FR3048527A1 - TEST DATA ANALYSIS SYSTEM OF AN AIRCRAFT ENGINE - Google Patents

Abstract

The invention relates to a method and a system for analyzing test data relating to a motor component, comprising: - an acquisition device (3) configured to acquire a multivariate time signal (15) of relative test data said motor component (9); - a processing device (5) configured to identify patterns (21) in said multivariate time signal and to associate tags with said patterns thereby transforming the multivariate time signal into a tag sequence (19); and a storage device (7) configured to store said label sequence (19) thereby forming an indexed database (17) for exploiting test data.

Description

TEST DATA ANALYSIS SYSTEM OF AN AIRCRAFT ENGINE TECHNICAL FIELD

The present invention relates to the field of data analysis systems of at least one engine equipment and more particularly, the exploitation of test data of an aircraft engine.

STATE OF THE PRIOR ART

Many tests are carried out on engines or parts of engines in development as well as on production engines.

During the development tests, several hundred sensors measure parameters relating to the engine and the test bench, thus generating a very large amount of data. For example, for a recording of 300 measurements at 10 Hz for one hour, we already have almost 11 million data. Thus many recordings are made during the same test at different frequencies and obtained from different sensors. It is thus impossible to manually analyze all these data.

Each test is carried out with precise and limited objectives defined by different consulting firms which are detailed in a register. During the test, the experts observe the predefined behavior indicators in the register and monitor threshold crossings and other specific behaviors (eg, very high energy, too high a temperature, too much vibration, or a sudden increase of these same parameters). The results are coded in logbooks provided during the test or in the form of pre-programmed analysis documents. At the end of the test, the data are stored in storage devices and made available to the different design offices that observe behaviors that interest them specifically. Therefore, if unusual behavior occurs that is not part of the preprogrammed analyzes by the consulting firms, the phenomenon is neither highlighted nor observed. Thus, it is estimated that less than 10% of the recorded data are actually analyzed.

Moreover, it is not possible to use an elementary metric on a space of such a large size to be able to compare all the data generated by the numerous sensors with old observations. The object of the present invention is to propose a system for analyzing test data relating to at least one piece of equipment of an engine making it possible to analyze virtually all the measurements systematically in order to identify unexpected behaviors and easily find any similar behavior already identified.

STATEMENT OF THE INVENTION

The present invention is defined by a test data analysis tool relating to a motor component, comprising: an acquisition device configured to acquire a multivariate time signal of test data relating to said engine component, a device processor configured to identify patterns in said multivariate time signal and to associate tags with said patterns thereby transforming the multivariate time signal into a tag sequence, and - a storage device configured to store said tag sequence thereby forming a base indexed data allowing the exploitation of test data.

Thus, the test as a whole is represented by a sequence of small tags stored in an indexed database for querying and comparing each new test to the previous tests, for example to know if a behavior has already been observed or not.

Advantageously, the processing device comprises: a parameterization module configured to define for each pattern a signature in the form of a vector called a signature vector comprising the values of the parameters describing the pattern according to a predetermined pattern, thus transforming the temporal signal multivariate into a sequence of signature vectors; - a classification module configured to classify the signature vectors into signature classes thereby forming a sequence of signature classes; and - an indexing module configured to assign a label to each class of signatures thus forming said label sequence.

Thus, the succession of the signature classes constitutes a compressed coding represented by the sequence of tags having an extremely smaller dimension than the multivariate signal. The replacement of the multivariate signal (or multiband) by a sequence of labels makes it possible to easily exploit the multivariate data.

Advantageously, the predetermined model can be selected from the following monovariate and multivariate mathematical models: piecewise linear model, oscillation model, autoregressive model, model of projections on a dictionary.

Thus, the predetermined pattern provides a framework for describing the pattern. The model can be selected according to its simplicity or its adaptability to the physical phenomena observed.

