Time series partitioning method

The partitioning method uses unsupervised learning and neural networks to group time series data systematically, addressing the need for prior knowledge in predictive maintenance and enhancing health monitoring processes across diverse databases.

FR3118246B1Active Publication Date: 2024-11-01SAFRAN ELECTRONICS & DEFENSE (FR) +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
FR2020013861
Authority / Receiving Office
FR · FR
Patent Type
Patents
Current Assignee / Owner
Filing Date
2020-12-21
Publication Date
2024-11-01
Estimated Expiration
2040-12-21

AI Technical Summary

Technical Problem

Existing predictive maintenance methods require a priori knowledge of equipment failure modes and are not systematic across different databases, limiting their applicability.

Method used

A partitioning method that uses unsupervised learning to group time series data based on similarity measures, generating images from distance matrices and applying neural networks for pattern recognition to define severity classes without requiring prior knowledge of failure modes.

Benefits of technology

Enables systematic partitioning of time series data across various databases, allowing for predictive maintenance without prior knowledge of equipment failure modes, improving the accuracy and applicability of health monitoring processes.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

Partitioning method comprising the steps of: acquiring an observation matrix (10) comprising time series (, ,…, ); for each time series: calculating a distance matrix comprising distance values ​​between the elements of the time series, then generating a primary image from said distance matrix; implementing a learning algorithm to segment the primary image to obtain a segmented image; defining a primary boundary signal representative of the boundaries from the segmented image; merging the primary boundary signals to obtain a global boundary signal (52), and defining classes from the global boundary signal (52). FIGURE FROM THE ABRIDGED: Fig. 2
Need to check novelty before this filing date? Find Prior Art

Description

Title of the invention: Method for partitioning time series

[0001] The invention relates to the field of Health-Monitoring processes.

[0002] BACKGROUND OF THE INVENTION

[0003] A Health-Monitoring process, commonly called PHM (for Prognostics & Health Management) although we also speak, but less frequently, of conditional maintenance, predictive maintenance or forecast maintenance, is used to anticipate the degradation of equipment or a system, particularly in the aeronautical field.

[0004] The predictive maintenance process thus makes it possible to avoid a breakdown, to optimize the service life, to anticipate and plan maintenance operations (and possible removal), and therefore to reduce the costs of maintenance operations for the equipment or system.

[0005] The predictive maintenance process comprises five main steps: a data acquisition step, a data processing (or pre-conditioning) step, a signature extraction step, a diagnostic step and a prognostic step.

[0006] The data acquisition step consists of acquiring data produced by sensors measuring parameters representative of the performance of the equipment or system. Data acquisition is, for example, carried out via tests, or by downloading onto the equipment or system when it is operational.

[0007] The data preconditioning step consists in particular of filtering the data to remove the measurement noise present in the data.

[0008] The signature extraction step consists of processing the data to extract one or more signatures representative of the state of the equipment or system and, possibly, of a degradation in the operation of the equipment or system.

[0009] The diagnostic step consists of determining the health status of the equipment or system at the current time from among a set of severity classes (failure modes) each associated with a severity level.

[0010] The prognosis step consists of estimating the residual life (or RUL, for Remaning Useful Life) of the equipment or system.

[0011] Note that it is possible to dispense with one or more steps, or to mix several steps. For example, if we only use algorithms deep learning for the diagnostic step and the prognostic step, it is not necessary to perform the signature extraction step (and it is even not recommended).

[0012] Patent application FR1907561, filed on 05 / 07 / 2019, describes a method for estimating the residual lifetime of equipment such as an electromechanical actuator. This document teaches a partitioning method in which the number of severity classes used is determined by the joint optimization of a temporal consistency criterion and an energy criterion. This method requires initially defining a number of classes in an “intuitive” manner, evaluating the temporal consistency and the level of convergence of the energy criterion associated with this initial number of classes, then, if necessary, repeating these steps until satisfactory temporal consistency and a level of convergence of the energy criterion are obtained.

[0013] This method is effective but requires relatively in-depth knowledge of the data in the database used and, in particular, of the failure modes of the equipment or systems taken into account.

[0014] SUBJECT OF THE INVENTION

[0015] The subject of the invention is a partitioning method which can be implemented more systematically on various databases, and without requiring a priori knowledge relating to the severity classes present in the observations of the database used. Summary of the invention

[0016] In order to achieve this goal, a partitioning method is proposed, implemented in a processing unit and comprising the steps of: • acquire an observation matrix produced from a database comprising observations of parameters carried out on at least one piece of equipment, the observation matrix comprising time series each comprising elements which are observations of a parameter; • for each time series: • calculating a distance matrix comprising distance values ​​between the elements of the time series, then generating from said distance matrix a primary image (20, 21) comprising pixels having levels representative of the distance values ​​of the distance matrix; • implement a learning algorithm to segment the primary image so as to obtain a segmented image comprising patterns delimited by boundaries, and produce a segmented matrix comprising pixel levels of the segmented image; • define from the segmented matrix a primary boundary signal representative of the boundaries; - merge the primary boundary signals to obtain a global boundary signal, and define classes from the global boundary signal.

[0017] The calculation of a distance matrix associated with each time series, and the representation of the distance matrices by images, makes it possible to extract group information in the time series in a systematic and fully automatic manner. The method according to the invention therefore does not require a priori knowledge relating to the failure modes of the equipment, and its use can be extended to different types of equipment or systems.

[0018] We further propose a partitioning method as previously described, in which the distance matrix is ​​a Gram matrix such that:

[0020] where Ej is a time series, the (xi ) are the elements of said series k'\ <k<n temporal, and where xi^ is a distance value between x] and xf„, with / < Z < n and 1 < m < n.

[0021] We further propose a partitioning method as previously described, in which the patterns are squares having a diagonal coincident with a portion of a diagonal of the primary image.

[0022] We further propose a partitioning method as previously described, in which the segmentation of the primary image consists of determining a probability of belonging of the pixels of the primary image to a first class representing the patterns or to a second class representing a background of the primary image.

