Technical field

The present description relates to techniques directed to automatically detect potential anomalies in the behavior of electrical appliances , by means of the analysis of their power consumption .

Various embodiments may apply e . g . to the detection of electrical anomalies and/or to the detection of behavior anomalies of electric appliances belonging to a monitored environment .

Description of the prior art

Anomaly detection can be defined as the task o f discerning the data which di f fers somehow from what is commonly defined as regular ( or standard, common) . This originates from the need to build a method able to automatically recogni ze anomalies in the behavior, consumption or usage of a device without any technical and manual intervention on the device itsel f . Every event which appears di f ferent from normal standards may represent an interesting event to monitor, irrespective that it is due to the appliance degradation over time or to a wrong use by the users . The capacity to detect by itsel f anomalies in the appliances ' usages may involve a signi ficant advantage in several applications .

The most interesting aspect of anomaly detection using such an approach of discerning the data which di f fers from standard or regular data is that anomaly detection may operate in a completely unsupervised fashion, namely without the necessity of an expert suggesting how a possible anomaly could mani fest itsel f .

Document US2013/204552A1 refers to an apparatus for detecting electrical anomalies in home appliances . These anomaly detection systems can work only i f electrical signals are received directly from the monitored appliances , therefore i f further information is needed further sensors are needed as well .

Obj ect and summary

An obj ect of one or more embodiments is to provide a method for anomaly detection in electric appliances that solves the drawbacks of the prior art which allows to obtain detailed information on the power consumption of a given appliance , without requiring appliance speci fic sensors .

According to one or more embodiments , that obj ect is achieved thanks to a method having the characteristics speci fied in Claim 1 . One or more embodiments may refer to a corresponding system performing the method and to a computer program product that can be loaded into the memory of at least one computer and comprises parts of software code that are able to execute the steps of the method when the product is run on at least one computer . As used herein, reference to such a computer program product is understood as being equivalent to reference to a computer-readable means containing instructions for controlling the processing system in order to coordinate implementation of the method according to the embodiments . Reference to " at least one computer" i s evidently intended to highlight the possibility of the present embodiments being implemented in modular and/or distributed form .

The claims form an integral part of the technical teaching provided herein in relation to the various embodiments .

According to the solution described herein, it is described a method for the detection of anomalies in a set of electric appliances comprised in a given monitored environment , said monitored environment comprising a main meter measuring electric power consumption by the set of electric appliances belonging to said given monitored environment , said method comprising performing a non-intrusive load monitoring comprising coupling one or more monitoring device to said main meter of the monitored environment measuring an aggregate load information representing the power consumption of the whole set of electric appliances belonging to said given monitored environment during a first given period, performing a non-event-based detection, detecting in said aggregate load information measured during a first given period time intervals of operation of selected target appliances operating in said monitored environment and outputting one or more activity windows indicating the time of start and the time of stop of said time intervals of operation of the selected target appliance , in particular by applying appliance models to said aggregate load information, then performing an event-based detection, extracting from time segments of aggregated load corresponding to said activity windows the disaggregated load corresponding to said selected target appliances , performing an anomaly detection procedure on the selected target appliance on the basis of said disaggregated load .

In variant embodiments , the anomaly detection procedure receives load information also from intrusive load monitoring comprising smart appliances and smart plugs acquiring the power consumption directly from a corresponding monitored appliance .

In variant embodiments , said non-event-based detection comprising comparing an appliance model representing the power consumption of the selected target appliance during a respective operation to said aggregate load stored in a library of appliance models , in particular said appliance model being neural network model , and said aggregated load being applied as input to said neural network model , said comparing operation outputting an activity window indicating the time of start and the time of stop of said time intervals of operation of the selected target appliance .

In variant embodiments , said event-based detection includes identi fying power events defined by a step change of the load of a given amount , obtaining an amplitude of the change and a time of the change , clustering said events according to the amplitude and the sign, comparing the amplitude to stored values of amplitude of model appliances assigning the cluster of a given amplitude to the model appliance with corresponding amplitude values , obtaining the disaggregated load on the basis of the assigned cluster .

In variant embodiments , the event-based detection comprising interacting with an end-user to obtain information regarding the operation of the set of appliances .

In variant embodiments , said anomaly detection procedure on the selected target appliance on the basis of said disaggregated load compares said disaggregated load to a re ference or standard load for said selected target appliance in order to detect di f ferences in amplitude and/or time and/or frequency which are identi fied as anomalies according to a set of criteria .

The solution here described refers also to a system for the detection of anomalies in a set of electric appliances comprised in a given monitored environment , comprising a main meter measuring electric power consumption by the set of electric appliances belonging to said given monitored environment , wherein said system is configured to operate according to the method of any of the previous embodiments .

