WO2017149324A1

WO2017149324A1 - Occupancy sensing

Info

Publication number: WO2017149324A1
Application number: PCT/GB2017/050581
Authority: WO
Inventors: Edward Hugh Owens; Joel CHANEY; David CORNE; Andrew Peacock
Original assignee: Heriot-Watt University
Priority date: 2016-03-04
Filing date: 2017-03-03
Publication date: 2017-09-08
Also published as: GB201603818D0

Abstract

A sensing system for use with at least one sensor having at least two defined states, the sensing system being adapted to: store distributions of sensor states about triggers associated with activities; receive signals indicative of sensor states from said at least one sensor; analyse the received sensor signals to produce distributions of sensor states for the received signals about triggers, and compare the distributions of sensor states for the received signals with the stored distributions of sensor states associated with activities to identify the activity or a change in the activity.

Description

Occupancy Sensing

Field of the Invention

The present invention relates to determining a change in occupancy in an environment using either a single or a multi-sensor system.

Background of the Invention

Most thermostats sold to the new build housing market are programmable. They are designed to be programmed by a user to match the user's lifestyle in order to reduce energy consumption and running costs. Programmable thermostats have proven to be ineffective due to public misconception about how they operate, the uninviting nature of user manuals and the complex nature of their user interfaces. There is a need to improve the functioning of a programmable thermostat.

Smart home technologies may play an increasingly important role in enhancing the lives of elderly people from health monitoring through to assistive control. These technologies can benefit from incorporating information pertaining to building occupancy. For instance, a thermostat can be automatically scheduled to maintain comfort when a property is occupied and to save energy at periods when it is not occupied. Likewise, lighting and appliances could be turned on and off when a property is occupied or unoccupied, or if there is a safety hazard. Security can also be maintained by locking doors when a property is unoccupied. By observing changes in activity, the well-being and health of occupants can be monitored. The effectiveness of implementing the above approaches is limited by the accuracy of occupancy detection. For sheltered accommodation for elderly residents, underperforming occupancy sensors could lead to an increase in vulnerability of residents. Therefore, there is a need for robust, consistent detection of living space usage.

Current approaches to occupancy detection are primarily focused at commercial buildings. Passive infrared sensors and ultrasound sensors are the most common commercially available and applied technologies. Typically, these are used in lighting control and act as passive devices i.e. switching on when motion is detected and switching off after a pre-determined interval has elapsed. In a study of commercial systems in offices, restrooms, classrooms and conference rooms the detection errors have been found to range between 0-28%, depending on the space in which the sensors are mounted [Neida et al. 2001 ]. For older people in a domestic setting with reduced levels of activity, technology that relies fundamentally on motion could incur false negatives, indicating that they are not in the space when in reality they are present but not moving. This problem becomes even more manifest for singly occupied elderly households where there tends to be fewer interactions. Furthermore, the systems tend to be based on single point detection and data is not available to be cross-checked or used by other control systems.

Other known approaches to occupancy detection include the use of video cameras and image processing; installing C0₂ sensors; analysing electricity data; monitoring MAC and IP addresses; using GPS on smartphones, and Wi-Fi fingerprinting. These approaches rely on continuous activity of an occupant or interactivity of an occupant with building infrastructure or equipment. Therefore, they are not applicable to the single elderly demographic of residential homes. Summary of the Invention

According to the present invention, there is provided a sensing system for use with at least one sensor having at least two defined states, the sensing system being adapted to: store distributions of sensor states about triggers associated with activities; receive signals indicative of sensor states from said at least one sensor; analyse the received sensor signals to produce distributions of sensor states for the received signals about triggers, and compare the distributions of sensor states for the received signals with the stored distributions of sensor states associated with activities to identify the activity or a change in the activity.

The sensing system may have a plurality of sensors, each sensor having at least two defined states. The plurality of sensors may comprise at least two different types of sensors. The system may be adapted for occupancy sensing.

The at least one sensor may comprise a door sensor that is operable to detect opening or closing of a door. The at least one sensor may comprise a proximity sensor and/or a motion sensor and/or an audio sensor and/or a temperature sensor.

The sensing system may be adapted to compare the distributions of sensor states for the received signals with the stored distributions of sensor states associated with activities in real time.

The sensing system may be adapted for use in a control system for controlling a heating, ventilation or cooling system for a building.

The system may be adapted for use a control system for controlling a lighting system in buildings.

The system may be adapted for use in a security system.

