Mye Sohn

Complex Activity Recognition Using Polyphonic Sound Event Detection.

In this paper, we propose a method for recognizing the complex activity using audio sensors and the machine learning techniques. To do so, we will look for the patterns of combined monophonic sounds to recognize complex activity. At this time, we use only audio sensors and the machine learning techniques like Deep Neural Network (DNN) and Support Vector Machine (SVM) to recognize complex activities. And, we develop the novel framework to support overall procedures. Through the implementation of this framework, the user can support to increase quality of life of elders’.

Transition activity recognition using fuzzy logic and overlapped sliding window-based convolutional neural networks

AbstractnIn this paper, we propose a novel approach that can recognize transition activities (e.g., turn to left or right, stand up, and travel down the stairs). Unlike simple activities, the transition activities have unique characteristics that change continuously and occur instantaneously. To recognize the transition activities with these characteristics, we applied convolutional neural network (CNN) that is widely adopted to recognize images, voices, and human activities. In addition, to generate input instances for the CNN model, we developed the overlapped sliding window method, which can accurately recognize the transition activities occurring during a short time. To increase the accuracy of the activity recognition, we have learned CNN models by separating the simple activity and the transition activity. Finally, we adopt fuzzy logic that can be used to handle ambiguous activities. All the procedures of recognizing the elderly’s activities are performed using the data collected by the six sensors embedded in the smartphone. The effectiveness of the proposed approach is shown through experiments. We demonstrate that our approach can improve recognition accuracy of transition activities.

Unsupervised and non-parametric learning-based anomaly detection system using vibration sensor data

In this paper, we propose an anomaly detection system of machines using a hybrid learning mechanism that combines two kinds of machine learning approaches, namely unsupervised and non-parametric learning. To do so, we used vibration data, which is known to be suitable for anomaly detection in machines during operation. Furthermore, in order to take into account various characteristics of abnormal data such as scarcity and diversity, we propose a novel method that can detect anomalous behaviors using normal patterns instead of abnormal patterns from the machines. That is, we first perform a machine learning of the normal patterns of the machines during operation, and if any of the operation patterns deviates from the normal pattern, we identify that pattern as abnormal. A key characteristic of our system is that it does not use any prior information such as predefined data labels or data distributions to learn the normal operation patterns. To demonstrate the superiority of our system, we constructed a test bed consisting of a washing machine and a 3-axis accelerometer. We also demonstrated that our system can improve the accuracy of anomaly detection for the machines compared to other approaches.

Situation reasoning framework for the Internet of Things environments using deep learning results

A goal of this paper is to suggest a framework to infer the situation using IOT sensor data. To do so, the framework adopts contexts which were derived from the learning results of multiple deep neural networks for IOT sensor data and carries out hierarchical clustering of contexts in terms of the spatio-temporality. With the learned dendrogram, the most appropriate situation is inferred from case-based reasoning depending on the similar time and location. The result of reasoning is stored in a case memory and this can contribute to learning of a case base. The primary contribution of this paper is the situation reasoning under consideration for spatio-temporality that is a characteristic of IOT sensor data. Also, we performed experiments to show the superiority of our framework. The experimental results are not bad for a first attempt. In further research, if the algorithms are improved, we can expect better results.

A Model-Data Framework for Dynamic Vehicle Routing Models

We propose a Model-Data Framework (MoDaF) to consider delivery-related static and dynamic contexts to support delivery service. In general, Static Vehicle Routing Models (SVRMs) usually use Static Modeling-Context (SMOC) such as Immediate Pickup and Delivery Requirements (IPDR) and resources related to idle vehicles at depots. On the other hand, Dynamic Vehicle Routing Models (DVRMs) need to consider Dynamic Modeling-Context (DMOC) such as unfulfilled Advanced Pickup and Delivery Requirements (APDR) on advanced vehicle routes. Therefore, it is getting more important to manage Modeling-Context (MOC) including both SMOC and DMOC and to specify the MOC into input data used in DVRMs to find a solution. Recently, the generation of input data for a specific IP (Integer Programming) model becomes a significant issue as the MOC is getting dynamic. Hence, it needs systematic approach. In this research, we focus on the generation of input data from the MOC and Input Data Inference Rules (IDIRs) based on forward chaining for a dynamic-MDPDPTW as an illustration of a specific DVRM. To prove the availability of a MoDaF, we implement a prototype to generate input data from MOC effectively and show that it is significantly useful to solve DVRM in the realistic vehicle routing system through the experimentation.