阅读说明：本技术 涡轮发动机叶片缺陷自动检测系统的训练方法 (Training method of automatic detection system for blade defects of turbine engine ) 是由 赫琳娜·沃罗比耶娃 西尔文·皮卡德 于 2020-04-03 设计创作，主要内容包括：一种用于训练用于自动检测涡轮发动机叶片(2)缺陷的系统(1)的方法,所述系统包括用于接收叶片的平台(3)、至少一个图像采集装置(4)、发光强度可以变化的至少一个光源(5)、至少一个移动装置(6),所述移动装置被配置为相对于所述至少一个图像采集装置(4)移动所述平台(3)或相对于所述平台(3)移动所述至少一个图像采集装置(4),以及处理单元(7),所述处理单元被配置为接收所采集的图像,并将其与以下信息项相关联,所述信息项与所述至少一个光源(5)的发光强度和在图像采集期间所述平台(3)相对于所述至少一个图像采集装置(4)的相对位置有关。(A method for training a system (1) for automatically detecting defects of a turbine engine blade (2), the system comprises a platform (3) for receiving the blade, at least one image acquisition device (4), at least one light source (5) with variable luminous intensity, at least one moving device (6), the moving device is configured to move the platform (3) relative to the at least one image acquisition device (4) or to move the at least one image acquisition device (4) relative to the platform (3), and a processing unit (7) configured to receive the acquired image, and to associate it with an item of information relating to the luminous intensity of said at least one light source (5) and the relative position of said platform (3) with respect to said at least one image acquisition device (4) during image acquisition.)

1. A method for training a system (1) for automatically detecting defects of a turbine engine blade (2), the system comprising a platform (3) for receiving a blade, at least one image acquisition device (4), at least one light source (5) whose luminous intensity can be varied, at least one moving device (6) configured to move the platform (3) with respect to the at least one image acquisition device (4) or to move the at least one image acquisition device (4) with respect to the platform (3), and a processing unit (7) configured to receive an acquired image and to associate it with an item of information relating to the luminous intensity of the at least one light source (5) and to the relative position of the platform (3) with respect to the at least one image acquisition device (4) during image acquisition, for each blade (2) to be studied in a group of blades having surface defects that have been identified, the training method comprises the following steps:

-defining a usable area (200) on the blade,

-for each individual relative position of the platform with respect to the at least one image capturing device, producing (210) a plurality of captured images of the usable area of the blade, each captured image of one and the same relative position being generated with a different luminous intensity,

-dividing (220) each image of the available region from one acquired image into a plurality of images of sub-regions,

-determining (230) pairs of pixel intensity boundaries, each pair comprising a minimum intensity and a maximum intensity,

-for each pair of borders:

a. for each sub-region of the available region, determining (241) the number of pixels contained in each image of said sub-region, the intensity of said pixels being between the minimum and maximum intensity of said pair of borders, and then

b. For each given relative position, for each sub-region of said relative position, selecting (242) a single image, the selected image corresponding to the image containing the most pixels whose intensity is between the minimum intensity and the maximum intensity of said pair of borders, and then

c. For each given relative position, for each sub-region, examining (243) each respective selected image of said sub-region by means of a classification convolutional neural network to detect the presence or absence of any defects, and then

d. Examining (244) the detections made by the classification convolutional neural network to determine whether the detected defect is true or false, an

e. Calculating (245) a detection efficiency for a given pair of boundaries within the usable area on the basis of all given relative positions, said detection efficiency depending on a ratio of a number of proven detections to a number of false detections,

-selecting (250) a pair of boundaries having the best detection efficiency.

2. Method according to the preceding claim, wherein if in step b at least two images comprise the same number of pixels, the image of the selected sub-area is the image whose mean value of the pixel intensities is closest to an intensity of 127.5.

3. The method according to any one of the preceding claims, further comprising generating additional images based on the acquired images, the additional images being usable in addition to the intensity images.

4. The method according to any of the preceding claims, wherein the different intensities used are the same for each individual relative position of the platform with respect to the at least one image acquisition device.

5. A method according to any of claims 1 to 4, wherein as many sub-areas without defects as sub-areas with defects are selected from all sub-areas of the available area of the blade under investigation.

Technical Field

The present invention relates to non-destructive testing of aerospace components, such as turbine blades.

Background

Inspection of turbine engine blades to detect surface defects, such as small dents, roughness, excess material, or lack of material in some places, is typically done by an operator customer.

One possible way of performing this inspection is to place the leaves to be inspected on a support, illuminate them and take a picture with good resolution. Based on these photographs, it is then necessary to determine whether there are defects on the blade surface. Grayscale photographs are sufficient to detect the aforementioned defect types.

