Method and device for predicting defect growth of gas turbine blade and application thereof

Abstract

The application provides a kind of gas turbine blade defect growth prediction method, device and its application, belong to artificial intelligence technical field, including: the video data of gas turbine is input to defect identification model, obtains the defect identification result of each blade of gas turbine, and based on defect identification result determines the target blade of existing defect;Based on video data, generate the three-dimensional point cloud data of target blade, and determine the current geometric dimension information of target blade defect area;Based on historical geometric dimension information and current geometric dimension information, construct defect growth prediction model;Predict the defect growth prediction result of target blade in future preset time period.The application realizes the automatic identification of blade defect and the accurate quantification of blade defect size by defect identification model and three-dimensional point cloud technology, and realizes the evolution prediction of blade defect based on defect growth prediction model, so as to improve the efficiency and accuracy of blade defect detection, and accurately predict the growth trend of defect.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, apparatus and application for predicting defect growth in gas turbine blades. Background Technology

[0002] Gas turbine blades are one of the core components of a gas turbine, and their condition directly determines the overall performance of the gas turbine. Defect inspection of gas turbine blades is an important step in ensuring the safe operation of the equipment.

[0003] Currently, existing technologies typically employ periodic borehole inspections to check for defects in gas turbine blades.

[0004] However, this method relies heavily on manual inspection, takes a long time, has low efficiency and low defect identification accuracy, and the defect size and growth trend of the blade can only be estimated based on human experience, making it difficult to achieve accurate quantification. Summary of the Invention

[0005] This invention provides a method, device, and application for predicting the defect growth of gas turbine blades, which solves the problems of low work efficiency and low defect identification accuracy caused by reliance on manual inspection in the prior art, as well as the difficulty in accurately quantifying the size and growth trend of blade defects. It realizes the automated identification, accurate quantification, and accurate prediction of defect growth trends of blade defects.

[0006] This invention provides a method for predicting defect growth in gas turbine blades, comprising the following steps: The video data of the gas turbine is input into the defect identification model, the defect identification result of each blade of the gas turbine is obtained from the defect identification model, and the target blade with defects is determined based on the defect identification result; Based on the video data, generate three-dimensional point cloud data of the target blade, and determine the current geometric dimensions of the defect area of ​​the target blade based on the three-dimensional point cloud data; Obtain historical geometric dimension information of the defect region of the target blade, and construct a defect growth prediction model based on the historical geometric dimension information and the current geometric dimension information; The defect growth prediction model is used to predict the defect growth of the target blade within a preset time period in the future. The defect identification model is trained based on blade samples labeled with defect locations and defect types.

[0007] According to a method for predicting defect growth in a gas turbine blade provided by the present invention, the step of generating three-dimensional point cloud data of the target blade based on the video data, and determining the current geometric dimension information of the defect region of the target blade based on the three-dimensional point cloud data, includes: The target blade is captured from the video data in a dual-view image, and the dual-view image is subjected to distortion correction and stereo correction based on the calibration parameters of the binocular camera to obtain a dual-view corrected image. Stereo matching is performed on the dual-view corrected image to obtain a pixel-level disparity map, and a depth map of the target blade and the three-dimensional point cloud data are generated based on the pixel-level disparity map and the calibration parameters. Based on the three-dimensional point cloud data, the defect region is extracted to obtain the current geometric size information of the defect region of the target blade.

[0008] According to a method for predicting defect growth in a gas turbine blade provided by the present invention, the step of constructing a defect growth prediction model based on the historical geometric dimension information and the current geometric dimension information includes: Based on the historical geometric dimension information and the current geometric dimension information, a time series of defect dimensions is constructed; The defect size time series is fitted with linear regression to obtain the defect growth prediction model.

[0009] According to the defect growth prediction method for gas turbine blades provided by the present invention, after obtaining the defect identification results of each blade of the gas turbine output by the defect identification model, the method further includes: Leaf detection is performed on each frame of the video data to obtain the position information of each leaf in each frame of the image; The location information is input into the trajectory tracking model to obtain the running trajectory of each blade output by the trajectory tracking model; The blade rotation direction is determined based on the direction of the running trajectory, and the crossing line is determined based on the blade rotation direction; Based on the relative positional relationship between the running trajectory and the crossing line, the total number of blades of the gas turbine is determined; Identify a specific blade among all the blades of the gas turbine, the specific blade being marked with a preset identifier; Based on the position of the specified blade in the gas turbine and the total number of blades, a mapping relationship between blade serial number and blade position is established, and the serial number of each blade in the gas turbine is determined based on the mapping relationship. The defect identification result of each blade of the gas turbine is associated with the serial number of each blade.

[0010] According to a method for predicting defect growth in gas turbine blades provided by the present invention, after associating the defect identification result of each blade of the gas turbine with the serial number of each blade, the method further includes: The serial number is used as an index to construct an information database; The static attribute information of the blade is stored in the information database, and the static attribute information includes one or more of the following: manufacturer, model and material information; The records of changes in the installation position of the blade at different times are stored in the information database; The operation and maintenance records of the blades are stored in the information database; The defect identification results of the blades are stored in the information database.

[0011] According to a method for predicting defect growth in a gas turbine blade provided by the present invention, after predicting the defect growth prediction result of the target blade within a preset time period using the defect growth prediction model, the method further includes: If the defect growth prediction result exceeds a preset safety threshold and / or the defect growth rate of the target blade exceeds a preset rate threshold, a first warning message will be issued.

[0012] According to the defect growth prediction method for gas turbine blades provided by the present invention, after obtaining the defect identification results of each blade of the gas turbine output by the defect identification model, the method further includes: Based on the defect identification results and the current geometric dimensions of the defect area of ​​the target blade, the defect level of the target blade is determined; A health assessment of the target blade is performed based on the defect level of the target blade to obtain the health assessment result of the target blade; If any of the health assessment results do not meet the preset assessment conditions, a second warning message will be issued.

[0013] According to the present invention, a method for predicting the defect growth of a gas turbine blade, after determining the defect level of the target blade, further includes: Obtain manual correction instructions, which include feedback on modifications to the defect level, defect type, and defect location; The defect identification result is updated based on the manual correction instruction; The updated defect identification results and their corresponding blade images are stored in the training sample library. The defect identification model is trained and updated based on the data in the training sample library.

[0014] According to the present invention, a method for predicting the defect growth of a gas turbine blade, wherein training and updating the defect identification model based on data from the training sample library includes: The training sample library is subjected to format validation and outlier filtering to obtain valid sample data; Monitor the increment of the valid sample data. If the increment exceeds a preset increment threshold, use the valid sample data to train and update the defect identification model.

[0015] The present invention also provides a defect growth prediction device for gas turbine blades, comprising the following modules: The defect identification module is used to input video data of the gas turbine into the defect identification model, obtain the defect identification result of each blade of the gas turbine output by the defect identification model, and determine the target blade with defects based on the defect identification result; The defect region determination module is used to generate three-dimensional point cloud data of the target blade based on the video data, and determine the current geometric size information of the defect region of the target blade based on the three-dimensional point cloud data; The prediction model building module is used to obtain historical geometric dimension information of the defect area of ​​the target blade, and to build a defect growth prediction model based on the historical geometric dimension information and the current geometric dimension information. The defect growth prediction module is used to predict the defect growth of the target blade within a preset time period using the defect growth prediction model. The defect identification model is trained based on blade samples labeled with defect locations and defect types.

