Construction method of learning model and inspection device

Abstract

To provide a construction method of a learning model and an inspection device capable of reducing overlooking of a defect and excessive detection of a non-defective product.SOLUTION: From inspection of an inspection object using a learning model, a result of overlooking of a defective product within a range of the feature amount of a defect or excessive detection (false report) of a non-defective product outside the range is acquired (S4), and distribution of the feature amount of the defect is updated based on the result to construct the learning model (S6 → S2).SELECTED DRAWING: Figure 9

Description

[Technical field]

[0001] The present invention relates to a method for constructing a learning model and an inspection device. [Background technology]

[0002] Patent Document 1 discloses a technology for inspecting the presence or absence of defects on the surface of an object to be inspected and determining whether the object is good or bad, based on a captured image of the surface of the object and the results of machine learning using multiple artificial defect images. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2011-214903 A Summary of the Invention [Problem to be solved by the invention]

[0004] In the above conventional technology, even if a human being designs the boundary between good and defective products and the distribution of defective products by looking at the distribution of defect features obtained from captured images, machine learning may perform pass / fail judgment by taking into account features other than the designed ones. As a result, there is a risk of a discrepancy between the range of feature values ​​of defects to be detected and the range of feature values ​​that the trained model determines as defects, resulting in overlooking defects or overdetection (false reporting) of good products. An object of the present invention is to provide a method for constructing a learning model and an inspection apparatus that can reduce overlooking defects and overdetection of non-defective products. [Means for solving the problem]

[0005] A method for constructing a learning model in one embodiment of the present invention generates artificial defects based on features of the defect to be inspected, a range of features for generating the artificial defects, and a distribution of the amount of artificial defects generated for the features in that range, constructs a learning model that learns the artificial defects, and from inspecting the object to be inspected using the learning model, obtains results of overlooking defective products that are within the range of the defect features or overdetecting good products that are outside the range, and updates the distribution based on the results to construct the learning model. Effect of the Invention

[0006] Therefore, in the present invention, it is possible to reduce overlooking defects and overdetection of non-defective products. [Brief description of the drawings]

[0007] [Figure 1] 1 is a schematic diagram of a visual inspection device 1 according to a first embodiment. [Diagram 2] 1 is an example of an initial shape E1 and a final shape E2 of an ellipse. [Diagram 3] This is an example of an image of generating defect elements for blowholes using the ellipses shown in Figure 2. [Figure 4] 1 is an example batch image of defective elements for blowholes. [Diagram 5] FIG. 3 is another example of a generated image of defect elements for blowholes and scratches using the ellipses shown in FIG. 2. [Figure 6] 13 is an example of a patch image of a defect element for a blowhole and a defect element for a scratch. [Figure 7] 13 is another example of an initial shape E1 and a final shape E2 of an ellipse. [Figure 8] FIG. 4 is a diagram illustrating a method for designing a distribution of feature amounts in the first embodiment. [Figure 9] 1 is a flowchart showing the flow of a learning model construction method according to the first embodiment. [Figure 10] FIG. 4 is a diagram illustrating a method for updating the distribution of feature amounts in the first embodiment. [Figure 11] FIG. 4 is a diagram illustrating a method for updating the distribution of feature amounts in the first embodiment. [Figure 12] FIG. 11 is a diagram illustrating a method for updating the distribution of feature amounts in the second embodiment. [Figure 13] FIG. 13 is a diagram illustrating a method for updating the distribution of feature amounts in the third embodiment. [Figure 14] FIG. 13 is a diagram illustrating a method for updating the distribution of feature amounts in the third embodiment. [Figure 15] FIG. 13 is a diagram illustrating a method for updating the distribution of feature amounts in the fourth embodiment. [Figure 16] FIG. 13 is a diagram illustrating a method for updating the distribution of feature amounts in the fourth embodiment. [Figure 17] FIG. 13 is a diagram showing a method for setting the distribution of feature amounts in the fifth embodiment. [Figure 18] FIG. 13 is a diagram illustrating a method for updating the distribution of feature amounts in the fifth embodiment. [Figure 19] FIG. 13 is a diagram illustrating a method for updating the distribution of feature amounts in the fifth embodiment. [Figure 20] 13 is an example of equipment data waveforms during normal operation and when a fault occurs. [Figure 21] FIG. 13 is a diagram illustrating a machine learning method according to a sixth embodiment. [Figure 22] 23 is a diagram illustrating a method for designing the distribution of feature quantities and a method for generating an artificial fault waveform in the sixth embodiment. FIG. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0008] [Embodiment 1] FIG. 1 is a schematic diagram of a visual inspection device 1 according to the first embodiment. The appearance inspection device 1 of the embodiment 1 includes a camera 2, a robot 3, and a computer 4. The camera 2 captures an image of the surface of an inspection object, which is an engine piston material or a finished processed engine piston product (hereinafter simply referred to as a piston) 5. The robot 3 is an articulated robot and has a hand 3a that holds the piston 5.

