FR2957170A1 - 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 - Google Patents

Abstract

The tool has an information processing unit (9) e.g. computer or logic controller, to calculate an output quality value of an input module corresponding to a data acquisition module, based on a transfer function associating imprecision of output in response to imprecision of input by using an intrinsic quality indicator, with the input module. The processing unit evaluates quantification of an equipment monitoring system (11) based on a quality value of a diagnosis result that corresponds to the quality value associated with an output module corresponding to an abnormality evaluation module. An independent claim is also included for a system for monitoring an equipment of an engine of an aircraft, comprising a data acquisition module.

Description

DESIGN TOOL FOR AN AIRCRAFT ENGINE MONITORING SYSTEM

TECHNICAL FIELD The present invention relates to the field of aircraft engine monitoring systems and more particularly to a tool for designing a surveillance system for an aircraft engine equipment.

STATE OF THE PRIOR ART The monitoring or the detection of anomaly of an aircraft engine is very important for the safety of the aircraft. Monitoring systems are used to verify the proper operation of the various aircraft engine equipment. For example, there is a monitoring system for analyzing the behavior of the engine during the ignition process, another for analyzing the gas trajectory, another for detecting filter clogging, and another for analyzing fuel consumption. oil, etc.

All of these monitoring systems improve the safety and reliability of aircraft engines. In particular, they make it possible to avoid or limit the IFSD (in-flight shutdown), to reduce delays or cancellation of flights (D & C delays and cancellations), and more particularly, facilitate engine maintenance by anticipating failures and identifying faulty or faulty components. However, the design of each monitoring system requires specific and complex means to solve the different technical aspects of the monitoring related to each particular engine equipment. To do this, it is necessary to build a model specific to the monitored equipment and to manage test and validation data corresponding to this engine equipment. Thus, for each equipment, it is necessary to deploy a lot of resources and time to model, develop, and validate the corresponding monitoring system. In particular, the validation of a monitoring system is a step that requires handling a very large number of data that require a significant amount of computing time to analyze them. In addition, it is important to choose the validation indicators in order to avoid that an anomaly is reported by mistake or that an element of the engine is in default without the event being detected.

In addition, it may happen that the validation level of a surveillance system is different from that of another surveillance system. This can complicate the analysis of data from different engine monitoring systems.

The object of the present invention is to propose a tool for designing an aircraft engine equipment monitoring system that makes it possible to define a modalization environment that is independent of the application context while allowing reliable validation in a manner that simple and 3 reuse elements for different stains.

PRESENTATION OF THE INVENTION The present invention is defined by a tool for designing a system for monitoring at least one equipment of an aircraft engine, said surveillance system being configured to receive measurements specific to said equipment and to delivering a result diagnosing the state of said equipment, comprising: means for constructing said monitoring system in a structure assembling a chain of modules dedicated to particular tasks of said surveillance system and comprising at least one input module, at least an intermediate module, and at least one output module, said modules being provided with input-output interfaces enabling them to communicate with each other and / or with said equipment and its environment, each of said modules being configured to receive via its interface input, an input and a DQV input quality value associated with said input, and to output on its own output interface, an output corresponding to a severity and a PQV output quality value associated with said output, - means for assigning a quality value to said specific measurements that correspond to the input associated with said at least one input module means for calculating, for each of said modules, the output quality value PQV 4 of said module according to a transfer function associating an inaccuracy of the output in response to an inaccuracy of the input by using at least one intrinsic quality indicator said module, and - means for evaluating a qualification of the monitoring system according to a quality value of said result which corresponds to the quality value of the PQV output associated with said at least one output module. Thus, the design tool makes it possible to validate the structure of generic modules independent of the application context and collaborating with each other to define a specific monitoring system for a particular piece of equipment of the engine. Advantageously, the design tool comprises means for weighting the output quality value of each of the modules and for determining an overall structural quality score as a function of said weighted values. Thus, more precision is obtained on the structural validation of the monitoring system. According to a first embodiment, said structural quality indicator corresponds to a variance estimation indicator of the imprecision of the output generated in response to a noisy input during a learning of said module. This makes it possible to validate, in a generic way, the structural quality of each module. According to a second embodiment, said structural quality indicator corresponds to an adequacy indicator AQV defining the measurement of a distance between the input of said module and a calibration input defined during a learning of said module. This AQV adequacy indicator measures the agreement between the current input data and the input data observed during the calibration of the module and thus gives an indication of its level of obsolescence. According to a second embodiment, said structural quality indicator corresponds to an MQV robustness indicator measuring a capacity of use of the module vis-à-vis new inputs, the robustness indicator being determined during a learning of said module. The robustness indicator is information of a priori quality which gives a measure of what the module is used to answer as quality. Advantageously, the PQV output quality value of said module is calculated by combining at least two of said indicators for estimating variance, AQV matching, and MQV robustness.

