold conditions.

[0023] This step, by constructing a deep neural network early warning model and combining it with correlation coefficients for multi-level early warning, enables early and accurate identification of runner failures, significantly improving the lead time and accuracy of early warnings, while reducing the false alarm rate, and providing a reliable triggering basis for subsequent defect confirmation and closed-loop optimization.

[0024] S103: When the multi-level early warning of the preset level is triggered, the surface image of the turbine runner is collected after the unit is shut down, and the defect is identified by the image recognition model to output the defect type and defect location. When a preset warning level is triggered, it indicates that the turbine runner may have exhibited an abnormal condition requiring further confirmation. In this case, under safe conditions after the unit is shut down, a visual inspection of the runner surface is necessary to obtain direct evidence of defects. Specifically, a mobile image acquisition device capable of accessing the turbine's flow passage is deployed. When the unit is shut down and a safe operating environment is available, multi-angle, high-resolution image acquisition is performed on key parts of the runner, such as blade surfaces, blade roots, outlet edges, and inlet edges—areas prone to defects. This yields raw image data containing information about the runner's surface condition. The acquired raw image data is then input into a pre-trained image recognition model. This model, built on a deep learning architecture, automatically extracts visual features related to typical defect types such as cracks, cavitation, and spalling from the images. Based on these features, it locates and classifies defect areas in the images, ultimately outputting the specific type of each identified defect and its location on the runner surface. As one implementation method, when a suspicious or dangerous warning is triggered, an underwater drone can be used to enter the tailrace pipe and photograph the turbine after the unit is shut down. The collected images are then input into an improved YOLOv8 model for defect identification. This model processes the images through its backbone network, neck network, and head network, and optimizes them based on the MPDIoU loss function, thereby outputting the defect type, confidence level, and bounding box location.

[0025] This step enables a leap from indirect early warning based on sensor data to direct confirmation based on visual images, associating early warning signals with specific physical defect morphology and location. This provides an intuitive and accurate basis for subsequent maintenance decisions, significantly improving the reliability and relevance of fault diagnosis.

[0026] S104: The early warning data confirming the defects is fed back as new negative samples to the deep neural network early warning model for retraining, and the image recognition model is continuously optimized through manual feedback and incremental learning to form a closed-loop optimization.

[0027] In the closed-loop optimization process of early warning and identification, confirmed defect early warning data is used as new negative samples and fed back to the deep neural network early warning model for retraining, thereby continuously improving the model's ability to identify pre-fault state features. Specifically, after the existence of a defect is confirmed by the image recognition model, the system associates the multi-dimensional feature vectors of the early warning trigger time and the time window before and after it with the corresponding fault type, severity, and other information to form new sample data with fault labels. These new samples are included in the training set of the early warning model and used together with the original normal state data and pre-fault state data for incremental training or complete retraining of the model, enabling the model to learn newly emerging fault symptom patterns and enhance its sensitivity to similar faults and early warning accuracy. At the same time, the image recognition model is also continuously optimized through human feedback and incremental learning mechanisms. The human feedback interface allows technicians to review and label the recognition results, correct misidentifications, or supplement new defect types. These manually confirmed image data are added to the training set. The incremental learning strategy is initiated after a certain number of new samples have been accumulated. It fine-tunes the model parameters with a small learning rate to avoid catastrophic forgetting, allowing the model to gradually adapt to new defect patterns and changing operating conditions while maintaining its original recognition capabilities. Through the collaborative optimization of the aforementioned early warning model and image recognition model, a closed-loop system of early warning-recognition-feedback-retraining is formed. This enables the entire fault diagnosis system to continuously accumulate knowledge and improve performance during actual operation, achieving adaptive evolution.

[0028] This technical step feeds back the early warning data that confirms defects as negative samples to the early warning model, and combines human feedback with incremental learning to optimize the image recognition model, forming a complete closed-loop optimization mechanism. This significantly improves the system's adaptability to new faults and changes in operating conditions, reduces the false alarm rate and the missed alarm rate, and continuously enhances the accuracy and timeliness of fault diagnosis in practical applications.

[0029] Based on the above embodiments, in some embodiments, the preprocessing and condition segmentation of the multi-source sensing data to extract multi-dimensional features includes: The raw data is cleaned to remove outliers; The operating data is divided into multiple detailed operating conditions based on the unit's active power, head, and guide vane opening. For each monitoring index under each working condition, the time domain characteristics, frequency domain characteristics, pressure pulsation characteristics, and temperature characteristics are extracted, and the correlation coefficient between the water conductance and active power is calculated as the physical correlation characteristics. All extracted features are combined into a multidimensional feature vector.