According to a first embodiment of the invention, the processing device comprises an interactive tool adapted to allow an operator to select a scale and a time window to identify a configuration and at least one pattern in said multivariate time signal according to bases. physical. The interactive tool allows an expert to view the multivariate time signal on a screen, choose a viewing scale and easily select a pattern and interest patterns by moving a time window on the multivariate signal.

According to a second embodiment of the invention, the processing device comprises: a selection module configured to select analysis configurations based on physical bases, each configuration being defined by a subset of single-variable signals comprising signals; exogenous and endogenous signals conditioned to said exogenous signals, - a detection module configured to detect behavioral breaks on each of the signals of said configuration and identify the instants of said breaks. an identification module configured to identify patterns in said subset of single-variable signals of said configuration by identifying stabilized and transient phases between said moments of breaks.

This makes it possible to systematically inspect all the test data automatically.

Advantageously, the break corresponds to a change in the configuration of the behavior model. For example, a change of trend according to a piecewise linear model, a change of frequency or amplitude according to an oscillation model, a mode change according to an autoregressive model, a cessation of the validity of a function relating to a multivariate model.

Advantageously, the processing device is configured to iteratively construct a plurality of tag sequences.

This allows you to progressively increase the labels in the indexed database. The labels can thus be completed or modified without taking up the whole compression.

Advantageously, the processing device is configured to compare each current multivariate time signal with the set of historized multivariate signals using an edit distance on the label sequences. The comparison makes it possible to classify between the signals.

Indeed, this makes it possible to classify the old tests by similarity of behavior during parts of the test and makes it possible to know if a behavior has already been observed or not. Thus, we can exploit not only production engines but also developing engines for which the data are less homogeneous.

Advantageously, the analysis system comprises a set of storage units associated with a corresponding set of processors, thus forming a distributed database.

The group of processors and storage units is used to process data in parallel. We can therefore realize different queries in parallel.

Advantageously, the processing device is configured to link the test data to identifiers relating to experts who have requested, performed or analyzed the tests.

Advantageously, the analysis system comprises a transmission device configured to transmit messages relating to specific behaviors to the experts interested in these behaviors.

This makes it possible to efficiently and automatically use the messages by highlighting any novelty detected during a test thereby providing relevant information to the appropriate experts. The invention also relates to a method for operating test data relating to a motor component, comprising the following steps: - acquiring a multivariate time signal of test data relating to said motor component, - identifying patterns in said temporal signal multivariate, - associating labels to said patterns thus transforming the multivariate time signal into a sequence of tags, and - storing said sequence of tags thus forming an indexed database for the exploitation of test data.

BRIEF DESCRIPTION OF THE DRAWINGS

Fig. 1 schematically illustrates hardware means implemented in the test data analysis system of at least one engine component, according to one embodiment of the invention;

Figs. 2A and 2B show a first example of a pattern describing the thrust of an aircraft engine, according to the invention;

Figs. 3A and 3B show a second example of a pattern describing the vibration of an aircraft engine according to the invention;

Fig. 4 schematically illustrates a multivariate signal comprising a set of analysis configurations, according to the invention;

Fig. 5 schematically illustrates the indexing of a multivariate signal, according to a preferred embodiment of the invention;

Fig. 6 schematically illustrates an interactive identification of patterns, according to a first preferred embodiment of the invention;

Fig. 7 schematically illustrates an automatic identification of patterns, according to a second preferred embodiment of the invention;

Figs. 8A-8B illustrate the detection of breaks on a monovariate signal according to a piecewise linear behavior model, according to the invention;

Fig. 9 illustrates the identification of a pattern related to a change of position of a thrust lever, according to the invention;

Fig. 10 schematically illustrates a method of operating a new test, according to a preferred embodiment of the invention; and

Fig. 11 schematically illustrates a test data analysis system of a motor component, according to another preferred embodiment of the invention.