[0009] The computer 4 is, for example, a personal computer, and includes a CPU 6 and a memory 7. The CPU 6 includes an inspection unit 8 and a learning unit 9. The inspection unit 8 includes an acquisition unit 8a, a defect determination unit 8b, and an image update determination unit 8c. The acquisition unit 8a acquires an image captured by the camera 2. The defect determination unit 8b inspects whether or not there are scratches or defects on the surface of the piston 5 based on the image acquired by the acquisition unit 8a and the learning results stored in the memory 7 (products with scratches or defects are NG, products without scratches or defects are OK). The image update determination unit 8c will be described later.

[0010] The learning unit 9 has a defect element generating unit 9a, an artificial defect image generating unit 9b, and a CNN learning unit 9c. The defect element generating unit 9a generates a defect element image. The defect element image may be generated by any method, and in the first embodiment, a defect element image is generated by overlapping a plurality of predetermined shapes. The artificial defect image generating unit 9b synthesizes the defect element image with a surface image of a sample product equivalent to the piston 5 to generate an artificial defect image that is a pseudo defective sample image. The CNN learning unit 9c performs machine learning using a plurality of artificial defect images. The machine learning is learning using a neural network, and in the first embodiment, deep learning using a CNN (Convolutional Neural Network) model is adopted. The memory 7 stores a plurality of defect element images generated by the defect element generating unit 9a, a plurality of artificial defect images generated by the artificial defect image generating unit 9b, and a learning result by the CNN learning unit 9c.

[0011] Next, a method for generating a defect element image by the defect element generating unit 9a will be described. The defect element generating unit 9a generates a defect element image by superimposing a plurality of basic shapes having different sizes and brightness. In the first embodiment, the basic shape is an ellipse, and the defect element image is generated by changing the input parameters. FIG. 2 shows an example of an initial shape E1 and a final shape E2 of an ellipse. The initial shape E1 in FIG. 2(a) has a major axis d1 and an aspect ratio a1. On the other hand, the final shape E2 in FIG. 2(b) has a major axis d2 and an aspect ratio a2.

[0012] The defect element is generated by generating one or more intermediate shapes that complement the initial shape E1 and the final shape E2, and overlapping each shape. Figure 3 shows an example of a generation image of a defect element for a blowhole using the ellipse shown in Figure 2, and the input parameters are the major axis (first parameter) d, the brightness (second parameter) L, the drawing path (third parameter) p(t), the rotation angle θ, and the number of ellipses n. The drawing path p(t) is expressed mathematically as a straight line, and the number of ellipses n=3. This makes it possible to generate a patch image of a defect element for a blowhole as shown in Figure 4.

[0013] Figure 5 shows an image of how defect elements for blowholes and scratches are generated using the ellipses shown in Figure 2, with the input parameters being the major axis d, brightness L, depiction path p(t), random rotation angle dθ, random coordinate shift amount dP, and number of ellipses n. The depiction path p(t) is expressed mathematically as a straight line, and the number of ellipses n = 4. By randomly rotating / translating the basic shape, it is possible to generate patch images of defect elements for blowholes, as shown in Figure 6(a), and defect elements for scratches, as shown in Figure 6(b). The same applies to patch images of defect elements for cast holes.

[0014] 7A and 7B are diagrams showing a method for generating an artificial defect image by the artificial defect image generating unit 9b, where FIG. 7A shows the method for generating an artificial defect image, and FIG. 7B shows the method for generating a patch image of an artificial defect. As shown in Fig. 7(a), a plurality of artificial defect images 12 can be generated by synthesizing patch images 10 of a plurality of defect elements with a patch image 13 of the piston 5 as a background image. On the other hand, as shown in Fig. 7(b), a plurality of artificial defect patch images 14 can be generated by synthesizing patch images 10 of a plurality of defect elements with a patch image 13 of the piston 5 as a background image.

