Supervised learning classification method for fluid consumption events in a monitored building, categorizing them into several classes of consuming elements.

A supervised learning method using simulated data from a large number of buildings enhances the precision of fluid consumption event classification, addressing the imprecision of existing methods and enabling effective detection and correction of excessive consumption.

FR3139918B1Active Publication Date: 2024-11-29SMART & BLUE
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
FR2022009520
Authority / Receiving Office
FR · FR
Patent Type
Patents
Current Assignee / Owner
Filing Date
2022-09-20
Publication Date
2024-11-29
Estimated Expiration
2042-09-20

AI Technical Summary

Technical Problem

Existing methods for classifying fluid consumption events in buildings are not precise enough to detect excessive consumption by consumer elements, as they rely on limited real-world data and fail to account for the diversity of consumption habits.

Method used

A supervised learning method using simulated and classified measurement signals from a large number of simulated buildings, combined with event and temporality parameters, to create a prediction model that accurately classifies fluid consumption events.

Benefits of technology

Improves the precision and recall of fluid consumption event classification, enabling effective detection of excessive consumption and allowing for user alerts and corrections.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 00000029_0000
    Figure 00000029_0000
  • Figure 00000029_0001
    Figure 00000029_0001
  • Figure 00000030_0000
    Figure 00000030_0000
Patent Text Reader

Abstract

The invention relates to a method for classifying, by supervised learning, fluid consumption events in a monitored building into several classes of consuming elements. The classification is performed using a predictive model parameterized from a database. This database includes, for each identified fluid event: event parameters PEv(i), temporal parameters PT(i), and the corresponding LEV(i) class of consuming element EC. Figure 1B (for the abstract)
Need to check novelty before this filing date? Find Prior Art

Description

Title of the invention: Method for classifying by supervised learning fluid consumption events by a building to be monitored among several classes of consumer elements Technical field

[0001] The field of the invention is that of the fluid consumption of a building to be monitored, for example that of the consumption of liquid water by a dwelling house, a building, a school, a hospital, a factory, etc.

[0002] The invention relates more specifically to the classification, by supervised learning, of fluid consumption events, present in a measurement signal representative of the temporal evolution of the overall consumption flow rate of the fluid of interest by the building to be monitored, and in particular by different consuming elements, such as for example taps, showers, flush toilets in the case of a residential house.

[0003] The classification amounts to determining the class (label, category, or label in English) of consumer element which is to be associated with each of the fluid events. The invention makes it possible in particular to detect excessive consumption of consumer elements of one or other of the predefined classes, and then makes it possible to indicate to the user whether consumer elements are being used too much or have a flow rate above a predefined reference. STATE OF THE PRIOR ART

[0004] A building to be monitored can be a house, a building, a school, a hospital, a factory, a company hotel, etc. The fluid of interest can be, depending on the type of building to be monitored, a gas or a liquid such as water, hydrogen, oxygen, nitrogen, etc.

[0005] The building to be monitored comprises several devices connected by a fluid circuit to a fluid source and which consume the fluid of interest. These devices will be called here "consumer elements". The consumer elements can be classified into different predefined classes. For example, in the case of a residential house where the fluid of interest is liquid water, the consumer elements may fall into the class of taps, showers, toilet flushers, dishwashers, or even washing machines.

[0006] The building to be monitored is usually equipped with a flow sensor adapted to measure a signal representative of the temporal evolution of the overall flow rate of the consumption of the fluid of interest by the building to be monitored, and more precisely by its consumer elements. This measurement signal includes fluid consumption events, which are defined as moments of non-zero flow rate.

[0007] There is then a need to be able to determine with improved precision the class of consumer elements to be associated with each of the fluid events linked to the consumption of the fluid of interest by the building to be monitored. This then makes it possible to monitor the consumption of the fluid of interest by the building to be monitored, and to be able to facilitate the detection, for example, of excessive fluid consumption of this or that class of consumer elements. Statement of the invention

[0008] The invention aims to propose a method which makes it possible to classify with improved precision the fluidic events linked to the consumption of the fluid of interest by the consuming elements of the building to be monitored, that is to say to determine the class of consuming elements to be associated with each of the events. fluidics.

[0009]

[0010] For this purpose, the subject of the invention is a computer classification method of fluidic events Ev^) consumption of a fluid of interest by elements EC consumers of a building to be monitored BS, among several predefined classes LEV(jjt) of EC consumer elements, the process comprising the following phases. First, a phase of generation of at least one database (BD1, BD2), comprising the following steps: - acquire measurement signals Il)mesBref(n) ( | representative of the l vEC ;Jn=i^ consumption flow rate of each of the EC consumer elements of N real reference buildings {Bref riJ ;N, with N>1, the measurement signals then being said to be classified, the EC consumer elements being the same classes than those of the building to be monitored BS; define consumption profiles Pstat|c m ' n "} 1 m=lJd of M so-called buildings simulated {Bsim(mJ}m_| ;M, with M>N, from the acquired classified measurement signals, each simulated building comprising EC consumer elements of the same classes as those of the building to be monitored BS; generate, by digital simulation, simulated signals ft) ] representative of the consumption flow of each the EC EC consumer elements of the simulated buildings {Bsim(m)}m=1;M having the defined consumption profiles, the generated simulated signals then being said to be classified; determining, for each of the fluid events {Ev(i)}i=i;NA identified in the generated classified simulated signals: so-called event parameters {PEv(i)}i=i;NA, representative of a duration and an elapsed volume of each of the fluid events; and so-called temporality parameters {PT(i)}i=1;NA, representative of a number of so-called similar fluid events, having a flow rate substantially identical to that of the fluid event considered and located in predefined successive time slots located before and after the fluid event considered; generate the database comprising, for each of the fluidic events {Ev(i)}i=i;NA identified: the event parameters {PEv(i)}i=1;NA, the temporality parameters {PT(i)}i=1;NA, and the corresponding class {LEV(î)}i=i;na of consumer element EC.

[0011] Then, a phase of parameterizing at least one prediction model by supervised automatic learning from the database.

[0012] Finally, a classification phase, comprising the following steps: - acquire a measurement signal SmesBS(t) representative of the consumption flow rate of the building to be monitored; - determine, for each of the fluidic events {Ëv^^i ;np identified in the acquired measurement signal, event parameters {Pgy^pj^i ;np and corresponding temporality parameters {;np; - predict, by the prediction model, for each of the fluidic events { ;np whose event parameters {Pgy^^i ;np and temporality parameters {Py^}j=i ;np form input data to the prediction model, the corresponding class {LEVQkp>i ;np of consumer element EC.

[0013] Some preferred but non-limiting aspects of this classification method are the following.

