Frugal predictive maintenance procedure, corresponding computer program product and computer-readable medium

Description

[0001] The technical field of the invention is that of the maintenance of a fleet of machines, in particular but not exclusively vehicles, the present invention relating more particularly to a method of frugal predictive maintenance of a fleet of machines, a corresponding computer program product and a corresponding computer-readable medium.

[0002] The maintenance of a fleet of machines, in particular a heterogeneous fleet of machines, is facilitated by predictive maintenance processes which allow parts to be replaced before a possible breakdown, thus reducing the repair costs induced by the consequences of the breakdown and the downtime of the machines in the fleet.

[0003] Current predictive maintenance techniques for machine parks in industry fall into two main categories: A data-driven approach, in which data from sensors placed on machines is collected and processed by machine learning algorithms. The aim of these algorithms is to identify common patterns that can predict failure. A model-driven approach, which is based on the concept of multiphysics simulation and an instrumentation plan whose objective is to define thresholds and thus make predictions based on lifespan and constraints caused by the external environment (vibrations, temperature, etc.).

[0004] The data-driven approach alone requires a large amount of data that must be properly correlated with technical facts in order to make a prediction. In the specific area of ​​military vehicle fleets, there are relatively few vehicles in the fleets, and utilization rates are low compared to a vehicle in a civilian fleet. The data-driven approach, which works in the civilian sector, is therefore difficult to transpose to the military. Obtaining a reliable prediction would require several years of data and multiple failure recurrences, which is incompatible with a rapid deployment plan.

[0005] The model-based approach is well-suited to a "functional chain" approach because wear indicators are representative of the monitored component. However, this approach is limited in its ability to correlate the wear or aging impacts of one sub-assembly on another. Furthermore, in the military domain, vehicle usage patterns are difficult to model due to the irregular, even chaotic, nature of military vehicle use. Therefore, while the model-based approach is promising, it proves insufficient for obtaining reliable predictions, particularly in the military field.

[0006] US patent application US2022 / 269259 A1 describes a predictive maintenance method for a device in the food industry. US patent application US2018257683 A1 describes a vehicle subsystem monitoring system. French patent application FR3092057 describes a predictive maintenance method applied to automobiles and based on a model-centric approach.

[0007] The present invention aims to combine the best of both approaches, data-driven and model-driven, to overcome the limitations indicated above.

[0008] The present invention therefore relates to a predictive maintenance method for a homogeneous fleet of machines, the machines in the fleet being of the same type, the method comprising the steps of: a) identify, for the given machine type, at least one machine part to be monitored that is affected by machine wear, b) identify the set of measurable parameters of at least one machine part to be monitored, c) model a digital twin of the given machine type comprising a digital model of at least one machine part to be monitored and, on a so-called drive machine of a given machine type having a predetermined degree of wear: d) have sensors available to measure each measurable parameter of the set of measurable parameters of at least one part of the machine to be monitored, e) simulate a set of predetermined incidents on the drive machine and record the measurement data from the sensors for each predetermined incident, f) enrich the digital twin with the measurement data from the sensors for each predetermined incident of the set of predetermined incidents, g) define, for each simulated predetermined incident of the set of predetermined incidents, a detection threshold, h) identify, among the sensors, for each simulated predetermined incident of the set of predetermined incidents, a group of so-called maintenance sensors capable of detecting the detection threshold with a predetermined probability, i) equip all the machines in the machine park with the group of maintenance sensors,j) periodically collect measurement data from the group of maintenance sensors equipping each machine in the machine park to enrich the digital twin; k) for each machine in the machine park, periodically collect measurement data from the group of maintenance sensors equipping it and apply the measurement data to the digital twin to deduce a maintenance requirement.

[0009] The method of the invention is frugal because it requires little data and uses an artificial intelligence (AI) model or an algorithmic model to process this data, thus achieving frugality. Unlike the dominant AI model, which relies on deep learning or machine learning, frugal AI does not require large volumes of data to train its model. The frugal predictive maintenance method according to the present invention is capable of learning from very little data. Through frugal AI, the invention aims to successfully initiate, in steps d) to h), a sufficiently efficient model initially and then to continuously learn, in step j), through human-machine interactions.Data acquisition is therefore no longer done solely upstream of the implementation of predictive maintenance, but over a long period of time, and with the advantage of not having to mobilize significant expert human resources.