[0015] In order to generate an artificial defect image, it is necessary to select the defect feature amount, and to design the range (NG range) and the distribution of the number of samples within the NG range. In the case of visual inspection, at least one of the brightness, size, and shape of the defect image is used as the defect feature amount. FIG. 8(a) is an example of a feature distribution when the feature amount of the captured image is the brightness. In this example, the brightness of a general defect is distributed in a range of 70 or less, and the brightness of a good product is distributed in a range of over 70. Since the boundary between a good product and a defect in terms of brightness is 70, as shown in FIG. 8(b), the range of brightness 0 to 70 is designed as the range of the defect feature amount (NG range), and the range of brightness over 70 is designed as the range of the good product feature amount (OK range). In addition, the number of samples (relative frequency) at each brightness within the NG range is, for example, a constant distribution. The artificial defect image generating unit 9b generates an artificial defect image based on the defect feature amount, the NG range, and the distribution of the number of samples within the NG range (hereinafter simply referred to as the distribution of the feature amount) designed in advance.

[0016] The CNN learning unit 9c makes the neural network learn defect images from the multiple artificial defect images and multiple artificial defect patch images generated by the artificial defect image generating unit 9b. Here, in machine learning, learning is performed using images generated by the artificial defect image generating unit 9b, not features designed by humans, so there is a discrepancy between the range of features that the trained model determines as defects and the NG range that humans want to detect. As a result, there is a risk of overlooking defects or overdetecting good products (hereinafter also referred to as false reports). Therefore, in the first embodiment, the results of overlooking defective products that are within the NG range and false reports of non-defective products that are outside the range are obtained from the inspection of the object using a learning model, and the distribution of the initially designed features is updated based on the results to construct a learning model.

[0017] FIG. 9 is a flowchart showing the flow of the learning model construction method of the first embodiment. In step S1, a distribution of feature quantities designed in advance is input to the artificial defect image generating unit 9b. In step S2, the artificial defect image generating unit 9b generates a plurality of artificial defect images based on the distribution of the input feature amounts. In step S3, the CNN learning unit 9c executes deep learning (machine learning) based on the generated artificial defect images to create a CNN model. In step S4, the defect determination unit 8b evaluates whether the captured image is a defective image (NG) or a non-defective image (OK) based on the created CNN model.

[0018] In step S5, the image update determination unit 8c determines whether the end condition of the artificial defect image update is satisfied. If YES, the process is terminated, and if NO, the process proceeds to step S6. Here, the image update determination unit 8c evaluates whether the captured image is NG or OK based on a predetermined algorithm different from the pass / fail determination by the defect determination unit 8b using the CNN model, and compares the result with the evaluation by the defect determination unit 8b. Then, if the image update determination unit 8c's own evaluation is NG and the evaluation by the defect determination unit 8b is OK, the defect determination unit 8b determines the defect image as an oversight (hereinafter also referred to as an oversight). On the other hand, if the image update determination unit 8c's own evaluation is OK and the evaluation by the defect determination unit 8b is NG, the defect determination unit 8b determines the defect image as a false report (hereinafter also referred to as a false report) of a non-defective image. In step S5, for example, if the oversight by the defect determination unit 8b is 0% and the false report is less than t%, the image update determination unit 8c determines that the end condition of the artificial defect image update is satisfied.

[0019] In step S6, the artificial defect image generating unit 9b updates the distribution of the feature amount, and the process proceeds to step S2. In step S2, a plurality of artificial defect images are generated based on the updated distribution of the feature amount so as to reduce oversights or false reports. Hereinafter, a method of updating the distribution of the feature amount will be described. In the first embodiment, as a method of reducing oversights or false reports, a method of adjusting a parameter a (slope: a≧0) for changing the weight of the number of samples (relative frequency) near the boundary of the feature amount, and a method of adjusting a parameter b (distance) for changing the range of the distribution of the feature amount to be detected are used, as shown in FIG. 10(b) (the initial distribution of the feature amount shown in FIG. 8(b) is set to a=0, b=0).