DETAILED PRESENTATION OF PARTICULAR EMBODIMENTS

The principle of the invention is to progressively index the data generated by engine test benches to optimize the analysis and exploitation of all these data.

Fig. 1 schematically illustrates hardware means implemented in the test data analysis system of at least one engine component, according to the invention.

For the sake of brevity, the expression "engine component" will be used hereinafter to designate the entire engine or only a part of the engine.

The analysis system 1 comprises an acquisition device 3, a processing device 5 and a storage device 7. The processing device 5 is a computer or computer adapted for executing one or more programs of computer comprising program code instructions, stored in computer memories and designed to implement the test data analysis system of a motor component, according to the invention.

During a test, the engine component 9 is mounted in the test bench 11 and a large number of sensors 13 (a few hundred) are used to measure parameters relating to the engine component 9 and the test bench 11. These many measurements (hundreds of thousands) are acquired at different frequencies. Some measurements (for example, vibration measurements) are recorded at a high frequency of the order of 50 kHz while other measurements (eg pressure measurements) are recorded at frequencies of about 1 Hz to 100 Hz. Thus, a test is represented by a succession of multivariate recordings of many measurements acquired at different frequencies. In the following, the multivariate time signal (or global signal) will be the succession of multivariate data recorded during a test.

In accordance with the invention, the analysis system 1 is configured to create an indexed database 17 by compressing each multivariate time signal 15 into a small label sequence 19. The indexed database 17 thus makes it possible to make queries and easily compare each new test to the previous tests.

Indeed, the acquisition device 3 is configured to acquire the multivariate time signal 15 of current test data relating to the engine component 9.

In addition, the processing device 5 is configured to identify patterns 21 in the multivariate time signal 15. A pattern 21 is an observation model defined over a time interval of interest and describing, for example, the behavior of a physical parameter. in relation to another (or other) physical parameter (s).

Figs. 2A and 2B show a first example of a pattern describing the thrust of an aircraft engine.

More particularly, FIG. 2A is a monovariable SI signal (i.e. a variable curve) representing the rotational speed of the motor shaft over a time interval and FIG. 2B is another monovariate signal S2 representing the temperature of the engine compressor over the same time interval. The time interval of interest starts notably with an increase in the level of the thrust (at the instant 20:17:35) and ends with a stabilization of the speed of rotation of the motor shaft as well as that of the temperature of the compressor.

According to this first example, the pattern thus models the variation of the thrust with respect to the speed of rotation of the shaft and the temperature of the compressor. In particular, the thrust enters the transient phase at the instant 20:17:35 and then the rotation speed reaches a stationary phase shortly after the instant 20:19:00. The temperature varies in a different way before also reaching a stationary phase. Thus, the pattern may advantageously be characterized by transient and stationary phases defined by parameters according to a distinctive mathematical model.

Thus, the pattern of Figs. 2A and 2B can be parameterized by the value of the increase of the thrust, the position of the thrust lever (or throttle) and possibly the initial values of the speed of rotation and the temperature.

Figs. 3A and 3B show a second example of another pattern describing the vibration of an aircraft engine.

More particularly, FIG. 3A illustrates a single-speed signal S3 of the rotational speed of the motor shaft over a time interval while FIG. 3B illustrates another monovariate signal S4 of the motor vibration over the same time interval. Here the pattern describes the variation of the vibration energy conditioned by the speed of rotation of the tree. In particular, it highlights a fact that can be regarded as surprising where the S4 vibration energy continues to increase for at least 10 minutes (between 17:45 and 17:55 times) despite the stabilization of the speed of rotation 53 over the same period. It may be interesting for an expert to look for similar reasons to better analyze this phenomenon.

