Corrosion amount prediction device

Abstract

To provide a corrosion amount prediction device capable of more quantitatively evaluating progression of rust of a rusted part from an image of the rusted part of a prediction target object.SOLUTION: A corrosion amount prediction device of the present invention is used for predicting an amount of corrosion of a rusted part of a prediction target object, and comprises a prediction unit configured to predict the amount of corrosion by acquiring the amount of corrosion from an image of the rusted part of the prediction target object using a prediction model, where the prediction model is a learning model constructed by performing machine learning using teacher data consisting of sets of images and corrosion amounts of the rusted part of the prediction target object.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to a corrosion amount prediction device that predicts the amount of corrosion of a rusted portion of a prediction object. [Background technology]

[0002] 2. Description of the Related Art In recent years, systems and devices have been developed that evaluate the state of rust on an object made of steel or the like from an image of the object.

[0003] As such a system or device, for example, an evaluation system that uses a captured image of the evaluation object to evaluate the degree of rust of the evaluation object is provided with an image acquisition unit that acquires the captured image, a correction unit that generates an evaluation image by correcting the captured image, an evaluation unit that performs evaluation based on the evaluation image, and an output unit that outputs the evaluation result by the evaluation unit, and the correction unit extracts an evaluation area, which is an image of a range having a predetermined area on the surface of the evaluation object from the captured image, and generates the evaluation image based on the evaluation area (Patent Document 1). Also, an information processing device is provided with an acquisition unit that acquires an image of a target distribution facility, a discrimination unit that identifies the rust area from the image acquired by the acquisition unit based on a first learning model that has learned the image of the distribution facility and the rust area of ​​the distribution facility in the image, and a judgment unit that judges the degree of deterioration of the target distribution facility from the data indicating the rust area identified by the discrimination unit based on a second learning model that has learned the data indicating the rust area and the deterioration degree of the distribution facility (Patent Document 2). [Prior art documents] [Patent documents]

[0004] [Patent Document 1] JP 2019-144013 A [Patent Document 2] Patent Publication No. 2021-086379 Summary of the Invention [Problem to be solved by the invention]

[0005] Conventional systems and devices for evaluating the occurrence of rust were unable to predict quantitative information such as height, which is an indicator of the progress of rust in a rusted part, from an image of the rusted part of an object. For this reason, there was a demand for a more quantitative evaluation of the progress of rust from an image of the rusted part of an object.

[0006] The present invention has been made in consideration of these points, and its purpose is to provide a corrosion amount prediction device that can more quantitatively evaluate the progression of rust in a rusted portion from an image of the rusted portion of the prediction object. [Means for solving the problem]

[0007] In order to solve the above problem, the corrosion amount prediction device of the present invention is a corrosion amount prediction device that predicts the amount of corrosion of a rusted portion of a prediction object, and is characterized in that it has a prediction unit that predicts the amount of corrosion by acquiring the amount of corrosion from an image of the rusted portion of the prediction object by using a prediction model, and the prediction model is a learning model constructed by performing machine learning using teacher data that is a set of an image of the rusted portion of the prediction object and the amount of corrosion. Effect of the Invention

[0008] According to the present invention, the progress of rust at a rusted portion of a prediction object can be evaluated more quantitatively from an image of the rusted portion. [Brief description of the drawings]

[0009] [Figure 1] FIG. 2A is a diagram illustrating a corrosion-degree prediction device according to an embodiment of the present invention, and FIG. 2B is a diagram illustrating a configuration of a computer that realizes the corrosion-degree prediction device according to an embodiment of the present invention. [Diagram 2] 1A is a flowchart of a method for constructing a corrosion-degree prediction device according to an embodiment, and FIG. 1B is a flowchart of machine learning in the method for constructing a corrosion-degree prediction device according to an embodiment. [Diagram 3]5A and 5B are diagrams illustrating the preparation of teacher data in a method for constructing a corrosion degree prediction device according to an embodiment. [Figure 4] FIG. 10 is a diagram illustrating a method for expanding training data in teacher data according to an embodiment. [Diagram 5] 1A is a flowchart of a method for predicting the amount of corrosion of a rusted portion of a prediction object according to one embodiment, and FIG. 1B is a flowchart of predicting the amount of corrosion using a method for predicting the amount of corrosion of a rusted portion of a prediction object according to one embodiment. [Figure 6] FIG. 13 is a scatter plot showing prediction errors for 26 images of rusted areas to be predicted. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0010] Hereinafter, an embodiment of the corrosion amount prediction device of the present invention will be described.