The system may be adapted for use in an assisted living system. The system may be adapted for use in an electric vehicle charging system. The system may be adapted for use in a power demand response system. Brief Description of the Drawings

Various aspects of the invention will now be described by way of example only, and with reference to the accompanying drawings, of which:

Figure 1 shows a block diagram of an occupancy sensing system;

Figure 2 is a flowchart of an occupancy sensing algorithm;

Figure 3 is a flowchart of an occupancy sensing algorithm;

Figure 4 is an example mean distribution for a particular environmental state constructed about an anchor point;

Figure 5 is a table of a state space of a three sensor system;

Figure 6 is a table of key activities and corresponding trigger events;

Figure 7 is a table of key activities and corresponding environmental state transitions; Figure 8 shows plots of mean state distributions for four key activities, (a) arriving, door opening; (b) leaving, door opening; (c) arriving, door closing and (d) leaving, door closing;

Figure 9 shows plots illustrating performance of classification algorithm;

Figure 10 is a table of performance of classification algorithm;

Figure 1 1 shows a decision tree indicating activity chains;

Figure 12 shows plots of results of a test analysis, and

Figure 13 is a table indicating performance of the test analysis. Detailed Description of the Drawings

Figure 1 shows an occupancy detection system that in this case has a plurality of different sensors S1 , S2, S3 and S4 for monitoring an environment, for example a house, and a monitoring and analysis system that includes a processor for analysing the outputs of those sensors. The sensors can be physically or wirelessly connected to the monitoring and analysis system. Each sensor is capable of seeing the environment and activities occurring within that environment from a unique point of view. Therefore, each different sensor provides distinct information of the environment.

The monitoring and analysis system is adapted to monitor activity in an environment. An activity is defined as an action carried out by an occupant in a space. Each activity generates an evolving sequence of signals from different combinations of sensors as the activity progresses. To interpret data from the individual sensors, the signals from the sensors need to be fused to maintain the temporal nature of the sequence. This allows simultaneous events to occur in any analysis.

Activities in the environment correspond to trigger events in collected sensor data. A distinction can be made between simple and complex activities that occur in an environment. Simple activities are those that can be identified using a single trigger. For example, an activity of 'movement in a room' may be classed as a simple activity and can be identified by a trigger defined by a single passive infrared sensor changing its reading from 0 to 1.

The output of each sensor in a defined system of sensors is limited to a set of defined states: s_t = {m₁, m₂, m₃ ... m_v}, where m is a particular sensor state and v, is the total number of states that a particular sensor /^', can exist in. If there are c sensors in the system, a unique combined system state, ψ , can be defined at any given time, t, as a function of the state of all the individual sensors: x j = |s₁, s₂, s₃, s₄ ... s_N) . Therefore, at a given time t, the state belongs to one of a discrete set: Vy,t = {Ψι,ψ2,Ψ3, - ΨΝ,}· N is the total number of possible system states and can be calculated by: N = Πί

An activity in an environment monitored by sensors will cause a system defined as above to transition between different states. Therefore, each activity corresponds to a dynamically evolving sequence of environmental states. Figure 2 is a flowchart 10 of an algorithm that can sense change in occupancy. This shows steps of an algorithm for determining a distribution of a single environment state. As an initial step 12, data corresponding to signals from individual sensors in the system is collected. The changing state of the environment being monitored can be found by sampling signals at a rate chosen to capture changes in the states occurring for a particular activity. An appropriate rate is dependent on a time scale over which identified activities tend to occur. The collected data is combined and/or fused together. The data may be represented as a vector, for example as a fused vector. The fused vector may be saved in a buffer. The combined data is a representation of the status of the environment being monitored. All possible combinations of sensor values correspond to an environmental state. A fused vector may represent all environmental states.

The combined data is continuously searched in order to identify a trigger event corresponding to a key activity 14. This may involve searching the fused vector. A trigger event may indicate the beginning of a key activity or may indicate the progression of a key activity. For example, the fused vector may cross a particular threshold or a distinctive change in vector values may occur thereby indicating that a particular activity is underway. The time at which a trigger event occurs defines an anchor point.

Once an anchor point is identified, a time-window of interest is defined around it 16. The time-window is a function of the time difference from the anchor point. The region within the time-window is referred to as the event space. The data corresponding to events in the event space are selected and further analysed. Then a distribution of occurrences of a particular environmental state in the time window is calculated 18. For a given trigger, a histogram is calculated. The histogram represents the distribution of the frequency of states occurring within the time window. An interval bin-size is chosen. The event space is divided into time-slices corresponding to the interval bin- size. For each interval, the number of times the environmental state occurred during that time is counted i.e. the frequency of occurrence of the environmental state is determined. Using the information from all the time intervals, a histogram can be drawn that represents the distribution of the frequency of the environmental state occurring within the time-window. The histogram is typically normalized to give a relative distribution of occurrence. Note also, that during this analysis a histogram for more than one environmental state can be calculated for the date in the time window.

