JP7292170B2 - Parts condition determination system - Google Patents

Description

The present invention relates to a parts condition determination system for determining the condition of collected used parts.

Patent Document 1 includes a first inspection camera that detects an abnormality on the upper side of the inspection object that is conveyed, and a second inspection camera that detects an abnormality on the upper side of the inverted inspection object. A visual inspection apparatus is disclosed that ejects defective products detected based on abnormality information from both inspection cameras from a conveying lane. The presence or absence of an abnormality on the surface of an object to be inspected is determined by comparing changes in color and brightness due to foreign matter, scratches, and dirt, and changes in surface unevenness, obtained by image processing, with preset discrimination criteria. It is determined.

Patent Document 2 discloses an inspection device that images and inspects a board after each process is performed at a site where a board on which electronic components are mounted through a plurality of processes is manufactured, and communication with each inspection device is possible. A substrate inspection system including an information processing device is disclosed. In this system, a plurality of programs are stored for processing to identify the causes of defects detected in the final form of the board. , the type of anomaly that occurred and the process in which the anomaly occurred. Furthermore, the cause of the defect is specified by executing an analysis program suitable for the recognition result.

Patent Literature 3 discloses an identification device that identifies the type of bread by recognizing a color image of bread captured by a camera. The identification device includes a color data extraction unit for obtaining feature amounts in respective color spaces of the inner area, the intermediate area, and the outer area of the bread from the color image of the bread, and a feature amount related to the outline of the bread from the color image of the bread. a contour data extracting unit for obtaining bread texture, a texture data extracting unit for obtaining a feature amount related to the texture of the bread from the color image of the bread using the luminance component, and a type of bread identified by the feature amount obtained by each extracting unit. and an identification unit.

JP 2011-117866 A JP 2006-339445 A Japanese Patent No. 5510924

The devices according to Patent Literatures 1 and 2 inspect for abnormalities in products in the manufacturing process in manufacturing factories, that is, for abnormalities in new products. Therefore, the forms of possible abnormalities are limited and can be presumed. Therefore, it is possible to prepare in advance a discrimination standard and an analysis program based on the estimated abnormality, and it is possible to discriminate the abnormality of the product by using the discrimination standard and the analysis program. The identification device disclosed in Patent Document 3 identifies the type of baked bread from its photographed image, and the type of bread to be identified, that is, the feature amount is known in advance. Therefore, it is possible to identify the type of bread appearing in the photographed image by using the feature amount of bread that has been confirmed in advance as an identification reference.

On the other hand, in a discriminating device that discriminates the state (normal or abnormal) of collected used parts, the collected parts to be discriminated show various forms due to changes over time. Good results are not expected if the anomaly discrimination technique is used as it is.

Accordingly, an object of the present invention is to provide a parts condition determination system that can easily and accurately determine whether collected used parts are normal or abnormal.

A parts condition determination system for determining the condition of a used collected part according to the present invention uses, as learning data, a collected photographed image of the collected part, a new photographed image of the collected part when it is new, or both photographed images. an AI unit learned to determine whether the collected part is normal or abnormal from the collected photographed image; a determination result notification unit for externally reporting a determination result of the AI unit; and the determination result notification unit. a re-learning command unit that commands the AI unit to re-learn based on the expert's judgment of the discrimination result notified by the unit ,
The AI unit includes an input data generation unit and an abnormality determination unit,
The input data generation unit generates input image data for the abnormality determination unit based on the recovered photographed image,
the abnormality determination unit is a convolutional neural network trained to determine whether the collected parts are normal or abnormal based on the input image data;
The input data generation unit includes a first image processing module that generates the input image data from the collected photographed image by paying attention to color change of the collected part over time, and a first image processing module that generates the input image data from the collected captured image by paying attention to the change in shape of the collected part over time. It has a second image processing module that generates the input image data from the photographed image, and a third image processing module that generates the input image data from the collected photographed image by paying attention to changes in the position of the collected parts over time.

With this configuration, the AI unit, which has been trained to discriminate whether the collected parts are normal or abnormal, is re-learned based on the expert's judgment of the discrimination result, so the state discrimination can be performed for many collected parts. While performing, the discrimination function is strengthened. As a result, even if collected parts change in various forms over time, the parts condition determination system can easily and accurately determine whether the used collected parts are normal or abnormal. . The morphology that changes with aging includes aging such as color change and shape change, and deformation (physical damage) due to sudden thermal shock, collision shock, and the like.

