Kazunobu Takahama

Human Activity Monitoring Using Fuzzified Neural Networks

Abstract For monitoring and estimating our daily activity, some kinds of devices are available. One of such kinds of monitoring devices is a MEMS based prototype which is developed by the Maenaka Human Sensing Fusion Project. We have developed a estimation method of human activity from three-axis acceleration data using the above-mentioned prototype. This method can estimate our unit activities, such as (1) walking, (2) running, (3) sitting, (4) lying, and (5) standing. In this paper, we propose a system that can find unusual situation from ECG data. Our proposed system is based on the fuzzified neural networks. The fuzzified neural network is trained by using sensing data with reliability grade. Since the fuzzified neural network learns normal state of the subject person, we can understand the ECG state of the subject when we analyze fuzzy outputs from the trained fuzzified neural network. This paper shows estimation results by using actual monitoring data which contains normal state, and artificial unusual data. From the results for the actual monitoring data, we can see that our proposed system was able to estimate the testing data as normal. From the results of estimating artificial unusual data, our proposed system can find the subject persons unusual situation.

Fuzzified Neural Network Based Human Physical Condition Monitoring Using MEMS Based Monitoring Devices

In order to understand our physical condition, we need to record the detail of physical condition data like the heart rate. However, for understanding such data, additional information such as what the subject is doing at that time is needed. We propose a combined system which consists of a fuzzified neural network based unusual condition detection and a standard neural network based action estimation. From experimental results, the effectiveness of our proposed system is shown for understanding our conditions.

A directed graph based feature definition for classifying nursing-care texts

In the aging society such as Japan, we feel large importance for improving the quality of nursing-care to keep our quality of life. Development of a computer aided evaluation system for improving the quality of nursing-care is our final goal. In order to evaluate the quality of actual nursing in wide areas in Japan, we have been collecting texts that are written by nurses using our Web based system. A SVM based classification system has been developed to classify such nursing-care texts, and a dependency relation based feature vector definition has also been proposed in our previous researches. When we train the SVM based classification system, pre-classified nursing-care texts by a few nursing-care experts are used as a training data set. Some texts in the training data are similar but classified into different classes. To classify the nursing-care texts with high accuracy, we need to extract numerical features that can express characteristics of the original text. In this paper, we explain some feature vector definitions and propose a directed graph based feature vector definition.

A matrix-based feature vector definition and a SVM-BDT-based classification system for classifying nursing-care texts

In this paper, we propose a method of nursing-care text classification. We have proposed some nursing-care classification methods using fuzzy systems, standard three-layer neural networks, and support vector machines. Also we have proposed several types of feature vector definitions for expressing free style Japanese texts into numerical vectors. This paper proposes a novel feature vector definition and a support vector machine utilizing a decision tree (SVM-BDT) based classification system. From experimental results, the effectiveness of both feature definition and SVM-BDT-based classification system is shown.

Soft class decision for nursing-care text classification using a k-nearest neighbor based system

In the aging society such as Japan, it is very important to improve the quality of nursing-care for keeping our quality of life. Our final goal is to develop a computer aided evaluation system to improve the quality of nursing-care. For evaluating the quality of actual nursing, we have been collecting texts that are written by nurses using our Web based system. In our previous works, a SVM based classification system has been developed to classify such nursing-care texts, and a dependency relation based feature vector definition has also been proposed. The training data are pre-classified texts by a few nursing-care experts. Some texts in the training data are similar but classified into different classes. To classify the nursing-care texts with high accuracy, we need to tackle such ambiguous class labels in the training data. In this paper, we propose a k-nearest neighbor based classification system which can classify into classes with certainty grade.

Fuzzy neural network based activity estimation for recording human daily activity

We proposed a standard three-layer feedforward neural network based human activity estimation method. The purpose of the proposed method is to record the subject activity automatically. Here, the recorded activity includes not only actual accelerometer data but also rough description of his/her activity. In order to train the neural networks, we needed to prepare numerical datasets of accelerometer which are measured for every subject person. In this paper, we propose a fuzzy neural network based method for recording the subject activity. The proposed fuzzy neural network can handle both real and fuzzy numbers as inputs and outputs. Since the proposed method can handle fuzzy numbers, the training dataset can contain some general rules, for example, “If x and y axis accelerometer outputs are almost zero and z axis accelerometer output is equal to acceleration of gravity then the subject person is standing.”