[0014] The temporal parameters PT(i) and P^ can also include, for each of the time slots, the cumulative volume of the fluid of interest flowed and the flow duration.

[0015] The PEv(i) event parameters may include, for each fluid event, the initial time of the fluid event, the flow duration, the elapsed volume and / or the average flow rate.

[0016] The database and the input data of the prediction model may include so-called comparison parameters, denoted respectively PEv|Bsim(i) and PEy|BS(jJ <i-*k))’ ct ^finis, pour chaque évènement fluidique Ev(i), Ev^y comme le rapport d’un débit moyen de l’évènement fluidique considéré sur un débit global moyen de tous les évènements fluidiques Ev(i), du bâtiment considéré.

[0017] The classification phase can be repeated, the measurement signals SmesBS(t) of the building to be monitored then being acquired over successive predefined monitoring durations.

[0018] In the comparison parameters P£v|Bs(jjd~*k) Pæscnts in the input data of the prediction model, the average global flow rate of all fluid events Ëvqy can take into account the average global flow rate of the previous iterations.

[0019] The generation phase may include the following steps: - determination of a first database where each fluidic event is defined by at least said event parameters and by said temporality parameters; - determination of a second database where each fluidic event is defined by at least said event parameters, by said temporality parameters, and by so-called comparison parameters by class defined, for each fluidic event of at least one class considered, as the ratio of an average flow rate of the fluidic event considered to an average overall flow rate of all the fluidic events of the class considered.

[0020] The parametric phase can then include the following steps: - setting up a first prediction mode by supervised machine learning from the first database; - setting up a second prediction mode by supervised machine learning from the second database;

[0021] The classification phase can then include the following steps: - determination, for fluidic events present in the measurement signal SmesBS(t), of the event parameters and the corresponding temporal parameters; then - prediction by the first prediction model, for each of the fluid events, which are defined by the event parameters and the temporality parameters which form input data of the first prediction model, of the corresponding consumer element class; then - determination, from the previously predicted classes, of corresponding comparison parameters per class; then - prediction by the second prediction model, for each of the fluidic events, which are defined by the event parameters, the pa temporality parameters, and the class comparison parameters that form input data of the second prediction model, of the corresponding consumer element class.

[0022] During the step of acquiring the measurement signals f ( t ) I , the N real reference buildings can be connected to a device for determining at least one prediction model, comprising fluid measurement sensors adapted to acquire the measurement signals for each of the consumer elements of the real reference building considered, and a computer connected to the fluid measurement sensors of the reference building considered and adapted to carry out the parameterization phase.

[0023] During the step of acquiring the measurement signal SmesBS(t) of the building to be monitored, the latter can be connected to a classification device comprising a fluid measurement sensor adapted to acquire the measurement signal SmesBS(t), and a computer connected to the fluid measurement sensor of the building to be monitored and adapted to carry out the classification phase.

[0024] Following the classification phase, the following steps can be carried out: - identification of a class of consumer elements for which the fluidic events have a value of a so-called monitoring parameter greater than a predefined threshold value; then - alerts the user indicating the class of consumer elements in question; then - correction by the user so that the consuming elements of the class in question then have a value of the tracking parameter lower than the predefined threshold value. Brief description of the drawings

[0025] Other aspects, aims, advantages and characteristics of the invention will appear better on reading the following detailed description of preferred embodiments thereof, given by way of non-limiting example, and made with reference to the appended drawings in which:

[0026] [Fig.1A] schematically illustrates an example of a building to be monitored, here a residential house, equipped with a device for classifying fluid events linked to the consumption of the fluid of interest by the building to be monitored;

[0027] [Fig.lB] illustrates an example of a measurement signal representative of the temporal evolution of the overall consumption flow rate of the fluid of interest by the building to be monitored in [Fig.lA];

[0028] [Fig.2] illustrates a flowchart of an event classification process fluids linked to the consumption of the fluid of interest by the building to be monitored, according to one embodiment;

[0029] [Fig.3A] schematically illustrates an example of a real reference building, here a residential house, equipped with a device for determining a prediction model by supervised learning;

[0030] [Fig.3B] illustrates examples of measurement signals representative of the local consumption flow rate by each of the EC consumer elements of the real reference building of [Fig.3A];

[0031] [Fig.3C] illustrates time slots located before and after the initial instant of a fluidic event, making it possible to define so-called temporality parameters representative of each of the fluidic events;

[0032] [Fig.4] illustrates steps implemented during the database generation phase, from simulated signals for so-called simulated buildings;

[0033] [Fig.5] illustrates a flowchart of a method for classifying fluid events linked to the consumption of the fluid of interest by the building to be monitored, according to another embodiment which implements two prediction models;

[0034] Figures 6A to 6C illustrate examples of cumulative distributions of flow rate, flow duration and elapsed volume, associated with fluid events from simulated signals for a simulated building, and compared to those from a real building;

[0035] Figures 7A to 7C respectively illustrate an example of the precision rate, recall rate and Fl-score rate, associated with the classification of fluid events of a building to be monitored carried out by classification methods according to different embodiments.

[0036] DETAILED DESCRIPTION OF PARTICULAR EMBODIMENTS

[0037] In the figures and in the remainder of the description, the same references represent identical or similar elements. In addition, the different elements are not shown to scale so as to enhance the clarity of the figures. Furthermore, the different embodiments and variants are not mutually exclusive and may be combined with each other. Unless otherwise indicated, the terms "substantially", "approximately", "in the order of" mean to within 10%, and preferably to within 5%. Furthermore, the terms "between ... and ..." and equivalents mean that the limits are included, unless otherwise stated.

[0038] The invention relates to the monitoring of fluid consumption within a building which consumes a fluid of interest via EC consumer elements of different LEv classes. It relates more specifically to the determination of the LEv class (label or tag) of EC consumer elements which is to be associated with each of the fluid events Ev linked to the consumption of the fluid of interest by the EC consumer elements of the building to be monitored BS.

[0039] The classification method is carried out using a computer (processing unit) which integrates at least one prediction model fl, f2 by supervised machine learning. The prediction model fl, f2 can be a decision tree model, a weak classifier boosting model, a linear or quadratic discriminant analysis model, a decision tree forest model, a k-nearest neighbor method, among others.

[0040] As detailed below, to obtain improved prediction performance, the classification method comprises a phase of learning the prediction model f1, f2 which is carried out from at least one database BD1, BD2 preprocessed in a particular manner.