[0010] The sensors can be of any type. They can contain memory that can be queried in the workshop, via wired or wireless connection, or remotely, and / or can be associated with communication systems to transmit the measured data, either at regular intervals or on demand. Thus, low-bandwidth networks such as SigFox® or LoRa® or similar can be used to transmit the measured data from the sensors to the digital twin, where it is injected to enhance the digital twin. These transmissions may preferably be encrypted. According to the invention, it is also envisaged, particularly during steps f) and j), to transmit the data to the digital twin at high bandwidth to provide the digital twin with a maximum amount of data.

[0011] The proposed method allows for the creation of a unique digital twin of the machines in the machine park. This digital twin is continuously enriched with feedback using a data-centric approach. Compared to existing methods, the method of the invention thus improves process inference.

[0012] The method of the invention makes it possible to obtain a real-time health status of each machine in the fleet. When the machines are vehicles, it also provides driver assistance by indicating to the driver the best driving practices to limit wear and tear based on the behavior of the digital twin. The method also makes it possible to determine whether a given vehicle can be operated, for how long, and under what conditions, particularly for starting or completing a mission.

[0013] The machine part to be monitored can be any assembly or subassembly of the machine, any component or group of components, any functional chain or group of functional chains.

[0014] When the machine is a vehicle, the machine part to be monitored may be, but is not limited to: a mechanical assembly or subassembly such as an engine, engine part, running gear, axle, tire, track, transmission, gearbox, braking system, weapon system, heating system, air conditioning system, opening / closing mechanism; a hydraulic assembly or subassembly such as brake fluid, hydraulic circuit, coolant, oil circuit, fuel circuit; or an electrical or electronic assembly or subassembly such as on-board electronics, battery circuit, headlight system, or power supply to a specific component. In particular, machine parts already equipped with sensors may be used. However, the invention is not limited in this respect, as it provides for the use of sensors to measure measurable parameters even when the machine is not equipped with them by default.

[0015] A measurable parameter is defined as any physical quantity that can be measured by a sensor.

[0016] The parts and parameters to be measured are chosen based on several criticality factors. A Reliability, Maintainability, Availability, and Safety (RMDS) approach can be used, for example, to select the parts and parameters, but other factors, such as machine downtime in case of failure or the cost of replacement parts, can also be considered. Advantageously, considering the entire functional chain limits the number of maintenance sensors required, thus meeting the necessary reliability needs. Ideally, sensors already installed on the machine should be used.

[0017] The digital twin replicates all or part of the machine. In particular, the digital twin can be, in a simplified aspect of the process of the invention, a digital model limited to the part of the machine to be monitored, but it is also envisaged according to the invention that the digital twin may be more complex, in particular a digital model including several parts of the machine to be monitored or even all the parts that can be monitored on the machine.

[0018] Digital twin modeling is a multi-physics numerical simulation that can be performed using software suites, such as, but not limited to, Altair®, SIMLab®, Roamaxtec®, Matlab Simulink®, or Simscape®. The granularity of the digital twin will depend on the level of maintenance required. The set of predetermined incidents in steps e) and f) includes all or some of the events that can influence the maintenance of a given type of machine, including, but not limited to, failures, shocks, or simply wear and tear.

[0019] The method of the invention thus allows, in a first training phase on a training machine, the construction of a digital twin already incorporating learning of potential maintenance incidents; then, in a second phase, the determination of the sensors necessary for detecting these incidents; and finally, in a third phase, the evolution of this digital twin with data from the operating data of the machines in the machine park, so that the digital twin closely resembles the machines whose maintenance is to be performed. It is then simply a matter of applying the data collected from the machines in operation to the digital twin to carry out predictive maintenance.After mathematical processing, it will be possible to inject the maintenance sensor data from a specific machine in the machine park into the digital twin, and use the corresponding data from the digital twin to deduce a degree of wear of the specific machine and, if necessary, program predictive maintenance.

[0020] The prediction can also be directly integrated into each machine, with feedback such as Remaining Useful Life (RUL) being indicated by the vehicle.

[0021] In the final stage of the process,

[0022] A fleet of machines equipped with predictive maintenance functions should improve the management of machine maintenance and usage.

[0023] The accuracy of forecasts made on monitoring functions has a significant impact on machine maintenance management.