This makes it possible to increase the accuracy of the internal quality of each module and thus to increase the reliability of the structural validation of the monitoring system. Advantageously, the design tool comprises means for applying at least one data quality indicator to a result of said monitoring system obtained from the known specific measures including faults known to verify the detection of said faults.

This qualifies the quality of the result and therefore, in combination with the structural qualification, allows complete validation of the monitoring system and gives an indication of the efficiency and level of maturation of the system. According to a first example, the data quality indicator is a measure of a POD detection probability defined as being the probability of detecting a fault when the monitored equipment has an anomaly, and which is determined according to a constraint of a predetermined false alarm rate PFA defined as the probability that the monitored equipment is healthy when the monitoring system detects a fault. The probability of detection POD makes it possible to validate with great reliability the result of the surveillance system. In a second example, the data quality indicator is a measure of a PCL location rate defined as being the probability that a location is good when an anomaly of the equipment is observed. The PCL location rate or probability is a classification quality information that makes it possible to validate the result of the monitoring system with great efficiency. Advantageously, the design tool includes means for calculating a CQV correlation quality indicator of said monitoring system by calculating a correlation between the output quality value PQV and a function related to said data quality indicator. This makes it possible to further increase the reliability of the validation of the monitoring system by performing a synthesis between its structural validation and its result validation.

Advantageously, the design tool comprises means for calculating a CVR fault coverage rate indicator defined as a function of a prior probability of occurrence of each fault and a volume of data necessary to cover said each fault. Thus, the fault coverage rate can be used to modulate some data quality indicators to confirm the validation of the monitoring system.

Advantageously, the design tool comprises means for calculating a genericity quality indicator measuring a level of genericity of each of said modules used in the composition of said monitoring system.

The genericity quality indicator provides additional information on the validation of the monitoring system. Advantageously, the design tool comprises means for calculating a global relevance indicator of the surveillance system according to a relevance score and a weighting associated with each of said modules. Overall relevance provides information on the performance of the surveillance system. The invention also relates to a system for monitoring at least one equipment of an aircraft engine designed by the design tool according to the above characteristics and comprising: at least one data acquisition module configured to receiving, via its input interface, specific measurements IN1 and a quality value DQV1 associated with said specific measurements, and for delivering on its output interface specific data OUT1 and an output quality value PQV1 associated with said specific data, a module of standardization configured to receive via said input interface, said specific data and said corresponding output quality value and to output on its output interface standardized data OUT2 and a quality value PQV2 associated with said standardized data, a detection module anomaly configured to receive via its input interface, said standardized data and the said quality value associated with the standardized data and for delivering on its output interface, an abnormality vector OUT3 and a quality value PQV3 associated with said anomaly vector, and a classification module configured to receive via its input interface , said abnormality vector and said quality value associated with the anomaly vector and for outputting via its output interface, a fault identification measurement OUT4 and a quality value PQV4 associated with said identification measurement. BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 illustrates hardware means implemented in a monitoring system of an aircraft engine equipment, according to the invention; Fig. 2 schematically illustrates a design tool of the monitoring system of FIG. 1; Fig. 3 schematically illustrates a module that can be used in the monitoring system of FIG. 1; Fig. 4 illustrates an example of distribution of probabilities of good and bad health of aircraft engine equipment, according to the invention; and FIG. 5 illustrates an example of a monitoring system of FIG. 1. DETAILED DESCRIPTION OF PARTICULAR EMBODIMENTS FIG. 1 illustrates hardware means implemented in the monitoring system of an aircraft engine equipment, according to the invention. These material means 1 comprise sensors 3 for measuring data relating to a particular piece of equipment of the aircraft engine 7 and its environment, and information processing means 9 such as a computer or computer for carrying out one or more of a plurality of computer programs comprising program code instructions stored in the storage means of the computer and adapted to implement a monitoring system 11 of the equipment 5. The monitoring system 11 is configured to receiving equipment-specific measurements 5 and providing a result diagnosing the state of this equipment 7. More particularly, the monitoring system 11 is configured to collect and digitize the specific measurements acquired by the sensors 3 on the equipment 5 of the engine 7 and its environment. The monitoring system 11 is also configured to define specific variables giving indications on physical or logical elements of the equipment 5.

The monitoring system 11 can also be configured to standardize or standardize these specific variables and to construct signatures or vectors of anomalies representative of the behavior of the equipment 5.

The monitoring system 11 is also configured to evaluate scores to diagnose or detect whether the anomaly signature reveals an abnormality. Depending on the value of the score relative to a predefined threshold, the monitoring system 11 is configured to generate or not an alarm indicating that an anomaly is detected. The monitoring system 11 may also be configured to identify or classify the detected anomalies.

Thus, depending on the type of application, the monitoring system 11 is configured to perform several tasks that may include data acquisition, data normalization, anomaly detection, and possibly the classification of detected anomalies.

