Method and device for predicting an upcoming failure of a vehicle component

Abstract

Method and device for predicting the imminent failure of a component of a vehicle. The method comprises receiving (11) first data representative of the first characteristics of the vehicle and predicting the next failure by inputting this data into a component failure prediction model.The component failure prediction model was learned from training data representative of similar characteristics of a set of vehicles, and comprises a sequence of classifiers (12), each classifier in the order of the sequence having respectively increasing specialization in determining a failure such that one classifier in the sequence determines the failure at one level of specialization and a subsequent classifier in the sequence determines the same failure at an increased level of specialization compared to the first classifier, the next predicted failure corresponding to the failure determined by the last classifier (12n) in the sequence. Figure 1.

Description

Title of the invention: Method and device for predicting an upcoming failure of a vehicle component Technical field of the invention

[0001] The present application generally relates to a method for predicting the failure of a component of a vehicle. The present application also relates to a method and an apparatus for controlling an on-board system of a vehicle based on a prediction of an upcoming failure, for example, an autonomous or semi-autonomous vehicle. The present invention also relates to a method for learning a component failure prediction model. State of the art

[0002] The purpose of this section is to introduce the reader to various aspects of the art, which may be related to various aspects of at least one exemplary embodiment of the present application that is described and / or claimed below. This discussion is deemed helpful to provide the reader with background information that will facilitate a better understanding of the various aspects of the present application.

[0003] Many modern vehicles are equipped with advanced driver assistance systems (AD AS). AD AS are passive and active systems designed to eliminate human error in driving all types of vehicles and / or to improve driving comfort, for example. ADAS systems use advanced technologies to assist the driver while driving, thereby improving performance. ADAS systems use a combination of sensor technologies to perceive the environment around a vehicle and its behavior, and then provide information to the driver or act on vehicle systems called active on-board systems by configuring their parameters or controlling them.

[0004] There are several levels of ADAS, acting in particular on the active on-board systems of the vehicle, for example adaptive cruise control, suspension control or power steering, by setting parameters to control certain actuators of each active on-board system. Some ADAS allow the vehicle to be autonomous, defining a level of autonomy. The level of autonomy of a vehicle, for example, varies from 0 to 5 (0 for a vehicle without autonomy and whose driving is under the total supervision of the driver, and 5 for a fully autonomous vehicle).

[0005] The 5 levels of autonomy used by the NHTSA (National Highway Traffic Safety Administration, the United States federal agency responsible for road safety) are as follows: - level 0: no automation, the driver having total control of the main vehicle functions (engine, accelerator pedal, steering, brakes); - level 1: driving assistance, automation is active for certain vehicle functions, the driver retaining overall control of the vehicle's driving; Cruise control is included in this level, as are other aids such as ABS (anti-lock braking system) or ESP (programmed electro-stabilizer); - level 2: automation of combined functions, where the control of at least two main functions is combined in the automation to replace the driver in certain situations; For example, adaptive cruise control combined with lane centering allows a vehicle to be classified as level 2, as does automatic parking assistance; - level 3: limited autonomous driving, where the driver can hand over complete control of the vehicle to the automated system, which will then be responsible for safety-critical functions; autonomous driving can only take place in certain environmental and traffic conditions (on motorways, for example); - level 4: fully autonomous driving under certain conditions, with the vehicle designed to perform all safety-critical functions on its own over a complete journey; the driver provides a destination or navigation instructions, but is not required to make himself available to take control of the vehicle; - level 5: fully autonomous driving without driving assistance in all circumstances.

[0006] The International Organization of Motor Vehicle Manufacturers classification is similar to that listed above, except that it comprises 6 levels, with level 3 in the American classification being divided into 2 levels in the International Organization of Motor Vehicle Manufacturers classification.

[0007] The AD AS embedded in a vehicle are supplied with data obtained from one or more on-board sensors such as, for example: • position and motion sensors such as speed, acceleration, gyroscope and tilt sensors that measure the position, speed and orientation of the vehicle, • proximity sensors that detect nearby objects to aid parking and collision avoidance, such as parking sensors and radars, • vision sensors such as cameras that provide visual information for driver assistance systems, traffic sign recognition, lane and pedestrian detection, • environmental sensors such as temperature, humidity, atmospheric pressure and air quality sensors, used to regulate climate control systems automation and monitor environmental conditions, • safety sensors such as airbag sensors that detect collisions and trigger safety devices, or tire pressure sensors, • engine and propulsion system sensors, including engine temperature, oil pressure and airflow sensors, which monitor and regulate engine performance, • comfort sensors such as ambient light and rain sensors to regulate lighting and automatic windshield wipers, and • connectivity sensors enabling integration with external devices, such as Bluetooth®, GPS and Wi-Fi® sensors for navigation and smartphone connectivity.

[0008] Furthermore, to improve vehicle reliability and passenger safety, vehicles require regular maintenance. Such operations follow a schedule published by a manufacturer, for example, in which it recommends how often certain mechanical parts should be replaced. These intervals are estimated based on average wear rates, for example, but do not necessarily take into account the actual condition and use of a particular vehicle. As a result, maintenance operations are performed more frequently, sometimes more often than necessary, to avoid any risks, making vehicle maintenance expensive. In addition, it may happen that the vehicle suffers unusual damage due to certain driving conditions. Unfortunately, these difficult conditions can only be identified by the users of the vehicle, and this depends on their own sensitivity.It is therefore difficult to determine the actual state of wear of a vehicle's components without specific analysis. Summary of the invention

[0009] The following section presents a simplified summary of at least one exemplary embodiment to provide a basic understanding of certain aspects of the present application. This summary is not an exhaustive overview of an exemplary embodiment. It is not intended to identify critical elements of an embodiment. The following summary merely presents certain aspects of at least one of the exemplary embodiments in a simplified form as a prelude to the more detailed description provided elsewhere in the document.

[0010] According to a first aspect of the present application, the invention relates to a method for predicting an upcoming failure of a component of a first vehicle, the method being implemented by a processor and comprising the following steps: - receiving first data representative of first characteristics of the first vehicle; - predicting said next failure by inputting said first data into a component failure prediction model, said component failure prediction model being learned in a learning phase from learning data representative of second characteristics of a set of second vehicles, the second characteristics being homogeneous with the first characteristics, and said component failure prediction model comprising a sequence of classifiers, each classifier of the classifiers sequentially traversed in the sequence having respectively an increasing specialization, determining a potential failure such that a classifier in the sequence determines the potential failure at one level of specialization and a following classifier in the sequence determines the same potential failure at an increased level of specialization compared to the first classifier, the next predicted failure corresponding to the potential failure determined by the last classifier in the sequence.