Furthermore, in the above configuration, the input data to the convolutional neural network for determining whether the collected part is normal or abnormal is not the collected photographed image itself, but the collected photographed image generated by the input data generation unit from the collected photographed image. Since the input image data is suitable for determining the state of the component, the accuracy of the determination result is improved. For example, if a collected part becomes abnormal due to a specific phenomenon caused by aging, it would be advantageous if the input image data was generated using a process that emphasizes the phenomenon.

Collected photographed images of collected parts that have deteriorated over time, including failure (damage), have the following three changes compared to the photographed images of new collected parts: (1) discoloration, (2) shape due to contraction and expansion ( size), and (3) change in attachment position due to peeling or displacement. Of course, these changes do not occur due to failure (damage), but can occur only due to simple aging (degradation). However, the degree of change in recalled parts that have failed (broken) is generally much greater than in recalled parts that have not failed. For this reason, we focus on at least one of (1) discoloration, (2) change in shape (size) due to contraction or expansion, and (3) change in mounting position due to peeling or slippage when determining the condition of collected parts. is important .
Therefore , in the present invention , the input data generation unit includes a first image processing module that generates the input image data from the collected photographed image by paying attention to the aging color change of the collected parts; A second image processing module for generating the input image data from the collected photographed image by paying attention to the shape change of the collected part over time, and generating the input image data from the collected photographed image by paying attention to the aged positional change of the collected part. and a third image processing module for generating.
In this configuration, using input image data generated using at least one of the first image processing module, the second image processing module, and the third image processing module improves the accuracy of the determination result.

A parts condition determination system for determining the condition of a used collected part according to the present invention uses, as learning data, a collected photographed image of the collected part, a new photographed image of the collected part when it is new, or both photographed images. an AI unit learned to determine whether the collected part is normal or abnormal from the collected photographed image; a determination result notification unit for externally reporting a determination result of the AI unit; and the determination result notification unit. a re-learning command unit that commands the AI unit to re-learn based on the expert's judgment of the determination result notified by the unit, wherein the AI unit includes an input data generation unit and an abnormality determination unit wherein the input data generation unit is a convolutional neural network configured to input the recovered photographed image and output a restored image similar to the new photographed image; and the abnormality determination unit includes: Based on the degree of similarity between the restored image and the new photographed image, it is determined whether the recovered part is normal or abnormal .
According to this configuration, the input data generation unit has an image restoration algorithm for outputting a restored image similar to the new photographed image of the collected part when new (a fake new photographed image) based on the collected photographed image. Since the image restoration capability of the image restoration algorithm is limited, when this image restoration algorithm is executed on the recovered photographed image that has undergone aging, the input image represents a restored image that is slightly different from the new photographed image. Data is output. In particular, when this image restoration algorithm is executed on collected photographed images of collected parts that have undergone aging and failure (damage), input image data representing a restored image with a greater difference than photographed images of new parts. is output. By utilizing this, the abnormality determination unit can determine whether the collected parts are normal or abnormal based on the degree of similarity between the restored image (input image data representing the restored image) and the photographed new product image (input image data representing the photographed new product image). can be discriminated.

In this configuration as well, it is preferable to generate a restored image from the collected photographed image by focusing on the three types of secular change described above. Accordingly, in one of the preferred embodiments of the present invention, the input data generation unit includes a first restoration module that outputs the restored image by paying attention to aging color change of the collected parts; It has a second restoration module that outputs the restored image by paying attention to the shape change over time, and a third restoration module that outputs the restored image by paying attention to the change in the position of the collected part over time. With this configuration, by using the input image data generated using at least one of the first restoration module, the second restoration module, and the third restoration module, the abnormality determination unit outputs a determination result with higher accuracy. be able to.

In one of the preferred embodiments of the present invention, the generation unit and the abnormality determination unit, which constitute the AI unit, are constructed by an integrated convolutional neural network. As a result, the AI unit is simplified, and as a result, a parts state determination system capable of simply and accurately determining the state of collected parts is realized.

Other features, functions and effects of the present invention will be made clear by the following description of the invention using the drawings.