[0041] This database BD1, BD2 is not formed, as it could be according to a natural approach, by classified (labeled) flow measurement signals coming from each of the EC consumer elements of several real reference buildings Bref. Indeed, this approach cannot be sufficiently representative of the diversity of the actual fluid consumption in real buildings. On the contrary, according to the invention, the database is formed from simulated and classified signals, representative of the flow rate of each of the EC consumer elements of simulated buildings B sim, and preferably of a number M of simulated buildings B sim much greater than the number N of real reference buildings Bref.The PstatBsim consumption profiles of the simulated buildings, from which the numerical simulation was carried out, were determined from the classified flow measurement signals from the real reference buildings, so that the simulated signals remain consistent with the actual consumption of the real buildings. Thus, it appears that the prediction quality is greatly improved.

[0042] In addition, the database is formed from the various fluidic events Ev of the fluidic consumption of the consumer elements EC of the simulated buildings. More precisely, it is formed from their representative parameters, including: - so-called event parameters PEv, representative of the event itself (flow duration, elapsed volume or equivalent (average flow rate for example), or even initial instant, etc.), and - so-called temporality parameters PT which indicate, for each of the fluid events Ev, the number of similar fluid events (in terms of average flow rate, or even flow duration), included in predefined time slots and located before and after the fluid event considered.

[0043] It appears that learning the prediction model from such a database improves prediction performance, in particular its precision, its recall, and therefore also its Fl-score (harmonic mean of the precision and recall).

[0044] Figure 1A illustrates an example of a building to be monitored BS comprising several consumer elements EC of different classes, and equipped with a classification device adapted to classify the fluid events Ëv linked to the consumption of the fluid of interest by the consumer elements EC, i.e. adapted to determine the class LEv of each of the fluid events of consumption by the building to be monitored BS.

[0045] The building to be monitored BS is one or more structures that consume a fluid of interest via its consumer elements EC, and whose consumption is to be monitored in order to detect, in particular, possible anomalies such as excessive consumption by the consumer elements EC of a certain class. Thus, the user is able to know the consumption of each class of consumer elements EC, and can thus know, for example, if the average flow rate of a class of consumer elements EC is greater than a predefined value. By consumption, it is meant that the building to be monitored BS receives the fluid of interest from a fluid source SF, and uses it (“consumes” it) according to different uses, which may be personal and / or professional.Such a building to be monitored BS can be, for example, a house, an apartment, a building with several apartments, a school, a factory, a hospital, a campsite, or other, and in general, is any type of structure or complex for personal and / or professional use. The building to be monitored BS can include several separate buildings, as for example in the case of a factory.

[0046] The fluid of interest may be a liquid or a gas, such as water, hydrogen, oxygen, nitrogen, helium, etc. Thus, the building to be monitored BS may be a factory that consumes, for example, hydrogen, oxygen or nitrogen, in liquid or gaseous phase. It may also be a training building (school) that uses liquid water. In the remainder of the description, the building to be monitored BS is a residential house, and the fluid of interest is liquid water.

[0047] The building to be monitored BS has several EC consumer elements which ensure the effective consumption of the fluid of interest. The EC consumer elements are of different L classes, for example a class R for taps (toilets, kitchen, bathroom, etc.), a class D for showers, a class C for toilet flushes, another class LV for dishwashers, and another class LL for washing machines. In this example, the classes are noted: L = {R, D, C, LV, LL}.

[0048] The building to be monitored BS also includes a fluid circuit which ensures the distribution of the fluid of interest from the fluid source SF to the consumer elements EC. These are distribution conduits, possibly equipped with valves. The SF fluid source of the fluid of interest can be a supply network, for example from the city, a reservoir, or other.

[0049] The building to be monitored BS is equipped with a classification device, which comprises a fluid measurement sensor CM for acquiring a measurement signal SmesBS(t) representative of the temporal evolution of the overall flow rate, at the input of the building to be monitored BS. The classification device also comprises a processing unit UT (computer) for determining the LEV classes of the fluid events Ëv linked to the consumption of the fluid of interest by the consumer elements EC and present in the measurement signal SmesBS(t). The processing unit UT integrates at least one prediction model fl by classification, the parameterization of which was carried out by supervised automatic learning from a predefined database BD1.

[0050] Note that the classification device is not able to know the temporal evolution of the flow rate of each of the consumer elements EC, but only to know the overall flow rate at the entrance to the building to be monitored BS. It is because of this absence of sensors dedicated to each of the consumer elements EC, and therefore the fact that only the overall flow rate information is accessible, that the invention provides for using a prediction model, here by supervised learning, to estimate the class LEV of consumer element EC which is to be associated with such and such fluidic event Ëv of overall flow rate.

[0051] The fluid measurement sensor CM is connected to the fluid circuit, and is located between the fluid source SF and the consumer elements EC. It is adapted to measure a measurement signal SmesBS(t) representative of the overall consumption flow rate of the fluid of interest by the building to be monitored BS. This CM fluid measurement sensor can be a flow meter water meter, a pulse water meter, among others.

[0052] In the remainder of the description, it is considered that the fluidic measurement sensor CM is a turbine flowmeter water meter, which acquires a measurement signal Smes^t), whose data are representative of the different fluid events Ev. A fluid event is defined as being a non-zero flow duration delimited in time by zero flow durations.

[0053] Figure 1B illustrates an example of measurement signal SmesBS(t), here in the form of a temporal evolution of the overall flow rate linked to the water consumption of a house to be monitored BS. The measurement signal SmesBS(t) presents several fluidic events Ey^, °ù j is an increment that goes from 1 to NP, with NP the total number of events fluidics Ev over a monitoring period Ats (for example 1 day). The measurement signal SmesBS(t) can be presented as a succession of electronic messages relating to each of the fluidic events: SmesBS(t) = {PevQ)} . Each j=l,NP fluidic event Ëvqj is characterized here by so-called event parameters, for example here its instant of start of flow t^, the duration of flow At^, and the average flow rate, or even the elapsed volume V(i): ; AIV(j)] • These parameters are indicated here as an example, and other parameters are possible. The measured information can be transmitted in real time to the processing unit, for example at a predefined frequency or as soon as a fluidic event Ëv<-\ cst

[0054]

[0055] completed, in the form of one or more electronic messages each containing the parameters PEvq) of the fluidic event Ëvq considered. Alternatively, the measurement signal SmesBS(t) may be a vector whose values ​​correspond to the flow rate measured at regular frequency. In the case of a pulse water meter, the measurement signal SmesBS(t) may be a vector comprising only the instant at which a pulse is emitted. This pulse is emitted when a predefined volume of the fluid of interest has flowed (e.g. 1 pulse per liter). The processing unit UT is connected to the fluid measurement sensor CM in a wired or wireless manner. It can be arranged in the building to be monitored BS or can be located remotely from it. It is a computer which comprises a calculator and at least one memory. It allows the implementation of the operations of the classification method making it possible to determine the class LgyQ of the fluid events Ëv^ present in the measurement signal SmesBS(t). The calculator comprises a programmable processor capable of executing instructions recorded on an information recording medium. The memory contains instructions for implementing the classification method. It is also adapted to store the information received by the fluid measurement sensor CM, and here comprises at least one prediction model fl (or fl, f2) having been parameterized during a supervised learning phase from at least one database BD1 (or BD1, BD2).