The above steps are carried out for one anchor point corresponding to a trigger event found in a data set. In order to determine a statistically accurate view of the relationship between the environmental states and the key activity, more than one occurrence of the trigger event must be analysed. Hence, steps 14, 16 and 18 are repeated for other occurrences of the trigger event found in the training data. For every trigger event found an anchor point is defined. The trigger event may be found in another data set. In more detail, for a second trigger event, the same time window function is used to delineate an event space around a second anchor point. In this way, the second event space corresponds directly with the first event space. By running the same histogram analysis on the environmental state data in the second event space, a comparison can be made between the first trigger event and the second trigger event. In particular, a comparison can be made between the first histogram and a second histogram corresponding to the second trigger event.

The above analysis is repeated 20 for a number of training examples that have been recorded. Then the mean, μ, and the variance, φ², of all training examples that have been analysed are calculated for each time interval. A mean distribution of the environmental state is calculated over the time window 22. An evolving distribution of the environmental state is calculated which represents the key activity of interest. Some steps can be carried out for more than one environmental state at a time. Also other ordering is possible. In particular, a distribution of occurrences for all states can be calculated for a given time window before mean distributions of states are calculated. Figure 3 is a flowchart of how the environment for other key activities of interest can be built up. Firstly, a key activity of interest and the corresponding trigger event to be searched for in a combined data set are determined. Then a distribution for all environmental state in a defined time window is calculated 22, as described above. An evolving mean distribution for all environmental states is calculated 23 where the mean is taken over all training examples containing an anchor point corresponding to the trigger event. The analysis is repeated 24 for more than one key activity of interest.

Training data is used to complete the analysis for the full range of desired activities to be classified from data associated with particular triggers. A table can be produced for each activity with a mean relative distribution of the occurrence of each state in time intervals and its associated variance across the training examples. The table gives an indication for each activity, of which states within a given time window, correlate positively with a particular activity being undertaken and which states have a weak or no correlation with different activities. The result of the above process is a mixture of distributions.

In general, a larger number of training examples leads to a better picture of the distribution of environmental states for different activities. Training examples can be collected from data corresponding to different training periods or from the same data set. The more stochastic an activity is the larger the number of possible system states and the greater the number of training examples is required to identify trends. In practice, it was found that, the system operates effectively with relatively few training examples.

Figure 4 shows a plot of an example mean distribution for a particular environmental state constructed about an anchor point. The x-axis represents time and in particular At which is the time from the anchor point. In Figure 4 the anchor point splits the x-axis into two domains corresponding to pre-anchor and post-anchor points. The y-axis denotes the frequency of occurrence of the environmental state for different times. The dark dashed line delineates an area in the plot and corresponds to a time window function. In this example, the time window function is a rectangular function defined by:

T is chosen for the window to be an appropriate size. Outside the window, the window function is 0 and no analysis is completed. Although a simple rectangular window function is chosen for this example, other window functions may be used. Figure 4 shows the window divided into time slices of size τ.

Figure 4 illustrates a histogram calculated about the anchor point. The interval bin size of the histogram is the time slice size τ. Each bin has a mean state count and a variance indicated. The mean state count of each bin is represented by the grey shaded rectangle and the variance of each bin is represented by the hatched rectangle. The indicated mean state count is the mean frequency of this state occurring in a given time slice relative to an anchor point where the average is taken over the anchor points analysed. The variance is the variance of the frequency of this state occurring in a given time slice relative to an anchor point over the anchor points analysed. The variance is a measure of how far the frequency of occurrence of the state spreads out over all the training examples analysed. In other words, a small variance indicates that data points of a sample tend to be close to the mean of the sample, and a large variance indicates that the date tend to be far from the mean.

Figure 4 has three intervals of interest marked. The interval marked 'A' has a small mean indicating that the frequency of occurrence of the state in this time interval is low. The interval 'A' also has a high degree of variance meaning that the frequency of occurrence of the state is spread out over the sample examples. The interval marked 'B' has a high mean state count meaning that the frequency of occurrence of this state in this time interval is high. The interval marked 'B' also has a low degree of variance indicating that the frequency of occurrence of this state tends to be close to the mean and therefore 'B' is a much stronger feature than TV. The interval marked 'C has no mean or variance plotted which indicates that there were no occurrences of this state in that time interval in any of the training examples sampled. Once a picture of typical distributions of states about selected triggers for key activities is established, these distributions can be used to classify unknown activities. To classify an unknown activity for a particular trigger, sensor data is collected for the unknown activity and compared to the collection of mean training distributions. A match score indicating the match between the unknown activity and the mean training distribution is calculated. Different methods of calculating the match score can be implemented.