Moreover, since the number of dimensions of the multivariate signal is very large, it is not possible to simultaneously analyze all the single-phase signals that make up this global signal 15. Then, the processing device 5 is advantageously adapted to allow segmentation ( in an interactive or automatic manner) of the overall signal into interesting small portions or pieces of small size called analysis configurations. Each analysis configuration comprises a small set of monovariable signals selected according to physical bases or expertise. The joint behavior of the single-variate signals on each analysis configuration is modeled by a corresponding pattern.

Indeed, FIG. 4 schematically illustrates a multivariate signal 15 comprising a set of analysis configurations 151. The global multivariate signal can thus be considered as a conjunction of the different analysis configurations 151. Each analysis configuration 151 corresponds to a selection of signals. monovariates (or measures) some of which are considered as exogenous and others endogenous. Note that depending on the case studies, some signals can be considered endogenous in one analysis configuration and exogenous in another analysis configuration.

It is recalled that an exogenous signal describes a context variable that is representative of the context. On the other hand, an endogenous signal describes an observable variable to follow and to analyze according to its context of observation. The identification of an endogenous or exogenous variable can be carried out according to technical criteria established by expertise according to for example the relevance of the physical correlations of the phenomena.

In the example of Figs. 2A and 2B, the analysis configuration comprises measurements of exogenous variables represented by the single-speed rotational speed and temperature signals as well as a measurement of an endogenous variable corresponding to the thrust. Moreover, in the example of Figs. 3A and 3B, the analysis configuration comprises an exogenous signal represented by the rotational speed measurements and an endogenous signal represented by the measurements of the vibration.

Thus, each analysis configuration 151 makes it possible to identify at least one pattern 21 on a limited number of monovariate signals including notably endogenous and / or exogenous type signals. The patterns 21 are advantageously identified by stationary or transient failure phases observed on the exogenous signals of the analysis configurations 151.

These phases are advantageously described by parameters representative of a signature of the phase following a predetermined mathematical model. The predetermined mathematical model can be a piecewise linear model, an oscillation model, an autoregressive model, or a model of projections on a learned or non-learned dictionary. The latter is a particular model using a database that can be extracted from the observations or that can be generic exploiting a mathematical dictionary of specific models.

In particular, the stationary phases can be parameterized by averages, slopes, initial values, correlations between variables and possibly autoregressive processes. As for transient phases, they can be parameterized by piecewise linear mathematical models or describing predefined evolutions modes that are coherent with physics, transfer functions highlighting correlations and stabilization speeds of endogenous variables, which depend on possibly exogenous conditions.

On the other hand, the monovariable signals of each analysis configuration 151 (shown in FIG. 7) can advantageously be smoothed and normalized in order to construct compressed and standardized patterns.

It should be noted that the construction of patterns 21 can be performed interactively (described below in connection with Fig. 6) or in a fully automatic manner (described hereinafter in connection with Fig. 7).

In addition, the processing device 5 is configured to associate tags 191 with the patterns 21 thereby transforming each current multivariate time signal into a tag sequence 19. Labels 191 are small names or labels that index the multivariate data of the test (see Fig. 4).

Fig. 5 schematically illustrates the indexing of a multivariate signal by the processing device, according to a preferred embodiment of the invention.

More particularly, the processing device 5 comprises parameterization modules 51, classification 53 and indexing modules 55 configured to create the labels 191.

The parameterization module 51 is configured to define for each pattern 21 a signature in the form of a vector called Vi signature vector comprising the parameters describing the pattern according to a predetermined mathematical model.

More particularly, the components of a signature vector Vi are the parameters defining the pattern 21. By way of example, the components of a vector Vi may comprise average values, variances, slopes, initial values, relationships between residues (correlations, temporal dependencies, etc.). The parameterization module 51 maps a signature vector Vi to each pattern 21 thus transforming the multivariate time signal 15 into a sequence of signature vectors Vi,...,... In other words, the test as a whole is compressed in this sequence of signature vectors Vi,... Vi, ... Vn.

