FR3066844A1 - METHOD FOR PREDICTING THE RUNNING OF A FAILURE ON AN APPARATUS IN TIME - Google Patents

Abstract

The invention relates to a method for predicting failures of a rotating machine comprising: A. a learning phase where: a. the values of the quantities supplied by the sensors are recorded until a time "tp" of occurrence of a failure b. identifying a magnitude or a mathematical processing thereof as a Marker of the failure, if there is a correlation between its evolution as a function of time and the occurrence of the failure, B. a prediction phase where: a. we observe the current evolution of the set of earlier markers as determined in the previous step, and b. if there is a match between the current values of the Previous Markers observed and that of the values of the Previous markers of the set of reference markers considered at a time "tp-x", an alarm is emitted indicating a date "x" of possible occurrence of the failure.

Description

METHOD FOR PREDICTING THE OCCURRENCE OF A DEVICE FAILURE OVER TIME

TECHNICAL FIELD OF THE INVENTION The invention relates to the field of predicting the occurrence of a failure on a device within a complex system, over time and a device for measuring operating variables allowing the implementation of the method according to the invention.

By complex systems is meant systems in which a multitude of devices of different natures cooperate to contribute to a common objective, these devices being able to be of the computer, electrical, electronic type, rotating machines powered by an energy source, or a combination of these.

These complex systems can thus consist of a group of devices constituting a vehicle (for example automobile, aircraft ...) or a sub-group of these devices intervening only on a particular functionality of this vehicle (engine, air conditioning, lubrication, resistive energy distribution network ...), or in a group of devices constituting a production unit of a factory, such as the electricity production unit of a nuclear power plant. The invention relates more particularly to the forecasting of the occurrence of a failure of at least one apparatus of a complex system, and to the quantification of the time remaining before its occurrence.

STATE OF THE ART

Conventionally, these kinds of complex systems are found in various economic sectors and in particular in aviation, industry or automation. Problems often arise during maintenance operations, since it can be very difficult to locate a faulty element of the system, which is the cause of a failure. The implementation of fault detection methods with additional detection equipment is also not always effective, leading, for security reasons, to the replacement of a set of elements. In all cases, maintenance operations which do not make it possible to locate a fault precisely or which involve additional detection equipment, generate an increase in maintenance costs.

There are already known diagnostic methods for locating a failure in a complex system, consisting in verifying the performance of the complex system based on operating information obtained from detection means, in determining from operating information an operating status. said operational, non-operational or degraded system, to compare, when the non-operational or degraded operating status has been determined, the operating information with predetermined data and to generate at least one hypothesis as to the location of the failure in the complex system.

These diagnostic methods, however, have a number of drawbacks.

Indeed, the known diagnostic methods are based, among other things, on a probabilistic analysis of failures.

This analysis generally makes it possible to indicate to the maintenance operator, one or more elements likely to be the cause of a failure, with a given degree of certainty. The latter is expressed using a corresponding probability calculation. The known diagnostic methods use a calculation algorithm in which most often there are almost arbitrary approximations and weights for failure messages from different tests. In addition, the diagnostic methods define arbitrary time windows, for example of a duration of ten seconds, during which the failure messages relating to distinct failures are taken into account. The results thus obtained by such algorithms intended to locate a failure in the complex system, cannot therefore be used as it is.

To overcome these drawbacks, the document FR 2 933 512 describes a method which consists in detecting failure occurrences and generating corresponding failure messages, filtering failure messages, determining a time window in which successive failure occurrences must occur. produce to be retained, perform contextual correlation analysis of failure messages, perform dynamic analysis on failure occurrences, and compare the results of dynamic analysis and contextual correlation analysis to predetermined data to locate the failure in the complex system.

However, in this process, the location of the failure is always based on a probabilistic analysis and there is no information to estimate how soon the failure will occur.

Document US 7,254,514 relates to a method for estimating the remaining operating time of a device before the occurrence of a failure, but achieves this by a probabilistic analysis which is thus not entirely satisfactory in terms of efficiency and details on the announced date of occurrence of a failure.

In order to overcome these drawbacks, the invention provides a non-probabilistic method for determining the date of occurrence of a failure, making it possible to best manage the preventive maintenance phases of the devices of a complex system.

OBJECT OF THE INVENTION To this end, and according to a first aspect, the invention proposes a method for predicting operating anomalies, failures, and / or breakdowns of a rotating machine from the monitoring of different quantities of operation of the machine such as its vibrations, its temperature, its ultrasound, its electromagnetic, physical or chemical parameters, etc., measured by a plurality of measurement sensors comprising: A. an initial learning phase during which: a. the values of the gross quantities characteristic of the operation of the machine supplied by the sensors are recorded in the form of historical data up to a time "tp" identified as that of the occurrence of at least one anomaly, failure and / or failure b . we identify the quantity or a mathematical treatment of a quantity, as a Relevant Marker of the occurrence of the failure, known as the Previous Marker, if its evolution as a function of time is progressive with a significant break occurring in a time interval located around the moment "tp" of the occurrence of the anomaly, failure and / or failure, different Markers thus identified for the same failure, forming a set of reference Markers, whose evolution of values over time is known until the moment the occurrence of the anomaly, failure and / or failure at the end of the learning phase, B. a phase of prediction of the date of occurrence of the failure for the machine in question during which: a. we observe the current evolution of the set of Earlier Markers as determined in the previous step, and if there is an adequacy between the current values of the Earlier Markers observed and that of the values of the Earlier Markers of the set of reference Markers considered at at an instant "tp-x", an alarm is emitted indicating a date "x" of possible occurrence of the failure.

