method for determining data representative of the risk of occurrence of a clogging event in a pumping station, device and corresponding program.

Abstract

The disclosure relates to a device and a method for determining data representing the risk of a pumping station clogging event at an industrial site, which includes the following steps: obtaining (S01) at least one time series Q of current environmental conditions (DRCEQ); determining (S02), within historical data in the form of at least one past time series X of environmental conditions (DRCEX), a correspondence between a variable in time series Q and a variable in said at least one time series X; when at least a partial correspondence is determined, obtaining (S03), using a time series Y of past pumping station clogging events, the risk data for the occurrence of a pumping station clogging event. Fig. 1

Description

Title of the invention: Method for determining data representative of the risk of a clogging event occurring in a pumping station, corresponding device and program. Scope of the invention

[0001] The invention lies in the field of prevention and mitigation of the risks of clogging affecting water pumping stations used for cooling industrial installations. Previous Art

[0002] Pumping stations are used in various industrial and environmental fields. They play important roles in diverse industrial sectors, such as nuclear power plants, where they make it possible to obtain large quantities of water used to cool installations, for example.

[0003] Pumping stations face various random situations that can affect their operation. One of the main threats is the massive influx of clogging materials (MCM). These sudden events, in which a large quantity of clogging materials (plant debris, algae, fish, etc.) is carried by the water, can obstruct pumping and filtration systems. These MCMs are often caused by natural phenomena such as tides, floods, or storms.

[0004] In addition, conditions such as stream flow, water level, and wind speed and direction can vary unpredictably. These fluctuations can lead to changes in the quantity and nature of clogging agents present in the water, thereby increasing the risk of clogging. Extreme weather conditions also pose a threat. Storms, heavy rainfall, floods, and droughts can affect the amount of debris carried by the water. These extreme weather events increase the risk of clogging by introducing additional debris into pumping systems. Seasonal changes also influence the amount of debris in the water. For example, spring floods can carry plant debris that has accumulated during the winter.Conversely, periods of low flow in summer allow debris to accumulate in waterways, thus increasing the risk of clogging during these low-flow episodes. Human activities also contribute to the presence of clogging agents in the water. Industrial discharges, construction work, and agriculture can introduce debris and sediment into the waterways. Watercourses. Finally, biological phenomena can lead to clogging events. The proliferation of certain species, such as algae or jellyfish, can be influenced by environmental factors such as water temperature and nutrient availability. These biological phenomena increase the risk of clogging and require constant monitoring.

[0005] The current situation for detecting and / or managing clogging problems relies primarily on business rules and human expertise. For example, rules such as "if the river flow rate exceeds X m / s, then the period is at risk" are applied by teams responsible for managing pumping stations. This approach can be supplemented by forecasts of certain important variables (such as future flow rate or wind speed) provided by specialized entities such as Météo-France or EDF DTG. However, no system for predicting the arrival of a massive clogging event exists. A technique is well described in patent application WO2017186603, but it is limited to a diagnosis with a delay of 6 hours before the feared event and is complex to implement.

[0006] There are also devices for measuring the presence of clogging agents, such as buoys for detecting "seaweed" (a collection of plant debris) or the use of drones to detect swarms of jellyfish. However, these devices do not provide predictive capabilities as such, but only spot measurements of the presence of clogging agents. Devices for cleaning and improving the filtration components of pumping stations also exist, but they do not allow for anticipating sudden or large-scale clogging events.

[0007] It is therefore necessary to have a solution which makes it possible to resolve at least partially this lack of medium- to long-term forecasting of future clogging situations within pumping stations. Summary of the invention

[0008] To overcome at least some of the drawbacks of the prior art, the invention combines data engineering methods, statistical learning algorithms and multivariate time series processing techniques to anticipate clogging events and provide predictive diagnostics.

[0009] More specifically, the disclosure relates to a method for determining data representing a risk of a clogging event occurring at a pumping station on an industrial site. Such a method is implemented using an electronic device comprising a memory and a processing unit. This method comprises at least one iteration of the following steps: - obtaining at least one time series Q of current environmental conditions of the pumping station; - determination, within historical data presented in the form of at least one time series X of past environmental conditions, of at least a partial correspondence between at least one variable of the time series Q and a corresponding variable of said at least one time series X; - when it is determined that at least a partial correspondence exists between said at least one variable of the time series Q and said at least one corresponding variable of said at least one time series X, obtaining, using a time series Y of past clogging events of the pumping station, the risk data for the occurrence of a clogging event of the pumping station.