[0056] Figure 2 is a flowchart of a method for classifying, according to one embodiment, the classes Lgv(j) of consumer elements EC to be associated with each of the fluidic events Ëv^j present in the measurement signal SmesBS(t). The method comprises a phase 100 of generating a database BD1, a phase 200 of learning, followed by a phase 300 of prediction.

[0057] In this example, the building to be monitored BS is a residential house with several inhabitants. The classification method is obviously not limited to this example. The building to be monitored BS comprises consumer elements EC of different known classes L, namely here classes: R (taps), C (flush toilets), D (showers), LL (washing machines) and LV (dishwashers).

[0058] Phase 100: Generation of the BDL database

[0059] The database BD1 is generated on the basis of measurement signals DmesBrTf <nJ < ( ti représentatifs du débit réel (i.e. du débit mesuré et non simulé) de each of the EC consumer elements of N real reference buildings Bref(n), with n ranging from 1 to N>1. These are therefore classified (i.e. labeled) measurement signals, given that we have a measurement signal for each of the EC consumer elements. The real reference building Bref(n) is identical or similar to the building to be monitored BS, in the sense that it has the same use (here a residential house) and includes EC consumer elements of the same class L (here: tap R, shower D, flush C, washing machine LL, and dishwasher LV).

[0060] Phase 100 of generating the database BD1 is illustrated in more detail in [Fig.4]. This phase 100 comprises the following steps: • a step 110 of acquisition, by experimental measurement, of measurement signals EJmes^c-pVi) (t)' representative of the real flow rate of each of the consumer elements EC of the N real reference houses Bref(n); • a step 120 of determining profiles, called statistics, of consumption Pstatgc-K CD each simulated building Bsim(m), with m ranging from 1 to M>N; • a step 130 of generation, by digital simulation, of simulated signals Dsinip^^'cD 0' representative of the flow rate of each of the consumer elements EC of the M simulated houses Bsim(mJ; • a step 140 of determining, for each of the fluid events Ev identified in the simulated signals, at least the event parameters P EV and the temporality parameters PT; • a step 150 of generation of the BDL database

[0061] During step 110, measurement signals DmesBref(tt) , (ti) are acquired, representing the actual flow rate of each of the consumer elements EC of the N houses actual reference numbers in Brief(n), over at least one Ats monitoring period (here at least one day). The number N can be equal, for example, to a few units, or even to a dozen. It is distinguished from the number M of simulated houses which can be equal to a few hundred, thousands, tens of thousands or even more.

[0062] [Fig.3A] schematically and partially illustrates a real reference house, Brief, identical or similar to the house to be monitored BS (here a residential house). The reference house is said to be real insofar as it is occupied by real users, and is therefore not a simulated house. It includes consumer elements EC whose classes are L = { R, C, D, LL, LV}. These are the classes of taps R, flush toilets C, showers D, washing machines LL, and dishwashers LV. The reference house Bref may or may not have the same number of inhabitants as the house to be monitored BS.

[0063] The N real reference houses Bref(n) are connected, directly or indirectly, to a device for determining at least one prediction model fl, f2. This device comprises fluid measurement sensors CM and a processing unit UT. In this example, the fluid measurement sensors CM acquire measurement signals DmesBref(,,) ( 0 which correspond to the temporal evolution of the flow rate of each of the consumer elements EC of the N real reference houses, and the transmit to the processing unit UT. In other words, each consumer element EC is equipped with a fluidic measurement sensor CM. The measurement signals DmesB1Tf(nJ . (t) are here vectors indicating the value of the flow rate measured at a predefined frequency, for example every second. They are noted: Dmes|^'' CJ) (t) = { Dmes| ief(1 ”(t) ; Dmes® ref(n) (t) ; Dmes^U'^t) ; DmesBæf'"'(t) DmesBvf(n,(t)}• Hs can obviously have other formats,

[0064]

[0065] such as a succession of electronic messages relating to each of the measured fluid events. [Fig.3B] illustrates examples of measurement signals from a faucet DmesR(t), a shower DmesD(t), a flush Dmesc(t), a washing machine DmesLL(t), and a dishwasher DmesLV(t). We note that it is possible to extract the fluid events Ev and the corresponding event parameters PEv, which include the initial instant te, the flow duration Ate, the elapsed volume V, the average flow rate, etc. We can also determine a statistical consumption profile associated with each reference house Bref(n), the maximum flow rate, etc., but also the number of uses, the average frequency and the standard deviation of uses, etc. Then, during step 120, statistical consumption profiles are determined (of the M simulated houses, from information from the 'EC' m=l^ acquired measurement signals | , and more precisely, from statistical consumption profiles n=I;N {PstatECef(n)} associated with the N houses of n=l;N reference which are determined from the acquired measurement signals.

[0066] Thus, a statistical profile of a building groups together information related to this building and its occupants / users, such as the number of occupants, the number and class of the different EC consumer elements. It also includes information statistical formations related to the fluid consumption habits of each of the occupants, such as the typical times of use of the different EC consumer elements, flow times, pauses between each fluid event (in particular for showers). Finally, it includes statistical information related to the EC consumer elements themselves: maximum flow rate, average flow rate, etc. This consumption information is called statistical insofar as it can include an average value and a variability (standard deviation) associated with a given distribution (normal, log-normal, exponential, bimodal, etc.).

[0067] Thus, for each reference house Bref(n), we obtain the information related to the house and its occupants. In addition, from the measurement signals acquired for each of the consumer elements EC, we determine the statistical consumption information. We thus define the corresponding statistical consumption profile Pstat|cfw-

[0068] The statistical consumption profiles Pstat|^nVm) are then determined for each of the M simulated houses Bsim(m), from the statistical consumption profiles of the reference houses Bref(n). As indicated previously, the number M can be very high compared to the number N, for example approximately 100,000 while N would be approximately 10. The statistical consumption profiles are determined by varying the different parameters of the profile. This will make it possible to avoid subsequently parameterizing the prediction model fl with a database BD1 whose values ​​would not be sufficiently representative of the possible effective variability of the fluid consumption of the real houses, and therefore of the house to be monitored BS.

