Rotating wheel defect detection and evaluation method and related device

Abstract

The invention belongs to a defect detection method, and provides a rotating wheel defect detection and evaluation method and a related device for solving the technical problems of insufficient accuracy and poor universality of a quality detection method of a rotating wheel in a water-turbine generator set at present, and the rotating wheel defect detection and evaluation method comprises the following steps: inputting external multi-dimensional information and internal structure multi-dimensional information of the rotating wheel into a feature extraction network; the feature extraction network fuses end-to-end super-division and feature extraction, and then inputs the defect feature vector into a self-adaptive defect classification model comprising a multi-task learning framework to obtain a defect type of the runner, a geometric parameter quantification result and a risk level evaluation result; and finally, combining the running state information in the external multi-dimensional information and the internal structure multi-dimensional information of the runner and the detection result, synchronizing the defect condition in the detection result into the digital twin of the runner, and calculating to obtain a runner defect comprehensive evaluation result. The rotating wheel defect detection speed and accuracy are improved, and potential faults can be prevented in advance.

Description

Technical Field

[0001] This application pertains to a defect detection method, specifically a method and apparatus for detecting and evaluating defects in a rotary wheel. Background Technology

[0002] With the continuous improvement of industrial automation, the quality of key components in hydro-generator units, such as the runner, is crucial to the safety and reliability of the power generation system. Traditional runner defect detection relies on manual visual inspection or simple machine vision systems, which are overly dependent on the experience and skill level of operators. This results in low repeatability, consistency, and reliability of inspection results; slow inspection speed, making it difficult to meet the needs of efficient production in modern manufacturing; limited ability to identify minute or hidden defects (such as internal cracks); and a lack of quantitative assessment of the impact of defects, failing to provide strong support for subsequent maintenance decisions.

[0003] While some automated inspection equipment has adopted image processing technology to some extent, it still has the following limitations: most solutions fail to fully utilize the latest deep learning achievements and perform poorly in feature extraction and pattern recognition; a complete closed-loop process from data acquisition to defect assessment has not been formed, which limits the overall performance of the system; it is difficult to adapt to different models and materials of rollers, and it is impossible to combine the equipment operating status in real time to classify the severity of defects and predict lifespan, resulting in a high rate of missed detections and poor versatility. Summary of the Invention

[0004] This application addresses the technical problems of insufficient accuracy and poor versatility in current methods for quality inspection of turbine runners in hydro-generator sets by providing a method and related apparatus for detecting and evaluating turbine runner defects.

[0005] To achieve the above objectives, this application adopts the following technical solution: Firstly, this application proposes a method for detecting and evaluating defects in turbine rotors, including: Obtain multi-dimensional information about the external environment and internal structure of the rotor; The external multidimensional information and internal structural multidimensional information of the wheel are input into the feature extraction network to obtain the defect feature vector; the feature extraction network integrates end-to-end super-resolution and feature extraction. The defect feature vector is input into an adaptive defect classification model that includes a multi-task learning framework to obtain the defect type, geometric parameter quantification results, and risk level assessment results of the wheel, which together constitute the detection results. By combining the operational status information from the external multidimensional information and the internal multidimensional information of the rotor, as well as the detection results, the defect information in the detection results is synchronized to the digital twin of the rotor, and the comprehensive evaluation result of the rotor defects is calculated.

[0006] Furthermore, the external multidimensional information and internal structural multidimensional information of the rotor include: surface image of the rotor, surface three-dimensional morphology, operating status, and internal multi-angle structural information.

[0007] Furthermore, after acquiring the multi-dimensional information about the external environment and the multi-dimensional information about the internal structure of the rotor, the process further includes: The external multidimensional information and internal structural multidimensional information of the rotor are subjected to noise reduction, edge enhancement and data alignment.

[0008] Furthermore, the feature extraction network includes a single-modal feature enhancement module and a multi-modal feature fusion module; The single-modal feature enhancement module is used to enhance the resolution of visual information in the multi-dimensional information of the external and internal structures of the wheel to obtain visual enhancement features. The multimodal feature fusion module is used to dynamically allocate the weights of each modal information in the external multidimensional information and internal structural multidimensional information of the wheel through an attention mechanism, and integrate the modal information into a fused feature vector.

[0009] Furthermore, the calculation method in the adaptive defect classification model includes: The defect feature vector is input into the student model after knowledge distillation to obtain preliminary feature signals; Based on the preliminary feature signals, through dynamic multi-task branching, the confidence level of the defect category, the geometric parameter quantification result, and the risk level assessment result are obtained, which together constitute the detection result; If the detection results include defect types that have not appeared before, manual verification is performed, and multiple samples of related defect types are obtained to fine-tune the student model.

[0010] Furthermore, the method for calculating the comprehensive evaluation result of the impeller defects includes: Mechanical impact assessment, lifespan prediction, and performance impact assessment are performed in the digital twin, respectively. Based on the results of mechanical impact assessment, life prediction, and performance impact assessment, a comprehensive assessment result of the impeller defects is obtained through integrated calculation.