[0020] As shown in Figure 10(c), if artificial defects with features similar to those of non-defective products are excessively taught in CNN learning, the number of false alarms will increase. In this case, as shown in Figure 11, at least one of the values ​​of parameters a and b is reduced. In other words, the weighting near the feature boundaries is reduced, or the NG range is narrowed. On the other hand, if defects near the features cannot be detected in CNN learning, the number of oversights will increase. In this case, at least one of the values ​​of parameters a and b is increased. In other words, the weighting near the feature boundaries is increased, or the NG range is widened.

[0021] As described above, in the first embodiment, the results of overlooking defective products that are within the range of defect features or overdetecting (false reports) good products that are outside the range are obtained from the inspection of the object using the learning model (S4 in FIG. 9), and the distribution of the features is updated based on the results to construct a learning model (S6→S2 in FIG. 9). This makes it possible to update the distribution of the features so that the NG range that humans want to determine as a defect matches the NG range of CNN learning (deep learning). As a result, it becomes possible to construct a learning model that can reduce overlooking defects and false reports of good products.

[0022] In the first embodiment, the feature distribution is updated by increasing the number of samples near the feature boundaries (FIG. 10(b)). This improves the ability to detect feature boundaries set by humans. Furthermore, the feature distribution is updated by updating the slope (parameter a) of the amount of artificial defects generated with respect to the change in feature so that the amount of artificial defects generated near the feature boundaries increases. In other words, the slope of the amount of artificial defects generated with respect to the change in feature is updated so that the amount of artificial defects generated decreases the further away from the feature boundaries. This increases the weighting of the number of samples near the feature boundaries, thereby further improving the ability to detect feature boundaries designed by humans.

[0023] Furthermore, the feature distribution is updated by expanding or shrinking the NG range so as to reduce overlooked defects and false positives (parameter b). This improves the consistency of judgment between the feature boundaries designed by humans and machine learning. For example, if the number of overlooked defective products increases, the NG range is expanded to within the current OK range. This reduces overlooked defects. On the other hand, if the number of false positives increases, the NG range is narrowed below the current NG range. This reduces false positives.

[0024] The learning model shown in the first embodiment is a method for constructing a learning model used in the visual inspection apparatus 1, the artificial defect is a composite defect image obtained by combining an image of an artificial defect with an image of a non-defective product, and the feature amount is at least one of the brightness, size, and shape of the image of the artificial defect. This allows the visual inspection apparatus 1 to accurately determine the presence or absence of scratches or defects on the surface of the detected object.

[0025] [Embodiment 2] The second embodiment differs from the first embodiment in that, as shown in Fig. 12, only parameter b that expands or contracts the NG range is adjusted as a parameter for updating the distribution of feature quantities. In other words, the distribution of the number of samples within the NG range is constant regardless of the luminance, as shown in Fig. 8(b). The method of adjusting parameter b is the same as in the first embodiment. In this way, by using only one parameter, it is possible to more easily correct the distribution than in the first embodiment.

[0026] [Embodiment 3] In the third embodiment, the distribution of the number of samples within the NG range is designed to be a mountain-shaped (approximately triangular with an upward convex shape) when the first artificial defect is generated, as shown in Fig. 13. As parameters for updating the feature amount, a method of adjusting parameter a for changing the feature amount with the largest number of samples (relative frequency) (changing the height of the apex of the mountain shape) and a method of adjusting parameter b for changing the range of the feature amount to be taught as NG are used. Note that when updating the feature amount, the mountain-shaped shape is maintained and the number of samples within the NG range is kept constant.

[0027] If a defect is overlooked due to the initial artificial defect generation, parameter a is increased and parameter b is enlarged, as shown in Fig. 14. On the other hand, although not shown, if a false report of a good product occurs, parameter a is decreased and parameter b is reduced. Since the actual distribution of the feature quantity is close to a mountain shape, by teaching the distribution of the feature quantity in a mountain shape as in the third embodiment, it is possible to improve the detection accuracy of the feature quantity boundary and intuitively correct the distribution of the feature quantity.

[0028] [Embodiment 4] The fourth embodiment differs from the first embodiment in that the distribution shape of the feature is not managed by a parameter, but is increased or decreased one sample at a time based on the learning result. A range A is set on the NG side and a range B is set on the OK side of the boundary of the defect feature at the time of the first artificial defect generation shown in FIG. 8(b), and if there is an oversight, one artificial defect sample is added randomly from range B (FIG. 15(b)). On the other hand, if a false report is generated, one artificial defect sample is added randomly from range A (FIG. 16(b)). In the fourth embodiment, the correction range is small (one sample at a time) without significantly changing the distribution of the feature all at once, so that it is possible to suppress a large fluctuation in the defect judgment range of deep learning every time the distribution of the feature is updated.

