Propagating sensor uncertainty to better infer office occupancy in smart building control

Occupant presence and behaviour in buildings is considered a key element towards building intelligent and pervasive environments. Yet, practical applications of energy intelligent buildings typically suffer from high sensor unreliability. In this work, we propose a layered probabilistic framework for occupancy-based control in intelligent buildings. We adopt a cascade of layers, where each layer addresses different aspects of the occupancy detection problem in a probabilistic manner rather than in a hard rule engine. We show that propagating uncertainty through each layer instead of standard hard decision outcomes improves the overall system performance. This ﬁnding suggests that smart building interfaces and communication data formats may need to input and output probabilistic data rather than simple discrete classiﬁcation outputs. System performance and user comfort were evaluated with real life radar sensor data, based on an algorithm that allows real-time (casual) processing. Energy savings of up to 30% were obtained, compared to baseline measurements, while maintaining user comfort. © 2018 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY license. ( http://creativecommons.org/licenses/by/4.0/ )


Introduction
Energy conservation is becoming an increasingly important topic due to the rising energy demands and diminishing energy resources.Since commercial office buildings account for the highest total energy consumption [1] of any building type, considerable effort has been dedicated to make buildings more energy efficient.Occupancy is considered a key element towards this direction and has long been used to control and dynamically adjust energy-related appliances and building systems (HVAC, lighting or other appliances).Those systems are based on motion or presencesensing modalities (sensor nodes) distributed within a building, for example to switch lights on or off, or to properly ventilate a building space.
Over the past decade, numerous systems have been proposed for smart building control.For example, the authors in [2] developed a system that tracks user movement in building spaces using a camera network solution called SCOPES [3] .A framework, based on closest distance Markov chains (CDMC) is used to capture the dynamics of user occupancy, with the aim to save energy in various areas, most notably HVAC and lighting.In [4] , building occupants are equipped with sensor badges, with which it is possible to achieve relatively accurate localization based on support vector machines (SVM).Those systems, however, are based on advanced sensor modalities, such as cameras, which are expensive and wearable devices, which may generate privacy concerns.The domain of building control can tolerate a loss in accuracy in favour of installation cost and privacy, thus, simple, binary sensors are preferred, that are easy to retrofit in existing buildings and comply with the existing privacy regulations.For instance, in [5] , a Hidden Markov Model based on PIR sensors was used to estimate the probability of home occupancy and automatically turn off the HVAC system when the occupants are away from home.In [6] , a similar model was proposed to improve presence detection for office lighting control.In [7] , authors fused information collected by a network of traditional PIR sensors in a Bayesian probabilistic framework, while authors in [8] developed an event-based pattern detection algorithm based on ambient sensing data, such as, lighting, acoustics, CO 2 , and temperature.Nonetheless, common sensing technologies still have limitations, in particular high percentages of failure to successfully detect a person that is sitting still, reading, typing, or watching a video, thus without making large movements with the arms or the body.
Radar sensing technology offers a promising approach to overcome the limitations of the common occupancy sensors.In recent years, there has been an expansion in the use of radar sensors for remote detection and noncontact measurement of humans and huhttps://doi.org/10.1016/j.enbuild.2018.08.039 0378-7788/© 2018 The Authors.Published by Elsevier B.V. This is an open access article under the CC BY license.( http://creativecommons.org/licenses/by/4.0/ ) man properties, including intruder detection and health diagnostics.One of the first demonstrations of the feasibility of remote respiration measurements with microwave radar was presented in [9] .Yet, the distance between target and radar sensor was limited to only 30 cm, much less than typical distances in building applications.Authors in [10] presented a more advanced Doppler radarbased method for the detection of vital signs, presence, and activity.However, in order to facilitate detection, shielded aluminium coating was used in the surfaces of the test room.The detection performance in real-life space was not demonstrated.The feasibility of radar sensing for the detection of small movements (i.e.typing) was demonstrated in [11] .However, the algorithm was not able to detect a person sitting still without being involved in any type of motion, a scenario rather relevant for building applications.In all cases, although microwave radar technology has gained increased popularity for activity detection, its feasibility and implementation in the context of smart building and control applications remains limited.
In summary, control strategies can be implemented using several available sensing technologies.Yet, most sensor based approaches suffer from high sensor unreliability, in particular in the detection of minor motions, such as typing.To this end, radar sensing technology offers a promising solution that still remains not fully explored.In order to reduce the measurement uncertainty, stochastic models like Bayesian networks, Gaussian mixture models and support vector machines have been implemented to improve occupancy detection.An alternative approach is to approximate office occupancy as a Markov process that moves between a number of states.Our challenge is to determine the current state of that process (e.g., presence or absence) from the sequence of observations made by (imperfect) sensors.Since office occupancy does not change very rapidly, preferably not only the latest sensor readings are considered, but also observations done in the (recent) past can improve the reliability.This concept is known as a Hidden Markov Model (HMM) and has been extensively explored for other applications.In our previous work [12,13] , we showed that HMMs provide an effective framework to reliably estimate the state of office occupancy, even if real human behaviour deviates from being a random Markov process.However, to be used with radar sensor data, we found that a single-layer HMM generated a prohibitively large parameters space and resulted in high classification errors, which made it not appropriate for smart building control applications.Therefore, inspired by the concept of Stacked Generalization, [22] , where the main idea is to combine classifiers in order to achieve higher predictive accuracy, we extend the typical HMM framework and propose and test a layered classifier representation.We refine the Stacked Generalization approach by letting each layer address different aspects of the occupancy detection problem.A low layer is added to model the uncertainties related to the sensor input and address the technological imperfections of the sensing technology.To this end, a Bayesian model is a suitable approach as a principled and pragmatic framework for representing the uncertainty related to the radar sensor input.The second, higher layer provides an appropriate framework to model the time-varying behaviour of office occupancy.Office occupancy may be a stochastic process, however, a chronological sequence of events can be observed (e.g.absence is followed by a walking event before any kind of desk work activity can be detected).Motivated by this, we construct a Hidden Markov Model (HMM) on the top of the first-layer.The following are the key contributions of our work: • A layered approach to occupancy-based control that achieves a logical separation of the measurement uncertainty related to the sensor network from human behaviour in an office context.The Bayesian probability framework offers a principled ap-proach to model the sensor measurement uncertainty, while, the Hidden Markov Model captures the time-varying behaviour of office occupancy.• We explicitly propagate uncertainties instead of hard classification decisions, enabling the decision layer to consider the consequences of plausible misclassifications and act accordingly.We adopt the notion of Bayesian probability as the formal quantification of subjective uncertainty.• We evaluate performance using real radar sensor data from a regular office environment.We explore the use of microwave sensors as occupancy sensors for indoor intelligent building applications.Radar sensors have been widely used for applications like ground surveillance and health monitoring, however, their application in smart building occupancy monitoring is not well studied.