Advantageously, the intersections of the time intervals relative to the different analysis configurations 151 form global intervals making it possible to simultaneously stack the set of signature vectors Vi,...,........ .Sub.n of these configurations into a vector of global signature.

Furthermore, since the signature vectors Vi, ... Vi, ... Vn are small, the processing device 51 is advantageously configured to measure the difference between these vectors in order to be able to measure distances between patterns 21. By way of example, the difference can be measured using a Euclidean distance if the parameters are normalized according to, for example, the methods described in the applicant's patent application FR2376988.

In addition, the classification module 53 is configured to classify the signature vectors Vi,...,...,... Vn (for example, by means of a segmentation algorithm) according to signature classes thus forming a sequence of classes Ci, ... Q, ... Cp of signatures. The succession of these classes Ci, ... Ci, ... Cp of signatures constitutes a compressed coding.

Advantageously, the processing module 5 is further configured to determine the distribution of similarities between the elements of each class C of signatures. For example, the median distance between signature vectors indicates the average difference between two vectors in the signature class. A quantile less than or equal to 25% provides information on the notion of proximity between two signature vectors.

Advantageously, the center of each class Q of signatures (for example the centroid of the signature vectors that compose it) makes it possible to detect the elements distant from their center. For example, elements beyond a 99% quantile can be considered as outliers.

Finally, the indexing module 55 (or labeling) is configured to assign a label 191 to each class C, signatures thus forming the label sequence 19. The label 191 of each class Ci of signatures is used to represent this class. The multivariate signal 15 is therefore coded in a class chain

Cl, ... Ci, ... Cp of signatures that can be seen at the highest level as a string of labels 19, while retaining more accurate information using the vector Vi signature of each class Q.

Thus, an "edit distance" distance can be easily used to measure the gap between two pieces of label strings. Then we can make rankings of the most similar to at least similar.

Advantageously, the processing device 5 is further configured to link the test data to identifiers relating to experts who requested, performed or analyzed the tests. In particular, each test can be annotated by the identifiers of the corresponding experts. Thus, at each new test, the experts who may be interested in the results of this new test are automatically identified.

Fig. 6 schematically illustrates an interactive identification of patterns by the processing device, according to a first embodiment of the invention.

According to this first embodiment, the processing device 5 comprises an interactive tool 57 comprising viewing interfaces 59 and control interfaces 61 (for example, interfaces combining touch screens and control points). Thus, the multivariate time signal (including test data and corresponding digital records) can be viewed and manipulated on the interactive tool 57.

Indeed, the interactive tool 57 is adapted to allow an operator (or an expert) to select a time window and a corresponding scale to identify according to physical considerations or expertise a configuration and at least one pattern in the temporal signal multivariate 15.

More particularly, once the operator has manually selected an analysis configuration 151 of interest comprising endogenous (s) and exogenous monovariable (s) signals, the interactive tool 57 displays on the screen 59 the curves 55, 56 of the selected signals. The operator can thus select an interval 11 of interest by applying temporal zooms on the curves. The interactive tool 57 includes an application allowing the operator to apply a windowing process to search for patterns of regressions of the endogenous signal (s) according to the context of the exogenous signal (signals). (s). The interactive tool 57 also includes an application for smoothing and normalizing the variables of the selected signals.

Thus, by choosing an appropriate scale and moving the time window, the operator can detect the interesting breaks and therefore identify the corresponding patterns 21.

In addition, the interactive tool 57 includes an application configured to discover possible relationships between a pattern 21 and other variables not belonging to the identified pattern thus allowing the expert to further enrich the latter. The interactive tool 57 also includes a recording tool 63 enabling the operator to record the identified patterns 21 as well as any analyzes or comments by the operator.

Once the patterns 21 are identified, the processing device 5 is configured to associate tags thereby transforming each multivariate time signal into a tag sequence as previously described in connection with FIG. 5.

