Defect inspection device and defect inspection method
Patent Information
- Application Number
- JP2022030636
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2022-03-01
- Publication Date
- 2025-03-05
- Estimated Expiration
- 2042-03-01
AI Technical Summary
Existing defect inspection methods struggle to accurately estimate the positions of various types of defects due to the lack of a versatile indicator and the reliance on pre-trained classifiers, which cannot recognize defects not learned during training, and are subjective to inspector skill levels.
A defect inspection device that utilizes tapping sound data to analyze and calculate feature values, perform pattern recognition through machine learning, and iteratively refine the estimation of defective areas by comparing with surrounding areas until convergence, reducing subjectivity and improving accuracy across different defect types.
The device enables precise estimation of various defect positions with reduced inspector variability and physical burden, achieving high accuracy in defect detection regardless of inspector skill and defect type.
Smart Images
Figure 00000000_0000_ABST
Abstract
Description
Technical Field
[0001] This application mainly relates to a defect estimation device that estimates defects in an inspection object based on impact sound data.
Background Art
[0002] Patent Document 1 discloses a method for determining the presence or absence of defects based on the impact sound spectrum when hitting a FRP (Fiber Reinforced Plastic) with a hammer. Specifically, a value of Δ = dB2 - dB1 is obtained. dB1 is the sound of a frequency characteristic when there is no defect. A portion of the FRP panel without defects often has a mountain-shaped spectrum in this frequency band and is characteristic. Also, dB2 is the sound of a frequency where a difference appears when there is a defect. A portion of the FRP panel with a defect often has a higher level in this frequency band compared to when there is no defect.
[0003] Patent Document 2 discloses a method for determining the presence or absence of defects based on the feature amount of impact sound data when hitting a structure with a hammer. In this method, the feature amount is extracted by obtaining the sound intensity every 200 Hz based on the impact sound data. Next, the feature amount is selected using clustering. In this method, a classifier for classifying the impact sound data of the determination target is generated based on the selected feature amount. In Patent Document 2, this classifier is used to classify the impact sound data of the determination target into the impact sound data of a structure without defects or the impact sound data of a structure with defects.
[0004] Non-Patent Document 1 discloses a method for estimating the approximate size and position of a cavity based on the sound pressure waveform when hitting a test piece containing a cavity with an impact jig. In this method, a clustering map is created based on the sound pressure waveform obtained from a test piece with a known cavity size and position in advance and the frequency characteristics to which the fast Fourier transform is applied. Next, the approximate size and position of the cavity are estimated by determining which group of the clustering map the impact sound waveform of the test piece to be determined is closest to.
Prior Art Documents
[0005] [Patent Document 1] Patent No. 4736501 [Patent Document 2] Japanese Patent Publication No. 2018-13348 [Non-patent literature]
[0006] [Non-Patent Document 1] Eiki Matsuoka, Yoichi Hirose, Takahiko Kurahashi, Yuki Murakami, Shigehiro Toyama, Fujio Ikeda, Tetsuro Iyama, and Ikuo Ihara, "Application of Adjoint Variable Method to Improve the Accuracy of Quantitative Defect Evaluation by Impact Testing," Journal of the Society of Materials Science, Japan, Vol. 67, No. 9, September 2018, pp. 860-876. [Overview of the Initiative] [Problems that the invention aims to solve]
[0007] The method in Patent Document 1 determines defects based on the Δ value, but the Δ value itself has no physical meaning and cannot be said to be a universally applicable indicator of defects. Therefore, depending on the type of defect, the Δ value of the defective part and the Δ value of the non-defective part may overlap. Consequently, the method in Patent Document 1 cannot estimate the location of various types of defects. The method in Patent Document 2 estimates the location of defects based on a classifier that has been trained and generated in advance. Therefore, the method in Patent Document 2 cannot recognize defects that have not been trained, and thus cannot estimate the location of various types of defects. Non-Patent Document 1 discloses a method for estimating the location of defects related to cavities, but does not disclose a method for estimating other types of defects.
[0008] This application has been made in view of the above circumstances, and its main purpose is to provide a defect inspection device capable of estimating the location of various types of defects. [Means for solving the problem]
[0009] The problem that this application aims to solve is as described above, and next, the means for solving this problem and their effects will be explained.