[0011] Such a method makes it possible to predict a component failure both quickly and accurately, based on a component failure prediction model learned from known data representative of similar component failures observed on a wide range of vehicles. Predicting a component failure thus makes it possible, for example, to plan a preventive maintenance operation, to set up a navigation system to guide a user to a workshop, or even to control the vehicle if it is autonomous, so that it goes to a garage or workshop to carry out the maintenance operation.

[0012] In an exemplary embodiment, said first characteristics comprising a first historical set of loads applied to said component of the first vehicle over the lifetime of said component.

[0013] In another embodiment, said second characteristics comprise, for each second vehicle of said set of second vehicles, a historical set of maintenance data of the second vehicle associated with a second historical set of loads applied to the component of a second vehicle, said component of the second vehicle corresponding to the component of said first vehicle.

[0014] In another embodiment, said learning phase comprises the following steps: - receiving second data (21b) representative of an overall target rate of true positives, of an overall target rate of false positives, and either of the maximum authorized number of levels in the sequence (N), or of at least a target rate of local true positives and of a target local rate of false positives. - determine (221) data representative of the maximum authorized number of levels (N) in the sequence, a target local true positive rate and a target local false positive rate from these second data (21b).

[0015] In another embodiment, said learning phase comprises the following steps: - training at each level of the next sequence a new classifier until the component failure prediction model comprising classifiers trained so far up to the current level of the classifier generates: • a local true positive rate lower than said target local true positive rate, or • a local false positive rate higher than said target local false positive rate, or • jointly an overall true positive rate higher than said target overall true positive rate, and an overall false positive rate lower than said target overall false positive rate; - Obtaining the decision of the trained component failure prediction model or a diagnostic output to determine why the model could not be trained.

[0016] In another exemplary embodiment, said component failure prediction model implements at each sequence level and algorithm from a set of algorithms comprising at least: - a logistic regression, - Bayesian modeling, - a survival analysis, - a decision tree and a deep neural network.

[0017] In another exemplary embodiment, the method further comprises: - the prediction of a maintenance operation of said first vehicle based on said next failure, - the control of an on-board system of said first vehicle as a function of said maintenance operation.

[0018] According to a second aspect of the present application, there is an apparatus for predicting the next failure of a first component of a vehicle, in which the apparatus comprises a memory associated with at least one processor configured to implement the method in accordance with the first aspect of the present application.

[0019] According to a third aspect of the present application, there is provided a vehicle, for example of the automobile type, comprising a device as described above according to the second aspect of the present invention.

[0020] According to a fourth aspect of the present application, there is provided a computer program product comprising instructions which, when the program is executed by one or more processors, causes one or more processors to execute a method according to the first aspect of the present application.

[0021] According to a fifth aspect of the present application, there is provided a non-transitory storage medium containing program code instructions for executing a method according to the first aspect of the present application.

[0022] The specific nature of at least one of the exemplary embodiments as well as other objects, advantages, features and uses of at least one of the exemplary embodiments will become apparent from the following description of examples taken in conjunction with the accompanying drawings. Brief description of the figures

[0023] It will now be necessary to refer, by way of example, to the attached drawings, which show exemplary embodiments of the present application and in which:

[0024] [Fig.l] illustrates a flowchart of the different steps of a process for predicting the imminent failure of a component of a vehicle, according to a particular and non-limiting embodiment of the present invention;

[0025] [Fig.2] illustrates a flowchart of the different stages of a process training the component failure prediction model used to predict the prediction of an upcoming failure of a component of a vehicle of [Fig.l], according to a particular and non-limiting example of the present invention; and

[0026] [Fig.3] illustrates a functional diagram of an example apparatus configured for predicting the imminent failure of a component of a vehicle, in accordance with at least one exemplary embodiment. Detailed description of the invention

[0027] At least one of the exemplary embodiments is described in more detail below with reference to the accompanying figures, in which examples of at least one of the exemplary embodiments are illustrated. An exemplary embodiment, however, may be embodied in numerous alternative forms and should not be construed as limited to the examples set forth herein. Accordingly, it should be understood that there is no intention to limit the exemplary embodiments to the particular forms disclosed. Rather, the disclosure is intended to cover all modifications, equivalents, and alternatives within the spirit and scope of the present application.

[0028] The present invention discloses methods and apparatus for predicting the imminent failure of a component of a vehicle. The method includes receiving first data representative of first characteristics of the vehicle and predicting the next failure by inputting this data into a component failure prediction model. The component failure prediction model component failure was learned from training data representative of the secondary characteristics of a set of vehicles, the second characteristics being homogeneous with the first characteristics, and comprises a sequence of classifiers, each classifier sequentially traversed in the sequence having respectively increasing specialization in determining a potential failure such that one classifier in the sequence determines the potential failure at one level of specialization and another The classifier in the sequence determines the same potential failure at an increased level of specialization compared to the first classifier, the next predicted failure corresponding to the potential failure determined by the last classifier in the sequence. The failure or predicted failure corresponds, for example, to the failure of a component associated with a deadline.For example, the right front wishbone will fail in 30 days, or the left front bearing will reach the end of its life in 2000 km.

[0029] [Fig.l] illustrates a flowchart 1 of the different steps of a process for predicting the next failure of a component of a vehicle, according to a particular and non-limiting embodiment of the present invention, for example of a component of a first vehicle.

[0030] The first vehicle corresponds for example to a vehicle with a thermal engine, with electric motor(s) or even to a hybrid vehicle with a thermal engine and one or more electric motors, for example a rechargeable hybrid vehicle. The first vehicle therefore corresponds for example to a land vehicle, for example a car, a truck, a bus, or a motorcycle.

[0031] The first vehicle may be equipped with various advanced driver assistance systems (ADAS) that are configured to assist the driver of the first vehicle in controlling the first vehicle. The first vehicle may be a semi-autonomous vehicle or an autonomous vehicle, for example, a vehicle traveling with an autonomy level that is, for example, greater than or equal to 2 in an autonomy range of between 1 and 5, where level 1 corresponds to a vehicle with a minimum autonomy level and level 5 corresponds to a vehicle with a maximum autonomy level, for example, a fully autonomous vehicle.