[0016] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the defect growth prediction method for gas turbine blades as described above.

[0017] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the defect growth prediction method for gas turbine blades as described above.

[0018] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the defect growth prediction method for gas turbine blades as described above.

[0019] The present invention provides a method, device and application for predicting the defect growth of gas turbine blades. It achieves automatic identification of blade defects and accurate quantification of blade defect size through defect identification model and three-dimensional point cloud technology, and predicts the evolution of blade defects based on defect growth prediction model, thereby improving the efficiency and accuracy of blade defect detection and accurately predicting the defect growth trend. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0021] Figure 1 This is one of the flowcharts of the defect growth prediction method for gas turbine blades provided by the present invention.

[0022] Figure 2 This is a flowchart illustrating the process of determining the geometric dimensions of the defective region of a blade, as provided by the present invention.

[0023] Figure 3 This is a schematic diagram of the process for constructing a defect growth prediction model provided by the present invention.

[0024] Figure 4 This is a flowchart illustrating the process of associating the defect identification results and serial numbers of blades provided by the present invention.

[0025] Figure 5 This is a schematic diagram of the process for assessing the health of leaves provided by the present invention.

[0026] Figure 6 This is a flowchart illustrating the process of updating defect identification results based on manual correction instructions provided by the present invention.

[0027] Figure 7 This is a full-cycle management architecture diagram of gas turbine blades provided by the present invention.

[0028] Figure 8 This is a schematic diagram of the process for precise blade identification and positioning provided by the present invention.

[0029] Figure 9 This is a schematic diagram of the binocular vision measurement process provided by the present invention.

[0030] Figure 10 This is a schematic diagram of the defect growth prediction device for gas turbine blades provided by the present invention.

[0031] Figure 11 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0033] It should be noted that, in the description of this invention, the terms "comprising," "including," or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0034] The terms "first," "second," etc., used in this invention are used to distinguish similar objects, not to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class, without limiting the number of objects; for example, a first object can be one or more. Furthermore, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0035] To facilitate a full understanding of the technical solution of this application, the following content is hereby introduced: As a core power source in aviation, power, and industrial drive industries, the reliability, efficiency, and lifespan of gas turbines directly impact system safety. Gas turbine blades are one of the core components of the entire system, and their condition directly determines the overall performance of the gas turbine. Operating in extreme high-temperature environments exceeding the melting point of metals and subjected to enormous centrifugal stress and complex vibrations, their aerodynamic design directly determines the unit's efficiency and power output. As the most valuable and technologically advanced component, the materials, manufacturing, and cooling technologies of gas turbine blades represent the pinnacle of modern industry. Their health is crucial to the safety and reliability of the entire system; failure could lead to catastrophic accidents and massive unplanned downtime losses. Therefore, meticulous life-cycle management of gas turbine blades is a key aspect of ensuring energy security.

[0036] Currently, the inspection of gas turbine blades mainly relies on periodic borehole inspections, but traditional methods have shortcomings: First, there are data silos, historical experience gaps and collaborative failures. Traditional data management methods cannot centrally store gas turbine blade data, resulting in scattered data that is difficult to centralize and reuse. Second, it relies heavily on manual inspection, which is inefficient and inaccurate. On average, manual inspection of gas turbine blade damage takes a long time, and the damage size can usually only be estimated. Defect classification is complex, leading to low personnel efficiency.

[0037] Furthermore, existing technologies include a blade temperature measurement system based on colorimetric thermometry. This system calculates the object's temperature by measuring the intensity of radiation at two different wavelengths to reduce the impact of the object's emissivity on measurement accuracy. It can measure blade temperature in real time, segment blade temperature data, and perform some analysis and storage. However, this system primarily focuses on temperature parameter measurement and lacks the ability to automatically identify and quantify surface defects in gas turbine blades (such as cracks, ablation, and corrosion), thus failing to meet the needs of comprehensive condition assessment.

[0038] Another existing technology discloses an intelligent detection and defect analysis method for gas turbine blades. This method acquires blade material properties, geometric structure, and operating load data, calculates the stress transfer matrix, and analyzes multi-node stress gradient changes to locate stress concentration points. Essentially, this technology uses visual images as an auxiliary verification or data supplement for stress analysis, rather than directly performing deep learning and interpretation of visual features. Its core remains a mechanical model, and visual data does not play a dominant role, potentially leading to insensitivity to defects that do not exhibit obvious stress anomalies but have existing surface damage.

[0039] In addition, existing technology discloses a compressor blade condition monitoring method and system based on a health management platform. This method primarily acquires real-time operating parameters of the gas turbine, calculates the performance parameter degradation, and identifies blade faults based on preset fault diagnosis criteria. This method indirectly infers blade faults through macroscopic degradation of system performance; it is a black-box or gray-box approach. Typically, blade damage can only be detected when it reaches a certain level and genuinely affects the overall performance of the unit. Furthermore, it is difficult to directly locate the specific faulty blade, let alone identify the specific type of defect.

[0040] In summary, existing technologies either focus on non-visual parameter analysis or are limited to traditional image processing with single functions, generally lacking the ability to detect early, minute defects and accurately quantify and track defects. Most technologies fail to fully utilize the powerful feature extraction and classification capabilities of deep learning in visual defect recognition, nor do they construct defect growth prediction models centered on visual evolution sequences, thus hindering the achievement of truly automated, intelligent, and precise blade defect identification and predictive maintenance.

[0041] This invention proposes a method for predicting the defect growth of gas turbine blades, which realizes automated, high-precision quantitative identification and measurement of blade defects, and constructs a prediction model based on historical data sequences to accurately estimate the remaining life of gas turbine blades, thereby upgrading from passive maintenance to proactive predictive maintenance and improving safety and reliability.

[0042] The following is combined with Figures 1-11 This invention describes the method, apparatus, and application of defect growth prediction for gas turbine blades.

[0043] Figure 1 This is one of the flowcharts illustrating the defect growth prediction method for gas turbine blades provided by the present invention, such as... Figure 1 As shown, the execution subject of the defect growth prediction method for gas turbine blades provided by the present invention can be an industrial control computer, a server, a cloud computing platform, or a computer capable of executing the method of the present invention, etc. Unless otherwise specified, the following embodiments will be described using an industrial control computer as an example.

[0044] As an optional embodiment, the defect growth prediction method for gas turbine blades mainly includes, but is not limited to, the following steps: Step 110: Input the video data of the gas turbine into the defect identification model, obtain the defect identification results of each blade of the gas turbine output by the defect identification model, and determine the target blade with defects based on the defect identification results; wherein, the defect identification model is trained based on blade samples labeled with defect location and defect type.

[0045] Video data of a gas turbine refers to a continuous sequence of images captured by image acquisition equipment during the shutdown and maintenance of the gas turbine, showing the internal blades. For example, a binocular camera installed at the end of a borescope can record a complete video of all rotating rotor blades in a turning state, thereby obtaining a high-definition video stream containing the surface conditions of all blades.

