• 专利附录
  • 专利说明
  • 权利要求
  • 类似技术
  • 同族专利
  • 引用文献
  • 法律状态
  • 优先权
    • 专利摘要:
    The invention relates to a decision support system (1) for the maintenance of a machine (2), comprising anomaly detection modules (31, 32, 33) configured to determine health indicators at from measurements of physical parameters of the machine, a calculator (5) configured to calculate an operating diagnosis from health indicators by applying a decision model capable of learning, and a human-machine interface (6) configured so as to enable at least one expert to consult the health indicators and to declare an expert diagnosis, characterized in that the computer (5) is further configured to compare an operating diagnosis calculated from a set of health indicators with at least one expert diagnosis declared after consultation of said set of health indicators, and to modify the decision model in the event of disagreement between the calculation of anomaly diagnosis and the at least one expert diagnosis declared. It extends to the method implemented by this system, as well as to a computer program product for the implementation of the method.
    公开号:FR3041326A1
    申请号:FR1558811
    申请日:2015-09-18
    公开日:2017-03-24
    发明作者:Tsirizo Rabenoro；Jerome Henri Noel Lacaille
    申请人:SNECMA SAS；
    IPC主号:
    专利说明:

    SYSTEM AND METHOD FOR DECISION SUPPORT FOR THE MAINTENANCE OF A MACHINE WITH LEARNING OF A DECISION MODEL SUPERVISED BY EXPERT ADVICE
    DESCRIPTION
    TECHNICAL AREA
    The field of the invention is that of systems for monitoring the state of health of a machine, such as a motor, in particular an aircraft engine. The invention more particularly relates to an automated decision support system for performing maintenance operations of a machine.
    STATE OF THE PRIOR ART
    Monitoring the state of health of a machine aims to improve its safety and reliability. In the case of aircraft engines in particular, this monitoring is intended to avoid or limit in-flight shutdown (IFSD), to reduce delays or cancellation of flights ("delays and cancellations"). , D & C), and more specifically, to facilitate engine maintenance by anticipating failures and identifying faulty or faulty components.
    In order to monitor the state of health of an aircraft engine, various monitoring or anomaly detection devices are used to verify the proper operation of the various components of the engine. There is, for example, a monitoring device for analyzing the ignition process behavior, another for analyzing the temperature of the gases, another for detecting filter clogging, and another for analyzing the consumption of oil and gas. fuel, etc.
    The data generated by these detection and monitoring devices are exploited by ground maintenance services using health monitoring algorithms ("health monitoring"). These algorithms raise alerts when they detect an anomaly. These alerts are then used by the experts of the ground maintenance teams to check the operational capabilities of the engine according to the alerts raised.
    The proper calibration of the health status monitoring algorithms is essential to raise relevant alerts when the engine is degraded and not to raise an alert during a false detection of abnormality for example.
    On the one hand, the aim is to immobilize the engines as little as possible in order to make their operation profitable, and, on the other hand, to judiciously anticipate the maintenance operations in order to avoid costly repairs. For example it may be interesting to change a certain room as soon as a damage rate is reached to limit the other parts in connection with the damaged part, the impact of this degradation. In another example, when warnings are raised unduly, this may bring as a safety measure to the immobilization of the engine which after verification by the experts will prove fit for theft.
    Nowadays, the calibration of the health status monitoring algorithms is imperfectly carried out for lack of data concerning any deterioration since the motors are repaired before damage can occur. In the absence of supervision data, it is necessary to use anomaly detection algorithms that only exploit normal cases and are not very suitable for identifying types of degradation. To achieve this identification, it is necessary to use signature models built by simulation from the expertise of aeronautical engineers but not necessarily observed.
    In addition, detection and monitoring devices produce many different indicators that can be correlated with each other. Experts may find it difficult to make a decision about the relevance of alerts by observing the indicators. They must therefore spend time on each specific case of anomaly while the large number of engines followed forces them to make decisions in increasingly limited time.
    DESCRIPTION OF THE INVENTION The object of the invention is to propose a decision-aid solution that is as automatic as possible, thus accelerating the work of the experts while increasing the rate of detection of anomaly without increasing the rate of false alarms. The invention proposes for this a system as described below.