Data collection
The data collection was carried out in Philips Lighting Turnhout, Belgium, using a commercially available CW radar sensor operating at the 24 GHz ISM band.A schematic drawing of the experimental set-up is shown in Fig. 1 together with a photo taken at the time of the experiments.The radar consists of two analog outputs, which provide the in-phase (I) and quadrature (Q) components of the received signal.The output power of the radar is 20 dBm and the gain of the antenna approximately 24 dB.The radar sensors were used to collect Doppler signatures of two human targets performing the following activities: • Walking, at different speeds towards and away from the sensor, at various distances from the sensor, ranging from 0.5 to 5.6 m (see blue square area in Fig. 1 a).
• Desk work at a distance of 0.5 m from the radar.The placement of the desk is shown in Fig. 1 a.The human targets were asked to perform typical desk work activities like typing, browsing the net etc.
The experiments were carried out over a 1-day period and all subjects were measured on the same day.Each target was measured for three sessions of ten minutes walking at various speeds and angles from the sensor, resulting in overall 3600 recordings of walking activity, each one second length.Three sessions of five minutes of the targets performing typical desk work were recorded, resulting in an overall of 1800 Doppler signature frames of desk work activity, each one second length.Additionally, the empty room was measured for a period of 20 min, to obtain the Doppler signature of target class absence.The I and Q components of the target echoes where digitized at a sampling rate F s = 44.1 kHz and memorised in a file for post processing.
In parallel to the radar sensor data collection, in order to collect information about typical occupancy profiles in an office, we implemented a baseline study in a typical office at Eindhoven University campus, with desks occupied with PhD students.Participants were asked to manually annotate their activities and walking in and out times on a pencil-and-paper form.Participants were asked to maintain their working style as usual.No activities were scripted in any form.In total, 10 participants were recorded for 15 days, resulting in a total of 150 daily 24-hour profiles of presence of occupants in their offices.