However, several problems arise in taking photographs that can be used to obtain grayscale photographs that can be used to compare and detect any defects on the blade surface.

First, the blade being inspected typically has a specular and textured coating. Thus, if the blade is illuminated with sufficient intensity to make certain coating details visible on a portion of the blade, there is a risk that other portions of the blade will be overexposed and corresponding portions on the obtained image will not be usable. Furthermore, the relief formed by the coating texture makes it difficult or even impossible to determine the optimal illumination (nearby pixels in the image will have different behaviors).

Second, some blades (e.g., turbine blades) have complex curvatures. Thus, if the turbine blade is illuminated, certain parts of the blade will be in shadow, which will cause corresponding parts on the obtained image to become unusable. Furthermore, it is not possible to use a grazing technique, considering that the curvature of the blade is present in all parts of the blade.

Third, the coating of the blades may be slightly different. This means that with the same illumination and the same position facing the image acquisition means, the leaves will appear on the image in a different way, possibly too brightly illuminated at the same place on the image, or not sufficiently illuminated.

Fourth, even with an accurate positioning system, the accuracy never reaches the total accuracy. This results in that the images of different blades acquired at theoretically the same position facing the image acquisition device are not exactly the same, resulting in noise in the positioning. This offset amplifies the aforementioned problem.

Furthermore, blade images taken with sufficient resolution to see defects are typically grayscale images, i.e. pixels take integer values between 0 and 255 and have a rather chaotic blade texture, a non-uniform surface appearance, which means that the available area contains many different pixel values, ranging from 0 to 255. The term "usable area" refers to an area that is not obscured, and is illuminated in such a way that the defect is visible.

It is rather easy to delineate a usable area on each captured image, each corresponding to a given pose of the blade facing the image capturing device, and the usable area is an area that is not blurred and the illumination can be adjusted in the correct way, i.e. without overexposure or too darkness. However, for a given photograph, the optimal illumination will not be the same for all available areas due to blade geometry and specular reflection. Furthermore, as described above, the optimal illumination for a given area and given pose is not necessarily the same from one blade to another. Therefore, the illumination of the available area is not a parameter that can be set manually.

A method of inspecting an object for surface defects is known in which photographs of the object are taken under one or different illuminations and then one or more images that may be defective are selected. The image is examined globally to determine if it has very bright areas or is in shadow (these areas are then features of a defect) and in this way see the predefined defect features. The image may be compared to a reference image. In this method, all images are subjected to a first processing step, regardless of their lighting conditions, and a first filter is used to eliminate images that do not contain any defects. The next additional step is to iteratively eliminate more and more images.

However, processing the images continuously to retain only the image with the best illumination for the defect detection task before any filtering of the defects represents a considerable time cost in the overall process of defect detection. In particular, the result is a longer calculation time for the defect detection task.

Furthermore, each time the known method processes the image completely.

A method for inspecting an object having a curvature is also known. In this method, two light sources are used to produce light that faces the object, referred to as "bright field illumination", and light that sweeps across the object, referred to as "dark field illumination". The light sources may be movable relative to the object. An inspection is then performed to look for defects based on the images acquired from these two viewpoints.

However, for objects with complex curvatures, varying continuously as much along the horizontal and vertical directions, it would be necessary to increase the different positions of the light sources to be at right angles to each small area.

In a manner quite similar to the previous method, a method is known in which two light sources positioned at precise angles with respect to the surface to be inspected are moved along a scan line to capture sub-images and reconstruct the final image to be inspected. In this method, it is therefore not necessary to use different light intensities. The luminous intensity used is already optimal because the angle of incidence is optimal.

However, this approach would be ineffective for complex object curvatures because an excessive number of different scan lines would need to be used.

Disclosure of Invention

The present invention is directed to an automated method of selecting optimal lighting within a given area for automatically controlling the surface of a turbine engine blade through a learning method, and more particularly to a method for training a system for automatically detecting defects in a turbine engine blade, the system including a platform for receiving the blade.

According to the subject matter of the present invention, a method is provided for training a system for the automatic detection of defects of blades of a turbine engine, the system comprising a platform for receiving the blades, at least one image acquisition device, at least one light source whose luminous intensity can be varied, at least one movement device configured to move the platform with respect to the at least one image acquisition device or to move the at least one image acquisition device with respect to the platform, and a processing unit configured to receive the acquired images and associate them with the following items of information: the information item relates to the luminous intensity of the at least one light source and the relative position of the platform with respect to the at least one image acquisition device during image acquisition.