[0011] Furthermore, after obtaining the comprehensive evaluation result of the impeller defects through comprehensive calculation, the process also includes: The impeller was maintained based on the comprehensive assessment results of its defects, and the maintenance results were recorded. Update the digital twin based on the maintenance results.

[0012] Secondly, this application proposes a wheel defect detection and evaluation system, comprising: The data module is used to acquire multi-dimensional information about the external environment and internal structure of the rotor. The feature extraction module is used to input multi-dimensional information about the external environment and the internal structure of the wheel into the feature extraction network to obtain defect feature vectors; the feature extraction network integrates end-to-end super-resolution and feature extraction. The detection module is used to input the defect feature vector into an adaptive defect classification model that includes a multi-task learning framework, and obtain the defect type, geometric parameter quantification results, and risk level assessment results of the wheel, which together constitute the detection results. The evaluation module combines the operational status information from the external multidimensional information and internal structural multidimensional information of the runner, as well as the detection results, to synchronize the defect information in the detection results to the digital twin of the runner, and calculate the comprehensive evaluation result of the runner defects.

[0013] Thirdly, this application proposes an electronic device, including: a memory and one or more processors; the memory is coupled to the processors; wherein the memory stores computer program code, the computer program code including computer instructions, and when the computer instructions are executed by the processor, the electronic device performs the steps of the above-described rotary wheel defect detection and evaluation method.

[0014] Fourthly, this application proposes a computer-readable storage medium storing a computer program, which is executed by a processor using the steps of the above-described wheel defect detection and evaluation method.

[0015] Compared with the prior art, this application has the following beneficial effects: This application proposes a method for detecting and evaluating defects in a turbine rotor. The method inputs multi-dimensional external and internal structural information of the rotor into a feature extraction network to obtain defect feature vectors. The feature extraction network integrates end-to-end super-resolution and feature extraction. The defect feature vectors are then input into an adaptive defect classification model incorporating a multi-task learning framework to obtain the rotor's defect type, geometric parameter quantification results, and risk level assessment results. Finally, combining the operational status information from the rotor's external and internal multi-dimensional information with the detection results, the defect situation in the detection results is synchronized to the rotor's digital twin, and a comprehensive defect assessment result is calculated. This application, combining a feature extraction network and an adaptive defect classification model, improves the speed and accuracy of rotor defect detection, reduces labor costs, and, through in-depth defect analysis, helps prevent potential failures in advance, ensuring the safety of equipment operation. Therefore, by leveraging deep learning, rapid and accurate defect detection is achieved, significantly improving work efficiency. Furthermore, it can detect not only surface defects but also internal damage and is applicable to rotors of various specifications and materials. In addition to providing defect detection results, it can also conduct in-depth risk analysis, which helps to provide preventive maintenance guidance, reduce operating costs, and extend the service life of the rotor.

[0016] This application also proposes a wheel defect detection and evaluation system, an electronic device, and a computer-readable storage medium, which possess all the advantages of the aforementioned wheel defect detection and evaluation methods. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the first process of the turbine defect detection and evaluation method of this application; Figure 2 This is a schematic diagram of the second process for the defect detection and evaluation method of the turbine in this application; Figure 3 This is a schematic diagram of the turbine defect detection and evaluation system of this application. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0020] Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0021] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0022] In the description of the embodiments of this application, it should be noted that if terms such as "upper," "lower," "horizontal," or "inner" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product of the invention is in use, they are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation on this application. In addition, terms such as "first" and "second" are only used to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0023] Furthermore, the use of the term "horizontal" does not imply that the component must be absolutely horizontal, but rather that it can be slightly tilted. For example, "horizontal" simply means that its direction is more horizontal than "vertical," and does not mean that the structure must be completely horizontal, but can be slightly tilted.

[0024] In the description of the embodiments of this application, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0025] In core industrial sectors such as energy, power, and water conservancy projects, hydro-generator units are key equipment for clean energy production, and their operational stability directly determines the security and continuity of power supply. The turbine runner, as the core working component of the hydro-generator unit, bears the important responsibility of converting water flow energy into mechanical energy. It operates under complex conditions of high pressure, high speed, water flow impact, and media corrosion, making it prone to various defects such as cracks, wear, corrosion, and internal inclusions. If these defects are not detected and addressed in a timely manner, they will continue to expand over time, potentially leading to decreased runner strength, reduced operating efficiency, and in severe cases, even runner fracture and other major safety accidents, causing huge economic losses and prolonged power supply interruptions.