The solution described here refers also to a computer program product that can be loaded into the memory of at least one computer and comprises parts of software code that are able to execute the steps of the method of any of the previous embodiments when the product is run on at least one computer .

Brief description of the drawings

The invention will now be described purely by way of a non-limiting example with reference to the annexed drawings , in which :

- Figure 1 represents schematically an architecture of a system implementing the method here described;

- Figures 2A, 2B, 2C represents time diagrams of electric loads which can be analyzed by the method here described;

Figure 3 represents a flow diagram showing a procedure of a module of the system of Figure 1 ;

- Figures 4A, 4B, 4C represent a flow diagram of the steps of an operation of the procedure of figure 3 and related diagrams ;

- Figure 5 represent a flow diagram of the steps of a further operation of the procedure of figure 3 ;

- Figures 6A to 6F represent time diagram showing signals used as input and output in the steps of Figure 5 ;

Figures 7A, 7B, 7C, 7D, 7E represents time diagrams related to sub-steps of a step of figure 5 ;

Figure 8 represents a time diagram with a magni fied portion representing an output of the procedure of figure 3 . Detailed description of embodiments

The ensuing description illustrates various speci fic details aimed at an in-depth understanding of the embodiments . The embodiments may be implemented without one or more of the speci fic details , or with other methods , components , materials , etc . In other cases , known structures , materials , or operations are not illustrated or described in detail so that various aspects of the embodiments will not be obscured .

Reference to " an embodiment" or "one embodiment" in the framework of the present description is meant to indicate that a particular configuration, structure , or characteristic described in relation to the embodiment is comprised in at least one embodiment . Likewise , phrases such as " in an embodiment" or " in one embodiment" , that may be present in various points of the present description, do not necessarily refer to the one and the same embodiment . Furthermore , particular conformations , structures , or characteristics can be combined appropriately in one or more embodiments .

The references used herein are intended merely for convenience and hence do not define the sphere of protection or the scope of the embodiments .

The solution here described refers to a method for detecting anomalies in electric appliances , which deals with electrical anomalies and behavioral anomalies .

Electrical anomalies involve power consumption values that are signi ficantly di f ferent from the expected consumption of the monitored appliance , caused either by a mal function of the appliance or by a misuse of the end-user .

Behavioral anomalies involve signi ficant variations in the usage patterns of household appliances/devices caused by a behavioural change of the end-user . The solution here described detects anomalies from such categories of anomalies by analyzing the disaggregated load of household appliances , that can be collected from the following sources : directly from appliances : the appliances themselves measure the time and frequency variations of ( or part of ) the previous list of parameters , by mean of embedded electrical circuits or add-on local measurement system, and are able to deliver the gathered information to a data collector ; obtaining a synthetic load profile : appliances ' disaggregated load is reconstructed through the analysis of electrical aggregated building signal , measured by a connected main meter to an electrical panel of a building or an loT third party add-on measurement system .

For load is intended the energy demand experienced on a system, e . g . a monitored environment compri sing a set of electric appliances using such energy . The load can correspond thus to an electric power consumption, an aggregated load corresponding to the power consumption of the set of appliances and a disaggregated load corresponding to the power consumption of a speci fic appliance , indicated also as target appliance in the following .

Obtaining a synthetic load profile requires a data analyzer for synthesi zing the required appliance load curve profiles fed to the anomaly detection module . This is performed by a Non- Intrusive Load Monitoring (NILM) process combining event-based techniques to decompose and extract the disaggregated load curve profiles of individual devices from the aggregated load signal and event-less machine learning based approaches for annotating the synthetic load curve profiles .

The anomaly detection system automatically generates an output signal delivered to multiple applications to systems which may comprise one or more of : a noti fication mechanism system : to automatically signal to a mobi le APP or site-web or short message system the user the imminent issue in real-time a data collector : a central gathering information collecting anomaly signals for further analys is or services an automatic controlling system : anomaly signals can automatically trigger the changing of operative status of appliances , commanding it through the postprocessed analysis of anomalies retrieved .

