Patent Application
20230408451
The patent document claims the priority and benefits of Korea Patent Application No. 10-2022-0075107, filed on Jun. 20, 2022, which is incorporated by reference in its entirety as part of the disclosure of this patent document.
Various embodiments of the present disclosure relate to an ultrasonic flaw-detection system and an ultrasonic flaw-detection method.
Recently, the application of high-tech new materials in the field of wind power generation system has gained popularity as a promising industry in new renewable energy fields. This composite material industry has a wide range of applications and exerts a significant influence on other fields of industry through leading technical innovations. Also, acquiring technical superiority in the composite material industry can be quite challenging.
In particular, the weight of a rotating blade itself is a key factor that significantly determines the efficiency of costly power generation facilities. Therefore, the industry has shifted from predominantly employing glass fiber composites in the past to manufacturing wind turbine blades using carbon fiber composites. The carbon fiber composite is expensive from a price standpoint due to the trend of larger size but can be about 40% lighter in weight than the glass fiber.
The lightweight nature of the carbon fiber composites can serve as a critical stepping stone that enables the realization of larger a generator capacity. As a result, the overall market for a wind turbine blade is expected to increase four to five times over the next ten years.
Meanwhile, a composite material blade may exhibit various internal damages such as debonding, delamination, and cracks, etc., due to complexities associated with materials and manufacturing methods. These damages are often challenging to detect visually.
These defects initially may be in a small size that does not affect structurally. However, as a crack propagates/progresses while being subjected to repeated load and impact load, the crack may become large enough to affect the structural safety of the blade over time, and thus, eventually, a big accident may occur. Therefore, if detection is carried out at an early fault stage and repair is made according to instructions, it is possible to prevent major accidents in advance and to reduce economic damage caused such accidents.
An ultrasonic flaw-detection method in which a detection target is scanned by using an ultrasonic wave and an inspector determines whether or not there is a defect in the detection target based on the obtained scan data is being mainly used as a blade structural safety detection method. In such a detection method, detection results may vary depending on inspector's skillfulness and subjective opinions, so that it is difficult to obtain the continuity and objectivity of the detection result.
In order to overcome the above-mentioned problems, there is a requirement for a detection method for minimizing a detection error by obtaining the objectivity of a result of an ultrasonic flaw-detection and by preventing human errors.
The purpose of the present disclosure is to provide an ultrasonic flaw-detection system capable of training artificial intelligence by using data determined to be defective by experienced inspectors and of objectively determining defects of a detection target by using the trained artificial intelligence.
The technical problem to be overcome by the present invention is not limited to the above-mentioned technical problems. Other technical problems not mentioned can be clearly understood from the embodiments of the present invention by a person having ordinary skill in the art.
One embodiment is an ultrasonic flaw-detection system including: an ultrasonic flaw-detection device configured to transmit an ultrasonic wave to a detection target, collect an ultrasonic echo signal reflected from the detection target, and then generate a signal data; a signal data preprocessor configured to preprocesses the signal data; a defect candidate group selection unit configured to select a defect candidate group based on the preprocessed signal data and generate defect candidate signal data based on the selection; an image data generator configured to generate image data based on the defect candidate signal data included in the defect candidate group; and a defect determination unit configured to determine whether there is a defect in the defect candidate group based on the image data.
The signal data preprocessor may remove noise from the signal data, may extract poles from the signal data, and may divide the signal data into a plurality of clusters having a certain size based on the pole.
The defect candidate group selection unit may determine whether a defect is included in the signal data belonging to each cluster based on a deep learning algorithm that uses each of the plurality of clusters as an input, and may select the cluster determined to include a defect as the defect candidate group.
The deep learning algorithm may be a variational auto encoder (VAE) or a residual neural network (ResNet).
The image data generator may generate a B-Scan image data and a C-Scan image data on the detection target, based on the signal data.
The image data generator may generate the B-Scan image data and the C-Scan image data on an area in which the defect candidate group is included in the detection target, based on the signal data included in the defect candidate group.
The defect determination unit may determine whether each of the defect candidate groups has a defect based on a deep learning algorithm using the image data as an input.
The deep learning algorithm may be a you only look once (YOLO) algorithm or a Faster R-CNN algorithm.
The defect determination unit may output whether there is a defect for each of the defect candidate groups and may output, when there is a defect, a bounding box that surrounds the corresponding defect.