System architecture
We developed a layered framework based on a cascade of classifiers, where, the inferential results of each layer are used as input to the consequent layer.Three types of motion (classification classes) are defined, as recommended by the NEMA standard for occupancy motion sensors, [14] namely  of a first-level Bayesian classifier are used as input for the second level HMM.We compare a framework built on traditional hard decision approach to a framework where soft classification decision is passed to consequent layers.Although both systems process the same initial sensory information, propagating uncertainty through the entire architecture results in better decisions in the final layer.
• Absence (Class H 1 ) The general architecture scheme of the proposed occupancy-based control strategy is presented in Fig. 2 .

Data preprocessing
The accuracy of motion recognition heavily depends on the selection of features that can best distinguish the different types of motion.The choice of feature extraction method depends strongly on the type of sensors that consist the monitoring system.The classification features used in this work are the filter-bank energy coefficients (FBEs).To implement this filter-bank, the signal is sliced into short-time frames, transformed using a Fast Fourier transform (FFT) and the magnitude is taken.The magnitude coefficients are then binned by combining them with a triangular filter.Here binning means that each FFT magnitude coefficient is multiplied by the corresponding filter gain and the results are accumulated.Thus, each bin holds a weighted sum representing the spectral magnitude in that filter bank channel.Fig. 3 illustrates the general form of this filter-bank.Normally the triangular filters are spread over the whole frequency range from zero up to the Nyquist frequency.However, band-limiting is often useful to reject unwanted frequencies or avoid allocating filters to frequency regions in which there is no useful signal energy.In our filter-bank analysis, lower and upper frequency cut-offs are set to zero and 300 Hz respectively.It makes no sense to design the upper frequency limit for detecting human beings as higher than 300 Hz, since this would correspond to speed of 6.8 km/h (pretty fast for office activity).The resulting N selected filter-bank energies define the feature vector x = [ x 1 , . . ., x N ] .

First layer Bayesian model
The raw sensory input, after the initial pre-processing step, is passed through the first-level Bayesian model with the goal to reduce the uncertainty related to the radar sensor measurements.Firstly, we construct the Naïve Bayes probability model [15] .We let the feature vector x = [ x 1 , . . ., x N ] represent the N selected filterbank energies ( N is the number of bins).For each target class, i.e., absence ( H 1 ), major ( H 2 ) and minor motion ( H 3 ), the selected features (FBEs) are assumed to be Gaussian distributed and statistically conditionally independent from each other given the target class H . Thus, their joint probability function is expressed as where, μ n | H k and σ 2 n | H k are the mean and standard deviation of the n th feature of target class H k , respectively.The mean value and standard deviation of each feature, for each target class, were estimated by using corresponding sequences belonging to a training set.Using Bayes theorem, the conditional probability In practice, there is interest only in the numerator of that fraction, as the denominator does not depend on the class H , and the values of the features x are given, so that the denominator is effectively constant.