[0032] The first vehicle is equipped with a detection module comprising a first set of detection devices, the objective of which is to process one or more time-stamped input signals, periodic or not, and not necessarily aligned at the origin in the time or frequency domain. The first set of detection devices comprises, for example, at least one sensor from among a: - pressure sensor, - force sensor, - strain sensor with strain gauge, - inertial sensors, - acceleration sensors, - vibration sensors, - wheel speed sensors, - steering wheel sensor, - radar, - lidar, - laser scanners, - cameras, - noise sensors, - rain sensor, - internal thermometer, - external thermometer, and - a weather station. Please note that this list is not exhaustive.

[0033] Such sensors can be used to determine both the environmental conditions in which the first vehicle is traveling and the intrinsic characteristics of the first vehicle.

[0034] The detection module periodically analyzes the recent part of the available signals, preprocesses them as needed, for example by sampling, filtering, imputation of missing values or removal of outliers, and aligns them, for example by interpolation, in order to obtain regularly spaced multidimensional data points, called time series.

[0035] According to an exemplary embodiment, the first vehicle is equipped with a communication unit, corresponding for example to a communication module of the TCU (Telematic Control Unit), the ATB (Autonomous Telematic Box) or the RFSB (Radio Frequency Service Box). Such a communication unit is advantageously connected to one or more antennas to, for example, transmit and / or receive data to and / or from a remote device, for example a cloud data center, or a mobile communication device, such as a smartphone or a tablet, via a wireless connection, according to OTA technology (Over The Air) for example.Wireless communication between the first vehicle and the remote device is established, for example, via a network infrastructure comprising, for example, one or more relay antennas or base stations, or in a direct communication mode (for example, when the mobile communication device is located at a . distance from the first vehicle 10 allowing such a mode of communication, for example at a distance of less than 10, 20 or 50 m). The wireless connection is based for example on one or more wireless communication protocols such as Bluetooth®, Wi-Fi® (based on IEEE 802.11), LTE (from the English "Long-Term Evolution", name of a mobile telephony standard), LTE-Advanced (name of a mobile telephony standard), 3GPP (from the English "3rd Generation Partnership Project", name of a mobile telephony standard) of the fourth or fifth generation, called 3GPP 4G or 5G, satellite communication such as Starlink®.

[0036] According to an exemplary embodiment, the first vehicle 10 is equipped with a second set of sensing devices including, for example, a thermometer, a pressure sensor, a force sensor or a strain sensor with strain gauge. The second set of sensors makes it possible to know the intrinsic parameters of the vehicle or to capture the physical reactions of the vehicle or its components in its environment. Please note that a sensor from the first set of sensing devices can also be included in the second set of sensing devices, since the recorded data can be used to know both the intrinsic parameters and the environmental conditions.

[0037] According to an exemplary embodiment, the first vehicle 10 is equipped with a navigation and geolocation system, also called a GNSS system (from the English “Geolocation and Navigation by a Satellite System”), for example a GPS type satellite positioning system. A geolocation signal is available to provide information on the momentary physical position of the vehicle.

[0038] The on-board systems, such as the communication unit, the ADAS and the road detection module, are each controlled by one or more computers, for example computers (electronic control unit). These computers form, for example, a multiplexed architecture for the realization of various services useful for the proper functioning of the first vehicle 10 and for assisting the driver and / or passengers of the vehicle in the control of the first vehicle 10 via the control of the on-board systems of the first vehicle 10. The computers communicate and exchange data with each other via one or more computer buses, for example a communication bus known as CAN, CAN FD, FlexRay (according to the ISO 17458 standard), LIN or Ethernet (according to the ISO / IEC 802-3 standard). A method for predicting the next failure of a component of a first vehicle is, for example, implemented by an ADAS processor of the first vehicle.

[0039] Firstly, the first data 11 are received. The first data are representative of the first characteristics of the first vehicle. The first data are representative of the first characteristics of the first vehicle. It is obtained from the time series produced by the detection module of the first vehicle. For example, it can be obtained from this time series in raw form or by temporal aggregation or by any other appropriate transformation.

[0040] In an exemplary embodiment, the first characteristics include a first historical set of loads applied to the component of the first vehicle during the lifetime of the component. This historical set of loads is, for example, acquired by the sensing module described above. The first characteristics may include additional static metadata of the car, which does not necessarily come from reading the sensors, such as the age of the vehicle, the make and model, the engine type, the weight, etc.

[0041] In a second step, the next failure is predicted by inputting the first data into a component failure prediction model. The component failure prediction model comprises a sequence of classifiers 12, each classifier 121, 122, 12n of the classifiers sequentially traversed in the sequence having respectively an increasing specialization in determining a potential failure such that a classifier in the sequence determines the potential failure at one level of specialization and a subsequent classifier in the sequence determines the same potential failure at an increased level of specialization compared to the first classifier. The next predicted failure corresponds to the potential failure 14 determined by the last classifier 12n of the sequence 12.

[0042] The entire sequence and each individual classifier at each level can produce a positive decision that indicates an imminent failure, or precisely a failure that is expected to occur at a predefined time or mileage horizon after that decision has been made. This horizon must be predefined before training a model, as it affects the composition of the training data.

[0043] In the flowchart 1, three classifiers of the sequence of classifiers 12 are represented, the sequence of classifiers 12 comprising N classifiers, each corresponding to a sequence level. However, the number N of classifiers is not limited to 3. The number of classifiers may be greater than or equal to 2, for example 2, 5, 10 or 100.

[0044] A first classifier 121 receives the first data 11 which are used to perform a calculation using a first algorithm and a result of this calculation is then transmitted to the decision function of the first classifier, the positive response of which is representative of the failure of a first component of the vehicle. If the response of the first classifier is positive, then the first data 11 are transmitted to a second classifier 122. As for the first classifier 121, this second classi 122 receives the first data 11 which are used to perform calculations using a second algorithm and a result of this calculation is then transmitted to the decision function of the second classifier, the positive response of which is representative of the failure of a first component of the vehicle. If the response of the second classifier is positive, the first 11 data are transmitted to a following classifier. These operations are repeated up to the last classifier 12n, the N-th classifier, which receives the first data 11 which are used to perform a calculation using an N-th algorithm and a result of this calculation is then transmitted to the decision function of the N-th classifier, the positive response of which is also representative of the failure of a first component of the vehicle.If the response of the N-th classifier is positive, then the positive decision of the last 12n classifier of the sequence 12 becomes the ultimate positive response of the entire sequence.