Indeed, the parameterization module 51 is configured to define for each identified pattern a signature vector comprising the parameters describing the pattern. The classification module 53 is configured to classify the signature vectors into different classes of signatures. Finally, the indexing module 55 is configured to assign a tag to each class thereby forming the tag sequence.

Fig. 7 schematically illustrates an automatic identification of patterns by the processing device, according to a second preferred embodiment of the invention.

According to this second embodiment, where the identification of the patterns is fully automated, the processing device 5 comprises a selection module 71, a detection module 73, and an identification module 75.

The selection module 71 is configured to select analysis configurations 151 based on physical considerations or expertise. Advantageously, the configurations are selected from a list predefined by the experts. This list can for example be built by the interactive tool 57.

Each analysis configuration 151 is defined by a subset of monovariate signals comprising exogenous signals and endogenous signals conditioned to the former. More particularly, each analysis configuration 151 is constructed by correlation calculations between different variables. These can be advantageously limited according to initial expert configurations. Indeed, the expertise accumulated during the selection of the signals by means of the interactive tool 57 (Fig. 6) makes it possible to avoid a combinatorial proliferation of "motifs" without physical foundation. Optionally, configurations can be built automatically by recognizing physical units of variables or sensor identifiers.

The detection module 73 is configured to automatically detect ruptures r1, r2 of the behavior model on each of the signals of the analysis configuration 151 and to identify the instants t1, t2 of these ruptures.

More particularly, the detection module 73 applies a pattern change detection algorithm that segments each single-piece signal into pieces corresponding to simple mathematical models (such as piecewise linear models). For example, for a piecewise linear behavior model, the break corresponds to a change of trend (i.e. variation of slope according to this model). For an oscillation model, the rupture corresponds to a change of frequency and / or amplitude. For an autoregressive model (i.e. a model that depends on the immediate past), the break corresponds to a change of model. For a multivariate model describing for example a correspondence between an endogenous variable and an exogenous variable, the rupture corresponds to a cessation of the validity of the correspondence between the two variables.

Figs. 8A-8B illustrate the detection of breaks on a monovariate signal 57 following a piecewise linear behavior model. More particularly, FIG. 8A illustrates a monovariate signal 57 of vibration over a time interval. Fig. 8B illustrates the segmentation of the vibration signal 57 into three pieces of straight lines d1, d2, d3 thus identifying the breaks r1, r2 and the times t1, t2 of these breaks. For example, the detection module 73 applies a PELT model change detection algorithm (Pruned Exact LinearTime) or a modified version of PELT to detect breaks. The PELT algorithm is based on linear regression easily and quickly calculable. Moreover, this kind of algorithm is flexible and can be easily modified to be applied to an autoregressive model.

In addition, the identification module 75 is configured to identify patterns 21 in the multivariate sub-signal (ie the subset of single-variated signals) of the analysis configuration 151 by identifying stabilized and transient phases between the instants of breaks detected by the detection module 73.

The multivariate sub-signal of the analysis configuration 151 is in fact modeled from the conjunction of the breaks (i.e. all the breaks) of the different single-component signals composing it taking into account the small delays between the breaks. The small delays in the correspondence of the breaks detected in the multivariate sub-signal provide information on the causality between the variables of the different single-variate signals and thus make it possible to identify interesting and recurrent patterns.

Fig. 9 illustrates the identification of a pattern related to a change of position of a lever thrust. This figure shows in particular the conjunction of the breaks between the modelizations of a signal 58 for detecting the position of the lever and a signal 59 of rotational speed. The stabilized and transient phases can easily be identified between the moments of rupture. In addition, in the time interval [0, 50], it is possible to clearly distinguish the causality delay between the instant of change of position of the lever and the instant of the beginning of decay of the speed of rotation.

In addition, as previously described in connection with FIG. 5, once the patterns 21 are identified, the processing device 5 is configured to associate tags with the different patterns.