[0030] In this embodiment, the specific implementation of preprocessing and classifying multi-source sensor data and extracting features to construct a multi-dimensional feature vector in step S101 is as follows: First, the raw data is cleaned, and outliers are identified and removed using the 3σ principle. That is, when the absolute value of the difference between a data point and the mean is greater than three times the standard deviation, it is determined to be an outlier and removed. Subsequently, the operating data is divided into multiple fine operating conditions according to the unit's active power P, head H, and guide vane opening Y. The active power is divided into 10 intervals at 0.1 times the rated power, the head is divided into 20 intervals at 0.05 times the maximum head, and the guide vane opening is divided into 20 intervals at 5%, for a total of 4000 operating conditions. Each operating condition is assigned a unique identifier Cijk. For each monitoring indicator under each operating condition, time-domain features, frequency-domain features, pressure pulsation features, and temperature features are extracted. The time-domain features include mean μx, variance σx², standard deviation σx, peak value xpeak, peak-to-peak value xpp, root mean square value xrms, kurtosis Kx, and skewness Sx, where mean μx = (1 / N)∑x(t), variance σx² = (1 / N)∑(x(t)-μx)², standard deviation σx = √(σx²), peak value xpeak = max(|x(t)|), peak-to-peak value xpp = max(x(t))-min(x(t)), root mean square value xrms = √((1 / N)∑x(t)²), and kurtosis Kx = (1 / N)∑(x(t)-μx). 4 / σx 4 The skewness Sx = (1 / N)∑(x(t)-μx)³ / σx³. Frequency domain features are extracted after obtaining the spectrum X(f) through Fourier transform of the signal. These include the spectral centroid fc = ∑f·|X(f)|² / ∑|X(f)|², the spectral dispersion fd = √(∑(f-fc)²·|X(f)|² / ∑|X(f)|²), the dominant frequency amplitude Amain = max(|X(f)|), and the harmonic amplitude Aharmonic = |X(2fmain)|. Pressure pulsation features include the pulsation amplitude pamp =ppp / 2, the pulsation frequency pfreq, and the pulsation energy penergy = ∫|p(t)-pmean|²dt. Temperature characteristics include the rate of temperature rise dT / dt = (T(t+Δt)-T(t)) / Δt, temperature difference ΔT = Tmax-Tmin, and temperature fluctuation amplitude Tamp = √((1 / N)∑(T(t)-Tmean)²). Furthermore, the correlation coefficient Rg,p between water conductivity and active power is calculated, with the formula Rg,p = ∑(Xi- (Yi- ) / √(∑(Xi- )²·∑(Yi- )²), where Xi is the water conductance swing of the i-th sampling, Let Yi be the average value of the power collected in the i-th sampling, and let Yi be the active power collected in the i-th sampling. The average value is calculated. Finally, all 128 extracted basic features are combined with one physical correlation feature Rg,p to construct a 129-dimensional feature vector F = [f1, f2, …, f129]T. Through the above processing, the cleaned data output from the previous sub-step is input into the working condition segmentation step, and the segmented working condition data is input into the feature extraction step. The final output feature vector provides accurate input for subsequent model construction and early warning analysis.

[0031] This specific implementation method significantly enhances the ability of the feature vector to represent the running state of the turbine by finely dividing the operating conditions and extracting comprehensive multi-dimensional features, especially by introducing the correlation coefficient between the water conduction swing and the active power as a physical correlation feature. This provides a more discriminative and physically interpretable input for the subsequent deep learning early warning model, thereby improving the accuracy and early warning of faults.

[0032] Based on the above embodiments, in some embodiments, the specific implementation of constructing a deep neural network early warning model based on normal state data and pre-fault state data in step S102 is as follows. First, from the normal state database... Feature vectors from each working condition are randomly selected as positive samples and labeled accordingly. ,in For operating condition indication, For timestamps, These are the feature vectors under normal conditions. Simultaneously, from the fault state database... Extract the time window before the fault The feature vectors are used as negative samples, and their labels are assigned. ,in The time elapsed since the fault occurred. Fault type The severity of the fault, Location of the fault. This represents the feature vector under fault conditions. The training set is constructed using the above method, and the number of positive and negative samples can be balanced using oversampling techniques, for example, taking 100,000 samples of each. The specific structure of this deep neural network early warning model includes: an input layer with a dimension of 129, used to receive 128 basic features and 1 physically related feature. eigenvectors Hidden layer 1 contains 512 neurons with ReLU activation and a dropout rate of 0.2; Hidden layer 2 contains 256 neurons with ReLU activation and a dropout rate of 0.2; Hidden layer 3 contains 128 neurons with ReLU activation and a dropout rate of 0.2; Hidden layer 4 contains 64 neurons with ReLU activation and a dropout rate of 0.2; Output layer contains 1 neuron with Sigmoid activation, used to output the fault probability. The model training uses the binary cross-entropy loss function $L = - \Big[ y \cdot \log(\hat{y}) + (1 - y) \cdot \log(1 - \hat{y}) \Big]$ for optimization, and the Adam optimizer is used for parameter updates. The update rule is as follows: first, calculate the first-order momentum. and second momentum ,in For the current gradient, , Then, the momentum is corrected for deviation to obtain... and Finally, update the model parameters. The learning rate , Through the above process, the trained early warning model can output accurate fault probabilities based on the input feature vectors.