[0023] We further propose a partitioning method as previously described, in which the learning algorithm is a convolutional neural network of the U-NET type.

[0024] A partitioning method as previously described is further proposed, further comprising the step of generating artificial primary images from simulation functions, and of training the learning algorithm using the artificial primary images.

[0025] We further propose a partitioning method as previously described, in which the simulation functions are piecewise continuous constant functions or increasing multi-slope functions.

[0026] A partitioning method is further provided as previously described, in which the definition of the primary boundary signal comprises the step of calculating a statistical function on sets of elements which each comprise elements of one of the anti-diagonals of the segmented matrix.

[0027] A partitioning method as previously described is further provided, further comprising the step of filtering the primary boundary signal using an iterative sliding window method.

[0028] Further provided is a partitioning method as previously described, wherein merging the primary boundary signals involves summing the primary boundary signals to obtain a summed boundary signal, and then estimating an empirical density function of the summed boundary signal to produce the overall boundary signal.

[0029] We further propose a partitioning method as previously described, in which, to calculate the empirical density function, a first formula is used in the case where the parameter observations have been carried out on the same equipment or system, and a second formula is used in the case where the parameter observations have been carried out on separate equipment or systems representing non-identically distributed realizations.

[0030] We further propose a partitioning method as previously described, in which the first formula is: [0031> dW =

[0032] where xi are elements of the summed boundary signal, K is a kernel function, h represents a bandwidth of the kernel function and n is a number of the elements of each time series.

[0033] We further propose a partitioning method as previously described, in which the second formula is:

[0034] ~ j [nm d(x\^x m ) = ^^11^- / /

[0035] where the are elements of the primary boundary signals, K is a kernel function, h^hj^hn represent a bandwidth of the kernel function, m is a number of non-identically distributed realizations and n is a number of the elements of each time series.

[0036] We further propose a partitioning method as previously described, in which the second formula is:

[0037] d(x^xm)

[0038] where the are elements of the primary boundary signals, Æ is a kernel function, the h\JljJln represent a bandwidth of the kernel function, m is a number of non-identically distributed realizations, n is a number of elements of each time series, and in which we have:

[0039] iej ]; mj te| gue = 1 •

[0040] A processing unit is further provided comprising at least one component of processing in which the partitioning process just described is implemented.

[0041] A processing unit as previously described is also proposed, comprising a GPU in which at least the learning algorithm is implemented for segmenting the primary images.

[0042] A computer program is also provided comprising instructions which cause the processing unit as previously described to execute the steps of the partitioning method which has just been described.

[0043] A computer-readable recording medium is further provided, on which the computer program just described is recorded.

[0044] The invention will be better understood in light of the following description of a particular non-limiting embodiment of the invention. Brief description of the drawings

[0045] Reference will be made to the attached drawings, among which:

[0046] [Fig.l] [Fig.l] represents a turbojet engine and electromechanical actuators of a thrust reverser;

[0047] [Fig.2] [Fig.2] represents steps of the partitioning process according to the invention;

[0048] [Fig.3] [Fig.3] represents a processing unit in which the partitioning process;

[0049] [Fig.4] [Fig.4] is a figure similar to [Fig.2], but more detailed;

[0050] [Fig.5] [Fig.5] represents a primary image obtained for the LLA cylinder;

[0051] [Fig.6] [Fig.6] represents a primary image obtained for the URA cylinder;

[0052] [Fig.7] [Fig.7] represents a “simple” discrete convolution product;

[0053] [Fig.8] [Fig.8] represents the architecture of a U-type convolutional neural network NET ;

[0054] [Fig.9] [Fig.9] represents a piecewise continuous constant function, a artificial primary image and the associated mask;

[0055] [Fig. 10] [Fig. 10] represents a segmented image obtained from the image of [Fig.5];

[0056] [Fig. 11] [Fig. 11] represents a segmented image obtained from the image of [Fig.6];

[0057] [Fig. 12] [Fig. 12] represents the images of figures 5 and 10 and a graph illustrating the result of the segmentation;

[0058] [Fig. 13] [Fig. 13] represents the steps implemented to define the primary boundary signals and the global boundary signal;

[0059] [Fig. 14] [Fig. 14] represents the steps implemented to obtain a primary boundary signal;

[0060] [Fig. 15] [Fig. 15] represents the filtering of a primary boundary signal;

[0061] [Fig. 16] [Fig. 16] represents a graph comprising a boundary signal primary (after filtering) obtained for the LLA cylinder;

[0062] [Fig. 17] [Fig. 17] represents a graph comprising the summed boundary signal obtained for the LLA cylinder;

[0063] [Fig. 18] [Fig. 18] represents two graphs comprising curves illustrating the kernel function density estimation of the summed boundary signal;

[0064] [Fig. 19] [Fig. 19] represents a graph comprising the signal of [Fig. 17] and a curve of the empirical density calculated with a Gaussian kernel function. DETAILED DESCRIPTION OF THE INVENTION

[0065] The invention is here implemented and has been experimentally validated on the cylinders (electromechanical actuators) of an electrical actuation system for the thrust reversers of an aircraft. This system is called ETRAS®.

[0066] The ETRAS® system makes it possible to open a portion of the nacelle of each turbojet engine of the aircraft to divert the airflow of said turbojet engine towards the front of the aircraft. This then results in braking of the aircraft when it is on the ground, which is added to the braking carried out by the brakes fitted to the braked wheels of the aircraft. Each brake of a braked wheel here comprises a stack of carbon discs and hydraulic actuators which apply a braking force to the stack of carbon discs.

[0067] The ETRAS® system is mainly used when landing the aircraft but can be used as a braking aid.

[0068] emergency during an aborted takeoff (or ATO for Aborted Take-Off).

[0069] Part of the ETRAS® system is shown in [Fig.l]. The turbojet 1 comprises a propeller 2 and a body 3. The nacelle 4 surrounds the turbojet 1. The ETRAS® system comprises for each nacelle 4 four cylinders 5 (only two are visible in [Fig.l]) as well as a harness 6.