[0011] [One embodiment] First, an outline of the corrosion amount prediction device according to the embodiment will be described by exemplifying the corrosion amount prediction device according to the embodiment. FIG. 1(a) is a diagram showing a schematic diagram of the corrosion amount prediction device according to the embodiment, and FIG. 1(b) is a diagram showing a schematic diagram of a computer for realizing the corrosion amount prediction device according to the embodiment. FIG. 2(a) is a flowchart of a method for constructing a corrosion amount prediction device according to the embodiment, and FIG. 2(b) is a flowchart of machine learning in the method for constructing a corrosion amount prediction device according to the embodiment. FIG. 3(a) and FIG. 3(b) are diagrams for explaining the preparation of teacher data in the method for constructing a corrosion amount prediction device according to the embodiment. FIG. 4 is a diagram for explaining a method for expanding training data in teacher data according to the embodiment. FIG. 5(a) is a flowchart of a method for predicting the amount of corrosion of a rusted portion of a prediction object according to the embodiment, and FIG. 5(b) is a flowchart of prediction of the amount of corrosion in the method for predicting the amount of corrosion of a rusted portion of a prediction object according to the embodiment.

[0012] (Corrosion amount prediction device) As shown in Fig. 1(a), a corrosion amount prediction device 1 according to one embodiment is a device that predicts the maximum height Hmax (amount of corrosion) of a rusted portion of an electrodeposition coated steel sheet (prediction target) after corrosion. The corrosion amount prediction device 1 includes an input unit 2, a memory unit 4, a processing unit 6 including a prediction unit 6a and a calculation unit 6b, and an output unit 8. The corrosion amount prediction device 1 is realized, for example, by a computer 100 shown in Fig. 1(b). The computer 100 includes a processing device (calculator) 110, a storage device 120, an input device 130, an output device 140, an input / output interface (I / F) 150, and the like.

[0013] The input unit 2 of the corrosion amount prediction device 1 inputs information such as images of rusted parts of the electrodeposition-coated steel sheet after corrosion. The memory unit 4 stores a prediction model (learning model) 4a, as well as other data 4b including teacher data described later, images of rusted parts, and a predicted value of the maximum height Hmax of the rusted parts. The prediction unit 6a of the processing unit 6 predicts the maximum height Hmax of the rusted part to be predicted by acquiring the maximum height Hmax of the rusted part (objective variable) from the image (explanatory variable) of the rusted part to be predicted of the electrodeposition-coated steel sheet after corrosion by using the prediction model 4a. The prediction model 4a is a learning model constructed by performing machine learning using teacher data that is a set of the image of the rusted part of the electrodeposition-coated steel sheet after corrosion and the maximum height Hmax. The output unit 8 outputs the prediction result (predicted value of the maximum height Hmax of the rusted part).