[0024] To estimate the RUL (RUL - acronym for Remaining Useful Life) of a predictive function, there are two distinct approaches.

[0025] A first, purely theoretical approach is based on a statistical approach linked to multi-physics modeling. This approach takes into account the theoretical mean time between failures (MTBF) as well as the machine usage profiles to determine their remaining lifespan.

[0026] A second approach is based on a practical approach directly linked to wear sensors or indicators and the multi-physics model.

[0027] This second approach works in the background of the first approach and takes priority over the first approach.

[0028] According to the first approach, to estimate the RUL of the predictive function, dynamic simulation based on lifetime models of the Weibull distribution type is used. These mathematical distributions are used to model the lifetimes and predictions made by the monitoring functions, the HUMS (Health and Usage Monitoring System).

[0029] To model the effects of predictive maintenance, it is necessary to model the predictions made by HUMS.

[0030] Event generation is based on lifetime: the Weibull distribution has two parameters, the scale factor (η) and the shape factor (β). For example, η=1000 and β=2 could be adopted. Other values ​​can be chosen, depending on the well-known practice of the person in the field.

[0031] To model the quality of forecasts, two parameters are used whose values ​​are randomly drawn (uniform model): an error rate and a forecast accuracy. These two parameters are expressed as percentages and are applied to the lifetime (LV) generated by the Weibull distributions.

[0032] Thus, a forecast error rate of 10% means that in 10% of cases, the forecast is greater than the actual ddv, and an accuracy relative to the actual ddv of 20% means that the error between the value of the actual ddv and the forecast is 20%.

[0033] The characterization of lifespan laws and their prediction relies primarily on the statistical analysis of data. Initially, the characterization of the predictive laws is based on assumptions and knowledge of simulation models. These laws are subsequently adjusted as experience data is accumulated. This experience is gathered in an accelerated phase (due to the frugal nature of the invention's process) during the functional and dysfunctional testing phase.

[0034] This approach makes it possible to estimate, through the simulation of analysis scenarios, the performance that monitoring functions must have, so that predictive maintenance improves machine availability, and more generally facilitates the management of the machine fleet.

[0035] The second, practical approach is directly linked to the functional and dysfunctional analysis phase of the frugal process according to the invention and allows, through sensor analysis (for example, wear indicators), the definition of 3 states: First state: everything is OK (no detection at the sensor level). In this case, the prediction is handled by the first statistical approach; second state: beginning of wear (detected by the sensors). The wear resistance will be correlated to the machine's usage profile (number of openings, temperature, vibration, etc.) and the knowledge of the expert systems. This wear resistance will be defined by a practical approach and correlated with multiphysics models. The prediction is then handled by the wear indicators with a simple threshold-type approach or a comparative analysis of the "signal signature" type. The wear resistance will be calculated directly in relation to the use case; third state: confirmed wear (detected by the sensors). In this case, replacement is necessary to avoid the risk of failure.

[0036] The method according to the present invention therefore has the advantage, compared to existing methods, of processing less data and conforming to a realistic model of the machine being maintained. It can thus be applied to fleets of vehicles that are used infrequently or have an irregular usage profile with significant variations, such as a fleet of military vehicles, but it can also be applied to fleets of civilian vehicles, which have both greater and more regular usage.

[0037] According to the invention, by machine is meant any dynamic system that can be modeled, the modeling being the association of a control system (electronic board associated with a control software) which controls an actuator (electric, hydraulic, pneumatic...), which actuator is at the initiative of the movement of a kinematic chain (mechanical, mechatronic, hydraulic, pneumatic, electrical...).

[0038] By way of non-limiting examples, a machine according to the invention can be a vehicle, in particular a military vehicle, an automated chain, a conveyor, an analytical automaton (medical, ...), a distribution system (coffee machine, ...), a power plant, a hydraulic power plant, an energy production device (tidal turbine, wind turbine...), machine tools... The invention is applicable to many types of machines.

[0039] According to one embodiment, the process further comprises, before step i), the steps of: i01) equip a restricted group of machines in the machine park with the maintenance sensor group, i02) periodically collect measurement data from the maintenance sensor group equipping each machine in the restricted group of machines in the machine park to enrich the digital twin.

