In construction sites, there occurs many industrial accidents. Also, with the aging of engineers and the lack of successors, worker’s burdens on jobs are increasing. Thus, it is concern about an increase in occupational accidents. In particular, it is thought that the probability of artificial mistake increases with decreasing concentration due to fatigue. Therefore, it is important to grasp states of workers constantly and to work out a safety measures. In existing research, there are some approaches to grasp the degree of fatigue and stress from blood and heart electric waves. However, these methods are not suitable for grasping the state of the operator in real time. On the other hand, with the development of sensor technology in recent years, smart watches that can measure heartbeat easily and inexpensively have attracted a public attention. If fatigue could be detected by smart watches, it would be possible to manage safety effectively. Therefore, in this research, we try to detect the degree of fatigue using heart rate data of smartwatch and verify its applicability.

Acquisition of location information using the GNSS sensor of the smartphone or tablet device identifies the user's own current location and supports movement to the destination without hesitation even in the first place. In this research, not only the GNSS sensor but also the Augmented Reality ( AR ) technology using the camera built in the mobile device establishes a technique to specify the three-dimensional position of the inspection object itself, and utilize it through actual river inspection work case examples are introduced.

In terms of research presentations, some presenters cannot put their thoughts together well or may not be confident about the contents of their presentations. With this research, we aimed to allow presenters to practice their preesenting skills using three Mugbot robots that could ask questions to presenters and praise their responses. In the results of an experiment using student participants conducting graduation research, the students did not gain confidence, but they seemed to be able to put their thoughts together.

Recently, educational-support robots, which support learning attract many people’s attentions. In previous research, the robot teaches how to solve some questions only. However, it is difficult for learners to improve their applied skill and inquiring mind, because the robot of previous research cannot prompt learners to deliberate. Thus, this study develops a robot which support based on cognitive apprenticeship. The previous study reported that teaching based on cognitive apprenticeship can prompt learners to deliberate in pedagogy. Therefore, the learner who was taught based on cognitive apprenticeship can be improved their applied skills and inquiring mind. In this paper, we investigates effects of educational support robots based on cognitive apprenticeship theory in collaborating learning with junior high school students.

The purpose of this research is to design a driving support agent that selects utterance strategy considering driver’s attributes and driving situation based on the politeness theory. In the politeness theory, utterance strategy for constructing good relationship between people is systematized. As a concrete approach, we designed agent linguistic behaviors based on the positive politeness strategies and negative politeness strategies, which are representative utterance strategies defined in the politeness theory. In the experiment, we used two kinds of movies which reflect scenes where the driver is supported by speech synthesis utterance from agents. The utterance contents of agents were designed by us. Experiment participants watched two movies in a row and evaluated impressions for agents. Here we discuss the impression evaluation results and future development. Experiment participants are university students or graduate students with a driving experience of three years or less.

Recently, it is called that Japan is a stress-filled society because various job stress are often seen in our activities at home and business. In addition, the mental health disorder may be caused by mental stress, which is produced by bad conditions for not only the mind but also the body, and also could be lead to many diseases. If we can monitor our own situation of mental stress easily, we can expect effective approaches for recognizing mental stress factors and prevention of mental stress on self-management. In this paper, we described system development to detect mental stress in real time, which informs us by means of unique material characteristic such as nasal skin temperature drops under stress condition. Furthermore, we are approaching to develop new detection system using a web camera and a thermographic camera.

For effectively and efficiently screening potential sleep-disordered patients, it is desired to establish simplified and easy-to-use techniques to measure the sleep state in a sufficient accuracy without elaborate facilities.Pressure sensor mats are a promising candidate but methods to estimate the vital signs of subjects from their sensing data are to be devised.The present paper focuses on the estimation of the heart rate from the high-dimensional time series of sleeping body pressure and proposes to apply a blind source separation technique named Generalized Morphological Component Analysis (GMCA) to the problem.We tested GMCA by data sets with two different sleeping positions of prone and supine, demonstrating that the latter is much harder to separate the heartbeat signal from the body pressure distribution time series.A detailed discussion is made on how to improve the estimation accuracy for the supine position.