[0014] (Method of constructing a corrosion amount prediction device) In the method for constructing a corrosion amount prediction device according to an embodiment, first, as shown in FIG. 2(a), teacher data is prepared. In the preparation of teacher data, first, as shown in FIG. 3(a), in an image of the surface of an electrodeposition-coated steel sheet sample after corrosion, a plurality of rusted parts to be the subject of machine learning are selected, and images of a plurality of rectangular regions of the same shape (rectangles whose length and width are parallel to the x direction and y direction of the xy plane parallel to the surface of the electrodeposition-coated steel sheet) containing each of the selected rusted parts are cut out. Then, a process is performed to convert the images of the cut-out regions into a plurality of images with a specified pixel ratio (a pixels × b pixels) without changing the aspect ratio, and the converted images are obtained as images of the rusted parts of a plurality of sets of teacher data. Next, as shown in FIG. 3(b), using a VR-5000 manufactured by Keyence Corporation, the height H [mm] of the rusted parts at each point on the xy plane parallel to the surface of the electrodeposition-coated steel sheet is measured in the rusted part imaged region on the surface of the electrodeposition-coated steel sheet where the image of the rusted part of each set of teacher data is captured. Then, the maximum value of the height H of the rust part in the rust part image capture area is obtained as the maximum height Hmax [mm] of the rust part, which is the correct label of the teacher data of each set. The electrodeposition coated steel sheet comprises a steel sheet, a zinc protective film provided on the surface of the steel sheet, and an electrodeposition coating film provided on the surface of the zinc protective film. The height H of the rust part at each point on the xy plane is the height of the rust part with the surface of the electrodeposition coated steel sheet (the surface of the electrodeposition coating film) as the reference (H=0). In this way, multiple sets of teacher data, which are pairs of images of the rust part and the maximum height Hmax of the rust part (correct label), are obtained and stored in the memory unit. In this way, multiple sets of teacher data are prepared.

[0015] Next, as shown in FIG. 2(a), machine learning is performed using multiple sets of teacher data by using ResNet (Residual Neural Networks), which is a type of deep learning as a machine learning method. In machine learning, first, teacher data is read as shown in FIG. 2(b). Specifically, multiple sets of teacher data stored in a storage unit are read. Next, the teacher data is divided. Specifically, the multiple sets of teacher data are divided into multiple sets of training data and multiple sets of validation data. Next, as shown in FIG. 2(b), learning is repeated to update the learning model for the teacher data including multiple sets of training data and multiple sets of validation data until the number of epochs (learning times) reaches a predetermined number.

[0016] In this case, as shown in FIG. 2(b), for each epoch (one learning), first, training is performed by having the learning model learn all the sets of training data in a predetermined iteration. In training, all the sets of training data are divided into multiple mini-batches, and the training data belonging to each mini-batch is trained in the learning model in order for the multiple mini-batches. In learning the training data belonging to each mini-batch, training data expansion, forward propagation, error calculation, back propagation, and weight update are performed in this order. Specifically, in training data expansion, as shown in FIG. 4, multiple pseudo-new rust images are generated from the rust image (reference image) of each training data belonging to each mini-batch by expansion methods such as blurring, cutting, brightness change, left-right inversion, up-down inversion, and sharpening. Then, multiple new sets of rust image and maximum height Hmax of rust are created by combining each of the multiple pseudo-new rust images with the maximum height Hmax (correct label) of the rust of the training data of the reference image, and these multiple sets are also used as training data to increase the variation of the training data and enable efficient training. In forward propagation, the prediction section of the processing section uses a learning model to predict the maximum height Hmax of the rusted areas by obtaining the maximum height Hmax of the rusted areas (objective variable) from the image of the rusted areas (explanatory variable) of each training data. In error calculation, the maximum height Hmax of the rusted areas predicted from the image of the rusted areas of each training data (predicted value) is compared with the maximum height Hmax of the rusted areas of each training data (correct label), and a prediction error is calculated by subtracting the correct label (actual value) from the predicted value. In back propagation and weight update, the prediction error calculated for each training data is reflected in the learning model, and parameters such as the weights of the learning model are updated so that the prediction error is reduced. In such training, the learning model after training is obtained by having the learning model learn the training data belonging to each mini-batch in order for multiple mini-batches.