In figure 1 it is shown a block schematics showing the architecture of an anomaly detection system according to the solution here described, indicated as a whole with the numeric reference 100 . The anomaly detection system 100 receives inputs from various electrical sources . With reference to Figure 1 , the anomaly detection system 100 is configured to analyze power consumption data coming from a set of electrical sources , which in the example includes three di f ferent types of power consumption monitoring sources , which are a set of smart appliances 110 , a set of smart plugs 120 and a non-intrus ive load monitoring (NILM) procedure 130 . The set of smart appliances 110 sends as input to an anomaly detection module 170 power consumption data SALi , where index i indicates the i-th smart appliance in the set 110 comprised in the monitored environment 105 , i . e . the set of appliances or power consuming devices which is being monitored, the set of smart plugs 120 sends as input to the anomaly detection module 170 power consumption data SPLj , where the index j indicates the j -th smart plug present in the monitored environment 105 , while the non-intrusive load monitoring (NILM) procedure 130 receives as input an aggregate load s ignal AS from a smart power meter 140 , which measures the overall power consumption, i . e . the aggregate load AS , of the monitored environment 105, i . e . of the set of appliances including also the appliances which neither belong to the set of smart appliances 110 nor to the set of smart plugs 120 , e . g . power consuming appliances without metering, i . e . without any type of power consumption metering capability either own or by an associated electric power meter . It is here speci fied that the main power meter may also operate in a distributed manner centrally collecting the measurements of electric power meter submodules associated to subnets or branches of the electric power network of the environment 105, and the smart meter 140 may be also distributed or applied to the central collection unit . The NILM procedure 130 then outputs disaggregated loads DALk, where index k indicates preferably the k-th appliance without metering in a set of appliances without metering, i . e . neither appliances of the set of smart appliances 110 nor of the set of smart plugs 120 in the monitored environment 105 , although in variant embodiments the index k may refer also to a set which includes also appliances of the set of smart appliances 110 nor of the set of smart plugs 120 in the monitored environment 105 . In figure 1 are shown a power LAL from the set 110 measured by the smart power meter 140 , a power LML from the set 120 measured by the smart power meter 140 and a power LNM from the set of appliances without metering measured by the smart power meter 140 , which is not however shown in figure 1 . In variant embodiments index k may in general indicate any k-th appliance in the set of appliances of the monitored environment 105 , in case the NILM procedure also provides disaggregated loads corresponding to appliances of sets 110 , 120 . As they are already provided directly to the anomaly detection module 170, this is usually not necessary .

The aggregated load signal AS may be in general defined as a sequence of instantaneous measurements acquired at the main power meter of the environment 105 by means of a smart meter device 140, where the measurements are defined as a list of time and frequency variations of active power, reactive power, current, voltage, current harmonics, harmonic distortion, voltage harmonics, power factor of incoming electrical signals supplying appliance loads in the monitored environment. Of course the main power meter can integrate the functions of the smart device 140, or the smart device 140 can be applied to a main power meter of the environment 105 which is not configured performing such functions, i.e. performing the sequence of instantaneous measurements just described.

Measuring electric power consumption thus means measuring one or more of different parameters representing the electric power consumption. The exemplary embodiment refers to the active power P, however active power and/or reactive power Q may be taken in account for instance, and/or the voltage and/or current harmonics number H and/or the Total Harmonic Distortion THD. Also exploitation of other power consumption related parameters is possible as indicated above, such as current, voltage, current harmonics, power factor.

In particular, although the exemplary embodiment refers to the active power P, representing the aggregated load signal AS as a single variable signal function of time t, the aggregate load signal AS may also be a n- variable function of time t, e.g. a vector [P(t) , Q(t) , H (t) , THD (t) ] . The anomaly detection module 170 on the basis of data SALi , SPLj , DAL is configured, for speci fic appliances or for each monitored appliance , to detect electrical anomalies EA, which involve power consumption values that are signi ficantly di f ferent from the expected consumption of the monitored appliance , caused either by a mal function of the appliance or by a misusage of the end-user; behavioral anomalies BA, which involve signi ficant variations in the usage patterns o f the appliances/devices in the monitored environment 105 , caused by a change of the behaviour of the end-user .

Smart appliances 110 include all those electric appliances that are capable of monitoring their power consumption and can communicate the corresponding power consumption data SALi to other devices through a communication link, e . g . a communication network, to the anomaly detection module 170 .

Smart plugs 120 comprise monitoring devices that are coupled between the appliance power supply and the wall socket , allowing to acquire the power consumption of the speci fic appliance and communicate the corresponding power consumption data SMLj through the communication link . Smart appliances 110 and smart plugs 120 adopt an intrusive monitoring approach to collect power consumption of the respective electric appliance . As this is not always feasible because of the high costs of the monitoring equipment and the potential large number of appliances to be monitored the system 100 comprises the non-intrusive load monitoring (NILM) 130 procedure to collect individual appliances consumption without installing multiple monitoring devices in the household .