Biodata is created by influence of various factors, it is difficult to identify factors, knowledge of experts is necessary. In this research, we try to visualize factors of bio data using machine learning and explain the outcome and problems.

In our previous researches, we discussed a method based on of extracting both of a similarities/style and differences/characteristic component from walking motions using four wearable motion sensors. This method suggested that data with segmented walking motion could be used to identify individuals. However, we did not discuss the physical meanings of the similarities/style and differences/characteristic component of a subject from the walking data. In this paper, we discuss a method of determining which segment data contributes to the similarities/style and differences/characteristic components, toward the understanding of the physical meanings of the similarities/style and differences/characteristic components of a subject from the walking data.

The annotation of classification process when classifying the situation in automatic driving technology is said to be indispensable for fulfilling accountability when automatic driving vehicle caused an accident. It is possible to classify the situation by the local region by making the self-organizing map into multiple layers. Cognitive processes can be visualized by annotating the classification process by tracing the winner from the classified situation.In this paper, we classify road conditions using multi-layered selforganizing map and show the possibility of visualization of cognitive process in automatic driving technique by automating its annotation.

Recently, activities for practical application of automatic drivinghave been rapidly progressing. In the future, it is considered necessary to enable interactive operation by natural language in order to easily operate autonomous cars.Therefore, in this research, we attempt to realize the correspondence relationship, i.e., grounding, between the driving instructions expressed in natural language and the objects in the real world recognized by the sensors attached to a car, and then convert the driving instructions into the particular spatial meaning description to operate the autonomous car. In this study, we particularly focus on the parking operation of the car.

Autonomous driving is performed by artificial intelligence (AI) on a computer. If an accident is occurred, the autonomous driving system should show the evidence of decision making to analyze the reasons of accident. In this report, we propose an integrated simulator for a knowledge-based AI that shows the evidence of decision making. To construct the simulator, we combine Autoware, Gazebo and the knowledge-based AI. Autoware is an open source software for autonomous driving, and Gazebo can simulate arbitarary road environments. By using the proposed simulator, we can verify the practicality of knowledge-based AI for autonomous driving many times. Experimental results show that the proposed simulator successfully presents the evidence of decision making.

As an international automated driving research project, Pegasus Research Project is driven by the German organization, BMWi, is a hot topic, which discussed four issues: 1) Requirements Definition & Convert for Database, 2) Data Processing, 3) Scenario Compilation / Database and 4) Assessment of Highly Automated Driving Function. In the present study, we explore a mesoscopic model to solve the problem on the standardization of metadata in the safety assessment of automated driving systems, especially focusing on vehicle interaction scenario.

Recently, research on Mini 4WD AI is getting active, and it is attracting attention. In developing a controller program for Mini 4WD AI , we need a compiling and a binary transfer, and this is a burden to (especially novice) programmers. The aim of our study is to develop a lightweight LISP-based programming system for Mini 4WD AI . To develop such a programming system, there seem two approaches --- one of them is to create a very compact LISP processing system that runs on the small microcomputer. This is a direct approach, but due to a constraint on memory capacity we must drastically reduce the functions of LISP. Another approach is to implement a LISP system on a PC that controls the microcomputer, where execution commands are transmitted through Bluetooth communications. In this presentation, we discuss the latter approach.

In the mini4WD-AI race, obstacles are placed in the course.In order to realize stable running over obstacles, speed control is important to run it through.Speed is determined by the position relations between the car and the obstacle.The system identifies the positions of obstacles on the course using cruise logs of several sensors as irregular values or unseen time-series data.

In the Mini4WD AI, speed control is necessary to safely pass obstacles in the course. We proposed the sensor-less revolution measurement, using electromotive force by motor revolution speed. This system need only 2 resistors and Analog-Digital Converter function on micro computer for motor control in hardware system. We explained the schematics for estimating the rotation speed and its measurement principle. And experimental results show the effectiveness of this estimation.

The game of Mini 4WD AI requests the high speed and stable traveling on the circuit course. In order to realize those requests, Mini 4WD has to run fast on the straight course and reduce speed to pass through the curve and jump ramps. In this paper, we propose a method for recognizing course of Mini 4WD by neural networks.