[0070] When the airbrake is opened, the part of the nacelle 4 secured to the nuts of the ball screws of the jacks 5 translates to reveal the grid 7 and allow the deflection of the air flow captured by the front of the turbojet 1.

[0071] The four cylinders 5 are named here by the following acronyms: URA (for Upper Right Actuator), ULA (for Upper Left Actuator), LRA (for Lower Left Actuator) and LLA (for Lower Left Actuator).

[0072] Accelerometers are installed on each of these jacks 5. The accelerometers measure the vibration levels experienced by the jacks 5.

[0073] The system was used during numerous opening / closing cycles, under different conditions of use, and in particular during the tests implemented for the qualification of the ETRAS® system. The data from the endurance cycles and therefore the signals measured during said cycles made it possible to construct the database used to train and validate the models which will be described.

[0074] The implementation of the predictive maintenance process requires characterizing the state of health of a piece of equipment or a system, i.e. its degree of degradation or its remaining lifespan.

[0075] First of all, with reference to [Fig.2], the general principle of the invention is described.

[0076] The health status of a piece of equipment (or a system comprising several pieces of equipment which interact) is characterized by a certain number p of parameters calculated from signals measured on said equipment.

[0077] The database used makes it possible to produce an observation matrix 10 (step E20).

[0078] The observation matrix 10 is derived from recordings of the measured signals, broken down into n operating cycles. The p parameters Stat ;, Stat 2, ..., Stat p are calculated on each operating cycle, thus modifying the time scale.

[0079] The concatenation of these new signals makes it possible to constitute the observation matrix 10: the observation matrix 10 is created by considering these p parameters calculated over the n successive operating cycles. The observation matrix 10 has the size n* P \ n rows and p columns.

[0080] We define a shape vector 11 as being a row of the observation matrix, that is to say the instantiation of all the parameters at the cycle ie K 1, nj . Each shape vector 11 is therefore of size 1 x p. We see in Figure 2 the shape vector 1 It instantiated at time f. The shape vector is a signature of the process.

[0081] To quantify the state of health of the system, we introduce the notion of the degree of severity of a fault, and we partition the successive values ​​of the shape vector according to this degree of severity. We use the hypothesis that a fault in the system gradually worsens, and that the system does not improve during its lifetime (no corrective maintenance).

[0082] This hypothesis makes it possible, for example, to predict the duration between two maintenance periods by considering the start of the period as being the initial state, or to estimate the lifespan of a system without corrective maintenance.

[0083] One of the major innovations brought by the invention is a new method of partitioning (or clustering) time series. This partitioning uses an unsupervised learning algorithm which groups points according to a measure of similarity or more precisely, a distance.

[0084] To do this, each parameter of the shape vector (namely time series) is encoded in the form of an image (step E21), and a learning algorithm is used to implement shape recognition to recognize patterns in each image as well as boundaries delimiting the patterns (step E22). A primary boundary signal representative of the boundaries is then produced, and the primary boundary signals are merged (step E23) to produce a global boundary signal from which the severity classes are defined (step E24).

[0085] The fusion of these primary boundary signals allows the creation of a meta-signature representing the likelihood of the boundaries of the severity classes.

[0086] All the operations described here are implemented in an electronic processing unit 14, visible in [Fig. 3], which comprises at least one processing component 15 adapted to execute instructions of a program to implement the partitioning method according to the invention. The program is stored in at least one memory 16 of the processing unit 14.

[0087] The processing component(s) 15 comprise, for example, a processor, a GPU (Graphics Processing Unit), a DSP (for Digital Signal Processor), a microcontroller, or a programmable logic circuit such as an FPGA (for Field Programmable Gate Arrays) or an ASIC (for Application Specified Integrated Circuit)•

[0088] The processing unit 14 acquires the measured signals which are stored in the database 17, implements processing on these measured signals, and stores the result of these processings in the memory 16. Initially, the database 17 is composed of unlabeled data from measurements carried out on test equipment (the jacks 5). These processings are therefore not implemented directly on equipment in operation (in service), but the results of the learning will make it possible to estimate the level of severity of defects and the residual service life of equipment in operation. The data obtained in operation will then be labeled and introduced into the database 17 to enrich it.

[0089] Each of these steps is now described in more detail, with reference to [Fig.4] and the following figures.

[0090] The processing unit 14 extracts the unlabeled signatures from the database 17, and separates the independent realizations (step E40) to produce the time series

[0091] We obtain a set of p time series ç= {E1? E2, -E / ,}, corresponding to the parameters of the shape vector allowing us to follow the state of the system.

[0092] Each time series contains n samples, so that we have, for je [[ 1; p] :

[0093] Xj = [ xlp x2r ' X»J ]T-

[0094] Each time series is therefore associated with one of the parameters and includes elements which are observations of said parameter. Each time series includes n observations (measurements) of said parameter.

[0095] The observation matrix therefore comprises p time series Xb X2....,XP which form the columns of said observation matrix.

[0096] The principle of data clustering is based on the minimization of an intra-class distance (objects sharing similar properties) and the maximization of an extra-class distance (objects with heterogeneous properties).

[0097] However, this local similarity measure must be extended to capture the global dynamics of the signal. A distance measure is therefore used between a point and all other points in the series (therefore for each of the n samples), n distinct distances can then be calculated. A signal sampled from n points thus generates the calculation of n2 different distances, which can be represented in a matrix form.

[0098] For each time series, the processing unit 14 calculates a distance matrix (step E41). The distance matrix is ​​a Gram matrix whose elements are distance values ​​between the elements of the time series.

[0099] For each time series, the distance matrix is ​​a square matrix of size nxn which csl q^c from a series with n samples (the measurements of the parameter) noted w ■

[0100] Each element of the distance matrix, indexed by the pair AA , is then obtained by calculating the scalar product between xi and xj. In other words, each element is the result of a scalar product between a point and its neighbors. The Euclidean norm associated with the previously defined scalar product is calculated by applying the square root to all elements. The distance matrix is ​​therefore made up of Euclidean distances.