One method to calculate a match score assumes that each state has a normal or Gaussian distribution of state counts in a time interval where the Gaussian distribution is defined by the mean and variance of the state in that time interval calculated as previously described. The score χ, for a given state ψ₇·, where j is the state number, in interval τ, is given by:

X^j.r = (K I β.0²)f(«, 4>² W(At { a £ M|a≥ 0 } (1) where κ is the measured number of counts of a given state in the corresponding time interval for the unknown activity, μ is the mean of the training distribution for the environmental state ψ₇· and 0²is the variance of the training distribution for the time interval, g is a Gaussian type function defined

Gaussian type function compares the measured value κ of the unknown activity with the training dataset distribution. For the unknown activity, if a state occurs P times within a time interval, and the total count of all the other states that occur within this time interval equals L, then κ = . f \s a function that gives a relative importance to the score as a function of the variance of the training data set. In this example, it is defined as f(a, 0²) = exp(-a0²). a is a constant free parameter. This function gives a variance of 0 a weight of 1 , and as the variance increases the score becomes proportionally less. Therefore, f is chosen to allow a greater weight to be given to states occurring in a particular time interval that have a stronger correlation with a particular activity i.e. low variance in the training set data. This is to be compared with states that appear to occur sporadically with weak correlation i.e. with a large variance in the training data. The value of the constant a, gives the magnitude of the weighting by f. Other weight functions can be chosen. The exponential is chosen for its simplicity and effectiveness. The window function W(At), defines the time window of the analysis, as described in relation to Figure 4. The total score, χ, is calculated by summing the score , x^_jiT, over all states and time intervals and dividing by a consistency factor:

X = _, ^¾1 "**/ > _{β {}β ε π\α≥ 0 } (2)

1+∑(χιΡρτ-Χψ] ) /M Where _τ is the score for state x _j in time window τ and is the mean score over all states and time intervals within the window W. ∑ is a

calculation of the overall variance of the distribution of scores. M is calculated by M = Nn where n is the total number of time intervals given by n = Τ/τ such that {n ε M} and N is the total number of possible states. The denominator is such that the overall score is lower when the total variance in the distribution of the individual scores is higher. The value of β gives the degree of this effect and is a free parameter. A low variance is favoured, which will occur when the score is consistent across the complete event space. Consequently a lower total score will result when there are a small number of matches perhaps corresponding to chance matches. On the other hand if there are no matches, or a very weak correlation, this might also result in a low variance in the score (which is close to 0 for most states), in this case the denominator would tend to 1 but the overall score would remain low. The probability of all the states in each time window matching consistently by chance is low if the features have been well chosen.

In summary, a series of training data sequences for the same activity has been used to define a collection of Gaussian distributions for a particular time interval, with each Gaussian distribution describing the likelihood of a given state occurring within a particular time interval. The variance, φ² , calculated from training data, takes into account the expected spread in the number of counts of a particular state for a specific activity in a given time window τ.

To test the effectiveness of the algorithm for occupancy sensing, a field trial was performed using a single household. Data was collected from an array of sensors in the household over a three week period. The collected data was split into three data sets: a training data set, a validating data set and a testing data set. The system included a passive infrared (PIR) detector, an ultra-sonic proximity sensor, a reed-switch, a light level sensor and a microphone for detecting sound level. The sensors were combined into a single box and mounted vertically above the principal door of the property, in this case, the principle door was the external kitchen door.

The PIR sensor is capable of detecting movement in a vicinity of the box. The proximity sensor is capable of detecting how close somebody is to the door. The reed switch was fitted to the door frame and signalled when the door was open or closed. The light level sensor is capable of detecting a significant change in light in the environment indicating when a light switch was turned on or off. The sound level sensor gives an output of a mean amplitude of sound at a particular moment. In the following analysis, only the data from the reed-switch door sensor, the ultrasonic sensor and the PIR sensor data is analysed.

A data collection device was manufactured on a printed circuit board using an ATmega328 with a 16 MHz Crystal. The data collection device was programmed to collect samples from the array of sensors at a rate of 4 Hz. A data protocol was programmed that sent data to a Raspberry Pi on request using an XRF wireless RF radio UART serial data module manufactured by Cisco. Data was processed on the Raspberry Pi and stored with timestamps using a Python Pandas DataFrame on a USB stick. Each of the reed-switch door sensor, the ultrasonic sensor and the PIR sensor data can exist in discrete environmental states. The PIR sensor is binary in nature and can only be in one of two states. These states can be represented by the values 0 or 1 , where 1 indicates movement and 0 indicates no movement. The door sensor is also binary and so can also only be in one of two states. The value 1 here may indicate the door is open, and 0 indicates the door is closed. The ultra-sonic proximity sensor gives an analogue signal. The output signal from the ultra-sonic proximity sensor is the distance between the sensed movement and the sensor. In the analysis, this output was discretized by representing distinct ranges of values to correspond to different states of the sensor. For example, the data corresponding to the ultra-sonic proximity sensor was assigned a value of 1 when someone was within close proximity of the door and a value of 0 when someone was not within close proximity of the door. This value can be allocated based on a comparison to a pre-determined threshold value. Therefore, the proximity sensor can be in one of two states. Since all three sensors have two possible states each, the combined three-sensor system has a state space of size eight. The following analysis is based on identifying and detecting patterns within this state space corresponding to movement in the monitored environment.