Thus, the parameterization module 51 is configured to define for each pattern a signature vector comprising the parameters describing the pattern according to a predetermined mathematical model. The classification module 53 is configured to classify the signature vectors into different classes of signatures. Finally, the indexing module 55 is configured to assign a tag to each class of signatures thus forming the tag sequence.

This second embodiment makes it possible to systematically inspect all the test data automatically.

Advantageously, a new sequence of labels is constructed at each new test. Indeed, the processing device 5 is configured to iteratively generate a plurality of label sequences 19. This makes it possible to progressively increase the number of tags 191 (with links to comments, documents or past records) in the indexed database 17. The labels 191 can thus be completed or modified without taking up the entire compression.

Indeed, FIG. 10 schematically illustrates a method of operating a new test, according to a preferred embodiment of the invention. Step El concerns the acquisition of a new multivariate time signal 15a.

Steps E2-E6 relate to the compression of the multivariate time signal 15a.

Indeed, in step E2, the processing device 5 decomposes the multivariate time signal 15a into several analysis configurations 151. In step E3, the processing device 5 identifies (automatically or interactively) the various patterns 21. Indeed, the joint behavior of groups of measures on each analysis configuration 151 is modeled by a corresponding pattern. In step E4, the processing device 5 represents the patterns 21 by a sequence of signature vectors Vi, ... V ,, ... Vn. Each signature vector has as components the parameters describing the corresponding pattern. In step E5, the processing device 5 classifies the signature vectors thus transforming the sequence of signature vectors into a sequence of classes Ci, ... Ci, ... Cp of signatures. In step E6, the processing device 5 transforms the sequence of classes Ci, ... Ci, ... Cp of signatures into a sequence 19a of labels by assigning a label 191 representative to each class of signatures. In step E7, the processing device 5 compares this new multivariate time signal 15a with the set of historized multivariate signals. Indeed, the processing device 5 is configured to use an editing distance to compare the new label sequence 19a with the label sequences 19 stored in the database 17. These comparisons will advantageously serve as a basis for a new classification of the tests as a whole. Thus, the different tests are classified by similarity of behavior during parts of each test. This allows to quickly know if a behavior has already been observed or not.

Moreover, the processing device 5 is configured to advantageously make rapid pre-comparisons between each current multivariate 15a temporal signal and the historized multivariate signals.

Indeed, once the multivariate time signal 15a of a new test has been coded as a class sequence, the processing device 5 uses the latter to perform a fast pre-query. Then, the processing device 5 identifies the class of each transient or stationary phase of the test and identifies the tests having similar class phases. This makes it possible to quickly have a first statistical result corresponding to the proportion of tests having phases of the same type. From this screening of tests and phases (indicating for each phase a list of similar tests), the processing device 5 easily performs a similarity search on the signature vectors. We can already identify similar past tests and from a measure of the average similarity (for example, the median intra-class distances) within a class, the processing device 5 checks whether these tests are really close to the new test. The processing device 5 is thus configured to identify phase by phase, the most similar tests. It is interesting to mention the list of similar tests especially if there are few similar tests.

In addition, the processing device 5 is configured to advantageously perform a similarity search on successions of labels on the sequence 19 of tags representing the test. Indeed, the processing device 5 looks in the database 17 if a particular sequence of tags 191 exists on another test. It is interesting to note the list of similar sequences especially if there are very few.

Fig. 11 schematically illustrates a test data analysis system 101 of a motor component, according to another preferred embodiment of the invention.

According to this embodiment, the analysis system 101 comprises (in addition to the hardware means described above) a set 81 of storage units 83 associated with a corresponding set of processors 85.

Advantageously, the different tag sequences 19 generated by the different tests can be stored (at least temporarily) on the set 81 of storage units thus forming a distributed database. Thus, one can very quickly perform parallel operations to access different data and exploit these data for statistical analysis 87 or to look for similar reasons.