[0040] This optional step allows, in a transitional phase, for further improvement of the digital twin by injecting measurement data from the maintenance sensors of the restricted group of machines into the digital twin, before the deployment of the process in the entire fleet of machines.

[0041] Here again, the data transfer from the restricted group of machines to the digital twin will be done at high speed to maximize the amount of data brought to the digital twin.

[0042] According to one embodiment, the restricted group of machines comprises between 10 and 100 machines, preferably between 20 and 50 machines, and more preferably 30 machines. However, the invention is not limited in this respect, and those skilled in the art will understand that the invention is applicable to a machine park larger than 100 machines.

[0043] According to one embodiment, the predetermined degree of wear corresponds to the degree of wear of a machine of the given type coming out of the factory.

[0044] This degree of wear can be obtained either from a new machine straight from the factory or from a refurbished machine. However, the invention is not limited to such a degree of wear; a higher degree of wear can be taken into account when modeling the digital twin. Thus, for a fleet of older machines, a minimum degree of wear can be determined, not corresponding to the factory acceptance wear level, but to the minimum wear level of the machine fleet, for modeling the digital twin. The method of the present invention is therefore intended to be applied to both new and older machine fleets. According to one embodiment, the predetermined probability is between 70% and 100%. %, preferably between 85% and 100 %,and is preferred in a way greater than 90%.

[0045] The predetermined probability will depend on the part being monitored, particularly its criticality in machine maintenance and the repair cost, including any spare parts. Advantageously, this predetermined probability can be modified based on the behavior of the digital twin model.

[0046] According to one embodiment, the measurable parameters are at least one of the following: temperature, vibration, pressure, deformation, current consumption, fluid consumption, resistivity, angle, acceleration, degree of pollution, humidity level and degree of wear.

[0047] According to one embodiment, in step j), measurement data from the group of maintenance sensors equipping each machine in the machine park are collected to enrich the digital twin at a frequency of between 1 and 90 days, preferably between 1 and 30 days, more preferably between 1 and 15 days.

[0048] The frequency of data collection from maintenance sensors for injection into the digital twin depends on machine usage: it must be correlated to the average usage rate of machines in the machine park.

[0049] Alternatively, it is of course also possible to transmit measurement data from maintenance sensors each time the machine is used, or after a predetermined number of uses. Each machine would then incorporate a computing device including a counter that counts machine uses and triggers the transmission of measurement data from the maintenance sensors to the digital twin after a predetermined number of uses. Each transmission of measurement data from the maintenance sensors to the digital twin resets the counter.

[0050] Sensor data collection can be performed locally or remotely, on demand or at regular intervals, with the sensors integrating or being associated, where appropriate, with memory and telecommunication capabilities. A human-machine interface can be provided for data collection, either on the machine or remotely, preferably with a means of data processing, for example, for formatting the data before injection into the digital twin.

[0051] Thus, sensors can be made up of or associated with one or more data processing devices, such as microcontrollers, microprocessors, processors, digital signal processors, or even FPGAs or ASICs, possibly associated with memory and / or telecommunications components, to enable the recording of measurement data.

[0052] According to one embodiment, the machine of a given type is a vehicle, in particular a military vehicle.

[0053] According to one embodiment, the vehicle component is at least one of the following: a weapons system, an air conditioning and heating system, a door, hatch, or window opening system, a rear ramp opening / closing system, a running gear, tires, tracks, an engine, on-board electronics, a CAN bus, a power distribution system, a gearbox, a braking system, a battery, an audible warning system, a visual warning system, a fluid level sensor, or a hydraulic circuit. The invention is not limited in this respect and can be applied to any subassembly comprising measurable parameters.

[0054] The invention also relates to a predictive maintenance method for a heterogeneous fleet of machines having machines of different types, the method comprising the implementation of the method as described above for each type of machine in the fleet of machines.

[0055] This gives us a digital twin for each type of machine in the fleet.

[0056] Advantageously, the process of the invention can be linked to the integrated management software package (ERP in English) of the company operating the machine park for the automation of order supply.

[0057] The process of the invention also makes it possible to plan maintenance operations, to smooth out the maintenance load, and to forecast the necessary stock of spare parts.

[0058] The method of the invention can be implemented centrally, with a single digital twin that serves for the machine park, or remotely on each machine, which then includes the necessary means in terms of computing, memory and means of communication to send its data to a server which updates the digital twin, receive an up-to-date copy of the digital twin and make a maintenance prediction.