Fig. 2 schematically illustrates a design tool of the monitoring system 11 of an equipment 5 of the aircraft engine 7, according to the invention. The design tool comprises the information processing means 9 comprising the means that are usually found in a computer: a central unit 9a, storage means 9b, input peripherals 9c and peripheral devices. 9d outlets. The processing means 9 are used to construct the monitoring system 11 according to an assembly of different components or elements that will be called "modules" 13a-13x dedicated to particular tasks of the monitoring system 11. In fact, the storage means 9b comprise a plurality of modules 13a-13x which can be accessed for example from an interactive graphical representation and where each module indicates the specific tasks or executable functions that it uses.

Thus, the monitoring system 11 can be constructed according to a structure assembling a chain of modules 13 selected from the plurality of modules 13a-13x according to the particular tasks of the monitoring system.

The module chain 13 comprises at least one input module 13a to 13c, at least one output module 12 13f and 13g, and optionally one or more intermediate modules 13d and 13e. There are modules for all tasks required for monitoring and troubleshooting. For example, some modules may be specialized for data acquisition, others for arithmetic tasks, others for filtering or for the exploitation of stochastic data, etc.

Fig. 3 schematically illustrates a module 13i that can be used in the monitoring system 11, according to the invention. Each module integrates algorithmic functions that are called when necessary. In fact, there may be a function to initialize the state of the module, another to do the calibration, other functions for the execution of the algorithm and possibly another to plot the results.

As a nonlimiting example that can be used in the context of the present invention, mention may be made of the structure of the modules described by the author of the present invention J. Lacaille in a publication entitled "A Maturation Environment to Develop and Manage Health Monitoring Algorithms". , PHM 2009, San Diego, CA. Each module 13i is provided with an input interface 13i1 and an output interface 13i2 enabling it to communicate with at least one other module and / or with the sensors 3 and / or with the output peripherals 9d. Each of the modules 13i is configured to receive, via its input interface 13i1, an input IN and a value of input quality DQV (Data Quality Value) associated with the input IN, and to deliver on its interface output 13i2, an Out (Output) output corresponding to a severity (Risk) and an output quality value (i.e., a quality of severity value) PQV (Predictive Quality Value) associated with the OUT output. Thus, the input-output interfaces 13i1-13i2 allow the modules 13 to communicate with each other or with the outside. More particularly, an input module 13a corresponds to a data acquisition module configured to receive via its input interface 13a1, an input IN1 corresponding to measurements specific to the equipment 5 and its environment from the sensors 3 The input module 13a is configured to also receive via its input interface an input quality value DQV1 associated with the specific measurements. In particular, the processing means 9 assign a predetermined quality value DQV1 to the specific measurements. For example, at the outset, the processing means 9 assign a quality value DQV1 equal to one (DQV1 = 1) to the specific measurements.

Each module is responsible for its output data. Thus, the input module 13a delivers on its output interface 13a2, an output OUT1 and a quality value of the output PQV1 which are retrieved by the next module 13d via its input interface 13d1. From module to module, we arrive at the last (or last) 14 module (s) 13f, 13g of the chain 13 corresponding to the module (s) output. More particularly, an output module may correspond to an anomaly evaluation module 13f or a classification module 13g. For example, the module 13f is configured to receive via its input interface 13f1, an input IN4 (corresponding to the output of the previous module: IN4 = 0UT3) and an input quality value DQV4 (corresponding to the quality value output of the previous module: DQV4 = PQV3) from the previous module 13e and to deliver via its output interface 13f2, an output OUT4 corresponding to the result of the monitoring system and a quality value of the result PQV4.