[0046] A defect identification model refers to a deep learning model that can automatically extract features from images and identify leaf defects. For example, a defect identification model can be a target detection model based on a convolutional neural network (CNN), such as the YOLO (You Only Look Once) series and Faster R-CNN (Faster Region-based Convolutional Neural Network).

[0047] As an optional embodiment, the defect identification model can also be a two-stage identification model. In the first stage, an object detection network (such as Yolox-nano) is used to detect all suspected defect areas in the video frame without classifying them, and then crop out small defect images. In the second stage, the cropped defect images are input into a classification network, such as a lightweight deep neural network (MobileNet-v2), for fine classification, thereby identifying various specific defect types such as tearing, burn-through, notch, pit, ablation, crack, curling, material loss, and coating loss.

[0048] Defect identification results refer to the detailed information about blade defects output by the defect identification model. For example, defect identification results may include the location coordinates of the defect in the image (such as bounding box coordinates), the confidence score of the defect, and the specific category label of the defect.

[0049] Determining a defective target leaf can be achieved by analyzing the defect identification results output by the defect identification model. For example, if the confidence of a detection box output by the defect identification model is greater than a preset threshold (such as 0.45), and the classification label corresponding to the detection box belongs to a preset defect type, then the corresponding leaf in the current image is determined to be a defective target leaf.

[0050] The defect identification model can be trained through the following steps: First, a large amount of historical image data of gas turbine blades is collected as a sample set; then, the sample set is manually labeled, and the labeling information includes defect type, defect level and location information, and is subjected to format verification and outlier filtering to remove records that do not meet the specifications; finally, the labeled valid sample data is input into the deep learning network to be trained for iterative training, and the model parameters are continuously optimized through the backpropagation algorithm until the defect identification accuracy reaches the preset requirements.

[0051] Step 120: Generate three-dimensional point cloud data of the target blade based on video data, and determine the current geometric dimensions of the defect area of ​​the target blade based on the three-dimensional point cloud data.

[0052] The three-dimensional point cloud data of a target blade refers to a set of discrete points that can characterize the spatial geometry of the target blade surface. For example, three-dimensional point cloud data contains the three-dimensional representation of each point on the blade surface. X , Y , Z Coordinate information can accurately reflect the curved shape of the blade and the depth of surface defects. The generation of three-dimensional point cloud data of the target blade can be achieved through binocular stereo vision technology.

[0053] The current geometric dimension information of the defect area of ​​the target blade refers to the numerical result of the defect after quantification in three-dimensional space. For example, the geometric dimension information may include specific parameters such as the length, width, depth, area and volume of the defect area, which can objectively reflect the severity of the defect at the current moment.

[0054] Step 130: Obtain historical geometric dimension information of the defect area of ​​the target blade, and construct a defect growth prediction model based on the historical geometric dimension information and the current geometric dimension information.

[0055] Historical geometric dimension information of the target blade defect area refers to the defect dimension data recorded at multiple past maintenance or monitoring points of the target blade. For example, historical geometric dimension information can be obtained by querying a pre-established blade life cycle asset management database, which stores quantitative data such as defect length and area obtained by the above-mentioned visual measurement methods during each maintenance of the blade, as well as their corresponding timestamps.

[0056] A defect growth prediction model refers to a mathematical or algorithmic model that can describe the change in defect size over time. For example, a defect growth prediction model can be a linear regression model built based on time series analysis, or a multinomial regression model or other machine learning prediction model, used to fit the growth trend of defects.

[0057] Considering that surface defects (such as cracks and corrosion) on gas turbine blades often tend to gradually expand over time under high temperature, high pressure, and centrifugal force, and that this growth usually follows a certain regularity, this invention constructs a defect growth prediction model based on historical and current geometric dimension information. Specifically, this invention arranges multiple historical measurement data and current measurement data in chronological order to form a time series of defect dimensions, and uses a regression algorithm to fit this series, thereby quantifying the evolution rate of the defect and providing a scientific basis for subsequent blade life prediction and maintenance decisions.

[0058] Step 140: Use the defect growth prediction model to predict the defect growth of the target blade within a preset time period in the future.

[0059] The future preset time period refers to a specific duration that is projected backward from the current detection time. This duration is usually set according to the gas turbine's maintenance cycle or operating plan. For example, the future preset time period can be the next 90 days, the next maintenance cycle, or the time span until the next planned shutdown.

[0060] The defect growth prediction result refers to the defect size and status that the target blade may reach at the end of or during a preset time period, based on the defect growth prediction model. For example, the defect growth prediction result may include the predicted defect length value, area value, or the specific time point at which the predicted defect size exceeds the safety threshold, as well as the defect growth rate (such as the monthly growth rate) calculated based on these predicted values, so as to help maintenance personnel determine in advance whether intervention or blade replacement is necessary.

[0061] The defect growth prediction method provided by this invention achieves automatic identification of blade defects and accurate quantification of blade defect size through a defect identification model and three-dimensional point cloud technology, and predicts the evolution of blade defects based on the defect growth prediction model, thereby improving the efficiency and accuracy of blade defect detection and accurately predicting defect growth trends.

[0062] Figure 2 This is a flowchart illustrating the process of determining the geometric dimensions of the defective region of a blade, as provided by the present invention. Figure 2 As shown, as another optional embodiment provided by the present invention, three-dimensional point cloud data of the target blade is generated based on video data, and the current geometric dimension information of the defect area of ​​the target blade is determined based on the three-dimensional point cloud data, including but not limited to the following steps: Step 210: Obtain dual-view images of the target blade from the video data, and perform distortion correction and stereo correction on the dual-view images based on the calibration parameters of the binocular camera to obtain dual-view corrected images.

[0063] Dual-view images of a target blade refer to a pair of images obtained by taking pictures of the same target blade from two different shooting angles at the same time using a binocular camera device. For example, dual-view images include a left view and a right view, which contain the texture and geometric information of the target blade from different perspectives.

[0064] The calibration parameters of a stereo camera refer to the set of parameters that describe the internal optical characteristics and external spatial positional relationship of the stereo camera. For example, the calibration parameters include the camera's intrinsic parameters (such as focal length and principal point coordinates) and extrinsic parameters (such as the rotation matrix and translation vector between the two cameras). These parameters are usually obtained in advance through a high-precision stereo calibration process.

[0065] Distortion correction and stereo correction refer to the processes of eliminating radial and tangential distortions in an image caused by lens perspective, and geometric transformations of the left and right images to make their optical axes parallel and their imaging planes coplanar. For example, distortion correction can restore the true geometric shape of an image, while stereo correction can make corresponding points in the left and right views lie on the same horizontal scan line, thereby simplifying subsequent stereo matching calculations.

[0066] Dual-view corrected images refer to standardized left and right image pairs output after correction processing. For example, in a dual-view corrected image, any point in the left image will have a corresponding point in the right image that is located at the same row coordinate, with only differences in column coordinates. Such dual-view corrected images are the basis for generating high-quality disparity maps.