[0069] Thus, concerning the parameters linked to the house and its occupants, houses are generated by varying the number of occupants, the number of consumer elements of each class, the number of classes present (thus some houses may not have a washing machine or a dishwasher). Statistical parameters linked to the consumption habits of each of the occupants are also varied, for example the typical schedules and the frequency of use of the different EC consumer elements. Thus, for example, for some occupants, the shower is taken in the morning, for others in the evening. For some, the duration of the shower will be long, for others short, etc. Finally, the parameters linked to the EC consumer elements themselves are varied in terms of average flow rate, maximum flow rate, standard deviation, etc. This gives the statistical consumption profiles Pstadg"' nv of the M simulated houses. m) / 1 \ 1 representative of the consumption flow rate of each of the 1 m=l:M {T.. . Bsini Dsim^ EC consumer elements of the M simulated buildings, based on the corresponding statistical consumption profiles. These simulated signals are said to be classified to the extent that a consumer element class is associated with each of the signals.

[0071] Thus, for each simulated house (and for each occupant), and taking into account the parameters of the corresponding statistical consumption profile, we simulate the temporal evolution of the flow rate Dsira^8™1'”'(t) each shower during a monitoring period Ats (here at least one day), as well as that ( t ) each flush and that ( t ) each faucet. We can also simulate that DsiniLL^Xt) ^es ^ave"^nSes and that DsimLyntm)(t) ^es lave'dishes- Of course, filters are present to prevent simulated signals of class D 'shower', for several occupants of the same simulated house, from presenting simultaneous EvD fluidic events... while the simulated house Bsim(mJ only has one shower.

[0072] Each simulated signal Dsim^^ ( t ) comprises one or more fluid events. For example, it has the format of a vector indicating the value of the flow rate at a frequency of, for example, one second over a period of one day, and preferably over a period of several days, or even weeks. The simulated signals can thus take the form of the signals illustrated in [Fig.3B].

[0073] Thus, a large number of realistic signals are obtained which are representative of the actual flow rate of the EC consumer elements in real houses, which will make it possible to generate a realistic database BD1 of large dimension. This great diversity of the simulated signals will make it possible to avoid prediction biases which could be induced by a database which would not be sufficiently representative of the possible actual diversity of consumption habits in real houses.

[0074] Then, during a step 140, parameters associated with each fluidic event associated with the fluidic consumption of the simulated houses Bsim(m) are determined, and more precisely event parameters PEv and temporality parameters Pt-

[0075] To do this, simulated global flow rate signals SsimBsim(m)(t) are determined, each of which is representative of the global flow rate of the corresponding simulated house Bsim(m), from the simulated signals Dsimæ^RCD (0 ' saS^ obviously classified (labeled) signals, insofar as they are associated with a signal LBsim(m)(t) indicating the corresponding class. This amounts to obtaining, for each simulated house Bsim(m) a signal whose format is identical to the measurement signal SmesBS(t) of the house to be monitored BS. In this example, the measurement signal SmesBS(t) is formed from a succession of fluid events |PEV^| and more precisely from a succession of parameters representative of the fluid events.

[0076] To do this, we add all the simulated signals ' D$im|set Dsim^^) relating to the same simulated house Bsim(m). We thus obtain, for each simulated house Bsim(m), the simulated overall flow signal SsirnBsim(m)(t) accompanied by the corresponding class signal LBsim(m)(t).

[0077] Thus, for each of the simulated global flow rate signals SsimBsim(m)(t), we identify the fluid events EvBsim(m) present and we associate them with the corresponding class t. Remember that the fluid events are events with non-zero flow rate ^Ev framed before and after by a period with zero flow rate.

[0078] However, a fluidic event EvBsim(m) of the simulated global flow signal Ssim Bsim(m)(0 pCU( may be the result of at least partial juxtaposition of several fluidic events from the simulated signals Dsiingcm <ny ( t ) ' Aussi, on peut attribuer la nouvelle classe ‘mixte’ ou ‘mélange’, notée M, à chaque évènement fluidique Ev d’un même signal simulé de débit global SsirnBsim(m)(t) et issu d’au moins deux éléments consommateurs.

[0079] Then, each of the fluid events Ev of the same simulated global flow signal SsirnBsim(m)(t) is characterized by event parameters PEv, namely for example the start time te, the flow duration Ate and the elapsed volume V or equivalent (average flow, etc.). We thus obtain a simulated global flow signal Ssim Bsim(m)(jj comprising a succession of fluid events and their characteristic parameters PEV, with the corresponding classes LEV: { PEv ; LEv}.

[0080] Thus, so-called temporality parameters PT are determined for each of the fluid events of the simulated global flow signals SsimBsim(m)(t). This involves enriching the information associated with the fluid events Ev of each simulated global flow signal SsirnBsim(m)(t), with correlation information with similar fluid events located before and after each of the fluid events considered. Indeed, it appears that enriching the database with such temporality parameters makes it possible to improve the prediction performance of the model.

[0081] As illustrated in [Fig.3C], we define a plurality of time slots C located before and after the initial instant te of a fluidic event Ev. We can thus define NC av time slots located before the initial instant te and NCap time slots located after the initial time te. In this example, NCav and NCap are identical but they could be different from each other, and here are each equal to 5. In addition, the upstream Cav and downstream Cap time slots have the same duration two by two, but they could have different durations. We can thus have: - two slots Qu and C(+ij ranging from [-tnci; te] and [te; +tnci], for example ranging from f. to more or less 7min; - two slots C(_2) and C(+2) going from [-tnc2; -tnci] and [+tnci; +tnc2], for example going from -30min to -7min, and going from +7min to +30min; - two slots C(3) and C(+3) going from [-tnc3; -tnc2] and [+tnc2; +tnc3], for example going from -60min to -30min, and going from +30min to +60min; - two slots C(.4) and C(+4) going from [-tnc4; -tnc3] and [+tnc3; +tnc4], for example going from -120min to -60min, and going from +60min to +120min; - two slots C(_5) and C(+5) ranging from [-tnc5; -tnc4] and [+tnc4; +tnc5], for example ranging from -240min to -120min, and ranging from + 120min to +240min;

[0082] The number of fluid events {Nc(P)}P= NCav;+NCaP similar to that considered in terms of flow rate (for example, same average flow rate plus or minus a predefined tolerance, for example equal to IL / min), whose initial instant is located in one or other of the time slots, is then determined. This makes it possible to establish the correlation between the fluid events of the same class. Indeed, it appears that the use of a consumer element, for example a tap or a shower, can present a succession of fluid events of the same flow rate, or even, in addition, of the same flow duration. This can be the case when using the shower, the tap for washing dishes, the cleaning program of a washing machine and a dishwasher, etc. The closest time slots allow, for example, the identification of showers, while the most distant time slots concern more the washing machines and dishwashers.On the other hand, some consumer elements may not exhibit such a temporal correlation between fluid events.

