Sensor platform based on environmental sensing and data fusion

Since a large portion of population spends >90% of their time indoors, it is important to measure and report indoor air quality to the users. These measurements of fine dust, carbon monoxide, volatile organic compounds, carbon monoxide are usually reduced to an air quality index to make reporting easier. The problem with such a metric is that the user does not know how much daily activities such as cooking, cleaning and smoking are influencing the air quality, so that they can improve their behaviour, or the smart-home can trigger signals to other devices to maintain a healthy environment when these activities are performed. As a step to develop such a system, this study focuses on developing a system that measures 9 different air quality parameters and is able to classify activities into one of the four classes (Background, Cleaning, Cooking, Smoking). The study contains a literature review of Machine learning techniques employed in similiar studies, and drawing influence from works in electronic noses and artificial olfaction for applications of food inspection. A few works also detected human activity and count. The selection of literature was done to include studies that use similar sensors, namely gas sensors of metal oxide semiconductor type, and optical solutions to measure CO2 and dust constrations. An alternative CFD simulation based approach is also seen in literature, but the focus was kept on sensor data analaysis. Rolling window feature extraction and normalization techniques are popular in this regard, before putting them through a dimentionality reduction algorithm such as PCA to finally classify it using popular classifiers. Some studies have attempted to eliminate the need of feature extraction using Deep Neural Networks but the works are scarce. Each study attempts to create its own dataset, with the unique selection of sensors, so techniques are highly application specific. To asses the feasibilibility in the context of Indoor Air Quality activity detection, an IoT platform is prepared for data acquisition complete with a sensor hub PCB and a mySQL server to collect important IAQ parameters at regular intervals. A dataset of 80 hours of air quality data is prepared and labelled in a test room. In the last section, the dataset is set as a supervised machine learning problem, time series feature extraction (Vector Array Normalization, Hurst Exponent) and classification techniques (Decision Trees, LDA, MLP) studied in literature are evaluated. The problem is evaluated in two scenarios, one is of binary classification where the algorithms are trained to distinguish between two classes, 'Background' and 'Event',. The second is the multiclass case where there are four classes 'Background', 'Cleaning', 'Cooking', and 'Smoking'. The best accuracy at the end of this study is obtained as 81% for the multi-class case, and 94% for the binary one . The recommendation for future work includes extension of the methodology to more uncontrolled scenarios and eventually an actual living space such as a home or office.