[0010] Thus, this approach makes it possible to provide predictive diagnoses based on historical events, thereby giving pumping station managers the opportunity to take preventive measures before clogging occurs. This improves the reliability and efficiency of pumping operations, reduces unplanned downtime, and minimizes the costs associated with emergency interventions and cleanups.

[0011] According to a particular feature, said at least one time series Q of current environmental conditions and said at least one time series Q of past environmental conditions each comprise at least one time series relating to the external environment QENy> ^ENV industrial site and at least one time series relating to the internal environment QTRA, Xtra of the pumping station and / or the industrial site.

[0012] Thus, the method makes it possible to capture a comprehensive picture of the factors influencing the risk of clogging. This multivariate approach allows for better identification of the interactions between external and internal conditions, thereby improving the accuracy of predictive diagnoses.

[0013] According to a particular feature, said at least one time series relating to the internal environment QTRA, %TRA of the pumping station includes variables that are independent of decisions and / or operating rules of the industrial site.

[0014] Thus, the process avoids biases introduced by human interventions or operational modifications.

[0015] According to a particular feature, the step of obtaining the time series Q of current environmental conditions includes at least one step of completing missing values ​​for at least some variables of the time series Q of current environmental conditions.

[0016] Thus, it is possible to provide a continuous character to data which may be in a discrete form.

[0017] According to a particular feature, the determination step comprises: - a first search step, based on a first current time series QEnv comprising at least one variable relating to the external environment of the industrial site, for at least a partial match with a time window W ENVi extracted from a first time series XEnv comprising at least one corresponding variable, the size of the time window Wen Vi being at most equal to the size of the first time series Q^y

[0018] Thus, it is possible to search more easily in the first time series XEnv • 'the quantity of data contained in this series is thus processed in successive windows which make it easier to search for a match.

[0019] According to a particular feature, when at least a partial match is found during the first search step, a second search step, based on a second current time series QTRA comprising at least one variable relating to the internal environment of the industrial site, finds at least a partial match with a time window IVTHAi extracted from a second time series comprising at least one corresponding variable, the size of the time window IVerai being at most equal to the size of the second time series QTRA-

[0020] According to a particular feature, the first step of searching for at least a partial match is implemented for a set of time windows. This allows for more efficient data processing, while also enabling the application of computational optimizations that facilitate obtaining a profile. This methodology is obviously applicable to the second Xtra-' time series, which thus produces the same effects.In a complementary implementation example, rather than first performing a search on the first time series XEnV, it is entirely conceivable and usable to perform the first search from the second time series Xra. In other implementation examples, a single search is performed on the time series X which includes both the first time series XenV and the second time series Xra, the set of variables composing these series then being combined into a single series. This is also applicable for Q (combining Q in v and ). Qtra^' as indicated in relation to the general presentation made previously, the set of these series being multivariate, for example.

[0021] According to a particular feature, the at least partial correspondence between two time series is obtained by calculating a minimum distance separating the values ​​of the different variables that correspond within the two time series.

[0022] In another aspect, the invention also relates to a device for determining a risk factor for the occurrence of a clogging event in a pumping station. This device comprises a memory and a processing unit. The processing unit is configured to execute at least one iteration of the following steps: - obtaining at least one multivariate time series of current environmental conditions of the pumping station; - determination, within historical data presented in the form of at least one multivariate time series of past environmental conditions, of at least a partial correspondence between the evolution of at least one variable of the multivariate time series and a corresponding variable of said at least one time series; - when it is determined that at least a partial correspondence exists between the evolution of said at least one variable of the multivariate time series and of said at least one corresponding variable of said at least one time series, obtaining, using a time series of past clogging events of the pumping station, the risk data of occurrence of a clogging event of the pumping station.

[0023] According to another aspect, the invention also relates to a computer program capable of implementing the described process and to a data carrier for recording this computer program.