[0033] This specific implementation method, through a clear positive and negative sample construction strategy, a refined deep neural network structure, and an efficient Adam optimization algorithm, ensures that the early warning model can fully learn the complex nonlinear correlation between the normal state and the pre-fault state, thereby significantly improving the accuracy and lead time of fault early warning and reducing the false alarm rate.

[0034] Based on the above embodiments, in some embodiments, constructing a deep neural network early warning model based on normal state data and pre-fault state data includes: The feature vectors under each working condition are extracted from the normal state database as positive samples, and the feature vectors within a predetermined time window before the fault are extracted from the fault state database as negative samples to construct a training set. The deep neural network early warning model is trained using a binary cross-entropy loss function. The deep neural network early warning model includes an input layer, multiple hidden layers, and an output layer connected in sequence, wherein the output layer uses the Sigmoid activation function.

[0035] Based on the above embodiments, in some embodiments, triggering multi-level early warnings in conjunction with the physical association features includes: Obtain the normal feature distribution based on the current operating conditions to calculate the comprehensive anomaly degree of the current real-time feature vector; The real-time feature vector is input into the deep neural network early warning model to obtain the fault probability; Based on the numerical range of the fault probability and the comprehensive anomaly degree, and combined with the threshold judgment of the physical correlation characteristics, different severity levels of warning are triggered.

[0036] In this embodiment, the specific implementation of triggering multi-level early warning in step S102 based on the correlation coefficient is as follows: During the online monitoring phase, the system receives and processes the real-time multi-dimensional feature vector constructed in step S101 in real time. This vector contains 128 basic features and one physical correlation feature, namely the correlation coefficient between the water conductance swing and active power. First, the system retrieves the corresponding feature statistical distribution, including the mean vector, from the normal state database based on the current operating conditions. and standard deviation vector ,in This represents the feature dimension index. Then, the anomaly score for each feature dimension is calculated. And further calculate the comprehensive anomaly degree. ,in In this embodiment, the total number of features is [number]. At the same time, the real-time feature vector The input is fed into a pre-trained deep neural network early warning model, which outputs a fault probability between 0 and 1. The system is based on failure probability. and comprehensive anomaly And combined with correlation coefficient It executes multi-level early warning logic. Specifically, when the conditions are met... or When the system determines it to be a level-of-concern warning, it will only record the information and issue a yellow alert, without triggering any audible or visual alarms. When the conditions are met... or And simultaneously satisfy the correlation coefficient When the condition is met, the system determines it to be a suspicious level warning, triggering an audible and visual alarm and issuing an orange alert. The introduction of the correlation coefficient in this condition aims to identify anomalies caused by hydraulic imbalance (such as uneven flow field in the impeller blades) to reduce the false alarm rate. or When the system determines the situation to be at a dangerous level, it triggers an emergency alarm and issues a red alert. The alarm level determination result is used as output to trigger subsequent steps of wheel image acquisition and defect identification.

[0037] This specific implementation method introduces a comprehensive anomaly degree. Physical correlation characteristics Threshold judgment enables the determination of fault probability. The supplementary verification and physical mechanism constraints effectively reduce the false alarms that may be generated by a single deep learning model. At the same time, the multi-level early warning mechanism provides differentiated response strategies for operation and maintenance personnel, significantly improving the reliability and practicality of the early warning system.

[0038] Based on the above embodiments, in some embodiments, the defect identification of the wheel surface image using an image recognition model includes: The acquired images of the wheel surface are input into an improved target detection model, which includes a backbone network, a neck network, and a head network. The backbone network includes an attention module for enhancing the focus on key features, the neck network adopts a multi-scale feature fusion structure and introduces an attention module, and the head network includes multiple detection heads. The model is optimized using a metric function based on the vertex distance and / or bounding box size between the predicted box and the ground truth box as the loss function.