Design implementation
The inferential results of the first layer are the input to the second layer HMM that models the time-varying behaviour of office occupancy.It is an open question what type of information should be passed on as input to the next layer.In a typical control system, the state-of-art output of a classification is the hard decision on state, suggesting that this can be a suitable input for the following layers.Yet, inspired by the popular concept of soft decision decoding [16] , we explore whether passing the conditional likelihood of each target class (soft decision), results in better performance.In information theory, a soft-decision decoder is a decoding method where, whereas a hard-decision decoder operates on data that take on a fixed set of possible values (typically 0 or 1 in a binary code), the inputs to a soft-decision decoder may take on a whole range of values in-between.This extra information indicates the reliability of each input data point, and is used to form better estimates of the original data.Similarly, in the context of a smart building architecture, since misclassification in an early stage affects the overall system performance, at the very least, each layer should pass on the limitations and uncertainties of its outputs.The implementation of both strategies is described as follows: Hard decision : The hard decision implementation is straightforward.A maximum a posteriori (MAP) decision rule assigns the class label ˆ y = H k to the most probable class as follows: The hard classification prediction of the first level Naïve-Bayes classifier ( ˆ y = H k ) will be the input for the second layer HMM.Soft Decision : The Bayesian probability theory provides a principled approach to model uncertainty.It allows the estimation of the probability of the feature vector belonging to a target class H .In this implementation, the first-level classifier does not pass the hard decision of the classification to the second layer, but the "likelihood" of belonging to specific class, expressed as the conditional probability P ( x 1 , . . ., x N | H k ) .This likelihood probability serves as a confidence measure for the prediction made in the first layer or simply, a degree of belief from a Bayesian point of view [23] .

Second layer-Hidden Markov Model
The Hidden Markov Model (HMM) of the second layer provides a framework to capture the time-varying nature of occupancy in an office environment.We extend our previous work in [13] , where human presence was modelled as an on-off process, by allowing a Markov process with three possible discrete user states, namely, q t = 1 represents the state that the user is absent, q t = 2 the state that the user is walking, whereas q t = 3 the state that the user is present in his desk.Those states represent a one-to-one correspondence to the NEMA recommended types of motion ( H k ) previously described.The possible user states are unknown to the system (hidden), but can only be observed through another set of stochastic processes that produce the sequence of observations, i.e., the radar sensing network, as shown in Fig. 4 .An HMM can be fully described by the following set of parameters, namely, the state transition probabilities, an observation model (emission probabilities), and the initial state distribution [17] : State transition probability matrix : A = { a i j } , a i j = P ( q t = j| q t−1 = i ) .The transition probabilities describe how the occupancy of a space changes over time.

Emission probability matrix
where, the sensor observation r t , which of course depends on the current state q t , corresponds to the physical output of the system being modelled.In essence, the emission probabilities are a metric that describes the quality of the sensor.In the typical HMM formulation, the observation symbols are a time sequence of the raw sensor signals.In this work however, the input to the HMM model is the inferential result from the previous layer.Depending on the input, hard or soft decision, the emission probability matrix is formulated as follows: • Hard decision : The observation vector is the hard decision output of the first layer classification ( r t = H k , k ∈ { 1 , 2 , 3 } ).The emission probability matrix, in this case, is estimated experimentally using labeled sequences of observations and states in a training step.
• Soft decision : The observation vector is the filter-bank energies coefficients r t = x 1 , . . ., x N and the emission probabilities are defined as the likelihood of belonging to each class, i.e., b k ( r t ) = P ( r t = x 1 , . . . ,x N | H k ) .In that case, the soft output of the Bayesian model immediately defines the emission probability.Initial state probability vector : π = { π i } .The initial state distribution specifies the occupancy probability at the initial time step t = 0 prior to any observation.
Based on the constructed HMM, we develop an approach to calculate the a posteriori probability of user state taking into account a priori knowledge on user behaviour, i.e., the series of sensor observations so far.In particular, for a given model λ = { A, B, π} , the forward algorithm [17] inductively calculates the forward variable α t ( q t = H k ) = P ( r 1 r 2 . . .r t , q t = H k | λ) , i.e., the probability of the observation sequence until time t and state at t , given the model.Any information about the history of the process needed for future inferences is reflected in the current value of the forward variable.
We can solve for α t ( q t = H k ) inductively as follows: Using the Bayes theorem, the conditional probability of state H k given the history of sensor observations, r 1 r 2 . . .r t , can be decomposed as There is interest only in the numerator of that fraction, as the denominator does not depend on the class H .The HMM algorithm chooses the most likely state H k at time t , given the history of observations, according to a maximum a posteriori (MAP) decision rule: In a typical hard rule engine implementation, this hard classification result would be propagated to the final decision layer.
Yet, we exploit the fact that the joint probability, α t ( q t = H k ) = P ( r 1 r 2 . . .r t , q t = H k | λ) , is a metric that reflects the system's degree of belief on each possible target class.It is a variable that encapsulates the estimation on user state together with the corresponding uncertainty, thus, can be a critical soft input for the follow-up decision layer that enables more accurate state estimation.