[0017] Next, as shown in FIG. 2(b), the learning model after training is verified using all sets of validation data. In the validation, validation data prediction, error calculation, and accuracy calculation are performed in this order. In the validation data prediction, the prediction unit of the processing unit uses the learning model to predict the maximum height Hmax of the rust by acquiring the maximum height Hmax of the rust from the image of the rust in each validation data (explanatory variable). In the error calculation, the maximum height Hmax of the rust predicted from the image of the rust in each validation data (predicted value) is compared with the maximum height Hmax of the rust in each validation data (correct label), and a prediction error is calculated by subtracting the correct label (actual value) from the predicted value. In the accuracy calculation, the MAE (Mean Absolute Error) is calculated as an accuracy index from the calculation result of the prediction error of all sets of validation data. In addition to the MAE, for example, RMSE (Root Mean Squared Error), RMSE of relative error, MAE of relative error, etc. may be calculated as an accuracy index.

[0018] Next, as shown in FIG. 2(b), the prediction model is updated. In updating the prediction model, for each epoch (one learning), it is confirmed whether the MAE of the accuracy index is smaller than the MAE in the previous epoch (learning), and if it is smaller, the prediction model is updated with the learning model after training. As described above, learning is repeated until the number of epochs reaches a predetermined number. In the process of repeating the learning, the prediction model is updated with the best learning model after training for each epoch (learning) in which the best learning model after training is obtained up to that point. As a result, when learning is repeated until the number of epochs reaches a predetermined number, the best learning model after training is obtained as the prediction model until the number of epochs reaches the predetermined number. Machine learning is performed as described above. In machine learning, the learning model after training is verified using verification data prepared separately from the training data, making it possible to distinguish learning (overlearning) specific to the training data, and overlearning of the prediction model is suppressed by using the learning model after training with the best accuracy index in the verification data as the prediction model.

[0019] Next, as shown in Fig. 2(a), the prediction model is stored. In storing the prediction model, the prediction model (the best learned model after training) is stored in the storage unit. In this manner, the corrosion amount prediction device is constructed by the construction method according to one embodiment.

[0020] (Method for predicting the amount of corrosion of the rusted part of the predicted object) In a method for predicting the amount of corrosion of a rusted part of a prediction target according to an embodiment, a corrosion amount prediction device according to an embodiment is used to predict the maximum height Hmax (amount of corrosion) of the rusted part of a prediction target of an electrodeposition-coated steel plate (prediction target) after corrosion. In this case, first, an image of the rusted part is prepared as shown in FIG. 5(a). In preparing the image of the rusted part, the rusted part of the prediction target is selected in an image of the surface of the electrodeposition-coated steel plate after corrosion, and an image of the area including the selected rusted part is cut out. At this time, the area from which the image is cut out is a rectangle of the same shape as the area from which the image of the rusted part of the teacher data is cut out. Then, a process is performed to convert the image of the cut out area into an image of a specified pixel ratio (a pixels × b pixels) without changing the aspect ratio, and the converted image is acquired as an image of the rusted part of the prediction target and stored in a memory unit. In this way, an image of the rusted part of the electrodeposition-coated steel plate after corrosion is prepared as the prediction target.

[0021] Next, as shown in FIG. 5(a), the amount of corrosion is predicted. In predicting the amount of corrosion, first, an image of the rusted part is read, as shown in FIG. 5(b). Specifically, an image of the rusted part to be predicted that is stored in the memory part is read. Next, forward propagation is performed. Specifically, the prediction part of the processing part predicts the maximum height Hmax [mm] of the rusted part to be predicted by acquiring the maximum height Hmax (objective variable) of the rusted part from the image of the rusted part to be predicted (explanatory variable) by using a prediction model (learning model). Next, the prediction result is output. Specifically, the output part outputs the predicted value of the maximum height Hmax of the rusted part.

[0022] Next, the prediction result is stored as shown in Fig. 5(a). In storing the prediction result, the predicted value of the maximum height Hmax of the rusted portion is stored in the storage unit. As described above, the maximum height Hmax of the rusted portion to be predicted is predicted by the method for predicting the amount of corrosion of the rusted portion according to one embodiment.