It is a schematic diagram explaining the basic principle of a components state determination system. BRIEF DESCRIPTION OF THE DRAWINGS It is a schematic diagram explaining the structure of 1st Embodiment of a components state determination system. It is a schematic diagram explaining the structure of 2nd Embodiment of a components state determination system. It is an explanatory view showing the flow of information in the parts state determination system of the second embodiment. FIG. 3 is a schematic diagram showing a neural network that constitutes an image restoration unit;

The basic principle of the component condition determination system will be described below with reference to FIG. Used and collected parts whose states are determined by this component state determination system are substrates built into the fuel cell power generator 1 installed at home. Various electronic components, connectors, and the like are mounted on the board. The fuel cell power plant 1, which is only schematically shown in FIG. 1, supplies electrical and thermal energy to the house. The fuel cell power generator 1 is inspected periodically or at the request of each household, and components such as substrates are replaced with new ones as necessary. The replaced boards and the like are carried into the service center C1 as collected parts. Further, even when the fuel cell power generation device 1 is discarded, substrates and the like are sent to the service center C1 as recovered parts from the dismantled fuel cell power generation device 1 . An identification code for identifying the collected parts is assigned to the received collected parts.

A parts management computer 100 that manages the parts of the fuel cell power generator 1 is installed in the service center C1. The parts management computer 100 can function as an abnormal parts discriminating device that discriminates the state of used collected parts. The parts management computer 100 functions as a parts state determination system, and includes a communication section 21 , an imaging section 22 , a data acquisition section 31 , an AI unit 10 , a determination result notification section 7 and a relearning instruction section 8 .

The data acquisition unit 31 acquires a photographed image (collected photographed image) of the collected parts created by the photographing unit 22, which is a color camera system. Further, a photographed image of the collected part when it was new (new article photographed image) is created by the parts manufacturing company C2, and the data acquisition section 31 acquires it via the communication section 21. FIG. Of course, the photographed image of the new product can also be created by the photographing section 22 . The recovered photographed image and the new photographed image acquired by the data acquisition unit 31 are stored so as to be extractable by the part identification code.

The AI unit 10 uses, as learning data, the photographed images of the collected parts, the photographed images of the collected parts when they are new, or both photographed images, and determines whether the collected parts are normal or abnormal from the photographed collected images. It is a machine learning model trained to discriminate. The AI unit 10 may be configured by being divided into the input data generation unit 5 and the abnormality determination unit 6 . At that time, the input data generation unit 5 generates appropriate input image data as input data for the abnormality determination unit 6 based on the recovered photographed image. The abnormality determination unit 6 outputs a determination result indicating whether the collected parts are normal or abnormal based on the input image data. In addition, the most basic abnormality determination unit 6 uses photographed images of normal new parts (the photographed images of used collected parts, or the photographed images of both of them) as learning data (teacher data). It can be constructed by a part state determination machine learning model generated by learning to determine an abnormality of the collected parts from the photographed images of the collected parts. By inputting a photographed image of a collected part (collected photographed image) into this machine learning model, a determination result as to whether the collected part is abnormal or normal is output.

A determination result notification unit 7 externally reports the determination result output by the AI unit 10 . By this external notification, for example, visual or auditory notification perceived by a person is performed. The relearning command unit 8 commands relearning to adjust the AI unit 10 that output such a determination result when the notified determination result is determined to be erroneous based on the expert's knowledge. .

FIG. 2 shows a first embodiment of a component condition determination system according to the present invention. In this first embodiment, the AI unit 10 consists of an input data generation unit 5 and an abnormality determination unit 6 . The input data generation unit 5 can apply multiple types of image processing to the recovered photographed image to generate multiple types of input image data for the abnormality determination unit 6 . Therefore, the input data generation unit 5 has an image processing section 51 , an image processing module selection section 52 and an image processing module storage section 53 . The image processing unit 51 uses an image processing module selected by the image processing module selection unit 52 from the image processing module group stored in the image processing module storage unit 53 to process the collected photographed image, and converts the input image data. Generate. In this embodiment, the image processing module group includes a first image processing module, a second image processing module, and a third image processing module. The first image processing module is a module for generating input image data from collected photographed images by paying attention to aging color changes of collected parts. The second image processing module is a module for generating input image data from collected photographed images by paying attention to changes in shape of collected parts over time. The third image processing module is a module for generating input image data from collected photographed images by paying attention to positional changes of collected parts due to aging.

The first image processing module generates input image data by applying image processing that emphasizes discoloration (color change over time) of the collected parts to the collected photographed image. Based on this input image data, the abnormality determination unit 6 determines the state (normal or abnormal) of the collected parts. The following symptoms are observed in recovered parts that are determined to be abnormal based on this input image data:
(a) Burning due to heat: turning brown to black,
(b) Missing component parts: become the base color,
(c) Expansion of components: contraction, deformation: changes in color balance occur.