[0010] According to a first aspect of this application, a defect inspection device having the following configuration is provided. That is, the defect inspection device comprises an acquisition unit and an estimation unit. The acquisition unit acquires impact sound data for each of a plurality of inspection positions in the inspection area from a sound receiving device. The estimation unit estimates the defect areas included in the inspection area based on the impact sound data acquired by the acquisition unit. The estimation unit performs the following processes (1) to (6): (1) Analyzes the impact sound data and calculates feature values that indicate the characteristics of the impact sound data for each inspection position. (2) Divides the inspection area into a provisional defect area and a provisional normal area which is the complement of the provisional defect area. (3) Performs pattern recognition by machine learning to distinguish the provisional defect area and the provisional normal area using the feature values. (4) Calculates the error of the pattern recognition for each inspection position. (5) Sets an inverted area based on the error and swaps the provisional defect area and the provisional normal area in the inverted area. (6) A detailed inspection is performed to estimate the defect region by repeating the processes from (3) to (5) until the results of the pattern recognition converge.
[0011] According to a second aspect of this application, a defect inspection device having the following configuration is provided. That is, the defect inspection device comprises an acquisition unit and an estimation unit. The acquisition unit acquires impact sound data for each of a plurality of inspection positions in the inspection area from a sound receiving device. The estimation unit estimates the defect areas included in the inspection area based on the impact sound data acquired by the acquisition unit. The estimation unit performs the following processes (1) to (3): (1) Sets a provisional inspection area and a plurality of comparison areas set based on the provisional inspection area within the inspection area. (2) Calculates feature values that indicate the characteristics of the impact sound data for the provisional inspection area and the comparison areas. (3) Estimates the defect areas based on the difference between the feature values of the provisional inspection area and the feature values of each comparison area.
[0012] According to a third aspect of the present application, the following defect inspection method is provided. That is, impact sound data for each of a plurality of inspection positions in an inspection area is acquired from a sound collection device. (1) Analyze the impact sound data to calculate a feature value indicating the feature of the impact sound data for each inspection position. (2) Divide the inspection area into a provisional defect area and a provisional normal area which is the complement of the provisional defect area. (3) Perform pattern recognition by machine learning so as to distinguish between the provisional defect area and the provisional normal area using the feature value. (4) Calculate the error of the pattern recognition for each inspection position. (5) Set an inversion area based on the error, and swap the provisional defect area and the provisional normal area in the inversion area. (6) Perform a detailed inspection to estimate a defect area by repeating the processes from (3) to (5) until the result of the pattern recognition converges.
Effect of the Invention
[0013] According to the present application, a defect inspection device capable of estimating the positions of various types of defects can be realized.
Brief Description of the Drawings
[0014] [Figure 1] Conceptual diagram showing a state of inspecting a panel using a defect inspection device. [Figure 2] Diagram showing a provisional inspection area and a comparison area of a panel in a preliminary inspection. [Figure 3] Flowchart of a preliminary inspection. [Figure 4] Diagram showing inspection positions, a provisional defect area, and a provisional normal area of a panel in a detailed inspection. [Figure 5] Flowchart of a detailed inspection. [Figure 6] Conceptual diagram showing the flow of a detailed inspection. [Figure 7] Diagram showing the result of an experiment using a defect inspection device. [Figure 8] Conceptual diagram showing a state of inspecting a panel using a defect inspection device of another embodiment. [[ID=3 / ID=37]]
Embodiments for Carrying Out the Invention
[0015] Next, embodiments of the present application will be described with reference to the drawings. FIG. 1 is a conceptual diagram showing a state in which a panel 21 is inspected using a defect inspection device 10.
[0016] The defect inspection device 10 is a device that estimates defects in the panel 21 based on impact sound data when the panel 21 is struck. The panel 21 is an inspection object of the defect inspection device 10. The panel 21 of the present embodiment is a multilayer panel having a structure in which both sides of a core material are sandwiched and adhered with skin materials. The core material is a member containing cavities such as a honeycomb core, for example. The skin material is a plate-like member thinner than the core material. Note that the defect inspection device 10 can inspect objects of any shape and material capable of impact sound inspection. Therefore, the inspection object of the defect inspection device 10 may be a solid member without cavities or an installed structure.
[0017] Examples of defects included in the panel 21 include a core recess where the core material is recessed compared to the surroundings, an adhesive non-applied portion where the core material and the skin material are not adhered, a backsheet application portion where a backsheet as a foreign object is included between the core material and the skin material, and the like. However, the defects detected by the defect inspection device 10 are not limited to the above defects.
