Reliable Measurement of Wireless Sensor Network Data for Forecasting Wellness of Elderly at Smart Home N. K. Suryadevara, S. C. Mukhopadhyay, R. Wang, R. K. Rayudu* and Y. M. Huang# Massey University, Palmerston North, New Zealand *Victoria University of Wellington, New Zealand # National Cheng-Kung University, Tainan, Taiwan Contact email: S. C.
[email protected] The computer based measurement systems proposed for human emotion recognition; and prediction models for a smart home based health care system such as [6-7] are based on the physiological parameter values. So far, smart home researches on elderly care are based on their individual requirements and have obtained meaningful results to certain extent [6-10]. However, there is still a lack in the confirmation of suitable technique and a low-cost solution. Modeling a human behavioral system is not simple, as understanding human behaviour and realizing into new situations with flexible hardware modules defined by computer algorithms is quite complex. One of the main objectives towards the development of a smart home monitoring system is to have an optimal or minimum number of sensors and intelligent data analysis. A home monitoring system based on a minimum number of sensors will lead to the development of a low-cost system. A low-cost system will be affordable to the elderly who are mostly retired or low-income earner. The sensors should be sufficiently providing information related to the identification of elderly activities to determine the wellbeing of elderly. Use of a large number of sensors will lead to a high-cost of the system and also not very easy to manage huge amount of continuous data coming out of the sensors. The detection of behavioral changes and for better prediction of object usages in a smart home can be realized with wellness determination process. “Wellness determination” a new framework in smart home research has been devised and extensive work is being carried out in this direction [11-13]. Wellness monitoring system for an elderly is précised to monitor the performance of the daily activities and decide whether the behavior is regular or irregular. Though a lot of works are reported on eldercare using different approaches, the concept of wellness is a new one. The wellness determination model is augmented with predictive ambient intelligence technique, in order to forecast the behavior of elderly under monitoring environment. Predictive Ambient Intelligence (PAI) environment gathers information from Wireless Sensor Networks (WSN) including environmental changes and occupants’ interactions with the objects in the monitoring environment [14]. The rest of this paper is organized as follows. The WSN of household appliances monitoring system and determination of wellness is described in Section II and the time series data processing for forecasting and behaviour detection in Section III. The system implementation and experiments results are presented in section IV and conclusion with future works in Section V.
Abstract In this paper, we present an engineering system for monitoring and forecasting wellness of an elderly person in relation to performance of daily activities. Complex behavioural changes of daily activities are captured in real time for reliable measurement of wellness operations. These tasks are realized with the sensor status of the household objects in use by the elderly in combination with prediction process of time series data processing algorithm. This will assist in determining the quantitative well-being of an elderly and alert if the daily activity behaviour is irregular. Keywords - Wireless Sensor Networks, Smart Home, Wellness, Time Series Analysis.
I.
INTRODUCTION
The performance of daily activities of normal elderly people is likely to be steady or unchanging. This implies that indicators of daily activity measurements can specify the quantitative wellbeing of the elderly person. Measuring “wellness”, we mean how well the person is able to perform basic daily activities. In general, basic Activities of Daily Living (ADL) behaviors are related to usage of household appliances. The daily home activities involving basic functions like preparing food, showering, walking, sleeping, watching television, reading books etc., are key indicators in determining the wellness of elderly home activity [1]. An intelligent, robust, low-cost and flexible engineering home automation system is very much required to record the basic home activities and respond immediately when there is a change in the regular activity of an elderly person, so that necessary care can be taken in advance. The necessity for an electronic system with intelligent mechanism for monitoring basic ADLs of elderly is increasing. Elderly people living alone can be assisted with the Information and Communication Technology (ICT) so that appropriate assistance can be provided at the right time by informing to the care taker. Measurement of how well the elderly is able to perform their basic ADLs can be done by monitoring the household appliances that are routinely used by the elderly in executing their everyday tasks. A variety of sensing systems have been proposed and developed in recent times such as usage monitoring of electrical devices in a smart home [2], context-aware smart home monitoring through pressure measurement sequences [3-4], low-power and low-cost implementation of Support Vector Machines (SVMs) for smart sensors and sensor networks for intelligent data analysis and pervasive computing [5].