[0067] Step 220: Perform stereo matching on the dual-view corrected images to obtain a pixel-level disparity map, and generate a depth map and three-dimensional point cloud data of the target blade based on the pixel-level disparity map and calibration parameters.

[0068] Stereo matching refers to the process of finding corresponding image points in a pair of dual-view corrected images. That is, for a point found in the left image, the corresponding physical point in the right image is found. For example, a semi-global block matching algorithm (SGBM) can be used for stereo matching. This algorithm constructs a cost cube (such as Census transform or absolute value of gray difference), combines multi-path dynamic programming to optimize the energy function, and uses left-right consistency detection and weighted least squares (WLS) filtering to optimize the edges, thereby accurately finding the correspondence between pixels in the left and right images.

[0069] A pixel-level disparity map is an image matrix that records the difference in the horizontal position of corresponding pixels in the left and right images. For example, the gray value of each pixel in the disparity map represents the difference in the horizontal coordinates of that point in the left and right views. The larger the disparity value, the closer the object is to the camera, and vice versa.

[0070] A depth map of a target blade refers to an image that describes the vertical distance of each point on the blade surface from the optical center of the camera. For example, each pixel value in the depth map directly corresponds to the depth value of that point in three-dimensional space.

[0071] Specifically, disparity can be converted into depth using the principles of triangulation. For example, the calculated disparity value can be... d Substitute into the formula In, among them, For depth value, For camera focal length, The baseline distance of the binocular camera. d To account for parallax, a depth map is generated. Then, based on the camera's intrinsic parameter model, each pixel in the depth map is mapped back to a 3D spatial coordinate system, ultimately generating a depth map containing... X , Y , Z Dense 3D point cloud data with coordinate information is used to reconstruct the 3D morphology of the blade surface.

[0072] Step 230: Extract the defect region based on the 3D point cloud data to obtain the current geometric size information of the defect region of the target blade.

[0073] Specifically, firstly, based on the depth abrupt change or curvature change characteristics in the 3D point cloud data, the point cloud belonging to the defect area is segmented from the background blade point cloud to complete the defect region extraction. Then, Principal Component Analysis (PCA) is used to analyze the extracted defect point cloud to determine the principal axis direction of the defect, and the maximum projection distance of the point cloud in the principal axis direction is calculated as the length of the defect. At the same time, the point cloud of the defect region is triangulated to construct a continuous triangular mesh model from the discrete point cloud. The surface area of ​​the defect is calculated by accumulating the areas of all the tiny triangles, thereby obtaining the current geometric dimension information including dimensions such as length and area.

[0074] The defect growth prediction method provided by this invention employs binocular stereo vision technology to perform rigorous distortion and stereo correction on dual-view images, and generates high-precision three-dimensional point cloud data based on pixel-level disparity maps. This effectively eliminates the influence of lens distortion and shooting angle on the measurement, restores the true three-dimensional morphology of the blade surface, and thus significantly improves the accuracy and reliability of defect geometric dimension measurement, providing a high-quality data foundation for subsequent trend prediction.

[0075] Figure 3 This is a schematic diagram of the process for constructing a defect growth prediction model provided by the present invention, as shown below. Figure 3 As shown, as another optional embodiment provided by the present invention, a defect growth prediction model is constructed based on historical geometric dimension information and current geometric dimension information, including but not limited to the following steps: Step 310: Construct a time series of defect dimensions based on historical and current geometric dimension information.

[0076] A defect size time series refers to an ordered set of data that arranges the defect size data of the same target blade at different historical moments in chronological order. For example, the defect size time series can be represented as a series of binary pairs (time point, defect size value), where the time point corresponds to the date of each maintenance, and the defect size value corresponds to the defect length or area measured during that maintenance.

[0077] Step 320: Perform linear regression fitting on the defect size time series to obtain the defect growth prediction model.

[0078] Linear regression fitting refers to a statistical analysis method that aims to find the best-fitting straight line that can describe the linear relationship between defect size and time as accurately as possible. For example, the slope and intercept of the line are calculated by the least squares method, where the slope represents the average growth rate of the defect.

[0079] Specifically, the regression equation can be calculated using a linear regression algorithm, with time points in the defect size time series as independent variables and defect size values ​​as dependent variables. y = kx + b The parameters, where y Represents the predicted defect size. x Represents time, k For the growth slope, b Using the initial intercept, a defect growth prediction model is obtained that can predict the defect size at any future time.

[0080] The defect growth prediction method provided by this invention establishes a defect growth prediction model by constructing a defect size time series and using linear regression fitting. It can make full use of historical evolution data to mine the inherent laws of defect growth, quantify the growth rate of defects over time, and thus achieve scientific extrapolation and accurate prediction of future defect states, providing data support for the formulation of reasonable maintenance plans.

[0081] Figure 4 This is a flowchart illustrating the process of associating the defect identification results and serial numbers of blades provided by the present invention, as shown below. Figure 4 As shown, as another optional embodiment provided by the present invention, after obtaining the defect identification results of each blade of the gas turbine output by the defect identification model, the following steps are also included, but are not limited to: Step 410: Perform leaf detection on each frame of the video data to obtain the position information of each leaf in each frame.

[0082] Leaf detection refers to the process of identifying leaf targets in video frames using object detection algorithms. For example, object detection models such as YOLOv8 can be used to process each frame of video image and identify all leaf regions contained in the image.

[0083] Location information refers to the specific area of ​​the detected leaf within the image coordinate system. For example, location information can be derived from the coordinates of the top-left corner of the target detection box (…). X min , Y min ) and the coordinates of the lower right corner ( X max , Y maxThe symbol ) is used to accurately locate the position of the blade in the current frame.

[0084] Step 420: Input the position information into the trajectory tracking model to obtain the running trajectory of each blade output by the trajectory tracking model.

[0085] A trajectory tracking model refers to an algorithm model that can associate the same target in consecutive video frames. For example, a trajectory tracking model can be the ByteTrack tracking algorithm, which associates and matches leaf targets detected in different frames, assigning the same leaf the same identity ID in different frames.

[0086] The trajectory refers to the set of position coordinates of the same blade in multiple consecutive frames of images. For example, the trajectory records the complete motion path of a blade from entering the video frame to leaving the video frame, and contains a series of ordered coordinate points.

[0087] Step 430: Determine the blade rotation direction based on the direction of the running trajectory, and determine the crossing line based on the blade rotation direction.

[0088] The blade rotation direction refers to the direction of rotation of the gas turbine rotor in a turning state. For example, the overall rotation of the blade can be determined by calculating the average direction vector of multiple running trajectories.

[0089] The crossing line refers to a virtual reference line used to assist in counting. For example, the crossing line can be a vertical bisector perpendicular to the direction of blade rotation and located in the center of the image. The counting logic is triggered when the blade trajectory crosses this line.

[0090] Specifically, the direction vectors from the starting point to the ending point of all trajectories can be calculated, and the 10 longest average directions can be selected as the final blade rotation direction. Based on this, a center line perpendicular to the blade rotation direction can be set as the crossing line.

[0091] Step 440: Determine the total number of gas turbine blades based on the relative positional relationship between the running trajectory and the crossing line.