[0029] [Embodiment 5] The fifth embodiment differs from the first embodiment in that the NG range is determined using two feature amounts. In the second embodiment, the feature amount of the defect is the brightness and the major axis of the ellipse when generating the defect element image. In the example of FIG. 17, the brightness of the defect is distributed in an area of ​​70 or less, and the brightness of the non-defective product is distributed in an area of ​​more than 70. In addition, the major axis of the defect is distributed in an area of ​​1.0 mm or more, and the major axis of the non-defective product is distributed in an area of ​​less than 1.0 mm. Since the boundary between the defect and the non-defective product in the major axis is 1.0 mm, and the boundary between the defect and the non-defective product in the brightness is 70, the range of the major axis over 1.0 and the brightness from 0 to 70 is set as the range of the feature amount of the defect (NG range), and the number of samples (relative frequency) in the NG range is designed to have a certain distribution, for example, as shown in FIG. 18(a). FIG. 18(a) shows the distribution z of the feature amount of the defect at the time of the first artificial defect generation. The size of the circle is the number of samples, and the larger the circle, the more the number of samples increases. The horizontal axis x is the coordinate of the brightness, and the vertical axis y is the coordinate of the major axis.

[0030] 18(b), it is assumed that in the evaluation by the defect determination unit 8b based on the initial artificial defect generation, the defect was overlooked due to its major axis and a false report was generated due to the defect brightness. Here, the distribution z can be calculated by the following formula (1). Z=a x x+a y y (x≦t x +b x , y≧t y +b y ) …(1) In addition, b x is the initial boundary of luminance t x Distance from b y is the initial boundary of the major axis t y is the distance from In formula (1), if there is an oversight, the parameters a and b are increased, and if there is a false alarm, the parameters a and b are decreased. This can reduce oversights and false alarms. For example, in the case of FIG. 18(b), an oversight occurred in the major axis and a false alarm occurred in the brightness, so in formula (1), x and b x Make a smaller y and b yAs a result, as shown in Fig. 19, the range of the major axis expands and the weighting increases as it approaches the boundary, and the range of brightness shrinks and the weighting decreases as it approaches the boundary. The distribution z is repeatedly updated until the range of defects that humans want to determine as defects matches the range of defects determined by the CNN learning model.

[0031] [Embodiment 6] The sixth embodiment is an example of constructing a machine learning model used in an inspection device that inspects an object for failure or abnormality based on waveform data generated by the object. In the sixth embodiment, the object to be inspected is a predetermined production facility, and predetermined facility data generated by the facility is used as the waveform data. Figure 20(a) shows the equipment data waveform when the equipment is operating normally from time t0 to t1, and the waveform changes gradually. On the other hand, Figure 20(b) shows the equipment data waveform when a fault occurs in the equipment just before time t1, and the fault occurs after the waveform fluctuation (differential) becomes large. In other words, the waveform trends are different when the equipment is normal and when a fault occurs, so it is possible to predict faults using an AI model.

[0032] As shown in Fig. 21, the AI ​​model performs machine learning using equipment data waveforms during normal operation and when a fault occurs, and performs fault prediction based on the actual equipment data waveform and the learning results. However, since equipment faults are rare, learning is performed using an artificial fault waveform, which is an equipment data waveform during an artificial fault, as in the first embodiment. Fig. 22(a) is an example of a feature distribution when the feature amount of the equipment data waveform is the maximum value of the differential of the waveform data. In this example, the maximum value of the differential of the waveform data is equal to or less than a predetermined value during normal operation and exceeds the predetermined value when a fault occurs, so the predetermined value is set as the boundary, and the range below the predetermined value is set as the OK range, and the range above the predetermined value is set as the NG range.