Decision layer
The final block in the control architecture is the decision strategy, i.e., how the system translates the a posteriori probability of every class into a meaningful decision.In the context of building control, the final decision depends on whether the user is present or absent, thus, the final decision is made between those two possible user states.We use a Majority Voting Decision Rule (MV) approach to the final classification problem.The MV framework proposes the transformation of k -class constraint classification into binary classification in higher dimension [18] .We adopt an AvA (all-versus-all) approach as it allows to perform log-likelihood ratio tests for discrimination between every pair of classes i and j Q i j = log The majority voting decision rule enables to adjust the thresholds in order to minimise the classification errors for each pair of target classes.Substituting (3) -( 4) , pair-wise comparisons can be re-formulated as The pair-wise decisions are combined by voting, that is, the class with the most pair-wise wins is selected.If the winner class is a presence class (minor or major motion), the final decision is presence.

Simulation results and discussion
In this section, we evaluate the classification performance of each intermediate classification layer.It is important to accurately quantify the uncertainty of intermediate layers in order to correctly propagate uncertainty to the following layers and make an informed decision.The metrics used to quantify the classification performance include precision -recall statistics and confusion matrices.Each classification layer should be able to distinguish between the three possible types of motion recommended by NEMA, namely, absence, walking (major activity), and desk work (minor activity).Representative radar signatures of each type of motion are presented in Fig. 5 .The corresponding spectrograms of these activities, presented in Fig. 6 , show that walking has a unique and easily distinguishable time-frequency characteristic.Similarly, we noticed that typical movements during desk work, like moving your head are easy to distinguish.However, minor motions, such as typing, are quite hard to distinguish from absence.True positive ( tp ) = correctly identified instances False positives ( fp ) = incorrectly identified instances False negative ( fn ) = incorrectly rejected instances

First layer classification performance
The confusion matrix and the corresponding precision -recall statistics of the first-level Bayesian classifier are presented in Tables 1 and 2 , respectively.For the classification tests, the system input was a series of signal frames to be classified as one of the possible target classes.Each classification frame is of 10 0 0 msec duration.A feature vector of 40 filter bank energy coefficients was extracted for each frame.Fourfold cross validation was performed with three folds used for training the Bayesian model and the rest to validate it.All simulations are performed in MATLAB.The classification results confirm the difficulty of Bayesian modelling to distinguish between minor motions and absence.In fact, 31.7% of desk work events were misclassified as absence.This poor performance of the simple first level classification motivated the addition of a higher level HMM layer.

Second layer classification performance
The proposed layered architecture, leverages the ability of the HMM to capture the temporal behaviour of office activity, while using the Bayesian model to capture the uncertainty in sensor observations.Our simulations are based on measured target activity trajectories (absence, walking from/ towards the sensor and desk work), thus, not necessarily matching any pre-assumed Markov process.During each state, pre-recorded Doppler signatures of each activity were fed into the detection algorithm to simulate fine grain sensor samples for the 150 daily occupancy profiles from the baseline study.The corresponding confusion matrices and precision-recall statistics are summarised in Tables 3-5 .
The additional HMM layer improves performance, i.e., the percentage of misclassifying minor events (desk work) as absence decreases from 31.7% to 20% and 14.7%, for hard and soft decision implementation respectively.Further improvement is expected by passing the soft decision result of the HMM model to the final layer as will be confirmed by the final layer performance.A decline in the precision of walking activity is reported since instances of walking are often misclassified as desk work.This stems from the fact that the walking state in the HMM model has a low prior probability, i.e., the system spends more time on desk work or absence states rather than the walking state, thus, the classification result tends to lean towards those states.Yet, since the final decision will be on presence/absence, this misclassification will have a minor impact in the overall system performance.