[0083] Note that it is also possible to identify similar “long” fluid events, for example, having a flow duration greater than a predefined value, such as 30 seconds, and / or similar “short” fluid events, for example, having a flow duration less than a predefined value, such as 10 seconds. It is also possible to calculate, for each of the time slots, the cumulative volume of all the fluid events identified in the time slot in question.

[0084] Thus, each of the simulated global flow rate signals SsirnBsim(m) comprises, per fluid event Ev, event parameters PEV representative of the fluid event in question (initial instant, flow duration, etc.) as well as PT temporal parameters.

[0085] Furthermore, it may be advantageous to determine, for each of the simulated global flow rate signals SsirnBsim(m), additional parameters called comparison parameters PEv|Bsim, where a flow rate (average flow rate for example) of each of the fluid events is compared to a global flow rate (average flow rate) of the simulated house Bsim(m). It is thus possible to calculate, for each fluid event, a normalized flow rate defined as being equal to the ratio of the average flow rate of the fluid event to the average global flow rate of the simulated house Bsim(m).

[0086] Finally, we generate the database BD1 as being the collection of all the fluid events of all the simulated global flow signals SsirnBsim(m), and more precisely of the event parameters PEv, the temporality parameters PT, and here the comparison parameters PEv|Bsim. For each event, we also have the corresponding class LEv.

[0087] Thus, we obtain a set of NA fluidic events Ev® with i ranging from 1 to NA, each fluidic event Evq being defined by the event parameters PEV(i), the temporality parameters PT(i) and the comparison parameters PEViBR(i), and by the corresponding class LEv(i). This set then forms the database BD1. The number NA can thus be very high, especially since it comes from the large number M of simulated houses Bsim(m).

[0088] Note here that the fluid events Ev® are mixed, in the sense that they are no longer associated with this or that simulated house Bsim(m) nor with this or that measurement day. On the other hand, temporal correlation information is present via the temporality parameters PT, which makes it possible to improve the prediction performance of the model.

[0089] Phase 200: Parameterization of the prediction model.

[0090] Phase 200 then consists of carrying out the training of the prediction model from the database BD1, that is to say of parameterizing the prediction model automatically so that from the input data which are here the parameters { PEv(i); PT(i); PEv|Bsim(i)}i=i;NA, it determines the output data which are the corresponding classes { LEv(i)}i=i;NA. Thus we note: Lev(i) = fl( PEv(i); PT(i); Pbvier® ), where fl is the parameterized prediction model.

[0091] As indicated previously, different prediction model algorithms can be used, but the inventors have found that the algorithm based on boosting type decision trees, and more precisely gradient boosting type, has good performance.

[0092] Phase 300: Classification.

[0093] Phase 300 consists of performing the prediction of the class (classification) of the fluidic events Ëv^ present in a representative measurement signal SmesBS(t) of the overall flow rate of the house to be monitored BS, this signal having been acquired by the fluidic measurement sensor CM (see fig.lA). More precisely, it is a question of estimating the class Lpy(j) of each of the fluidic events Ëv^y

[0094] Also, during a step 310, the measurement signal SmesBS(t) is acquired by the sensor of CM fluidic measurement. It corresponds to a succession of event parameters PEV(jk) representative of fluidic events Ev^k) that it contains: Smesæjt^ {P EvW ) • Event parameters ^Evfjk) are identical to the PEv event parameters from the BD1 database, and include in this example, the initial time of the fluidic event, its flow duration and the elapsed volume. The measurement signal SmesBS(t) is acquired by the fluidic measurement sensor CM over a monitoring period Ats(k), for example over one day. When the monitoring period k is over, the fluidic measurement sensor acquires a new measurement signal SmesBS(t) over the following monitoring period Ats(k+i), and so on.

[0095] Then, during a step 320, the additional parameters are determined, here the temporality parameters PTQk) and the comparison parameters P£v^jj^ky so that the measurement signal SmesBS(t) forms input data for the model of fl prediction having the same parameters (same dimensions) as during the learning phase 200.

[0096] The temporal parameters Py^ are determined in the same way as that described previously, by identifying the number of fluid events similar to that considered and located in the upstream and downstream time slots.

[0097] Where appropriate, the comparison parameters PEv|B$ jki-»k) are determined in the same way as that described previously, by calculating the normalized flow rate of each fluid event Ëv^y ie- 'c ratio of the average flow rate of each fluid event Ëv^ to 'c average overall flow rate of fluid consumption of the house to be monitored BS. Note that the average overall flow rate of the house to be monitored BS is advantageously the average overall flow rate of all the measurement signals Smes^t) having been acquired previously, from the initial acquisition (index ki). Thus, the average overall flow rate of the house to be monitored BS will be more precise as the acquisitions of the measurement signal SmesBS(t) follow one another.

[0098] Finally, during a step 330, the classes (classification) of the fluid events Ëvqjq present in a measurement signal SmesBS(t) are predicted (classified). The input data provided to the prediction model fl are PEvjBSQki-k)) ' and the output data are the estimated classes L^) for each fluid event Ey^, so that we have: LevOe) = f 1 ( { PEvak> M PEvMik^k)} ) • When the prediction is carried out for the measurement signal SmesBS(t) of the monitoring duration of index k, phase 300 is repeated for a new measurement signal SmesBS(t) acquired over the monitoring duration of index k+1.

[0099] Thus, the classification process is able to determine the element class consumer for each fluid event present in the signal of SmesBS(t) measurement representative of the overall consumption flow rate of the fluid of interest of the building to be monitored BS. Due to the fact that the input data includes the temporal parameters, it appears that the prediction model has improved accuracy.

[0100] Note that the prediction step 330 may include constraints that block the allocation of one or other of the classes to certain fluid events ÊV(^) and allocate a class called 'Other' when the estimated class is prohibited. Thus, the events of each class comply with conditions, for example, on the flow duration, the average flow rate, the number of similar successive fluid events, etc. Also, for example, the class 'Shower' could not be allocated to a fluid event whose flow duration would be less than 5 seconds or greater than 900 seconds.

[0101] Finally, note that the classification process can make it possible to detect excessive consumption of the consuming elements of one or other of the classes. predefined, and can then indicate to the user if consumer elements are overused or have a flow rate above a predefined reference.

[0102] In this respect, the method may include an additional step in which a predefined monitoring parameter (average flow rate, average flow duration, etc.) of the consumer elements of each class is compared with a predefined threshold value. Then, the class whose consumer elements have a monitoring parameter value greater than the threshold value is identified. It is then possible to identify consumer elements of the same class which have excessive usage, or a flow rate greater than a predefined threshold. The method may then include a step of alerting the user indicating the class in question, then a step of correction by the user so that the consumer elements of the class in question then have a monitoring parameter value lower than the predefined threshold value. For example, this may result in a replacement of the shower heads, or even by replacing the consuming elements in question with less consuming elements.