978-1-4673-4623-8/13/$31.00 ©2013 IEEE
16
II.
SYSTEM DESCRIPTION
The overall structure of the system consists of two important modules: i) Wireless Sensor Network (WSN) with ZigBee modules and ii) Intelligent home monitoring software system to collect sensor data and perform data analysis. Exploration of the sensor data involves measuring the wellness and detecting behavioural changes of an elderly. Fig.1 depicts the block diagram of the wellness measurement system.
The developed sensing system is noninvasive, flexible, lowcost and safe to use. The system does not use any motion and/or vision based sensor. The rationale for observing usage of household appliances is based on the fact that these are regularly used by the elderly in various situations like preparation of food, relaxing, toileting, sleeping and grooming activities. They are useful to determine the wellness of the person in performing these activities.
ADC Electrical Sensor units Household Appliances
Force Sensors Units
Contact Sensors
ADC ZigBee
Wellness Measurement
N/W
Digital
Figure 1. Block diagram of Computer Based Wellness Measurement system
A. Design of the Sensing Units:
Figure 3. The schematic electrical appliance monitoring unit
The WSN setup used for monitoring smart home consists of fabricated electrical sensing units[12] for Room Heater, Water Kettle, Toaster, Microwave, TV, and Dishwasher, force sensing units[12] for Bed, Couch, Chair, Toilet and contact sensors for Grooming Cabinet, Fridge. These are installed at an elderly home to monitor their daily activity behaviour in terms of household object usages and execute effectively machine learning process. Fig.2 shows the electrical sensing units connected to various household appliances. The designed electrical sensing unit schematic is shown in fig.3.
The advantage of electrical sensing schematic [12] is that fabricated electrical appliance monitoring unit can connect two different electrical appliances on a single power inlet, having the intelligence to detect which particular device is ON and how long it is used. The intelligent bed monitoring system is based on a Flexi Force sensor; it is an ultra-thin, flexible, nonobtrusive printed circuit. The Flexi-Force sensors were strategically placed below the mattress of the bed, to determine if a force is being exerted on the bed i.e., someone is on the bed. Fig. 4 shows the force sensing unit connected to various household appliances.
Figure 4. Sensing units connected to Bed,Chair appliance Figure 2. Electrical appliance monitoring units connected to various household appliances
17
The electrical, force and contact sensing units are integrated with ZigBee modules to collect the household appliances usages at a centralized computer based measurement system. Captured raw sensor data are effectively processed for recording the statuses of appliances based on the threshold values as active or inactive. Processed sensor data is then stored in the processing unit in the form of event based activity. This is most efficient technique, as it reduces the size of storage to a large extent and more flexible for measuring the wellness of elderly in real-time.
III.
TIME SERIES MODELING AND FORECASTING
Time is a fundamental element in our daily life and will provide us a vital source of information for smart home monitoring system. Moreover, livelihood activities are cyclic and evaluation of daily activities will indicate performance behaviour of an elderly. Hence, monitoring daily usage of household appliances (i.e.) object monitoring in smart home will help us to recognize the habitual nature of person and thereby we can know how “well” the elderly is able to perform his essential daily activities.
In addition, to the fabricated sensing units, emergency help and deactivate operations are made-up with ZigBee modules to facilitate the corresponding operations during the real-time activity monitoring of the elderly. Since, we are concerned with how “well” elderly is able to perform their basic ADLs, required number of sensing devices that correspond to the daily object usages are used in the present system.
The normal daily activities changes can be easily known with respect to the time i.e. regular usage duration with allowable residuals of certain objects can indicate the regular behaviour of the elderly and if there are any changes to the usage duration then we can say that there is an irregular activity. Therefore, we model the measurement system in terms of time series analysis for effective forecasting the usage duration of appliances.
The system interprets all the essential elderly daily activities such as preparing breakfast/lunch/dinner, showering, rest room use, dining, sleeping and self-grooming. Basically, the wellness of the elderly is entirely determined on the study of the usage of appliances. Importance of the system is that it has been designed and developed for using in an existing elderly home rather than a newly constructed house or test bed scenario.