[0059] The invention also relates to a computer program product comprising instructions which, when the program is executed by a computer, cause the computer to implement a process as described above.

[0060] The invention also relates to a computer-readable medium comprising instructions which, when the program is executed by a computer, lead the computer to implement a process as described above.

[0061] To better illustrate the object of the present invention, preferred embodiments are described below, by way of illustration and not limitation, together with the accompanying drawings. These drawings show: [ Fig. 1 [ ] is a block diagram of the process according to one embodiment of the invention. ] Fig. 2 ] is a block diagram of the process according to another embodiment of the invention.

[0062] If we refer to the Figure 1 , we can see that we have represented a predictive maintenance process according to the present invention.

[0063] To illustrate the invention with a specific example, the method will be implemented on a vehicle-type machine, in particular a military vehicle. Those skilled in the art will understand that this example is purely illustrative and in no way limits the application of the invention to vehicles; the invention can be applied to any type of machine, and the vehicle example does not introduce any limitation that would prevent the invention from being applicable to other machines.

[0064] The process includes the following steps: Step a: identify, for the given vehicle type, at least one vehicle part to be monitored that is affected by vehicle wear.

[0065] In this step, each vehicle component to be monitored is selected based on the intended predictive maintenance objective, by determining which vehicle parts to monitor. Factors such as component criticality, downtime in case of failure, or replacement part cost can be taken into account. The component can be a single element, such as a tire or the starter, or comprise several components, such as a wheel or the engine, or even a functional chain, such as the drivetrain, the heating / air conditioning system, the complete electrical system, or several interacting parts, such as the transmission and the chassis. It is, of course, possible to select multiple components.

[0066] Step b: identify all measurable parameters of at least one part of the vehicle to be monitored.

[0067] In this step, once each part to be monitored has been identified, the measurable parameters relating to each part to be identified are listed.

[0068] Step c: model a digital twin of the given vehicle type including a digital model of at least one vehicle part to be monitored.

[0069] In this step, the digital twin of the given vehicle type is modeled, and includes a digital model of each part to be monitored.

[0070] The following steps are carried out on a so-called training vehicle of the given type having a predetermined degree of wear. This degree of wear will preferably be a factory-recommended (new vehicle) wear level.

[0071] Step d: have sensors available to measure each measurable parameter of the set of measurable parameters of at least one part of the vehicle to be monitored.

[0072] This step involves over-instrumenting the area to be measured using an instrumentation plan to ensure that all measurable parameters are available, particularly those that can detect wear or deviations from normal operation. The existing sensors on the vehicle must be taken into account, and any necessary external sensors must be added. The instrumentation plan allows for the characterization and specification of the required type and characteristics of the sensors to be implemented (amplitude, bandwidth, accuracy, repeatability, etc.). Several instrumentation plans with sensors having different characteristics can be implemented to optimize the measurement of the parameters of the monitored area. Ideally, one sensor should be used to monitor several components of the area being monitored.Ideally, the information used to define the remaining useful life should come from several measurable parameters in order to make the prediction more robust.

[0073] Step e: Simulate a set of predetermined incidents on the training vehicle and record the measurement data from the sensors for each predetermined incident.

[0074] Predetermined incidents can be breakdowns, known because they have already been encountered, or unknown, shocks or wear and tear (tire, corrosion...).

[0075] Step f: enrich the digital twin with sensor measurement data for each predetermined incident in the set of predetermined incidents.

[0076] The measurement data from the training vehicle are retrieved and injected into the digital twin so that its behavior is as close as possible to that of a real vehicle in the fleet of vehicles for which predictive maintenance is to be ensured.

[0077] The artificial intelligence engine, or learning algorithm, used to process this data from the digital twin can include, but not exclusively, libraries used in data science, including Python ® language libraries used in artificial intelligence, including but not limited to Keras, for creating neural network models used with the TensorFlow library to optimize calculations, PyTorch, also for creating neural network models, or Scikit-Learn, implementing parameterized algorithms, or in R language with the Caret package, for example, which allows calling many machine learning methods by offering a unified interface and which includes various utility functions.

[0078] The data will preferably be stored in a data lake, which is a system or directory for storing raw data. A data lake can store various types of data and allows for the storage of very large quantities of raw data in its native format for an indefinite period. This storage method facilitates the coexistence of different data schemas and formats.