    BRIEF DESCRIPTION OF THE DRAWINGS Other aspects, objects, advantages and characteristics of the invention will appear better on reading the following detailed description of preferred embodiments thereof, given by way of non-limiting example, and made in reference to the accompanying drawings in which: - Figure 1 schematically illustrates a decision support system for the maintenance of a machine according to a possible embodiment of the invention; - Figure 2 schematically illustrates a decision support method for the maintenance of a machine according to a possible embodiment of the invention.
    DETAILED DESCRIPTION OF PARTICULAR EMBODIMENTS The invention generally provides a tool capable of automatically providing useful information for decision making relating to maintenance operations of a machine. The invention is advantageously applied to the maintenance of an aircraft engine, without this being for all that limiting.
    In this context, the invention proposes a decision support system for the maintenance of a machine which automatically calculates a diagnosis of operation of the machine. Functional diagnosis means an indication as to the presence or absence of a particular function (such as
    The principle of the invention is to correlate the operating diagnoses resulting from automatic calculations with diagnoses made by experts. The automatic calculations are performed using a decision model capable of learning, and this learning is performed iteratively so as to minimize the diagnostic error between the system response and the experts' responses.
    With reference to FIG. 1, the invention relates more particularly to a decision support system 1 for the maintenance of a machine, such as an aircraft engine 2.
    The system 1 comprises an anomaly detection unit 3 which comprises various anomaly detection modules 31, 32, 33. These modules receive measurements of physical parameters relating to different components of the machine, and are configured to determine indicators. from these measures. These health indicators are provided to a database 4.
    The health indicators produced by the modules 31, 32, 33 are calculation results, generally algorithm outputs, or statistical test results. These include, for example, outputs of health status monitoring algorithms or summaries or results of test analyzes.
    Typically, for the surveillance of aircraft engine fleets, the measurements made during each flight are analyzed by a fleet manager. These analyzes are performed by the anomaly detection modules 31, 32, 33 which implement algorithms for detecting failure in the behavior of the flight engine. There are several kinds of breaks (frank or progressive), several ways to observe them (long or short term), and a hundred signals to look separately or in relation. There is therefore a variety of anomaly detection modules (several thousand) that all provide, after each flight, their health indicators.
    The system 1 further comprises a computer 5 configured to calculate an operating diagnosis from health indicators provided by the anomaly detection modules 31, 32, 33. This calculation is made by applying a decision model which includes a set of rules for merging health indicators and synthesizing merged information to provide a functioning diagnostic. By way of example, the computer 5 can take as input the last 10 results of all the anomaly detection modules 31, 32, 33. It outputs a single piece of information, for example a piece of information indicative of an absence of breaking or information indicative of a break of the given type (performance, compressor, turbine, vibration, etc.), on a signal or several signals measured using successive detections to confirm the information thus produced.
    In the context of the invention, the decision model is capable of supervised learning, i.e. it can automatically develop its rules from a learning database containing health indicators. and validated operational diagnostics, here the database 4.
    This learning can be done by different methods. For example, it can exploit a naive Bayesian type classification algorithm or a "random forest" type of forest. It can also rely on regression methods (such as neural networks), or even exploit a reasoner using fuzzy logic.
    The decision model can thus be a Bayesian naive classification or a classification resulting from the application of a decision tree forest algorithm. The naive Bayesian-type algorithm gives results easily apprehended by the experts, and is preferably used at the initialization of the process. It is indeed more didactic for the experts, which allows them to have a good confidence in the invention. Then, once the initialization phase is completed, a "random forest" type algorithm can be exploited: it gives better results but these are more difficult to interpret. The invention proposes to iteratively correct this model so as to improve the quality of the operational diagnostic information. A re-learning of the decision model is achieved by reconsidering the decision rules developed during the learning and relearning them in order to provide a more reliable diagnosis of operation. At the end of each re-learning, a new decision model is generated. In other words, the decision model is modified at each iteration of a relearning.
    In order to initiate learning of the decision model, subsets of available health indicators can be selected at random (whose size and type depend on the precision sought, as well as sets of health indicators for which already has a validated diagnosis of functioning). An expert analyzes each subset of data and determines whether or not there is an anomaly and what type (labeling). At his expert diagnosis, he can add a level of quality or level of confidence. This diagnosis and this level of quality are then linked to the subsets of data in the database 4.