[0092] The relative positional relationship refers to whether the blade's trajectory intersects with the crossing line. For example, it involves calculating the coordinates of the intersection point between the trajectory and the crossing line to determine whether the starting and ending points of the trajectory are located on opposite sides of the crossing line.

[0093] Specifically, the distance between the intersection of the trajectory and the crossing line can be calculated. If the crossing condition is met and the direction is consistent with the rotation direction, the counter is incremented by 1. The final total number obtained is the total number of blades of the gas turbine.

[0094] Step 450: Identify a specific blade among all the blades of the gas turbine. The specific blade is marked with a preset identifier.

[0095] A designated blade refers to a specific blade that serves as the positioning reference for the entire rotor blades. For example, the designated blade is usually the first blade on the rotor and is called the reference blade or the first blade.

[0096] Preset identifiers refer to unique symbols or codes engraved or marked on the base of a specified leaf. For example, a preset identifier may be an engraved number "R1-C1-S01" or a specific geometric feature point, used to distinguish the leaf from other leaves that look similar.

[0097] Step 460: Based on the specified blade position in the gas turbine and the total number of blades, establish a mapping relationship between blade serial number and blade position, and determine the serial number of each blade in the gas turbine based on the mapping relationship.

[0098] Specifically, starting from the identified designated blade (e.g., number 1), and combining the determined blade rotation direction and the total number of blades N, the subsequent blades are sequentially numbered incrementally, thereby establishing a one-to-one correspondence between physical location and logical number (1 to N). For example, if the designated blade is R1-C1-S01, then the next blade is automatically marked as R1-C1-S02, and so on.

[0099] Step 470: Associate the defect identification results of each gas turbine blade with the serial number of each blade.

[0100] Considering the large number of gas turbine blades and their similar appearance, without precise numbering and association, defect data cannot be located to specific physical blades, resulting in data silos. Therefore, this invention binds the defect identification result of each blade to its unique serial number through the above-mentioned trajectory tracking, counting, and reference blade identification steps, thereby enabling precise positioning and full life-cycle tracking management of each specific blade asset.

[0101] Specifically, when a defect is identified in a blade, the defect will be automatically recorded and archived in the information database corresponding to the blade number based on the current counting status and mapping relationship. For example, the identified crack information will be directly written into the database record of blade R1-C1-S05.

[0102] The defect growth prediction method provided by this invention combines blade detection, trajectory tracking, rotation direction determination, and cross-line counting technologies to achieve automatic counting of gas turbine blades. By using designated blades with preset labels as reference anchor points, a precise mapping relationship between physical location and logical sequence number is established. This method can solve the problems of data silos and difficult positioning in traditional methods, and accurately associate defect information with each specific physical blade, thereby achieving refined asset management and full life cycle traceability at the blade level.

[0103] In another embodiment of the present invention, after associating the defect identification result of each gas turbine blade with the serial number of each blade, the method further includes: Use serial numbers as indexes to build an information database.

[0104] Specifically, a structured data table or document collection is established using the unique serial number of each blade (such as R1-C1-S01) as the primary key. This is used to centrally store all lifecycle data related to the blade. For example, a dedicated digital archive folder is created for each blade.

[0105] The static attribute information of the blades is stored in the information database. The static attribute information includes one or more of the following: manufacturer, model, and material information.

[0106] Static attribute information is usually entered when the blade leaves the factory or enters the warehouse. It serves as the basic identity data of the blade and supports subsequent queries and statistical analysis.

[0107] Records of changes in the installation location of the blades at different times are stored in the information database.

[0108] Installation location change records refer to the historical trajectory information of the blade being installed in different units, different rotor stages, or different slot numbers throughout its entire life cycle. For example, it records the entire process of a blade being removed from the first-stage rotor of unit A, repaired, and reinstalled on the second-stage rotor of unit B, thus enabling full traceability of the installation history.

[0109] The operation and maintenance records of the blades are stored in the information database.

[0110] Operation and maintenance logs refer to detailed logs of all maintenance activities performed on the blade, including manual or automated operation and maintenance actions such as the time of each inspection, cleaning records, coating repair operations, grinding treatment, and replacement of parts.

[0111] The defect identification results of the blades are stored in the information database.

[0112] Considering that gas turbine blades are high-value assets with a long lifespan and complex operating conditions, the lack of unified data management will lead to information fragmentation. Therefore, this invention constructs an information database indexed by blade serial numbers to centrally link and store multi-dimensional data such as static attributes, installation location changes, operation and maintenance, and defect identification results. This breaks down data silos and enables transparent insight and refined management of each blade from its entry into the warehouse to its scrapping.

[0113] The defect growth prediction method provided by this invention, by constructing an information database indexed by blade serial number and integrating static attributes, installation change records, operation and maintenance records and dynamic defect data, can break the data silos and fragmentation problems of traditional management models, realize centralized storage and correlation analysis of multi-dimensional blade data, and thus provide comprehensive and complete data support for the whole life cycle asset management, status traceability and scientific decision-making of blades.

[0114] In another embodiment of the present invention, after predicting the defect growth prediction result of the target blade within a preset time period using the defect growth prediction model, the method further includes: if the defect growth prediction result exceeds a preset safety threshold and / or the defect growth rate of the target blade exceeds a preset rate threshold, then a first warning message is issued.

[0115] The preset safety threshold refers to the upper limit of defect size set according to the gas turbine design specifications or safe operation standards. For example, for crack length, the preset safety threshold can be 3 mm. Once the predicted value exceeds this value, it means that the blade is at risk of failure.

[0116] The defect growth rate of the target blade refers to the increase in defect size per unit time calculated based on the defect growth prediction model. For example, it can be represented by calculating the slope of the prediction curve or the monthly growth percentage, such as a monthly growth rate greater than 15%.

[0117] The preset rate threshold refers to the maximum speed limit at which defects are allowed to grow. For example, when a defect grows extremely fast even though it is still small in size (such as a monthly growth rate of more than 15%), it indicates that the defect is in an unstable expansion period. The preset rate threshold is used to capture such anomalies of rapid deterioration.

[0118] The first warning message refers to the prompt signal automatically generated by the system when it detects the above-mentioned risks. For example, the first warning message may be a red highlighted alarm that pops up on the monitoring interface, accompanied by a text message or APP notification pushed to the mobile terminal of the relevant person in charge, reminding the operation and maintenance personnel that the blade may have an over-limit failure within a specific time in the future and that they need to pay attention immediately.

[0119] Figure 5 This is a schematic diagram of the process for assessing leaf health provided by the present invention, as shown below. Figure 5 As shown, as another optional embodiment provided by the present invention, after obtaining the defect identification results of each blade of the gas turbine output by the defect identification model, the following steps are also included, but are not limited to: Step 510: Based on the defect identification results and the current geometric size information of the defect area of ​​the target blade, determine the defect level of the target blade.

[0120] The defect level of the target blade refers to the classification description of the severity of the defect. For example, the defect can be divided into three levels: severe, moderate and minor, each corresponding to a different risk coefficient.

[0121] Specifically, the defect level can be determined by preset classification rules. For example, by combining the defect type (such as cracks being more serious than coating peeling) and size (such as length greater than 5mm being serious), defects can be classified as serious, moderate, or minor.