Based on the time series of past data, a suitable method considered for predicting the near future values was “Seasonal (cyclic) Decomposition” [15]. It is used primarily as a preliminary tool when attempting to analyze trend. It is also suitable for exhibiting seasonal pattern which may be existing in the series and useful for forecasting process. Trend component is estimated by using the principle of moving average. The Exponential Moving Average considered in the analysis is given by the equation:
B. Wellness Determination: The Wellness functions as presented in [13], determine indices values based on known duration values of the different activities of elderly under care. Activity information is obtained from the interview of a respective elderly during the installation of the system. The obtained sensor activity data are amended while the system is running. For example, if the person gets up from bed at different time in the morning, the latest time is taken as the normal time of rise in the morning. It does not take into account of the duration the elderly is sleeping.
MAt+1 = αXt + (1-α) MAt
(1)
MAt+1-Moving average prediction MAt - Previous Moving average α -- Smoothing Constant Xt -- Observed quantity at time‘t’ The smoothing constant (‘α’), is derived from the number of sensor observations from the start of the system to the recently observed value. The advantage of this moving average is that a smaller smoothing constant gives more relative weight to the observations in the more distant past and a larger smoothing constant provides more weight to the most recent observation. The basic features like trend and seasonality describe a time series by its degree. After estimating the internal components like trend and seasonality of a time series, we have extracted errors by de-trending. Smoothed Trend Curve (STC) for various household usage durations is derived by applying eq (1).
The wellness function calculated in this manner [13] does not take into account day of the week, weekly, monthly and seasonal variation. Though the daily activities of an elderly are assumed to be remaining constant, but the variation of the seasonal weather can have a strong influence on it. Moreover, with time the daily activities of the elderly also get changed. It is important to take notes of those changes and included in the model. Instead of taking fixed maximum time of over-usage (Tn) and no-usage (T) as given in [13] for calculating wellness functions, it will be more practical to use updated maximum time duration of the appliances including seasonal factors to determine β1 and β2 effectively. In order to determine the updated time parameters and record the past maximum durations, Time Series modeling is applied. This will also provide a trend on the usage of appliances and perform near forecast based on the past usages. The appliance usage duration tendency is very much helpful in prediction as well in determining the behavior of the elderly. Potential advantage is to update the run-time duration usages of a particular appliance in monitoring state. This will enable to consider the updated maximum usage duration of appliances as stated in [13].
The seasonally adjusted factor is resulted from the decomposition process considering one week as one season (cycle). Following the additive method, for the forecasting process, the most appropriate fitted curve is computed by adding smoothed trend curve and seasonally adjusted factors. For demonstration purpose, we considered forecasting using additive model which identifies the excessive amount in the sensor duration dependent variable with respect to the time. Implementation of forecast model is further discussed in the following section.
18
IV.
EXPERIMENTAL RESULTS
Fig. 5 shows the front end of the developed software system indicating sensing signal of the household appliance is either active or non-active. Real-time sensing activity at the corresponding hour of the day is recorded simultaneously in the respective files of the centralized measurement system for determining wellness indices. The developed Graphical User Interface reveals the sensing signal status with their respective symbols and information like: number of times used every day, minimum usage duration, average usage duration, maximum usage duration, last effective time used and non-active duration since last used. Also, console window show the sequence of receptors usage.
(a): Subject #1: Bed usage durations
(b): Subject #1 Toilet usage durations.
Figure 5. GUI of the wellness measurement system
(c) Subject #1.Dining chair usage durations Figure 6: Subject #1 household appliances usage durations and their corresponding trends. (Green color:Trend, Blue color: Actual Observations)
For illustrating the forecasting process, we have presented the non-electrical appliances usage durations and their trends. This will elucidate the exact behavior of the elderly person in utilizing the household appliances. Some of the electrical appliances such as water kettle, microwave, and laundry machines are preprogramed and auto control. Their past usage durations may not be useful to infer how well the elderly person is able to perform their ADLs. Based on the eq(1), usage duration of the appliances and their corresponding trend are plotted. Fig. 5 shows some of the household appliances sequence plots of sensing activity durations and trend for eight weeks at an elderly house living alone.