The result OUT4 of the monitoring system 11 is a diagnosis of the anomaly; it is a measurement which reflects an abnormal behavior of the equipment 5. For example, if the value of the result OUT4 exceeds a predefined threshold, an anomaly detection appears and possibly generates an alarm on the output devices 9d. To this result is associated a quality of output PQV4 indicating the reliability of the result OUT4. This quality PQV4 which represents a precision of each output can be used to modulate all the decision system that comes after. For example, if an OUT4 result exceeds the predefined threshold indicating the existence of an anomaly, but the quality of the PQV4 result is low, then it may be assumed that the detection of the anomaly is not very reliable. Since the quality of the result depends on the quality of each of the modules, the processing means 9 are configured to calculate for each module 13i, the output quality value PQV of the module 13i according to a transfer function associating an inaccuracy of the output. OUT in response to inaccuracy of the IN input using at least one intrinsic quality indicator 15i to the module 13i. The intrinsic quality indicator 15i corresponds to a sensitivity measurement of the module. For example, in order to generically measure the structural quality of each module 13i, the structural quality indicator 15i may correspond to a variance estimation indicator of the imprecision of the output generated in response to a noisy input when learning a module 13i. In practice, during the validation phase, the neighborhood of an IN entry point can be noisy (for example, by mixing data while respecting the distribution of the laws of the inputs). In response, the output OUT of the module 13i has a variation with respect to the usual output, which makes it possible to calculate the variance of this result indicating the variability of the information of the output. Thus, for each of the points in a learning database, a corresponding variance can be calculated and an estimate (for example, based on a regression model) of the variance. In this case, this variance estimation function is the intrinsic quality indicator 15i which also corresponds to the transfer function of the module. According to another example, the structural quality indicator 15i may correspond to an Adequancy Quality Value (AQV) indicator defining the measurement of a distance between the input IN of the module 13i and a calibration input defined during the analysis. learning a module 13i. In this case, the AQV adequacy indicator translates the imprecision of the input data into an imprecision on the output data. Here, as a first approximation, the transfer function is equal to the adequacy indicator AQV (i.e., PQV = AQV x DQV). This AQV adequacy indicator measures the agreement between the current IN input data and the input data observed during the calibration of the module 13i and thus gives an indication of its level of obsolescence. For example, if the IN inputs of a module 13i no longer resemble the observations used during its calibration, the module 13i needs to be reworked. It should be noted that the AQV can be calculated for any type of module. The AQV adequacy indicator is the adequacy of an input data. Thus, in a practical manner, it suffices to record the data used for calibrating the module 13i so that its suitability can then be measured by comparing these calibration data with new data that arrives during the execution of the module. 13i. In other words, the distance between the new inputs and the usual calibration inputs is measured. For example, a Kullback-Lieber distribution distance can be used. 17 The AQV adequacy indicator can also be calculated as an adjustment of the last observations. The p-level of a statistical comparison distribution test can be used with or without Ansari-Bradely, Kolmogorov-Smirnov, Wicoxon, or other hypotheses. As indicated previously, the AQV can be defined as the probability that the current distribution of the inputs corresponds to the distribution of observations used for the calibration. Then, after the calibration, the adequacy is equal to one (AQV = 1). However, when the measurement context changes, the AQV tends to zero naturally. One can either warn the user or update the module 13i automatically. In the latter case, a selection of observations is implemented for re-learning purposes. In another example, the structural quality indicator 15i corresponds to a Model Quality Value (MQV) robustness indicator measuring a capacity of use of the module 13i in front of new inputs. This robustness indicator MQV is determined during a learning module 13i. In this case, the inaccuracy of the output data depends on the robustness of the module 13i (i.e., the ability to use the module for new input data), and the transfer function is here equal to the robustness indicator MQV (that is, PQV = MQV x DQV) The robustness indicator MQV thus gives information on the operation of the module 13i with respect to general data and then indicates what should be the average quality of the module 13i when new data is applied. The MQV indicates the quality of an internal state of the module 13i calculated during the calibration, thus during the development of the module 13i on test data and thus, at the end of the calibration, there is an intrinsic quality to the module 13i. In other words, it is a quality information a priori that gives a measure of what the 13i module is used to answer as quality. Note that in general, a module 13i is calibrated from a limited set of measurements and then the estimate of the robustness indicator MQV can be calculated using a cross-validation model or a bootstrap technique. The MQV can be estimated for any module with a quantifiable output. Of course, the PQV output quality value of a module 13i can be calculated by combining at least two structural quality indicators 15i among the variance estimation, AQV matching, and MQV robustness indicators. This makes it possible to take into account several components of the internal state of the module 13i that may depend on the measurements of the input and the internal states of the module 13i. By default, the output quality value PQV of the module 13i can be calculated by performing a normalized multiplication between the AQV adequacy indicator and the MQV robustness indicator.