For learning the Mini 4WD AI, some sensor data are needed. It takes a lot of time and cost to acquire them on Mini 4WD.In order to decrease the time and cost, in this study, we construct Mini 4WD simulator by Unity. In this experiment, we compare of real Mini 4WD to the movement of Mini 4WD simulator.

In the elderly people, the sense of balance hebetates and the walking function declines. This causes the walking rhythm to be disturbed and causes a fall, leading to injuries such as fractures. Final goal of this is prediction of a fall by measuring foot pressure during walking. To realize that, a wearable plantar pressure measurement system was developed by using pressure sensors. Then, the developed system was applied three persons to obtain foot pressure data in moving straight and turning. Reliability of the system was verified based on the collected data.

Performing competitions such as tennis and baseball can damage medial collateral ligament (MCL) of the elbow joint．This study aims to develop an evaluation system with high repeatability and reproducibility for the evaluation of MCL．We use a stress measurement system for quantifying stress load and three-dimensional electromagnetic sensor for acquiring three-dimensional position and attitude information for testing equipment． The developed system measured valgus stress test 5 times for three subjects, including detachment of instruments． Finally, the valgus stress can be measured with high reproducibility.

In sports using racquets such as tennis, the triangular fibrillary cartilage complex (TFCC) can be injured by the action of towing, compression, twisting of the wrist. To treat triangular fibrocartilage complex (TFCC), we aim to develop a device that can quantitatively evaluate the behavior during wrist rotation. In this research, we develop a system to measure the positional relationship between the radius and the ulna with wrist rotation by attaching an electromagnetic sensor to the wrist. In the experiment, the behavior of the left and right wrists when turning was evaluated in three healthy adults.

In this paper, multilayer clustering based on T-norms for high-dimension low-sample size (HDLSS) data is proposed. In order to reduce the number of variables of HDLSS data, a variable selection method based on fuzzy clustering is employed. Numerical examples show better performance of the proposed method.

Due to Information becomes enormous and complicated, data processing technology has developed rapidly. In recent years, among the data analysis methods which are means for extracting useful information from such enormous information, topological data analysis which analyzes by focusing on the structure of data with reference to topological geometry has attracted attention. On the other hand, clustering is one method of data analysis of unsupervised learning method, and it is used in various fields including information science. In this paper, we focus on the structure of the cluster, not the data, and propose a clustering algorithm whose clustering result is homotopy equivalent with weighted alpha complex. In addition, we show the mathematical properties of the proposed algorithm and examine the effectiveness of the proposed method through numerical examples.

The Louvain method is one of the typical network clusterings. It is well-known that the Louvain method obtains better cluster partition in short time. However, there are several network data which are not obtained better cluster partition by the Louvain method. One of the reasons for the above is that the Louvain method focuses on only edge connection. We proposed the method which focuses on node size. The proposed method optimizes the objective function of k-medoids by solving linear programming problem under the constraints on node size. We verified the usefulness of the proposed method in the viewpoint of calculation time and accuracy with unweighted network datasets. The numerical examples show that the proposed method is faster and obtains better cluster partition than the Louvain method with the datasets which consist a number of small node size clusters.

Most network data obtained from the real world consists only edge connection. To analyze network data using clustering, weight or dissimilarity between nodes are required. Although various weighting methods have been proposed, it has not been discussed which weighting is suitable. In this study, we used two weighting methods which are calculated from edge connection to verify the suitable weighting method to unweighted network data. One is the Euclidean distance based on adjacency matrix and the other is the diffusion kernel. Next, the k-medoids method and the Louvain method are executed to obtain network cluster partition. After that, obtained network cluster partition is evaluated by cluster validity criteria including the Modularity. The result showed that the diffusion kernel is effective.

A self-organizing map (SOM) is known as a good tool to discover and visualize underlying rules in a given data set. In this study, it is applied to analyzing 840 company information extracted from job-offering documents sent to Saga University. According to the preceding study, its essence is not just data size, but a ratio of map size to data size. Then, training performance of each SOM, whose size of the competitive layer is 35x35, 40x40, 45x45, 50x50, or 55x55, is investigated through some computer simulations. As a result, a smaller SOM cannot develop an appropriate feature map successfully, because the competitive layer is too small to develop the overall data set property. Moreover, a common winner for different applied samples is observed, the number of such common winners is quite big, and its distribution is not uniform. In contrast, a bigger SOM has an enough space to develop, so above-mentioned problems are solved.