[0026] In current practices for defect detection on rotary turbines, traditional inspection methods still suffer from numerous unresolved pain points. Manual visual inspection is the most common and basic method, where inspectors examine the turbine surface with the naked eye or with the aid of simple magnifying glasses or endoscopes. This method relies entirely on the operator's experience and judgment, resulting in high labor intensity, low efficiency, and the fact that the skill level, sense of responsibility, and even visual state of different personnel directly affect the inspection results. This leads to extremely poor repeatability and consistency in defect identification, and the same defect may be missed, misjudged, or have inaccurate defect level determinations in different inspection scenarios. For minute defects on the turbine surface (such as microcracks less than 0.1 mm wide) or defects hidden internally (such as internal cracks, pores, inclusions, etc.), manual vision cannot effectively identify them. These hidden defects often become significant safety hazards in subsequent operation. Meanwhile, traditional inspection methods can only qualitatively determine the existence of defects, failing to accurately measure geometric parameters such as the length, depth, and area of ​​the defects. Furthermore, they cannot quantitatively analyze the impact of defects on the turbine's mechanical performance, operating efficiency, and remaining lifespan. This makes it difficult for maintenance personnel to make scientific and reasonable maintenance decisions, forcing them to adopt passive strategies such as over-maintenance or delayed maintenance, increasing maintenance costs or allowing risks to accumulate. In addition, traditional inspection requires prolonged unit shutdowns, severely impacting power generation efficiency, which is seriously incompatible with the requirements of modern industry for continuous and efficient production.

[0027] To address the shortcomings of traditional manual inspection, some automated inspection attempts have gradually emerged in the industry. Some solutions incorporate machine vision systems, using industrial cameras to capture images of the turbine surface and then analyzing them using simple image processing techniques, achieving preliminary automated defect identification. Other solutions combine non-destructive testing technologies such as ultrasonic testing and radiographic testing to detect internal defects in the turbine. Ultrasonic testing utilizes the propagation characteristics of ultrasound in a medium, determining the location and size of defects by receiving reflected signals. Radiographic testing uses X-rays or gamma rays to penetrate the turbine, forming defect images based on differences in ray attenuation. These technologies reduce reliance on human experience to some extent, improving inspection efficiency, especially in the detection of internal defects, showing significant progress compared to manual vision. Furthermore, some automated equipment also possesses basic defect parameter measurement functions, enabling the preliminary acquisition of defect dimensional information, providing some reference for maintenance decisions.

[0028] However, existing automated inspection solutions still have many limitations and fail to fundamentally solve the core problem of wheel defect detection. First, most solutions use traditional image processing algorithms, failing to fully utilize advanced artificial intelligence technologies such as deep learning. Their feature extraction capabilities are weak, resulting in low accuracy in identifying minute defects and irregularly shaped defects in complex backgrounds, leading to high false negative and false positive rates. Second, existing technologies are mostly single-stage inspection tools, lacking a complete closed-loop process from multimodal data acquisition, intelligent defect identification, classification and quantification to risk assessment, lifespan prediction, and maintenance recommendations. Data from each stage is isolated, preventing collaborative optimization and limiting the overall inspection performance of the system. Third, existing solutions struggle to adapt to the inspection needs of wheels made of different materials and with different structural types, lacking dynamic adaptability to changes in operating conditions. They cannot combine real-time parameters such as wheel rotation speed, vibration, and temperature for defect hazard classification, leading to a disconnect between risk assessment and actual operating conditions. Meanwhile, existing technologies are not precise enough in terms of defect quantification assessment, especially in predicting defect expansion trends and remaining wheel life. They have failed to establish an assessment model that combines physical simulation with data-driven approaches, making it difficult to provide comprehensive and reliable technical support for operation and maintenance decisions. In addition, some solutions involve bulky equipment, complex operation, and high testing costs, lacking versatility and practicality, making it difficult to promote and apply them on a large scale in the industry.

[0029] In summary, existing turbine defect detection technologies have significant shortcomings in terms of detection accuracy, efficiency, intelligence level, and closed-loop capability throughout the entire process. Therefore, developing a highly efficient, accurate, intelligent turbine defect detection, analysis, and evaluation method with full lifecycle assessment capabilities has become an urgent need for industry development.

[0030] Based on the above, this application proposes a method and related apparatus for detecting and evaluating defects in a turbine. The following detailed description of this application is provided in conjunction with the embodiments and accompanying drawings.

[0031] like Figure 1 The diagram shown is a flowchart of one method for detecting and evaluating defects in a rotary wheel according to this application, which may include: S101, acquire multi-dimensional information about the external environment and internal structure of the rotor.

[0032] This application, based on the concept of full-dimensional data coverage, utilizes various sensing technologies and detection devices to collect data from multiple dimensions, including the external surface, three-dimensional morphology, operational dynamics, and internal structure of the rotor. This ensures that no defect-related information is missed, providing a complete and high-fidelity data source for subsequent feature extraction and evaluation. Thus, it achieves full-dimensional coverage of defects on the rotor's external surface, three-dimensional geometric features, dynamic operating status, and internal structural defects. In practical applications, the resolution and accuracy of the collected data meet the detection requirements for minute and internal defects. Furthermore, the data can be processed to maintain spatiotemporal consistency across multiple modalities, avoiding the information bias caused by a single data type.