[0039] In this embodiment, when a suspicious or dangerous level warning is triggered, images of the turbine surface are collected by an underwater drone after the unit is shut down. The collected images are then input into an improved YOLOv8 model for defect identification. This improved YOLOv8 model includes a backbone network, a neck network, and a head network. Specifically, the backbone network comprises 5 Conv modules, 4 C2f modules, and 1 SPPCSPC+SimAM module. The SimAM attention module is used to calculate attention weights, and its calculation formula is as follows: ,in , This module enhances the model's focus on key features. The Neck network employs an FPN+PAN structure and introduces four GAM attention modules to fuse multi-scale features and improve the detection capability for small targets. The Head network contains four Detect heads, at least one of which is a high-resolution head used to output defect prediction results at different scales. During model training, MPDIoU is used as the loss function, and its calculation formula is... ,in , , , The total loss function is Using this loss function, the model can more accurately regress the positional deviation between the predicted bounding box and the true bounding box. The model ultimately outputs the defect type, confidence level, bounding box position, and size estimate for each defect, thereby achieving accurate identification of defects such as cracks, cavitation, and spalling on the turbine surface.

[0040] This specific implementation enhances the model's ability to extract key defect features by introducing SimAM and GAM attention modules, while optimizing the bounding box regression accuracy using the MPDIoU loss function, thereby significantly improving the accuracy of defect identification and localization.

[0041] Based on the above embodiments, in some embodiments, the continuous optimization of the image recognition model through manual feedback and incremental learning includes: Provides a human feedback interface for receiving verification and annotation of the recognition results; When the number of new samples reaches a preset threshold, incremental training is initiated to update the model parameters while maintaining the original recognition ability. Images whose recognition confidence level is within a preset range or that have been manually verified as incorrect are identified as difficult examples and given higher sampling weights in subsequent training.

[0042] In this embodiment, the specific implementation of continuously optimizing the image recognition model through manual feedback and incremental learning in step S104 is as follows. First, the system constructs a manual feedback interface. This interface is used by technicians to review and annotate the defect recognition results after receiving them from the image recognition model. Technicians use this interface to confirm whether the recognition results are correct. If the recognition results are incorrect, the correct defect type, location, or annotation information is fed back. Simultaneously, if there are defects in the image that the model failed to recognize, technicians can also add defect types and annotate them. These manually reviewed image data and their corresponding annotation information are collected by the system as new sample data and stored in the incremental training sample library. When the number of new samples in the incremental training sample library reaches a preset threshold... At this time, the system automatically triggers the incremental learning process. During the incremental learning process, the system retrieves new sample data from the incremental training sample library. And based on the parameters of the current image recognition model The gradient descent method is used for parameter updates, and the update rule follows the formula. ,in, The learning rate is set to 0.0001. For the model on new sample data The gradient of the loss function. Through this incremental training, the model parameters change from... Updated to This allows the model to adapt to newly emerging defect features without forgetting previously learned knowledge. Furthermore, to further enhance the model's ability to identify difficult samples, the system also introduces a difficult example mining mechanism. Specifically, this mechanism is used to assess the confidence level of the image recognition model's output. Images with recognition results falling within the range [0.4, 0.6], or images confirmed as incorrectly identified through manual review, are classified as difficult examples by the system and added to a dedicated difficult example library. During subsequent incremental training, when the system samples from this library, it assigns these difficult examples higher sampling weights. Specifically, the sampling weights of the difficult examples... Set as normal sample sampling weight 3 times, that is This weighted sampling strategy allows the model to focus more on samples that are difficult to identify correctly during training, thus effectively improving its performance in recognizing complex, ambiguous, or rare defects. Through the synergistic effect of human feedback, incremental learning, and hard example mining, the image recognition model can be continuously optimized in practical applications, with its recognition accuracy and robustness constantly improving.

[0043] This specific implementation provides an automated and continuous online optimization path for image recognition models by introducing a human feedback interface, setting incremental learning trigger thresholds and learning rates, and implementing a difficult example weighted sampling mechanism based on confidence intervals and human review results. This significantly improves the model's adaptability to novel and complex defects and its recognition accuracy, ensuring the long-term effectiveness of the model in real-world operating environments.

[0044] Based on the above embodiments, in some embodiments, the method further includes: The contribution of each feature to the model output is calculated using the SHAP value method. The contribution is used to quantify the degree of influence of the feature on the fault probability prediction. Based on the contribution level, fault symptom patterns are mined to generate an interpretable list of abnormal indicators and fault type inferences.