The NILM procedure 130 is configured to extract the power consumption of individual appliances , in particular of those without metering, through the disaggregation of the whole measured power consumption of the monitored environment 105 , e . g . a household, on which the system 100 operates , and comprises one or more , preferably a single one , monitoring device which is coupled to the main meter or main metering system of the household or environment , indicated as smart meters 140 . It is used here the term 'household' to refer to the environment comprising the appliances on which the system for anomaly detection operates , i . e . the monitored environment 105 , which can be of course a di f ferent type of building with respect to a house , e . g . of fice building, a factory, or also simply a set of appliances which is desired to monitor, coupled by a communication link with the anomaly detection module 170 , which include the NILM 130 and possibly also the sets 110 and/or 120 .

Smart meters 140 include monitoring devices coupled or natively incorporated in the main meter of the monitored environment 105 , capable of communicating through the link, in particular a network, the power consumption of the whole monitored environment 105 .

The granularity of the three sources of consumption data, 110 , 120 , 130 may be di f ferent . Smart meters 120 and smart appliances 110 may acquire the power consumption directly from the monitored appliance , thus allowing to capture small electrical anomalies thanks to the high resolution of the acquired information . NILM 130 although may extract the power consumption of di f ferent appliances with an acceptable level of detail , which still allows to capture electrical anomalies in the overall appliance consumption .

The module configured to perform the NILM procedure 130 and the anomaly detection module 170 can be embodied by the same module or comprised in a same processing device, e.g. a microprocessor or a microcontroller, which comprises communication links to the smart meter 140 and to the set of smart appliances 110 and of smart plugs 120. Such communication links may be represented by one or more communications networks, using cables or operating wireless, by Wi-Fi or by Bluetooth for instance. In variant embodiments, modules 130 and 170 can be separate modules, each with its own communication links .

Figures 2A-2C shows three examples of power consumption, i.e. diagrams which shows the power consumption as a function of time t. In figure 2A a plot 210 shows the typical power consumption, in watt W, of a dishwasher acquired by means of a smart plug 120, i.e. power data SMLj , or of a smart appliance 110, i.e. power data SALi, as a function of time t in seconds. In figure 2B a plot 220 instead shows the disaggregated load DALk of a dishwasher obtained by means of non-intrusive load monitoring techniques, i.e. procedure 130. Finally, in figure 2C a plot 230 shows the whole power consumption of the monitored environment 105, i.e. aggregated load AS, as function of the time t. Time ranges in the example are of 10000 seconds respectively, from 70000 to 80000 seconds .

The power consumption data supplied from the blocks 110, 120, 130 to the anomaly detection module 170 are of the type shown in figure 2A-2C, i.e. power consumption values at given time instants, in particular at given sampling or measuring instants.

In figure 3 it is shown a block schematics of the NILM procedure 130, which is configured to extract disaggregated load curve profiles, i.e. power consumption values as function of time, DALk, of the appliances operating in the monitored environment 105 by processing the aggregated load signal AS received from the main meter of the house through the smart device 140 .

The NILM procedure 130 comprises two main modules operating in sequence , a non-event-based module 320 and an event-based module 330 .

The non-event-based module 320 is configured for performing a first detection operation, in the fol lowing also called not event-based detection meaning that is not based on events as opposed to a second detection operation 330 which is instead based on events E . The first detection operation or not event-based detection detects the activation of the various appliances operating in the monitored environment 105 during a given period, e . g . during the day . In detail , the non-eventbased module 320 receives as input the aggregated load AS from the main power meter of the monitored environment 105 through the smart device 140 and generates a set of annotations AT , i . e . a report comprising for each k-th appliance a start time and stop time of every operation, i . e . activity windows AWk of each k-th appliance , i . e . of each target appliance of which the non-event module 320 is trying to output an associated activity window AWk - The target appliance may be one in the set of appliances without metering . The first detection operation is thus an operation of detection of the time period or periods of activation of a given target appliance .

The event-based module 330 is then configured for performing a second detection operation, or event-based detection, extracting the disaggregated load DALk o f the appliances of the monitored environment 105 from the activity windows AWk extracted by the non-event-based module 320 .

The aggregated load AS data may comprise values measured in time of an active power P and also of a reactive power Q, collected with a given sampling period, e.g. 1 second. The aggregate load AS data can be expressed also through current and voltage values rather than power values, or other equivalent electric quantities. In addition, the aggregated load AS can be sampled at lower periods (e.g. corresponding to sampling frequencies in the range of kHz) or lower frequencies (e.g. period of 15 minutes) with respect to the standard frequency of 1 Hz or period of 1 second mentioned above.