Another embodiment is an ultrasonic flaw-detection method by the ultrasonic flaw-detection system. The ultrasonic flaw-detection method includes: transmitting an ultrasonic wave to a detection target, collecting an ultrasonic echo wave reflected from the detection target, and then generating a signal data; preprocessing the signal data; selecting a defect candidate group based on the preprocessed signal data and generate defect candidate signal data based on the selection; generating image data based on the defect candidate signal data included in the defect candidate group; and determining whether there is a defect in the defect candidate group based on the image data.
The features, advantages and method for accomplishing the present invention will be more apparent from referring to the following detailed embodiments described as well as the accompanying drawings. However, the present invention is not limited to the embodiment to be disclosed below and may be implemented in various different forms. While the embodiments bring about the complete disclosure of the present invention and are provided to make those skilled in the art fully understand the scope of the present invention, the present invention is just defined by the scope of the appended claims. The same reference numerals throughout the disclosure correspond to the same elements.
When one component is referred to as being “connected to” or “coupled to” another component, the one component may be directly connected or coupled to the another component. However, the one component may be indirectly connected to the another component and there may be an intervening component interposed between and connecting them. Meanwhile, what one component is referred to as being “directly connected to” or “directly coupled to” another component indicates that another component is not interposed between them. The term “and/or” includes each of the mentioned items and any combination of the mentioned items thereof.
Terms used in the present specification are provided for description of only specific embodiments of the present invention, and not intended to be limiting. In the present specification, an expression of a singular form includes the expression of plural form thereof unless specifically stated otherwise. In the present disclosure, the terms “comprises”, “comprising”, and the like may indicate the presence of features, steps, operations, elements, and/or components, but do not preclude addition of one or more other functions, steps, operations, elements, components, and/or combinations thereof.
While terms such as the first and the second, etc., can be used to describe various components, the components are not limited by the terms mentioned above. The terms are used only for distinguishing one component from the other.
Therefore, the first component to be described below may be the second component within the spirit of the present invention. Unless differently defined, all terms used herein including technical and scientific terms have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention belongs. In addition, terms defined in dictionaries generally used should be construed to have meanings matching contextual meanings in the related art.
A term “part” or “module” used in the embodiments may mean software components or hardware components such as a field programmable gate array (FPGA), an application specific integrated circuit (ASIC). The “part” or “module” may perform certain functions. However, the “part” or “module” is not meant to be limited to software or hardware. The “part” or “module” may be configured to be placed in an addressable storage medium or to restore one or more processors. Thus, for one example, the “part” or “module” may include components such as software components, object-oriented software components, class components, and task components, and may include processes, functions, attributes, procedures, subroutines, segments of a program code, drivers, firmware, microcode, circuits, data, databases, data structures, tables, arrays, and variables. Components and functions provided in the “part” or “module” may be combined with a smaller number of components and “parts” or “modules” or may be further divided into additional components and “parts” or “modules”.
Methods or algorithm steps described relative to some embodiments of the present invention may be directly implemented by hardware and software modules that are executed by a processor or may be directly implemented by a combination thereof. The software module may be resident on a RAM, a flash memory, a ROM, an EPROM, an EEPROM, a resistor, a hard disk, a removable disk, a CD-ROM, or any other type of record medium known to those skilled in the art. An exemplary record medium is coupled to a processor and the processor can read information from the record medium and can record the information in a storage medium. In another way, the record medium may be integrally formed with the processor. The processor and the record medium may be resident within an application specific integrated circuit (ASIC).
Referring to
The ultrasonic flaw-detection device 10 may transmit an ultrasonic signal to a detection target through a probe, may detect the returned ultrasonic signal, and then may generate signal data. The returned ultrasonic signal may be referred to as an echo signal. The generated signal data may be transmitted to the signal data preprocessor 20 and/or the image data generator 40.
According to an embodiment, the ultrasonic flaw-detection device 10 may use a longitudinal ultrasonic wave and may use a pulse reflection method in which ultrasonic pulses having a short duration are repeatedly generated and reflected pulse signals are analyzed.
As shown in
Here, the signal data can also be referred to as A-Scan data. In the signal data, the magnitude of the ultrasonic signal reflected from the detection target can be represented by the amplitude of the ultrasonic signal. The amplitude of the ultrasonic signal may change over time. This can be represented through a graph displaying a signal waveform. the signal waveform may be displayed by a display of the ultrasonic flaw-detection system 100. In the graph, the X axis represents the progression of time and the Y-axis represents the magnitude of the amplitude of the reflected ultrasound signal.