By “adequacy between the current values of the Markers of the set of Markers and those of the values of the Markers of the set of reference Markers”, one understands that at least half of the Markers of the set of Markers currently observed reaches the values of the corresponding Markers in the set of Reference Markers.

By "Relevant marker" is meant a Marker which, for different occurrences of the same failure, presents in at least 80% of cases, a significant rupture occurring at the same time "tp" of the occurrence of the failure (correlation rate greater than 0.8 in absolute value).

Advantageously, the invention may have one and / or the other of the following characteristics: it comprises a step of identifying the moment of occurrence of the failure consisting of starting from the historical data delivered by said sensors, to be characterized: periods normal operation of said electric machine, periods of voluntary interruption of service of said electric machine - abnormal periods of interruption of said electric machine, the time of occurrence of a failure being determined from the detection of the end period of a normal operating sequence of said electric machine followed by a period of abnormal interruption of said electric machine. characterization of periods of normal operation and periods of voluntary interruption of service is carried out: from the identification in historical data of cyclical operating data from an operating schedule and / or from the comparison with the operation of other machines in the same assembly. it includes a step of identifying the moment of occurrence of the failure consisting in integrating external data such as the data of a CMMS software or a data entered by an operator the sensors are controlled to carry out measurements of the operating quantities with a high initial sampling frequency, such as 10,000 data per second, and if no variation is detected between two acquisitions, the sampling frequency for the data concerned is reduced, and / or if a variation is detected important between two acquisitions, the sampling frequency for the data concerned is increased if during the learning phase, different sets of Prior Markers are identified as relevant to characterize the same failure, a probability of occurrence of the failure is calculated for each set of relevant markers during the prediction phase, we continue to record the evolutions in time of the various Previous Markers identified during the learning phase and we modify according to the evolution of the sets of Reference Anterior Markers characterizing a failure during the prediction phase, we continue to record the values of the quantities characteristic of the operation of the machine by the sensors in the form of historical data until the occurrence of at least one failure and the nature of the Previous Markers of the set of Reference Markers characterizing the failure is modified according to the learning phase and / or during the prediction phase: o the variables characteristic of the operation of the machine by the sensors are recorded in the form of data also logged beyond the occurrence of at least one failure and beyond the occurrence of a repair thereof, where we observe the evolution of the relevant Markers after the occurrence of a repair on the machine, known as Posterior Markers, o and if the signature characteristic of the anomaly has disappeared in the Posterior Markers, information indicating positively the repair carried out is indicated or o if the date of possible occurrence of a failure determined from of the Later Markers, is: greater than the date of possible occurrence of a failure determined from the Previous Markers, information indicating positively the repair carried out is indicated equal to or less than the date of possible occurrence of a failure determined from the Markers An indication is given which qualifies negatively the repair carried out. The invention also relates to a system for predicting failures of a rotating machine on the basis of monitoring the various operating variables of the machine such as its vibrations, its temperature, its ultrasound, its electromagnetic settings ticks, physical or chemical ..., measured by a plurality of measurement sensors, implementing the method of any one of the preceding claims, and comprising: a server provided with a computer processor capable of collecting in real time data from different sources among sensors, capable of performing mathematical processing of this data, capable of comparing this data or of mathematical processing thereof with recorded reference changes, and provided with a memory for storing this data, of their mathematical processing, and of their evolution over time, a group of sensors provided with a means of remote communication with the server and with a local memory for temporary storage of values measured by the sensors, the measured data being stored up to 'when transmitted to the server, then deleted a terminal for the user in communication with the server provided with an interface this display of the alarm of the possible date of occurrence of a failure on a machine.

Ideally, the processor also integrates the data supplied by sensors intrinsic to the machine and / or the data from CMMS software.

Preferably, the group of sensors is brought together in a single box with magnetic attachment to the machine, and includes at least an ultrasonic sensor, a vibration sensor producing vibration spectra, a temperature probe.

According to another interesting characteristic, the alarm display interface is the interface of the CMMS software. The invention also relates to a method for automatically detecting the occurrence of a failure from monitoring the historical data supplied by a multitude of sensors for monitoring the operation of a device, which consists in characterizing: - periods of normal operation of said electric machine, - periods of voluntary interruption of service of said electric machine - abnormal periods of interruption of said electric machine, the moment of occurrence of a failure being determined from the detection of the end of period d 'a normal operating sequence of said electric machine followed by an abnormal interruption period of said electric machine.

According to an advantageous characteristic of this process, the characterization of the periods of normal operation and periods of voluntary interruption of service is carried out: from the identification in the historical data of cyclic operating data from a planning of operation and / or from comparison with the operation of other machines in the same assembly.