The disaggregated load DALk comprises detailed power consumption data, e.g. values measured in time of an active power P and also of a reactive power Q, of a given operation OP of a given k appliance, detected by the nonevent-based module 320. Therefore it can take the form of a P(t, OP, k) where the time has the length of the activity window AW -

In figure 3 it is also indicated a block representative of an end-user 350, i.e. means which allows a user to interact or interface with event-based module 330, which can interact with the event-based module 330, for instance in pull-tag mode and push-tag mode .

In the pull-tag mode the NILM procedure 130 is configured to ask the end-user 350 to confirm that the operation of the k-th appliance identified by the nonevent-based module 320 is correct, in order to save house-specific parameters HSP for the detected appliance, in particular in a house-specific parameters database HSP. The house-specific parameters HSP may include the power levels of the various events composing the disaggregated load DALk of the k-th target appliance, where power may mean active power or a more extended set of power consumption representing variables, e.g. active power P, reactive power Q, voltage harmonics number h, THD, as per the example of Table 1 below, or different sets of power consumption representing variables as indicated above ( chosen among active/reactive peak power, active/reactive ready state power, current/voltage harmonics , Total Harmonic Distortion, etc . ) .

In the push -tag mode the end-user 350 can proactively noti fy the NILM procedure 130 on the start time and stop time of an operation of an unknown appliance , which is saved for improving the appliance models .

The end-user 350 can communicate with the NILM procedure 130 in the system 100 through either mobile and/or web applications , therefore in that case the system 100 is configured to communicate over mobile communications channels and/or networks accessing the World Wide Web .

With reference to figure 4A which represents a block schematics of the non-event-based module 320 , the nonevent-based module 320 is configured to detect the operations of the appliances in the monitored environment 105 which contributes to an aggregated load AS received as input , i . e . power values sampled at given time instants , over a given interval of time , e . g . 24 hours . The aggregated load AS over 24h is better shown in Figure 4B .

The non-event-based module 320 comprises a library 450 of appliance models AM, each one trained to identi fy the activity of a speci fic appliance in the aggregated load 410 . In figure 4A are indicated three models corresponding to a dishwashing machine DSW, a washing machine WM, a fridge FRG, respectively . With 420 is indicated applying an appliance model AM of the library 450 , corresponding to the k-th target appliance , for instance implemented as a neural network model , mainly based on deep learning techniques in combination with classical statistical inference methods, the aggregated load AS an input. The appliance models AM are trained offline in a supervised way using as input a large dataset of operations of the corresponding appliances, either received from the end-users or manually generated by expert personnel. The output of the application of the appliance model 420 is a sequence of probability estimates 430, i.e. probability values between zero and one covering the entire input time window received from the aggregated load 410, e.g. 24h.

As better shown in the figure 4B, the probability estimates 430 contains a probability window PWk in the probability estimate as a function of time t which the probability estimate, from 0 to 1, for the appliance to which the operation 420 refers is one, i.e. the model of the k-th target appliance in the monitored environment 105. The probability window PWk is thus a time interval of activity, from start to stop, in the example 13:45- 15-45.

Then, the probability estimates 430 are processed to extract activity windows AWk from the aggregated load AS, i.e. in the aggregated load AS as a function of time, over the given period, are selected the values corresponding to sub-interval of time of the aggregated load AS corresponding to the one or more probability window PWk in the probability estimates 430, as indicated in the diagram 440. The one or more activity windows AWk are given as input to the subsequent event--based module 330, which is configured to extract the true disaggregated load of the monitored appliance. In figure 4C the time diagram of the aggregated signal AS with the activity window AWk as selected in the aggregated load AS over 24h and the zoom diagram of a windowed signal ASk extracted from activity window AWk over its time window (13:45-15:45 in the example) is shown. Thus , summing up a probability window PWk is a time interval on which the sequence of probability estimate , calculate on the aggregate load signal AS by applying 420 the model AM corresponding to the target appliance , has maximum value , the activity window AWk is the segment of the aggregate signal AS corresponding on the time scale to the time interval of the probability window PWk -

The final output of the non-event-based module 320 can contain multiple windows of operation 440 for each appliance that can be detected by the appliance models 420 . As shown in figure 5 in diagram 440 the activity window AWk shown comprises values of aggregated load AS , speci fically active power P, over an interval of two hours from 13 : 45 to 15 : 45 , which the non-event-based module 320 has identi fied as corresponding to the interval of operation of a given k-th target appliance, for instance the dishwashing machine DSW, i . e . , the interval in which the target appliance has been operating .