The second image processing module generates input image data by applying image processing that emphasizes a change in size (change in shape over time) of the collected parts to the collected photographed image. Based on this input image data, the abnormality determination unit 6 determines the state (normal or abnormal) of the collected parts. The following symptoms are observed in recovered parts that are determined to be abnormal based on this input image data:
(a) thermal expansion: parts become larger;
(b) shrinkage on drying: the part becomes smaller;
(c) missing components: some dimensions are reduced;
(d) Dissolution, Elution: Target parts disappear (=part dimensions become zero).

The third image processing module generates input image data by performing image processing that emphasizes a change in the mounting position of the collected part (change in position over time) on the collected photographed image. Based on this input image data, the abnormality determination unit 6 determines the state (normal or abnormal) of the collected parts. The following symptoms are observed in recovered parts that are determined to be abnormal based on this input image data:
(a) Deformation due to contact: Mounting position shifts,
(b) Swelling of parts: beyond the installation range.

The abnormality determination unit 6 receives the input image data generated by the input data generation unit 5 and outputs a determination result as to whether the collected parts corresponding to the input image data are normal or abnormal. The input data generation unit 5 has an image preprocessing function for performing various processes on the captured image, and input image data is generated through this image preprocessing function. Such image preprocessing functions include functions for converting RGB data into grayscale data, HSV data, etc., and functions for performing image processing filters such as image reduction and edge detection. The anomaly discrimination unit 6 is composed of a convolutional neural network, preferably deep learning. The learning of the abnormality determination unit 6 is performed by the learning management section 33 . The learning data used at the time of this learning are a new photographed image for which the discrimination result is normal, a recovered photographed image for which the discrimination result is normal, and a recovered photographed image for which the discrimination result is abnormal, and determination of whether the recovered photographed image is normal or abnormal. is done artificially. Furthermore, the input image data may be a front shot image of the collected part, a side shot image of the collected part may be used as the input image data, or an obliquely shot image of the collected part may be used as the input image data. Furthermore, an abnormality determination unit 6 may be prepared for each input image data.

The determination result notification unit 7 notifies the determination result output by the AI unit 10 through a notification device (not shown), further links the determination result with the identification code of the collected part, and records it. The determination results output by the AI unit 10 are randomly checked by a skilled expert. If the expert check reveals that the abnormality determination unit 6 misidentified a normal collected part as abnormal, the relearning command unit 8 instructs the relearning command unit 8 to determine that the collected part is normal. Re-learning is requested.

Next, a second embodiment of the component condition determination system according to the present invention will be described with reference to FIGS. 3 and 4. FIG. FIG. 3 is a schematic diagram for explaining the configuration of the second embodiment of the component condition determination system. FIG. 4 is an explanatory diagram showing the flow of information in the parts condition determination system of the second embodiment.

Also in this second embodiment, the AI unit 10 is composed of the input data generation unit 5 and the abnormality determination unit 6 . In this embodiment, the input data generation unit 5 generates input image data to be input to the abnormality determination unit 6 from the recovered photographed image and the new photographed image acquired by the data acquisition section 31 . The input data generation unit 5 has various pre-processing functions for the image data. The preprocessing functions include a function of converting RGB data into grayscale data, HSV data, etc., and a function of executing image processing filters such as image reduction and edge detection. The abnormality determination unit 6 is composed of an image restoration section 6A and an abnormality determination section 6B. The image restorer 6A is constructed by a convolutional neural network, preferably deep learning, and is trained to restore images as will be described in detail below. FIG. 5 schematically shows such deep learning. The input image data is vector data consisting of its pixel values, and is indicated by Z1, Z2, . . . in FIG.

The image restoration unit 6A receives input image data generated based on the photographed image of the collected part (collected photographed image), and outputs restored image data similar to the new photographed image of the collected part when it is new. At this time, the worse the condition of the collected parts due to failure, breakage, or the like, the less similar the restored image will be to the new photographed image. When input image data based on photographed images of new products is input, restored image data showing a degree of similarity of approximately 100% is output. is input, the restored image data showing a similarity lower than 100% is output. Therefore, the abnormality determination unit 6B compares the restored image data of the collected part with the new image data based on the new photographed image of the collected part when it is new, calculates the degree of similarity, and based on this degree of similarity, determines the collected part. Determine normal or abnormal.