BRIEF DESCRIPTION OF THE FIGURES Other objects and advantages of the invention will appear during the description which follows, given with reference to the appended figures, in which: FIG. 1 schematically represents the architecture of the system implementing the method according to the invention. FIG. 2 represents a schematic perspective view of an embodiment of a monitor for capturing current values of the relevant operating variables of a device. FIG. 3 represents an interface for signaling the next occurrence of a failure on an apparatus as provided by the method according to the invention FIG. 4a represents a graph illustrating the vibrations measured as a function of time by a sensor positioned on an apparatus (spindle of a machining center) FIG. 4b represents a Relevant marker defined by the method according to the invention from the vibration quantity illustrated in FIG. 4a, represented nt the evolution of average vibrations measured over a short period (for example a week) Figure 5a represents a graph illustrating on a first curve (circles) the temperature T, and on a second curve the vibrations or an average of these over a short period (for example a week), obtained for a rotation of a hydroelectric group, FIG. 5b represents a graph illustrating the curves of FIG. 5a with a leveling of the bearings that they each define. FIG. 5c represents a graph illustrating a Relevant Marker which can be defined from the magnitudes vibration and temperature represented in FIGS. 5a and 5b which is constituted by an artificial sum of the numerical values “temperature” and “vibration” as measured at the same times, where as averaged at the same time intervals.

DETAILED DESCRIPTION OF THE FIGURES

In connection with FIG. 1, there is described a method of predicting / predicting failures, anomalies or breakdowns of an electric, pneumatic, belt-driven, hydraulic or mechanical drive machine using a plurality of sensors comprising sound, or vibrational, or electromagnetic or physical or chemical sensors.

It is advantageously implemented within groups of devices operating in synergy or not, within the same unit, in real conditions, such as the devices of a factory, of a power plant or of a nuclear power plant, of a research, manufacturing, production laboratory, as in laboratory conditions for the various devices of a pilot, or any other group of devices.

And this, in different sectors of industry including energy (production, distribution), transport (real-time monitoring of vehicles, functional analysis, optimal use), and different types of infrastructure (elevators, building framework , ventilation system).

This process mainly allows the detection of anomalies occurring on a machine, from the monitoring of different operating parameters of this machine (vibrations, operating temperature, generated ultrasound, fluid levels (fuel, oil).

Thanks to the method according to the invention, can be obtained: a photograph of the anomalies existing at an instant t in a group of devices a photograph of the anomalies which will occur in the more or less near future, with a statistical weighting of the date of occurrence an anomaly a definition of one or more successions of situations (known as scenarios) which can lead to the occurrence of an anomaly with a statistical weighting for each identified scenario. The idea being to better predict for each device, the date of inevitable occurrence of a failure, in order to plan maintenance of this device: as close as possible to the occurrence of the failure but before that, in order to take advantage of the most long period of operation without repairing the device in question, and ideally in time slots dedicated to the maintenance of the site in question, so as not to impact the general operation of the site by a specific maintenance shutdown for a device. The invention can be implemented to predict the occurrence of breakdowns, failures or anomalies in devices of the rotary machine type.

An example of implementation of the invention is given in the following for this type of machine, encountered mainly in machining centers, for purely illustrative purposes, without these examples being limiting in nature. of devices to which the invention may apply.

In rotating machines, the typical failures encountered mainly occur at the spindles, axes of rotation, translation, tilt that drive movements of various parts.

Among the most frequently encountered failures, one can cite: 1) faults linked to the movement of an axis of displacement in translation of a part, and coming from: a bad positioning of the axis according to a given setpoint (precision on the position to be reached) of a bad supply current of the servomotor controlling the axis for a given setpoint (precision on the current) of a range of setpoints for which positioning faults appear or supply current (precision on the set point) faults appearing in the service life of the axis (natural drift due to wear or deformation of the materials due to operation) 2) faults linked to movement of a rotating axis or spindle and originating from: alignment errors of the rotating axis with respect to an additional part, such as a bearing of balancing defects of the axis itself or of a part attached to it in rota tion (mechanical defect) of precision faults in the angular positioning of the axle or of a part which is integral with it in rotation of faults appearing in the service life of the axle (natural drift due to wear or material deformation due to operation)

General presentation of the architecture (figure 1)

As shown in FIG. 1, the prediction system according to the invention 1 mainly comprises: a server 2 implementing a computer learning machine processor: o able to collect in real time data from different sources such as sensors 6 extrinsic to the machine studied, sensors 4 intrinsic to it, or even data from CMMS 5 software capable of performing mathematical processing of this data, and choosing digital treatments according to predefined criteria mathematics to be retained as Markers of a given failure on a given device, o able to record the evolutions of the Markers retained as a function of time in the form of reference developments, o able to compare data acquired or being acquired , or mathematical treatments of these, to recorded reference changes, o and provided with a memory of st storage of these data, their mathematical processing, and their evolution over time, - a group of external sensors 6 in the form of a monitor with magnetic fixation and antenna for remote communication with the server, containing different sensors (by example for monitoring a rotating machine type device: an ultrasonic sensor, a vibration sensor producing vibration spectra, a temperature probe, and for monitoring a device of the driveshaft type, a the speed of rotation of the shaft, and a vibration sensor performing vibration spectra), provided with a local memory for temporary storage of values measured by the sensors, the measured data being stored until it is transmitted to the server, then deleted (flash memory) - a terminal 7 for the user in communication with the server 2 provided with an interface for displaying the alarm of the possible date of occurrence of a machine failure.