[0024] The electronic device for determining data representative of the risk of a pumping station clogging event has the architecture of a computer. It is equipped with one or more processors capable of executing all types of computer programs, from operating systems to application software, written in compiled or interpreted languages. The various components of the electronic device for determining data representative of the risk of a pumping station clogging event are interconnected by a communication bus. The electronic device for determining data representative of the risk of a pumping station clogging event may optionally be Equipped with a communication system to communicate via protocols such as Bluetooth, Ethernet, or WiFi with other systems and connect to mobile or fixed telecommunications networks. The electronic device for determining data representative of the risk of a pumping station clogging event also includes memory components that record the data and programs necessary for the device's operation.The electronic device for determining data representative of the risk of a pumping station clogging event is further modified so that it can perform the operations of obtaining current environmental data via sensors, organized into multivariate time series; searching for partial matches between this current data and historical data also in the form of time series; and when a partial match is found, using time series representing past clogging events to determine the risk of a pumping station clogging event.

[0025] Data carriers can be any entity or device capable of storing programs. For example, the carriers can include a storage means, such as a ROM, for example a CD-ROM or a microelectronic circuit ROM, or a magnetic recording means such as a hard drive, or more commonly, flash memory. Alternatively, the carriers can be transmissible media such as an electrical or optical signal, which can be transmitted via an electrical or optical cable, by radio, or by other means. The programs according to the invention can, in particular, be downloaded from a network such as the Internet. Alternatively, the information carrier can be an integrated circuit in which the program is incorporated, the circuit being adapted to execute or to be used in the execution of the process in question. Brief description of the figures

[0026] Other features and advantages of the invention will become more apparent from the following description of a particular embodiment, given by way of simple illustrative and non-limiting example, and the accompanying drawings, among which:

[0027] - [Fig.1] represents the main steps of the disclosure process;

[0028] - [Fig.2] represents the implementation of a distance profile calculation for a example of implementation;

[0029] - [Fig.3] illustrates a device for implementing the disclosure process.

[0030] Description of an embodiment

[0031] As previously stated, one purpose of this disclosure is to provide a mechanism for determining the risk of an event occurring Clogging of a pumping station. The study of the clogging phenomenon has shown that multiple timescales exist. Periods of high water increase the amount of debris in circulation, as plant debris is carried into the watercourse. During periods of low flow, this stock of material settles to the bottom of the watercourse. During periods of high flow, some of this stock is re-moved, which, during low tides for example, leads to a high concentration of debris in the water column, increasing the risk of clogging. The combination of these factors results in a lag between the (multiple) causes and the effect (i.e., clogging). The difficulty in predicting such effects can be compounded by the fact that measurements can be taken not only at different times but also at different locations (water flow sensors may be located far from the pumping station itself, for example).

[0032] The inventors have developed a method for searching for similarities in past environmental conditions (which have been the subject of previous measurements), these past environmental conditions having led (or not) to subsequent clogging situations (which have also been documented, i.e., recorded). One of the inventors' motivations is that experts in the field suggest that the clogging phenomenon is influenced by the accumulation of sedimentary deposits forming in certain locations during periods of insufficient water flow (i.e., without flooding in winter / spring). These stored deposits can be released in the event of flooding or high tides the following year, for example.

[0033] This method is implemented iteratively, using current environmental conditions data (CECD), which are obtained from sensors located within, around, and / or at a distance from one or more pumping stations. Ideally, the current environmental conditions data (CECD) include both environmental data around the pumping station and operating data from the pumping station, such as data from internal water circulation circuit sensors or water pressure in filtration or flow devices. The current environmental conditions data (CECDq) are organized as multivariate time series (Q), also called queries.The variables in these multivariate time series represent the evolution of values ​​measured by the different sensors of environmental conditions (precipitation, water level, wind, current strength around the pumping station and water pressure, water level, opacity, within components of the pumping station, etc.).

[0034] One or more (potentially partial) correspondence searches between this current data (Q) and past historical data (DRCEX), also in the form of multivariate time series (X), are carried out, in particular by calculating minimizations of the similarity functions, as explained later. Of course, the variables of current environmental conditions (DRCEq) and historical data (DRCEX) relate at least to identical or at least similar environmental data.

[0035] When a match (partial or complete) is found, a risk data point for the occurrence of a clogging event can be searched for and / or obtained using time series (Y) that include past clogging (or non-clogging) events. In other words, this method makes it possible to anticipate clogging risks by comparing current environmental conditions Q (current) with similar situations from the past X, based on a multivariable approach. Thus, by finding similar environmental conditions in the past, and similar operating conditions (of the pumping station) during these past periods, the disclosure technique makes it possible to provide the pumping station operator with a limited set of known past situations, these situations being as close as possible to the period of interest in terms of medium / long-term clogging risk.