    Still with reference to FIG. 1, the decision support system 1 comprises a man-machine interface 6 configured so as to enable at least one expert to consult the health indicators stored in the database 4 and to declare an expert diagnosis. Experts can observe the signals that interest them (depending on their skills for example) and provide their expert diagnosis (performance problem, compression problem, too much vibration, etc.).
    The computer 5 is furthermore configured to compare an operating diagnosis calculated using the decision model from a set of health indicators with at least one expert diagnosis declared after consulting said set of health indicators. health. In the event of a discrepancy between the calculated operating diagnosis and the at least one declared expert diagnosis, the decision model can be modified, that is to say, it is relearned in order to minimize the error diagnosis.
    FIG. 2 schematically illustrates an example of a sequence of different steps that can lead to the modification of the decision model in a method according to one possible embodiment of the invention.
    The method comprises a prior calibration step during which one or more experts study the health indicators initially stored in the database 4 to provide their expert diagnoses and allow the initialization of learning the decision model.
    Then during a step "DIAG-EXP", each expert of a panel of experts, via the human-machine interface 6, consult and assess a set of health indicators stored in the database 4 and in return, declares his expert diagnosis (containing his classification of observed functioning and the associated level of confidence).
    During a step "DIAG-CAL", the calculator 5 calculates an operating diagnosis from said set of health indicators using the previously learned decision model. During a "DIFF" step, the operating diagnosis calculated automatically by the computer is compared with the expert diagnoses. In case of agreement ("NI"), a level of expertise of the system, in particular a level of expertise to detect a given operation, can be strengthened.
    In case of disagreement ("01"), experts are asked during a "CONF DIAG-EXP" stage to confirm their judgment while possibly modifying the level of confidence they place in their judgment. To judge a disagreement, the expertise levels of the experts to identify the functioning object of the diagnosis can be taken into account, for example via a weighting of the importance of the diagnosis of each expert.
    If yes ("02"), ie if at least one case of disagreement is confirmed, for example if several cases of disagreements are confirmed that allow a sufficient level of confidence to be reached in the judgment of the disagreement, it is proceeded during a step "MOD" to the modification of the learning model. This change is a re-learning of the rules of the model exploiting the contents of the database in which said set of health indicators is now associated with the confirmed expert diagnoses.
    If not ("NI"), that is, if the experts accept the proposal of the system (they consider that the diagnosis of functioning is correct and at least for some, invalidate their initial judgment), it is done during a step "REC" at the registration of this diagnosis in the database 4 where it is associated with said set of health indicators.
    In order to achieve this confirmation / reversal of their initial judgment, each panel expert can, via the man-machine interface, consult the expert diagnosis declared by each of the other panel experts, as well as the calculated operating diagnosis.
    Each expert can present a level of expertise, and the calculator 5 can be configured, when modifying the decision model, to weight the importance of a confirmed expert diagnosis according to the level of expertise of the expert. expert.
    An expert can have several skills, the level of expertise of an expert can be declined into a set of levels of expertise, each relating to the detection of a given operation. When an expert consults the diagnoses of other experts and the one calculated by the system, expert and system expertise levels can be provided.
    The experts can make an initial declaration of their competences, that is to say of their capacities to detect a functioning (an anomaly for example) given. When an expert diagnosis is evaluated by the panel of experts, it can be labeled as relating to one or more skills. It is thus possible, for a given expert, to count the positive expert diagnoses (validated by the panel) for the detection of a given operation, and thus to measure a success rate representative of the level of expertise for detection. said operation.
    In the same way, the system also becomes an expert over time and so it is possible to assess one's skills and thus to monitor one's maturity. As an example, the system calculates a functional diagnosis for a given operation A. To reach this conclusion, its decision model used past experiences to identify this operation. So he already had the answers of experts with a competence for functioning A and he had thus himself acquired a level of expertise for this operation A. This level of expertise can serve as an indication of the relevance of the response of the system. This relevance is even higher if the system has a good level of expertise for the detection of other operations. For example, he has a good level of expertise for operation B, and he judges that the operation is not type B.