In this case, the quality of the output data depends both on the degree of obsolescence and on the robustness of the module 13i as follows: PQV = (MQV OO AQV) xDQV In particular, the geometric mean can be taken. JMQVxAQV if you have no idea about the dimensions of the intrinsic state of module spaces and data. The output quality PQV measures the quality of each module and as the input data of each of the intermediate (s) 13d, 13e and 13f, 13g output modules are calculated by previous modules (i.e., each intermediate task has a quality conditioned by each previous task), then the quality of the result PQV4 (or PQV5) gives a good indication of the qualities of all the previous modules 13a to 13e (or 13a to 13f) of the chain of modules 13 This makes it possible to validate the monitoring system 11 as a global function indicating that when data is entered into this system, an OUT4 result (or OUTS) with some quality information PQV4 (or PQV5) is obtained on these results. . The processing means 9 are configured to evaluate a qualification of the monitoring system 11 as a function of the quality value PQV4 (or PQV5) of the result OUT4 (or OUTS) which corresponds to the quality value of the output associated with the module (s). ) output 13f, 13g and depends on the intrinsic quality of each of the modules. Thus, the structural qualification of the system is a validation of the monitoring system 11 as a selection of different modules that collaborate with each other to solve the different operational tasks of the system. In other words, the different functions of the various modules 13a to 13g necessary for the monitoring of the equipment 5 (for example, functions concerning the acquisition of the measurements, the standardization of the data, the detection of the anomalies, the classification and the identification faults) are validated independently by calculating the internal quality of each module collected from the module's local experiments. Advantageously, the processing means 9 are configured to weight the output quality value of each of the modules and to determine an overall structural quality score as a function of these weighted values. This makes it possible to trace information at the level of the monitoring system 11 by taking into account a quality score of each of the modules of the chain 13 according to its importance. After the validation of the structure or the architecture of the monitoring system 11, it is advantageous to validate the quality of its result. Thus, the processing means 9 are configured to apply at least one data quality indicator 17 (see Fig. 2) to an OUT4 or OUTS result of the monitoring system 11 obtained from the known specific measures including faults known to verify the detection of faults. In other words, specific measurements are introduced into the monitoring system 11 from a database of faults generated by simulation, test benches, and sometimes (but fortunately very rarely) records of actual operational events. Next, one or more data quality indicators 17 are applied to the result of the monitoring system 11 to verify the detection of faults. This gives a result score in parallel with the structural note which makes it possible to validate the monitoring system 11 on a specific data set, on a specific equipment 5 of the aircraft engine 7, and in a specific context. According to a first example, the data quality indicator 17 is a measure of a probability of detection P0D (Probability of Detection) defined as being the probability of detecting a fault when the monitored equipment has a fault (or anomaly). . The probability of detection POD is determined according to a constraint of a predetermined false alarm rate PFA (Probability of False Alarm) defined as the probability that the monitored equipment is healthy when the monitoring system 11 detects a fault. The probability of detection POD (Probability Of Detection) which is a quality information of the result also called the detection power 1-p, makes it possible to validate with great reliability the result of the surveillance system 11. It will be noted that the probability of detection POD mainly applies to fault diagnosis or fault assessment modules. 22 The probability of POD detection can be calculated from a database of faults generated by simulation, test benches, and sometimes (but very rarely) records of actual operational events. In this case, the POD detection probability can be estimated as the ratio of the number of faults detected to the total number of faults. It should be noted that the rate or probability of PFA false alarm is a very important criterion in the aeronautical field. Indeed, the event that the equipment is healthy (that is, in good health) when the surveillance system detects a fault is a phenomenon that must be limited because it can permanently change the credibility image that the user can have of the monitoring system 11. Thus, it is very advantageous that the rate or probability of false alarm PFA is low. The probability of PFA false alarm that is defined by the probability of being healthy when detected, can be calculated from Bayesian rules. Let P (detected) the marginal probability that an anomaly is detected by the monitoring system 11, and P (healthy) the marginal probability that the equipment 5 is healthy. Then, the false alarm rate PFA is the probability that the equipment 5 is a posteriori healthy knowing that a fault is detected. This can be represented by the following conditional probability: PFA = P (healthy / detected) = P (detected / healthy) P (healthy) P (detected) 23 The probability of detecting a fault when the equipment is healthy corresponds to the detection error a = P (detected / healthy) FIG. Figure 4 illustrates an example of a probability distribution of good and bad health. More particularly, the curve C1 represents the distribution of a probability of good health (healthy equipment) and the curve C2 represents the distribution of a probability of poor health (equipment at fault) compared to a predefined threshold of diagnosis D. L detection error (i.e., the probability of detecting a fault when the equipment is healthy) corresponds to the part of the curve C1 which exceeds the diagnostic threshold D. The error (it is that is, the complement of the detection probability POD) corresponds to the part of the curve C2 before the diagnostic threshold D. Thus, the false alarm rate PFA can be obtained as a function of the probability of detection POD (c that is, 1-P = P (detected / fault)) as well as a prior probability of having a fault P (fault) and the detection error a, as follows: PFA = a (1p (fault)) a (1p (fault) + PODxP (fault) Advantageously, in order to reach a low predetermined rate of false alarm PFA (for example less than 1%), we can operate a confirmation method using for example several successive detections. Thus, by confirming an information corroboration anomaly, the false alarm rate that behaves as a product of probabilities will of course be decreased. In addition, the detection error can be adjusted for PFA and POD data, according to the following equation: a = PODx PFA x P (fault) 1-PFA 1-P (fault) This can be interesting because a module is calibrated using the detection error knowing that the POD detection power is obtained as a quality information. This equation shows that since the false alarm rate PFA and the probability a priori of having a fault P (fault) are very small, the detection error that behaves like their product, will be even lower.