Figure 5 shows a table summarizing how a unique value was attributed to each combined environmental state. Each individual sensor state of the three sensors, door, PIR and proximity sensor was allocated a unique natural prime number. In detail, the state of the PIR sensor corresponding to no movement detected was allocated a value of 1 , and the state of the PIR sensor correspond to movement detected was allocated a value of 2. The state of the door closed was allocated a value of 3 and the state of the door open was allocated a value of 1 1 . The state of the proximity sensor corresponding to nobody near door was allocated a value of 5 and the state of the proximity sensor corresponding to somebody near door was allocated a value of 7. By multiplying the state values together each environmental state of the state space is allocated a unique value. Figure 6 is a table showing key activities in an environment chosen for detection in the study. Four of these key activities relating to a door are: opening a door when arriving, closing a door after arriving, opening a door to leave and closing a door to leave. Other activities listed in Figure 6 are movement within the home and no movement for an extended time. Each activity in Figure 6 involves one, two or three of the sensors of the three sensor system. Each activity in Figure 6 has a corresponding trigger event. For the activities of opening a door when arriving or opening the door to leave the trigger event is the door opening. For the activities of closing a door after arriving and closing a door to leave the trigger event is the door closing. For the activity of movement within the home, the trigger event is that movement is detected for 60% of 3 seconds by the passive infrared sensor. For the activity of no movement for an extended time the corresponding trigger event is no movement is detected for greater than 60 seconds by the passive infrared sensor. Figure 6 also indicates abbreviations given to each activity.

Figure 7 is a table of environmental state transitions corresponding to trigger events of particular activities in an environment. The transition in environmental state from system state ψι to system state x j is represented in Figure 7 by {i u i j}. As can be seen in Figure 7, each trigger event has one or more environmental state transitions associated with it. As an example, a trigger event corresponding to the activity of opening door when arriving is the transition between system state "door closed, no movement detected, nobody near door" to "door open, no movement detected, nobody near door". The number of activities identified within a property could be extended according to different sensor data available.

Ground truth data was also recorded independently from the three sensor system array during the experimental trial. A large green button was positioned in the household adjacent to the main entranceway. The position was chosen to make it easy for the occupant to switch the button on or off when they entered and left the house. The occupant switched on the button to record when they were occupying the house. The button glowed when switched on. The button was also designed to be large enough that it was hard to miss or for the occupant to forget it was there. Data was collected and sent to the Raspberry Pi in a similar manner to other sensor data. The ground truth data was also stored with timestamps in a Python Pandas Data Frame. The ground truth data allows for an objective data of occupancy to be recorded. In the study, presence (PS) and no presence (NPS) are completely defined by their respective triggers. In order to differentiate between events "arriving, opening door", "leaving, opening door", "arriving, closing door", "leaving, closing door" the algorithm described above was trained as a binary classification algorithm. Data recorded in an initial two week period was searched for the door related trigger events shown in Figure 5. Using the ground truth data collected during the initial period data was also isolated and classified by hand into the 4 defined categories. In the two week time period 23 samples of each of the 4 activities were identified. The sample data was divided into two parts. For each activity, 10 samples were used as a training data set giving a total of 40 samples altogether. The remaining 13 samples for each activity, giving a total of 52 samples, were allocated to an independent validation data set. As described with reference to Figures 1 , 2 and 3, the training set of data was used to calculate evolving mean distributions about key triggers for environmental states. The validation data set was then used to evaluate and validate the algorithm against the ground truth date. In the analysis the following parameters were used: window period T=25, time interval τ=1 , weight parameters a=5.8 and β=λ . The results of the analysis are shown in Figure 8.

Figure 8 shows plots of mean distributions generated by the algorithm. The aim is to identify the four key activities: arriving, door opening; arriving, door closing; leaving, door opening and leaving, closing. The mean distribution for each key activity is shown in Figures 8(a), (b), (c) and (d). Each plot has a y-axis indicating the eight unique environmental states of the three-sensor system. The x-axis of each plot indicates the time before and after the trigger event measured in seconds. Only the time-window selected is shown, which is 50 seconds in total. The time window of each environment state is divided into 50 time intervals, each of a second in duration, represented by a rectangle. For each state, a rectangle is shaded to indicate the normalized frequency of occurrence of the state in that particular time interval. A darker colour indicates a higher normalized frequency. A lighter colour indicates a lower normalized frequency. An absence of colour indicates a zero normalised frequency of occurrence.