In addition, data from different sources will be stored and made available through a cloud interfacing to third-party systems through its APIs.

The data acquired and the analysis are made available on a platform via web and mobile applications.

The sensors are controlled to carry out measurements of the operating quantities with a high initial sampling frequency, such as 10,000 data per second, and if no variation is detected between two successive acquisitions, the sampling frequency for the data concerned is reduced (for example by 50%, and / or if a significant variation between two acquisitions is detected, the sampling frequency for the data concerned is increased (for example by 50%). Changing the frequency Sampling can also be done manually.

The sensors

The monitor 6 as shown in FIG. 2 is a flexible multi-measurement acquisition system in the definition of the sensors that it integrates.

It embeds several selected sensors in a thematic logic: for an electromechanical device, it would for example be relevant to measure vibrations, ultrasound, contact temperature while for so-called ambient supervision it would be more useful to monitor a gas content specific (flammable, toxic, ...) or ultrasonic signals.

Generally, these sensors are on non-intrusive measurements:

Tribology, pH, alignment, vibrations, temperature, speed, geolocation, fuel level, ultrasound, power factor.

Some examples of on-board sensors within the monitor: temperature, vibration (spectra) sensor, real-time oil analyzer, pressure sensor, current sensor) ...

The flexibility offered by this monitor therefore makes it possible to completely personalize it (materials, sensors, details), in order to collect the most relevant data.

These monitors are chosen robust, resilient and waterproof. The type of materials the constituents can adapt to the use for which they are intended (very high temperatures, rooms subject to flooding, high pressures, etc.). In addition, their geolocation facilitates their use on geographically dispersed and / or difficult to access assets.

This monitor constitutes a wireless acquisition system, high precision and bidirectional, communicating on a broadband network, medium range, low consumption, but also on long range networks (LoRaWAN) for geographically distributed assets.

A monitor 6 has an internal space and a data storage memory allowing it to accommodate four separate on-board sensors or to be connected to thirty-two separate external sensors in multiplexed mode (probes).

Other features:

Vibration analysis up to 10 kHz, data transmitted by the monitor on the 868 MHz band, on periodic request from the server (for example one request for vibration spectrum per day) or automatically when a predefined number of spectrum is recorded or a particular storage capacity reached, possibility of acquiring measurements at the same time as these measurements or other measurements are transmitted to the server acquisition of the data that is stored in an optimized local storage memory (flash memory) , then deletion of the data transmitted ultrasound 40 to 400 kHz temperature sensor multichannel sensor: which allows to have several sensors at the same time battery management by the monitor who will administer the sensors each monitor has a unique address (ID) and is administered by a bridge which verifies its ID and automatically manages the encryption key when the monitor is switched on, a 30 ms time window opens. 300 ms during which it starts listening (a spectrum request by the server, or the reception of an encryption key by the bridge).

Learning phase

A multitude of data from different sensors is recorded over a long period (a few months, for example three or four).

These data are recorded in a database located at the level of the sensor, temporarily stored in this database and deleted as soon as they have been transmitted to the server.

These data are analyzed to extract metadata or Markers which will be used to establish the prediction of the date of occurrence of the failure.

Markers or Metadata

The raw data supplied by the same sensor are made up of a time series of values marked by the date on which each was acquired (time stamped) representing the measurements made for the quantity measured by the sensor.

A marker or a metadata within the meaning of the invention will be made up: either by the raw time-stamped data measuring the same quantity, which are then organized in point clouds by the method of nearest neighbors, in order to define an analytical function representing the evolution of these point clouds as a function of time either by digital processing of raw data coming from the same sensor or from different sensors such as: o a mathematical processing brought to the point clouds coming from raw time-stamped data (for example integral, derivative, square or other of the analytical function representative of these raw data) followed as a function of time o an "intelligent" combination of raw data with time stamps of different quantities (preferably limited to a combination of three quantities) (for example a sum of analytical functions representative of two quantities) followed as a function of time o the monitoring of a prem 1st quantity as a function of another quantity

Raw data representative of a quantity or any digital processing thereof, will also be considered as a Marker or Metadata relevant for the characterization of a failure if this raw data (or the analytical function representing it) or any digital processing of this, evolves over time with a break (slope break, appearance of a peak, different behavior, appearance of a plateau ...) occurring at the same time as that of the detection of the failure and presents an interaction consistent with the failure considered.

By consequent interaction between a Marker and a given failure, we mean a Marker which for different occurrences of the same failure, presents in at least 80% of cases, a significant rupture occurring at the same time "tp" of the occurrence of the failure ( correlation rate greater than 0.8 in absolute value).

Identification of the date “tp” of occurrence of failure on revolution markers

An important step of the device according to the invention consists in identifying the moment of occurrence of the failure from the historical data delivered by said sensors.

This identification is carried out by analyzing the raw data supplied by the sensors to identify: - periods of normal operation of said electric machine, - periods of voluntary interruption of service of said electric machine (weekends, at night, the public holidays, during scheduled maintenance periods ...) - abnormal periods of interruption of said electric machine due to the occurrence of a fault on the device.

The time of occurrence of a failure is determined according to the invention from the detection of the end of the period of a normal operating sequence of said machine followed by an abnormal interruption period of said electric machine.