In figure 5 it i s shown a block schematics of the event-based module 330 , which is configured to extract the disaggregated load DALk of a given k-th target appliance of the monitored environment 105 , given as input an activity window AWk which allows to select and extract a windowed aggregate signal ASk, i . e . the aggregated signal AS in the corresponding time interval comprising the complete operation of a given k-th target appliance in set 105 . In detail , the event-based module 330 is configured to extract from the windowed aggregate signal ASk only power events E belonging to such k-th target appliance , while ignoring all the events belonging to other appliances or devices in the set 105 which may be also active in the activity window AWk - In other words , the event-based module 330 acts as a filter for the k-th target appliance , whose load or power consumption is to be disaggregated from the aggregated load AS .

Thus , the non-event-based detection 320 comprises comparing, in block 420 , to said aggregate load AS an appliance model representing the power consumption of the selected target appliance during a respective operation stored in a library of appliance model s 450 , in particular said appliance model being a neural network model , and said aggregated load AS being applied as input to said neural network model , said comparing operation outputting probability estimates PWk as a function of time over the given first period, e . g . a day, indicating the probability of a matching between the appliance model and the aggregated load AS , then associating an activity window AWk indicating the time of start and the time of stop of said time intervals of operation of the selected target appliance to a time interval in which the probability estimates PWk is greater over the given first period, in particular above a threshold or corresponding to a maximum value .

Figure 5 describes the di f ferent processing stages of the event-based module 330 , starting from the activity window AWk, i . e . the aggregated load received as input to the final disaggregated load of the targeted appliance .

Thus , the event-based module 330 comprises first a preprocessing stage 510 , which is configured to perform a robust filtering of the windowed aggregate signal AS k in the activity window AWk, received as input , to remove noisy and outlier spikes that do not correspond to actual step changes in power of appliances turn-on or turn-of f events . In figure 6A it is shown a time diagram of the windowed aggregate signal AS k as aggregated active power measured by the smart meter 140 over the monitored environment 105 over time t . In f igure 6B it is shown a time diagram of the filtered windowed aggregate signal AFSk as, in the example, aggregated active power measured by the smart meter 140 over the monitored environment 105 over time t, at the output of the preprocessing stage 510.

Then in an Event Detection 520 block on the filtered windowed aggregate signal AFSk it is performed detection of events E, defined as a significant step change, positive or negative in such signal, due to the electrical activity of certain appliances. With this criterion are detected transient states and start-up features of appliances, that are then used to extract further signatures. A signature is a sequence of power events E characterizing the operation of an appliance. Real and reactive power, P(t) and Q(t) respectively are fed to the inputs, although in figure 6A-6F only active power is shown for simplicity. In figure 6C it is shown a marked aggregate signal AMSk in which the detected power events E are indicated by full circles. As mentioned, a power event E can be identified if there is a significant step change, i.e. a variation over a given amount (or percentage of the steady state power) of the filtered windowed aggregate signal AFSk, i.e. of the power or load, over a given sampling time interval, e.g. a different quotient of the power with respect to the sampling interval. For instance, indicating with AP the power change in the filtered windowed aggregate signal AFSk and At the sampling interval, which in the example is one second, a power event occurs if AP/At is above a given threshold value. The sampling interval At may be shorter in time than the time interval over which the change, or transient, takes place, thus several events E can be associated to a transient. The filtered windowed aggregate signal AFSk as a function of time t is thus transformed in a sequence of power events E at given instants of time connected by segments ET, which are substantially constant power segments and segments with a steep increment or decrement of power, denoting a transient .

Then in a Feature Extraction block 530, once active windows are marked, i.e. on the basis of events E obtained at stage 530, obtained at the stage 530, features F are extracted, from the transient portion of the filtered windowed aggregated load signal AFSk- Of course, this operation may be performed also on the original unfiltered windowed signal ASk- Given the events E, considered as time-power value coordinate pairs, (t, AP) , as shown in figure 6D a feature F corresponds in general to an event E at the same time coordinate t, while the power change AP is taken with sign, i.e. an increase is associated to a positive value of power change AP, a decrease in power to a negative value of power change AP, thus estimating positive and negative peak power amplitudes. In addition, in variant embodiments, for the determination of features F estimation of other quantities such as spike widths, gradients, peak to peak amplitude and peak amplitude to A(P, Q) ratios may also be used in a processing of the unfiltered signal, i.e. the windowed aggregate signal ASk- In figure 6D the identified features F are shown as vertical bars, in correspondence of the time of the peak power event E. As better seen in Figure 6E, this representation is applied also to the power change with small amplitude.