The resulting mean distributions follow logic. Figure 8(a) shows the mean distribution for system states for the activity arriving, door opening. The system state 15 corresponding to door closed, no movement detected, no proximity to door has high occurrence frequency before the trigger for arriving, door opening. System state 15 is dominant before the trigger. Figure 8(d) shows the mean distribution for system states for the activity leaving, door closing. System state 15 has high occurrence frequency after the trigger event and is dominant in Figure 8(d). It may be expected that system state 42 corresponding to door closed, movement detected, somebody near door have occurrence for the activity leaving, door opening and the activity arriving, door closing. It can be seen from Figure 8(b) and Figure 8(c) that this is indeed the case.

These are two distinct and important methods of reporting accuracy of the classification algorithm especially when there is an unbalance in the number of samples in a particular class. If only the proportion of correct classifications is calculated, it might lead to a false appearance of a good performance in the algorithm. This is important when reporting results in evaluating the algorithm in its performance over a whole day. As an example, if it is assumed that there are only 2 leaving events in a day that both last 10 minutes and there is a prediction stating that the home was occupied for the whole day then the prediction would have a high overall accuracy (23 hour 40 min / 20 minutes) where in fact the home was not occupied for 20 minutes. Therefore, performance evaluation must be performed separately for occupied and unoccupied states. It is useful for performance evaluation to take into account the following four parameters: (1 ) when the home is correctly predicted as occupied (true positives); (2) when the home is correctly predicted to be unoccupied (true negative); (3) when the home is incorrectly predicted to be unoccupied when it is in fact occupied (false negatives); (4) when the home is incorrectly predicted to be occupied when it is empty (false positive).

Two parameters, Recall and Precision are defined. These parameters are helpful in interpreting performance of the algorithm. From the cases in a validation set, Recall ls a measure of the fraction of positive cases identified by the classifier out of the total of all of the positive cases in the validation set. Recall is the total number of true positives divided by the sum of the number of true positives and false negatives. From all the examples the classifier has labelled, Precision is a measure of what portion has been labeled correctly. Precision is calculated as the total number of true positives divided by the sum of false positives and true positives.

As described elsewhere there is a number of free parameters that can be varied in the algorithm. The algorithm was run more than once, while varying these parameters to determine the dependence of performance of the algorithm on the free parameters. Figure 9 shows three plots that illustrate how the Recall of the algorithm varies as a function of the following set of parameters: (a) window size (7), (b) the weight parameter(a) and (c) the time interval (τ). Algorithm performance is clearly dependent on the set of parameters. In particular, there is a significant effect for the activity arriving, door closing as indicated by the square points on the plots.

Figure 9(b) shows the dependence of classification accuracy (Recall) on the free parameter weight for the four key activities. For arriving, door closing activity a fall-off in Recall as weight is increased is observed. The fall-off can be explained by observing that a higher degree of variation in the arriving, door closing activity is present compared to other activities. The activity arriving, door opening has a high classification accuracy because this activity corresponds to a distinct change in sensor state from nothing occurring to an activity occurring. The activities leaving, door opening and closing, door opening also tend to follow a similar pattern to the activity arriving, door opening. The observed variation in the classification accuracy for the activity arriving, door closing is due to the nature and variation in the way an occupant arrives at a house. A higher weight constant a is linked to less consideration given to system states occurring with a higher variability in total count in any given time interval i.e. system states with a larger standard deviation. This is achieved through the choice of f(a,0²) in Equation (1 ). In light of these variations, it is important to select sensible values for different parameters. Validation data can be utilized to select values for the algorithm parameters. It is possible to optimize the algorithm depending on the particular situation by varying the different parameters and monitoring the performance metrics. Figure 10 is a table showing measured classification performance. In particular, the recall and precision is calculated for the trained algorithms performance on a validation data set. Two classification problems were analyzed: (i) distinguishing between arriving, door opening and leaving, door opening and (ii) distinguishing between arriving, door closing and leaving, door closing. For simplicity, all parameters were set to same values for both classification problems: these are window period T=25, time interval τ=1 , weight parameters a=5.8 and β=λ . As can be seen from Figure 9, the parameters chosen perform very strongly in identifying arriving, door opening and leaving, door opening events.

Figure 1 1 shows a collection of decision trees derived from the training data. The decision trees are a map of a complete set of paths between observed activities in the set of training data. The activities are labelled by abbreviations according to the abbreviations found in Figure 7. The decision trees break down classification of occupancy into a series of logical choices based on activities that have taken place. Each rectangle of Figure 1 1 represents a node of the decision tree. The shaded nodes indicate a root of each tree. Each root represents an activity which may be monitored in a home. Each root is connected to a number of possible leave nodes. The leaves either correspond to another root node or an occupancy prediction.