The method according to the invention can provide for a characterization of the periods of normal operation and of periods of voluntary interruption of service: - from the comparison of the cyclical data of operation of a machine such as deduced from the recording of the data from sensors during the learning period and / or - from an operating schedule for the device or group of devices, and / or - from comparison with data from other machines the same set.

Assessment of the relevance of the Marker or Metadata From the relevant raw measurements or information from the sensors and recorded as a function of time, one or more metadata is generated in the following manner: we choose relevant quantities to be measured by the device's sensors according to the invention as a function of the apparatus studied o for example in the aforementioned case of a rotating axis controlling the displacement in translation of a block, these quantities will be the current position Ai of the block along the axis X, the current Bi supplied to the drive motor in rotation of the rotating part, the setpoint Ci supplied to the motor) thanks to an analysis of the data of these quantities measured by the sensors, and / or to an indication external to the sensors (maintenance report , CMMS computer-assisted maintenance management software), we identify the date tp of occurrence of the failure or anomaly and possibly the subsequent date tR of its repair o example: sudden imbalance of a rotating part around an axis which results in the monitoring of the angular position of the rotating part as a function of time, by a difference between the current position of the rotating part (or the block to be moved) and the dictated setpoint, deviation which begins at time tp and then increases. The repair, if it is correctly carried out, will be observed by bringing the current position of the rotating part or block back into line, and the setpoint dictated from an instant tr if one of the quantities measured, or a mathematical treatment of this (its integral or its derivative, squared ...) or a combination of two quantities (sum, followed by the evolution of a first time-stamped quantity, according to a second quantity with the same time-stamp ) or a mathematical treatment of such a combination, presents an evolution as a function of time which presents a "rupture" (peak, different slope, different behavior, appearance of a plateau ...) at the same time as that of the occurrence of the fault, this quantity, or its mathematical treatment, could constitute a relevant Marker: o In the above example, the current supplied to the drive motor of the rotating part to position it at the requested setpoint e, may present from the instant tp of occurrence of the imbalance, a slope or a form of evolution different from those before the instant tp, and may for this reason be considered as a Metadata o for example a combination of quantities, may consist of monitoring the evolution of time-stamped measurements of a first quantity as a function of the corresponding time-stamped measurements of the second quantity (for example monitoring of the current position of the block as a function of the detected current of the motor actuating l 'axis of its displacement), or squared of the measures of this second quantity, or any other relevant mathematical processing of the measures of this second quantity; to the sum of these quantities as a function of time ... o another combination of quantities may consist of monitoring the evolution as a function of time, of the average time for a block or an axis to reach a current position respecting a position setpoint given to detect drifts

If mathematical functions such as the exponential, the logarithm of the gross quantity, a combination of quantities or the mathematical treatments of these present as a function of time a "rupture" at the same time as that of the occurrence of the failure, these mathematical functions could constitute a relevant marker.

To guarantee the selection of a really relevant set of Markers, that is to say whose different markers each correctly characterize the particular anomaly considered, without being redundant and being, on the contrary, the least correlated possible with each other, a maximum correlation coefficient to be observed between two Markers in the same game. For example, this maximum correlation coefficient could be set between 0 and 0.5 in absolute value, preferably 0.5.

A quantity or a mathematical treatment of this could be considered as a Relevant Marker if in its evolution as a function of time the significant rupture occurs not at the same time "tp" of the occurrence of the failure, anomaly, failure but in an interval located around this moment "tp". This interval may include the different moments of failure occurring resulting from different scenarios leading to the same failure identified during the learning phase.

Exploitation of Markers or Metadata

Several different Markers or Metadata A, B, C, ... I, ..., N can be identified for the same failure / anomaly and their evolution as a function of time A (t), B (t), C (t ), ... I (t), ..., N (t) until the occurrence of the fault and beyond, can be reconstructed from the data of the raw measured quantities time-stamped by the sensors, and recorded as a benchmark evolution during the learning phase.

This reference evolution may highlight: a set of values of these markers at the precise moment of the occurrence of the failure at time "tp": A (tp), B (tp), C (tp), .. .I (tp), ..., N (tp), which will constitute an identifier or a Signature characterizing the failure concerned on this device any set of values previous or set of Markers prior to the occurrence of the failure at the moments “tp- x ”: A (tp-x), B (tp-x), C (tp-x), ... I (tp-x), ..., N (tp-x), which give at each instant previous, a previous state of the relevant Metadata to characterize the failure considered on the device considered, among: o the instant before the failure / anomaly Atp. „Btp„ Ctp., o earlier before the failure A ,,. , », B ,,.,», C /., »O any time before the failure Atp.P, B,

o long before the failure / anomaly A ,,., ™, B ,,., ™, CN —- tp-1000 of the temporal sequences of relevant Markers characteristic before the failure, any set of values after the occurrence of the failure , including when a repair has been made to this failure, for example at time "tR": A (tR), B (tR), C (tR), ... I (tR), ..., N (tR)

And the states following this repair, which give at each instant following the repair, a subsequent state of the relevant markers to characterize the failure considered on the device considered at the instant "tR + y": A (tR + y), B (tR + y), C (tR + y), ... I (tR + y), ..., N (tR + y) (Posterior Markers): o Am, B „Ct„, o a certain time after the repair AM „, Bt„, „, CM„ o long after the repair A MOO „, BM0M, C MM0 and time sequences of relevant Markers after the Repair.