[0018] In the present embodiment, an inspector performs an impact sound inspection using a hammer 22. Specifically, the inspector strikes each position of the panel 21 using the hammer 22 in accordance with a predetermined procedure. Note that an impact sound inspection may be performed using a percussion hammer or the like instead of the hammer 22. Also, as will be described later, the panel 21 may be automatically struck using a striking device.
[0019] As shown in FIG. 1, the defect inspection device 10 includes a sound collecting device 11 and a computer 12.
[0020] The sound-collecting device 11 is a microphone that receives the sound generated when the panel 21 is struck and converts it into an electrical signal. Hereinafter, this electrical signal will be referred to as sound data. The sound data shows the change in amplitude over time, i.e., the sound waveform. The sound-collecting device 11 may be of the diaphragm type or the piezoelectric type. The sound-collecting device 11 may be configured to be directional in a specific direction (for example, the direction in which the panel 21 is located) or it may be omnidirectional. The sound-collecting device 11 outputs the sound data to the computer 12 by wire or wireless connection.
[0021] The computer 12 comprises an arithmetic unit such as a CPU, a storage device such as an HDD, SSD, or flash memory, and a communication device such as a wired module or a wireless module. By executing a program stored in the storage device, the computer 12 can perform various processes related to the defect inspection device 10. Through the cooperation of the above hardware and software, the computer 12 can be operated as an acquisition unit 12a and an estimation unit 12b.
[0022] The acquisition unit 12a acquires impact sound data from the sound collection device 11. The acquisition unit 12a receives the impact sound data using the connector or antenna of the computer 12 and performs amplification and signal conversion, etc.
[0023] Furthermore, the acquisition unit 12a stores the sound data in a memory device, associating it with the position where the hammer 22 struck the panel. There are various methods for associating sound data with position. For example, the sound data can be associated with position by pre-determining the position and sequence in which the hammer 22 strikes the panel 21. Alternatively, the sound data can be associated with position by having the inspector strike the panel 21 with the hammer 22 in response to voice or video instructions from the computer 12. Alternatively, the sound data can be associated with position by having the inspector register the striking position by operating a keyboard or mouse, and then actually striking the panel 21 with the hammer 22.
[0024] The estimation unit 12b estimates the defect area of the panel 21 based on the impact sound data acquired by the acquisition unit 12a. The estimation of the defect area by the estimation unit 12b may be performed in parallel with the process in which the inspector strikes the panel 21 with the hammer 22. Alternatively, the estimation of the defect area by the estimation unit 12b may be performed after the process in which the inspector strikes the panel 21 with the hammer 22 is completed. Details of the process performed by the estimation unit 12b will be described later.
[0025] In tapping inspections, it is common for skilled inspectors to make judgments based on the timbre of the tapped sound. However, this method results in significant variability in inspection accuracy among inspectors. Furthermore, training inspectors requires considerable time and effort, and can be physically demanding for them. In addition, methods described in prior art documents rely on identifying or learning specific characteristics, making it difficult to accurately estimate the location of various types of defects. In this regard, the defect inspection device 10 of this embodiment reduces variability in inspection accuracy among inspectors by quantifying the sound collection conditions, and can more accurately estimate the location of various types of defects by focusing on characteristic errors that do not depend on specific characteristics. The defect inspection method of this embodiment will be described in detail below.
[0026] The inspection performed using the defect inspection device 10 includes a preliminary inspection to estimate the approximate location of the defect and a detailed inspection to estimate the shape of the defect. In this embodiment, after performing the preliminary inspection, a detailed inspection is performed using the results of the preliminary inspection. However, the preliminary inspection can also be omitted. If the preliminary inspection is omitted, both the location and shape of the defect are estimated during the detailed inspection.
[0027] The outline inspection will be explained below with reference to Figures 2 and 3. Figure 2 is a diagram showing the preliminary inspection area and comparison area of panel 21 in the outline inspection. Figure 3 is a flowchart of the outline inspection.
[0028] As mentioned above, panel 21 contains various types of defects. Furthermore, the sound produced when tapping a defective area differs depending on the type of defect. Therefore, it is difficult to learn all the sounds in advance. On the other hand, the sound produced when tapping a defective area differs from the sound produced when tapping the normal area surrounding the defect. In the preliminary inspection, this point is taken into consideration, and areas where the sound differs from other areas are identified and these areas are targeted for detailed inspection. The specific process is described below.