[0036] The iterative process developed by the inventors thus consists of searching, within historical data, for data close to or similar to that of the query. The query itself can include data spanning a longer or shorter period, depending in particular on the operational implementation conditions. For example, the query data can cover a period of one week, one month, one year, or even several years. Among the events that can influence the size of the query Q are, for example, the date of the last cleaning of the pumping station or other events affecting both the pumping station itself and the surrounding watercourses, for example.

[0037] In any event, starting from the query data (which can therefore be very numerous and voluminous), it is necessary to search for similar data in the historical data (X). This search can be performed using a sliding window (VE), for example, where the size of the window (VE) in the historical data (X) is equal to the size of the query (or a fraction thereof). This search can also be performed using a sliding time window, meaning that it is not the quantity of query data (Q) that is used, but the time period covered by the query: such a possibility implies, according to the present document, that a preliminary data transformation process must take place to adapt the input data (X, Q, Y) to this search strategy. Other search possibilities are conceivable, as detailed below.

[0038] If no similarity is found, the window (Wj) is shifted (shifted by a time step or shifted to the next values ​​of the historical data X) and the search is performed on the next portion (Wj) of historical data. If similarity is found, a process involving the search for events within the series of events (X) associated with this next portion (Wj+i) of historical data (X) is carried out: the presence or absence of a pumping station clogging event possibly associated with the portion of historical data is searched for, and so on, either until all the historical data has been reached, or according to a predetermined stopping parameter. [Fig. 1] illustrates the different steps of the process implemented.

[0039] More specifically, the process comprises at least one iteration: - a prior acquisition step (SOI), from at least one sensor (CAP0,... CAPZ), of representative data of current environmental conditions (DRCEq) relating to the environment of the pumping station, these environmental condition data (DRCEq) being organized in the form of at least one multivariate time series (Q); - at least one search step (S02), within historical data presented in the form of at least one multivariate time series (X) representative of past environmental conditions (DRCEX), of at least a partial correspondence between the evolution of at least one variable of the multivariate time series (Q) and a corresponding variable of said at least one time series (X); - when it is determined that at least a partial correspondence exists between the evolution of said at least one variable of the multivariate time series (Q) and of said at least one corresponding variable of said at least one time series (X), obtaining (S03), using a time series (X) representative of past clogging events of the pumping station, the risk data of occurrence of a clogging event (DRSEC) of the pumping station.

[0040] The determination is therefore carried out in two stages. Depending on the implementation, it is also possible, within the search stage (S02), to divide this search (S02) into two separate searches, particularly to limit the calculations to be performed. Various other computational optimizations are proposed by the inventors so that this method can be implemented more simply, more quickly, and with less energy consumption.

[0041] Thus, the proposed method allows the pumping station manager (and beyond, the production site manager) to provide a sufficiently early view of a Probable clogging event – ​​anticipated in the sense that the operator has time to consider ways to limit the consequences of the probable event, for example by temporarily reducing pump power during the risk period without stopping production, or possibly switching to another cooling source. This is achieved by using regularly collected data describing the pumping station's environment (weather data, river flow rate, tidal coefficient, etc.).Using internal data related to the operation of the pumping station (pressure differentials and / or water level differentials in circuits supplying filter drums or in recirculation circuits, for example, this data providing direct or indirect indications of the congestion of the pumping station's internal components and therefore its capacity to efficiently pump the water volumes necessary for effective cooling), the proposed method makes it possible to produce predictive diagnoses, from the Y series, over a given future period (future within the correspondence found in X with respect to Q) (for example, a full week). Thus, the proposed method produces an AMC susceptibility diagnosis through pure statistical learning, from multivariate time series.

[0042] The search step (S02) itself may be preceded by a possible division of the historical and current data into two distinct types of data: historical and current data relating to the external environment of the pumping station (Qv and Xenv^) and historical and current data relating to the internal environment of the pumping station (QTRA and XyRA). This division may pre-exist and is therefore optional.