[0103] [Fig.5] illustrates a flowchart of a classification method according to another embodiment, which differs from the method of [Fig.2] essentially in that it uses two prediction models fl and f2 successively. The principle is to use the first prediction model fl to carry out a first level of classification, and thus deduce additional parameters, which will then be added to the input data of the second prediction model f2 which will then carry out a second level of classification. This makes it possible to further improve the performance of the classification method.

[0104] The classification method comprises a phase 100 of generating two databases BD1 and BD2, then a phase 200 of learning the two prediction models f1 and f2, and finally a phase 300 of two-level prediction.

[0105] During phase 100 of generation of the databases BD1 and BD2, steps 110 to 140 are identical or similar to those described previously and are not described again here.

[0106] Step 150 consists of determining (sub-step 151) the event parameters PEv, the temporality parameters PT (and here the comparison parameters PEviBsim) to obtain the first database BD1. Here, we operate in an identical or similar manner to the method of [Fig.2]. The database BD1 therefore includes these parameters PEv, PT and here PEv|Bsim, as well as the LEV classes assigned to each fluid event Ev.

[0107] Step 150 also consists of determining (sub-step 152) additional parameters PEv|L called 'comparison by class' (whereas the parameters PEv|Bsim are comparison parameters with respect to the simulated house Bsim), and then obtaining the second database BD2. Thus, the database BD2 includes the event parameters PEv, the temporality parameters PT, the comparison parameters PEv|Bsim, and the comparison parameters by class PEv,L. It also includes the classes LEv assigned to each fluidic event Ev.

[0108] Thus, for each of the simulated global flow signals SsirnBsim(m), the additional parameters PEv|L are determined, where the average flow rate of each of the fluid events is compared to the average global flow rate of the fluid events of each of the classes. It is thus possible to calculate, for each fluid event, the normalized class flow rate defined as being equal to the ratio of the average flow rate of the fluid event to the average global flow rate of the fluid events of the same class, for the simulated house considered.

[0109] Then, during phase 200, the two prediction models fl and f2 are parameterized by supervised machine learning. More precisely, during step 201, the

[0110] [YES]

[0112]

[0113]

[0114] prediction model fl is parameterized from database BD1, and in step 202, prediction model f2 is parameterized from database BD2. Finally, during phase 300, the steps of acquisition 310 and preprocessing 320 of the measurement signal SmesBS(t) are carried out in a manner identical or similar to the steps described in connection with [Fig.2]. Then, during a step 331, a first level of prediction of the fluidic events Ëv^) present in a measurement signal SmesBS(t) is carried out, using the first model fl. The input data provided to the model fl are {PEvW P^ PEv|BS( jki~*k)} ' and the output data are the classes summed Mjk) for each fluidic event Ëv^y so that we have: The v(jk) - f 1 ( { PEv(jl> PT(j,k> PEv|BSCiki-*k)} ) ' Then, during a step 332, the comparison parameters per class PEv]L(ik)' are determined. This step is possible to the extent that the previous fluid events Ë^j^i-^ p have been classified during the previous iterations, either by the model fl or by the model f2. It is then possible to determine the characteristic flow rate (e.g. average, median flow rate, etc.) of each class of consumer elements of the house to be monitored BS, and then to compare the determined flow rate of the fluid event to be classified with the characteristic value of each class (ratio of the average flow rate of the fluid event considered to the average (or median) flow rate of all the fluid events that fall within the class considered). And so on for all the classes. Of course, it is possible to determine the comparison parameters per class Pp^j^j for each of the classes, or for only one or the other of them.We could therefore limit ourselves to considering only the class of flush toilets. Finally, during step 333, the second level of prediction of the fluidic events Ëv^) present in a measurement signal SmesBS(t) is carried out using the second model f2. The input data provided to the model f2 are {PEv(ik> ; PEv|L(w} 'and the output data are the classes estimated for each fluid event Ëv^y so that we have: LEMjk) = f2( {PEvük); P^ PEvMpâ-k): PeW)1 ) •11 it appears that the classification method according to this embodiment has improved performance. In this respect, Figure 6A, Figure 6B, and Figure 6C illustrate changes in the distribution function (also called the cumulative distribution function, or Cu mulative Distribution Function in English) of the flow rate (fig. 6A), the flow duration (fig. 6B) and the elapsed volume (fig. 6C) for the fluid events resulting from the simulated global flow signal SsimBSim whose statistical consumption profile Pstat^"1 is identical to the profile Pstat^ of a real reference house. These curves are compared to those resulting from the real house considered.

[0115] Note here that the flow distribution function CDFD(d) corresponds to the probability of obtaining a value of the average flow less than or equal to d: CDFD(d) = P(D <d). cette probabilité p(d<d) est égale à la proportion %ats,d de durée d’écoulement cumulée atec|d au cours surveillance ats (ici journée) pour laquelle le débit moyen d non nul et plus égal d. la fonction répartition cdfd(d) comprise entre 0% 100%, varie zéro maximal mesuré par les signaux mesure dmesec(t).

[0116] Thus, it appears that the simulated signals DsimEC(t), generated during step 130 described previously, have a very good correspondence with the measurement signals DmesEC(t). As a result, the simulated overall flow rate signal Ssim(t), generated from the simulated signals DsimEC(t), is particularly realistic. This is therefore also the case for the database.

[0117] Finally, [Fig.7A], [Fig.7B], and [Fig.7C] illustrate the prediction performances for different classification methods, and here respectively the precision, the recall and the Fl-score. The precision is defined as being equal to the ratio of the number of fluid events correctly classified in a given class to the total number of fluid events classified in the class in question. The recall is defined as being equal to the ratio of the number of fluid events correctly classified in a given class to the total number of fluid events actually belonging to the class in question. Finally, the Fl-score corresponds to the harmonic mean of the precision and the recall.

[0118] Here we consider the results of a model fO which would have been parameterized by the PEV event parameters alone and not also by the PT temporality parameters nor by the PEv|Bsim- parameters. We also consider the results of the model fl described with reference to [Fig.2], and those of the two-level model fl and f2 described with reference to [Fig.5]. The classes considered here are those of the flush toilets C, that of the showers D, that of the washing machines LL, those of the dishwashers LV, and finally that of the consumer elements R* which it is not possible to distinguish from the others (here essentially the taps R and the 'mixed' class M with other consumer elements). It appears that the performances of the models fl and fl+f2 are indeed better than those of the model fO.

[0119] Particular embodiments have just been described. Different variants and modifications will be apparent to those skilled in the art.