[0101] Note that, although the Euclidean distance is used here, other mathematical distances can be used, depending on the dynamics to be highlighted. For example, the Mahalanobis distance could be used.

[0102]

[0103]

[0104]

[0105] We therefore construct, for each time series, a matrix representation of said time series. The objective is to group the measurements together according to similarities which take into account both the local and global dynamics of the time series. The distance matrix can be written as follows: x'i) xQ ... d^ xj ; ) where is a time series, the (xî ) are the elements of said time series, and where is a distance value between x} and xlm, with 1 -

[0106] Some fundamental properties can be deduced from the structure of the distance matrix.

[0107] First, the distance matrix is ​​symmetrical and consists of positive elements.

[0108] Indeed, the distance matrix F is made up of scalar products. The scalar product being commutative, we have: [°109] l (X^j)

[0110] The distance matrix F is therefore symmetrical. Moreover, by definition of the Euclidean distance, all the parameters of F are positive.

[0111] Furthermore, the distance matrix being symmetrical, the diagonal is an axis of symmetry of the matrix and it is made up of null elements.

[0112] Here, the processing unit 14 normalizes all the elements of the distance matrix in the interval [0; 1],

[0113] The processing unit 14 then generates from the distance matrix a primary image comprising pixels having levels representative of the distance values ​​of the distance matrix.

[0114] The pixel levels here are gray levels.

[0115] Since the primary images are representations of the distance matrices, they are single-channel.

[0116] The primary image 20 of [Fig.5] is obtained from averages of the vibration signal (the parameter) measured on the LLA over 1130 cycles.

[0117] The primary image 21 of [Fig.6] is obtained from minima of the vibration signal (the parameter) measured on the URA over 1130 cycles.

[0118] The symmetrical structure is clearly visible in [Fig.5], and it is visible in [Fig.6].

[0119] The signals used are noisy. This noise is the origin of the lines in the image resulting from the distance matrix.

[0120] In the image of [Fig.5], two groups of pixels are visible, forming two squares 22 separated by a border located between 700 and 800 cycles. These squares 22 represent the zones of distances close to 0. They therefore testify to the low variation of the value of the parameter over time. The contrasting borders, for their part, testify to the sudden variation of the signal. Two squares 23 can also be seen in [Fig.6].

[0121] Except for intra-group variance, the vibration signals have the dynamics of piecewise continuous constant functions. The continuity jump occurs when the equipment (the cylinder) moves from one degree of fault severity to another. Returning to a lower degree of severity is impossible if no maintenance is carried out on the cylinder.

[0122] The assumption is made that the size of the zones decreases as aging progresses since the latter accelerates. This information could be integrated into the prognostic algorithm subsequently.

[0123] It is noted that by using the representation of a matrix as an image, it is possible to extract information from groups in a time series. This matrix being symmetrical, it is advantageous to use an information compression method to minimize the computational load of such a method. For example, one could keep only the upper triangular part (and in particular by transforming a triangular matrix into a square matrix). The results of detecting the degrees of severity are not impacted.

[0124] This symmetry will be exploited below for the generation of artificial data to compensate for the lack of real data, during a design phase for example.

[0125] A learning model is then used to recognize one or more patterns in each primary image, and to produce a primary boundary signal representative of boundaries delimiting said patterns (step E42).

[0126] The learning model implements the principle of semantic segmentation. This uses a computer vision method, which consists of associating each pixel of a primary image with a class. Here, two classes are distinguished:

[0127] — a first class represents square patterns having a coincident diagonal with a portion of the diagonal of the primary image;

[0128] — a second class which represents the background of the primary image, and therefore the noise.

[0129] The definition of these two classes makes it possible to extract useful information from the image primary.

[0130] The learning model used is a deep neural network that is particularly suited to pattern recognition. More specifically, the network used here is a U-NET type convolutional neural network. It should be noted that another type of neural network could be used, and even another learning model.

[0131] In industry, and in particular in the aeronautical field, it is difficult to obtain the data necessary for the training phase of a neural network. We therefore use a neural network architecture which makes it possible to carry out image segmentation for a small population of samples with regard to the size of the databases usually used for training deep neural networks.

[0132] The neural network used is made up of discrete convolution products (therefore essentially multiplications and accumulations). The network is specialized in pattern extraction.

[0133] Let IG M„ ( P ) be a square matrix of size n. To perform a discrete convolution, a matrix of size k <n inférieure à 1 appelée noyau est introduite.

[0134] The convolution kernel is applied to the image matrix by successive and ordered translations, which makes it possible to preserve the local structures of the patterns. [Fig.7] illustrates a “simple” discrete convolution product produced by a convolution kernel 24 applied to the matrix 25 of the primary image.

[0135] It is noted that the structure of such a network adapts to any image, a fortiori, to any square-sized matrix. Thus, the method is not constrained by the size of the series considered. Here, the examples are developed with series having 1130 observations, but it is of course possible to extend the method to series having any number of observations.

[0136] We of course choose a kernel adapted to the desired result, that is to say to the detection of boundaries separating patterns.

[0137] The basic architecture of the network used is described in [Fig.8].

[0138] The network 26 consists of two parts.

[0139] A part 27 (on the left in [Fig.8]) allows the information to be compressed and a part 28 (on the right in [Fig.8]) allows the same information to be decompressed. At the output of the network, a segmented image whose pixel value represents, after normalization, the probability of belonging to one of the two classes defined upstream, is obtained. Indeed, the last layer of the network 29 is a softmax layer (normalized exponential function) which allows any number to be transformed into empirical probability.

[0140] The neural network 26 used operates in a similar way to those described in the document Ronneberger, Olaf, Philipp Fischer and Thomas Brox (May 2015). "U-Net: Convolutional Networks for Biomedical Image Segmentation ». In : arXiv:1505.04597 [es]. arXiv : 1505.04597 [es].