The authors classified serum uric acid values in health examination data by SOM significance method. Kohonen's LVQ (Learning vector quantization) is a method for classification. LVQ does not require a normal distribution of data. On the other hand, the multivariate discriminant analysis method is based on a normal distribution of data, In the case of the classification where data is not normally distributed, LVQ could have a better accuracy. Parameter setting such as the number of codebook vectors is arbitrary for LVQ. It is an advantage but could be a disadvantage oppositely. In this study, we report the basic properties of LVQ by using model data.

Iris data is composed of setosa 50 plants, vergicolor 50 plants, virginica 50 plants, and total 150 plants. This iris data is used as basic benchmark data of cluster classification. However, when analyzing with spherical SOM, misclassification was found at the boundary between vergicolor and viginica. For the analysis, the distance calculation on the sphere, and the LVQ method were used.　Further error sum of squares was also used as an aid.

In recent years, many laser measurement instruments for measuring the three-dimensional shape of real space as point cloud data have been developed. Ministry of Land, Infrastructure, Transport and Tourism is pushing “i-Construction” to improve productivity by utilizing the three-dimensional information obtained by these lasers measuring instruments in construction work sites. Especially, in river management, survey results are accumulated in river management offices and development bureau throughout the country originating from a flood safety assessment project (LP project) utilizing airborne laser survey. Therefore, the opportunity to use point cloud data is increasing. However, the existing application for processing point cloud data has a problem that does not have any functions for the river management. Therefore, in this research, we propose a point cloud editor suitable for river management.

Recently, standardization of the as-built management using point cloud data for river banks and expansion of engineering species to be managed are being studied. However, in the case of using point cloud data, boundary line (break line) of change in cross section of structure required for the asbuilt management is ambiguous. Therefore, it is difficult to estimate break line. In the existing research, there is a method to estimate the intersection lines at the crown surface and the slope face using Digital Mapping (DM) data or manually entered line. But it is impossible to apply the existing method to the completion of making bank and concrete block with DM data. Also, it is difficult to extract points measured break line needed for as-built management. In this work, we develop a technique for extracting break line using the shape of structure written on the design drawings, and then we evaluate its effectiveness.

River structures play a role of preventing flooding in the event of a disaster, and its shape continues to change constantly. It is necessary to inspect and grasp its shape. In order to improve the efficiency of river maintenance, the authors proposed a river maintenance system using three-dimensional point cloud data obtained by unmanned aerial vehicle (UAV). However, since it was measurement from the sky over the river, it was difficult to construct three-dimensional data of revetment and bridges. In this research, three-dimensional data was constructed using UAV and terrestrial laser scanner to solve the problem of the preceding research. As a result, the environment was constructed that it not only grasps the river shape in three dimensions, but also manages periodic inspection records and photographs at accurate positions in three dimensional space.

In Japan, MMS is used for maintenance and management of roads and rivers and generation of highly accurate three-dimensional map for automatic driving. However, MMS is very expensive and it is difficult to deploy to government offices nationwide. For this reason, the Ministry of Land, Infrastructure, Transport and Tourism aims to establish vehicle-based sensing technology using inexpensive equipment. Meanwhile the use of inexpensive equipment decreases the accuracy of measurement data, so establishment of a technique for generating three-dimensional information with high precision is desired. Therefore, the authors have designed a vehicle mounted sensing unit combining inexpensive sensors. In this research, the sensing unit is prototyped and wide area point cloud data is generated by two methods. Then, we consider the accuracy and challenges of each method.

Information on what a user can see and what he/she cannot see in the surrounding environments is one of the important clues to estimate his/her subjective states and predict the next action. Our previous study has proposed a system for estimating the user’s three-dimensional field of view by using multiple RGB-D cameras cooperatively. However, in the previous system, positions of the cameras must be known and fixed. Therefore, the system cannot estimate the user’s view from moving cameras. In this study, we propose a view estimation system using only one omnidirectional stereo camera to achieve a wide range estimation from a moving camera.