[0033] S102, input the multi-dimensional information of the external environment and the multi-dimensional information of the internal structure of the wheel into the feature extraction network to obtain the defect feature vector; the feature extraction network integrates end-to-end super-resolution and feature extraction.

[0034] It should be noted that end-to-end super-resolution refers to integrating super-resolution reconstruction and feature extraction into a single end-to-end trained network architecture, which does not require separate super-resolution preprocessing and can directly output enhanced defect feature vectors.

[0035] This application addresses the problem of blurred features and low pixel proportion of minute defects in raw data. It utilizes end-to-end super-resolution technology to improve data resolution and enhance the feature representation of minute defects. Furthermore, by leveraging the hierarchical feature extraction capabilities of deep learning networks, key defect features can be automatically captured from multimodal data, achieving synergistic optimization of super-resolution enhancement and feature extraction, and avoiding feature loss caused by traditional separate processing.

[0036] This application can effectively improve the feature recognition of minute defects, solve the problem that traditional feature extraction networks are not good at capturing minute defects and low-contrast defects. Multimodal feature fusion makes the defect feature vector more comprehensive, providing high-quality input for subsequent classification and quantization. The end-to-end architecture simplifies the processing flow and improves the feature extraction efficiency.

[0037] S103, the defect feature vector is input into an adaptive defect classification model including a multi-task learning framework to obtain the defect type, geometric parameter quantification results and risk level assessment results of the wheel, which together constitute the detection results.

[0038] It's important to note that a multi-task learning framework refers to a model that simultaneously learns multiple related tasks, such as classification, regression, and evaluation. It's a deep learning architecture that improves the overall performance of the model by sharing information between tasks. An adaptive defect classification model, on the other hand, can automatically adjust its parameters based on changes in data distribution or the emergence of new defect types, achieving accurate classification.

[0039] This application, based on task collaboration and adaptive evolution, achieves simultaneous output of defect classification, geometric quantization, and risk assessment through a multi-task learning framework. It utilizes knowledge distillation to improve the accuracy of the lightweight model and combines meta-learning and incremental learning mechanisms to enable the model to quickly adapt to new defect types, avoiding the low efficiency and poor generalization ability of traditional single-task models. In practical applications, as an example, a multi-task architecture with a shared backbone and multiple branches can be adopted. The classification branch can output defect types such as cracks, wear, and corrosion, along with their confidence levels, through fully connected layers. The quantization branch outputs defect geometric parameters through a regression head, and the risk assessment branch integrates geometric parameters with real-time operational status parameters to output the risk level. During model training, a knowledge distillation mechanism can be used, employing a high-precision, large-scale model as the teacher model to optimize the student model. Meta-learning is then used to initialize model parameters, and an elastic weight consolidation algorithm is combined to achieve incremental learning of new defect types.

[0040] S104, combining the operating status information from the multi-dimensional information of the external and internal structures of the rotor, as well as the detection results, synchronizes the defect information in the detection results to the digital twin of the rotor, and calculates the comprehensive evaluation result of the rotor defects.

[0041] This application leverages the virtual-real mapping and simulation prediction capabilities of digital twins to synchronize detection results and real-time operational status information to a virtual mirror. Combined with physical simulation models and data-driven models, such as long short-term memory networks, it achieves a comprehensive assessment of the impact of defects on the mechanical performance, operational efficiency, and remaining lifespan of the turbine runner. Therefore, it enables dynamic and accurate quantitative assessment of defect impact, comprehensively improving the match between severity scores and actual defect hazards. The digital twin can support visualized simulation of defect expansion trends, providing intuitive references for maintenance personnel. The comprehensive assessment results provide a scientific basis for maintenance decisions, avoiding over-maintenance or under-maintenance, reducing maintenance costs, and improving the safety and reliability of turbine runner operation.

[0042] This application first collects comprehensive data on the external surface, 3D morphology, operating status, and internal structure of the rotor using multiple high-precision sensing devices to ensure no defect information is missed. Then, it utilizes a fusion end-to-end super-resolution feature extraction network to enhance the representation of minute defect features and extract multimodal fusion feature vectors. Next, a multi-task adaptive classification model synchronously outputs the defect type, geometric parameters, and risk level, achieving accurate defect identification and quantification. Finally, based on a digital twin, combined with physical simulation and a data-driven model, a comprehensive assessment of defect severity, remaining lifespan, and performance impact is completed, forming a complete closed-loop detection and evaluation system. This represents a leap forward in rotor defect detection, moving from manual reliance to intelligent autonomy, from single-detection to comprehensive evaluation, and from static analysis to dynamic prediction.

[0043] like Figure 2The diagram shown is a second flowchart of the wheel defect detection and evaluation method of this application, which may include: S201, Multimodal Data Acquisition.