[0045] In this embodiment, the specific implementation method of using the SHAP value method for fault symptom pattern mining is further defined. Specifically, after completing the training of the deep neural network early warning model, the system takes the trained model and the feature vectors from the normal state database and fault state database used for training as inputs to initiate the SHAP value analysis process. The core of this analysis process lies in calculating the marginal contribution of each feature to the model output, i.e., the SHAP value. For the i-th feature in the feature vector, its SHAP value is... The calculation is based on a combination formula that iterates through all feature subsets: In this formula, The set representing all 129 features. yes The middle does not contain features any subset, This subset The number of features, The total number of features (i.e., 129). Indicates using only subsets When considering features in a model, the model is applied to specific input samples. The predicted value is calculated by adding features under all possible subset combinations. The changes in the predicted values ​​before and after the model are calculated and weighted according to the subset size to obtain the final feature. SHAP value This value quantifies the contribution of the feature to the model's output failure probability. A positive value indicates that the feature value leads to a higher failure probability output by the model, while a negative value has the opposite effect. After calculating the SHAP value of each feature for all samples, the system performs a summary analysis and outputs a feature importance ranking list. This list clearly indicates the top few features that contribute the most to failure early warning, such as the root mean square value of water conductance swing and the amplitude of tailrace pressure pulsation, and records their corresponding SHAP values. Subsequently, the system combines the changing trends of these high-contribution features within different time windows before the failure (such as -48h, -24h, -12h) and establishes trend equations for the feature changes over time by fitting quadratic or exponential functions, thereby uncovering specific failure symptom patterns. For example, "the root mean square value of water conductance swing in the X direction begins to rise continuously 48 hours before the failure, and the fitted equation is..." These discovered patterns are stored in a symptom pattern database, which is used in subsequent online monitoring to provide maintenance personnel with a more interpretable list of abnormal indicators and possible fault type inferences when an alert is triggered.

[0046] By employing the SHAP value method for fault symptom pattern mining, this embodiment quantifies the specific contribution of each feature to the early warning model's decision-making, transforming the originally "black box" deep learning model output into interpretable, physically meaningful feature importance rankings and trend patterns. This not only significantly improves the interpretability of fault diagnosis, enabling operations and maintenance personnel to intuitively understand the core abnormal indicators leading to early warnings and their development processes, but also provides a knowledge foundation for subsequent early warning-identification closed-loop optimization. This allows the model to learn from each confirmed fault case and update the symptom pattern library, thereby continuously improving the accuracy and lead time of early warnings.

[0047] Based on the above embodiments, in some embodiments, after the defect is confirmed by the image recognition model, the system further calculates the risk coefficient and generates corresponding decision suggestions. Specifically, the system first obtains the defect size s output by the image recognition model, which is obtained through bounding box estimation, specifically... Where k is a scaling factor used to convert pixel size to actual physical size. Simultaneously, the system obtains the warning lead time corresponding to this defect from the warning records. The time parameter is the time difference, in hours, between the moment the warning is first triggered and the moment the system shuts down upon final confirmation of the defect. Additionally, the system obtains the recognition confidence level *c* of the image recognition model for the defect, with a value ranging from [0,1]. After obtaining these three parameters, the system calculates the risk coefficient according to a preset formula. Calculate the risk coefficient R, where This is the critical size. , , The weighting coefficients are satisfied. The risk coefficient R comprehensively reflects the ratio of defect size to critical size, the length of the early warning time, and the level of confidence in identification; a higher value indicates a higher risk. Finally, the system generates decision suggestions based on the calculated risk coefficient R: when In such cases, continued monitoring is recommended, with monthly inspections. In such cases, planned maintenance is recommended, with repairs scheduled within 3 months; when In such cases, it is recommended to arrange for a shutdown as soon as possible; when If this occurs, it is recommended to shut down the machine immediately. This decision recommendation, along with information such as defect type, location, and size, will be integrated into the final fusion diagnostic report.

[0048] Through the aforementioned risk coefficient calculation and decision suggestion generation mechanism, this embodiment can quantify and integrate information from three dimensions: early warning time, defect size, and identification confidence level. This provides maintenance personnel with scientific, graded, and actionable handling suggestions, avoiding the one-sidedness of making decisions based on a single indicator. It significantly improves the timeliness and accuracy of fault handling and reduces the risk of equipment damage or unnecessary downtime losses caused by improper decision-making.

[0049] Based on the above embodiments, this embodiment first constructs a multi-dimensional sensing data acquisition system. Multiple types of sensors are deployed at key locations within the unit, including vibration and sway monitoring units, pressure monitoring units, temperature monitoring units, and operating parameter acquisition units. Specifically, a +X direction horizontal vibration sensor, a -Y direction horizontal vibration sensor, and a +X direction vertical vibration sensor are installed at the top cover location; a +X sway sensor and a -Y sway sensor are installed at the water guide bearing location. For pressure monitoring, pressure sensors are installed at the volute inlet, the bottom of the top cover, the tailrace inlet, the tailrace outlet, and the rear of each guide vane. For temperature monitoring, temperature sensors are installed at the water guide bearing, thrust bearing, upper guide bearing, and lower guide bearing. All sensors are synchronously acquired via a high-speed data acquisition instrument, with sampling frequencies set according to signal type: 1kHz for vibration and sway signals, 100Hz for pressure signals, and 1Hz for temperature and operating parameter signals.