Claims

1. Claims Method for computer classification of fluidic events Ëv^ of consumption of a fluid of interest by consumer elements EC of a building to be monitored BS, among several predefined classes Lgv(jk) of consumer elements EC, the method comprising the following phases: a generation phase (100) of at least one database (BD1, BD2), comprising the following steps: • acquire (110) measurement signals | Dmesæf 'n! ( t ) I representative of the flow rate of 1 ' 1 n=IiN consumption of each of the EC consumer elements of N real reference buildings {Bref riJ(ri-i ;N, with N>1, the measurement signals then being said to be classified, the EC consumer elements being of the same classes as those of the building to be monitored BS; • define (120) consumption profiles (pstat88^"0) of M so-called simulated buildings i ' EC WhM {Bsim(m)}m=i ;M, with M>N, from the acquired classified measurement signals, each simulated building comprising EC consumer elements of the same classes as those of the building to be monitored BS; • generate (130), by digital simulation, simulated signals | t ) I representative of the l ' m= IM consumption flow rate of each of the EC consumer elements of the simulated buildings {Bsim(m)} m=i;M having the defined consumption profiles, the generated simulated signals then being said to be classified; • determine (140), for each of the fluidic events {Ev(i)}i=i;NA identified in the generated classified simulated signals: • so-called event parameters {PEv(i)}i=1;NA, representative of a duration and a volume elapsed time of each of the fluid events; and • so-called temporal parameters {PT(i)}i=1;NA, representative of a number of so-called similar fluidic events, presenting a flow rate substantially identical to that of the fluidic event considered and located in predefined successive time slots located before and after the fluidic event considered; • generate (150) the database (BD1; BD2) comprising, for each of the fluidic events {Ev(i)}i=i;NA identified: the event parameters {PEv(i)}i=i;na, the temporality parameters {PT(i)}i=1;Na, and the corresponding class {LEV(î)}i=i;na of consumer element EC; a parameterization phase (200) of at least one prediction model (fl, f2) by supervised automatic learning from the database (BD1; BD2); a classification phase (300), comprising the following steps: • acquire (310) a measurement signal SmesBS(t) representative of the consumption flow rate of the building to be monitored (BS); • determine (320), for each of the fluidic events {ËVf^^i ;np identified in the acquired measurement signal, event parameters {Pgy^)} j=i ;NP and corresponding temporality parameters {Py^Jj^i ;np; • predict (330), by the prediction model (fl; f2), for each of the fluidic events {ËV( jjc)}j=i;NP whose event parameters {PEyQj.pH;np and temporality parameters {P^j^)}j=i;np form input data to the prediction model, the corresponding class {LEy(^}j=i;np of consumer element EC.

2. Classification method according to claim 1, in which the temporal parameters PT(i) and also comprise, for each of the time slots, the cumulative volume of the fluid of interest flowed and the flow duration.

3. Classification method according to claim 1 or 2, in which the event parameters PEv(i) and comprise, for each fluid event, the initial instant of the fluid event, the flow duration, the elapsed volume and / or the average flow rate.

4. Classification method according to any one of claims 1 to 3, in which the database and the input data of the prediction model comprise so-called comparison parameters, denoted respectively PEv|Bsim(i) and PEvjBSüki^k))' and defined, for each fluidic event Ev(i), Ëv^y as the ratio of an average flow rate of the fluidic event considered to an average overall flow rate of all the fluidic events Ev®, ÊV(y,) of the building considered.

5. Classification method according to any one of claims 1 to 4, in which the classification phase (300) is repeated, the measurement signals SmesBS(t) of the building to be monitored (BS) then being acquired over successive predefined monitoring durations.

6. Classification method according to claims 4 and 5, wherein, in the comparison parameters P^-|bs( present in the input data of the prediction model, the average global flow rate of all fluid events Ëv^) takes into account the average global flow rate of the previous iterations.

7. Classification method according to any one of claims 1 to 6, in which: • the generation phase (100) comprises the following steps: • determination (151) of a first database (BD1) where each fluidic event (Ev(i)) is defined by at least said event parameters (Pev®) and by said temporality parameters (PT(o); • determination (152) of a second database (BD2) where each fluidic event (Ev®) is defined by at least said event parameters (PEV®), by said temporality parameters (PT®), and by so-called comparison parameters by class (P evil®) defined, for each fluidic event (Ev®) of at least one class considered, as the ratio of an average flow rate of the fluidic event considered to an average overall flow rate of all the fluidic events of the class considered (Ev®); the parametric phase (200) comprises the following steps: • parameterization (201) of a first prediction mode (fl) by supervised machine learning from the first database (BD1); • parameterization (202) of a second prediction mode (f2) by supervised automatic learning from the second database (BD2); the classification phase (300) includes the following steps: determination (320), for fluidic events (ËV(ÿ-)) present in the measurement signal SmesBS(t), event parameters (Pe^-)) and pa temporality meters ) correspondents; then prediction (331) by the first prediction model (fl), for each of the fluid events (Ey^), which are defined by the event parameters ( PEv(j,k) ) and the temporal parameters form input data of the first prediction model (fl), of the element class (Lg^)) corresponding consumer (EC); then determination, from the previously predicted classes, of comparison parameters per class (P^m^)) corresponding; then prediction (331) by the second model of prediction (f2), for each of the fluidic events (Ëv<-i®X which are defined by the pa- event meters (PgvQk'p' 'cs parameters of temporality (pT .,3, and the comparison parameters by class (PeW)) which form data input of the second prediction model (f2), of the corresponding consumer element (CE) class (Lg^p).

8. Classification method according to any one of claims 1 to 7, wherein, during the step of acquiring (110) the measurement signals | DmeSpæt(n) ( t ) | , the N real reference buildings are l J Jn=kN connected to a device for determining at least one prediction model, comprising fluid measurement sensors (CM) adapted to acquire the measurement signals for each of the consumer elements (EC) of the real reference building considered (BR), and a computer connected to the fluid measurement sensors of the reference building considered (BR) and adapted to carry out the parameterization phase (200).

9. Classification method according to any one of claims 1 to 8, in which, during the step of acquiring (310) the measurement signal SmesBS(t) of the building to be monitored (BS), the latter is connected to a classification device comprising a fluid measurement sensor adapted to acquire the measurement signal §mesBS(t), and a computer connected to the fluid measurement sensor of the building to be monitored (BS) and adapted to carry out the classification phase (300).

10. Classification method according to any one of claims 1 to 9, comprising, following the classification phase (300), the following steps: • identification of a class of consumer elements for which the fluid events have a value of a so-called tracking parameter greater than a predefined threshold value; then • alerting the user indicating the class of consumer elements in question; then • correction by the user so that the consumer elements of the class in question then have a value of the tracking parameter less than the predefined threshold value.