[0122] Step 520: Perform a health assessment on the target blade based on its defect level to obtain the health assessment result of the target blade.

[0123] Health assessment results refer to comprehensive indicators that quantitatively reflect the current health status of the blades. For example, health assessment results can be the Blade Health Index (BHI), which ranges from 0 to 100 points. The higher the score, the better the health status.

[0124] Specifically, health assessment results can be generated by a preset algorithm based on the unit or blade. For example, a weighted calculation can be performed based on the defect level, with a weight of 5 points for severe defects, 3 points for moderate defects, and 1 point for minor defects. A health score between 0 and 100 points can be obtained through comprehensive calculation, and this score is the health assessment result of the blade.

[0125] Step 530: If any health assessment result does not meet the preset assessment conditions, a second warning message is issued.

[0126] Preset assessment criteria refer to the minimum passing standard that the health assessment results must meet. For example, the preset assessment criteria could be a health score higher than 60 points, and a score lower than that would be considered as being in a sub-healthy or malfunctioning state.

[0127] The second warning message refers to the alarm prompt issued when the health score is too low. For example, the system interface may display that the health of the blade is not up to standard and suggest that the machine be shut down for maintenance or the blade be replaced to prevent the fault from escalating.

[0128] The defect growth prediction method provided by this invention determines the defect level by combining defect type and geometric size, and constructs a quantitative health assessment system accordingly. It can transform complex defect data into intuitive health scores and automatically trigger an early warning when the health level is lower than the preset standard, thereby helping maintenance personnel to quickly grasp the overall status of the blade and realize the transformation from passive maintenance to proactive health management.

[0129] Figure 6 This is a flowchart illustrating the process of updating defect identification results based on manual correction instructions provided by the present invention, as shown below. Figure 6As shown, as another optional embodiment provided by the present invention, after determining the defect level of the target blade, the following steps are included, but not limited to: Step 610: Obtain manual correction instructions, which include feedback on modifications to the defect level, defect type, and defect location.

[0130] Specifically, the system can provide an interactive interface for professional engineers to review the results of automatic identification. If the professional engineer finds that the system identification is incorrect (such as misreporting oil stains as cracks, or overestimating the defect level), they can manually adjust it through the interface tools. For example, they can modify the position of the defect boundary box, correct the defect category label, or reset the defect level. The system receives the instructions generated by these operations as manual correction instructions.

[0131] Step 620: Update the defect identification results based on manual correction instructions.

[0132] Updating defect identification results refers to overwriting the original automatically identified information with accurate information confirmed by manual review. For example, the status of the record in the database is marked as "manually reviewed" and the corrected defect type and size data are saved.

[0133] Step 630: Store the updated defect identification results and their corresponding blade images into the training sample library.

[0134] The training sample library refers to a database used to store high-quality labeled data for subsequent model learning. For example, images that have been manually corrected (especially "difficult samples" that the system originally misidentified or missed) are added to the sample library as positive or negative samples, while retaining the accurate labeling information after manual correction.

[0135] Step 640: Train and update the defect identification model based on the data in the training sample library.

[0136] Specifically, the original defect identification model is iteratively trained using the updated sample set, so that the defect identification model can learn new defect features or correct previous misconceptions, thereby achieving continuous evolution of the algorithm's capabilities.

[0137] The defect growth prediction method provided by this invention introduces a manual review and correction process, transforming expert experience into high-quality labeled data stored in the training sample library. Based on this, the model is continuously trained and updated. This effectively solves the problem of false alarms or missed alarms that may occur in the model under complex working conditions, enabling the model performance to continuously evolve over time and ensuring high reliability under long-term operation.

[0138] As an optional embodiment, the present invention also provides a complete model management and update mechanism. First, a three-level annotation data management system is established, specifically by setting up three storage levels in the project root directory: original annotation data, re-annotated data, and annotation metadata. Through hierarchical management, strict control and traceability of data versions are achieved, thereby realizing version control of original annotation data and re-annotated data and ensuring the traceability of data iteration.

[0139] Secondly, a model repository is built, with a storage structure that includes current and historical versions. Each version directory stores the native PyTorch model files, TFLite model files adapted for mobile inference, input / output format configuration files, and detailed changelogs to meet the deployment needs of different terminals. Furthermore, when a user logs into their terminal device, the system automatically sends a version check request to the server. If a new version is detected, the system downloads a difference file (approximately 3-5MB) generated based on model parameter comparisons. This process supports resuming interrupted downloads, and an update log pop-up (e.g., indicating "Added coating peeling detection, accuracy improved by 3%)" informs the user of the update content upon completion. If no new version is found, the system directly enters the main detection interface, thus achieving efficient and incremental model updates.

[0140] In another embodiment of the present invention, training and updating the defect identification model based on the data of the training sample library includes: performing format verification and outlier filtering on the data of the training sample library to obtain valid sample data; monitoring the increment of valid sample data, and if the increment exceeds a preset increment threshold, using the valid sample data to train and update the defect identification model.

[0141] Format validation refers to checking whether the structure and fields of input data conform to predefined specifications. For example, using JSON Schema to validate annotation files ensures that necessary fields such as image_id, category_id, and bbox are included and that the data types are correct, while removing records with corrupted formats or missing key information.

[0142] Outlier filtering refers to using statistical methods to identify and remove obviously unreasonable labeled data. For example, calculating the standard score of defect size and considering data with a standard score exceeding 3 times the standard deviation as outliers, or directly filtering out noisy data with labeled size exceeding 30% of the blade body length to prevent incorrect labeling from contaminating the model.

[0143] Valid sample data refers to the set of images and labeled data that are of reliable quality and have a standardized format, which are retained after the above verification and filtering processes. For example, these valid sample data represent high-value training resources after cleaning and can truly reflect the defect characteristics of the blades.

[0144] The increment of valid sample data refers to the number of valid samples that have been added to the training sample library and passed the verification since the last model training. For example, the increment can be the number of newly added "manually reviewed" and "verified" images that the system counts in real time.

[0145] The preset incremental threshold refers to the minimum number of samples required to trigger automatic retraining of the defect recognition model. For example, when 500 or more new valid labeled images are added, the system will automatically start a background training task to update the defect recognition model using the accumulated new data.

[0146] The defect growth prediction method provided by this invention establishes a manual review and feedback mechanism to strictly verify the format of the training sample library data and filter outout values. When the effective sample increment reaches a preset threshold, the model is automatically retrained, which enables the defect recognition model to continuously adapt to new working conditions and defect morphologies, thereby continuously improving recognition accuracy and generalization ability.

[0147] As an optional embodiment, the present invention also provides a visual hierarchical navigation function, specifically by constructing a hierarchical navigation interface and establishing an application Geographic Information System (GIS) map. Users can enter through this interface by clicking level by level, with the path covering the unit, rotor, number of revolutions and down to a single blade. When any blade is clicked, the system can pop up an integrated view to display the specific data of the blade, thereby realizing intuitive management and quick query of blade assets.

[0148] Figure 7 This is a full-cycle management architecture diagram of gas turbine blades provided by the present invention, as shown below. Figure 7 As shown, the closed-loop logic of the self-operating architecture for full lifecycle management is demonstrated, which mainly includes four core modules: asset data acquisition, data analysis and processing, data storage and application, and sample library training.