[0141] It is noted that the passage through images to characterize the time series, but also the fact that the operations implemented in the neural network used are essentially multiplications and accumulations, make the use of a GPU particularly relevant to implement the invention. Consequently, in an advantageous embodiment of the processing unit 14 at the hardware level, it will comprise a GPU to implement the processing of the primary images (neural network), as well as possibly one or more other processing components to implement all or some of the other steps of the partitioning method described here.

[0142] However, as we have just seen, it is possible that the quantity of data available for training is insufficient. We therefore increase the data by generating primary images artificially, which improves the training of the network 26.

[0143] The processing unit 14 therefore uses an artificial generator to generate artificial images.

[0144] The artificial generator will instantiate a mathematical model of simulation functions, which therefore correspond to the dynamics of the signals studied, to simulate the behavior of a real time series. Here, the simulation functions are piecewise continuous constant functions. The simulation functions could be functions of different types, and for example increasing functions with multiple slopes.

[0145] The piecewise continuous constant functions each comprise a series of steps of varying widths and varying amplitudes. The widths give the dimensions of the square patterns and the amplitudes give the pixel values. This process simultaneously generates the pattern for training (with shades of gray) and the binary target pattern (ground truth).

[0146] The encoding used above (creation of the distance matrix) is illustrated with artificial signals with reference to [Fig.9].

[0147] A piecewise continuous constant function 31 is generated. The function 31 comprises four steps of variable widths (x 0, xi, x 2, x 5) and variable amplitudes. The function 31 is encoded by calculating a Gram matrix of said function. An artificial distance matrix and an artificial primary image 32 are obtained from said artificial distance matrix. A mask 33 of the image is also produced.

[0148] A few thousand images are thus generated for training the neural network. The training data is generated in gray levels. The mask 33 represents the ground truth that the learning architecture should produce. At each iteration of the learning, the network's result will be compared to mask 33.

[0149] The U-NET network will then modify the value of its weights according to the error rate calculated between the “truth” and its output.

[0150] After the training phase, an architecture capable of discerning the desired patterns is created. The patterns are squares having a diagonal coincident with a portion of the primary image.

[0151] [Fig. 10] represents a segmented image 35 resulting from the network's inference on the primary image 20 of [Fig. 5], while the U-NET network was trained with purely artificial data. [Fig. 11] represents a segmented image 36 resulting from the network's inference on the primary image 21 of [Fig. 6], while the U-NET network was trained with purely artificial data. The square-shaped patterns 22 and 23 appear distinctly in Figures 10 and 11.

[0152] [Fig. 12] represents the primary image 20 of [Fig.5] at the input of the network and the segmented image 35 at the output of the network, as well as a graph 37 representing the result of the segmentation.

[0153] The noise obtained in the segmented image could be further reduced by improving the training phase and in particular by generating noisy data for learning. The process could also be completed by increasing the learning base with real data from process monitoring. It would also be possible to reduce the size of the learning base when the aggregation of new real data could reduce the relevance of certain pre-existing segments constructed artificially. A phase of pruning and strengthening the database 17 would thus be carried out as the device is operated.

[0154] It is noted that learning a deep structure with purely synthetic data allows for a wider adoption of artificial intelligence methods in areas with sparse data.

[0155] We have therefore just explained how the processing unit 14 transforms each time series into a primary image, then produces from a primary image a segmented image which makes the square patterns visible.

[0156] With reference to [Fig. 13], when these steps have been carried out, signal processing operations (step E130) are implemented on the segmented matrices 35 to recover the boundaries of the groups of pixels corresponding to the patterns (step E131 and step E43 in [Fig. 4]), and to filter the results obtained (step E132). These steps are carried out for all the segmented images each associated with one of the parameters. These steps are followed by merging operations (step E133) which consist of merging the information (step E134 and step E44 in [Fig. 4]) and estimate a density by kernel function (step E135 and step E45 in [Fig.4]) to obtain a meta-signature (step E136).

[0157] Note that steps E41 to E43 of [Fig.4] are repeated for each of the time series.

[0158] Boundary extraction consists of producing from each segmented image 35 a primary boundary signal representative of the boundaries of the groups of pixels of said segmented image, then filtering the primary boundary signal.

[0159] With reference to [Fig. 14], the processing unit 14 produces from each segmented image associated with a parameter a segmented matrix 40 which comprises the levels of the pixels of the segmented image. The construction of the primary boundary signal consists first of all in calculating a statistical function on sets of elements 41 which each comprise the elements of one of the anti-diagonals 42 of the segmented matrix 40 associated with said parameter (step E140).

[0160] The statistical function is for example the median function.

[0161] The interest of such a direction of travel is to be able to carry out a dimension reduction and to obtain a discriminant function of groups (or clusters).

[0162] The segmented matrix 40 is therefore traversed by increasing index along the opposite diagonal to obtain sets of successive elements 41 of increasing then decreasing length.

[0163] The statistical function is calculated for each set of 4L elements. It then takes a non-zero value at the location of a group of pixels and zero at its borders (or vice versa depending on the chosen convention).

[0164] We then obtain for each segmented matrix 40 a primary boundary signal 43, representative of the boundaries of the segmented matrix.

[0165] The primary boundary signal 43 thus obtained is extremely noisy and sees its abscissa scale dilated by a factor of 2. In the case studied, the number of cycles increases from 1130 to 2259 (indeed, the diagonal of the segmented matrix is ​​not counted twice, the abscissa is dilated by a factor of two minus one to be exact), a direct consequence of the traversal method used. For the 4x4 matrix represented, the number of sets of elements 41 is equal to 7 (= (2x4)-1).

[0166] The relevant information in the primary boundary signal 43 (i.e., the statistic calculated for each opposite diagonal) is contained in the peaks initiating a region of abrupt variance change. The primary boundary signal 43 should be filtered to increase the signal-to-noise ratio and produce a signal containing only said peaks. However, standard filtering methods require the non-trivial adjustment of intrinsic parameters such as the sliding window, i.e., the number of points to which the filter will be applied in a simultaneous. This window allows you to control the sensitivity of the algorithm to local or more global dynamics.