[0029] As shown in Figure 2, in the preliminary inspection, the surface of panel 21 is divided into multiple square regions. The length of one side of each region is length L1. In the preliminary inspection, the inspector strikes a point in one of these regions once or multiple times with a hammer 22. Preferably, the position where the inspector strikes with the hammer 22 is the center of each region, as shown by the midpoint in Figure 2. The sound collection device 11 converts the sound generated by striking with the hammer 22 into sound data and outputs it to the computer 12. The sampling frequency when collecting the sound and converting it into sound data is 44.1 kHz.
[0030] As shown in Figure 3, the acquisition unit 12a of the computer 12 acquires impact sound data for each position from the sound collection device 11 (S101). Next, the estimation unit 12b of the computer 12 sets the initial preliminary inspection area (S102). The preliminary inspection area is the area to be targeted for estimation of the presence or absence of defects using a preliminary inspection. In Figure 2, the preliminary inspection area is shown as hatched lines. In the preliminary inspection, all areas are inspected in order.
[0031] Next, the estimation unit 12b sets up multiple comparison areas based on the preliminary inspection area (S103). The comparison area is an area for comparing the preliminary inspection area with the impact sound data. In Figure 2, the comparison area is shown as halftone hatching. Here, areas adjacent to the preliminary inspection area are highly likely to have the same defects as the preliminary inspection area. Also, areas far from the preliminary inspection area may have different impact sound data even if they are normal areas. Therefore, in this embodiment, the comparison area is set to be two areas away from the preliminary inspection area. However, any area can be selected as the comparison area as long as it is different from the preliminary inspection area. For example, depending on the characteristics of the panel 21 and the size of the length L1, it may be preferable to select an adjacent area or an area three or more areas away.
[0032] Next, the estimation unit 12b calculates the difference between the feature value of the provisional inspection area and the feature value of each comparison area (S104). Feature values are values that indicate the characteristics of the impact sound data. In this embodiment, the AR coefficient of the impact sound data is used as the feature value. As described above, the impact sound data shows an impact sound waveform. If the same position on the panel 21 is struck multiple times with the hammer 22, the multiple impact sound waveforms may be correlated and multiple impact sound waveforms with a correlation coefficient of a certain value or higher may be extracted and averaged. The Nth order AR coefficient is estimated for this impact sound waveform. The AR coefficient can be estimated using known methods such as the Yule-Walker method, least squares method, maximum likelihood method, or Burg method. The estimated AR coefficient for the impact sound waveform at coordinate (i,j) is shown by the following equation (1).
number
[0033] In this embodiment, the difference in feature values is calculated by subtracting the AR coefficient of the comparison region from the AR coefficient of the provisional test region and squaring the result. If the coordinates of the provisional test region are (i1, j1) and the coordinates of a certain comparison region are (i2, j2), then the difference d AR This is shown by the following equation (2). The difference in the AR coefficient, which is a characteristic value, can be considered a value that indicates the difference in the timbre of the sound produced.
number
[0034] Next, the estimation unit 12b calculates the sum of the differences between the feature values of the provisional test area and the feature values of each comparison area P. AR Based on this, it is decided whether or not to include the set provisional inspection area in the detailed inspection (S105). Total value P AR Based on equation (2) above, this is expressed by the following equation (3).
number
[0035] The preliminary examination involves identifying areas for preliminary examination that will be the subject of the detailed examination. For example, the sum of the magnitudes of the differences P. AR The larger the value, the higher the probability that it is a false defect region, and the sum of the magnitudes of the differences P AR The smaller the value, the higher the probability that it is in the pseudo-normal range. Therefore, the estimation unit 12b calculates the sum of the magnitudes of the differences P. AR Based on this, the system estimates whether the relevant preliminary inspection area is a preliminary defect area or a preliminary normal area. If the estimation unit 12b estimates the preliminary inspection area to be a preliminary defect area, it will include this preliminary inspection area in the detailed inspection. Furthermore, if the estimation unit 12b estimates the preliminary inspection area to be a preliminary normal area, it will exclude this preliminary inspection area from the detailed inspection. However, the estimation unit 12b may also determine that preliminary normal areas near a preliminary defect area are subject to detailed inspection. A threshold may be used when deciding whether or not to include an area in the detailed inspection. If a threshold is used, the threshold can be determined appropriately based on experiments or experience. Alternatively, for example, the difference between the feature value of the preliminary inspection area and the feature value of the comparison area may be calculated individually for each of the 16 comparison areas in Figure 2, and the decision on whether or not to include an area in the detailed inspection may be made based on the number of locations where the difference exceeds the threshold.