    And as described above, if the operation diagnosis for the given operation A is confirmed by the experts, then the level of expertise of the system for the detection of this operation A is reinforced. The invention extends not only to the decision support system 1, but also to the decision support method for the maintenance of an engine implemented by such a system. It also relates to a computer program product comprising code instructions for executing the steps of the method according to the invention, when said program is executed on a computer.
    By directly taking into account the opinion of the experts, the invention overcomes the lack of supervision of existing systems by creating a perennial knowledge base.
    The experts also benefit from the invention through the statistical study that can be done of their respective intelligence histories. Which positively involves them in the process. The invention uses this same information to calibrate itself and takes the opportunity to evaluate the skills of each group of experts. The invention makes it possible to calibrate much more finely the algorithms for learning the decision model. The automated diagnosis is then of better quality, which makes it possible to anticipate the degradation of the parts by changing them rapidly as soon as a diagnosis indicates a degradation of the part. In fact, the earlier the deterioration of a part is identified, the less likely it is that this part will damage the other parts of the engine with which it is interacting. Thus, the invention makes it possible to improve the effectiveness of the repairs and the cost thereof by limiting, on the one hand, the number of parts to be changed, and, on the other hand, by limiting the number of inspections to carried out by the experts following the reception of an alert.
    权利要求:
    Claims (11)
    [1" id="c-fr-0001]
    A decision support system (1) for maintaining a machine (2), comprising anomaly detection modules (31, 32, 33) configured to determine health indicators from measurements of physical parameters of the machine, a computer (5) configured to calculate an operating diagnosis from health indicators by applying a decision model capable of learning, and a man-machine interface (6) configured to allow at least one expert to consult the health indicators and to declare an expert diagnosis, characterized in that the computer (5) is further configured to compare (DIFF) an operating diagnosis calculated from a set of health indicators with at least one expert diagnosis reported after consultation of said set of health indicators, and to modify (MOD) the decision model in case of disagreement between the calculated abnormality diagnosis and the at least one expert diagnosis declared.
    [2" id="c-fr-0002]
    2. System according to claim 1, wherein the computer (5) is configured to modify the decision model after declaration, via the human-machine interface (6), of a confirmation of the at least one diagnosis of Expert declared.
    [3" id="c-fr-0003]
    3. System according to claim 2, wherein the computer (5) is configured not to modify the decision model after declaration, via the human-machine interface, a reversal of the at least one expert diagnosis. declared.
    [4" id="c-fr-0004]
    The system of one of claims 1 to 3, wherein the computer (5) is configured to compare the abnormality diagnosis calculated from said set of health indicators with a plurality of expert diagnoses each declared after consulting said set of health indicators by an expert belonging to a panel of experts, and wherein the human-machine interface (6) is further configured to allow each panel expert to consult the diagnosis of expert declared by each of the other panel expert (s) and the calculated anomaly diagnosis.
    [5" id="c-fr-0005]
    5. System according to one of claims 1 to 4, wherein the computer (5) is further configured, when modifying the decision model, to weight the importance of the at least one expert diagnosis declared by an expert by a level of expertise of the expert.
    [6" id="c-fr-0006]
    6. System according to claim 5, wherein the level of expertise is associated with a given operation.
    [7" id="c-fr-0007]
    7. System according to one of claims 1 to 6, wherein the computer is configured to strengthen a level of expertise of the system in case of agreement between the calculated operating diagnosis and the at least one expert diagnosis declared .
    [8" id="c-fr-0008]
    8. System according to one of claims 1 to 7, comprising a database (4) for the storage of health indicators and declared expert diagnoses.
    [9" id="c-fr-0009]
    9. System according to one of claims 1 to 8, wherein the decision model is a Bayesian naïve classification or a classification resulting from the application of a decision tree forest algorithm.
    [10" id="c-fr-0010]
    10. Decision support method for the maintenance of a machine, comprising the determination of health indicators from measurements of physical parameters of the machine, the calculation (DIAG-CAL) of a diagnosis of operation at a machine, from health indicators by applying a decision model capable of learning, the consultation of health indicators by at least one expert and the declaration of an expert diagnosis (DIAG-EXP) by the at least one expert , characterized in that it comprises the steps of comparison (COMP) of an operating diagnosis calculated from a set of health indicators with at least one expert diagnosis declared after consulting said set of health indicators. health, and modification (MOD) model decision in case of disagreement between the calculated operating diagnosis and the at least one expert diagnosis reported.