[0167] To overcome this uncertainty and make the result more systematic, an iterative method with a sliding window of increasing width (otherwise called dynamic batch peak detection) is developed.

[0168] This method uses a filter function which is a simple statistical function. Here, the filter function is a maximum function, which could be modified depending on the dynamics to be highlighted in the primary boundary signal.

[0169] With reference to [Fig. 15], this method comprises the following steps:

[0170] - a current table 45 is defined, the size of which corresponds to the number of samples contained in the primary boundary signal (11 samples in the simplified example of [Fig. 15]), i.e. in the series of samples on which the statistic is calculated (in total, 2259 samples in our case);

[0171] - the current table 45 is initialized to 0 (all its elements are null);

[0172] - a “current width” variable is initialized to 2;

[0173] - windows 46 having the current width as their width are applied to the series consisting of the 47 samples of the primary boundary signal, without overlap between the windows;

[0174] - the maximum value is calculated for each window 46. The maximum value is obtained for one of the samples of said window 46. The maximum value is added to the element of the current table 45 whose position in the current table 45 corresponds to the position of said sample in the series;

[0175] - the current width is incremented by one unit;

[0176] - the previous steps are repeated, until the current width is equal to the size of the series.

[0177] At the first iteration, each window 46 therefore covers two samples then, at each iteration, the size of the windows 46 is incremented by one unit. The maximum value obtained on each window 46 is added to the element of the current table 45 corresponding to the cycle of the window where it is obtained, while the values ​​of the other elements are retained.

[0178] Thus, following the first iteration, the table 45a is obtained. It can be seen that in each window 46 covering two samples 47, the maximum value between the two samples is preserved and is assigned to the element of the table 45a whose position in the vector corresponds to the position in the series of samples of the sample which has the maximum value of the window.

[0179] For example, for window 46 which includes samples x0 and xb the maximum corresponds to the value y2 associated with xp Sample xi is in second position in the series, so that the second box of table 45a takes the value y2 (= 0 + y2).

[0180] The zero value is added to the other element of the array 45a (the first) which remains of zero value.

[0181] Following the second iteration, table 45b is obtained.

[0182] Following the third iteration, table 45c is obtained.

[0183] This sequence is repeated for window sizes incremented by one unit each time until the window covers the entire series, i.e., the primary boundary signal.

[0184] This filtering algorithm amplifies the local dynamics at the beginning of the loop and then tends to amplify the more global dynamics.

[0185] It is noted that, when filtering the signal by sliding window with incremental size, the maximum was chosen in the method developed here. It is possible to select another statistic, for example the median, to highlight other dynamics of the signal.

[0186] By dividing the abscissa scale by two to compensate for the dilation generated by the method of [Fig.14], the filtered primary boundary signal 47 visible in [Fig.16] is obtained. The peaks stand out clearly.

[0187] The primary boundary signal obtained in [Fig. 16] has zero values ​​at locations where the algorithm's statistics have never been verified. The indices of the non-zero values ​​are then recovered.

[0188] As it stands, the latter have values ​​between [ 0; 2*h - 1 ] or [ 0; 1]. To ensure that the abscissa scale of the signal is representative of reality, the integer part of the obtained values ​​divided by two is calculated. The assumption is made that a precision of two cycles is tolerated for the calculations. Indeed, when the cycle number is not a multiple of 2, this trivial reduction will have the effect of translating this point between two integer cycles.

[0189] Subsequently, an array of size less than n, i.e. 1130, containing these cycle values, is created.

[0190] An empty array of size exactly n is then initialized with zero values ​​and unit values ​​are added to the box designated by each of the previous cycle numbers. The final array is therefore made up of binary elements.

[0191] Note that the cycles visible in the primary boundary signal represent probable group boundaries.

[0192] The primary boundary signal fusion algorithm involves summing the primary boundary signals to obtain a summed boundary signal, and then estimating an empirical density of the summed boundary signal to produce the overall boundary signal.

[0193] The empirical density function can be oversampled. Also, theoretically, a continuous function will be created to estimate the shape of the likelihood group boundaries. Therefore, the loss of precision caused by the previous method has anecdotal effects on the general method.

[0194] We see in [Fig. 16] that clear boundaries appear around 800 / 2 = 400, 2000 / 2 = 1000 and 1100 cycles.

[0195] At this stage, the method can have two possible orientations depending on the nature of the descriptor.

[0196] For a given descriptor:

[0197] — either the amplitude is normalized between 0 and 1, which gives a weight to the peak in its neighborhood;

[0198] — either the value 1 is assigned to each peak, which gives the same weight to all the peaks.

[0199] The steps just described are performed for each parameter independently, and therefore p times in our case. The primary boundary signals, representative of the boundaries defining the square patterns, were therefore produced and then filtered for all parameters.

[0200] A new matrix gathering the concatenated boundary signals for each descriptor is obtained: ffff ] J 11 J 12 ' 13 7 lj» = AAAA, eM^P) ffff L^nl J ni J ni JnPi

[0201] The processing unit 14 will then produce a global boundary signal from the primary boundary signals.

[0202] The processing unit 14 uses an aggregation function to create new information by combining the previously calculated data. The processing unit 14 uses a kernel density estimation method (Kernel Density Estimation or KDE) to merge this information.

[0203] Hypotheses are then formulated regarding the nature of the manipulated data (which are the time series previously calculated for each descriptor). The signals obtained are considered as realizations of sampled random variables. As a result, their distribution provides information on their theoretical probability density. The kernel function makes it possible to construct an empirical density that converges towards its theoretical density (subject to certain hypotheses on the kernel functions).

[0204] As we have seen, the data in the database used in the example described here come from measurements of p parameters (statistical descriptors) carried out on four pieces of equipment: the four cylinders URA, ULA, LLA, LRA.

