1 Introduction
1.1 Background
1.2 Related work
CASAS dataset apartment id | Accuracy |
---|---|
HH101 | 64.18% |
HH102 | 74.57% |
HH103 | 61.98% |
Averaged | 66.91% |
1.3 Motivation
1.4 Contribution to research
1.5 Data availability
2 Methods
2.1 Hypothesis
Event timestamp | Sensor | Event | Label | Location |
---|---|---|---|---|
2012-07-20 11:09:59.128578 | MA015 | OFF | Toilet="end" | Bathroom |
2012-07-20 11:50:04.029817 | M008 | ON | Relax="begin" | Lounge |
2012-07-20 11:55:40.958827 | M008 | OFF | Relax="end" | Lounge |
2012-07-20 12:00:43.445554 | MA015 | ON | Personal_Hygiene | Bathroom |
2012-07-20 12:00:44.634408 | MA015 | OFF | Personal_Hygiene | Bathroom |
2012-07-20 12:00:45.129838 | MA015 | ON | Personal_Hygiene="begin" | Bathroom |
2012-07-20 12:05:30.659402 | MA015 | OFF | Personal_Hygiene="end" | Bathroom |
2.2 Dataset
2.3 Research method
Time window | Range |
---|---|
Morning | 0700–1100 |
Afternoon | 1200–1700 |
Evening | 1800–2100 |
Night | 2200–0600 |
Dataset | Time_window | Predicted_next_label | Inf_score | Actual_next_label |
---|---|---|---|---|
hh101 | 22:00:00–06:59:59 | Personal_Hygiene | 0.348281 | Evening_Meds="begin" |
hh101 | 22:00:00–06:59:59 | Evening_Meds="end" | 1.05856 | Evening_Meds="end" |
hh101 | 22:00:00–06:59:59 | Personal_Hygiene | 0.305946 | Dress="begin" |
hh101 | 22:00:00–06:59:59 | Dress="end" | 1.09628 | Dress="end" |
hh101 | 22:00:00–06:59:59 | Personal_Hygiene | 0.44534 | Personal_Hygiene |
hh101 | 22:00:00–06:59:59 | Personal_Hygiene | 0.899305 | Personal_Hygiene |
hh101 | 22:00:00–06:59:59 | Personal_Hygiene | 0.899305 | Personal_Hygiene="begin" |
3 Results
4 Conclusion
Test phase | Apartment dataset | Accuracy (%) |
---|---|---|
Time window with no contextual data | HH101 | 3.2557 |
HH102 | 27.2411 | |
HH103 | 9.8706 | |
Average accuracy: 13.456 | ||
Time window with added contextual data | HH101 | 65.0407 |
HH102 | 82.3991 | |
HH103 | 76.2162 | |
Average accuracy: 74.552 |
CRAFFT benchmark | Semantic blocks model (SBM) | |
---|---|---|
HH101 | 64.18 | 65.0407 |
HH102 | 74.57 | 82.3991 |
HH103 | 61.98 | 76.2162 |
Average | 66.91 | 74.552 |
5 Future work
-
Automated dynamic clustering of the smart environment data, so that the most optimal time windows are generated, bespoke, for each environment based on previously observed data. Exploring the use of machine learning clustering algorithms would be a good starting point for this work.
-
Adding more contextual data. The SBM has been designed to be extensible. Such data may include person metadata (age, adult/child, gender, religion, state/emotions), calendar metadata (daily events, calendar events, ritual/cultural/geographical events) that could influence a person’s activities, and multi-occupancy environment data where interactions with other people are taken into account.
-
Experiment with a data completeness coefficient, that would alter the prediction calculation based on how many of the contextual sets of data were present and complete.
-
Pairing with a DBN to account for causality, and shortlisting of possible next activities.
-
Pairing with a Deep Learning (DL) model for shortlisting of possible next activities.
-
Comparing the SBM results to a modified CRAFFT model that includes time-windowed data.
-
Analyse the SBM performance when transferring and applying it to other smart home datasets.
-
Determine the SBM accuracy for a range of parameters including size of the time window, and frequency of activity events.