By using multiple RGB-D cameras, the method for estimating a user’s attention based on the saliency using the distribution of visual acuity in the user’s estimated first person view has been proposed. However, in this method, the calculation does not take the human’s visual characteristics into account. It is considered that the human’s visual attention is affected by not only color but also distance and three-dimensional shape of the object. In this study, with the aim of estimating a user’s attention more accurately, we propose a system for converting information obtained by the RGB-D cameras to point cloud data and reconstructing the user’s first person view.

Discussion on information efficiency of Konohen fuzzy systems.

The purpose of this study is to comprehensively analyze open data on local governments and to construct a local government evaluation model using emotional words by combining psychological models. In the method proposed in this paper, principal component analysis and fuzzy inference are performed on prefecture data of 65 indicators used in “Happiness index Ranking”. As a result, the evaluation values of “Comfortable” and “Active” of each prefecture are obtained, and evaluation values are located in the circumplex model of Russell. Each prefecture is evaluated with emotional words of the Russell model.

Research on communication is perfomed from a multifaceted point of view and research results on pragmatics have given various knowledge to research and development on human-robot communication including discussion on nonverbal communication. In this research, we aim to clarify the constructive theory of interdependent temporal learning process of body exercise ability and communication ability in cognitive development. Therefore, in this paper, we propose a method of learning behavior using multi layer perceptron from data obtained by human motion recognition and measurement using Spiking Neural Network.

In this paper, we discuss an improvement of extraction method of generalized dynamic reducts (GDRs).For a given set of subtables generated from the original decision table,we have proposed and implemented a method for extracting attribute subsetsthat can classify all classifiable objects in at least$100\times(1-\epsilon)\%(0\leq\epsilon<1)$of the given subtables, however, we have not confirmed whether the extracted attribute subsets are relative reducts in the given subtables.In this paper, we introduce a new parameter $\beta (0\leq\beta<1)$ and propose a method that extracts attributes subsets that can classify all classifiable objects in at least $100\times(1-\epsilon)\%$ of the subtables and be relative reducts in at least $100\times(1-\beta)\%$ of the subtables.

Rough-set-based interrelationship mining, proposed by the authors, aims at mining characteristics based on comparison of attribute values between different attributes with respect to some relationships between the attributes. However, in our formulation, the relationships between attributes are given and the relationships were not the target of mining. In this paper, we discuss a possibility of mining relationships between attributes.

Privacy protection is absolutely imperative for data release while the utilization of public data and big data is getting popular. In this paper, data anonymization methods using rough set-based rule induction are investigated. It has been shown that many rules with imprecise conclusions can improve the classification accuracy of the rule-based classifier. Data anonymization methods utilizing rules with imprecise conclusions are proposed. The data tables anonymized by one the proposed methods can preserve the classification accuracy of rules induced from them. The proposed methods as well as the conventional data anonymization methods are compared from both viewpoints of the classification accuracy of rules induced from the anonymized data table and the preservation of data anonymity. The results show the usefulness of the proposed methods.

This paper overviews a collection of problems and solutions posed in the new framework of inverse problems of rough sets, whereby a boundary or boundary-like set is to be found given a lower and upper approximations of ‘an unknown set’. We consider simple examples of the inverse problem as well as a family of existing and new classification and clustering problems.

In the production line management of a food factory, a video recording device is installed and a human works on confirming the products for the captured image. It takes a lot of efforts because there are a lot of images to check. We, therefore, apply object recognition and detection technology using neural network for a production line management system to decrease the effort needed for the work. The objects are images of packed meat to be sold at stores, and we identify from the image the kind of meat (beef, pork etc.) and the cuts of meat (barbecue, shabu-shabu, mince etc.). For the merchandising label attached on the pack, we detect the area of label and identify the contents written on the label.