According to a second example, the data quality indicator 17 is a measure of a PCL (Probability of Class Location) detection rate defined as being the probability that a location is good when an anomaly of the equipment 5 is observed. The PCL location rate or probability, which is classification quality information, is essentially applied to the classification modules (for example, the module 13g) and makes it possible to validate the result of the monitoring system 11 with great efficiency. makes it possible to locate the precise element (for example, alternator, cable, harness, etc.) of the equipment 5 having an anomaly. In general, the probability of detection POD makes it possible to detect whether the equipment 5 comprises an abnormal element, and then the PCL localization rate indicates the efficiency with which this element can be located or identified. The PCL localization rate when applied to a 13g classification module has a similar meaning to that of the POD detection probability. After detection of a fault, a diagnostic module 13f can provide an anomaly signature and a classification module 13g associates a fingerprint with this signature. This footprint may be representative of a specific fault or even of a physical component that has been impacted by the detected degradation. This classification can be deterministic (that is, the footprint is clearly identified) or probabilistic (that is, a probability of impact is associated with each footprint). The PCL localization rate calculates the proportion of good classifications among all faults. PCL = P (well classified / fault) = P (C / F) As previously, the PCL localization rate can be estimated from a database of fault measurements (by simulations, test benches, or rare events). The PCL localization rate can be Bayesian deduced from the previous indicators.

PCL = P (C / F) = P (C / FD) P (D / F) + P (C / FD) P (D / F) P (C / D) = P (C / DF) P (F) / D) + P (C / DS) P (S / D) PCL = P (C / D) POD 1pFA 26 The letter C means "well classified", the letter F means "fault", the letter D means "detected" The symbol D means "not detected", the letter S means that the equipment is "healthy" (in other words, S = F). Knowing that P (C / D) = 0 (because there can be no classification when no detection is triggered), that P (C / S) = 0 (because if the system is healthy, any possible attempt to classification will be an error), that PFA = P (S / D) = 1-P (F / D), and that POD = P (D / F), then we get: Note that each of the two indicators PCL and POD or possibly, any other data quality indicator 17 may be used independently of one another or in combination depending on the type of application of the monitoring system 11. Advantageously, the processing means 9 are configured to calculate a correlation quality indicator CQV (Correlation Quality Value) of the monitoring system 11 by calculating a correlation between the output quality value PQV and a function related to the data quality indicator 17. Of course, it is possible to use as an indicator of data quality 17 the probability of detection POD or the PCL location rate or a to another quality indicator. The data quality indicator (e.g. POD) corresponds to a measurement of a posterior error E 27 between a predicted output y and an estimated output y (i.e., E = yÿÿ). It should be noted that the correlation between PQV and E must be zero, otherwise there would have been a method to improve the estimate. However, the correlation between PQV and E2 is not zero and corresponds to a negative value that is even closer to -1 when PQV is a "good" information. Thus, for example, the correlation quality indicator CQV can be calculated as the opposite of the correlation between the square of the error E and the output quality value PQV. The QQV correlation quality indicator qualifies the relevance or accuracy of the PQV output quality and thus makes a synthesis between the structural validation and the result validation of the monitoring system 11. Advantageously, the processing means 9 are configured to calculating a Coverage Quality Value (CVR) coverage rate indicator defined as a function of a prior probability of occurrence of each fault and a volume of data necessary to cover this fault. The coverage rate corresponds to the data available to validate each type of fault. In particular, the faults not covered by the surveillance system are thus identified. To calculate the coverage, a priori knowledge is used on the occurrence of each fault (that is, the probability of occurrence of the fault). fault) and the amount of data needed to cover this fault. The expert in the field of monitoring of an aircraft engine therefore associates a percentage of coverage per fault (for example, 50% if only half of the volume of data necessary to cover this particular fault is available) . We then make a weighted average by the occurrences Oi of these covers Ci to obtain the coverage rate CVR: CVR = L OZ x Cl with E OZ = 1.

For example, the PCL location rate which is a good classification rate can be modulated using the information that made it possible to calculate the CVR coverage rate. In particular, it is possible to modulate the result of each class i by the occurrence Oi and / or the coverage Ci. Thus, the CVR fault coverage rate can be used to modulate certain data quality indicators 17 enabling confirmation of the validation. the surveillance system 11.

In addition, the processing means 9 are advantageously configured to calculate a Generic Level Genericity Level (GCL) quality indicator measuring a level of genericity of each of the modules 13 used in the composition of the monitoring system 11. The quality indicator of GCL is a structure flag that can be used for all modules and indicates the number of times a module is used.

