Recognition of daily living activities is important to realize advanced services that increase QoL (Quality of Life) of residents by providing context-aware appliance control, health monitoring and so on. Our previous work already achieved recognition of nine types of daily living activities with an accuracy of 68% by applying Random Forest machine learning algorithm to the data collected with ECHONET Lite appliances and motion sensors in a home. ECHONET Lite is a communication protocol for the control and the sensor network of networked appliances used in a home and has been standardized as ISO/IEC-4-3. Many appliance manufacturers are developing ECHONET Lite appliances and introducing them into the market. In addition, motion sensors are already widespread and attached to some appliances such as lighting devices and IH cooking heaters so that they are automatically turned off when no one exists nearby. In this paper, we propose a new method of daily living activity recognition by introducing time series data analysis with LSTM of Deep Learning. We collected data with ECHONET Lite appliances and motion sensors attached to the appliances over 12 days in our smart home facility where four participants spent usual daily life for three days each. As a result of time series analysis of the collected data with LSTM, we achieved the recognition accuracy of 85% for 9 different activities.

In recent years, the number of people using online recipe services for cooking has increased with the spread of the smartphone. In a recent survey, more than 60% of housewives answered that they use the opportunity to use online recipe sites more often. It is difficult to match food taste to user 's preference as each online recipe page shows a recipe to realize just one taste even though there are countless numbers of recipes in an online recipe service. As a preliminary experiment, we investigated the degree of individual difference in taste preference. As a result, one-third of the subjects preferred to adjust the amount of hot water used for a Miso Soup by -20% or + 20% of the specied amount of the original recipe. The results of the two conducted experiments lead to the decision, that there is a need for a system that can help to adjust the optimal amount of seasoning to personal taste and to support the addition of the exact amount of seasoning. The purpose of this research is, therefore, to increase daily meal satisfaction through the realization of a system, that improves the taste of food based on an online recipe closer to the user's preferable taste without burdening the user. To realize this system, the following three problems are approached: (1) building an individual preference model for food taste, (2) determining the seasoning quantity based on the individual preference model. Existing Research focuses on recommending recipes according to personal taste. However, there is no research on the usage of a system which supports the adjustment of the recipe based on personal preference. Therefore, we develop a system that extracts the user's preference by assessing the individual taste through a questionnaire in a learning model and adjusting the amount of seasoning in the base recipe accordingly, which solves the problem (1) and (2). As a validating experiment, we analyzed the user preference models based on feedback before and after adjusting the seasonings. The results showed that the preference model estimation using a two-week timeframe whose results show that the longer timeframe also increased the accuracy of the preference model, to a point, that the individual preference model can be built correctly.