[0036] Feature values are not limited to AR coefficients; they may be calculated using other methods. For example, a power spectrum calculated by performing a Fourier transform on a percussion sound waveform may be used. A power spectrum is a value that indicates the intensity of each frequency of a percussion sound waveform. Specifically, the power spectrum can be divided into multiple frequency bands, and the energy values of each frequency band can be quantified and used as feature values.
[0037] Next, the estimation unit 12b determines whether the inspection of all preliminary inspection areas has been completed (S106). If there are any preliminary inspection areas that have not been inspected, the estimation unit 12b sets the next preliminary inspection area (S107) and repeats the process from step S103 onwards. The setting of the next preliminary inspection area may be done by an inspector or the like, instead of the estimation unit 12b. If the inspection of all preliminary inspection areas has been completed, the estimation unit 12b terminates the overview inspection.
[0038] By performing the above-described overview inspection, it is possible to estimate the area containing defects, regardless of the type of defect. Furthermore, since the estimation is objective and does not depend on the inspector's subjectivity or experience, it is possible to estimate the area containing defects regardless of the inspector's skill level.
[0039] Next, we will explain the detailed inspection process with reference to Figures 4 to 6. Figure 4 is a diagram showing the inspection position, provisional defect area, and provisional normal area of a panel during a detailed inspection. In Figure 4, the inspection position of the panel is indicated by the midpoint, the provisional defect area by diagonal grid hatching, and the provisional normal area by diagonal hatching. Figure 5 is a flowchart of the detailed inspection process. Figure 6 is a conceptual diagram showing the flow of the detailed inspection process.
[0040] In this embodiment, the detailed inspection is performed after the preliminary inspection. The detailed inspection is an inspection to estimate the shape of the defect in the provisional defect area estimated in the preliminary inspection. Therefore, the length L2, which indicates the interval between the positions where the sound-collecting device 11 is struck in the detailed inspection shown in Figure 4, is often shorter than the length L1, which indicates the interval between the positions where the sound-collecting device 11 is struck in the preliminary inspection shown in Figure 2.
[0041] First, the estimation unit 12b sets the inspection area (S201). The inspection area is the area of the unit to be inspected in the detailed inspection. The inspection area is set based on the provisional defect area estimated in the general inspection. In this embodiment, the inspection area is set to an area where one side of the provisional defect area is three times longer. That is, if the length of one side of the provisional defect area is denoted by L3 and the length of one side of the inspection area is denoted by L4, then 3 × L3 = L4 holds true. However, there are various ways to set the inspection area; the inspection area may be set to an area where one side of the provisional defect area is a predetermined multiple, for example, 2 times or more and 4 times or less, or an area of a predetermined size and shape may be set to the inspection area.
[0042] Next, the computer 12 instructs the inspector to tap the inspection area at intervals of length L2. Hereafter, the positions within the inspection area that are tapped for detailed inspection will be referred to as inspection positions. The acquisition unit 12a of the computer 12 acquires tapping sound data for each inspection position from the sound collection device 11 (S202).
[0043] Next, the estimation unit 12b of the computer 12 calculates feature values for each inspection location (S203). In this embodiment, the AR coefficient is used as a feature value in the detailed inspection as well as in the general inspection. The method for estimating the AR coefficient is the same as in the general inspection. Also, as in the general inspection, the power spectrum can be used as a feature value. Note that, for the entire panel 21, there may be large differences between AR coefficients in the normal region at distant locations, but within a narrow inspection area, it can be assumed that the AR coefficients in the normal region are almost the same.
[0044] Next, the estimation unit 12b sets the provisional defect region and the provisional normal region based on the results of the preliminary inspection (S204). In this embodiment, the provisional defect region and the provisional normal region are set based on the results of the preliminary inspection. The inspection region is classified into either the provisional defect region or the provisional normal region. In other words, the provisional normal region is the complement of the provisional defect region. Note that if the preliminary inspection is omitted, any region may be set as the provisional defect region. This is because, as will be described later, any incorrect setting of the provisional defect region will be corrected.