[0023] (effect) As described above, in the corrosion amount prediction device according to the embodiment, the maximum height Hmax of the rust part can be predicted from the image of the rust part of the electrodeposition-coated steel plate (prediction object) after corrosion by using a prediction model constructed by performing machine learning using teacher data that is a set of an image of the rust part of the electrodeposition-coated steel plate after corrosion of a sample and the amount of corrosion. Therefore, the progress of rust in the rust part of the electrodeposition-coated steel plate after corrosion can be easily detected from an image of the rust part that has slightly occurred on the surface of the electrodeposition-coated steel plate, so that the progress of rust in the rust part can be detected at an early stage. Furthermore, the progress of rust in the rust part of the electrodeposition-coated steel plate after corrosion can be inspected without destructive inspection or contact, and the progress of rust in the rust part can be inspected quickly. Therefore, according to the corrosion amount prediction device according to the embodiment, as in the above embodiment, the amount of corrosion, such as the maximum height of the rust part, which is quantitative information that serves as an index of the progress of rust in the rust part, can be predicted from the image of the rust part of the prediction object. In other words, the progress of rust in the rust part can be more quantitatively evaluated from the image of the rust part of the prediction object. Furthermore, since the progress of rust in the rusted portion of the predicted object can be easily detected from an image of the rusted portion that has slightly occurred on the surface of the predicted object, the progress of rust in the rusted portion can be detected at an early stage. Furthermore, the progress of rust in the rusted portion of the predicted object can be inspected without destructive inspection or contact, and the progress of rust in the rusted portion can be inspected quickly.

[0024] [Details of the embodiment] Next, the corrosion degree prediction device according to the embodiment will be described in more detail.

[0025] The corrosion amount prediction device is not particularly limited as long as it is a device that predicts the amount of corrosion of the rusted part of the prediction target. The prediction target is not particularly limited as long as the rusted part occurs due to corrosion of the constituent material, and the amount of corrosion of the rusted part can be predicted by acquiring the amount of corrosion of the rusted part from an image of the rusted part using a prediction model. For example, members such as steel plate, galvanized steel plate, and electrocoated steel plate are listed. Among them, members used in vehicles are preferable, and members used under the floor of the vehicle are particularly preferable. The rust of the vehicle members, especially the rust of the members under the floor of the vehicle, greatly affects the durability and safety of the vehicle, but it is difficult to detect and quantitatively measure the progress of the rust, and it is also difficult to plan preventive measures, so that the effect of the corrosion amount prediction device can be obtained significantly. Specifically, the corrosion amount, which is a quantitative index of the progress of rust in the rusted part, can be predicted from the image of the rusted part of the member by using the prediction model, so that the progress of rust in the rusted part of the member under the floor of the vehicle can be detected at an early stage by a quick inspection, and the durability of the vehicle can be improved by performing appropriate maintenance and repair of the vehicle. In other words, the corrosion amount prediction device makes it possible to predict the amount of corrosion from images of rusted areas, which was previously difficult, making it possible to manage the components of the vehicle, particularly those under the floor, and improving the safety of the vehicle.

[0026] The galvanized steel sheet is a steel sheet comprising a steel sheet and a zinc protective film (plating film) provided on the surface of the steel sheet, while the electrodeposition-coated steel sheet is a steel sheet comprising a steel sheet, a zinc protective film provided on the surface of the steel sheet, and an electrodeposition coating film provided on the surface of the zinc protective film.

[0027] The amount of corrosion of the rusted portion of the prediction object is not particularly limited as long as it is quantitative information that is an index of the progress of rust in the rusted portion of the prediction object, and examples thereof include the maximum height of the rusted portion, the raised volume of the rusted portion, etc. The maximum height of the rusted portion is the maximum value of the height of the rusted portion based on the surface of the prediction object, and for example, in the case where the prediction object is an electrodeposition-coated steel plate as in the above embodiment, it is the maximum value of the height of the rusted portion based on the surface of the electrodeposition coating. The raised volume of the rusted portion is the volume of the region above the surface of the prediction object in the raised portion of the rusted portion, and for example, in the case where the prediction object is an electrodeposition-coated steel plate as in the above embodiment, it is the volume of the region above the surface of the electrodeposition coating in the raised portion of the rusted portion.

