We propose a technique for detection of surface defects on steel sheet products by texture analysis using Gabor filters considering the defect orientation. This technique enables detection of defects with a low signal-to-noise ratio on inspection images, which are difficult to detect by conventional methods. In this technique, texture features are extracted by a group of Gabor filters with multiple scales and orientations. By statistically analyzing the extracted texture features, it is possible to detect defects as pixels with statistically abnormal texture features. The abnormality of these texture features is evaluated by Mahalanobis' distance. In this technique, sub-band decomposition adapted to the defect orientation is introduced in Gabor filtering to improve sensitivity to linear defects, which are elongated in the longitudinal (rolling) direction of steel sheets and common in steel sheet products. Experiments in which this technique was applied to images of real defects on galvanized steel sheets confirmed that sensitivity was improved compared to the case without sub-band decomposition.

In this paper, we developed and evaluated the effectiveness of a line-tracing robot teaching material using reinforcement learning for students in the Department of Mechanical Engineering. Currently, most AI and reinforcement learning materials are based on simulations.However, it is important to have teaching materials that allow students to learn not only how to simulate, but also how to adapt it to the actual environment. In this study, taking a line-trace robot as an example, we developed teaching materials that also allow students to learn about the reality gap that occurs between the simulated and real environments. In order to evaluate the teaching materials, we verified through lessons using the developed line-tracing robot whether the teaching materials were comprehension and whether they were effective in promoting understanding of Q-learning. An analysis of the questionnaire conducted after the lecture confirmed developed material was effective in promoting understanding of the reality gap.

In social systems that are hierarchically constructed from various decision-makers, consensus building is becoming more difficult due to the complex relationships among various systems, and decision support by information technology is required. In social simulation, which is the technology we are focusing on, models that can reflect the intentions of various stakeholders are required based on the concept of multi-stakeholder processes. However, conventional modeling methods are applied independently at each level, making it difficult to achieve both a macro evaluation of the entire system and a micro evaluation of the detailed parts. Therefore, this paper proposes a multi-scale modeling method that can model each system component at the spatio-temporal level. Specifically, we propose a method to organize decision-makers in a social system, construct a decision model on an appropriate spatio-temporal scale, and then integrate them based on two types of relationships among decision-makers: codes of conduct as their constraint relationship and social consciousness shared by the entire society. As a case study, we apply the proposed method to the COVID-19 policy, and verify through computer experiments that the proposed model structure is capable of representing stakeholder decision-making regarding COVID-19.

Trajectory-tracking control is an essential technology for the automatic operations of automobiles. However, convergence performance heavily depends on the velocity of the target trajectory. This study introduces a design method for a trajectory tracking control law with a preliminarily defined convergence performance independent of the target velocity for a target trajectory with a variable velocity. Moreover, the effectiveness of the proposed method is confirmed by computer simulation.