[0045] Next, the estimation unit 12b performs pattern recognition using machine learning to distinguish between the false defect region and the false normal region using feature values (S205). In this embodiment, pattern recognition is performed in the following way. Here, the false defect region is shown as A, and the false normal region is shown as B. A and B each contain multiple inspection positions. The AR coefficients, which are feature quantities of these inspection positions, are linearly separated in linear space. That is, the separation function shown in equation (4) below is used, and as shown in equation (5), the false defect region is set to 1 and the false normal region to 0. Then, the weights w of the separation function are set to minimize the squared error I shown in equation (6). k Determine k=1, ..., N.
number
[0046] Furthermore, the method of distinguishing between pseudo-defective regions and pseudo-normal regions using pattern recognition based on feature values is not limited to the method described above. For example, pattern recognition may be performed using techniques such as nonlinear separation or support vector machines.
[0047] Next, the estimation unit 12b calculates the pattern recognition error for each inspection position (S206). This error is an index for evaluating the accuracy of pattern recognition, and specifically is the squared error I described above.
[0048] Next, the estimation unit 12b determines whether the pattern recognition results have converged (S207). The method for determining whether convergence has occurred is arbitrary, and various methods can be used. For example, the estimation unit 12b determines whether the pattern recognition results have converged based on the smallness of the pattern recognition error in the inspection area, for example, the value of the squared error I. In detail, a threshold is set for the pattern recognition error, and the pattern recognition results are determined to have converged when there are no longer any regions where the error exceeds the threshold. However, the pattern recognition results may be determined to have converged even if there are still small regions where the error exceeds the threshold.
[0049] If the estimation unit 12b determines that the pattern recognition result has not converged, it sets an inverted region based on the pattern recognition error and swaps the provisional defect region and provisional normal region within the inverted region (S208). The inverted region can be set, for example, using a threshold. That is, if a region where the pattern recognition error exceeds the threshold was a provisional defect region, it is changed to a provisional normal region, and if a region where the pattern recognition error exceeds the threshold was a provisional normal region, it is changed to a provisional defect region. In the example shown in Figure 6, the region where the pattern recognition error exceeds the threshold is the inverted region. Furthermore, the inverted region can be set to a region limited to the boundary between the provisional defect region and the provisional normal region (assuming that the defect region is one region) among the regions with large pattern recognition errors, or to a region with a shape similar to a specific pattern (such as a frequently occurring defect shape).
[0050] Subsequently, the estimation unit 12b repeats the processes from steps S205 to S207. That is, after swapping the provisional defect region and the provisional normal region, it performs pattern recognition and error difference calculation using machine learning again. The estimation unit 12b repeats these processes until the pattern recognition results converge. When the estimation unit 12b determines that the pattern recognition results have converged, it estimates the region where the separation function shown in equation (4), in other words, the value calculated according to the feature values, exceeds a threshold, as the defect region (S209). Alternatively, the provisional defect region at the time the pattern recognition results converge may be estimated as the defect region.
[0051] Figure 7 shows the results of an experiment in which defects were estimated by performing a preliminary inspection and a detailed inspection using the defect inspection device 10 of this embodiment. In this experiment, defects with sizes of 10 mm, 20 mm, and 30 mm were created on multiple panels, and these were estimated. The types of defects created were the areas where adhesive was not applied and the areas where the back sheet was applied, as described above. In addition, the skin material placed on the upper side of the panel used in the experiment is called the upper skin, and the skin material placed on the lower side is called the lower skin. The surface of the upper skin contains irregularities. The surface of the lower skin is flat. In Figure 7, three photographs are arranged in the left-to-right direction. In each photograph, the size of the created defects is different (10 mm, 20 mm, and 30 mm from left to right). The panel, surface, and defects used in the inspection are indicated on the left side of the leftmost photograph. Also, the rectangles indicated in the center of each figure in Figure 7 indicate the extent of the actually created defects. In each figure in Figure 7, the area enclosed by the curve is the defect area estimated by the estimation unit 12b. For defects that the inspector was unable to detect when inspecting without the defect inspection device 10, an "X" mark is placed in the upper right corner of the photograph. As shown in Figure 7, when inspector A inspected without the defect inspection device 10, a defect with a size of 10 mm on panel 1 was not detected. However, by using the defect inspection device 10, this defect was successfully detected. Similarly, when inspector B inspected without the defect inspection device 10, defects with sizes of 10 mm, 20 mm, and 30 mm on panel 3 were not detected. However, by using the defect inspection device 10, these defects were successfully detected. As shown in Figure 7, by using the defect inspection device 10 of this embodiment, defects can be estimated with high accuracy regardless of individual differences among inspectors or the size of the defects.