[0028] The prediction unit of the corrosion amount prediction device is not particularly limited as long as it predicts the amount of corrosion by acquiring the amount of corrosion from an image of the rusted portion of the prediction object by using a prediction model. The image of the rusted portion of the prediction object is not particularly limited as long as it is an image from which the amount of corrosion of the rusted portion can be acquired by using a prediction model, and for example, as in the above embodiment, an image in which a region including the rusted portion is cut out from an image obtained by capturing the surface of the prediction object can be included.

[0029] The prediction model is not particularly limited as long as it is a learning model constructed by performing machine learning using teacher data that is a set of an image of the rusted part of the prediction target and the amount of corrosion, but examples of the learning model include learning models constructed by performing machine learning using deep learning such as ResNet (Residual Neural Networks), Convolutional Neural Network (CNN), GoogleNet, SVM (Support Vector Machine), Random Forest, etc. Among them, learning models constructed using deep learning, etc. are preferred, and learning models constructed using ResNet, etc. are particularly preferred. EXAMPLES

[0030] Hereinafter, the corrosion degree prediction device according to the embodiment will be described more specifically with reference to examples.

[0031] [Example] As the corrosion amount prediction device of the embodiment, an example of the corrosion amount prediction device according to the embodiment was constructed by an example of the construction method of the corrosion amount prediction device according to the embodiment. In this case, first, in the preparation of the teacher data, a corrosion test was carried out to simulate the corrosion including rust erosion on a sample of an electrodeposition-coated steel plate simulating a member under the floor of a vehicle, and the electrodeposition-coated steel plate after corrosion (prediction target) was obtained, and teacher data of 196 sets of images of rusted parts (explanatory variables) and the maximum height Hmax [mm] of the rusted parts (objective variable) was prepared from the electrodeposition-coated steel plate after corrosion. Next, in the machine learning, first, in the division of the teacher data, the 196 sets of teacher data were divided into 146 sets of training data and 50 sets of verification data. Next, learning was repeated for the teacher data including the 146 sets of training data and the 50 sets of verification data until the number of epochs (learning times) reached 1000 times (predetermined number of times).

[0032] In this case, for each epoch (one learning session), first, training was performed by having the learning model learn all the sets of training data in a specified iteration. In the training, all the sets of training data were divided into multiple mini-batches, and the training data belonging to each mini-batch was trained into the learning model in order for each of the multiple mini-batches. In learning the training data belonging to each mini-batch, training data expansion, forward propagation, error calculation, back propagation, and weight update were performed in this order. In addition, in the training data expansion, multiple pseudo-new images of rust areas were generated under the following conditions from the images of rust areas (reference images) of each training data belonging to each mini-batch using the expansion methods of blurring, clipping, brightness change, left-right inversion, up-down inversion, and sharpening.

[0033] (Conditions for extension measures) Blur... Randomly select a target set from the training data and perform the blurring under the following conditions: kernel size in x direction: 3, kernel size in y direction: 3, standard deviation in x direction: 3, and standard deviation in y direction: 3. If the standard deviation in the x and y directions is small, the influence on the surrounding pixels will be large, making the image smoother, so the strength of the blurring is adjusted according to the standard deviation. · Cropping...A target set is randomly selected from the training data. The image is cropped by deleting 1 to 10 pixels from the vertical and horizontal edges of the base image. The cropping location is set randomly. The aspect ratio of the cropped image is made the same as that of the base image. Brightness change: This is performed on one out of every two sets of training data. The brightness of the reference image is changed by a factor of 0.8 to 1.2. Left-right flip: This is performed on one out of every two sets of training data. · Upside down flip · · · This is performed on one out of every two sets of training data. Sharpening: This is performed on a random set of subjects from the training data. The edges of various color regions in the reference image are made twice as bright.