Then, in Event Clustering 540 the features F are grouped into separate clusters CO, Cl, ..., Ci,...Cn according to the value of their amplitude. In this stage 540, a non-parametric clustering algorithm, mean-shift clustering, is used because it is not available the knowledge of the number of existing appliances in advance . The great benefit of the mean-shi ft clustering algorithm is that it is non-parametric which is independent of the underlying distribution and entails a mode-seeking algorithm . In other words , the advantage involved by the non-parametric clustering is that there is no need to set in advance the number of clusters to find, ( as it happens in the K-means algorithm) , but the number of clusters is estimated during the execution of the algorithm itsel f . This because usually it i s not known a priori how many clusters pertaining di f ferent events are present in the activation window . Once all clusters have been identi fied, only those clusters containing the power levels of the speci fic appl iance that it is desired to disaggregate are selected . This is obtained by using a set of "house-speci fic parameters" HSP ( such as active/reactive peak power, active/reactive ready state power, current/voltage harmonics , Total Harmonic Distortion, etc . ) describing the power levels of each appliance of the monitored environment 105 , as better described with reference to Figures 7A-7E . In figure 6E are shown with full circles the events E belonging to a first cluster CO and with hollow circles the events belonging to a second cluster Cl . In this case the clustering distinguishes simply between positive amplitude power features ( CO ) and negative amplitude power features ( Cl ) .

Finally, in an Events Pairing stage 550 , in the matching process , the features F, in pairs of positive power ( CO ) and negative power ( Cl ) , are compared to the values of specific appliances in the house-speci fic parameters HSP database so as to infer the usage interval of each appliance , obtaining the- corresponding disaggregated load DAL -

Figures 7A-7E , describe in detail how the housespeci fic parameters HSP are used in the event clustering module 540 and Events Pairing stage 550 , by showing time diagrams of the relevant quantities interconnected by arrows representing the relevant steps . In particular, it is shown that the house-speci fic parameters HSP allow to disambiguate between multiple appliances operating in the same window AWk of aggregated load ASk - Indeed, most of the times the aggregated load AS presents overlapping operations from di f ferent appliances . For example , the aggregated load AS k in Figure 7A contains both the activity of a fridge FRG and the activity of a dishwasher DSW .

The corresponding house-speci fic parameters HSP (which is shown in a simpli fied form in figure 7B ) are shown in Table 1 here below .

Table 1

In the first column it is indicated the appliance , DSW or FRG, in the second column the electric parameter, and in the third the value of the electric parameter . All the values are expressed as a range between a positive value and a negative value .

In figure 7A it is shown the diagram of the load aggregate signal AS , over an interval of 4 hours . For simplicity' s sake the aggregate signal AS is here represented only by the active power P, although other parameters may be part of a n-variable load aggregate signal AS, as mentioned, (among active/reactive peak power, active/reactive ready state power, current/voltage harmonics, Total Harmonic Distortion, etc.) and Table 1 itself shows four parameters, P, Q, H, THD, thus load signal may be in embodiments a four variable vector [P[t) , Q(t) , H(t) , THD (t) ] . Then in figure 7B a representation of the house-specific parameters HSP table - which is a simplified for graphical reason representation of Table 1 above - is shown, representing the house-specific parameters HSP database .

After the aggregate signal AS of figure 7A undergoes steps 510-540, determining a set of clusters (C0-C4 as shown in figures 7B, 7C) , in step 550 the house parameters HSP are used to select the clusters with power level (or one or more other power-related quantity, as reactive power, number of harmonics, THD (Total Harmonic Distortion) ) matching that associate in the table HSP to a given appliance. Thus, in figure 7C it is shown the diagram DSW_C representing the power level taken from the dishwasher DSW (dashed lines) the house-specific parameters HSP which match the clusters CO, Cl obtained in step 540, the same in the diagram FRG_C where the power level of the fridge FRG is used, obtaining cluster of the fridge FRG C2, C3, distinguished on the basis of the power P is shown. Cluster C4 is also shown, indicated with a cross, that, since it does not match any power level in the house-specific parameters HSP database it is classified as an outlier.

Thanks to the house-specific parameters HSP database, thus the stage 550 is able to select only the power events belonging to the appliance it is desired to disaggregate. As shown in figure 7C, from the clusters Ci obtained are extracted the clusters CO , Cl ( triangle up and triangle down symbols respectively) pertaining the positive and negative clusters matching the housespeci fic parameters HSP of the dishwater DSW, speci fically with the power P ± 1950 W, and the clusters C2 , C3 ( square , circle and cross symbol respectively) , speci fically with the power P ± 90 W . Cluster C4 is classi fied as an outlier since does not match any power level in the house-speci fic parameters HSP table .