[0045] Conversely, if at some level of the sequence the response of the classifier's decision function is negative, i.e. not representative of the failure of a first component of a vehicle, then an absence of failure 13 is determined and the processing no longer propagates. In this case, the process ends and no potential failure is detected. In this way, a first vehicle almost certainly showing symptoms of a next failure will be rejected at the beginning of the sequence. The more ambiguous a specific case is, the more levels it will pass through in sequence before being assigned an ultimate model decision.

[0046] According to an exemplary embodiment, the component failure prediction model implements at each sequence level an algorithm from a set of algorithms comprising at least: - logistic regression, - Bayesian modeling, - survival analysis, - decision tree and - deep neural network.

[0047] In another exemplary embodiment, the method further comprises predicting the maintenance operation, the maintenance operation being that of the first vehicle based on the next failure. For example, the maintenance operation is deemed necessary before the immobilization of the first vehicle and must therefore be carried out within a given time, the time being determined, for example, by an estimated distance in kilometers or an estimated time in days. To carry out this maintenance operation, an on-board system of said first vehicle is then controlled based on the maintenance operation. The on-board system is, for example, a navigation system, in which the address of a garage or workshop is automatically entered as a destination in order to guide the driver of the first vehicle, or in another example, the first vehicle is autonomous and automatically goes to this destination to carry out the maintenance operation or to check the need for this maintenance operation.

[0048] For many types of vehicle component faults at any given time, only a very small fraction of the entire field population will typically require servicing. This is either because the faults are rare, such as battery cell contamination at the manufacturing stage, or because the level of component wear considered hazardous has not yet been reached, such as for tires or brake pads. Therefore, when evaluating a typical population of vehicles using the component failure prediction model above, most will be dismissed as unlikely to fail with only few calculations performed at the early levels of the sequence. A larger computational load is only required for a fraction of the vehicles being monitored.

[0049] In real-life scenarios, predictive vehicle monitoring is a recurring process. If, for example, a given fleet of vehicles needs to be monitored at a certain cadence, it may happen that due to noise in the vehicle data and measurements, the model does not always produce consistent decisions.

[0050] In an exemplary embodiment, the component of a first vehicle may be labeled as likely to fail for several consecutive scoring sessions, e.g., three sessions, then labeled as unlikely to fail on the fourth score, and again labeled as likely to fail on the fifth score. In such cases, temporal integration of the scoring results is provided.

[0051] Various embodiments of the invention may use per-vehicle model decision disambiguation techniques, such as simple averaging of output failure probabilities, exponentially weighted averaging to better account for more recent outputs, or majority voting among "hard" sequence decisions (class labels). In addition, a more liberal decision disambiguation rule may be applied, for example, deciding that the component will fail if its failure was predicted in at least one of a predefined number of scoring sessions.

[0052] It should be noted that the learning phase of the method for predicting the next failure of a component of a first vehicle can be implemented on a first machine, for example with large storage, computing and access capacities for the learning data, such as a server. Once the component failure prediction model has been trained, it can then be put into works on a second machine, such as a mobile device, or directly installed in the on-board system of the first vehicle.

[0053] If the failure prediction model of a component is stored on a device remote from the first vehicle, the first vehicle transmits the first data to this remote device, for example via a wireless communication link, the remote device implementing the method of predicting the next failure of a component of a first vehicle.

[0054] Such a method makes it easy to identify the first vehicle requiring maintenance based, for example, on the driving conditions it has experienced. Preventive maintenance is then performed according to the needs of the first vehicle, and not just according to a maintenance schedule. In this way, maintenance operations are more appropriate, and a user of the first vehicle then uses the first vehicle in better conditions, the first vehicle being safer and more comfortable.

[0055] [Fig.2] illustrates a flowchart 2 of the different stages of a process training the component failure prediction model used to predict the next failure of a component of a vehicle, for example of the component of the first vehicle, according to a particular and non-limiting example of the present invention.

[0056] The component failure prediction model is learned in a learning phase from the learning data 21a representative of the second characteristics of a set of secondary vehicles, the second characteristics being homogeneous with the first characteristics.

[0057] By way of example, the training data 21a comprises data at least homogeneous or similar with the first data, representative of the historical loads applied to the component of a second vehicle, the component of the second vehicle corresponding to the component of the first vehicle. The historical loads may correspond, for example, to a certain measure of stress applied to the components of the vehicle. For each second vehicle, these historical loads are associated with a historical set of maintenance data of the second vehicle, so that the second characteristics comprise, for each second vehicle of the set of second vehicles, a historical set of maintenance data of the second vehicle associated with a second historical set of loads applied to the component of a second vehicle, the component of the second vehicle corresponding to the component of the first vehicle.

[0058] The performance of the component failure prediction model is estimated using performance indicators or measures that characterize the naturally competing objectives of the component failure prediction model, that is, its ability to successfully capture the true failure of the second vehicle component and reject the second vehicle component without failure. Examples of such competing metric pairs are the true positive rate (also known as recall) and the false positive rate, or precision and recall.

[0059] In a first exemplary embodiment, the learning phase comprises a step of receiving the second data 21b representative of an overall target rate of true positives A, an overall target rate of false positives B and a number N of classifiers in the sequence. The target rate of true positives A and the target rate of false positives B express the targeted predictive performance of the sequence. For example, the target value of the rate of true positives A is between 80 and 90%. In another more complex context, such as in the automotive industry, such a rate of true positives A is often out of reach. However, in certain contexts, for example to identify vehicles that must be recalled to the garage to replace a non-critical part, a target rate of true positives A of less than 50% could prove sufficient.On the other hand, the target false positive rate B should be chosen taking into account the possible practical consequences of false predictions, such as the excessive cost of towing the vehicle or unfounded repairs, or the possible negative experience of car owners warned of non-existent problems. Reasonable target values of the false positive rate B seem to be limited to the 10% level.

[0060] In this first exemplary embodiment, the learning phase comprises a step of determining 221 data representative of a target local true positive rate a and a target local false positive rate [3 of the global true positive rate objective, the global false positive rate objective and the number N of classifiers in the sequence.