[0079] Several learning opportunities are envisaged, preferably in a complementary manner: Supervised learning: based on a representative model for training, it consists of learning on an annotated database; unsupervised learning: the right representative model is sought by machine learning, without having annotated data.

[0080] At the post-processing level, it is therefore possible to bring out new fault correlations not imagined when setting up the digital twin of the part to be monitored: for example, driving with the front right brake partially blocked causes wear on the gearbox.

[0081] Normative concepts (notably FMDS) related to equipment aging can be integrated into prediction laws to improve them.

[0082] We can see that the training dataset is built progressively from data from the training vehicle.

[0083] Step g: define, for each simulated predetermined incident in the set of predetermined incidents, a detection threshold.

[0084] Here again, the detection threshold will depend on the criticality of the component being monitored. For a highly critical component, the detection threshold can be very precise, while for a less critical component, it can be less precise. The detection threshold can also be positioned closer to or further from the expected failure point, depending on the anticipated repair time and downtime.

[0085] Step h: identify among the sensors, for each simulated predetermined incident of the set of predetermined incidents, a group of so-called maintenance sensors allowing the detection threshold to be detected with a predetermined probability.

[0086] This step is a data processing step that allows limiting the number of sensors needed to detect the desired detection thresholds.

[0087] From the next step, maintenance sensors are installed on all vehicles in the fleet for fleet-wide deployment of the process. Step i: Equip all vehicles in the fleet with the group of maintenance sensors.

[0088] Step j: Periodically collect measurement data from the group of maintenance sensors equipping each vehicle in the fleet to enrich the digital twin. This step allows for long-term improvement of the digital twin. Once again, the training data for the digital twin is built up incrementally, this time across all vehicles in the fleet.

[0089] Step k: For each vehicle in the fleet, periodically record the measurement data from the group of maintenance sensors equipping it and apply the measurement data to the digital twin to deduce a maintenance need.

[0090] This is the step in the process that enables predictive maintenance. Each vehicle in the fleet is compared to its digital twin to deduce its remaining useful life (RUL) and therefore the probability of a breakdown occurring, thus triggering predictive maintenance. The prediction step can be performed on the vehicle itself, in which case it must have the necessary computing, memory, and communication resources. The digital twin is integrated into the computing resources and updated as the fleet's digital twin evolves via communication channels. These channels can therefore be used both to retrieve data from a centralized digital twin of the fleet and to update a copy of the centralized digital twin of the fleet on the vehicle.The prediction step can also be performed remotely on a computer that manages the digital twin of the vehicle fleet. This will be particularly useful when it is complicated to implement a copy of the centralized digital twin on each vehicle.

[0091] Given the long learning period required in the later stages of the invention process, it is understood that the process will become increasingly accurate as the deployment time increases over a fleet of vehicles.

[0092] It should be noted that a digital twin deployed on a fleet of vehicles of a given type according to the method of the invention, regardless of its state of maturity, can be deployed on another fleet of vehicles of the same type, without necessarily having to reproduce steps a to h, with minimal adaptation.

[0093] In the embodiment shown in the Figure 2 , we find the same steps as in the steps described in Figure 1and which will not be described in more detail here. Steps i01 and i02 are introduced before step i, to improve the digital twin before the deployment and operation of the process across the entire vehicle fleet: Step i01: Equip a limited group of vehicles in the fleet with the maintenance sensor group. Step i02: Periodically collect measurement data from the maintenance sensor group equipping each vehicle in the limited group of vehicles in the fleet to enrich the digital twin.

[0094] These additional steps make the digital twin more robust before its deployment across the entire fleet of vehicles.

[0095] In practical terms, the method of the invention can be implemented using a HUMS kit (Health and Usage Monitoring System), which is an electronic unit that records the existing CAN bus, conditions and records signals from added sensors, and executes a predictive algorithm. In this case, the centralized digital twin of the vehicle fleet will be updated periodically, and the prediction will be performed remotely on the vehicle.

[0096] As mentioned above, the invention is not limited to vehicle fleets, and can be applied to all fleets of machines, and more generally to all fleets of dynamic systems that can be modeled, the modeling being the association of a control system (electronic card associated with a control software) which controls an actuator (electric, hydraulic, pneumatic...), which actuator is at the initiative of the movement of a kinematic chain (mechanical, mechatronic, hydraulic, pneumatic, electrical...).