[0044] It can utilize integrated high-resolution HD cameras, laser scanners, vibration sensors, and other sensing devices to acquire omnidirectional surface images, three-dimensional topographic data, operating status parameters, and multi-angle images of the internal structure of the rotor. Specifically: Surface image acquisition can utilize a 4K industrial CMOS high-definition camera to ensure clear imaging of minute surface defects. Three-dimensional topography data can be obtained using a laser triangulation scanner to acquire three-dimensional point cloud data. Operating status parameters are acquired through piezoelectric vibration sensors, Hall effect speed sensors, and temperature sensors to obtain triaxial vibration acceleration, rotational speed, and temperature, respectively. Internal structure images can be obtained using an industrial CT scanner, acquiring slice images and a three-dimensional reconstruction model of the internal structure through tomographic scanning. During data acquisition, a GPS timing module can be used to add a unified timestamp to the data from each device to ensure spatiotemporal consistency.

[0045] S202 introduces advanced image denoising and edge enhancement algorithms, and uses adaptive data augmentation technology to build a defect sample library, solving the problem of insufficient defect samples in real-world scenarios.

[0046] Lightweight convolutional neural network models can be designed, embedding attention mechanisms to prioritize the features of defective regions, reduce background interference, and ensure the quality of data input into AI models.

[0047] For surface and CT images, adaptive weighted bilateral filtering combined with db4 wavelet thresholding is used for denoising, effectively removing Gaussian and salt-and-pepper noise while preserving defect edge details. For vibration signals, empirical mode decomposition and wavelet packet transform are combined to eliminate power frequency interference and environmental noise. For edge enhancement, an improved Canny edge detection algorithm is used, which adaptively calculates the threshold based on local image variance instead of a fixed threshold, accurately enhancing defect edges. For 3D point cloud data, normal estimation and curvature calculation highlight the contour features of the defect region. In the data alignment stage, SIFT (Scale-Invariant Feature Transform) feature point matching is used to achieve multi-view image stitching, and ICP (Iterative Closest Point) algorithm is used to complete the registration of the 3D point cloud and CAD model (registration error ≤ ±0.05mm). Frame synchronization alignment of vibration signals and image data is achieved based on timestamps.

[0048] Then, defect samples of different types of runners such as water turbines, steam turbines, and electric motor runners can be collected, covering 6 major categories and 23 subcategories, including cracks, wear, corrosion, internal inclusions, and porosity. Each subcategory contains more than 1,000 original samples and LabelMe annotation data, which can specifically annotate the defect location, type, size, etc., to build a standardized defect sample library.

[0049] Furthermore, if sample imbalance exists, online adaptive data augmentation strategies can be employed to generate new samples from image data through random rotation, flipping, scaling, brightness / contrast adjustment, Gaussian blurring, and the addition of random noise. 3D point cloud data is expanded through random sampling, translation, rotation, and the addition of slight perturbations. Vibration signals are transformed into feature samples of different frequency bands through time-frequency domain transformations such as Fourier transform and wavelet transform, ultimately increasing the sample library size by more than 10 times.

[0050] Then, using MobileNetV3-Small as the base network, depthwise separable convolutions are used instead of traditional convolutions, compressing the number of parameters to 1 / 8 of that of traditional CNNs, thus meeting the real-time processing requirements at the edge. A CBAM (Convolutional Block Attention Module) is embedded, which strengthens the channel weights of defect features through a Squeeze-Excitation mechanism in the channel dimension, and focuses on the defect region through a convolutional attention mechanism in the spatial dimension, effectively suppressing background interference such as wheel surface texture and scratches. The model input is a 3-channel fused image: the original image, the edge-enhanced image, and the grayscale-normalized image. The output is a feature map focusing on the defect features.

[0051] S203 effectively captures subtle defect features through a deep feature extraction network.

[0052] This application designs a convolutional neural network (CNN) or other suitable deep learning architecture specific to the wheel defect detection task, which can automatically and effectively capture subtle defect features from preprocessed images.

[0053] As an example, this application aims to adapt to the detection of minute defects. It selects a convolutional neural network (CNN) as the basic architecture, optimizes the network structure according to the morphological characteristics and data types of wheel defects, and adjusts the convolutional kernel size, network depth and feature map dimensions to ensure the network's ability to perceive minute features and avoid feature redundancy caused by excessive pursuit of network complexity.

[0054] As a preferred approach, a feature pyramid structure is introduced into the network. Features are extracted from shallow to deep layers through different levels of convolutional layers. Shallow features focus on details such as defect edges and textures, while deep features capture the overall shape of the defect and its contextual relationships. Through top-down feature fusion and lateral connections, multi-scale feature information is integrated to ensure that features of both small and relatively large defects can be effectively captured.

[0055] Furthermore, feature enhancement mechanisms can be embedded in key layers of the network to address the issue of weak feature signals from subtle defects. Adaptive weight adjustments can strengthen the feature response of defect areas and suppress invalid signals from background areas. Simultaneously, the network loss function can be optimized by increasing the weight of subtle defect samples during training, guiding the network to prioritize learning the feature patterns of subtle defects and improving the sensitivity of feature capture for such defects.