[0043] Assuming that such a partitioning is carried out, the search step (S02) includes: - the determination of a current window (W ENVi) from the historical data (XEnv^ the size of the current window (W ENVi) is at most equal to the size of the query (Qp^y); - the calculation of a current distance (d^NVi^ cntrc 'cs data from the query variables (Q) and the corresponding variable data from the JC 1 yv current window (W ENVi);

[0044] When the current distance (d^^yî) between 'cs data of the query variables ^ENV^ Ct 'CS data of the corresponding variables of the window (IVENVi^) is greater than a predetermined threshold provided as a parameter, the next window ENVi+1^ is accessed

[0045] When the current distance (dENyj) is less than the predetermined threshold, it is then determined whether the internal environmental conditions of the pumping station (QTra ct XTRA>) are also similar. To do this, this second search includes: - the determination of a current window (W^r^) from the historical data (X jj^). the size of the current window (VVyj^y) is at most equal to the size of the query (QTR^); - the calculation of a current distance (dRR^}) between the data of the query variables (QTRA) and the data of the corresponding variables of the current window (Wrraj);

[0046] When the current distance (d-^R^j) is greater than a predetermined threshold (which is also a parameter of the process), then we move to the next external environmental conditions window (W envî+1)-

[0047] When the current distance (dj-R^j) is less than the predetermined threshold, the determination of a possible occurrence of a clogging situation is sought in the time series (Y) representing past clogging events of the pumping station. In its simplest form, this time series (Y) of clogging events explicitly indicates the dates on which these events occurred, and it is a binary series: 1s represent the days (or hours) corresponding to cloggings, while 0s represent the days (or hours) corresponding to the absence of cloggings. Other types of series Y can be considered, such as series including probabilities of clogging occurrence or other data allowing clogging prediction. In an example embodiment, a future period (future with respect to the window (Wtra^) in the time series (Y) of clogging events) is determined using a parameter of the method.For example, when the window (Wtrai) spans a period of 1 month, we will look for what happens one week after that month in the time series (Y). Thus, the risk data for the occurrence of a pumping station clogging event can take several forms depending on the implementation.

[0048] It is understood that the proposed division between external environmental data, internal environmental data, and clogging event data is a specific, non-limiting implementation intended to facilitate the explanations provided. It is entirely possible that a single multivariate time series of historical data X may be used and that, within the framework of the steps presented above, only certain variables may be used to perform the calculations.

[0049]

[0050]

[0051]

[0052]

[0053]

[0054]

[0055]

[0056]

[0057]

[0058]

[0059]

[0060]

[0061]

[0062]

[0063]

[0064]

[0065]

[0066]

[0067]

[0068]

[0069]

[0070]

[0071]

[0072]

[0073]

[0074]

[0075]

[0076]

[0077] The following pseudocode illustrates the proposed methodology in an example implementation. This pseudocode performs a calculation for four pumping stations (TRAlà4), which explains the division into four series. -♦Entry**: **Exit**: y GP 1. Requirements: - 1 S le.nv + 1113Xi t&nv, i - m > 1 2. Extract the subsets: - Xnv, Atra «- extract_subset ■ Xenv, Xtral, Xtra 2, X tra 3, X tra 4 extract _subsets(x) ~ Q&nv, Qtral, Qtra2, Qtra 3, Qtra 4 OXtC aCt SUbSetd^ 3. Initialize Penv <- zeros_array(in-l+1) 4. For each i G [1, e]: - Pi * P^Qenv, Xenv, , lenv, tenv, i, Aenv) XX ~ Penv Penv + Pi 5. Extract the regions below the threshold: - Renv * extract regions_below(Penv, min^Pem} xr) 6. Initialize the arrays: - y <- zeros_array(4) - distances <- zeros_array(4) 7. For each slice i G [1,4]: - Initialize Ptra infnty_array(4, m-1+1) - For each slice j G [1,4]: - For each region, region G Renv: - Ptn[j, region] «- P(Qtrai, Xtraj, region, hra, 0, - Extract Rira <- extract_regions_bellow(Ptrm, mii^PtrJ) xt) - y[i] 1 / || SjII^II max(y[.Rtr4j]]) - distance^!] <- mii^PIrai) 8. Select the strategy: - If strategy == 'best', return y[argmir(distaaces)]

[0078]

[0079]

[0080]

[0081]

[0082]

[0083]

[0084]

[0085]

