Using high temporal resolution sensor data to support independent living

我们将探索家庭内传感器数据模式与弱势居民健康模式之间的联系。我们将监测室内环境（温度、湿度、空气质量）和用电情况，并使用模式中的特征来检测异常事件。我们将使用参与者的健康和幸福数据来评估检测到的通常事件是否与家庭中的潜在问题有关。一旦传感器数据和潜在健康状况之间建立了联系，我们将致力于提前预测事件，以便提前或先发制人地提供支持。为了确保这种方法的相关性，我们将在使用协同设计方法的过程中让最终用户参与进来。我们已聘请了一个公众参与和参与小组，并将建立一个由保健和保健提供者代表组成的利益攸关方小组。我们将招募50名参与者，他们是弱势群体或有现有健康状况的人。我们将利用综合Smartline数据集（包括长期和高频时间序列环境传感器数据和四年的用电量）的分析技术经验。我们将对数据进行特征化描述，并检测和预测家庭中有关健康和幸福问题的变化。如果成功，这项可行性测试将支持早期干预，从而维持独立生活。我们将使用以下方法从数据中提取特征。傅里叶分析将确定传感器数据中的主导频率。自回归模型将确定以前的读数对当前和未来读数的影响程度。长短期记忆神经网络将用于预测读数。我们还将使用神经网络和支持向量机在异常发生之前预测异常，并使用聚类分析对具有不同类型特征的天进行分类。

We will explore the links between patterns of sensor data within the home and health patterns of vulnerable residents. We will monitor internal home environment (temperature, humidity, air quality) and electricity usage over time, and use features in the patterns to detect unusual events. We will use health and wellbeing data from participants to assess whether the usual events detected relate to underlying issues in the home. Once connections between sensor data and underlying health are established, we will aim to predict events in advance to allow earlier or pre-emptive support. To ensure the relevance of this approach we will involve end users throughout using a co-design approach. We have engaged a public involvement and engagement group, and will establish a stakeholder group of representatives of health and care providers. We will recruit 50 participants, who are vulnerable or have existing health conditions. We will draw on our experience of analysis techniques with the comprehensive Smartline data set (including long-term and high-frequency time-series environmental sensor data and electricity usage for four years). We will characterise data, and detect and predict changes in the home suggesting health and wellbeing issues. If successful, this test of feasibility will support early intervention and thus maintaining independent living. We will extract features from the data using the following methods. Fourier analysis will determine dominant frequencies in the sensor data. Autoregressive models will establish the extent of influences from previous readings to current and future readings. Long short-term memory neural networks will be used to predict readings. We will also use neural networks and support vector machines to predict anomalies in advance of them occurring, and cluster analysis to categorise days that have different types of features.