Final layer performance
In the context of building control, the final decision is typically made between presence and absence.The system exploits information on occupancy to turn the lights on or off, or to adjust other energy related appliances.Besides the general goal to maximise the potential energy savings, user comfort is an essential success criteria that should remain a critical design aspect in building control.In order to optimise this tradeoff between energy savings and user comfort, we need to consider appropiate performance evaluation metrics.To this end, we choose the false negative rate (FNR), i.e., the total time that the algorithm wrongly assumed non-presence normalised presence time, as a metric that reflects user annoyance (appliance turns off during presence).The false positive rate (FPR), defined as the total time that the algorithm wrongly assumed that a user is present normalised over total absence time, can be interpreted as a metric of energy cost (appliance would remain on while the user is absent).The classification performance is subject to a trade-off between false positive and false negative rates that can be influenced by the threshold setting ( δ ij ) in the final decision layer.A threshold setting that is reasonable for illumination control (low probability of erroneously switching of the light), but may be less appropriate for other smart building applications.Also, during periods of energy scarcity the false negative rate may need to be jeopardised a bit to more aggressively save energy by eliminating more false positives.This can be achieved by appropriate choice of thresholds.A more generic comparison can be made by Receiver Operating Curves (ROCs), which plot the selection of all achievable combinations of FNR and FPR, depending on the choice of threshold.The ROC curve of Fig. 7 graphically illustrates the trade-off between user annoyance and energy cost of the suggested architecture for various choices of δ ij thresholds.As expected, a softdecision implementation performs better in the presence of noisy sensor data than its hard-decision counterpart, suggesting that the uncertainties associated with each observation vector should be propagated to the next layer instead of simply passing the hard classification outcome.This finding is better illustrated with a simple visual example in Fig. 2 .In this scenario, a hard decision in the first stage of the architecture propagates confidence that the user is absent, resulting in turning the lights off.In contrast, the soft decision implementation, still makes a mistake on the most probable state, but propagates its uncertainty to the next layer.At the final decision stage, the control component reacts to this uncertainty, resulting in a completely different outcome.

Open-plan office example
We wish to address the potential energy savings of a control system based on the suggested architecture with an openplan office simulation.We focus on lighting control as an example that demonstrates the potential reduction in lighting energy consumption.In lighting control, two factors are of major importance, namely, energy savings and user comfort.We assess the potential energy savings by estimating the expected annual energy consumption (in kWh) of the lighting system.We use the number of false negatives, i.e., the number of times that the control system erroneously turned the lights off when the user is present, as a metric that reflects user annoyance.In typical lighting control, high level requirements suggest keeping the number of false negatives less than one in 160 hours.In that case, the accepted percentage of false negatives lies in the range of 10 -6 .Therefore, in most practical lighting control applications, controllers have a pre-set time delay, i.e., the duration of time from the last time a motion was detected until the controlled lighting loads are deactivated.Projected annual energy consumption of lighting system.The suggested layered architecture is compared against manual and optimal control.Typically, the energy consumed during unoccupied hours in commercial buildings is more than 40% of the estimated load.
We compare the performance of the suggested layered architecture to the baseline manual control system.Manual control considers the situation where users would operate the lighting system, assuming that the first person that enters the room will switch the lights on, and the last person that leaves the room will switch the lights off.It should be noted however that according to recent empirical studies, [19,20] the actual after-hour energy consumption in buildings is much higher than expected, often more than 40-80% of occupied load.Thus, the actual energy consumption by a useroperated lighting system is expected to be higher than the manual control consumption calculated.The performance of a hypothetical genius classifier is also compared.This scenario corresponds to optimal control performance.
We consider an example lighting system in an indoor office with 12 LED dimmable luminaires [25 W].An example illustration is shown in Fig. 8 .The workspace plane is assumed to be divided into 12 logical zones, as shown in Fig. 8 b.Each luminaire consists of a local controller, an LED dimmable luminaire, a radar sensor and a communication module.The presence sensing region is assumed to be constrained to a zone.In practice, it may extend further and the presence of occupants outside the zone might affect detection.The dimming level of each luminaire is determined by the associated local controller, such that the resulting illumination achieves target illuminance values in the corresponding zone in the workspace plane.Each node runs the layered presence detection algorithm and makes an estimate of the local occupancy state.The local information on zone occupancy state is communicated between neighbouring nodes.Based on sensing information and coordination with neighbouring controllers, each controller determines the dimming level of its luminaire.
For the simulations, we use sample occupancy profiles from our baseline study to generate a randomised annual occupancy profile for all desks in the open-plan office.The expected annual energy consumption is simply the power consumed by electrical lighting (in kWh) during the year assuming that lights in a zone provide 500 lx when the zone is occupied, dim to 300 lx when only the neighbouring zones are occupied, and switch off automatically as soon as the occupancy sensor estimates absence in a zone after a fixed delay period, following recommended norms for office lighting.[21] According to our simulation analysis presented in Fig. 9 , the suggested solution was able to achieve energy savings of up to 30%, compared to baseline manual control, without too much energy wasted during un-occupied periods (false positives).Savings are expected to be higher when considering the energy wasted during non-occupied hours in commercial buildings.The high-level requirements for false negative ratio in the range of 10 -6 was achieved with setting the time delay to five minutes.