In addition, the analysis system 101 includes a transmission device 89 configured to automatically transmit messages relating to specific or unusual behaviors to experts 91 singularly interested in these behaviors.

Indeed, since the tests are requested, carried out and studied by design offices or experts, the processing device 5 is configured to identify specific applicants, operators and authors of analyzes in order to annotate each test. using the identifiers of the consulting firms and experts concerned. The annotation thus takes into account the authors of the specifications before the test, the authors of the logbooks during the test as well as those of the studies or analyzes after the test. Each of the documents used includes names of authors or services as well as specific themes corresponding respectively to the types of tests, the anomalies detected during the test and the types of behavior observed. In addition, a proximity graph between experts can be established according to location considerations (same design offices) or similar areas of interest (experts working on similar tests), etc.

Thus, when analyzing a new test, when the processing device 5 finds similar tests, the transmission device 89 directly sends an information message to the identifiers or experts 91 associated with the similar tests.

It will be noted that the analysis system according to the invention makes it possible to exploit data of any kind generated not only during tests on production engines but also on developing engines for which the data are less homogeneous.

The analysis system according to the invention can be applied to all types of engines and in particular to aircraft engines.

Claims (10)

A system for analyzing test data relating to a motor component, characterized in that it comprises: an acquisition device (3) configured to acquire a multivariate time signal (15) of relative test data said motor component; - a processing device (5) configured to identify patterns (21) in said multivariate time signal and to associate tags with said patterns thereby transforming the multivariate time signal into a label sequence (19), and - a storage device (7) configured to store said label sequence (19) thereby forming an indexed database (17) for exploiting test data. 2. Analysis system according to claim 1, characterized in that the processing device comprises: a parameterization module configured to define for each pattern a signature vector comprising parameters describing the pattern according to a predetermined pattern, thereby transforming the pattern; multivariate time signal in a sequence of signature vectors; - a classification module configured to classify the signature vectors into signature classes thereby forming a sequence of signature classes; and - an indexing module configured to assign a tag to each class of signatures thus forming said label sequence. 3. Analysis system according to claim 2, characterized in that the predetermined model is selected from the following mathematical models: piecewise linear model, oscillation model, autoregressive model, and projection model on a dictionary. 4. Analysis system according to any one of the preceding claims, characterized in that the processing device comprises an interactive tool adapted to allow an operator to select a scale and a time window to identify a configuration and at least one pattern in said multivariate time signal according to physical considerations. 5. Analysis system according to any one of claims 1 to 3, characterized in that the processing device comprises: a selection module configured to select analysis configurations based on physical bases, each configuration being defined by a subset of monovaried signals comprising exogenous signals and endogenous signals conditioned to said exogenous signals; a detection module configured to detect behavioral breaks on each of the signals of said configuration and to identify the instants of said breaks; identification configured to identify patterns in said subset of single-valued signals of said pattern by locating stabilized and transient phases between said break times. Analysis system according to Claim 5, characterized in that the break corresponds to a change of trend according to a piecewise linear model, to a change of frequency or of amplitude according to an oscillation model, to a change of following an autoregressive model, to a cessation of the validity of a function according to a multivariate model. 7. Analysis system according to any one of the preceding claims, characterized in that the processing device is configured to iteratively construct a plurality of tag sequences. The analysis system of claim 7, characterized in that the processing device is configured to compare each current multivariate temporal signal with the set of historically multivariate signals using an edit distance on the label sequences. 9. Analysis system according to any one of the preceding claims, characterized in that it comprises a transmission device configured to transmit messages relating to specific behaviors to experts interested in these specific behaviors. 10. A method for analyzing test data relating to a motor component, characterized in that it comprises the following steps: -acquire a multivariate time signal of test data relative to said motor component, - identify patterns in said multivariate temporal signal, associating labels with said patterns thus transforming the multivariate time signal into a sequence of tags, and storing said sequence of tags thus forming an indexed database allowing the exploitation of test data.