[0205]

[0206]

[0207]

[0208]

[0209]

[0210]

[0211]

[0212]

[0213]

[0214]

[0215]

[0216]

[0217]

[0218]

[0219]

[0220]

[0221] The implementation of the method according to the invention then differs depending on whether the measurements of the p parameters were obtained on the same equipment (or the same system) or on different equipment (or systems). In the case where the measurements of the p parameters were obtained on the same equipment, that is to say in this case on the same cylinder, the p descriptors are considered as p realizations of the same random variable, themselves independent and identically distributed (iid). In the case where the measurements of the p parameters were obtained on different devices, each descriptor is considered to be a different random variable. The realizations are still independent but no longer identically distributed. We therefore obtain two different methods of merging information depending on the reality to be modeled. We first consider the case where the p descriptors are considered as p realizations of the same random variable. The previously calculated values ​​related to the peaks relative to each of the p descriptors can be summed: the processing unit 14 adds the filtered primary boundary signals, to obtain a summed boundary signal. Thus, considering the previously defined matrix, a new vector is therefore calculated such that: ' 1 7 lâ<« tj The primary boundary signals are therefore summed to obtain a summed boundary signal forming a new time series E = ( Xj ) . The signal of The summed boundary 48 obtained for the LLA is visible in [Fig. 17]. The empirical density of this summed boundary signal 48 is then calculated using the following formula: d(x) = r where xi are the summed boundary signal elements, K is the kernel function, represents a bandwidth of the kernel function and n is a number of the elements of each time series. K represents the kernel function applied to the variable centered at the point of the sampled and scaled series. The kernel function can be of different types. Here, we use the Gaussian kernel function which is of the form: Æ(x) - ^==exp4*2. Thus, xi with ie [[ 1; n J represents the sequence of values ​​taken by the random variable andx the current point of the estimated density. The signal density can be approximate. The sum of the kernels is allowed by the assumption that the samples are identically distributed. The normalization factor A is then added to the expression. The element represents the bandwidth of the kernel function. It is proportional to the support of the latter. Note that in the case of a Gaussian kernel function, the support is infinite but h defines the spread of the Gaussian.

[0222] [Fig. 18] shows an example with a simplified discrete signal comprising six peaks 49. The above equation results in a sum of densities 50 located at the points of the simplified discrete signal and generates a resulting signal 51.

[0223] By applying this principle to the summed boundary signal 48 visible in [Fig. 17], we obtain the empirical density 52 visible in [Fig. 19], which constitutes the overall boundary signal.

[0224] Note that the method requires the choice of two parameters: the kernel function and the bandwidth value.

[0225] For the kernel function, the Gaussian is selected for its theoretical properties.

[0226] Other kernel functions could be used, such as binomial functions like the &Epanechnikov kernel. These would minimize the computational load of the method.

[0227] The bandwidth is defined by the rule stated in the document Silverman, Bernard W. (Apr. 1986). Density Estimation for Statistics and Data Analysis, en. CRC Press, isbn: 978-0-412-24620-3. for a Gaussian kernel namely h = 0.9Anwith A = min(a, inter-quartile range / 1.34) • With the inter-quartile range the difference between the third and the first quartile in a series, in other words a measure of data dispersion.

[0228] The boundaries between the severity classes are defined by the maxima of the empirical density, i.e. of the global boundary signal 52. We therefore distinguish three classes in [Fig. 19] (on either side of the maxima).

[0229] Using the properties of densities, it is possible to calculate the probability of occurrence of a group change and therefore of a boundary. To do this, simply integrate the signal up to the desired boundary.

[0230] The empirical density, that is to say the global boundary signal, then represents a meta-signature which could subsequently characterize a piece of equipment or a system studied in the manner of a descriptor.

[0231] We can therefore evaluate the state of an equipment or an on-board system in operation using the global boundary signal, then classify the global boundary signal

[0232]

[0233]

[0234]

[0235]

[0236]

[0237]

[0238]

[0239]

[0240]

[0241] according to the classes defined in [Fig. 19], and thus assess the health status of the equipment or system and its remaining lifespan. In the previous step, it was considered that the URA, ULA, LRA and LLA cylinders represented the realizations of the same random variable. Now, let us consider that the 4 cylinders each represent a random variable; the boundary signals of each can no longer be summed in this way because there is no guarantee that the realizations of the random variables considered are identically distributed. By extension, we can consider that the 4 cylinders represent 4 different systems to be monitored, therefore 4 different realizations. The following formulas will be generalized, to take into account the future plurality of equipment, for m realizations, therefore m dimensions for the estimation of density by kernels, then called multivariate. Before presenting the general case, the density estimation by multivariate kernels in two dimensions will be presented for reasons of clarity. Let: P2 P be the density function to be estimated in the case of two boundary signals and (X| PJ Yn) a sample of the density d. Using modeling by produced kernels, the following estimate is obtained: Thus, the variables x and belong to the support of the estimated density and, in the case of dimension 2, it can be represented by a matrix. It is then possible to generalize this formulation in dimension m in the case that interests us by taking as a basis the following formula applied to the elements of the matrix 9?: / V ] w, fl [ *—rfH l : I* \ ] , d(x\^xm) = [FI^Ay — ) | where the are elements of the primary boundary signals, £ is a kernel function, the h\JljJln represent a bandwidth of the kernel function, m is the number of non-identically distributed realizations and n is a number of the elements of each time series. This empirical density forms the global boundary signal in the case of interest to us. From this equation, it becomes possible to enrich the density calculation carried out earlier by assuming that the realizations of random variables considered come from different distributions. For the ETRAS® system, it would then be interesting to make modeling assumptions at several levels, depending on whether the system is considered as a complete ETRAS® system or a set of cylinders. In our example, if each of the four cylinders is considered as a system, then it is not trivial to consider that the values ​​of the vibration signals of each of the cylinders follow the same distribution law. Indeed, the cylinders are not worn uniformly across the system. To improve the reliability of the information obtained with the meta-signature of [Fig. 19], it is then appropriate to merge the information from the four different cylinders. A weighting can then be applied to the descriptors of each of the electromechanical actuators while the estimated multivariate density will be considered as a mixture of Gaussians (in the case of a Normal kernel function).