In this embodiment, the main state changes due to aging and failure (damage) are (1) discoloration, (2) shape (size) changes due to contraction and expansion, and (3) peeling and displacement. In order to individually check the change in the mounting position due to the We have prepared three deep learning modules. Therefore, the image restoration section 6A includes a restoration module execution section 61, a restoration module selection section 62, and a restoration module storage section 63. FIG.

The deep learning modules (or set values of weights and biases) selected from the restoration module storage unit 63 by the restoration module selection unit 62 and loaded into the restoration module execution unit 61 are a first restoration module, a second restoration module, and a second restoration module. 3 recovery module.

The first restoration module has an image restoration function of restoring a collected photographed image of a collected part having discoloration (color change over time). If the image restored by this image restoration function clearly differs from the restored image of the new product, the abnormality determination unit 6B determines that the discoloration of the recovered part is abnormal and that the discoloration is due to a part failure. Assume there is. A possible example here is:
(a) Burning due to heat: turning brown to black,
(b) Missing component parts: become the base color,
(c) expansion, contraction, deformation of components: changes in color balance occur;
is mentioned.

The second restoration module has an image restoration function that restores a collected photographed image of a collected part having a change in size (change in shape over time). If there is a clear difference between the image restored by this image restoration function and the restored image of a new product, the abnormality determination unit 6B determines that the collected part has an abnormal change in size and that the change indicates a part failure. presumed to be based on A possible example here is:
(a) thermal expansion: parts become larger;
(b) shrinkage on drying: the part becomes smaller;
(c) missing components: some dimensions are reduced;
(d) dissolution, elution: target parts disappear (=part dimensions become zero),
is mentioned.

The third restoration module has an image restoration function that restores a collected photographed image of a collected part having a change in mounting position (change in position over time). If there is a clear difference between the image restored by this image restoration function and the restored image of a new product, the abnormality determination unit 6B determines that the collected part has an abnormal change in the mounting position, and that change is a part failure. presumed to be based on A possible example here is:
(a) Deformation due to contact: Mounting position shifts,
(b) part expansion: exceeding the installation range;
is mentioned.

The abnormality determination unit 6B restores restored image data, which is a restored image of a recovered part, output from each of the first restoration module, second restoration module, and third restoration module, and new restoration data, which is a restored image of a new article ( It is also possible to generate directly from the photographed image), and it is determined whether the collected part is normal or abnormal based on the degree of similarity. At that time, the similarities using the restored image data by the first restoration module, the second restoration module, and the third restoration module are combined, for example, their added value (including weighted added value) or average value (weighted (including the average value), or if any similarity exceeds the allowable range, it may be determined to be abnormal. Furthermore, it is also possible to designate an abnormal location and perform normality/abnormality discrimination of individual components mounted on the board.

The first restoration module, the second restoration module, and the third restoration module can also be integrated into one deep learning module and incorporated in the restoration module execution unit 61 . In that case, the restoration module selection unit 62 and the restoration module storage unit 63 are omitted.

Further, in this embodiment, the plurality of restoration modules are configured to perform image restoration by aiming at different causes of abnormalities in the collected parts. However, the direction in which the collected parts are observed may be important when determining an abnormality in the collected parts. For this reason, as a restoration module, a front photographed image of the collected part is used as input image data, a side photographed image of the collected part is used as input image data, and an obliquely photographed image of the collected part is used as input image data. may be provided.

In this embodiment as well, the determination result by the abnormality determination unit 6 is checked by a skilled expert at random. If the expert check reveals that the abnormality determination unit 6 misidentified a normal collected part as abnormal, the relearning command unit 8 instructs the re-learning command unit 8 to restore the image restoration unit 6A so that the collected component is determined to be normal. Re-learning is requested. Similarly, when the expert check reveals that the abnormality determination unit 6 misidentified an abnormal collected part as normal, the re-learning command unit 8 instructs the re-learning command unit 8 to determine that the collected part is normal. Re-learning of 6A is requested.

As described above, it is possible to determine whether a collected part is normal or abnormal, and whether a specific portion of the collected part is normal or abnormal using the component state determination system according to the present invention. The service life of the collected parts can be estimated from the discrimination result of the collected parts and the mode of use of the parts (usage time, usage load, usage environment, etc.), and improvement points of the collected parts can be found.

[Another embodiment]
(1) In the above-described embodiment, the image restoration unit 6A is configured using deep learning. It is also possible to configure using an image processing module such as technology.