From the Fig. 6, it was observed that the time series is not stationary. It is obvious that human behavior is complex and the activity durations may not be constant. In order to have a reasonable ninth week forecast value from the past sensor activity durations, we have investigated Seasonal-Auto Integration with Moving Average(S-ARIMA) and observed that this method is apt for forecasting and measuring wellness. The values of the S-ARIMA process such as trend, seasonal adjusted factor and residuals of the regression parameters are derived by using the Autocorrelation (ACF) and Partial Autocorrelation (PACF) functions of the time series.
19
Figure 8 : Residual autocoreraltion and partial autocorrelation function of time series of the sensing durations for Dining Chair sensing unit
Figure 9: Residual autocoreraltion and partial autocorrelation function of time series of the sensing durations for bed sensing unit(Sleeping activity)
The implemented process in the system is SARIMA (ps, ds, qs) (Ps, Ds, Qs), where ps: is order of process AR, Ps: is order of seasonal process AR, qs: is the order of process MA, Qs: is the order of MA, ds: is the order of difference, Ds: is the order of seasonal difference. On the basis of residual ACF spike at lag 1 and declining toward zero and residual PACF spike at lag1 and is also declining toward zero from lag2, this guide us to select SARIMA (1, 1, 0) (1, 1, 0)7 model for forecasting process. Following the additive method for the forecasting process, the most appropriate fitted curve is computed by adding smoothed trend curve and seasonally adjusted factors i.e. Forecast=Trend + SAF.
Two instances were rightly identified as irregular usage of the bed as the elderly has wake up early at one instance and slept for longer duration in the second instance. Fig.10 shows the 9th week predicted bed usage duration. The significance levels of the residuals are less than 2%. Suitability of the predicted curve with respect to observation sequence is also verified by implementing One-Sample Kolmogorov-Smirnov Test (KS-test) [15] for normal distribution of the errors existing in the predicted curve. Fig.11 indicates the errors of predicted fitting as a normal distribution. Table. 1: Prediction of 9th week Bed,Toilet and Dining Chir usage durations based on SARIMA (1,1,0)(1,1,0)7 process th
Seasonally Adjusted Factor (SAF) is resulted from the decomposition process considering one week as one season (cycle). Accordingly the fitted curve and the forecast are shown in fig.10. Considering 95% confidence interval, the residuals prevailing in the prediction curve are computed by twice the standard deviation. Table. 1 show the forecast range values for ninth week based on eight week sleeping durations and comparing with the actual values. Except at two instances, remaining five days bed usage durations have matched correctly.
Bed Usage Forecast Actual Max Min 8:22:11 4:42:11 8:16:17 9:24:57 5:44:57 4:38:13 8:47:36 5:07:36 8:17:46 8:00:02 4:20:02 7:38:17 8:34:19 4:54:19 7:51:38 8:07:15 4:27:15 8:15:01 7:43:36 4:03:36 7:40:32
20
9 Week Duration(hh:mm:ss) Toilet Usage Dining Chair Forecast Forecast Actual Actual Max Min Max Min 0:11:32 0:07:45 0:10:51 0:31:29 0:26:20 0:28:45 0:10:21 0:06:15 0:09:10 0:28:45 0:24:10 0:25:24 0:11:25 0:07:20 0:08:55 0:29:10 0:25:38 0:27:49 0:11:50 0:07:35 0:07:10 0:28:10 0:24:48 0:24:20 0:10:48 0:06:45 0:09:54 0:27:35 0:23:45 0:26:36 0:12:10 0:08:28 0:10:25 0:29:20 0:24:46 0:28:20 0:11:50 0:07:26 0:10:45 0:30:10 0:26:50 0:27:10