[0056] In addition, if the input data contains multimodal information such as images and 3D point clouds, the features of each modality are extracted through the corresponding sub-networks during the feature extraction stage. Then, a fusion module is designed to integrate the defect features of different modalities, complementing the advantages of visual and geometric features, and further improving the recognition of subtle defect features.

[0057] S204 trains a classifier to distinguish different types of defects through transfer learning and online update mechanisms.

[0058] This application introduces a multi-task learning framework to address different types of defects, such as cracks, wear, and corrosion. It simultaneously outputs defect type classification, geometric parameter quantification, and risk level assessment, enabling the classifier to continuously optimize its parameters, adapt to newly emerging defect types, and provide probability scores.

[0059] In practical applications, mature deep learning models pre-trained on industrial defect detection datasets or general image datasets are selected as the base model. The parameters of the underlying general feature extraction layer are frozen, and the top-level network structure and parameters are fine-tuned only for specific tasks such as classification, quantification, and evaluation of wheel defects. Combined with the existing wheel defect sample library, the model is trained through a small number of iterations to quickly adapt to the feature patterns of wheel defects, shortening the training cycle and improving initial performance.

[0060] Then, a network structure with a shared backbone and multiple task branches is constructed based on the basic model. The shared backbone network is responsible for extracting general defect features. On this basis, three parallel task branches are built: a classification branch focuses on different defect types such as cracks, wear, and corrosion; a quantization branch outputs geometric parameters such as the length, depth, and area of ​​defects through regression calculation; and a risk assessment branch determines the risk level of defects by combining defect type, geometric parameters, and real-time operating conditions. Finally, a unified loss function is designed to balance the training priorities of each task, ensuring synchronous optimization of multiple tasks and consistent results.

[0061] In some embodiments of this application, a model parameter update trigger mechanism can be established to automatically initiate the update process when a new defect sample is detected or a certain amount of new working condition data is accumulated. An incremental learning strategy is adopted to fine-tune relevant model parameters using only new data, while protecting existing defect identification knowledge, thus preventing the model from forgetting old knowledge. The updated model parameters are synchronized to each edge detection device via the cloud, achieving unified evolution of the global model.

[0062] In practical applications, probabilistic calculation mechanisms can be introduced into the output layer of each task branch. These mechanisms, using methods such as the Softmax activation function or Bayesian inference, can output the probability distribution of defect types and risk levels, as well as the error range of the geometric parameter quantification results. Furthermore, based on historical inspection data and manual review results, the probability scoring threshold is periodically calibrated to ensure that the reliability of the results matches the actual inspection accuracy.

[0063] S205, Comprehensive Assessment.

[0064] A cloud-edge system is built, where real-time defect detection is performed at the edge, and incremental model training and lifespan prediction are conducted in the cloud. The severity of each defect is calculated based on the classification results. Combining physical simulation and statistical models, each detected defect is comprehensively evaluated, and a detailed inspection report is generated, including but not limited to the defect location, size, shape, type, and its prediction of the impact on the overall performance and lifespan of the impeller.

[0065] As an example, assuming a 0.08mm microcrack is detected on-site, the specific method for comprehensive evaluation could be: The input data is synchronized to the constructed digital twin mirror. Inputs include "microcrack, 0.08×0.06mm, level 4 risk," real-time operating parameters "rotation speed 3000rpm, vibration 5g, temperature 120℃," and the initial CAD model of the rotor and stainless steel material parameters. The three-dimensional coordinates of the microcrack are precisely marked on the 1:1 virtual model. Vibration and temperature data are synchronized to simulate the real-time stress distribution around the defect, resulting in a virtual model completely synchronized with the real rotor, where the location, shape, and surrounding working conditions of the defect are clearly visible.

[0066] Then, accurate diagnosis is performed based on the virtual model: Mechanical impact assessment: The stress concentration factor was calculated to be 3.2 and the fracture risk factor to be 0.75 using the XFEM algorithm. Lifetime prediction: Due to large fluctuations in operating conditions (vibration 5g), dynamic weight allocation (data model 0.7, physical model 0.3), combined with historical LSTM data and the Paris formula, the remaining lifetime is predicted to be 180 days (S_life=80 points). Performance impact assessment: Through virtual model simulation, it was calculated that microcracks lead to a 2.3% decrease in runner efficiency and a 5.1% decrease in stiffness (S_perf=85 points).

[0067] Calculate the overall score S: S = 0.4 × 75 + 0.3 × 80 + 0.2 × 85 + 0.1 × 80 = 79 points The overall assessment is at a medium-to-high risk level.

[0068] Therefore, comprehensive evaluation results can be output to visualization tools: Overall score: 79 (medium to high risk); remaining life: 180 days; efficiency loss: 2.3%; stiffness decrease: 5.1%; stress concentration factor: 3.2; fracture risk: 0.75.