[0061] In a second exemplary embodiment, the learning phase comprises a step of receiving second data 21b representative of a global target true positive rate A, a global target false positive rate B and at least one of the local targets of a true positive rate and a target local false positive rate [3. In this second embodiment, the learning phase comprises a step of determining 221 data representative of a target local true positive rate a and a target local false positive rate [3 if they have not been received previously and the number N of classifiers in the sequence from the secondary data 21b.

[0062] The training phase further comprises the steps of training at each level of the next sequence a new classifier until the component failure prediction model, including the classifiers trained up to present up to the current level classifier, generates: • a local true positive rate lower than said target local true positive rate, or • a local false positive rate higher than said target local false positive rate, or • jointly an overall true positive rate higher than said target overall true positive rate, and an overall false positive rate lower than said target overall false positive rate; and a step of producing the decision of the trained component failure prediction model or a diagnostic output for determining why the model could not be trained.

[0063] In a third exemplary embodiment, the training phase comprises a step of receiving the second data 21b representative of an overall target rate of true positives A, an overall target rate of false positives B, a third parameter specifying the maximum complexity of the model in sequence. By default, it can be expressed with the maximum allowed number of levels in sequence (first exemplary embodiment), but other options include the minimum improvement in the accuracy of the prediction on a set maintained between the current level and the previous level, the maximum total training time of the model or, given that each subsequent sequence level receives a decreasing number of training instances, the minimum number of second vehicles providing input data for this level. If the complexity requirement of the model is provided, the algorithm obtains a natural guarantee to terminate in finite time.Otherwise, to be useful in practice, the algorithm must resort to internal stopping conditions to ensure that it terminates at a given point in time. An example of a stopping condition requiring such an internal stopping criterion is a situation where the set of secondary vehicles classified as failure-prone at the current sequence level and transferred to the input of the next sequence level degenerates with respect to the feature space used by the local classifier. In this case, all training instances have exactly the same characteristics, making it impossible to add additional discriminative power to the model, regardless of the training algorithm. In conclusion, and to simplify other considerations, it can be assumed that the complexity bound is always defined, either explicitly by the system manager or implicitly, in which case it may not be known a priori.

[0064] At each sequence level, the current model is trained using the available historical observations. [Fig.2] illustrates a sequence model with N classifiers, therefore. The first level is trained in a first training operation 231. After the first training operation, the overall performance of the component failure prediction model is compared to the global requirements, so that, for example, the overall true positive rate after the first TPRi classifier is compared to A and the overall false positive rate after the first FPRi classifier is compared to B. If the result of this first comparison 241 indicates that the model is sufficiently discriminating, for example when TPRi>A and FPRi <B, alors l'entraînement du modèle de prédiction de défaillance d’un composant est terminé dans une opération complète 28. Inversement, si les conditions globales ne sont pas vérifiées, les conditions locales sont analysées. Les indicateurs locaux sont comparés aux cibles locales, par exemple le taux de vrais positifs tpri généré par le premier classificateur est comparé à a et le taux de faux positifs fpri généré par le premier classificateur est comparé à [3.If the result of this second comparison 251 indicates that the first classifier is sufficiently discriminating, for example when tpri>a and fpri<[3, then the training of the first classifier ends, and the second classifier is trained in a second training operation 232. If the local requirements cannot be satisfied, the first indiscriminate cases are extracted in an extraction operation 261 and the training data 21a and the second data 21b are analyzed in a first analysis operation 271. If the training data 21a can be refined and / or the requirements of the second data 21b adjusted to achieve the target performances in an adjustment operation 222, the training phase restarts with refined training data and / or with adjusted requirements.If the training data 21a cannot be refined and the requirements of the second data 21b cannot be adjusted to achieve the target performance, the component failure prediction model cannot be trained and the root causes must be analyzed, as part of an investigation operation 29, paying particular attention to the first indiscriminate cases extracted during the extraction operation 261.

[0065] In the same way, the second classifier is trained during a second learning phase 232. After the second training operation, the overall performance of the component failure prediction model is compared to the overall requirements, so that, for example, the overall true positive rate after the second classifier TPR2 is compared to A and the overall false positive rate after the second classifier FPR2 is compared to B. If the result of this third comparison 242 indicates that the model is sufficiently discriminating, for example when TPR2>A and FPR2 <B, alors l'entraînement du modèle de prédiction de défaillance d’un composant est terminé dans une opération complète 28. Inversement, si les conditions globales ne sont pas vérifiées, les conditions locales sont analysées.Local indicators are compared to local targets, for example the true positive rate tpr2 generated by the second classifier is compared to a and the false positive rate . positive cases fpr2 generated by the second classifier is compared to [3. If the result of this fourth comparison 252 indicates that the second classifier is sufficiently discriminating, for example when tpr2>a and fpr2<[3, then the training of the second classifier is finished, and the next classifier is trained in another training operation. If the local requirements cannot be satisfied, the second indiscriminate cases are extracted in an extraction operation 262 and the training data 21a and the second data 21b are analyzed in a second analysis operation 272. If the training data 21a can be refined and / or the requirements of the second data 21b adjusted to achieve the target performance in the adjustment operation 222, the training phase restarts with refined training data and / or with adjusted requirements.If the training data 21a cannot be refined and the requirements of the second data 21b cannot be adjusted to achieve the target performance, the component failure prediction model cannot be trained and the root causes must be analyzed, as part of an investigation operation 29, paying particular attention to the indiscriminate cases of the second extraction extracted in the extraction operation 262.

[0066] These operations are then repeated for each classifier i in the sequence of classifiers. The i-th classifier is trained during an i-th training phase. After the i-th training operation, the overall performance of the component failure prediction model is compared to the overall requirements, so that, for example, the overall true positive rate after the i-th classifier TPRi is compared to A and the overall false positive rate after the i-th classifier FPRi is compared to B. If the result of this fifth comparison 24i indicates that the model is sufficiently discriminating, for example when TPR;>A and FPRi <B, alors l'entraînement du modèle de prédiction de défaillance d’un composant est terminé dans une opération complète 28. Inversement, si les conditions globales ne sont pas vérifiées, les conditions locales sont analysées.The local indicators are compared with the local targets, for example the true positive rate tpp generated by the i-th classifier is compared to a and the false positive rate fpp generated by the i-th classifier is compared to [3. If the result of this sixth comparison indicates that the i-th classifier is sufficiently discriminating, for example when tpr;>a and fpr;<[3, then the training of the i-th classifier is completed, and the next i+1 classifier is trained in another training operation. If the local requirements cannot be satisfied, the i-th indiscriminate cases are extracted in an extraction operation and the training data 21a and the second data 21b are analyzed in an i-th analysis operation. If the training data 21a can be refined and / or the requirements of the second data 21b adjusted for . achieve the target performance in the adjustment operation 222, the training phase restarts with refined training data and / or with adjusted requirements. If the training data 21a cannot be refined and the requirements of the second data 21b cannot be adjusted to achieve the target performance, the component failure prediction model cannot be trained and the root causes must be analyzed, as part of an investigation operation, paying particular attention to the i-th non-discriminative cases extracted during the extraction operation.