An activity chain has a first step on a root node. The second step is chosen from one of the possible connected nodes. For example, a first step of AO has a choice of NPE, AC, PE and PE as a second step. NPE, AC or PE are themselves root nodes and a choice of NPE, AC or PE results in moving to a node that is a root node. In this way an activity chain in the decision tree is continuous. If the second step in the activity chain is chosen as PS, then this correspond to an occupancy prediction. In this case PS indicates occupancy of a value 1 .

Zero occupancy can only occur when a particular serious of event takes place. For example, if an initial state of the system is assumed to be occupancy equal to 1 , with the door closed, then a detection of an activity involving a door opening is required. In this case, the activity chain must be as follows: an activity involving door opening, door opening followed by activity of leaving, door closing (LC) followed by activity no presence detected. It is assumed that if a door is left open, followed by activity not present detected then the person is still close the home. It is assumed that the door is shut when a person leaves the home completely.

Occupancy of an environment at a given moment in time is a function of activities that have taken place. The household used in the experimental study is single occupancy and therefore no attempt was made to count number of occupant at a given time. The method described can be extended to a method for determining occupancy in a multi- occupancy home through appropriate training data for simultaneous multiple entry.

The classification algorithm combined with the decision tree was evaluated based on a testing data set collected over a five day period. This data set was not used in any other stage of the analysis. Results were compared to ground truth occupancy data collected during the same five day period. Figure 12 shows plots of results of one out of the five days of the five day test period. The x-axis of all plots indicates time elapsed over a day from 5.30 am to 8.30 pm. This range corresponds to the time first left the house and the last time the left they entered the house, according to recorded ground truth data. The range is chosen to give a realistic view of performance during the day while occupancy is changing. During the night, occupancy did not change and was correctly predicted by multi-sensor system for each of the five days of evaluation.

Figure 12(a) is a plot of occupancy throughout the day. The occupancy either has value 1 indicating occupied or value 0 indicating unoccupied. The measured occupancy is occupancy measured by the algorithm and may also be referred to as predicted occupancy. Figure 12(b) is a plot of ground truth occupancy throughout the day derived from recorded ground truth data. Likewise, ground truth occupancy is either value 1 indicating occupied or value 0 indicating unoccupied. Figure 12(c) is a plot of system state throughout the day.

Figures 12(d) and 12(e) are plots of alternative predictions of occupancy using only data measured by the passive infrared sensor. Figure 12(d) is a plot of a simple function using passive infrared sensor data with a 10 minute timeout. In other words, an unoccupied state corresponds to 10 minutes without the passive infrared detector detecting any movement in the room. Figure 12(e) is a plot of a simple function using passive infrared sensor data with a 20 minute timeout. In other words, an unoccupied states corresponds to 20 minutes without the passive infrared detector detecting any movement in the room. Both Figures 12(d) and 12(e) have either value 1 indicating occupied or value 0 indicating unoccupied.

Figure 13 shows a table indicating performance of the algorithm over the five day testing period. The system and algorithm is able to predict when the house is occupied with a near 99% Recall and Precision, and able to predict when the house is unoccupied with over 90% Recall and Precision. Unoccupied performance is lower than occupied performance because there were a few instances in the five days when an occupant recorded leaving the house but did not shut the door. In addition there were also a few incorrect classifications.

Overall, as can be seen in Figure 13, the predictions by the multi-sensor system are considerably more accurate than the predictions by the passive infrared sensors alone. This is due to occurrence of significant periods when the house is occupied but the passive infrared sensor does not detect this corresponding to a false negative. The passive infrared sensor has only 53% Recall for 10 minutes timeout. The passive infrared sensor has a higher Recall for a longer timeout as a greater portion of the occupied period is then predicted correctly. The passive infrared sensor predicts that the house is occupied with a relatively high precision of almost 82% for both timeouts. Therefore, the sensor is poor at predicting when the house is not occupied because the occupant spends a considerable amount of time at home. When the occupant leaves the house, the passive infrared sensor will wrongly mark the room as occupied for a considerable period after a leaving event due to the timeout periods. Due to relative proportions this leads to a bad overall performance of the passive infrared sensor correctly predicting an empty house.

The system has been shown to perform well in a single occupancy household. The occupancy sensing approach applied here has achieved the goal of considerably reducing the number of false negatives compared to passive infrared sensors. This makes it suitable for controlling residential heating control systems correcting the shortcomings of programmable thermostats as previously discussed.

The increased accuracy makes the system applicable for use in an assisted living environment. Its output would be likely to engender a significantly higher level of trust, particularly as a consequence of the reduced number of false negatives. In this case, the methodology could be used to interpret sensor data to identify changes in activity that might signal something to be concerned about. This can be done in a discrete way, without the occupant having to wear a pendant, allowing the system to contribute to increased feeling of independence and general well-being.