[0069] After maintenance based on the above comprehensive evaluation results, feedback data can be collected, such as the complete elimination of cracks and restoration of efficiency to 99% of the original level after actual repair. This feedback data can be used to fine-tune the material parameters of the virtual model and the weights of the coupling model within a preset timeframe. When similar microcracks are encountered again in the future, the evaluation accuracy can be effectively improved, and the remaining life prediction error will be significantly reduced.

[0070] S206, Generate a visual report.

[0071] The feature pyramid network can separate the morphological and physical features of defects, automatically adjust the defect judgment threshold, reduce the false alarm rate, develop a user-friendly interface, intuitively display the detection results, and provide detailed maintenance suggestions.

[0072] The defect feature data from the comprehensive evaluation phase can be input into a feature pyramid network. Through multi-scale feature hierarchical processing, the morphological and physical features of defects can be separated, clarifying the distribution patterns and correlations between the two types of features. Based on the feature distribution characteristics, a threshold adjustment model is constructed. Combining the defect signal strength, background interference level, and historical false alarm / missed alarm records in real-time detection data, the defect judgment threshold is automatically optimized, ensuring that the threshold always adapts to the current detection scenario and achieving dynamic control of the false alarm rate.

[0073] As an example, when displaying information through a visual interface, the interface architecture can be designed according to the usage requirements of industrial operation and maintenance scenarios, dividing it into three core modules: basic information display area, defect visualization area, and evaluation result area.

[0074] Specifically: the basic information display area clearly presents key background data such as the rotor model, inspection time, and operating conditions; the defect visualization area uses a 3D model overlay annotation method to intuitively display the location, shape, and size of defects, supporting interactive operations such as rotation, scaling, and sectioning, while different colors and icons are used to distinguish defect types and risk levels; the evaluation results area uses a combination of charts and text to display core evaluation conclusions such as defect severity scores, performance impact quantification data, and remaining life prediction, ensuring clear and efficient information transmission.

[0075] In practical applications, maintenance decision-making logic can be constructed based on assessment results such as defect type, severity, performance impact, and remaining lifespan, combined with industry operation and maintenance standards, turbine design parameters, and historical maintenance cases. Personalized maintenance recommendations, including maintenance priorities, suggested maintenance windows, recommended repair processes, and subsequent monitoring plans, can be generated for different risk levels and defect types.

[0076] The impeller defect detection and evaluation method proposed in this application not only improves the speed and accuracy of impeller defect detection and reduces labor costs, but also helps prevent potential failures through in-depth defect analysis, ensuring the safety of equipment operation. Leveraging the powerful pattern recognition capabilities of deep learning, it achieves rapid and accurate defect detection, significantly improving work efficiency. Furthermore, it can detect not only surface defects but also internal damage, and is applicable to impellers of various sizes and materials. It also possesses self-learning capabilities, continuously improving its detection level over time, reducing reliance on manual intervention. Therefore, it provides more than just simple detection results; it also includes in-depth risk analysis and preventative maintenance guidance, helping to reduce operating costs and extend equipment lifespan.

[0077] like Figure 3 The diagram shown is a schematic of one possible system for detecting and evaluating defects in a rotary wheel according to this application, which may include: The data module is used to acquire multi-dimensional information about the external environment and internal structure of the rotor. The feature extraction module is used to input multi-dimensional information about the external environment and the internal structure of the wheel into the feature extraction network to obtain defect feature vectors; the feature extraction network integrates end-to-end super-resolution and feature extraction. The detection module is used to input the defect feature vector into an adaptive defect classification model that includes a multi-task learning framework, and obtain the defect type, geometric parameter quantification results, and risk level assessment results of the wheel, which together constitute the detection results. The evaluation module combines the operational status information from the external multidimensional information and internal structural multidimensional information of the runner, as well as the detection results, to synchronize the defect information in the detection results to the digital twin of the runner, and calculate the comprehensive evaluation result of the runner defects.

[0078] It should be noted that, in the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of each block is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple blocks may be combined or integrated into another device, or some features may be ignored or not executed. The modules described as separate components may or may not be physically separated. The components shown as modules may be one or more physical units, that is, they may be located in one place or distributed in multiple different places. Some or all of the modules can be selected to achieve the purpose of the solution in this embodiment according to actual needs.

[0079] Furthermore, in the various embodiments of the present invention, the modules can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit. The integrated unit described above can be implemented in hardware or as a software functional unit.

[0080] This application also provides an electronic device, which may include one or more processors, memory and communication interfaces.

[0081] The memory, communication interface, and processor are coupled together. For example, the memory, communication interface, and processor can be coupled together via a bus.

[0082] The communication interface is used for data transmission with other devices. The memory stores computer program code. This computer program code includes computer instructions, which, when executed by the processor, cause the electronic device to perform the steps of the aforementioned wheel defect detection and evaluation method.