By analyzing the device under consideration over a long period, it is possible that the same failure occurs with different causes and thus different groups or sets of Metadata or Markers (first set of Metadata leading to the failure p: Atp, B, C /; second set of Metadata leading to the failure p: D, Εφ, FJ. The study of the data provided over this long learning period may in this case make it possible to define: different scenarios or sets of Metadata leading to the same type of failure and statistics of occurrence of failure according to the scenario considered.

Prediction phase of the date of occurrence of a future anomaly

Once the initial learning phase is complete, and that different sets of Markers (at least one) are identified for the same failure, different reference evolutions of relevant Marker sets are recorded and ideally different scenarios likely to lead to the same anomalies are identified and weighted by statistical coefficients of their occurrence, it is possible to predict the date of occurrence of a future anomaly by monitoring the data delivered by the various sensors. To this end, according to the invention: a. the learning processor analyzes the current evolution of the relevant Marker sets determined in the previous step, from the measurements transmitted to the server by the sensors b. the microprocessor compares: i. the current values of the different markers of the same set of Markers characteristic of an anomaly with the values recorded during the learning step for the same set of Reference Markers or ii. a temporal sequence observed on the current evolution of the relevant Marker sets with a temporal sequence of the reference evolution recorded during the learning phase c. and transmit instructions for issuing an alarm indicating the occurrence within a duration “x” of a failure if: i. the current values of the various markers of the same set of Markers characteristic of an anomaly are identical at least in part (for example two markers out of four) to the values of previous Markers of a given failure at an instant tp-x before the occurrence of the failure, or if ii. the temporal sequence observed on the current evolution of the relevant Marker sets reproduces the evolution of a temporal sequence recorded during the learning phase iii. and if the set of markers considered characterizes a scenario of occurrence of the failure having a probability of occurrence greater than a threshold (for example 30%). d. otherwise, the analysis of the raw data is continued with the same set of markers and / or a second set of markers characteristic of the same or another failure is analyzed.

Constant learning

During the prediction phase, we continue to record: the changes over time of the various relevant markers identified during the learning phase and we modify the reference changes characterizing a failure as a function, also, ideally, we continue to record the variables characteristic of the operation of the machine by the sensors in the form of historical data until the occurrence of at least one fault and the relevant markers and reference changes characterizing a fault are modified accordingly

Quality of a repair made

During the learning phase and / or during the prediction phase: the variables characteristic of the operation of the machine by the sensors are recorded in the form of historical data also beyond the occurrence of at least one failure and beyond of the occurrence of a repair thereof, the evolution of the relevant markers is observed after the occurrence of a repair on the machine, called posterior markers, and an alarm is emitted indicating the possible date y of occurrence of the failure from the time difference between the corresponding moment (tp-y) on the reference evolution, to the set of values of the relevant relevant markers, and the moment (tp) of the occurrence of the failure on the evolution of reference to determine the date of occurrence of the next failure after repair.

If the repair is correct, the signature characteristic of the anomaly will have disappeared in the Later Markers or the remaining time "y" before the occurrence of the next anomaly should be greater than the remaining time predicted before the repair.

The disappearance of the signature characteristic of the anomaly or the indication of a new remaining duration greater than the remaining duration initially calculated, may allow the quality of the repair carried out to be assessed.

Example

The method according to the invention has in particular been implemented in a pellet press in the food industry. In the case of this press, the monitoring instrumentation concerns the bearings and the drums (motor at 1500 rpm). The monitor 6 has been specifically configured for monitoring this pellet press (and of any pellet press of the same type) and is fixed by its magnetic base in this case of use near the engine and the belt 8. From when the monitor is started, the measurements are collected and instantly available on the platform. On each of the instrumented subcomponents, relevant measurements are acquired in order to have continuous monitoring of the degradation of the press: vibration (global level and spectrum), infrared temperature, ultrasound, which are measured by sensors intrinsic to the monitor. 6 and the intensity of the current (thanks to an ammeter sensor external to the monitor located in the electrical cabinet).

Once acquired, the data is analyzed by integrating the thresholds defined by the maintenance technician from the platform or those defined by the existing standards on the machine concerned (conditional maintenance) in the manner indicated above to define remaining usage times before the occurrence of a breakdown, quantified in hours, day, and possibly weighted by a statistical coefficient representative of the probability of accuracy of the duration announced.

The alerts relating to these durations of failure occurrences are then sent to the cloud according to the rate chosen.

The bidirectionality of the monitors 6 makes it possible, on measurements such as vibration, to acquire spectra which are analyzed and interpreted automatically with detailed monitoring of the development of the anomalies detected. These analyzes as well as the history of the data are available and can be exported in several formats (XML, CSV or PDF).

The method according to the invention, thanks to its automatic learning processor tracks several fault signatures on the machines in order to automatically extract reliable fault scenarios therefrom. The analysis of these markers makes it possible to detect and monitor the failure at the first proven signs of anomaly.