In practice, a weight can be associated with each module 13i according to its importance in the composition of the monitoring system 11. Thus, the level of genericity of the monitoring system 11 can be calculated as a weighted average of the GCL genericity quality indicators. . When a module 13i is used more often, its quality is increased faster and therefore, the more the generic generic modules 11 monitoring system is the reliability of its validation. Moreover, the processing means 9 are advantageously configured to calculate an indicator of relevance relevance of methodology (ROM) of the monitoring system 11 as a function of a relevance score and a weighting associated with each of the 15 modules 13. Indeed, the monitoring system 11 is composed of modules 13 and each module 13i performs a particular task using a specific algorithm. The choice of modules used in the composition of the monitoring system 11 can be validated by the expert by assigning a score of relevance for each module 13i. According to the weighting associated with each module of the monitoring system 11, a note of global relevance ROM is finally deduced. It indicates whether it is still possible to make technical improvements in the choice of modules 13 to improve the monitoring system 11. FIG. 5 illustrates an example of a monitoring system 11 consisting of a chain of modules 13, according to the invention. This example shows that the input modules 13a, 13b comprise a normal data acquisition module 13a and a simulated data acquisition module 13b. One or the other of the acquisition modules 13a, 13b communicates with a normalization module 13d. During a calibration phase, the simulated data acquisition module 13b communicates with the normalization module 13d. On the other hand, in execution or normal operation, it is the normal data acquisition module 13a which communicates with the normalization module 13d. The normalization module 13d is connected to a diagnostic or anomaly detection module 13f (scoring) which is in turn connected to a classification module 13g to identify faults. The monitoring system 11 of FIG. 5 can be used for example to analyze the behavior of the aircraft engine 7 during the first seconds of the ignition process.

During the calibration or training phase, this monitoring system 11 can be validated structurally using the intrinsic quality indicators 15i described above. In addition, the result of the monitoring system 11 can be validated using the data quality indicators 17 and the other indicators described above. In execution, the normal data acquisition module 13a is configured to receive, via its input interface 13a1, time specific IN1 measurements acquired by the sensors 3 on the aircraft engine 311 and to deliver on its output interface 13a2. specific data or variables OUT1 giving indications on physical or logical elements of the equipment 5. These data are extracted from the raw time measurements (temperatures, pressures, fuel flow, rotations of trees, etc.). By way of example, the specific data may correspond to the time required for a shaft of the motor 7 to reach maximum acceleration after each starting of the engine, the gradient of the exhaust gas temperatures of the engine 7, the time of ignition, valve opening time, speed trend, etc. In addition, the data acquisition module 13a is configured to receive, via its input interface 13a1, a quality value DQV1 of the specific measurements and to deliver via its output interface 13a2, a value of output quality PQV1 associated with the data. specific OUT1.

The normalization module 13d is configured to receive via its input interface 13d1, the specific data OUT1 (IN2 = OUT1) and the corresponding output quality value PQV1 (DQV2 = PQV1) from the acquisition module 13a to normalize. these specific data and to deliver on its output interface 13d2 standardized data OUT2 independent vis-à-vis the external context and a quality value PQV2 associated with the standardized data.

From this standardized data, it is now necessary to diagnose an anomaly and then deduce a specific failure and possibly the physical component concerned. Thus, the anomaly detection module 13f which corresponds to a first output module is configured to receive via its input interface 13f1, the standardized data OUT2 (IN3 = OUT2) and the quality value PQV2 (DQV3 = PQV2 ) associated with the standardized data from the normalization module 13d to construct a signature or anomaly vector representative of the behavior of the engine 7 and to diagnose if the anomaly vector reveals an anomaly. The anomaly detection module 13f can detect an anomaly by calculating, for example, the standard of the anomaly vector knowing that the normal signatures are rather flat whereas the abnormalities are represented by important and easily interpretable variations. In addition, the anomaly detection module 13f is configured to deliver on its output interface 13f2, the abnormality vector or risk OUT3 and a quality value PQV3 associated with the anomaly vector. In case the abnormality vector or score reveals an abnormality with a relevant quality value and a relevant result quality indicator, an alarm is then triggered. More particularly, if the value of the score OUT3 exceeds a first predefined diagnostic threshold, if the quality value PQV3 associated with the score exceeds a second predefined threshold, and if for example the probability of detection POD is relevant, then a detection appears which generates an alarm. This makes it possible to trigger precise alerts based on a good detection criterion. The classification module 13g which corresponds to a second output module is configured, only in case the abnormality vector OUT3 reveals an abnormality, to receive via its input interface 13g1, the anomaly vector OUT3 (IN4 = OUT3 ) as well as the quality value PQV3 (DQV4 = PQV3) associated with the anomaly vector from the anomaly detection module 13f. In response, the classification module 13g is configured to deliver on its output interface 13g2, a failure identification measurement OUT4 and a quality value PQV4 associated with the identification measurement.

After having identified the failures by calculating for example, for each one a probability of occurrence, one can use this one to detect the faulty components. For this, we can exploit a decision grid defined by expertise and giving for each of the analyzed physical components a probability of fault when a specific failure is observed.