VI. REFERENCES 1.Smarr C.A, Fausset C. B and Rogers W. A, “Understanding the potential for robot assistance for older adults in the home environment”, Technical Report-HFA-TR-1102, School of Psychology, Human Factors and Aging Laboratory-Georgia Tech- Atlanta, http://hdl.handle.net/1853/39670 , 2011. 2.Rahimi S, Chan A.D.C, Goubran R, “Usage Monitoring of Electrical Devices in a Smart Home”, Proceedings of the IEEE International Conference on Engineering in Medicine and Biology-EMBC11, Boston, U.S.A, pp. 5307-5310, Sep-2011. 3.Arcelus A, Veledar I, Goubran R, Knoefel F, Sveistrup H and Bilodeau M, “Measurements of Sit-to-Stand Timing and Symmetry from Bed Pressure Sensors,” IEEE Transactions on Instrumentation and Measurement, vol. 60, no. 5, pp. 1732-1740. 4.Amaya A, Rafik G, Heidi S, Martin B, Frank K, “Context-Aware Smart Home Monitoring Through Pressure Measurement Sequences, Proceedings of the IEEE International Workshop on Medical Measurements and Applications, Ottawa, Canada, pp. 32 – 37, April 2010. 5.Boni A, Pianegiani F, Petri D, “Low-Power and Low-Cost Implementation of SVMs for Smart Sensors”, IEEE Transactions on Instrumentation and Measurement, vol. 56, no.1, pp.39-44. 6.Kwang E. K, Hyun C.Y, Kwee B. S, “Emotion Recognition using EEG Signals with Relative Power Values and Bayesian Network”, International Journal of Control, Automation and Systems, Vol. 7, Issue. 5, pp. 865-870, Oct-2009. 7.Jakkula V.R, Cook D, Jain G, “Prediction models for a smart home based health care system”, Proceedings of the 21st International Conference on Advanced Information Networking and Applications, pp. 761 - 765, 2007 . 8.Arabnia H.R, Wai C.F, Changhoon L, Yan Z, “Context-Aware Middleware and Intelligent Agents for Smart Environments,” IEEE Intelligent Systems, Vol.25, Issue:2, pp.10 – 11, 2010. 9.Jae H.S, Boreom L, Kwang S.P, “Detection of Abnormal Living Patterns for Elderly Living Alone Using Support Vector Data Description”, IEEE Transactions on Information Technology in Biomedicine, Vol. 15, No. 3, Page(s):438-448, May 2011. 10.Tibor B, Mark H, Michel C.A.K, Jan T, “An Ambient Agent Model for Monitoring and Analyzing Dynamics of Complex Human Behavior”, Journal of Ambient Intelligence and Smart Environments, Vol 3, No. 4, Page(s): 283-303, 2011. 11.Suryadevara N.K, Gaddam A, Rayudu R.K, Mukhopadhyay S.C, “Wireless Sensors Network Based Safe Home to Care Elderly People: Behaviour Detection”, Elsevier-Sensors and Actuators: A: Physical, Vol.186, Pages 277-283, 2012. 12.Gaddam A, Mukhopadhyay S.C, Gupta G.S, “Elder Care Based on Cognitive Sensor Network”, IEEE Sensors Journal, Vol. 11, No. 3, Page(s): 574 – 581, 2011. 13.Suryadevara N.K, Mukhopadhyay S.C, “Wireless Sensor Network based Home Monitoring System for Wellness Determination of Elderly”, IEEE Sensors Journal, Vol: 12, No: 6, pp.1965 – 1972, 2012. 14.Susan M, Juan Y, Lorcan C, Bleakley C, Dobson S, “Activity Recognition using Temporal Evidence Theory”, Journal of Ambient Intelligence and Smart Environments, Vol. 2, No. 3, pp.253-269, 2010. 15.Brockwell P.J and Davis R.A, “Introduction to Time Series and Forecasting”, Springer, 2nd edition, pp.326-330, 2001.
9th week Forecast
Figure 10: Eight week sleeping observations and Ninth week predicted sleeping durations. (Green color: Fitted and Forecast, Blue color: Actual Observations)
Figure 11: K-S test result for Normal distribution of predicted sleeping durations
V.
CONCLUSION
In this paper, we presented measurement of daily activities through usage of household appliances sensor data. Predicting the behavior of an elderly person was based on past sensor activity durations. Combination of sensing system with time series data processing enabled us to measure how well an elderly person is able to perform their daily activities in real time. So far, the forecasting process was able to rightly measure the wellness indices related to use of non-electrical appliances. Hence, some of the basic elderly daily activities such as sleeping, toileting, dining and relaxing are rightly assessed by the wellness measurement system. Since, most of the electrical appliances usage durations are predefined; validation for activities such as preparing food is limited. However, additional data processing method such as sensor sequence activity pattern analysis was able to rightly measure the occurrences of activities such as preparing breakfast, lunch, dinner and snacks. The next step will be to devise a robust forecasting method including outliers in the wellness measurement system.
21