[0097] By way of non-limiting examples, a machine according to the invention may be a vehicle, in particular a military vehicle, an automated assembly line, a conveyor, an automated analyzer (medical, etc.), a dispensing system (coffee machine, etc.), a power plant, a hydroelectric plant, an energy production device (tidal turbine, wind turbine, etc.), machine tools, etc.

Claims

1. A predictive maintenance method for a homogeneous machine fleet, the machines in the machine fleet being of the same given type of machine, the method comprising the steps consisting in: a) identifying, for the given type of machine, at least one machine part to be monitored which is affected by wear of the machine, b) identifying the set of measurable parameters of the at least one machine part to be monitored, c) modeling a digital twin of the given machine type comprising a digital model of the at least one machine part to be monitored, and, on a so-called training machine of the given machine type having a predetermined degree of wear: d) providing sensors to measure each measurable parameter of the set of measurable parameters of the at least one machine part to be monitored, e) simulating a set of predetermined incidents on the training machine and recording the measurement data from the sensors for each predetermined incident, f) enriching the digital twin with the measurement data from the sensors for each predetermined incident of the set of predetermined incidents, g) defining, for each simulated predetermined incident of the set of predetermined incidents, a detection threshold, h) identifying among the sensors, for each simulated predetermined incident of the set of predetermined incidents, a group of so-called maintenance sensors making it possible to detect the detection threshold with a predetermined probability, i) equipping all the machines in the machine fleet with the group of maintenance sensors, j) periodically recording the measurement data from the group of maintenance sensors equipping each machine in the machine fleet to enrich the digital twin, k) for each machine in the machine fleet, periodically recording the measurement data from the group of maintenance sensors equipping it and applying the measurement data to the digital twin to deduce a maintenance need.

2. The predictive maintenance method according to claim 1, characterized in that it further comprises, before step i), the steps consisting in: i01) equipping a restricted group of machines from the machine fleet with the group of maintenance sensors, i02) periodically recording the measurement data from the group of maintenance sensors equipping each machine in the restricted group of machines in the machine fleet to enrich the digital twin.

3. The predictive maintenance method according to claim 2, characterized in that the restricted group of machines comprises between 10 and 100 machines, preferably between 20 and 50 machines, more preferably 30 machines.

4. The predictive maintenance method according to one of claims 1 to 3, characterized in that the predetermined degree of wear corresponds to the degree of wear of a machine of the given type ex factory.

5. The predictive maintenance method according to one of claims 1 to 4, characterized in that the predetermined probability is between 70% and 100%, preferably between 85% and 100%, and is more preferably greater than 90%.

6. The predictive maintenance method according to one of claims 1 to 5, characterized in that the measurable parameters are at least one of temperature, vibration, pressure, deformation, current consumption, fluid consumption, resistivity, angle, acceleration, degree of pollution, humidity level and degree of wear.

7. The predictive maintenance method according to one of claims 1 to 6, characterized in that in step j), the measurement data from the group of maintenance sensors equipping each machine in the machine fleet are recorded to enrich the digital twin at a frequency of between 1 and 90 days, preferably between 1 and 30 days, more preferably between 1 and 15 days.

8. The predictive maintenance method according to one of claims 1 to 7, characterized in that the machine of the given type is a vehicle, in particular a military vehicle.

9. The predictive maintenance method according to claim 8, characterized in that the vehicle part is at least one of a weapons system, the air conditioning and heating system, a door, hatch or window opening system, a rear ramp opening / closing system, a running gear, tires, tracks, the engine, on-board electronics, CAN bus, power distribution, gearbox, braking system, battery, an audible warning system, a light warning system, a liquid level, a hydraulic circuit.

10. A predictive maintenance method for a heterogeneous machine fleet having machines of different machine types, characterized in that it comprises the implementation of the method according to one of claims 1 to 9 for each type of machine in the machine fleet.

11. A computer program product comprising instructions which, when the program is executed by a computer, cause the computer to implement a method according to one of claims 1 to 10.

12. A computer-readable medium comprising instructions which, when the program is executed by a computer, cause the computer to implement a method according to one of claims 1 to 10.