[0067] The last classifier in the sequence, the N-th classifier, is trained during an N-th learning phase 23n. After the N-th learning operation, the overall performance of the component failure prediction model is compared to the overall requirements, so that, for example, the overall true positive rate after the N-th TPRN classifier is compared to A and the overall false positive rate after the N-th FPRN classifier is compared to B. If the result of this seventh comparison 24n indicates that the model is sufficiently discriminating, for example when TPRN>A and FPRN <b, alors l'entraînement du modèle de prédiction défaillance d’un composant est terminé dans une opération complète 28. inversement, si les conditions globales ne sont pas vérifiées, n-ièmes cas non discriminants extraits d'extraction et données d'apprentissage 21a deuxièmes 21b analysées n-ième d'analyse. si peuvent être affinées ou exigences la deuxième donnée ajustées pour atteindre performances cibles l'opération d'ajustement 222, phase redémarre avec des ajustées. que cibles, le peut entraîné causes profondes doivent analysées, d'investigation, en accordant attention particulière aux indiscriminés d'extraction.

[0068] It is recommended to keep weak learners as simple as necessary to meet the level-specific criteria, thus overall favoring the cheapest and fastest models. In fact, it is still essential that the trained model reasonably exploits the available historical knowledge about a second vehicle and its usage, and that it allows flexible tuning of the classifier's operating point to adjust its performance to the predefined criteria. The latter, a and [3, express the minimum true positive rate and the maximum false positive rate specific to the level accordingly, or their equivalents, as indicated above. Since in the sequence classifier structure, the equations below are valid as long as N is known before sequence learning begins:

[0069] [Math.l] TPRV = n^tpr l— 1 f 100701 FPRy =

[0071] a and fi can be automatically deduced as follows:

[0072] [Math.2]

[0073] p =

[0074] With: • TPRn the global true positive rate (after the N-th classifier), • FPRn the global false positive rate (after the N-th classifier), • tpp the local true positive rate for the i-th classifier, • fpq the local false positive rate for the i-th classifier, • N the number of classifiers in the sequence, • Has the overall target rate of true positives, • B the overall target rate of false positives, • has the local target rate of true positives, and • P the target local false positive rate.

[0075] If N is not known, then the currently obtained best tpp level (no worse than the globally expected A) and the derived fpq level can be accepted unconditionally and the training process can continue, efficiently with local performance checks bypassed but global checks still active. Alternatively, with more advanced users in mind, a separate embodiment of the invention allows for explicitly providing at the input the thresholds A, B, as well as at least one of the level-specific thresholds a, p. Then, by solving N with one of the above equations, one can determine the number of classifiers in the sequence and, if applicable, the other level-specific threshold.

[0076] The most important element of the sequence learning procedure is that training instances at level i that are classified as negative by the joint model integrating all levels in sequence up to i are irreversibly rejected. This follows a conservative principle that a given level should eliminate only the most obviously negative cases, leaving the discrimination mination of the remaining most ambiguous cases at later, more specialized levels of the sequence. Training of the i+1-th level in sequence, if allowed, therefore proceeds only with instances judged positive at the i-th level. In practice, however, since the overall true positive rate requirement is usually high, the level-specific rates are even higher, which inevitably has to come at the cost of equally high false positive rates, which delays the potentially positive decision of the entire sequence. It is therefore to be expected that, when growing the component failure prediction model in sequence, it is perfectly legitimate to expect a transition of most training instances from a given level to the next level that encompasses deep classifier designs.

[0077] It should be noted that, despite the fact that each level-specific classifier may be relatively low in terms of allowable false positive rates, its complexity, e.g., the number of parameters used or data dimensions exploited, is expected to increase with the number of levels. This is because the deeper in the sequence the sequence, the more difficult it is to discriminate a mixture of positive and negative instances encountered at the input. Moreover, since the level-specific recall thresholds are very high, i.e., 99%, most positive cases will typically survive to the last level in sequence, while many negative cases will not. This means that at subsequent levels in the sequence, the classification problem to which the component failure prediction model must be adapted becomes increasingly balanced.It may then be appropriate to use different learning strategies at different sequence levels, e.g., at each subsequent level, reset the class / instance weights to correctly reflect the progressive balancing of class proportions in the classifier cost function, or even consider completely different learning algorithms.

[0078] For example, for a lifetime linked to the mechanical stresses undergone by a suspension element, a first level would include a simple mathematical calculation of the stresses, while a following level would include a calculation by coarse finite elements of coating, and finally a last level would include a calculation by finite elements, so that the mesh would be very fine.

[0079] In an exemplary embodiment, an interface is provided to assist vehicle preventive maintenance model builders in successfully managing the construction of models to meet customer desired performance criteria. The system internally utilizes the sequence classifier training procedure described above and may be implemented as a software module running on any suitable computing resource, from a device mobile or a single standard personal computer (PC), to server farms or a cloud environment. In practice, the end-user interface part of the system alone is not computationally intensive, as it merely interprets client inputs and provides simple outputs. As such, it can be offloaded to a lightweight device and run there as an application. However, this application must also trigger the underlying training and evaluation algorithms in sequence. For computational efficiency and following the principle of co-location of storage and computing resources, these must be geographically close to where the data originally resides or can be transferred over the network.

[0080] More precisely, a model builder receives input performance requirements to be satisfied, which can be understood as a kind of contract to be fulfilled, and triggers the process. The requirements can usually be expressed in a non-rigorous way, for example using natural language, so that at the first stage they are analyzed and converted into the sequenced parameters described above. Then the training process begins. Once the first sequence level is formed, global checks against thresholds A and B are performed. If the target performance is met at this stage, the contract is considered fulfilled and the model ready to be delivered to the customer who requested it.Otherwise, if the local conditions involving thresholds a and [3 are met, it means that the training process is on track to achieve the target predictive performance and the next sequence level can be trained. At this point, the system operator can see the training progress indicator displayed meaningfully. The same process continues on the next sequence levels as needed.