[0050] The collected raw data underwent preprocessing and operating condition classification. First, data cleaning was performed, using the 3σ principle to identify and remove outliers; if the absolute value of the difference between a data point and its mean is greater than three times the standard deviation, it is considered an outlier and removed. Then, based on the unit's active power, head, and guide vane opening, the operating data was divided into multiple fine-grained operating conditions. Active power was divided into 10 intervals, head into 20 intervals, and guide vane opening into 20 intervals, for a total of 4000 operating conditions, each assigned a unique identifier. For each operating condition, time-domain features, frequency-domain features, pressure pulsation features, and temperature features were extracted for each monitoring indicator. Time-domain features included mean, variance, standard deviation, peak value, peak-to-peak value, root mean square value, kurtosis, and skewness. Frequency-domain features included spectral centroid, spectral dispersion, dominant frequency amplitude, and harmonic amplitude. Pressure pulsation features included pulsation amplitude, pulsation frequency, and pulsation energy. Temperature characteristics include the rate of temperature rise, temperature difference, and temperature fluctuation amplitude. In addition, the correlation coefficient between water conductance and active power was calculated, with a value ranging from [-1, 1]. When its absolute value is greater than or equal to 0.6, a significant linear positive correlation between water conductance and active power is considered, which is an important indicator of uneven flow field in the turbine blades. All extracted features were combined into a 129-dimensional feature vector.

[0051] Establish normal state and fault state databases. Data from the initial commissioning phase or confirmed fault-free operating periods are selected as the normal state data source. Features are extracted from each monitoring indicator under each operating condition to construct a normal state feature library, and the mean and standard deviation vectors of the feature vectors under each operating condition are calculated. Operational data from periods when the unit experienced faults such as runner cracking, cavitation, and spalling are collected and annotated by professionals, including fault type, severity, time of occurrence, and location. Data from different time windows prior to the fault are extracted to construct a fault state feature library.

[0052] A deep neural network early warning model was constructed. Feature vectors under various operating conditions were randomly sampled from the normal state database as positive samples, and feature vectors from the fault state database for the 48 hours, 24 hours, and 12 hours prior to the fault were extracted as negative samples to construct the training set. The deep neural network early warning model consists of an input layer, four hidden layers, and an output layer. The input layer has a dimension of 129. Hidden layer 1 contains 512 neurons with ReLU activation and a Dropout rate of 0.2. Hidden layer 2 contains 256 neurons with ReLU activation and a Dropout rate of 0.2. Hidden layer 3 contains 128 neurons with ReLU activation and a Dropout rate of 0.2. Hidden layer 4 contains 64 neurons with ReLU activation and a Dropout rate of 0.2. The output layer contains one neuron with Sigmoid activation, outputting the fault probability. The model was trained using a binary cross-entropy loss function, with Adam as the optimizer, a learning rate of 0.001, and a batch size of 64. After training, the SHAP value method is used to calculate the contribution of each feature to the model output and to uncover the patterns of fault symptoms.

[0053] During the online monitoring phase, data from various sensors are collected in real time, and features are extracted using the aforementioned method to form a real-time feature vector. The current operating condition is determined based on the current active power, head, and guide vane opening. The normal feature distribution under this condition is retrieved from the normal state database, and the anomaly score and overall anomaly score of each feature are calculated. The real-time feature vector is input into a trained deep learning model to calculate the fault probability. Multi-level early warnings are set based on the fault probability and overall anomaly score: when the fault probability is greater than or equal to 0.3 and less than 0.5, or the overall anomaly score is greater than 2, a level of concern warning is triggered, the system records the information, and a yellow alert is issued; when the fault probability is greater than or equal to 0.5 and less than 0.7, or the overall anomaly score is greater than 3, and the correlation coefficient between the guide vane swing and active power is greater than or equal to 0.6, a level of suspicion warning is triggered, an audible and visual alarm is issued, and an orange alert is issued; when the fault probability is greater than or equal to 0.7, or the overall anomaly score is greater than 4, a level of danger warning is triggered, an emergency alarm is issued, and a red alert is issued.

