Host: The Japan Society Cookery Science
Name : 2021Annual Meeting of The Japan Society Cookery Science 2021
Location : [in Japanese]
Date : September 07, 2021 - September 08, 2021
[Purpose]
The handling of hand tools, including the operation of knives, requires skill. The purpose of this study is to investigate the factors that contribute to the acquisition of knife skills. We have been collecting and analyzing detailed time-series data on knife manipulation in order to find out the factors. In the 2019 report, we performed machine learning using support vector machines to estimate the factors that affect the mastery of knife manipulation, and selected parameters corresponding to the pitch of knife raising and lowering, the roll of knife twisting, and the blurring of the blade edge. In this paper, we report on the implementation of more effective machine learning and the features revealed by the analysis of the classifier.
[Methods]
From the records of knife operations by 6-axis motion sensors (3-axis angular velocity, 3-axis acceleration), we calculated a total of 234 feature points, including the mean value, standard deviation, quartiles, and frequency characteristics of each. Based on these feature values, a principal component analysis was performed to aggregate the dimensions. We then used SVM (Support Vector Machine), a machine learning method, to classify the learners and experts. The classifiers and factor loadings were then analyzed.
[Results and Discussion]
When we increased or decreased the number of principal components used for training and classified the learners and experts, the third and later components did not contribute to the classification. The third and later components did not contribute to the classification. When we checked the first and second principal components, we found that the first component had a large score for motion in the cutting direction, while the second component had a positive score for the blurring component and a negative score for the cutting direction.
When the support vectors were analyzed, the factor depth of the first component for the learners was negative, and the percentage was positive for the skilled users, but the second component was negative only for the skilled users in many cases.
These results indicate that the features that make up the first principal component are factors of proficiency, while the second principal component may be a factor of immaturity.
This research was supported by JSPS Grant-in-Aid for Scientific Research JP17K19942.