Conclusions and future directions
In this work, we respond to the need for improved detection of building occupants by developing a layered architecture for smart building control.A layered representation allows for the separation of the architecture into separate building blocks, where, cleanly separated algorithms may run on different devices / entities.The suggested solution combines the advantages of a Hidden Markov Model that naturally captures the temporal structure of office occupancy with Bayesian modelling, which captures the physical properties of the sensor network.The inferential results of each layer are used as input to the consequent layer.We discuss the limitations with making hard decisions, asserting that soft decisions should be propagated throughout the decision layer.Our results seem to justify that typical communication standards and message formats may have to be extended.We recommend soft decision values, in particular a posterior probabilities of the NEMA states, as a suitable candidate for sensor node messages.This also enables an appropriate partitioning of algorithms into decentralised and centralised parts with limited communication requirements.We explore the use of microwave sensors as occupancy sensors and evaluate performance using real data from a regular office environment.Our results showed significant energy saving potential in a typical open-plan office environment without sacrificing user comfort.Ongoing work includes our efforts to fuse information from multiple sensor sources within the building in order to improve the detection performance.A layered representation in combination with a communication protocol based on propagating a posterior values offers a flexible framework for combining information from heterogeneous sensor networks and ensures coherence.In fact, when combining data from different sensing modalities, which result in a different feature space, the a posterior values could be a versatile parameter in the interface.

Fig. 1 .
Fig. 1.Experimental set-up: (a) Sketch of layout.Sensors placed at positions F1-F4, at height 2.7 m from the floor.The targets were asked to perform walking activity inside the area defined by the square blue dotted line (total area is 16 m 2 ).Desk work activity performed behind the desk, at a distance of 0.5 m from the sensor, (b) Photo with Target walking towards the radar sensor.

Fig. 2 .
Fig.2.Suggested layered architecture.The initial raw sensor input is preprocessed to obtain the feature vector.Two layers of classifiers are suggested.The inferential results of a first-level Bayesian classifier are used as input for the second level HMM.We compare a framework built on traditional hard decision approach to a framework where soft classification decision is passed to consequent layers.Although both systems process the same initial sensory information, propagating uncertainty through the entire architecture results in better decisions in the final layer.

Fig. 3 .
Fig. 3. Filter-bank implementation: Each bin holds a weighted sum representing the spectral magnitude in that filter bank channel.

Fig. 4 .
Fig. 4. HMM model with states H 1 , H 2 , H 3 corresponding to NEMA classes.The observation symbol is the input from the previous layer.The Markov process evolves in time, taking a certain next state according to the transition probabilities a ij .

Fig. 6 .
Fig. 6.Spectrograms of Doppler signatures of different types of motion.

Fig. 8 .
Fig. 8. (a) Lighting system in an example office room, (b) Depiction of logical zones in the workspace plane (blue rectangles depict luminaires).

Fig. 9 .
Fig.9.Projected annual energy consumption of lighting system.The suggested layered architecture is compared against manual and optimal control.Typically, the energy consumed during unoccupied hours in commercial buildings is more than 40% of the estimated load.

Table 4
Layer 2 Confusion matrix-Soft decision in layer 1.