    [11" id="c-fr-0011]
    A computer program product comprising code instructions for performing the process comparison and modification steps of claim 10 when said program is run on a computer.
    类似技术:
    公开号 | 公开日 | 专利标题
    EP3350660B1|2021-01-27|Decision aid system and method for the maintenance of a machine with learning of a decision model supervised by expert opinion
    EP2376989B1|2019-10-16|Identification of defects in an aircraft engine
    EP2912526B1|2017-12-06|System for monitoring a set of components of a device
    US20170284896A1|2017-10-05|System and method for unsupervised anomaly detection on industrial time-series data
    WO2010057846A1|2010-05-27|Method for recognising sequential patterns for a method for fault message processing
    Rögnvaldsson et al.2018|Self-monitoring for maintenance of vehicle fleets
    FR2966616A1|2012-04-27|METHOD, DEVICE AND COMPUTER PROGRAM FOR AIDING THE DIAGNOSIS OF A SYSTEM OF AN AIRCRAFT USING GRAPHICS OF REDUCED EVENTS
    US20110296248A1|2011-12-01|Systems and methods for restoring machine state history related to detected faults in package update process
    EP2874106A1|2015-05-20|System and method for aircraft failure diagnosis
    WO2020015471A1|2020-01-23|Automotive diagnostic method and device, and storage medium
    EP3215903B1|2019-12-18|Tool for validating a system for monitoring an aircraft engine
    WO2017077247A1|2017-05-11|System and method for monitoring a turbomachine, with indicator merging for the synthesis of an alarm confirmation
    FR2957170A1|2011-09-09|Equipment monitoring system designing tool for engine of aircraft, involves processing unit evaluating quantification of monitoring system based on quality value of result corresponding to output quality value associated with output module
    FR3012636A1|2015-05-01|METHOD FOR NON-REGRESSION OF A DESIGN TOOL OF AN AIRCRAFT ENGINE MONITORING SYSTEM
    FR3035232A1|2016-10-21|SYSTEM FOR MONITORING THE HEALTH CONDITION OF AN ENGINE AND ASSOCIATED CONFIGURATION METHOD
    FR3065995B1|2019-07-05|SYSTEM AND METHOD FOR MONITORING A TURBOMACHINE WITH ANOMALY DETECTION CORRECTED BY WEAR FACTOR
    Khan et al.2012|No Fault Found, Retest OK, Cannot Duplicate or Fault Not Found?–Towards a standardised taxonomy
    Giordano et al.2022|Data-driven strategies for predictive maintenance: Lesson learned from an automotive use case
    Langer et al.2014|A self-learning approach for validation of communication in embedded systems
    FR3076634A1|2019-07-12|METHOD FOR ANALYZING PLATFORM FAILURE SYMPTOMS, AND SYSTEM THEREOF
    WO2021069824A1|2021-04-15|Apparatus, method and computer program for monitoring an aircraft engine
    EP3559767A1|2019-10-30|Method for characterizing one or more faults in a system
    同族专利:
    公开号 | 公开日
    CN108027611A|2018-05-11|
    CN108027611B|2021-05-04|
    US20180253664A1|2018-09-06|
    EP3350660A1|2018-07-25|
    WO2017046530A1|2017-03-23|
    EP3350660B1|2021-01-27|
    FR3041326B1|2019-05-10|
    引用文献:
    公开号 | 申请日 | 公开日 | 申请人 | 专利标题
    US5132920A|1988-02-16|1992-07-21|Westinghouse Electric Corp.|Automated system to prioritize repair of plant equipment|
    US7587296B2|2006-05-07|2009-09-08|Applied Materials, Inc.|Adaptive multivariate fault detection|
    EP2634660A1|2010-10-28|2013-09-04|Hitachi, Ltd.|Anomaly diagnostic device and industrial machine|
    US20120151290A1|2010-12-09|2012-06-14|GM Global Technology Operations LLC|Graph matching system for comparing and merging fault models|CN111965995A|2020-08-03|2020-11-20|国营芜湖机械厂|Three-in-one aircraft avionics system health assessment and prediction system and method thereof|CN1811474A|2004-07-25|2006-08-02|天水长城电力仪器设备厂|Transformer dynamic stability state parameter tester and diagnostic software|