[0081] If the local and global thresholds are not reached at some level i of the sequence, this means that training cannot continue and the model cannot be built given the current training data and performance criteria. To facilitate the diagnosis of why the model could not be built, the data vectors associated with the training instances that managed to “survive” up to the current level are extracted, corresponding to the i-th indiscriminate cases extracted in the extraction operation. As a reminder, given the principles of sequence accumulation described previously, in subsequent levels such cases will mainly include almost all true positives, and initially a large but gradually decreasing number of false positives, thus constituting a mixture that is increasingly difficult to distinguish from the training instances.The process operator is then asked whether to attempt to modify the target criteria or refine the dataset. If not, the . process stops with the "hard" training cases just extracted and which serve as auxiliary output to help understand why the training objective could not be achieved.

[0082] Performance criteria or data set refinement are additional possibilities to be considered in separate embodiments of the invention. In the first case, given the gap between the target performance and the combined performance achieved by the sequence up to the k-th level where training stopped, the algorithm may, with the agreement of the system user, lower the global threshold A and / or increase the global threshold B while keeping the number of sequence levels N unchanged, recalculate the local thresholds a and [3 accordingly, and then restart the sequence creation process with these new parameters. Alternatively, A and B may remain unchanged, but N increased and, a and [3 modified accordingly to allow less strict discrimination between positive and negative cases at each sequence level once training is restarted.

[0083] Refinement of the input dataset is a type of additional intervention in the original training configuration that can be considered to recover from the potential premature termination of the sequence learning process. It can be used with or without the aforementioned modification of the sequence parameters. There are a range of techniques to combat insufficient discriminative performance of machine learning models.Potential methods to apply here include simply increasing the size of the training population, which is often useful on its own, different variants of bottom-up minority class sampling, such as SMOTE (Chawla, Bowyer, Hall, & Kegelmeyer, 2002), or boundary-SMOTE (Han, Wang, & Mao, 2005) where only the most “difficult” minority class instances around the decision boundary are upsampled, or changing the feature representation of the data (e.g., considering additional dimensions, feature extraction or selection). In all cases, since the dataset changes, the training procedure must be started from scratch.

[0084] The model training process ends either when the model is correctly fitted and meets the input criteria, or with an error message informing that these criteria could not be met, supported by a list of training instances that the last activated sequence level could not discriminate well enough. Ultimately, both of these results need to be communicated to the target recipients, who are not necessarily familiar with machine learning jargon. Therefore, an additional sequence translation module may be necessary to obtain commonly understood information, e.g., a short passage in natural language.

[0085] [Fig. 3] shows a block diagram illustrating an exemplary system or apparatus 3 in which various exemplary aspects and embodiments are implemented.

[0086] The apparatus 3 may be integrated as one or more devices comprising the various components described below. In various embodiments, the apparatus 3 may be configured to implement one or more of the aspects described in the present application, the apparatus 3 being at least configured to predict the imminent failure of a component of a vehicle.

[0087] Examples of equipment that may constitute all or part of the apparatus 3 include a calculator, an electronic control unit (ECU), a data processing device, or other communication devices. The elements of the apparatus 3, alone or in combination, may be incorporated into a single integrated circuit (IC), multiple integrated circuits, and / or discrete components. For example, in at least one embodiment, the processing and encoder / decoder elements of the apparatus 3 may be distributed across multiple integrated circuits and / or discrete components. In various embodiments, the apparatus 3 may be communicatively coupled to other similar systems, or other electronic devices, via, for example, a communication bus 320 through dedicated input and / or output interfaces 32.

[0088] The apparatus 3 may include at least one processor 30 configured to execute instructions loaded therein to implement, for example, the various aspects described in the present application. The processor 30 may include integrated memory, an input and output interface, and various other circuits known in the art. The apparatus 3 may include at least one memory 31 (e.g., a volatile memory device and / or a non-volatile memory device). The apparatus 3 may include a storage device, which may include non-volatile memory and / or volatile memory, including, but not limited to, electrically erasable programmable EEPROM, read-only memory (ROM), programmable read-only memory (PROM), random access memory (RAM), dynamic random access memory (DRAM), static random access memory (SRAM), flash memory, magnetic disk drive, and / or optical disk drive.The storage device may include an internal storage device, a connected storage device, and / or a network accessible storage device, by way of non-limiting examples.

[0089] The apparatus 3 may include an encoder / decoder module configured, for example, to process image data to provide encoded / decoded text data, and the encoder / decoder module may include its own processor and memory. The encoder / decoder module may represent one or more modules that may be included in a device to perform the encoding and / or decoding functions. decoding. As is known, an apparatus may include either or both of the encoding and decoding modules. In addition, the encoder / decoder module may be implemented as a separate element or may be incorporated into the processor 30 as a combination of hardware and software, as is known to those skilled in the art.

[0090] The program code to be loaded onto the processor 30 or the encoder / decoder to perform the various aspects described in the present application may be stored in a storage device and then loaded onto the memory 31 for execution by the processor 30. In accordance with the various embodiments, one or more of the processors 30, the memory 31, the storage devices, and the encoder / decoder modules may store one or more of various elements during execution of the processes described in the present application. These stored elements may include, but are not limited to, data representative of the audio content, data representative of the video content, haptics-related data, a bitstream, matrices, variables, and intermediate or final results of processing equations, formulas, operations, and operational logic.

[0091] In several embodiments, memory within the processor 30 and / or the encoder / decoder module may be used to store instructions and provide working memory for processing that may be performed during data processing, encoding, or decoding.

[0092] In other embodiments, however, memory external to the processing device (e.g., the processing device may be the processor) may be used for one or more of these functions. The external memory may be memory 31 and / or the storage device, e.g., dynamic volatile memory and / or non-volatile flash memory. In at least one embodiment, fast external dynamic volatile memory such as RAM may be used as working memory for processing data.