[0086] - If strategy == 'max', return max(y) - If strategy = = 'average', return meaiiy} The notation used in this pseudocode example is as follows. The algorithm takes as input one (or more) historical data elements g, where m is the size of the element and d is the number of sensors. As mentioned earlier, it also uses a history of the clogging risk score YG Rm, which can be constructed from the history of clogging events. Next, a query qg is considered, where 1 is the size of the query (with 1 < m), representing the data from which the possible future occurrence of a clog is to be estimated. In the most common case, Q contains the last 1 data points emitted by the system (the system viewed as a set of sensors), in order to provide an assessment at the present time. The query Q typically contains data from the previous N years (for example, 1 or 2 years) to account for phenomena that are sometimes slow and sometimes rapid in their dynamics. In other situations, which might be related to different geographical areas, for example, the query Q may include data spanning shorter periods (from a few months to a few days). To account for these phenomena, as explained previously, the proposed methodology is broken down into two steps: 1. Calculating the environmental profile between QE^y and ^ENV' cIu' aims to find situations similar to Q in X in terms of environmental conditions (wind, water height, flow rate, etc.). 2. the calculation of tranche profiles between QTRA and which seeks to identify, within the similar environmental situations of step 1, similar situations in terms of internal operating conditions of the pumping station. A very simplified example of calculating a profile between and qg is illustrated by Figure 2. For easier understanding of the example in Figure 2, m = 10, 7 = 3 and d = 1. This is a definite operation such that: [Math.l] = ...,p ,, J With knowing a dissimilarity function between two vectors (e.g., the dissimilarity function is a Euclidean distance) and PV=[x- +J i] R^' 'a subsequence of size 1 starting at index 1 in X. A situation W, from the history, is considered similar to Q if p.< T.minP^Q, xj' avec T is a tolerance parameter.

[0087]

[0088]

[0089]

[0090]

[0091]

[0092]

[0093]

[0094]

[0095]

[0096] We can also calculate a distance profile with a set of subqueries defined by time constraints [tp te] and a length 1' < 1 with the constraint such that 1' + maXjtj < 1. In this configuration, we define: [Math.2] With : [Math.3] Pij=4Q'lJ ' ') = [xi+t, ....Xj+t.+r-!]e.R''d 0^ = (^,...,^)6^ In view of these explanations, in the example in Figure 2, the fourth window (W4) is the one that presents the smallest distance from the query Q. It is also possible to assign a weight of 2 to each subquery to establish its importance in the similarity search, such as: [Math.4] Pr%PtJÀJ These profile calculations are applicable, as explained above, to the series ^TRA' ^ENV' ^TRA' ^ENV l°rsQue ceüe methodology for calculating profiles is preferred. Of course, the example in [Fig.2] is not representative of reality. On the one hand, the length of the series is much longer and can represent several thousand values, both for current environmental series and for historical environmental series, and this across several dimensions (multivariate). Processing such quantities of data is therefore not considered feasible mentally or manually, notably because of the length of the calculations required, which would result in such a long computation time that the result obtained would probably occur after the clogging event, which is unthinkable, and on the other hand because of the risk of errors.It must be noted, however, that given the length of these series, even the computerized processing of the calculations presented above may require optimization, both to increase the speed of the calculations and to reduce the resources allocated to them. For these reasons, as previously stated, . Computational optimizations can be implemented. Thus, to increase the efficiency of the proposed method, the following optimizations have been successfully implemented: - Sliding calculation of the mean and standard deviation: instead of recalculating the mean and standard deviation for each subsequence (W), sliding sums are used. This reduces the time complexity from to ° where m is the size of the time series and l is the length of the subsequence (W);

[0097] More particularly, given a time series Xm} of size Given m and a parameter of length \, we wish to extract the mean and standard deviation of all subsequences (yn-f+1) such that = To avoid having To calculate Pf &i independently for each W j, the inventors propose to keep two sliding sums to calculate U? such that:

[0098] [Math.5] x

[0099] So

[0100] [Math.6] Xi = 2M + wi+1A-X] = X^ +

[0101] And

[0102] [Math.7]