[0052] Next, other embodiments will be described with reference to Figure 8. In the defect inspection device 10 described above, an inspector strikes the panel 21 with a hammer 22. Alternatively, in the defect inspection device 100 of the other embodiment, the panel 21 is automatically struck using a striking device 32.
[0053] The striking device 32 is equipped with a hammer at its tip, for example, and can strike the panel 21 by moving the hammer back and forth. The striking device 32 is also configured to be movable along one side of the panel 21, for example. The panel 21 is placed on a conveyor 31 whose transport amount can be adjusted. Therefore, the conveyor 31 and the striking device 32 can strike any position on the surface of the panel 21. With this configuration, the striking device 32 can strike a predetermined position or a position determined by the computer 12. By including the striking device 32 in the defect inspection device 100, the physical burden on the inspector can be reduced.
[0054] The striking device 32 may be movable in two orthogonal directions in a plan view. In this case, the conveyor 31 can be omitted. The defect inspection device 100 may also integrate the sound pickup device 11, the computer 12, and the striking device 32.
[0055] The sound pickup device 11 and computer 12 of the defect inspection device 100 are the same as those in the embodiment described above, so their explanation will be omitted.
[0056] The defect inspection apparatus 10,100 of this embodiment comprises an acquisition unit 12a and an estimation unit 12b. The acquisition unit 12a acquires impact sound data for each of multiple inspection positions in the inspection area from the sound collection device 11. The estimation unit 12b estimates the defect area included in the inspection area based on the impact sound data acquired by the acquisition unit 12a. (1) The impact sound data is analyzed to calculate feature values that indicate the characteristics of the impact sound data for each inspection position. (2) The inspection area is divided into a provisional defect area and a provisional normal area which is the complement of the provisional defect area. (3) Pattern recognition is performed by machine learning to distinguish between the provisional defect area and the provisional normal area using feature values. (4) The error of pattern recognition is calculated for each inspection position. (5) An inversion area is set based on the error, and the provisional defect area and provisional normal area of the inversion area are swapped. (6) Detailed inspection is performed to estimate the defect area by repeating the processes from (3) to (5) until the pattern recognition results converge.
[0057] This allows for the estimation of the location of various types of defects by predicting defect areas based on comparisons with the surrounding area, rather than relying on pre-created judgment models based on training data.
[0058] In the defect inspection apparatus 10,100 of this embodiment, the estimation unit 12b determines that the pattern recognition results have converged when there are no longer any regions where the error exceeds a threshold.
[0059] This allows for accurate estimation of the defect area.
[0060] In the defect inspection apparatus 10,100 of this embodiment, the acquisition unit 12a acquires impact sound data of a provisional inspection area and impact sound data of multiple comparison areas located at different positions from the provisional inspection area from the sound collection device 11. The impact sound data is analyzed to calculate the characteristic values of the impact sound data of the provisional inspection area and the characteristic values of the impact sound data for each comparison area. Based on the sum of the magnitude of the difference between the characteristic values of the provisional inspection area and the characteristic values for each comparison area, a preliminary inspection is performed to divide the provisional inspection area into a provisional defect area and a provisional normal area.
[0061] By comparing the feature quantities of impact sound data from different locations, the pseudo-defect region can be accurately defined.
[0062] In the defect inspection device 10,100 of this embodiment, the interval between the positions where the object to be inspected is struck during detailed inspection is shorter than the interval between the positions where the object to be inspected is struck during general inspection.
[0063] A preliminary examination provides an overview, while a detailed examination allows for a more thorough assessment.
[0064] The defect inspection device 10,100 of this embodiment is equipped with a striking device 32 that automatically strikes multiple inspection locations.
[0065] This reduces the workload for inspectors.
[0066] Although preferred embodiments of this application have been described above, the above configuration can be modified as follows, for example.
[0067] In the above embodiments, a method of performing both a general inspection and a detailed inspection, and a method of performing only a detailed inspection were described. Alternatively, a general inspection may be performed and the detailed inspection omitted. When the detailed inspection is omitted, the area estimated to be a provisional defect area in the general inspection of the above embodiments is treated as a defect area.
[0068] The flowchart shown in the above embodiment is just one example, and some processes may be omitted, some processes may be modified, or new processes may be added.