The system of the invention can be applied to all emerging applications where occupancy sensing is used, for example smart home or home automation systems. The increased accuracy of the occupancy sensing of the invention will improve control outcomes accordingly. Systems that can be controlled include the Heating, Ventilation and Cooling (HVAC) systems for both residential and non-residential buildings, lighting controls in buildings and security systems.

The invention could also be used in demand response technology that switch loads on or off based on electricity network signals, e.g. tariffs. Demand response refers to schemes in the energy market whereby end-users are encouraged to reduce or shift their demand at peak times. Occupancy is closely tied with household energy consumption, and so understanding occupancy patterns across a large number of dwellings can potentially be used to improve demand forecasting through identifying the periods in different households when demand and by association load shift opportunity is high .

Knowledge about the occupancy patterns of a block of buildings can be used to create a more firm description of the aggregated demand response potential.. Information on identified relevant load shifting opportunities can be contextualised using discrete occupancy data before being communicated to end-users. Combining typical occupancy patterns of individual households with their consumption data over a defined community, allows identification of households that have the greatest potential to participate in, and make a contribution to overall community demand response. These households can then be targeted.

By installing additional technology, it is possible to remotely actuate loads and achieve load shifting without the need for the occupant to take action. This actuated demand response can be enhanced by knowledge of occupancy patterns; which provides constraints on periods when user comfort should be maintained, and allows the selection of actuated loads that bring occupant benefit, rather than leading to wasting energy and increasing end-users bills (e.g. when the occupant is not at home and not predicted to be at home). Real time occupancy sensing in the context of demand response is of benefit for both user-initiated and actuated load shifting. Control technology is starting to emerge that optimises the use of thermal storage devices to maximise the utilisation of renewable generation. These thermal storage devices may provide space heating or hot water for buildings. Accurate building occupancy sensing of the type described here can provide an input to this control technology that will increase its performance.

The sensing system of the present invention has potential to be used in any environment where sensors can be used to detect changes in the surrounding environment. As one example, the invention could be used for counting and identifying wildlife. For example, the analysis could be triggered by a certain movement or a noise and the animal identified from the activity around the trigger, such as its call. One of the strengths of the invention is that it does not rely on the song being identical, but rather uses the distributions of key features.

In a sequence of time series data, detected by a sensor(s), the approach could be used to pick out abnormalities, which could trigger a notification. For example, this might be used in trains/cars to provide forewarning of a potential hazardous situation, or with water pumps, biogas systems, solar panels in the developing world to identify when they are not performing as well as they should. The system of the invention could be used to identify selected events from data coming from in built mobile phone sensors. For instance, it could be used to map how you use time during a day (when you leave the home, drive, cycle etc).

A skilled person will appreciate that variations of the disclosed arrangements are possible without departing from the invention. Accordingly, the above description of a specific embodiment is made by way of example only and not for the purposes of limitations. It will be clear to the skilled person that minor modifications may be made without significant changes to the operation described.

Claims

1 . A sensing system for use with at least one sensor having at least two defined states, the sensing system being adapted to: store distributions of sensor states about triggers associated with activities; receive signals indicative of sensor states from said at least one sensor; analyse the received sensor signals to produce distributions of sensor states for the received signals about triggers, and compare the distributions of sensor states for the received signals with the stored distributions of sensor states associated with activities to identify the activity or a change in the activity.

2. A sensing system as claimed in claim 1 comprising a plurality of sensors, each sensor having at least two defined states.

3. A sensing system as claimed in claim 1 or claim 2 wherein the plurality of sensors comprises at least two different types of sensors.

4. A sensing system as claimed in any of the preceding claims adapted for occupancy sensing.

5. A sensing system as claimed in claim 4 wherein said at least one sensor comprises a door sensor that is operable to detect opening or closing of a door.

6. A sensing system as claimed in any of the preceding claims wherein the at least one sensor comprises a proximity sensor and/or a motion sensor and/or an audio sensor and/or a temperature sensor.

7. A sensing system as claimed in any of the preceding claims adapted to compare the distributions of sensor states for the received signals with the stored distributions of sensor states associated with activities in real time.

8. A sensing system as claimed in any of the preceding claims adapted for use in a control system for controlling a heating, ventilation or cooling system for a building.

9. A sensing system as claimed in any of the preceding claims adapted for use a control system for controlling a lighting system in buildings.

10. A sensing system as claimed in any of the preceding claims adapted for use in a security system.

1 1 . A sensing system as claimed in any of the preceding claims adapted for use in an assisted living system.

12. A sensing system as claimed in any of the preceding claims adapted for use in an electric vehicle charging system.

13. A sensing system as claimed in any of the preceding claims adapted for use in a power demand response system.