[0103] This avoids recalculating the sums Xi and ^ for each W because they share the values ​​ / -1 with the sums of thus reducing the time complexity very significantly. - Euclidean distance optimization: The Euclidean distance is reformulated to avoid recalculating redundant operations. The sum of squares of the query elements (Q) is calculated only once. The sum of squares of the sliding window elements (W) is calculated using a sliding sum. Cross-correlation is calculated using the convolution theorem, resulting in significant performance gains for large queries and sliding windows. More specifically, the Euclidean (squared) distance can be expressed as follows:

[0104]

[0105]

[0106]

[0107]

[0108]

[0109]

[0110] [YES]

[0112] [Math. 8] dist(Q, W) = ^(<7 / wj)2 = l1J=1(QJ)2+l1j=1( In the context of a distance profile P(Q, X) between a series X and a query Q, this way of expressing the operation has certain advantages that make it possible to avoid recalculating many operations when calculating distances with sliding windows (for example dist^Q, Wj and dist(Q, IVi+])): / rr 12 should only be calculated once for any dist(Q ^=1^) sL( wj) 2 can be calculated using a sliding sum, yl q . is an element of the cross correlation Q*X. Using the theorem By convolution, we can calculate the cross correlation in a manner equivalent to a convolution in the frequency domain with a point-by-point multiplication, allowing a significant performance gain on large time series. - Cross-correlation for sliding queries: When calculating distance profiles for sliding queries, the cross-correlation is updated linearly from the previous correlation. This avoids recalculating the entire correlation for each new query. Indeed, considering two series X, Y of size in. When it is decided to calculate the distance profiles P X, Q, PX, j, ... with Qi = [^ï' of which we use a sliding window on Y to obtain the queries Q.. It is then possible to avoid recalculating the cross correlation Q^*X for each new query; Considering the following cross-correlation dn r. l ^i-1 tGl' calculated for query Q, the following cross-correlation for query Qi (q *X = je C ' i J) Can be obtained in linear time from the ^X such that: [Math.9] C 'j =ciA- Using this method, it is sufficient to recalculate the first element (C'p each time the Q query is updated. Similarly, it is also possible to parallelize the calculation of the matrix profile of all distance profiles. - Optimizations for distance profiles with standard normalization: The Euclidean distance with standard normalization can also be similarly reformulated to include mean and deviation terms. type. This allows the use of methods similar to simple Euclidean distance optimizations. Thus, the distance can be reformulated as follows:

[0113] [Math. 10] / QW-lp,11. \

[0114] Further, more advanced optimizations are also possible using graphics processing units (GPUs) to perform the calculations. Optimizations for other distances, such as dynamic time warping, are also possible and allow for a reduction in complexity similar to that obtained for Euclidean distance. Furthermore, time series indexing methods can be implemented in conjunction with, or as a replacement for, the methods described above.

[0115] Depending on the implementation, the historical data (X) and query data (Q) may also be incomplete, excessive, or insufficient. For example, data may be missing (because the time intervals for data collection are different or because the sensors are defective). Therefore, in such cases, preprocessing of this data may be important. In an industrial site such as those described above, there is a large amount of sensor data. To avoid bias, the inventors selected data that is not influenced by the decisions or operating rules of the industrial site. For example, the activation of washing pumps or changes in the speed of specific mechanisms (such as a filter drum) were excluded.

[0116] According to the present embodiment, the retained variables (data) undergo completion, for example, linear interpolation when the number of successive missing data points is less than or equal to a predetermined threshold (for example, starting from 3 missing values), and interpolation via a regression model in other cases. For example, to predict the missing water level over certain intervals, it is possible to use water flow data and the distance of the moon and the sun (data that influences tides) to predict the missing values. Also, some data may exhibit a discrete rather than continuous nature over certain time intervals (possibly due to interpolation). A moving average can then be applied to prevent the re-identification of these intervals by the method of the invention, because of this discrete nature.

[0117] In relation to [Fig.3], the computerized device (100) for implementing the method for determining the risk of a clogging event occurring at a pumping station comprises several interconnected components to ensure the efficient processing of environmental and operational data.

[0118] The device (100) includes a central processing unit (CPU) (110) that executes the instructions of the computer program. This central processing unit is connected to a main memory (120) that stores the data and programs necessary for the operation of the device. The main memory (120) may include random access memory (RAM) modules and non-volatile memory modules (ROM, SSD).

[0119] To accelerate intensive calculations, the device (100) is equipped with graphics coprocessors (GPUs) (130). These GPUs are used to perform parallel calculations, in particular for cross-correlation and convolution operations, thus reducing the processing time of multivariate time series.