[0149] First, in the asset data acquisition phase, the system collects image and video data, uses AI visual recognition and positioning technology to count and identify blades, and creates blade asset files, thus binding physical blades with digital assets. The collected data then enters the data analysis and processing phase. The system automatically detects various defects on the blade surface using an AI defect visual recognition model and generates 3D point clouds to quantify defect sizes using binocular visual dimensional measurement technology. It also provides a manual review and fine-tuning function, allowing engineers to correct the recognition results and ensure data accuracy. The processed data is then transmitted to the data storage application module, where blade health management is performed. Defect trend analysis is conducted based on historical data, and the blade status is displayed through a multi-dimensional visualization interface to assist in operation and maintenance decisions.

[0150] Finally, the high-quality data, after manual review, flows to the sample library training module. The system preprocesses the data (such as format verification and outlier filtering), uses the accumulated valid samples to continuously train and update the AI ​​model, and distributes the optimized model, thereby continuously improving the system's recognition accuracy and generalization ability, forming a data-driven self-evolutionary closed loop.

[0151] Figure 8 This is a flowchart illustrating the precise blade identification and positioning process provided by the present invention, as shown below. Figure 8 As shown, this process details the complete operational steps from image acquisition to data archiving. After the process begins, image acquisition is performed first to obtain video or image data of the internal blades of the gas turbine. Subsequently, four key tasks are executed in parallel: first, damage on the blade surface is detected using a defect identification model to generate defect data; second, the three-dimensional geometric parameters of the defects are calculated using binocular measurement technology to generate defect size data; third, AI-based identification of unique markers is performed to locate the reference blade with a specific code (e.g., R1-C1-S01); and fourth, AI-based blade counting is performed to count the total number of blades and determine the direction of rotation. Based on the results of unique marker identification and blade counting, the system can successfully extract the blade number and establish a mapping between physical location and logical sequence number.

[0152] Next, the generated defect data, defect size data, and extracted blade numbers are integrated to associate the blades with the data, ensuring that each defect record can be accurately located to a specific blade asset. The associated data then enters a manual review stage, where engineers confirm or correct the identification results. Finally, the entire process is completed, the task is finished, and the data is accurately archived.

[0153] Figure 9 This is a schematic diagram of the binocular vision measurement process provided by the present invention, as shown below. Figure 9 As shown, this process demonstrates the specific steps for accurately quantifying blade defects using binocular vision technology. First, binocular calibration is performed to obtain the camera's intrinsic and extrinsic parameters, providing a benchmark for subsequent calculations. Next, image acquisition is performed, obtaining left and right view images of the target blade. Then, distortion correction is applied to the acquired images using calibration parameters to eliminate lens distortion, and stereo correction is performed to align the left and right images. Based on this, a stereo matching algorithm is executed to calculate the correspondence between the left and right images, generating an initial disparity map, which is then optimized to improve its accuracy and edge quality. The optimized disparity data is converted into depth information, completing the depth map generation. Next, defect segmentation is performed based on the depth map, separating the defect region from the background, and point cloud extraction is performed to reconstruct a 3D point cloud model of the defect region. Size calculations are performed based on the point cloud data to obtain the defect's length, area, and other geometric parameters. Finally, an error compensation mechanism is used to correct the calculation results, further improving measurement accuracy, and the final quantification result is output.

[0154] Figure 10 This is a schematic diagram of the defect growth prediction device for gas turbine blades provided by the present invention, as shown below. Figure 10 As shown, it mainly includes, but is not limited to: The defect identification module 1010 is used to input video data of the gas turbine into the defect identification model, obtain the defect identification results of each blade of the gas turbine output by the defect identification model, and determine the target blade with defects based on the defect identification results. The defect area determination module 1020 is used to generate three-dimensional point cloud data of the target blade based on video data, and to determine the current geometric size information of the defect area of ​​the target blade based on the three-dimensional point cloud data. The prediction model building module 1030 is used to obtain historical geometric size information of the defect area of ​​the target blade, and to build a defect growth prediction model based on the historical geometric size information and the current geometric size information. The defect growth prediction module 1040 is used to predict the defect growth of the target blade within a preset time period using a defect growth prediction model. The defect identification model is trained based on blade samples labeled with defect locations and defect types.

[0155] It should be noted that the defect growth prediction device for gas turbine blades provided by the present invention can execute the defect growth prediction method for gas turbine blades described in any of the above embodiments during specific operation, which will not be elaborated in this embodiment.

[0156] The defect growth prediction device for gas turbine blades provided by this invention realizes automatic identification of blade defects and accurate quantification of blade defect size through defect identification model and three-dimensional point cloud technology, and realizes the evolution prediction of blade defects based on defect growth prediction model, thereby improving the efficiency and accuracy of blade defect detection and accurately predicting defect growth trend.

[0157] Figure 11 This is a schematic diagram of the structure of the electronic device provided by the present invention, such as... Figure 11As shown, the electronic device may include: a processor 1110, a communications interface 1120, a memory 1130, and a communication bus 1140, wherein the processor 1110, the communications interface 1120, and the memory 1130 communicate with each other through the communication bus 1140. The processor 1110 can call logical instructions in the memory 1130 to execute a defect growth prediction method for gas turbine blades. The method includes: inputting video data of the gas turbine into a defect identification model, obtaining the defect identification result of each blade of the gas turbine output by the defect identification model, and determining the target blade with defects based on the defect identification result; generating three-dimensional point cloud data of the target blade based on the video data, and determining the current geometric size information of the defect area of ​​the target blade based on the three-dimensional point cloud data; obtaining historical geometric size information of the defect area of ​​the target blade, and constructing a defect growth prediction model based on the historical geometric size information and the current geometric size information; and using the defect growth prediction model to predict the defect growth prediction result of the target blade in a future preset time period; wherein the defect identification model is trained based on blade samples labeled with defect locations and defect types.

[0158] Furthermore, the logical instructions in the aforementioned memory 1130 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0159] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the defect growth prediction method for gas turbine blades provided by the above methods. The method includes: inputting video data of the gas turbine into a defect identification model, obtaining the defect identification result of each blade of the gas turbine output by the defect identification model, and determining the target blade with defects based on the defect identification result; generating three-dimensional point cloud data of the target blade based on the video data, and determining the current geometric size information of the defect area of ​​the target blade based on the three-dimensional point cloud data; obtaining the historical geometric size information of the defect area of ​​the target blade, and constructing a defect growth prediction model based on the historical geometric size information and the current geometric size information; and using the defect growth prediction model to predict the defect growth prediction result of the target blade in a future preset time period; wherein, the defect identification model is trained based on blade samples labeled with defect locations and defect types.