[0069] The functions of each element, including the estimation unit 12b disclosed in this application, can be performed using a circuit or processing circuit, including a general-purpose processor, a dedicated processor, an integrated circuit, an ASIC (Application Specific Integrated Circuit), a conventional circuit, and / or a combination thereof, configured or programmed to perform the disclosed functions. A processor is considered a processing circuit or circuit because it includes transistors and other circuits. In this disclosure, a circuit, unit, or means is hardware that performs the enumerated functions, or hardware programmed to perform the enumerated functions. The hardware may be hardware disclosed herein, or other known hardware that is programmed or configured to perform the enumerated functions. If the hardware is a processor, which is considered a type of circuit, then the circuit, means, or unit is a combination of hardware and software, and the software is used to configure the hardware and / or the processor. [Explanation of symbols]
[0070] 10,100 Defect Inspection Equipment 11. Sound collection device 12 Computers 12a Acquisition part 12b Estimation part 21 panels 22 Hammer 31 Conveyor 32 Impact device
Claims
1. an acquisition unit that acquires hammering sound data for each of a plurality of test positions in the test area from a sound collection device; an estimation unit that estimates a defective area included in the inspection area based on the hammering sound data acquired by the acquisition unit; Equipped with The estimation unit (1) analyzing the hammering sound data to calculate a feature value indicating a feature of the hammering sound data for each of the inspection positions; (2) Dividing the inspection area into a provisional defect area and a provisional normal area which is a complement of the provisional defect area; (3) performing pattern recognition by machine learning to distinguish between the provisional defect area and the provisional normal area using the feature value; (4) calculating an error in the pattern recognition for each of the inspection positions; (5) setting a reversed region based on the error, and swapping the temporary defective region and the temporary normal region in the reversed region; (6) A defect inspection apparatus characterized in that detailed inspection is performed to estimate the defect area by repeating the processes (3) to (5) until the result of the pattern recognition converges.
2. 2. The defect inspection device according to claim 1, The defect inspection device is characterized in that the estimation unit determines that the result of the pattern recognition has converged when there is no longer any region in which the error exceeds a threshold value.
3. 3. The defect inspection device according to claim 1, the acquiring unit acquires, from the sound collecting device, hammering sound data of a provisional inspection area and hammering sound data of a plurality of comparison areas located at positions different from the provisional inspection area; The estimation unit analyzing the hammering sound data to calculate the feature value of the hammering sound data of the provisional inspection area and the feature value of the hammering sound data of each comparison area; A defect inspection device characterized by performing a general inspection to divide the provisional inspection area into the provisional defect area and the provisional normal area based on the difference between the feature value of the provisional inspection area and the feature value of each comparison area.
4. 4. The defect inspection device according to claim 3, A defect inspection device characterized in that the intervals between positions at which the inspection object is struck during the detailed inspection are shorter than the intervals between positions at which the inspection object is struck during the general inspection.
5. 5. The defect inspection device according to claim 1, A defect inspection device comprising a striking device that automatically strikes a plurality of said inspection positions.
6. an acquisition unit that acquires hammering sound data for each of a plurality of test positions in the test area from a sound collection device; an estimation unit that estimates a defective area included in the inspection area based on the hammering sound data acquired by the acquisition unit; Equipped with The estimation unit (1) setting a provisional inspection area and a plurality of comparison areas set based on the provisional inspection area within the inspection area; (2) calculating a feature value indicating a feature of the hitting sound data for the provisional inspection area and the comparison area; (3) A defect inspection apparatus characterized in that a defective area is estimated based on the difference between the feature value of the provisional inspection area and the feature value of each of the comparison areas.
7. acquiring hammering sound data for each of a plurality of test positions in the test area from a sound collection device; (1) analyzing the hammering sound data to calculate a feature value indicating a feature of the hammering sound data for each of the inspection positions; (2) Dividing the inspection area into a provisional defect area and a provisional normal area which is a complement of the provisional defect area; (3) performing pattern recognition by machine learning to distinguish between the provisional defect area and the provisional normal area using the feature value; (4) calculating an error in the pattern recognition for each of the inspection positions; (5) setting a reversed region based on the error, and swapping the temporary defective region and the temporary normal region in the reversed region; (6) A defect inspection method characterized by performing detailed inspection to estimate a defect area by repeating the processes (3) to (5) until the result of the pattern recognition converges.