According to an embodiment of the present disclosure, a deep learning algorithm may be applied in order to determine the presence of a defect in the detection target. In order to apply deep learning algorithms, it is often necessary to process different data into constant input data. This processing can be performed by the signal data preprocessor 20.
The signal data received from the ultrasonic flaw-detection device 10 may include noise data. The signal data preprocessor 20 may first remove noise data from the signal data. According to an embodiment, in order to select a defect candidate group from the signal data, it may be necessary to extract signal poles from the signal data. Once the noise data is removed from the signal data, the signal poles can be more clearly and accurately extracted from the signal data. The signal data, from which the noise data is removed by the operation of the signal data preprocessor 200 may be referred to as noise-processed signal data. According to an embodiment, the signal data preprocessor 20 may remove noise data from the signal data by using Wavelet Denoising, which is one of Python libraries. The Wavelet Denoising can remove noise data present in the signal data through a process of decomposing the wavelet, calculating a universal threshold, and then reconstructing by using a threshold coefficient.
The signal data preprocessor 20 may detect peaks, each having an amplitude greater than a threshold value from the noise-processed signal data after the noise data is removed from the signal data. Then, the signal data preprocessor 20 may perform a pole extraction function that automatically extracts poles by controlling the minimum distance parameter between each of the detected peaks.
When a surface of the detection target is coated or an inner end portion of the detection target is bonded, those portions of the detection target may be detected or considered as discontinuous portions. Large signals may be reflected from those discontinuous portions. A pole may be formed in a discontinuous portion. In particular, the above-described discontinuous portion may have a higher peak than other poles. Then, in the signal data, a discontinuous portion by coating and discontinuous portion by bonding may be found in the signal data, and clustering may be performed to generate data having the corresponding portion as start and end points. Then, one cluster may include information on the detection target in a thickness direction at one point. According to an embodiment, the clustering may be performed such that a start point and an end point of a cluster are determined based on a discontinuous portion by coating and/or a discontinuous portion by bonding.
Finally, the signal data preprocessor 20 may adjust each cluster to have the same size. According to the embodiment, the number of data that each cluster has may be adjusted to be the same. According to an embodiment, the signal data preprocessor 20 may perform clustering or adjust clusters such that portions of the detection target represented by each of the clusters are in the same size.
The signal data preprocessor 20 may transmit the preprocessed signal data to the defect candidate group selection unit 30. Signal data transmitted from the signal data preprocessor 20 to the defect candidate group selection unit 30 may be referred to as a preprocessed signal data.
The defect candidate group selection unit 30 may select a defect candidate group present in the detection target based on the signal data provided from the ultrasonic flaw-detection device 10. Here, the defect candidate group may indicate a signal data or a location within the detection target, which is likely to have a defect. According to the embodiment, when the signal data is divided into a plurality of clu sters by the above-described signal data preprocessor 20, the defect candidate group selection unit 30 may determine whether a defect exists in each cluster, and the defect candidate group may be a set of clusters in which defects may exist. According to an embodiment, the defect candidate selection unit 30 may determine whether or not each cluster from the plurality of clusters is likely to have a defect therein and generate a defect candidate group including clusters found likely to have a defect therein. Signal data representing, corresponding to or included in the defect candidate group may be referred to as defect candidate signal data.
According to the embodiment, the defect candidate group selection unit 30 may select or determine or generate the defect candidate group based on an artificial intelligence or deep learning algorithm by using the signal data preprocessed by the signal data preprocessor 20 as an input.
According to an embodiment, defect candidate group selection unit 30 may use, as an artificial intelligence algorithm, a variational auto encoder (VAE) or a residual neural network (ResNet) that is a multilayer neural network classification algorithm.
The VAE may be an unsupervised anomaly detection algorithm. The VAE may create a probability distribution by reducing the dimension of the input signal data, and perform an operation of generating data again by performing sampling through the created probability distribution. Since the VAE transmits data while learning a transfer function together, various and similar result values can be generated. Here, if a sliding window concept is applied, time series data analysis may be possible. Here, the size of the window may be the same as the size of one cluster described above, and may be designated as 500 according to an embodiment.
According to the embodiment, the VAE may perform learning based on unsupervised learning. The unsupervised learning may be a method of learning only with actual signal data without including a result value indicating whether the defect candidate group exists.
The RestNet may be a type of network in which the number of hidden nodes is further increased by adding convolutional layers based on a VGG-19 artificial intelligence algorithm. In Restnet, the neural network is further deepened, and a shortcut connection 310 between nodes is added.