According to a general feature, the training method comprises, for each blade to be studied of a group of blades having surface defects already identified, the following steps:

-defining a usable area on the blade to be studied,

-for each individual relative position of the platform with respect to the at least one image capturing device, producing a plurality of captured images of the usable area of the blade, each captured image of one and the same relative position being generated with a different luminous intensity,

-dividing each image of the available area from one acquired image into a plurality of images of sub-areas,

-determining a plurality of pairs of pixel intensity boundaries, each pair comprising a minimum intensity and a maximum intensity,

-for each pair of borders:

a. determining the number of pixels contained in each image of the sub-region, the intensity of which is between the minimum intensity and the maximum intensity of the pair of borders, and then

b. Selecting a single image for each sub-region at a given relative position, the selected image corresponding to the image containing the most pixels with an intensity between said minimum and maximum intensities of the pair of borders, and then

c. Examining each sub-region of a given relative position by classifying a convolutional neural network to detect the presence or absence of any defects, and then

d. Examining detections made by the classified convolutional neural network to determine whether the detected defect is true or false, an

e. Calculating the detection efficiency of a given pair of boundaries within the usable area on the basis of all given relative positions, the detection efficiency depending on the ratio of the number of proven detections to the number of false detections,

-selecting a pair of boundaries having the best detection efficiency.

The method according to the invention comprises an efficient pre-processing enabling to select from images acquired at the same location at different luminous intensities sub-regions which will later be used for detecting anomalies. The preprocessing first includes configuring the sub-regions according to the number of pixels contained in the different minimum and maximum boundary pairs. During the adjustment phase, which is performed only once, the best boundary pair is selected in such a way as to provide the best possible performance for the detection algorithm to be used later in the processing line. In the production line phase, the optimal sub-area is selected directly with respect to the end set at the time of adjustment, which is fast and may not increase the number of sub-areas to be tested in the test.

The present invention thus provides a method for training an automatic inspection system so that the inspection method implemented by said system only keeps the image with the best illumination for the defect inspection task before any filtering of the defects, and therefore the calculation time of the defect inspection task can be saved.

Furthermore, the invention provides for processing regions of the image independently, not necessarily the entire image.

Furthermore, the method may avoid that the light source is repositioned at a certain angle relative to the object. An image is available as long as the brightness of a certain area is sufficient and one of the intensities is used.

By means of the invention, a large area can be illuminated at a single position of the light source, which makes it possible to machine blades with complex curvatures without increasing the number of different positions of the light source to find right angles with respect to each small area.

Thus, the selected pair of borders may be selected so as to be able to select a single image for each sub-area from different available images acquired at different intensities, in anticipation of future training steps or using a system for automatically detecting blade defects.

According to a first aspect of the training method, if in step b the at least two images comprise the same number of pixels, the image of the selected sub-area is the image whose average of the pixel intensities is closest to an intensity of 127.5.

According to a second aspect of the training method, the method may further comprise generating additional images based on the acquired images, which additional images may be used in addition to the intensity images.

According to a third aspect of the training method, the different intensities used are the same for each individual relative position of the platform with respect to the at least one image acquisition device.

According to a fourth aspect of the training method, as many non-defective sub-areas as there are defective sub-areas can be used, from all sub-areas of the available area of the blade under study.

Thus, some sub-regions may not be used so that the number of defective and non-defective sub-regions is the same. If there are fewer defective sub-areas, it is possible to randomly select the same number of non-defective sub-areas as defective sub-areas from all available non-defective sub-areas. If there are fewer defect-free sub-areas, it is possible to randomly select the same number of defective sub-areas as defect-free sub-areas from all available defective sub-areas. Other numbers of sub-regions may also be selected. This aspect of the training method can be done on a leaf-by-leaf basis, or all leaves together.

In another subject of the invention, a system for automatic detection of defects of a turbine engine blade is provided, the system comprises a platform for receiving the blade, at least one image acquisition device, at least one light source, which may vary in luminous intensity, at least one moving device configured to move the platform relative to the at least one image acquisition device or to move the at least one image acquisition device relative to the platform, and a processing unit, configured to receive the acquired image and associate it with an item of information relating to the luminous intensity of the at least one light source and the relative position of the platform with respect to the at least one image acquisition device during image acquisition, for each acquired image, a plurality of acquired maps are taken with the luminous intensity of the at least one light source for each of the relative positions.

According to a general feature of the invention, the processing unit is configured to define a usable area on the blade, divide the acquired image from the usable area into a plurality of sub-areas, and process each sub-area.