[0054] When a suspicious or dangerous warning is triggered, images of the runner surface are collected by an underwater drone after the unit is shut down. Specifically, after the unit is shut down, the rear valve of the dual-valve isolation chamber is opened, the underwater drone is placed into the isolation chamber and connected by a wired cable, the rear valve is closed, and the front valve is opened to equalize the pressure in the isolation chamber and the tailrace pipe. The drone then exits the isolation chamber and enters the tailrace pipe. The operator remotely controls the drone to photograph various parts of the runner, collecting an image set. The collected images are preprocessed, including median filtering for noise reduction, histogram equalization enhancement, cropping to remove irrelevant background, and size normalization. An improved YOLOv8 model is used for defect identification. This model includes a backbone network, a neck network, and a head network. The backbone network contains 5 Conv modules, 4 C2f modules, and 1 SPPCSPC+SimAM module. The neck network adopts an FPN+PAN structure and introduces 4 GAM attention modules. The head network contains 4 Detect heads. The model loss function is MPDIoU, and its calculation formula is as follows: ,in , , , The model outputs the defect type, confidence level, bounding box location, and size estimate for each defect.

[0055] A self-learning mechanism is established to continuously optimize the image recognition model through manual feedback and incremental learning. Technicians review the recognition results and provide feedback. Incremental training is initiated when the number of new samples reaches a threshold of 100, with a learning rate of 0.0001. Images with confidence levels between 0.4 and 0.6, or those with errors in manual review, are added to a difficult example library, with a sampling weight three times that of normal samples. Early warning data confirming defects are used as new negative samples to retrain the early warning model, forming an early warning-recognition closed-loop optimization. The risk coefficient is calculated using the following formula: ,in , For fault dimensions, To provide advance warning time, To identify confidence levels, , , The risk coefficient is used as the weighting factor. Decision recommendations are generated based on the risk coefficient: When the risk coefficient is less than 0.3, continue monitoring and conduct monthly inspections; when the risk coefficient is greater than or equal to 0.3 and less than 0.6, conduct planned maintenance; when the risk coefficient is greater than or equal to 0.6 and less than 0.8, arrange for immediate shutdown; when the risk coefficient is greater than or equal to 0.8, shut down immediately.

[0056] This embodiment achieves earlier and more accurate turbine runner fault warning through multi-source data fusion and deep learning, and continuously optimizes through a self-learning mechanism, significantly improving the safety and economy of turbine operation.

[0057] Please refer to Figure 2 , Figure 2 This is a structural block diagram of a turbine runner fault early warning and self-learning identification device provided in an embodiment of the present invention; the specific device may include: The data acquisition and feature extraction module is used to acquire multi-source sensor data during the operation of the water turbine, preprocess the multi-source sensor data and classify the operating conditions to extract multi-dimensional features. The multi-dimensional features include time domain features, frequency domain features and physical correlation features. The physical correlation features are used to characterize the correlation between water conductance and active power. The early warning module is used to construct a deep neural network early warning model based on normal state data and pre-fault state data, input the multi-dimensional features into the deep neural network early warning model to output the fault probability, and trigger multi-level early warnings in combination with the physical correlation features. The image recognition module is used to collect images of the turbine surface after the unit is shut down when the multi-level early warning of the preset level is triggered, and to use the image recognition model to identify defects in the turbine surface images in order to output the defect type and defect location. The closed-loop optimization module is used to feed back the early warning data of confirmed defects as new negative samples to the deep neural network early warning model for retraining, and to continuously optimize the image recognition model through manual feedback and incremental learning to form a closed-loop optimization.

[0058] The turbine runner fault early warning and self-learning identification device of this embodiment is used to implement the aforementioned turbine runner fault early warning and self-learning identification method. Therefore, the specific implementation of the turbine runner fault early warning and self-learning identification device can be found in the embodiment section of the turbine runner fault early warning and self-learning identification method above. For example, the data acquisition and feature extraction module 100, the early warning module 200, the image recognition module 300, and the closed-loop optimization module 400 are respectively used to implement steps S101, S102, S103, and S104 in the above-mentioned turbine runner fault early warning and self-learning identification method. Therefore, its specific implementation can be referred to the description of the corresponding embodiments, and will not be repeated here.

[0059] A specific embodiment of the present invention also provides a turbine runner fault early warning and self-learning identification device, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of the above-mentioned turbine runner fault early warning and self-learning identification method.

[0060] A specific embodiment of the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described method for early warning and self-learning identification of turbine runner faults.