The structure of the ResNet can solve problems of existing artificial intelligence networks or deep learning networks such as vanishing gradient or overfitting. The greater the depth of the layer, the more learning efficiency increases. The ResNet may be named in the sense of learning a residual F(x) which is a difference between an output value H(x) and an input value x. The ResNet may intend to make the residual F(x) zero through learning.
The ResNet may be trained based on supervised learning, and the supervised learning may train a neural network while notifying learning data together with defect candidate group information included in the learning data.
Referring back to
The image data generator 40 may generate image data based on the signal data of the defect candidate group received from the defect candidate group selection unit 30. The generated image data may be a B-Scan image and/or a C-Scan image used in the ultrasonic flaw-detection. A method for generating such image data may adopt a method for generating B-Scan and/or C-Scan images. B-scan is a two-dimensional cross-sectional imaging technique that provides a vertical representation of a detection target. C-scan is a planar imaging technique that provides a two-dimensional representation of a surface in a certain depth of the detection target. According to an embodiment, the image data generator 40 may generate the image data for the entire detection target, however, according to an embodiment, it may be sufficient to generate image data only for portions corresponding to the defect candidate group received from the defect candidate group selection unit 30.
Referring to
Referring to
According to an embodiment, the ultrasonic flaw-detection system 100 may further include a display. The display may be configued to display the B-Scan image 410 and/or the S-Scan image 420 generated by the image data generator 40.
Referring back to
The defect determination unit 50 may finally determine whether there is a defect in the defect candidate group of the detection target based on the images provided from the image data generator 40.
According to the embodiment, the defect determining unit 50 may finally determine the defect based on an artificial intelligence or deep learning algorithm using the image data provided from the image data generating unit 40 as an input.
According to the embodiment, the defect determination unit 50 may use, as an artificial intelligence algorithm, a you only look once (YOLO) algorithm or a Faster R-CNN algorithm which is effective in distinguishing objects within an image.
The YOLO algorithm extracts features from one image and simultaneously creates a bounding box and divides classes, so that defects can be quickly determined.
The Faster R-CNN algorithm may be a two-stage object detection algorithm, comprising a first step and a second step. The first step is detecting an object by extracting features from an image, and the second step is calculating a defect probability of the detected object and box coordinates of the object, and finally of reading whether or not there is a defect.
According to the embodiment, in the first step of the Faster R-CNN, features are extracted and objects are detected based on the ResNet 510. The second step of the Faster R-CNN is performed with a region proposal network (RPN) 520 and a region of interest (RoI) pooling 530. Thereby, the defect determining unit 50, using the Faster R-CNN, may determine a probability that a calculated object candidate region belongs to a defect class or a background noise class. The defect determination unit 50 may determine the class of each object (whether it is a defect class or a background noise class) based on a threshold value, that is, reading the defect.
The defect determination unit 50 may output the bounding box as a result of the detection and reading process. The bounding box may be displayed on the images provided to the defect determination unit 50. The bounding box may indicate a defect portion on the lower surface.
The defect determination unit 50 may detect and read the defect based on the deep learning algorithm such as the Faster R-CNN. The effective learning may be required in order to improve the performance of such a deep learning algorithm. In order to train the deep learning algorithm employed by the defect determination unit 50, the defect determination unit 50 may perform data labeling (annotation) to include information on the defect portion in the image data. That is, the defect determination unit 50 employs a deep learning algorithm that utilizes supervised learning. This algorithm is trained using image data that has been reprocessed to include information about the location of the defect in the image data that includes the defect as an object. Through such learning, the performance of the deep learning algorithm employed by the defect determination unit 50 can be improved.
As described above, the ultrasonic flaw-detection system 100 proposed in the present disclosure is capable of determining whether the detection target has a defect or not. It achieves this by selecting a defect candidate group based on the signal data and subsequently determining the presence of a defect based on the image data obtained from the selected defect candidate group. Furthermore, the ultrasonic flaw-detection system 100 proposed in the present disclosure has the capability to enhance its performance by employing artificial intelligence for defect detection.
Referring to
According to various embodiments of the present disclosure, in step S10, the ultrasonic flaw-detection system 100 utilizes the ultrasonic flaw-detection device 10 to emit ultrasonic waves toward the detection target 200. It then detects and captures the signal reflected from the detection target 200. According to the embodiment, the ultrasonic flaw-detection device 10 may generate ultrasonic pulses having a short duration in a repeated manner. These pulses are emitted towards the detection target 200, and the ultrasonic flaw-detection device 10 may obtain the signal data by detecting the pulse signal reflected from the detection target 200.