(2) In the above-described embodiment, the abnormality determining section 6B determines abnormality of the collected parts based on the restored image data output from the image restoring section 6A. When the input image data is integrated and the input image data is input, the normal/abnormal determination result of the collected parts can be output. Such an integrated anomaly determination unit 6 can be constructed using, for example, a generative adversarial network or an autoencoder. At that time, the hostile generation network generates a pseudo-new photographed image from the photographed image of the collected part, and determines whether the collected part is normal or abnormal based on the difference between the pseudo-new photographed image and the true photographed image of the new article, and the degree of authenticity. be done.

(3) The AI unit 10 can also be built in a mobile terminal such as a tablet computer with a camera. By using such a portable terminal, a maintenance inspector can check the normality/abnormality of the removed collected parts while patrolling each household. If the recovered part that has been removed is normal, it is also possible to restore the recovered part.

(4) In the above-described embodiment, the substrates inside the fuel cell power generator 1 installed at home were treated as collected parts. Home appliances other than the device 1, for example, home security system equipment, home appliance system equipment, cooking system equipment, air conditioning system equipment, etc. Normal/abnormal determination of substrates etc. installed in various system equipment that manages home life It can also be configured to do so.

(5) The input data generation unit 5 and the abnormality determination unit 6 can be integrated with a convolutional neural network, that is, the AI unit 10 can be configured with a single convolutional neural network.

INDUSTRIAL APPLICABILITY The present invention can be applied to a system for determining normality/abnormality of recovered parts (replacement parts) of equipment in which various substrates and component units are installed.

1: fuel cell power generator 5: input data generation unit 6: abnormality determination unit 6A: image restoration unit 6B: abnormality determination unit 51: image processing unit 61: restoration module execution unit 62: restoration module selection unit 63: restoration module storage unit 52: image processing module selection unit 53: image processing module storage unit 7: determination result notification unit 8: re-learning command unit 10: AI unit 21: communication unit 22: imaging unit 31: data acquisition unit 33: learning management unit

Claims (4)

A parts state determination system for determining the state of used collected parts,
Determining whether the collected part is normal or abnormal from the collected photographed image using the collected photographed image of the collected part, the new photographed image of the collected part when it is new, or both photographed images as learning data. an AI unit that has been trained to
a determination result notification unit that externally reports the determination result of the AI unit;
a re-learning command unit that commands the AI unit to re-learn based on an expert's judgment on the discrimination result notified by the discrimination result notification unit;
The AI unit includes an input data generation unit and an abnormality determination unit,
The input data generation unit generates input image data for the abnormality determination unit based on the recovered photographed image,
the abnormality determination unit is a convolutional neural network trained to determine whether the collected parts are normal or abnormal based on the input image data;
The input data generation unit includes a first image processing module that generates the input image data from the collected photographed image by paying attention to color change of the collected part over time, and a first image processing module that generates the input image data from the collected captured image by paying attention to the change in shape of the collected part over time. A part state having a second image processing module that generates the input image data from a photographed image, and a third image processing module that generates the input image data from the collected photographed image by paying attention to the positional change of the collected part over time. discrimination system. A parts state determination system for determining the state of used collected parts,
Determining whether the collected part is normal or abnormal from the collected photographed image using the collected photographed image of the collected part, the new photographed image of the collected part when it is new, or both photographed images as learning data. an AI unit that has been trained to
a determination result notification unit that externally reports the determination result of the AI unit;
a re-learning command unit that commands the AI unit to re-learn based on an expert's judgment on the discrimination result notified by the discrimination result notification unit ;
The AI unit includes an input data generation unit and an abnormality determination unit,
The input data generation unit is a convolutional neural network configured to input the recovered photographed image and output a restored image similar to the new photographed image;
The abnormality determination unit determines whether the collected parts are normal or abnormal based on the degree of similarity between the restored image and the photographed image of the new product. The input data generation unit includes a first restoration module that outputs the restored image by paying attention to color change of the collected part over time and a second restoration module that outputs the restored image by paying attention to shape change of the collected part over time. 3. The parts condition determination system according to claim 2, further comprising a module and a third restoration module for outputting the restored image by paying attention to changes in position of the collected parts over time. 4. The parts condition determination system according to any one of claims 1 to 3, wherein the input data generation unit and the abnormality determination unit are constructed by an integrated convolutional neural network.

Images