[0160] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements a method for predicting the defect growth of gas turbine blades provided by the methods described above. The method includes: inputting video data of the gas turbine into a defect identification model; obtaining the defect identification result of each blade of the gas turbine output by the defect identification model; and determining the target blade with defects based on the defect identification result; generating three-dimensional point cloud data of the target blade based on the video data; and determining the current geometric dimension information of the defect area of ​​the target blade based on the three-dimensional point cloud data; obtaining historical geometric dimension information of the defect area of ​​the target blade; constructing a defect growth prediction model based on the historical geometric dimension information and the current geometric dimension information; and using the defect growth prediction model to predict the defect growth prediction result of the target blade within a future preset time period. The defect identification model is trained based on blade samples labeled with defect locations and defect types.

[0161] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0162] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0163] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting defect growth in gas turbine blades, characterized in that, include: The video data of the gas turbine is input into the defect identification model, and the defect identification results of each blade of the gas turbine are obtained from the output of the defect identification model. Leaf detection is performed on each frame of the video data to obtain the position information of each leaf in each frame of the image; The location information is input into the trajectory tracking model to obtain the running trajectory of each blade output by the trajectory tracking model; the blade rotation direction is determined based on the direction of the running trajectory, and the crossing line is determined based on the blade rotation direction; Based on the relative positional relationship between the running trajectory and the crossing line, the total number of blades of the gas turbine is determined; Identify a specific blade among all the blades of the gas turbine, the specific blade being marked with a preset identifier; The designated blade is a specific blade that serves as the positioning reference for the entire rotor blade; Based on the position of the specified blade in the gas turbine and the total number of blades, a mapping relationship between blade serial number and blade position is established, and the serial number of each blade in the gas turbine is determined based on the mapping relationship. This includes: taking the specified blade as the starting point, combining the blade rotation direction and the total number of blades, sequentially incrementing the number of the blades that pass through, establishing a one-to-one correspondence between physical position and logical serial number, and determining the unique serial number of each blade in the gas turbine. The defect identification result of each blade of the gas turbine is associated with the unique serial number of each blade; and the target blade with the defect is determined based on the defect identification result. Based on the video data, generate three-dimensional point cloud data of the target blade, and determine the current geometric dimensions of the defect area of ​​the target blade based on the three-dimensional point cloud data; Obtain historical geometric dimension information of the defect region of the target blade, and construct a defect growth prediction model based on the historical geometric dimension information and the current geometric dimension information; The defect growth prediction model is used to predict the defect growth of the target blade within a preset time period in the future. The defect identification model is trained based on blade samples labeled with defect locations and defect types.

2. The method for predicting defect growth in gas turbine blades according to claim 1, characterized in that, The step of generating three-dimensional point cloud data of the target blade based on the video data, and determining the current geometric dimensions of the defect region of the target blade based on the three-dimensional point cloud data, includes: The target blade is captured from the video data in a dual-view image, and the dual-view image is subjected to distortion correction and stereo correction based on the calibration parameters of the binocular camera to obtain a dual-view corrected image. Stereo matching is performed on the dual-view corrected image to obtain a pixel-level disparity map, and a depth map of the target blade and the three-dimensional point cloud data are generated based on the pixel-level disparity map and the calibration parameters. Based on the three-dimensional point cloud data, the defect region is extracted to obtain the current geometric size information of the defect region of the target blade.

3. The method for predicting defect growth in gas turbine blades according to claim 1, characterized in that, The defect growth prediction model constructed based on the historical geometric dimension information and the current geometric dimension information includes: Based on the historical geometric dimension information and the current geometric dimension information, a time series of defect dimensions is constructed; The defect size time series is fitted with linear regression to obtain the defect growth prediction model.

4. The method for predicting defect growth in gas turbine blades according to claim 1, characterized in that, After associating the defect identification result of each blade of the gas turbine with the serial number of each blade, the method further includes: The serial number is used as an index to construct an information database; The static attribute information of the blade is stored in the information database, and the static attribute information includes one or more of the following: manufacturer, model and material information; The records of changes in the installation position of the blade at different times are stored in the information database; The operation and maintenance records of the blades are stored in the information database; The defect identification results of the blades are stored in the information database.

5. The method for predicting defect growth in gas turbine blades according to claim 1, characterized in that, After predicting the defect growth of the target blade within a predetermined time period using the defect growth prediction model, the method further includes: If the defect growth prediction result exceeds a preset safety threshold and / or the defect growth rate of the target blade exceeds a preset rate threshold, a first warning message will be issued.

6. The method for predicting defect growth in gas turbine blades according to claim 1, characterized in that, After obtaining the defect identification results for each blade of the gas turbine output by the defect identification model, the method further includes: Based on the defect identification results and the current geometric dimensions of the defect area of ​​the target blade, the defect level of the target blade is determined; A health assessment of the target blade is performed based on the defect level of the target blade to obtain the health assessment result of the target blade; If any of the health assessment results do not meet the preset assessment conditions, a second warning message will be issued.

7. The method for predicting defect growth in gas turbine blades according to claim 6, characterized in that, After determining the defect level of the target blade, the method further includes: Obtain manual correction instructions, which include feedback on modifications to the defect level, defect type, and defect location; The defect identification result is updated based on the manual correction instruction; The updated defect identification results and their corresponding blade images are stored in the training sample library. The training sample library is subjected to format validation and outlier filtering to obtain valid sample data; Monitor the increment of the valid sample data. If the increment exceeds a preset increment threshold, use the valid sample data to train and update the defect identification model.

8. A defect growth prediction device for gas turbine blades, characterized in that, include: The defect identification module is used to input video data of the gas turbine into the defect identification model and obtain the defect identification results of each blade of the gas turbine output by the defect identification model; Leaf detection is performed on each frame of the video data to obtain the position information of each leaf in each frame of the image; The location information is input into the trajectory tracking model to obtain the running trajectory of each blade output by the trajectory tracking model; the blade rotation direction is determined based on the direction of the running trajectory, and the crossing line is determined based on the blade rotation direction; Based on the relative positional relationship between the running trajectory and the crossing line, the total number of blades of the gas turbine is determined; Identify a specific blade among all the blades of the gas turbine, the specific blade being marked with a preset identifier; The designated blade is a specific blade that serves as the positioning reference for the entire rotor blade; Based on the position of the specified blade in the gas turbine and the total number of blades, a mapping relationship between blade serial number and blade position is established, and the serial number of each blade in the gas turbine is determined based on the mapping relationship. This includes: taking the specified blade as the starting point, combining the blade rotation direction and the total number of blades, sequentially incrementing the number of the blades that pass through, establishing a one-to-one correspondence between physical position and logical serial number, and determining the unique serial number of each blade in the gas turbine. The defect identification result of each blade of the gas turbine is associated with the unique serial number of each blade; and the target blade with the defect is determined based on the defect identification result. The defect region determination module is used to generate three-dimensional point cloud data of the target blade based on the video data, and determine the current geometric size information of the defect region of the target blade based on the three-dimensional point cloud data; The prediction model building module is used to obtain historical geometric dimension information of the defect area of ​​the target blade, and to build a defect growth prediction model based on the historical geometric dimension information and the current geometric dimension information. The defect growth prediction module is used to predict the defect growth of the target blade within a preset time period using the defect growth prediction model. The defect identification model is trained based on blade samples labeled with defect locations and defect types.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the defect growth prediction method for gas turbine blades as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the defect growth prediction method for gas turbine blades as described in any one of claims 1 to 7.

Images