Takuma Akiduki

Toward classification of swimming style by using underwater wireless accelerometer data

The use of technical refinement is playing an important role in the process of optimizing training efficiency and improving results of athletes. Therefore, we aim to develop an online monitoring and feedback tool to follow the swimming training using a single waterproofed and wireless motion sensor attached on the swimmers center of body (back). In this paper, as the first step, we investigated the feature quantities to discriminate four swimming styles using the data of a single three-axis accelerometer. We found that there was a possibility that the combination of simple feature quantities could be used to discriminate four swimming styles.

Associative memories with multi-valued cellular neural networks and their application to disease diagnosis

Cellular neural networks (CNNs) are one type of interconnected neural network and differ from the well-known Hopfield model in that each cell has a piecewise linear output characteristic. In this paper, we present a multi-valued CNN model in which each nonlinear element consists of a multi-valued output function. The function is defined by a linear combination of piecewise linear functions. We conduct computer experiments of auto-associative recall to verify our multi-valued CNNs ability as an associative memory. In addition, we also apply our multivalued CNN to a disease diagnosis problem. The results obtained show that the multi-valued CNN improves classification accuracy by selecting the output level q properly. Moreover, these results also show that the multi-valued associative memory can expand both the flexibility of designing the memory pattern and its applicability.

Realization of multi-valued associative memory with cellular neural network

In this paper, a design and implementation method of cellular neural networks with a multi-valued saturation function is discussed. The multi-valued saturation function is defined as a linear combination of piecewise linear functions that consist of a saturation range and a linear-slope range. The length of these ranges determine the output characteristics of each neuron, and is expressed by two parameters. We investigate the relation between these output characteristics and the ability of the multi-valued associative memory with computer simulation. Finally, we propose a design guideline for deciding its parameters from these results.

Study on bridge floor crack detection using the 2-dimensional complex discrete wavelet packet transform

Bridge inspection usually is done visually on the spot. For this reason, inspection is very difficult in the case where humans can not enter, and it has become a problem. Therefore, in order to overcome this problem, we propose a new method in which pictures are taken by Drone and cracks are detected by image processing using a wavelet transform. Since bridge crack include many directional components, the characteristics of directional selection is useful for feature extraction. In this study, a 2-Dimensional Complex Discrete Wavelet Packet Transform with excellent directional selection is applied to bridge floor crack detection, and encouraging results are obtained.

A New Design Method of Multi-Valued Cellular Neural Networks for Associative Memory

It has been reported in the literature that Cellular Neural Networks (CNN) are effective as associative memories and they have been applied to many kinds of pattern recognition tasks. Flexibility of their design can be increased by expanding the output function from being 2-valued to being multi-valued. As a design method for associative memories, SVD(singular value decomposition) is popular, but a design procedure which uses LMI(linear matrix inequality) was proposed and obtained excellent results. In this paper, we propose a new design method expanded from the 2-valued output CNN to a multi-valued output CNN by using the LMI method and confirm the effectiveness of it.

A Study of a new wavelet neural network for deep learning

As development of smart devices progresses, voice control is becoming more and more important. The first step of voice control is to separate noise and the voice. In recent years, as the data size is increasing, deep learning has become a very useful tool in data processing. The use of deep learning to separate signals is becoming increasingly popular. However, deep learning has the disadvantage that the output of the pre-processing part used in the multi-layer neural network suffers from oscillation when applied to a regression problem. In this study, a new wavelet neural network has been proposed to improve this phenomena. The wavelet neural network applies a discrete wavelet transform in the hidden layer of the traditional multi-layer neural network. In our simulation experiments, encouraging results were obtained by the proposed method.

Consideration of composing method of the optimized real-signal mother wavelet

The main drawback of abnormality diagnosis using the wavelet transform (WT) is that it needs much time. In order to improve this problem, a new real signal mother wavelet (RMW) that is made from all objective frequency components has been suggested. Thus the RMW can be greatly shorten the WT calculation time, and this WT is named the Wavelet Instantaneous Correlation (WIC). However, in order to construct the conventional RMW, it is essential to gather the target sounds, namely “sound samples” beforehand. In general, these are gathered by professionals using sensors. Furthermore it is possible that the abnormality diagnosis accuracy may decrease, because of a frequency component difference between the actual sound and the “sound sample”. In this study, therefore, we propose an improved method in which the expected characteristic frequency is extracted from actual sounds automatically, without requiring the “sound sample” to be obtained by professionals.

A new approach to diagnose Sleep Apnea Syndrome using a continuous wavelet transform

Currently, it is said that potential sufferers of Sleep Apnea Syndrome (SAS) account for up up to about 2% of the population in Japan. Not only does SAS cause lack of concentration during the day, it may also cause complications such as hypertension and heart failure, and it has been called a modern disease. However, there is a problem that it is impossible to decide if one suffers from it And diagnosis is difficult if a patient does not go to hospital, because diagnosis requires many resources. Therefore, we propose a method that can easily diagnose SAS by the continuous wavelet transform (CWT) using a vocal sound signal, and obtain encouraging results.

Human Motion Analysis from Inertial Sensor Data Based on Nonlinear Dynamics

Abstract This paper discusses the problem of human motion analysis from inertial sensor (accelerometer and gyroscope) data, which is the time sequence data obtained from human motion. The inertial sensors are widely used in the wearable sensing field for understanding human activities. Human motion can be divided into simple movements, such as swinging the arms or legs. These simple movements are fundamentally periodic, and they can be modeled with dynamical systems having periodic attractors. On the other hand, more complex motion like walking can be represented as a sequence of simple movements. To address the problem of analyzing human motion data, we propose a new framework based on a dynamical system. The human motions observed by the inertial sensors are divided into simple movements, and they can be often described as periodical signals. The periodical signals are described by a dynamical system, which can store the periodical or transitional trajectory in a state space by using basin of attraction. The attractors abstract a kind of proto-symbol. The dynamical system having periodical attractors is shown to characterize human motion effectively by exploiting spatiotemporal continuity, because it describes the flow of its transitions into the state space. Moreover, in the designed symbol space from the attractors, the information of human motion dynamics is described in the placement of the points. Thus, each point or transition between points in the symbol space corresponds to a specific type of motion; it can act as a notation method of motion characteristics.