Gradually, the prediction is refined; the characterization of the failures corresponding to this becomes more and more formal. The measures concerned as well as the type of failures characterized upstream therefore make it possible to act in a targeted manner and to optimize actions and resources: act early enough to avoid heavy and costly failures, but never too early to avoid expenses and unnecessary stops.

Example of Figure 4: Relevant marker from a single quantity (vibrations)

The curve of FIG. 4a illustrates the vibrations measured during the initial learning phase of the method according to the invention, by a sensor of a system according to the invention, illustrating the degradation of a spindle over 10 weeks, and which s '' resulted in a spindle failure.

The critical vibration threshold is reached around the 13th week from which the frequency of observation of vibrations above the critical threshold (shown by the dashed line) increases enormously. A safety interval x not to be reached can be defined as starting at this 13 ”week and ending at the 20 week, week of failure. The evolution of these vibrations over time (average values as a function of time) can be modeled by an exponential model of equation: y = 20.879e0.1419x which presents a rupture at the start of the safety interval x predefined in meaning that the curve exceeds the maximum critical threshold existing at the start of the safety interval.

This model can therefore be used as a relevant marker to characterize the failure of this spindle.

During the prediction phase, monitoring the numerical values of the evolution of the vibrations over time, knowing the model "y = 20.879e0.1419x e" that this evolution follows as revealed by the learning phase, will allow to deduct for a current value from an average of vibrations yi, the remaining time xi before reaching the safety interval xs, at any time.

This model makes it possible to anticipate the end of life of a spindle with a period of 15 weeks for a certainty rate of 90%.

Example of Figure 5: Relevant marker from two quantities (vibration and temperature)

FIG. 5a represents in arbitrary units, the curves representing the changes in real time during the learning phase of the temperature (circles) and of the vibrations (crosses) of the bearings, until an instant of occurrence of a failure, such than qualified by an operator. These bearings can be those of a hydroelectric group of an alternator turbine assembly, for example those involved in the rotation of the alternator.

We note on each of the two curves the presence of a plateau respectively at the beginning of the curve for the vibrations (duration of approximately 14 weeks, bounded by the two oblique lines) and at the end of the curve for the temperature (duration of approximately 14 weeks, bounded by the two double oblique lines), which make them unsuitable for the definition of a relevant Marker since on these levels, the numerical values of the corresponding quantities are practically constant as a function of time and do not allow a moment to be identified precise on the curve but a time interval of several months, too wide to serve as a prediction of the date of occurrence of a breakdown / failure / anomaly.

These gross quantities are identified by the processor as unsuitable for the definition of relevant markers because they would imply uncertainties in terms of remaining operating time of ten weeks, which is unsatisfactory to say the least.

The processor decides that it is necessary to carry out a digital processing of these quantities in order to define Relevant Markers as understood by the invention. The analysis of Figure 5b on which the two curves are brought together on the ordinate to make their bearings coincide, show that the first operating drifts are detected thanks to the temperature when the vibration is still in "flat phase" (step not marked degradation observed). When the temperature reaches its “flat phase” (bearing), we begin to observe the operating drifts at the level of the vibration curve until reaching the end of the bearings' life.

The combination of the “temperature” (early phase wear model) and “vibration” (end of life phase wear model) markers is identified by the processor as enabling the definition of a model with linear progression over time over a wide area. period (28 months) without plateau of more than a week or even more than 2 days, and therefore without uncertainty in terms of remaining operating time linked to the taking into account of the gross quantities.

The combination (in this case the sum) of these two markers is illustrated in FIG. 5 and makes it possible to locate the functioning of the bearing over time in order to determine with precision the remaining life because at any value yi of the artificial sum “vibration + temperature” will correspond to a precise time (and no longer a time interval) including when the isolated gross quantities temperature or vibration are constant, and therefore a time remaining before precise breakdown “xi” also precise.

The curve resulting from the artificial sum of the temperature and vibration data follows a quasi-linear progression over 3 distinct time intervals phase l: y = 0.2x phase 2: y = O.lx + 0.1 phase 3: y = 1.2x - 4.8

This “vibration + temperature” sum therefore makes it possible to highlight a model for monitoring the overall life cycle of the bearings forming a Relevant Marker and thus to generate a prediction as soon as the first drifts are detected. The forecast analysis of bearing failures and its RUL (time remaining before the appearance of a failure) will therefore be carried out on the basis of the marker made up which presents quasi-linear phases of degradation (3 phases on the example of Figure 5) can be expressed as follows depending on the phases: y = ax + b where x is time or a and b are real and a non-zero.