Time series is a sequence of data in a temporal axis. We have proposed a method for partitioning time series data into several periods whose trends are different from the adjacent with a hierarchical clustering. In the real-time environment, we need to do clustering every time when a new data are added. In this paper, we apply the method to an online environment. First, we partition the initial data into several clusters. Then, we determine the first cluster to be a period hereafter the data in the period are not considered for clustering, and store the size of the second cluster as the threshold. Every adding a new data to the time series, we do clustering for the time series data and we determine the first cluster to be a period with updating the threshold if the size of the first cluster is successively greater than the threshold several times. We repeat these processes. This method makes it possible to apply our partition method to an online environment because the number of data is decreased significantly in the clustering.

Hanshin Expressway Engineering Company Limited has been conducted damage analysis of the road pavement by using images taken with a line scan camera mounted on a road surface property measuring vehicle. And then analysis worker detects crack by own eyes and calculates “crack rate”.However, the efficiency of the analysis work is not good and judgment result varies depending on the characteristics of the worker, because the crack detection works rely on human inspection. In this research, the system evaluates the statement of a crack by using DL-FCM classifiers that is combined deep learning and Fuzzy C-Means classifiers is proposed.

Stable combustion for long periods of time is required in waste incineration facilities to achieve stable power generation and emission control. The crane operator aims to 1) keep the waste pit flat for safe crane operation, and 2) “mix” the waste several times using the crane, in order to break the plastic waste bag and homogenize the waste before combustion. In this research, we developed an automatic crane operation scheduler and waste pit simulator. We proposed an improved genetic algorithm for time-series optimization in crane operation scheduling. A fuzzy function was used for the fitness calculation in the genetic algorithm to evaluate the created crane operation schedule.

This paper discusses a class incremental problem based on Fuzzy Genetics-based Machine Learning (GBML). In general, a learning algorithm of Fuzzy GBML is designed as a batch learning. Therefore, it requires a re-learning with a whole data if the Fuzzy GBML attempts to learn a new class data. We propose a method to efficiently generate rules corresponding to new class by selecting and relearn less importance rules out of rules classifying existing classes. From experimental results, the proposed method is able to perform the class incremental learning efficiently.

This paper discuss the on-line learning of fuzzy classification system. The on-line learning algorithm is formulated using Exact Soft Confidence Weighted (SCW) learning, which has been proposed in the machine learning community. Under the situation where only a single training pattern is available at a time, a fuzzy classification system is incrementally constructed using SCW. We will discuss the characteristic feature of the on-line learning by a series of computational experiments.

Information aggregation has a long history of studies and is being used in decision-making, sensor fusion, affective intelligence, Web Intelligence, and many other fields for combining certainties, reliabilities, sentiments and other degrees to judge something from uncertain information in the real world. The paper discusses a method to combine uncertainties represented by Certainty Factors, which is a traditional and practical representation of uncertainty used in many intelligent systems, but has been criticized due to lack of a sound mathematical theory. The paper tries to establish a theory for sound interpretation of Certainty Factors and proposes a new interpretation using Possibility theory. It also proposes a few methods to combine Certainty Factors with good theoretical basis, one of which is exactly the same as the one criticized for long time.

Pursuit problem is one of the benchmark problems of multi-agents. For more realistic model, we have proposed a staged view and then a staged binocular view for hunters. Simulation results with these views show that hunters can even learn and capture the prey with a fuzzy Qlearning. In the previous study, we have proposed hunters with a communication function and compared the capture ability in several cases of sending information with a fuzzy Q-learning, where an action from Q-table for visual information and that for communication information are combined with the constant weights. The results were worse than that for only visual information. In this study, we dynamically change the weights of combination for a situation, where we learn the weight with a fuzzy Q-learning. The result gets better than that for only visual information.

In railways, brake control is the important factor on safety, punctuality and etc. However, the slip is likely to occur during brake control since the wheels and rails are made of metal. The slip causes extension of the brake distance and damage to the wheels. The condition of the slip is greatly due to various factors such as the load, the state of the contact surface between the wheels and the rails, so the slip becomes a high uncertainty matter. Actual train uses the cooperation control of electric brake and air brake for wheel-slide-protection (WSP), but it can be more efficient to divide braking force because each of them performs independent. It is known that fuzzy inference shows high effectiveness for the object with high uncertainty. In this paper, we will compared the cooperation control with WSP using only air brake through numerical simulation, and evaluate of the cooperation control of electric brake and air brake.