According to an aspect of the automatic detection system, for processing each sub-region, the processing unit may be configured to select only a single sub-region image from a plurality of additional images generated based on images from the acquisition, the selection being performed using a predefined pair of borders in accordance with the following steps:

a. determining the number of pixels contained in each image of the sub-region, the intensity of which is between the minimum intensity and the maximum intensity of the pair of borders, and then

b. A single image is selected for each sub-region at a given relative position, the selected image corresponding to the image containing the most pixels with an intensity between the minimum and maximum intensities of the pair of borders.

Drawings

FIG.1 is a schematic view of a system for automatically detecting turbine engine blade defects according to an embodiment of the present invention.

FIG.2 illustrates a flow diagram of a method for training the system of FIG.1 for automatically detecting defects in turbine engine blades, according to an embodiment of the present invention.

Detailed Description

FIG.1 schematically represents a system for automatically detecting defects in a turbine engine blade.

The system 1 for automatic detection of defects in a turbine engine blade 2 comprises a movable platform 3 on which the blade 2 to be studied is placed, a fixed camera 4 and a fixed light source 5 whose luminous intensity can be varied. The blade 2 rests vertically against the platform 3. In other words, the blade 2 comprises a blade root 22, a blade body 24 and a blade tip 26, the blade body 24 extending between the blade root 22 and the blade tip 26, the blade root 22 resting on the platform 3 in such a way that the blade body 24 extends in a direction orthogonal to the surface of the platform 3 on which the blade root 22 rests, said orthogonal direction corresponding to the vertical direction.

In the embodiment shown in fig.1, the platform 3 is movable and can be moved relative to the camera 4 and the light source 5, both of which are kept in fixed positions due to the moving means 6 comprised in the platform 3. In order to capture images around the blade 2, the platform 3 may be rotated about the rotation axis and moved vertically and horizontally.

The detection system 1 further comprises a processing unit 7 coupled to the platform 3, the camera 4 and the light source 5. The processing unit 7 is configured to receive the captured image sent by the camera 4 and to associate it with information relating to the luminous intensity of the light source 5 and the position of the platform during image capture.

FIG.2 schematically represents a flow chart of a training method implemented by a system for automatically detecting defects of the turbine engine blade of FIG.1 to optimize its operation during a defect search on an unknown blade.

In order to train the inspection system 1, a set of known blades is used, the defects of which are known and recorded in the processing unit 7.

For each leaf to be studied in the leaf set, the training method performs the following steps.

In a first step 200, the camera 4, the light source 5 and the platform 3 are adjusted to acquire an image such that the entire blade 2 can be covered by the available area. On each image thus acquired, a usable area is delimited.

In the next step 210, for each position of the platform 3, a plurality of images of each available area of the blade 2 is acquired using the camera 4, each image of one and the same position of the platform 3 being acquired with a different luminous intensity of the light source 5. The detection system 1 may be configured to acquire the same number of images for each position, for example twenty images for one and the same position of the platform 3, each acquired image having a different intensity. Thus, for each position of the platform 3, there are twenty or so acquired images of the usable area of the blade 2 under investigation, each image having a different intensity, namely twenty different luminous intensities.

In a following step 220, each image acquired by the camera 4 is sent to a processing unit which divides the acquired image into images of a plurality of sub-areas.

In a next step 230, pairs of boundaries of pixel intensities are input, each pair of boundaries comprising a minimum intensity and a maximum intensity.

Then, in step 240, the most efficient pair of boundaries is determined by performing the following steps for each pair of boundaries.

In a first sub-step 241, for each image of a sub-region, the number of pixels comprised in the image having an intensity between the minimum intensity and the maximum intensity of the pair of borders is determined.

Then, in a second sub-step 242, a single image is selected from the different images of the same sub-area of a given position of the platform 3, these different images being taken with different luminous intensities. The selected image corresponds to the image containing the most pixels with an intensity between the minimum and maximum intensities of the pair of boundaries. If at least two images contain the same number of pixels, the image of the selected sub-area is the image with the average of the pixel intensities closest to 127.5 intensities.

Then, in a third sub-step 243, for each position of the platform 3, each sub-region is examined using a classification convolutional neural network to detect the presence or absence of any defects.

Then, in a fourth substep 244, the detection by the classified convolutional neural network is compared to a table of records of known defects that locate the leaf to determine whether the detected defect is real or false.

Then, in a fifth substep 245, the detection efficiency is then calculated for a given pair of boundaries on the available area, taking all the positions of the platform 3. The detection efficiency depends on the ratio of the number of verified detections to the number of false detections.

Finally, in step 250, a pair of boundaries having the best detection efficiency is selected.

Accordingly, the present invention provides a method for training a system for automatically detecting defects in turbine engine blades, the system including a platform for receiving the blades to automatically control the surface of the turbine engine blades through a learning method.