[0061] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0062] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0063] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0064] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0065] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for early warning and self-learning identification of turbine runner faults, characterized in that, Includes the following steps: Multi-source sensor data is collected during the operation of the water turbine. The multi-source sensor data is preprocessed and divided into operating conditions to extract multi-dimensional features. The multi-dimensional features include time-domain features, frequency-domain features, and physical correlation features. The physical correlation features are used to characterize the correlation between water conduction swing and active power. A deep neural network early warning model is constructed based on normal state data and pre-fault state data. The multi-dimensional features are input into the deep neural network early warning model to output the fault probability, and the physical correlation features are combined to trigger multi-level early warning. When the multi-level early warning at the preset level is triggered, the surface image of the turbine runner is collected after the unit is shut down. The image recognition model is used to identify defects in the surface image of the turbine runner to output the defect type and defect location. The early warning data confirming defects are fed back as new negative samples to the deep neural network early warning model for retraining. The image recognition model is continuously optimized through manual feedback and incremental learning to form a closed-loop optimization.

2. The method according to claim 1, characterized in that, The preprocessing and condition segmentation of the multi-source sensor data to extract multi-dimensional features includes: The raw data is cleaned to remove outliers; The operating data is divided into multiple detailed operating conditions based on the unit's active power, head, and guide vane opening. For each monitoring index under each working condition, the time domain characteristics, frequency domain characteristics, pressure pulsation characteristics, and temperature characteristics are extracted, and the correlation coefficient between the water conductance and active power is calculated as the physical correlation characteristics. All extracted features are combined into a multidimensional feature vector.

3. The method according to claim 1, characterized in that, The deep neural network early warning model constructed based on normal state data and pre-fault state data includes: The feature vectors under each working condition are extracted from the normal state database as positive samples, and the feature vectors within a predetermined time window before the fault are extracted from the fault state database as negative samples to construct a training set. The deep neural network early warning model is trained using a binary cross-entropy loss function. The deep neural network early warning model includes an input layer, multiple hidden layers, and an output layer connected in sequence, wherein the output layer uses the Sigmoid activation function.

4. The method according to claim 1, characterized in that, The method of triggering multi-level early warnings by combining the physical correlation features includes: Obtain the normal feature distribution based on the current operating conditions to calculate the comprehensive anomaly degree of the current real-time feature vector; The real-time feature vector is input into the deep neural network early warning model to obtain the fault probability; Based on the numerical range of the fault probability and the comprehensive anomaly degree, and combined with the threshold judgment of the physical correlation characteristics, different severity levels of warning are triggered.

5. The method according to claim 1, characterized in that, The defect identification of the wheel surface image using an image recognition model includes: The acquired images of the wheel surface are input into an improved target detection model, which includes a backbone network, a neck network, and a head network. The backbone network includes an attention module for enhancing the focus on key features, the neck network adopts a multi-scale feature fusion structure and introduces an attention module, and the head network includes multiple detection heads. The model is optimized using a metric function based on the vertex distance and / or bounding box size between the predicted box and the ground truth box as the loss function.

6. The method according to claim 1, characterized in that, The process of continuously optimizing the image recognition model through human feedback and incremental learning includes: Provides a human feedback interface for receiving verification and annotation of the recognition results; When the number of new samples reaches a preset threshold, incremental training is initiated to update the model parameters while maintaining the original recognition ability. Images whose recognition confidence level is within a preset range or that have been manually verified as incorrect are identified as difficult examples and given higher sampling weights in subsequent training.

7. The method according to claim 1, characterized in that, Also includes: The contribution of each feature to the model output is calculated using the SHAP value method. The contribution is used to quantify the degree of influence of the feature on the fault probability prediction. Based on the contribution level, fault symptom patterns are mined to generate an interpretable list of abnormal indicators and fault type inferences.

8. A turbine runner fault early warning and self-learning identification device, characterized in that, include: The data acquisition and feature extraction module is used to acquire multi-source sensor data during the operation of the water turbine, preprocess the multi-source sensor data and classify the operating conditions to extract multi-dimensional features. The multi-dimensional features include time domain features, frequency domain features and physical correlation features. The physical correlation features are used to characterize the correlation between water conductance and active power. The early warning module is used to construct a deep neural network early warning model based on normal state data and pre-fault state data, input the multi-dimensional features into the deep neural network early warning model to output the fault probability, and trigger multi-level early warnings in combination with the physical correlation features. The image recognition module is used to collect images of the turbine surface after the unit is shut down when the multi-level early warning of the preset level is triggered, and to use the image recognition model to identify defects in the turbine surface images in order to output the defect type and defect location. The closed-loop optimization module is used to feed back the early warning data of confirmed defects as new negative samples to the deep neural network early warning model for retraining, and to continuously optimize the image recognition model through manual feedback and incremental learning to form a closed-loop optimization.

9. A turbine runner fault early warning and self-learning identification device, characterized in that, include: Memory, used to store computer programs; A processor is configured to execute the computer program to implement the steps of the turbine runner fault early warning and self-learning identification method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the turbine runner fault early warning and self-learning identification method as described in any one of claims 1 to 7.

Images