[0120] The device (100) also includes interfaces for obtaining representative data of current environmental conditions (CREC) from the sensing devices (CAP0, CAP1,..., CAPx) (140). These sensors are, for example, connected to the device via communication interfaces (150) such as USB ports, network interfaces (Ethernet, WiFi), or wireless communication protocols (Bluetooth).

[0121] The captured data are organized into multivariate time series (Q) and stored in a database (160). This database can be hosted locally on a hard disk drive (HDD) or an SSD, or remotely on a cloud server accessible via a network connection.

[0122] The device (100) may also include a dedicated or non-dedicated data processing unit (170) that executes the similarity search and distance profile calculation algorithms. This processing unit may be integrated into the central unit (110) or be a separate module optimized for time series calculations.

Claims

Demands

1. Method for determining data representing a risk of occurrence of a clogging event of a pumping station of an industrial site, method implemented by means of an electronic device comprising a memory and a computing unit, method comprising at least one iteration of the following steps: - obtaining (SOI) at least one time series Q of current environmental conditions (DRCEq) of the pumping station; - determining (S02), within historical data in the form of at least one time series X of past environmental conditions (DRCEX), at least a partial correspondence of at least one variable of the time series Q and a corresponding variable of said at least one time series X;- when it is determined that at least a partial correspondence exists between said at least one variable of the time series Q and said at least one corresponding variable of said at least one time series X, obtaining (S03), using a time series Y of past clogging events of the pumping station, the risk data for the occurrence of a clogging event of the pumping station.;

2. Method of determination according to claim 1 characterized in that said at least one time series Q of current environmental conditions (DRCEq) and said at least one time series X of past environmental conditions (DRCEX) each comprise at least one time series relating to the external environment ^ENV of the industrial site and at least one time series relating to the internal environment QTRA, T RA of the pumping station and / or the industrial site.

3. A method for determining the conditions of claim 2, characterized in that said method determines at least one time series relating to the internal environment QTRA, ^TRA of the pumping station includes variables that are independent of decisions and / or operating rules of the industrial site.

4. Method of determination according to claim 1 to 3, characterized in that the step of obtaining (SOI) the time series Q of current environmental conditions (DRCEq) comprises at least one step of completing missing values ​​for at least some variables of the time series Q of current environmental conditions (DRCEq).

5. A method for determining the value of claim 1, characterized in that the determination step (S02) comprises: - a first search step, based on a first current time series QENV comprising at least one variable relating to the external environment of the industrial site, for at least a partial match with a time window WENVi extracted from a first time series XE^-y comprising at least one corresponding variable, the size of the time window WENVi being at most equal to the size of the first time series QE^y

6. A method for determining according to claim 5, characterized in that when at least a partial match is found during the first search step, a second search step, based on a second current QTRA time series comprising at least one variable relating to the internal environment of the industrial site, finds at least a partial match with a WERAi time window extracted from a second XERA time series comprising at least one corresponding variable, the size of the WTRAi time window being at most equal to the size of the second QTRA- time series.

7. Method of determination according to claim 5 or 6, characterized in that the first step of searching for at least a partial match is implemented for a set of time windows (WENV1, ..., WENV1.niV}).

8. A method for determining the time series according to claims 1 to 7, characterized in that at least a partial correspondence between two time series is obtained by calculating a minimum distance separating

9.

10. the values ​​of the different variables that correspond within the two time series. Device for determining data representing a risk of occurrence of a clogging event at a pumping station, device comprising a memory and a processing unit, said processing unit being configured to execute at least one iteration of the following steps: - obtaining at least one time series Q of current environmental conditions (DRCEQ) of the pumping station; - determination, within historical data presented in the form of at least one time series X of past environmental conditions (DRCEX), of at least a partial correspondence between at least one variable of the time series Q and a corresponding variable of said at least one time series X; - when it is determined that at least a partial correspondence exists between said at least one variable of the time series Q and said at least one corresponding variable of said at least one time series X, obtaining, using a time series Y of past clogging events of the pumping station, the risk data for the occurrence of a clogging event of the pumping station. Computer program comprising instructions for the implementation of a method according to any one of claims 1 to 8, when said instructions are executed by a processor of a computer processing circuit.

Images