[0242] In the previous equation, the information provided by each descriptor is weighted identically. It could be interesting to assign a specific weight to certain descriptors. Each kernel function would be weighted by (pp ÎE [ 1; m J such that X"m (p — ] to maintain a probability mass function. The multivariate kernel density estimation could then be written as

[0243] The presented information fusion makes it possible to create a meta-signature representing the group membership function of the data considered. This method makes it possible to take into account the contribution of data from different systems (for several ETRAS®) or even the contribution of data for several actuators (if we consider the level of the cylinders).

[0244] As we have seen, and returning to [Fig.4], the determination of the maxima (step E46 in [Fig.4]) makes it possible to obtain the number of severity classes. The processing unit 14 stores the results in the memory 16 (step E47).

[0245] Thus, it is understood that the data fusion step 45 makes it possible to produce a meta-signature for each piece of equipment (here for each URA, ULA, LRA and LLA cylinder), but also a global meta-signature which takes into account the observations produced for each of these pieces of equipment.

[0246] We therefore improve the estimation of the state of health and the residual lifespan of a subject equipment (or a subject system) by using observations obtained on different equipment (or systems) which are close but not necessarily identical to said subject equipment (and which may differ from each other).

[0247] It is of course advantageous to format the labels for each point of each time series (step E48), so as to produce labeled data. The labeled data are then reintroduced into the database 17 (step E50).

[0248] Of course, the invention is not limited to the embodiment described but encompasses any variant falling within the scope of the invention as defined by the claims.< / n>

Claims

Claims

1.

2. Partitioning method, producing results making it possible to evaluate the state of health and the residual lifespan of equipment in operation, the partitioning method being implemented in a processing unit (14) and comprising the steps of: acquiring an observation matrix (10) produced from a database (17) comprising observations of parameters carried out on at least one piece of equipment (5), the observation matrix comprising time series (Xb V2,..Xp) each comprising elements which are observations of a parameter; for each time series: • calculate a distance matrix comprising distance values ​​between the elements of the series te physical, then generating from said distance matrix a primary image (20, 21) comprising pixels having levels representative of the distance values ​​of the distance matrix; implementing a learning algorithm (26) to segment the primary image so as to obtain a segmented image (35, 36) comprising patterns (22, 23) delimited by boundaries, and producing a segmented matrix (40) comprising pixel levels of the segmented image; define from the segmented matrix a primary boundary signal (47) representative of the boundaries s ; merge the primary boundary signals to obtain a global boundary signal (52), and define classes from the global boundary signal (52). The partitioning method of claim 1, wherein the distance matrix is ​​a Gram matrix such that: ' d^, X0 44' x*2) ■. • d(x\, xQ ' d^x 1 ^ x']) 44, X0 ... d{xi2, 4) ■444) d^x^ ... 44,4)- where E(- is a time series, the La) are the elements of said k'] <k<n série temporelle, et où       xi^ est une valeur de distance entre x\ xlm, avec 1 l n m < n.

3. A partitioning method according to claim 2, wherein the patterns are squares (22, 23) having a diagonal coincident with a portion of a diagonal of the primary image.

4. Partitioning method according to one of the preceding claims, in which the segmentation of the primary image (20, 21) consists of determining a probability of belonging of the pixels of the primary image to a first class representing the patterns or to a second class representing a background of the primary image.

5. Partitioning method according to one of the preceding claims, in which the learning algorithm (26) is a convolutional neural network of the U-NET type.

6. A partitioning method according to one of the preceding claims, further comprising the step of generating artificial primary images (32) from simulation functions (31), and training the learning algorithm using the artificial primary images.

7. A partitioning method according to claim 6, wherein the simulation functions are piecewise continuous constant functions (31) or multi-slope increasing functions.

8. A partitioning method according to one of the preceding claims, wherein defining the primary boundary signal comprises the step of calculating a statistical function on sets of elements (41) which each comprise elements of one of the anti-diagonals (42) of the segmented matrix (40).

9. The partitioning method of claim 8, further comprising the step of filtering the primary boundary signal using an iterative sliding window method.

10. A partitioning method according to one of the preceding claims, wherein merging the primary boundary signals (47) comprises summing the primary boundary signals to obtain a summed boundary signal (48), and then estimating an empirical density function of the summed boundary signal to produce the overall boundary signal (52).

11. Partitioning method according to claim 10, wherein, to calculate the empirical density function, a first formula is used in the case where the parameter observations were carried out on the same equipment or system, and a second formula is used in the case where the parameter observations were carried out on separate equipment or systems representing non-identically distributed realizations.

12. The partitioning method of claim 11, wherein the first formula is: 2(x) where the xi are elements of the summed boundary signal, K is a kernel function, h represents a bandwidth of the kernel function and n is a number of the elements of each time series.

13. A partitioning method according to claim 11, wherein the second formula is: d(X1, ^xm) = E.=i{A\~h~ ) ) where the are elements of the primary boundary signals, K is a kernel function, the represent a bandwidth of the kernel function, m is a number of non-identically distributed realizations and n is a number of the elements of each time series.

14. A partitioning method according to claim 11, wherein the second formula is: / X 1 fi—\ 1 d ( X1, xm ) = ( — ) j, where the are elements of the primary boundary signals, is a kernel function, the hydlj.hn represent a bandwidth of the kernel function, m is a number of non-identically distributed realizations, n is a number of the elements of each time series, and wherein we have: [[ 1; m ] such that L = 1 '

15. Processing unit (14) comprising at least one processing component (15) in which the partitioning method according to one of the preceding claims is implemented.

16. Processing unit according to claim 15, comprising a GPU in which at least the learning algorithm is implemented for segmenting the primary images.

17. Computer program comprising instructions which cause the processing unit according to one of claims 15 or 16 to execute the steps of the partitioning method according to one of claims 1 to 14.

18. A computer-readable recording medium on which the computer program according to claim 17 is recorded.