[0033] The distribution of the number of samples (relative frequency) within the NG range may be constant for each waveform data differential maximum value, or may be weighted more heavily as it approaches the boundary. As a method for updating the distribution of the waveform data differential maximum value (distribution of feature quantities) when a fault occurs, a method for adjusting a parameter a for changing the weight of the number of samples near the boundary and a method for adjusting a parameter b for changing the NG range are used, as in the first embodiment, as shown in FIG. 22(b). In machine learning, learning is performed by randomly generating an artificial fault waveform so that the increase and decrease are gradual in the time period t0 to t' and the increase and decrease are gradual in the time period t' to t (FIG. 22(c), FIG. 22(d)). This allows the artificial fault waveform to be closer to the equipment data waveform when a fault occurs. The flow of the learning model construction method is the same as that of the first embodiment shown in Fig. 9, and it is determined whether there is an oversight or false report in the failure evaluation based on the AI ​​model, and the distribution of the feature amount is updated so that there are no oversights or false reports. As a result, it is possible to reduce oversights and false reports of equipment failures in a machine learning model used in an inspection device that inspects production equipment failures based on waveform data generated by the production equipment.

[0034] Other Embodiments The above describes an embodiment for carrying out the present invention, but the specific configuration of the present invention is not limited to the configuration of the embodiment, and design changes and the like that do not deviate from the gist of the invention are also included in the present invention. For example, in the first to fifth embodiments, the object to be inspected is a piston, but the object to be inspected is not limited to a piston. In the sixth embodiment, an example was shown in which the object to be inspected was production equipment, but the object to be inspected may be other equipment, products, or devices. The learning results are not limited to neural networks or deep learning, but can be any machine learning results. [Explanation of symbols]

[0035] 1...visual inspection device, 2...camera, 5...piston, 7...memory, 8...inspection unit

Claims

1. A method for constructing a learning model for use in an inspection device that inspects an object for defects, comprising: generating an artificial defect based on a feature amount of a defect to be inspected, a range of the feature amount for generating an artificial defect, and a distribution of the generation amount of the artificial defect with respect to the feature amount in the range; Constructing a learning model that learns the artificial defect; Obtaining a result of overlooking a defective product within a range of the defect feature amount or overdetecting a non-defective product outside the range from the inspection of the object using the learning model; updating the distribution based on the result to construct the learning model; How to build a learning model.

2. The method for constructing a learning model according to claim 1, updating the distribution increases the amount of the artificial defects generated near the boundary of the feature amount; How to build a learning model.

3. The method for constructing a learning model according to claim 2, The updating of the distribution includes updating a gradient of the generation amount of the artificial defects with respect to a change in the feature amount so that the generation amount of the artificial defects near a boundary of the feature amount increases. How to build a learning model.

4. The method for constructing a learning model according to claim 2, The distribution is updated so that the amount of generated artificial defects decreases as the amount of generated artificial defects increases away from the boundary of the feature amount. How to build a learning model.

5. Embodiment The method for constructing a learning model according to claim 1, The distribution is updated based on the result to increase or decrease the amount of artificial defects generated in a specific feature amount so as to reduce the oversight of the defective product or the overdetection of the non-defective product. How to build a learning model.

6. The method for constructing a learning model according to claim 2, The updating of the distribution includes expanding or shrinking a range of the feature quantity that generates the artificial defect based on the result so as to reduce overlooking of the defective product or overdetection of the non-defective product. How to build a learning model.

7. The method for constructing a learning model according to claim 6, In updating the distribution, the range of the feature quantity for generating the artificial defect is expanded to the non-defective product side. How to build a learning model.

8. The method for constructing a learning model according to claim 1, The learning model is a method for constructing a learning model for use in a visual inspection device, The artificial defect is an artificial defect image obtained by combining an image of an artificial defect with an image of a non-defective product, The feature amount is at least one of the brightness, size, and shape of the image of the artificial defect.

9. An inspection apparatus that inspects an object for defects using a learning model, comprising: An inspection device that uses the learning model construction method according to claim 1 as the learning model construction method.

10. 1. A method for constructing a learning model used in an inspection device that inspects an object for failure based on waveform data emitted by the object, comprising: A generation of an artificial fault waveform, which is waveform data at the time of an artificially generated fault occurrence, is performed based on a feature of a fault to be inspected, a range of the feature for generating the artificial fault waveform, and a distribution of the generation amount of the artificial fault waveform for the feature in the range; Constructing a learning model that learns the artificial defect waveform; Obtaining a result of overlooking a fault occurring within a range of a feature amount of the artificial fault waveform or an overdetection of a non-fault occurring outside the range from an inspection of the object to be inspected using the learning model; updating the distribution based on the result to construct the learning model; How to build a learning model.

Images