The event pairing module 550 then uses the selected events , i . e . the clusters of each appliance , to reconstruct the original disaggregated load SALk of the target appliance as shown in figure 7E , i . e . once the clusters of a target appliance are identi fied, the corresponding disaggregated load SALk is obtained as a polyline which segments connect the corresponding power events E . In figure 7E the polyline corresponding to the dishwater DSW is indicated for instance as disaggregated load SALk of the target appliance , although also the polyline corresponding to the disaggregated load of the fridge ERG is shown and can be an output of step 550 as well , together with the load of the dishwater DSW or separately .

However, besides the house-speci fic parameters HSP database , the system may also rely on some more general parameters defined for the di f ferent appliances , which enable the disaggregation task also in the absence of the confirmation of an end-user . By way of example , it may be known that the events related to the dishwater generally stay between 1700 W and 2200 W of active power P . Once the end-user confirms the operation detected by the non-event-based module , it is possible to update the events E with more precise estimates from the monitored environment . Subsequently, on the basis of the disaggregated load SALk for a given target appliance , the anomaly detection module 170 can perform a comparison of such di saggregated load SALk with an expected load for the target appliance and identi fy, on the basis of this comparison, electrical anomalies and/or behavioral anomalies in the operation as the method 300 provides profiles of the load of the speci fically targeted appliance , i . e . , disaggregated loads SALk, in function of time t with a high accuracy . The accurate extraction of appliance disaggregated load curve profiles provides a set of detailed information on the anomalies otherwise impossible to obtain j ust from the aggregate signal or other compact representations of such appliance . Indeed, information such as the type of anomaly and its potential causes are necessary in order to suggest targeted interventions to the end-user . To this regard, in figure 8 it is shown the disaggregated load of a washing machine DSW, in which a main area MA under the power curve represents the water heating stage and the minor spikes , magni fied in a zoom window ZW correspond to the spin cycles . Figure 8 shows the active power measured over a range of 4000 seconds , and the zoom window ZW over 100 seconds of time and between 1600 and 1950 W of active power W . As shown each transient due to a spike is circa ten seconds in time width .

In general , the anomaly detection module 17 ' i s configured to operate on the basis of an expected load for the target appliance , which is based on a standard or reference behaviour of the appliance power profile . The standard behaviour may represent a status of the appliance operating in its optimal conditions . In most cases , the standard behaviour of a certain appliance is obtained by analyzing its past operations within the same monitored environment , creating a model which is representative of the collected appliance operations . In other cases, it is possible to rely on a general model obtained from multiple houses in order to identify outliers across different users. Once the standard behaviour for the target appliance has been obtained, the anomaly detection module 170 is configured to monitor the disaggregated load of the target appliance, looking for outliers presenting significant deviations from the expected load.

Thus, the anomaly detection module 170 is configured on the selected target appliance on the basis of said disaggregated load to compare said disaggregated load to a reference or standard load for said selected target appliance in order to detect differences in amplitude and/or time and/or frequency which are identified as anomalies according to a set of criteria.

The anomaly detection module 170 can be configured to check the overall health status of the monitored appliance, by comparing the last extracted appliance signature, i.e. disaggregated load SALk, with past appliance signatures in order to detect significant changes in power consumption.

Also, once it is known the standard behaviour of an appliance, e.g. by analyzing the combinations of all its past operations, it is possible to estimate the duration and power consumption of the different stages composing the overall appliance operation. This information may be used to detect electrical faults concerning either the single appliance or the main meter of the household as well, e.g., if the disaggregated load SALk shows that a phase is not completed or performed.

In the same way, the analysis of the target appliance power profile on a given activity interval provides useful information about the general efficiency of the monitored appliance. Also, the power consumption of the appliance can be compared with the power consumption of s imilar appliances in other houses , in order to estimate the overall appliance ef ficiency with respect to devices operating in equivalent conditions . The high resolution of the extracted power signatures permits also to distinguish the di f ferent operations modes (programmes ) of the appliance and perform further analysis on the basis of the associated power consumption values , e . g identi fy the most ef ficient operation modes for the monitored appliance .

Also , given that the disaggregated load allows to precisely detect the start time and the stop time of single operations of the appliance , it is possible to perform an accurate estimate of the duration and the time of usage of a single appliance operation . This can be used to detect anomalous behaviors in the end-users , for instance concerning an excessive usage of one or multiple devices , i . e . too many operations of a given appliance .

The solution according to the various embodiments here described allows to obtain the following advantages .

The solution described allows to obtain detailed information on the power consumption of a given appliance , as disaggregated load, without requiring appliance speci fic sensors .

Of course , without prej udice to the principle of the embodiments , the details of construction and the embodiments may vary widely with respect to what has been described and illustrated herein purely by way of example , without thereby departing from the scope o f the present embodiments , as defined the ensuing claims .