In step S20, the ultrasonic flaw-detection system 100 may preprocess the signal data obtained in step S10. The preprocessing operation may include removing noise included in the signal data, extracting poles from the signal data with noise removed, and dividing the signal data into a plurality of clusters based on the extracted poles. Here, the plurality of clusters may be adjusted to have the same size so that each of clusters having the same size become input data to be input to the deep learning algorithm in the next step.
In step S30, the ultrasonic flaw-detection system 100 may obtain a defect candidate group based on the preprocessed signal data. According to the embodiment, the ultrasonic flaw-detection system 100 may determine whether a defect is likely to exist in each of the plurality of clusters generated in step S20. The ultrasonic flaw-detection system 100 may generate the defect candidate group such that the defect candidate group includes clusters determined to be likely to have a defect. Accordingly, the defect candidate group may be a set of clusters in which defects may exist.
According to the embodiment, the ultrasonic flaw-detection system 100 may select or determine or generate the defect candidate group based on an artificial intelligence or deep learning algorithm using the signal data preprocessed in step S20 as an input. The ultrasonic flaw-detection system 100 may use, as an artificial intelligence algorithm, a variational auto encoder (VAE) or a residual neural network (ResNet) that is a multilayer neural network classification algorithm. The deep learning algorithm that the ultrasonic flaw-detection system 100 can use is not limited thereto, and it is also possible for the ultrasonic flaw-detection system 100 to use other deep learning algorithms.
The deep learning algorithm utilized by the ultrasonic flaw-detection system 100 to obtain the defect candidate group can be optimized through prior learning before its actual application. Here, learning can be performed by using the signal data selected as having defects by experts.
As a result of step S30, the ultrasonic flaw-detection system 100 may obtain the defect candidate group determined to likely or potentially have defects. Here, the defect candidate group may be a set of clusters expected to include or likely to have a defect.
In step S40, the ultrasonic flaw-detection system 100 may generate image data based on the defect candidate group obtained in step S30. The image data may be data for an image that has been conventionally referred to as B-Scan and/or C-Scan used in the ultrasonic flaw-detection. A method for generating such image data may adopt a method for generating conventional B-Scan and/or C-Scan. According to the embodiment, it is not necessary to generate the image data for the entire detection target, and it may be sufficient to generate the image data only for portions related to the signal data selected as having defects in step S30. In other words, it may be sufficient to generate image data only for portions corresponding to the defect candidate group received from the defect candidate group selection unit 30.
In step S50, the ultrasonic flaw-detection system 100 may determine a defect based on the generated image data. According to the embodiment, the ultrasonic flaw-detection system 100 may finally determine the defect based on an artificial intelligence or deep learning algorithm using the image data generated in step S40 as an input.
According to the embodiment, the ultrasonic flaw-detection system 100 may use, as a deep learning algorithm, a you only look once (YOLO) algorithm or a Faster R-CNN algorithm which is effective in distinguishing objects within an image.
The ultrasonic flaw-detection system 100 may provide information indicating the object obtained as a result of step S50. The ultrasonic flaw-detection system 100 may determine whether the class of the object is a defect class or background noise class. The ultrasonic flaw-detection system 100 may provide a bounding box. If the class of the object is determined to be a defect, the bounding box may mean a rectangular box surrounding the object. The bounding box may be in any geometry shape, such as a circle, oval, rhombus, which can encircle or identify the object, identified as a defect.
Although the present invention has been described with reference to the embodiment shown in the drawings, this is just an example and it will be understood by those skilled in the art that various modifications and equivalent thereto may be made. Therefore, the true technical scope of the present invention should be determined by the spirit of the appended claims. Also, it is noted that any one feature of an embodiment of the present disclosure described in the specification may be applied to another embodiment of the present disclosure.
According to the embodiments of the present disclosure, the ultrasonic flaw-detection is performed by using a trained artificial intelligence, so that the objectivity of a detection result can be obtained and human errors can be prevented.
According to the embodiments of the present disclosure, through an ensemble model to which a signal model and an image model are applied together, defect extraction performance can be improved during automatic evaluation and the reliability of defect analysis results can be enhanced.
Advantageous effects that can be obtained from the present disclosure are not limited to the above-mentioned effects. Further, other unmentioned effects can be clearly understood from the following descriptions by those skilled in the art to which the present disclosure belongs.
| Number | Date | Country | Kind |
|---|---|---|---|
| 10-2022-0075107 | Jun 2022 | KR | national |