[0034] In such training, the training data belonging to each mini-batch was trained in the learning model in order for a plurality of mini-batches, thereby obtaining a learning model after training. Next, the learning model after training was verified using all sets of verification data. Next, the prediction model was updated. As described above, learning was repeated until the number of epochs reached 1000. Machine learning was performed as described above. Next, in the case of saving the prediction model, the prediction model (the best learning model after training) was stored in the storage unit.

[0035] [evaluation] An example of the method for predicting the amount of corrosion of the rusted portion according to the above embodiment was carried out, and the prediction accuracy of the corrosion amount prediction device constructed in the embodiment was evaluated. In this case, when the prediction method was carried out, in preparation of images of the rusted portions, 26 images of the rusted portions of the electrodeposition-coated steel plate (prediction object) of the vehicle after corrosion were prepared. Then, in predicting the amount of corrosion, for each image of the rusted portion to be predicted, the prediction unit of the processing unit of the corrosion amount prediction device constructed in the embodiment used a prediction model to obtain the maximum height Hmax (objective variable) of the rusted portion from the image of the rusted portion to be predicted (explanatory variable) to predict the maximum height Hmax [mm] of the rusted portion.

[0036] Furthermore, to evaluate the prediction accuracy, for the 26 rusted parts to be predicted, the height H [mm] of the rusted parts was measured at each point on the xy plane parallel to the surface of the electrodeposition-coated steel sheet in the rusted part image area on the surface of the electrodeposition-coated steel sheet where the image of each rusted part was captured, and the maximum value of the height H of the rusted parts was calculated as the actual measured value [mm] of the maximum height Hmax of the rusted parts. Then, for the 26 rusted parts to be predicted, the predicted value and the actual measured value of the maximum height Hmax of the rusted parts were compared, and the prediction error [mm] was calculated by subtracting the actual measured value from the predicted value. Figure 6 is a scatter diagram showing the prediction error for the 26 rusted parts to be predicted. Furthermore, from the calculation results of all prediction errors, the following accuracy indices were further calculated: R2 (coefficient of determination), RMSE (Root Mean Squared Error), MAE (Mean Absolute Error), sigma (standard deviation of absolute error), RMSE_ratio (RMSE of relative error), MAE_ratio (MAE of relative error), sigma_ratio (standard deviation of relative error), acc_thres0.025 (proportion of inferred data with absolute error of 0.025 mm or less), acc_thres0.05 (proportion of inferred data with absolute error of 0.05 mm or less), acc_thres0.1 (proportion of inferred data with absolute error of 0.1 mm or less), and acc_thres0.25 (proportion of inferred data with absolute error of 0.25 mm or less). These accuracy indices are shown in Table 1 below.

[0037] [Table 1]

[0038] From the results shown in Fig. 6 and Table 1, it is believed that the corrosion amount prediction device constructed in the embodiment can predict with high accuracy the maximum height Hmax of the rusted part from the image of the rusted part of the electrodeposition-coated steel plate (prediction target) of the vehicle after corrosion. Specifically, since the MAE is about 0.1 mm, it is believed that the maximum height Hmax of the rusted part can be predicted with high accuracy with a prediction error of about 0.1 mm.

[0039] Although the embodiment of the corrosion amount prediction device of the present invention has been described in detail above, the present invention is not limited to the above embodiment, and various design modifications can be made without departing from the spirit of the present invention described in the claims. [Explanation of symbols]

[0040] 1: Corrosion amount prediction device, 2: Input section, 4: Memory section, 4a: Prediction model (learning model), 4b: Other data, 6: Processing section, 6a: Prediction section, 6b: Calculation section, 8: Output section

Claims

[Claim 1] A corrosion amount prediction device for predicting a corrosion amount of a rusted portion of a prediction object, comprising: a prediction unit that predicts the amount of corrosion by acquiring the amount of corrosion from an image of the rusted portion of the prediction object by using a prediction model; The corrosion amount prediction device is characterized in that the prediction model is a learning model constructed by performing machine learning using training data which is a set of images of the rusted portion of the prediction object and the amount of corrosion.

Images