Example of prediction presentation The interface available to an operator, as shown in Figure 3, informs him during the prediction phase, if the observation of the current evolution of the Relevant Markers reveals a set of Previous Markers of a fault identified during the learning phase: a) the type of fault concerned "Failure" YL axis fault "as it results from the set of Prior Markers concerned b) to d) the equipment concerned: b) : type of equipment "XXXX" c) location of equipment "XXXX XXXX" d) Relevant Marker measurements allowing the identification of the possible occurrence of a breakdown "Measures concerned: Motor temperature YL axis , YL Axis Encoder Temperature "e) the date of intervention to be planned (information deduced from the remaining time before failure x resulting from the observation of the Relevant Markers concerned" Intervention to be planned within less than 6 hours "f) a ideal intervention range from the maintenance calendar and the planned intervention date deducted ": 03/05/2017 13:00:00 -03/05/2017 14: 00: 00" g) one or more prescriptions ( s) useful repair if the learning phase or the previous test phase made it possible to identify, with an indication of the expected success rate: i. Waiting for temperature drop: 100% disappearance of markers ii. Cleaning / return to origin: 100% disappearance of markers

Claims (10)

1. Method for predicting operating anomalies, failures, and / or breakdowns of a rotating machine from monitoring different operating variables of the machine such as its vibrations, its temperature, its ultrasound, its electromagnetic parameters , physical or chemical, measured by a plurality of measurement sensors comprising: A. an initial learning phase during which: a. the values of the gross quantities characteristic of the operation of the machine supplied by the sensors are recorded in the form of historical data up to a time "tp" identified as that of the occurrence of at least one anomaly, failure and / or failure b . we identify the quantity or a mathematical treatment of a quantity, as a Relevant Marker of the occurrence of the failure, known as the Previous Marker, if its evolution as a function of time is progressive with a significant break occurring in a time interval located around the moment "tp" of the occurrence of the anomaly, failure and / or failure, different Markers thus identified for the same failure, forming a set of reference Markers, whose evolution of values over time is known until the moment the occurrence of the anomaly, failure and / or failure at the end of the learning phase, B. a phase of prediction of the date of occurrence of the failure for the machine in question during which: a. we observe the current evolution of the game of Previous Markers as determined in the previous step, and b. if there is an adequacy between the current values of the Previous Markers observed and that of the values of the Previous Markers of the set of reference Markers considered at an instant "tp-x", an alarm is issued indicating a date "x" of possible occurrence of the failure. 2. Method according to the preceding claim, characterized in that it comprises a step of identifying the moment "tp" of occurrence of the failure consisting of starting from the historical data delivered by said sensors, to characterize: - periods of normal operation of said electric machine, periods of voluntary interruption of service of said electric machine - abnormal periods of interruption of said electric machine, the time of occurrence of a failure being determined from the detection of the end of period d 'a normal operating sequence of said electric machine followed by an abnormal interruption period of said electric machine. 3. Method according to the preceding claim, characterized in that the characterization of the periods of normal operation and periods of voluntary interruption of service is carried out: from the identification in the historical data of cyclic operating data - from an operating schedule and / or from comparison with the operation of other machines in the same assembly. 4. Method according to claim 2 or 3, characterized in that it comprises a step of identifying the moment of occurrence of the failure consisting in integrating external data such as the data of a CMMS software or a data entered by a operator. 5. Method according to one of the preceding claims, characterized in that the sensors are controlled to carry out measurements of the operating quantities with a high initial sampling frequency, such as 10,000 data per second, and if it is not detected no variation between two successive acquisitions, the sampling frequency for the data concerned is reduced, and / or if a significant variation is detected between two acquisitions, the sampling frequency for the data concerned is increased. 6. Method according to one of the preceding claims, characterized in that if during the learning phase, different sets of Prior Markers are identified as relevant for characterizing the same failure, a probability of occurrence of the failure is calculated for each set relevant markers. 7. Method according to one of the preceding claims, characterized in that during the prediction phase, we continue to record the changes over time of the various previous markers identified during the learning phase and we modify accordingly the evolution of the sets of previous reference markers characterizing a failure 8. Method according to one of the preceding claims, characterized in that during the prediction phase, the values of the quantities characteristic of the operation of the machine by the sensors are continued to be recorded in the form of historical data until the occurrence of d '' at least one failure and the nature of the Previous Markers in the set of Reference Markers characterizing the failure is modified 9. Method according to one of the preceding claims, characterized in that during the learning phase and / or during the prediction phase, the quantities characteristic of the operation of the machine by the sensors are recorded in the form of data historically recorded. in addition to beyond the occurrence of at least one failure and beyond the occurrence of a repair thereof, the evolution of the relevant Markers is observed after the occurrence of a repair on the machine, known as Posterior Markers, and - if the signature characteristic of the anomaly has disappeared in the Later Markers, information indicating positively qualifying the repair is indicated or if the date of possible occurrence of a failure determined from the Later Markers, is: o greater than the date of possible occurrence of a failure determined from the Previous Markers, information indicating positively the repair is indicated. carried out o equal to or less than the date of possible occurrence of a failure determined from the Previous Markers, information indicating negatively the repair carried out is indicated 10. System for predicting failures of a rotating machine based on the monitoring of various operating variables of the machine such as its vibrations, its temperature, its ultrasound, its electromagnetic, physical or chemical parameters, measured by a plurality of sensors. measurements, implementing the method of any one of the preceding claims, and comprising: - a server provided with a computer processor capable of collecting in real time data from different sources among sensors, capable of performing mathematical processing of these data, able to compare these data or mathematical treatments thereof with recorded reference changes, and provided with a memory for storing these data, their mathematical processing, and their evolution over time, a group of sensors provided with a means of remote communication with the server and a local memory of temporary storage of values measured by the sensors, the measured data being stored until it is transmitted to the server, then deleted - a terminal for the user in communication with the server provided with an interface for displaying the alarm of the possible date of occurrence of a machine failure.