    CN101266279B|2008-05-09|2011-03-16|东北大学|Electric network failure diagnosis device and method|
    US8380642B2|2008-12-03|2013-02-19|Schlumberger Technology Corporation|Methods and systems for self-improving reasoning tools|
    US8949668B2|2011-05-23|2015-02-03|The Boeing Company|Methods and systems for use in identifying abnormal behavior in a control system including independent comparisons to user policies and an event correlation model|
    CN102981108B|2012-11-29|2015-03-11|重庆大学|Transformer internal insulation aging diagnosis method based on multi-feature information fusion technology|
    CN203066979U|2013-02-27|2013-07-17|安徽恒源煤电股份有限公司|GPRS remote technology-based monitoring device for maintenance of coal mining machine|
    CN103245861B|2013-05-03|2016-06-08|云南电力试验研究院（集团）有限公司电力研究院|A kind of transformer fault diagnosis method based on Bayesian network|
    CN104506338A|2014-11-21|2015-04-08|河南中烟工业有限责任公司|Fault diagnosis expert system based on decision tree for industrial Ethernet network|
    CN104751305A|2015-04-28|2015-07-01|国网上海市电力公司|Trouble analysis and repair-based intelligent interaction system and control method thereof|US10878385B2|2015-06-19|2020-12-29|Uptake Technologies, Inc.|Computer system and method for distributing execution of a predictive model|
    US10176279B2|2015-06-05|2019-01-08|Uptake Technologies, Inc.|Dynamic execution of predictive models and workflows|
    FR3089501B1|2018-12-07|2021-09-17|Safran Aircraft Engines|COMPUTER ENVIRONMENT SYSTEM FOR AIRCRAFT ENGINE MONITORING|
    CN110705957A|2019-08-31|2020-01-17|曲阜师范大学|Suggestion result feedback method of man-machine cooperative type weak artificial intelligence cloud system|
    FR3110721A1|2020-05-20|2021-11-26|Thales|Method and electronic device for determining a list of maintenance action , associated computer program|
    法律状态:
    2016-09-02| PLFP| Fee payment|Year of fee payment: 2 |
    2017-03-24| PLSC| Search report ready|Effective date: 20170324 |
    2017-05-17| PLFP| Fee payment|Year of fee payment: 3 |
    2018-08-22| PLFP| Fee payment|Year of fee payment: 4 |
    2018-09-14| CD| Change of name or company name|Owner name: SAFRAN AIRCRAFT ENGINES, FR Effective date: 20180809 |
    2019-08-20| PLFP| Fee payment|Year of fee payment: 5 |
    2020-08-19| PLFP| Fee payment|Year of fee payment: 6 |
    2021-08-19| PLFP| Fee payment|Year of fee payment: 7 |
    优先权:
    申请号 | 申请日 | 专利标题
    FR1558811A|FR3041326B1|2015-09-18|2015-09-18|SYSTEM AND METHOD FOR DECISION SUPPORT FOR THE MAINTENANCE OF A MACHINE WITH LEARNING OF A DECISION MODEL SUPERVISED BY EXPERT ADVICE|
    FR1558811|2015-09-18|FR1558811A| FR3041326B1|2015-09-18|2015-09-18|SYSTEM AND METHOD FOR DECISION SUPPORT FOR THE MAINTENANCE OF A MACHINE WITH LEARNING OF A DECISION MODEL SUPERVISED BY EXPERT ADVICE|
    EP16785185.6A| EP3350660B1|2015-09-18|2016-09-15|Decision aid system and method for the maintenance of a machine with learning of a decision model supervised by expert opinion|
    US15/759,394| US20180253664A1|2015-09-18|2016-09-15|Decision aid system and method for the maintenance of a machine with learning of a decision model supervised by expert opinion|
    CN201680053869.XA| CN108027611B|2015-09-18|2016-09-15|Decision assistance system and method for machine maintenance using expert opinion supervised decision mode learning|
    PCT/FR2016/052333| WO2017046530A1|2015-09-18|2016-09-15|Decision aid system and method for the maintenance of a machine with learning of a decision model supervised by expert opinion|
    [返回顶部]