[0083] The processor can be a processor or controller, such as a Central Processing Unit (CPU), a general-purpose processor, a Digital Signal Processor (DSP), an Application-Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with this disclosure. The processor can also be a combination that implements computing functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc. The processor can be used to support an electronic device in performing the method steps provided in the above embodiments.

[0084] The bus can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. These buses can be categorized as address buses, data buses, control buses, etc.

[0085] This application provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the steps of the above-described wheel defect detection and evaluation method.

[0086] The computer-readable storage media involved in this application include random access memory (RAM), memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disks, removable disks, CD-ROMs, or any other form of storage media known in the art.

[0087] The above are merely preferred embodiments of this application and are not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for detecting and evaluating defects in a rotary wheel, characterized in that, include: Obtain multi-dimensional information about the external environment and internal structure of the rotor; The external multidimensional information and internal structural multidimensional information of the wheel are input into the feature extraction network to obtain the defect feature vector; The feature extraction network integrates end-to-end super-resolution and feature extraction; The defect feature vector is input into an adaptive defect classification model that includes a multi-task learning framework to obtain the defect type, geometric parameter quantification results, and risk level assessment results of the wheel, which together constitute the detection results. By combining the operational status information from the external multidimensional information and the internal multidimensional information of the rotor, as well as the detection results, the defect information in the detection results is synchronized to the digital twin of the rotor, and the comprehensive evaluation result of the rotor defects is calculated.

2. The method for detecting and evaluating defects in a rotary wheel according to claim 1, characterized in that, The external multidimensional information and internal structural multidimensional information of the rotor include: surface image of the rotor, surface three-dimensional morphology, operating status, and internal multi-angle structural information.

3. The method for detecting and evaluating defects in a rotary wheel according to claim 1, characterized in that, After acquiring the multi-dimensional information about the external environment and the multi-dimensional information about the internal structure of the rotor, the process further includes: The external multidimensional information and internal structural multidimensional information of the rotor are subjected to noise reduction, edge enhancement and data alignment.

4. The method for detecting and evaluating defects in a rotary wheel according to claim 1, characterized in that, The feature extraction network includes a single-modal feature enhancement module and a multi-modal feature fusion module; The single-modal feature enhancement module is used to enhance the resolution of visual information in the multi-dimensional information of the external and internal structures of the wheel to obtain visual enhancement features. The multimodal feature fusion module is used to dynamically allocate the weights of each modal information in the external multidimensional information and internal structural multidimensional information of the wheel through an attention mechanism, and integrate the modal information into a fused feature vector.

5. The method for detecting and evaluating defects in a rotary wheel according to claim 1, characterized in that, The calculation method in the adaptive defect classification model includes: The defect feature vector is input into the student model after knowledge distillation to obtain preliminary feature signals; Based on the preliminary feature signals, through dynamic multi-task branching, the confidence level of the defect category, the geometric parameter quantification result, and the risk level assessment result are obtained, which together constitute the detection result; If the detection results include defect types that have not appeared before, manual verification is performed, and multiple samples of related defect types are obtained to fine-tune the student model.

6. The method for detecting and evaluating defects in a rotary wheel according to claim 1, characterized in that, The method for calculating the comprehensive evaluation result of the turbine runner defects includes: Mechanical impact assessment, lifespan prediction, and performance impact assessment are performed in the digital twin, respectively. Based on the results of mechanical impact assessment, life prediction, and performance impact assessment, a comprehensive assessment result of the impeller defects is obtained through integrated calculation.

7. The method for detecting and evaluating defects in a turbine according to claim 6, characterized in that, After obtaining the comprehensive evaluation result of the turbine defect through comprehensive calculation, the following is also included: The impeller was maintained based on the comprehensive assessment results of its defects, and the maintenance results were recorded. Update the digital twin based on the maintenance results.

8. A defect detection and evaluation system for a rotary wheel, characterized in that, include: The data module is used to acquire multi-dimensional information about the external environment and internal structure of the rotor. The feature extraction module is used to input multi-dimensional information about the external environment and the internal structure of the wheel into the feature extraction network to obtain defect feature vectors. The feature extraction network integrates end-to-end super-resolution and feature extraction; The detection module is used to input the defect feature vector into an adaptive defect classification model that includes a multi-task learning framework, and obtain the defect type, geometric parameter quantification results, and risk level assessment results of the wheel, which together constitute the detection results. The evaluation module combines the operational status information from the external multidimensional information and internal structural multidimensional information of the runner, as well as the detection results, to synchronize the defect information in the detection results to the digital twin of the runner, and calculate the comprehensive evaluation result of the runner defects.

9. An electronic device, characterized in that, include: A memory, one or more processors; the memory is coupled to the processors; wherein the memory stores computer program code, the computer program code including computer instructions, and when the computer instructions are executed by the processor, the electronic device performs the steps of the wheel defect detection and evaluation method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the wheel defect detection and evaluation method as described in any one of claims 1-7.

Images