[0093] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting. It will also be understood that the terms "includes / comprises" and / or "comprising / comprising", when used in this specification, may specify the presence of indicated features, elements and / or components, for example, but do not exclude the presence or addition of one or more other features, elements, components and / or groups thereof. In addition, when an element is said to be "reactive" or "connected" to another element, it may be directly reactive or connected to the other element, or intermediate elements may be present. In contrast, when an element is said to be "directly reactive" or "directly connected" to another element, there are no intermediate elements. present.

[0094] It should be noted that the use of any of the symbols / terms " / ", "and / or" and "at least one of", for example, in the cases of "A / B", "A and / or B" and "at least one of A and B", may be intended to encompass the selection of the first listed option (A) only, or the selection of the second listed option (B) only, or the selection of both options (A and B).As a further example, in the cases of "A, B and / or C" and "at least one of A, B and C", this formulation is intended to encompass the selection of the first listed option (A) only, or the selection of the second listed option (B) only, or the selection of the third listed option (C) only, or the selection of the first and second listed options (A and B) only, or the selection of the first and third listed options (A and C) only, or the selection of the second and third listed options (B and C) only, or the selection of all three options (A and B and C). This may be extended, as is clear to one of ordinary skill in this and related arts, to as many items as there are enumerations.

[0095] Various numerical values may be used in the present application. Specific values may be exemplary and the aspects described are not limited to these specific values.

[0096] It will be understood that, although the terms first, second, etc. may be used herein to describe various elements, such elements are not limited by these terms. These terms are used only to distinguish one element from another. For example, a first element could be called a second element, and, likewise, a second element could be called a first element without departing from the teachings of this application. No order is implied between a first element and a second element.

[0097] Reference to "an exemplary embodiment" or "an exemplary embodiment" or "an implementation" or "an implementation", as well as other variations thereof, is frequently used to indicate that a particular feature, structure, characteristic, etc. (described in connection with the embodiment / implementation) is included in at least one embodiment / implementation. Thus, occurrences of the phrase "in an exemplary embodiment" or "in an exemplary embodiment" or "in an implementation" or "in an implementation", as well as any other variations, appearing at various places in the present application do not necessarily all refer to the same embodiment.

[0098] Similarly, reference herein to "in accordance with an exemplary embodiment / example / implementation" or "in an exemplary embodiment / example / implementation", as well as other variations thereof, is frequently used to indicate that a particular feature, structure, or characteristic (described in connection with the exemplary embodiment / example / implementation) may be included in at least one exemplary embodiment / example / implementation. Thus, occurrences of the phrase "in accordance with an exemplary embodiment / example / implementation" or "in an exemplary embodiment / example / implementation" in various places in the specification do not necessarily all refer to the same exemplary embodiment / example / implementation, and separate or alternative exemplary embodiments / examples / implementations are not necessarily mutually exclusive of other exemplary embodiments / examples / implementations.

[0099] The reference numerals in the claims are given for illustrative purposes only and do not have a limiting effect on the scope of the claims. Although not explicitly described, the embodiments / examples and variations present may be used in any combination or subcombination.

[0100] A number of implementations have been described. Nevertheless, it will be understood that various modifications may be made. For example, elements of different implementations may be combined, supplemented, modified, or deleted to produce other implementations. In addition, a person of ordinary skill in the art will understand that other structures and processes may be substituted for those disclosed and that the resulting implementations will perform at least substantially the same function(s), in at least substantially the same manner, to achieve at least substantially the same results as the disclosed implementations. Accordingly, these and other implementations are contemplated in the present application.

Claims

Claims

1. A method for predicting an upcoming failure of a component of a first vehicle, the method being implemented by a processor and comprising the following steps: - receiving first data (11) representative of first characteristics of the first vehicle; - predicting said upcoming failure by inputting said first data into a component failure prediction model, said component failure prediction model being learned in a learning phase from learning data (21a) representative of second characteristics of a set of second vehicles, the second characteristics being homogeneous with the first characteristics, and said component failure prediction model comprising a sequence of classifiers (12), each classifier (121, 122,12n) classifiers sequentially traversed in the sequence having respectively increasing specialization, determining a potential failure such that a classifier in the sequence determines the potential failure at one level of specialization and a following classifier in the sequence determines the same potential failure at an increased level of specialization compared to the first classifier, the next predicted failure corresponding to the potential failure (14) determined by the last classifier (12n) of the sequence (12).,

2. The method of claim 1, wherein said first characteristics comprise a first historical set of loads applied to said component of the first vehicle over the lifetime of said component.

3. The method of claim 1 or 2, wherein said second characteristics comprise, for each second vehicle of said set of second vehicles, a historical set of maintenance data of the second vehicle associated with a second historical set of loads applied to the component of a second vehicle, said component of the second vehicle corresponding to the component of said first vehicle.

4. A method according to any one of claims 1 to 3, wherein

5.

6.

7. The said learning phase includes the following steps: - receiving second data (21b) representative of an overall target rate of true positives, of an overall target rate of false positives, and either of the maximum authorized number of levels in the sequence (N), or of at least a target rate of local true positives and of a target local rate of false positives. - determining (221) data representative of the maximum authorized number of levels (N) in the sequence, a target local rate of true positives and a target local rate of false positives from these second data (21b). Method according to claim 4, wherein said learning phase comprises the following steps: - training at each level of the next sequence a new classifier until the component failure prediction model comprising classifiers trained so far up to the current level of the classifier generates: • a local true positive rate lower than said target local true positive rate, or • a local false positive rate higher than said target local false positive rate, or • jointly an overall true positive rate higher than said target overall true positive rate, and an overall false positive rate lower than said target overall false positive rate; - Obtaining the decision of the trained component failure prediction model or a diagnostic output to determine why the model could not be trained. Method according to any one of claims 1 to 5, in which said component failure prediction model implements at each sequence level and algorithm from a set of algorithms comprising at least: - a logistic regression, - Bayesian modeling, - a survival analysis, - a decision tree and a deep neural network. A method according to any one of claims 1 to 6, further comprising: - the prediction of a maintenance operation of said first vehicle based on said next failure, - the control of an on-board system of said first vehicle as a function of said maintenance operation.

8. A computer program product comprising program code instructions for executing the method of any one of claims 1 to 7, when said program is executed on a computer.

9. Apparatus (3) for predicting the next failure of a component of a vehicle, wherein said apparatus comprises a memory (31) associated with at least one processor (30) configured to implement the method according to any one of claims 1 to 7.

10. A vehicle comprising the apparatus (3) according to claim 9.

Images