Claims (14)

REVENDICATIONS1. Tool for designing a monitoring system (11) of at least one equipment (5) of an aircraft engine (7), said surveillance system being configured to receive measurements specific to said equipment and to deliver a result diagnosing the state of said equipment, characterized in that it comprises: - means (9) for constructing said monitoring system (11) according to a structure assembling a chain of modules (13) dedicated to particular tasks of said monitoring system and comprising at least one input module (13a-13c) and at least one output module (13f, 13g), said modules (13) being provided with input-output interfaces (13i1-13i2) allowing them to communicating with each other and / or with said equipment (5) and its environment, each of said modules being configured to receive via its input interface (13i1), an input (IN) and an associated input quality value (DQV) at said entrance, and to deliver on its output interface (13i2), an output corresponding to a severity and an output quality value (PQV) associated with said output, - means (9) for assigning a quality value to said specific measurements that correspond to the input associated with said at least one input module (13a to 13c), - means (9) for calculating for each of said modules the output quality value (PQV) of said module (13i) according to a transfer function associating a imprecision of the output in response to inaccuracy of the input using at least one intrinsic quality indicator (15i) to said module, and - means (9) for evaluating a qualification of the monitoring system (11) as a function of a quality value of said result which corresponds to the quality value of the output (PQV) associated with said at least one output module (13f, 13g). 2. Design tool according to claim 1, characterized in that it comprises means (9) for weighting the output quality value of each of the modules (13) and for determining an overall structural quality score according to said values. weighted. 3. Design tool according to claim 1 or 2, characterized in that said structural quality indicator corresponds to a variance estimation indicator of the imprecision of the output generated in response to a noisy input during a learning of said module. 4. Design tool according to any one of the preceding claims, characterized in that said structural quality indicator (15i) corresponds to an adequacy indicator (AQV) defining the measurement of a distance between the input 36 of said module. (13i) and a calibration input defined during a learning of said module. 5. Design tool according to any one of the preceding claims, characterized in that said structural quality indicator (15i) corresponds to a robustness indicator (MQV) measuring a capacity of use of the module facing new inputs, the robustness indicator being determined during a learning of said module. The design tool according to claim 5, characterized in that the output quality value (PQV) of said module is calculated by combining at least two of said variance estimation, adequacy (AQV) indicators, and Robustness (MQV). 7. Design tool according to any one of the preceding claims, characterized in that it comprises means (9) for applying at least one data quality indicator (17) to a result of said monitoring system (11) obtained from the known specific measures comprising known faults to verify the detection of said faults. A design tool according to claim 7, characterized in that the data quality indicator (17) is a measure of a probability of detection defined as being the probability of detecting a fault when the monitored equipment exhibits a fault. anomaly, and which is determined according to a constraint of a predetermined false alarm rate defined as being the probability that the monitored equipment is healthy when the monitoring system detects a fault. A design tool according to claim 7 or 8, characterized in that the data quality indicator (17) is a measure of a location rate defined as the probability that a location is good when an anomaly equipment is observed. 10. Design tool according to any one of claims 7 to 9, characterized in that it comprises means (9) for calculating a correlation quality indicator of said monitoring system (11) by calculating a correlation between the value quality of output (PQV) and a function related to said data quality indicator. 11. Design tool according to any one of claims 7 to 10, characterized in that it comprises means (9) for calculating a fault coverage rate indicator defined according to a prior probability of occurrence of each fault and a volume of data necessary to cover the said fault. 12. Design tool according to any one of the preceding claims, characterized in that it comprises means (9) for calculating a genericity quality indicator measuring a level of genericity of each of said modules (13) used in the composition. said monitoring system (11). 13. Design tool according to any one of the preceding claims, characterized in that it comprises means (9) for calculating an overall relevance indicator of the surveillance system (11) according to a rating of relevance and reliability. weighting associated with each of said modules (13). 14. System for monitoring at least one equipment (5) of an aircraft engine (7) designed by the design tool according to any one of the preceding claims, characterized in that it comprises: at least one data acquisition module (13a, 13b) configured to receive via its input interface (13a1) specific measurements (IN1) and a quality value (DQV1) associated with said specific measurements, and to output on its output interface (13a2) specific data (OUT1) and an output quality value (PQV1) associated with said specific data, a normalization module (13d) configured to receive via said input interface (13d1), said specific data and said value corresponding output quality and to output on its output interface (13d2) standardized data (0UT2) and a quality value (PQV2) associated with said standardized data, an anomaly detection module (13f) configured é to receive via said input interface (13f1), said standardized data and said quality value associated with the standardized data and to deliver on its output interface (13f2), an anomaly vector (0UT3) and a quality value (PQV3) associated with said anomaly vector, and a classification module (13g) configured to receive via its input interface (13g1), said anomaly vector and said quality value associated with the anomaly vector and to deliver on via its output interface (13g2), a failure identification measure (0UT4) and a quality value (PQV4) associated with said identification measure.