Method for detecting anomalies in acquired time series, and associated devices
The method uses statistical techniques and non-parametric methods to efficiently detect anomalies in time series with reduced resource consumption and minimal training data, addressing the limitations of existing methods by enhancing accuracy and reducing false positives.
Patent Information
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- THALES SA
- Filing Date
- 2025-12-17
- Publication Date
- 2026-06-25
AI Technical Summary
Existing anomaly detection methods in time series, such as those based on expert knowledge, classifiers, and neural networks, are limited by the need for extensive training data and suffer from high false positives or missed anomalies due to small or large latent space dimensions.
A method involving statistical techniques like elastic alignment algorithms and k-nearest neighbors to determine representation parameters and abnormality thresholds, reducing the reliance on external resources and requiring minimal training data, while using a non-parametric method to classify anomalies in time series.
This approach efficiently detects anomalies with reduced computational resources and minimal training data, minimizing false positives and ensuring rapid, accurate classification of normal and abnormal behaviors.
Smart Images
Figure EP2025087728_25062026_PF_FP_ABST
Abstract
Description
[0001] Method for detecting anomalies in acquired time series and associated devices
[0002] The present invention relates to a method for detecting anomalies in acquired time series. The present invention also relates to a computer program product and an associated readable information storage medium.
[0003] In the field of signal processing, and more specifically time-domain signal processing, signals from measurements contain errors originating, for example, from the measurement, transmission, reconstruction, or failure of the transmission system from which the signals originate.
[0004] Therefore, it is necessary to be able to detect abnormal behaviors of time signals in order to post-process them and anticipate a possible failure of the transmission system.
[0005] To meet this need, it is known to use anomaly detection devices based on methods using latent spaces, classifiers, expert knowledge or, more recently, neural networks.
[0006] However, in practice, it turns out that all these techniques have constraints that are still too limiting.
[0007] Expert knowledge-based techniques are limited by the fact that many anomalies are not characterized by an explicit rule, while clustering approaches are themselves dependent on this same expert knowledge.
[0008] Processes based on classifiers and neural networks are limited by the substantial need for examples they require for their training.
[0009] Finally, latent spaces still induce too many false positives for latent spaces of too small dimensions and, conversely, anomalies too often ignored for latent spaces of too large dimension.
[0010] There is therefore a need for a more efficient anomaly detection process in time signals with an implementation that consumes fewer external resources, particularly in terms of training data.
[0011] To this end, the present description relates to a method for detecting anomalies in time signals acquired by an anomaly detection device and comprising the following phases:
[0012] - training the anomaly detection device on a plurality of training time series comprising normal training time series, normal calibration time series and abnormal time series, the training phase comprising the steps of:
[0013] - application of a statistical technique on each of the normal training time series to obtain a plurality of training representation parameters, a representation parameter being a vector representation of a time series, in a predetermined basis, said plurality of representation parameters comprising at least one representation parameter representing an amplitude and one representation parameter representing a phase,
[0014] - application of the statistical technique to each of the normal and abnormal calibration time series to obtain a plurality of calibration representation parameters,
[0015] - calculation, for each calibration representation parameter, of a list of training distances between said calibration representation parameter and each of the training representation parameters, and
[0016] - determination of an abnormality threshold based on consistency rates associated with each of the normal calibration and abnormal time series, the consistency rate associated with a time series being determined from the lists of training distances associated with said time series,
[0017] - Acquisition of measurement points of a physical quantity over time by at least one sensor,
[0018] - obtaining a time series acquired from a plurality of acquired measurement points,
[0019] - Anomaly detection in the obtained time series, including the following steps:
[0020] - application of the statistical technique to the time series obtained to obtain a plurality of representation parameters,
[0021] - calculation, for each representation parameter, of a list of distances between said representation parameter and each of the training representation parameters,
[0022] - classification of the resulting time series as abnormal if the consistency rate between the lists of distances associated with each of the representation parameters is below the abnormality threshold. According to other advantageous aspects of the invention, the detection method comprises one or more of the following features, taken individually or in any possible combination:
[0023] - during the application steps, the statistical technique is also applied to a weighted average of normal training series.
[0024] - During the training phase, the abnormality threshold is determined by an operator.
[0025] - the statistical technique is an elastic alignment algorithm.
[0026] - the statistical technique is a square root velocity function.
[0027] - the consistency rate associated with a time series is determined by applying a non-parametric method to the lists of training distances associated with said time series, the non-parametric method advantageously being a k nearest neighbors technique.
[0028] - the detection phase further includes a step of determining a calculation index lower than the number of measurement points of the plurality of measurement points, the detection phase then being applied to a time series obtained limited to a restricted number of measurement points, the restricted number being equal to the calculation index.
[0029] - The classification step includes the implementation of the following inequality:
[0030] {{ / i'” . ''} n 0 '' . r 0 "} < fc} where:
[0031] - denotes the training function corresponding to the i-th distance associated with the amplitude representation of the time series f,
[0032] - denotes the training function corresponding to the i-th distance associated with the phase representation of the time series f,
[0033] [f^ a ...,f NN (a:> denotes the degree of consistency between the amplitude representation and the phase representation (classified as such < fi+i and k denotes the abnormality threshold. A time series is classified as abnormal if the inequality is not validated by said time series.
[0034] - The method further includes a preprocessing phase following the time series acquisition phase, in which the obtained time series is preprocessed, for example, during a data smoothing, filtering, segmentation, or derivation step. - During the time series acquisition phase, a configurable time window is applied to the acquired measurement points, with measurement points outside this window being excluded to form the time series.
[0035] - the process includes an alert phase in which an external electronic medium alerts the operator if a time series from the anomaly detection device is classified as abnormal.
[0036] The present description relates to a computer program product comprising a readable information carrier, on which is stored a computer program comprising program instructions, the computer program being loadable onto a data processing unit and adapted to cause the implementation of a process as defined above when the computer program is implemented on the data processing unit.
[0037] This description also relates to a readable information carrier containing program instructions forming a computer program, the computer program being loadable onto a data processing unit and adapted to drive the implementation of a process as defined above when the computer program is implemented on the data processing unit.
[0038] In the following description, a quantity is substantially equal to a value when the quantity is greater than or equal to 90% of the value and the quantity is less than or equal to 110% of the value.
[0039] The invention will become more apparent upon reading the following description, given solely by way of non-limiting example and with reference to the drawings in which: Figure 1 is a schematic representation of an anomaly detection device monitoring the behavior of time signals from a system, Figure 2 is a flowchart illustrating an example of implementation of an anomaly detection method by the anomaly detection device, Figure 3 is a flowchart illustrating an example of implementation of a training phase of an anomaly detection method, and Figure 4 is a flowchart illustrating an example of implementation of a detection phase of an anomaly detection method.
[0040] Figure 1 schematically represents a system 2, a set of sensors 4A, 4B and 4C, a processing circuit 5 comprising a plurality of signal shaping units 6A, 6B and 6C and an analog-to-digital converter 8, an anomaly detection device 10 and an external electronic support 12. System 2 is a device characterized by a physical quantity that varies over time.
[0041] System 2 is configured to emit a time signal in the form of measurement points emitted at regular or irregular time intervals.
[0042] For example, system 2 is an aircraft, an electronic device, or any other device characterized by a physical quantity that varies over time.
[0043] Typically, the physical quantity that varies over time is a position in a predetermined frame of reference, a heading, a speed, a temperature, or any other physical quantity from system 2.
[0044] Sensors 4A, 4B and 4C are configured to capture the physical quantity characterizing system 2 in the form of an analog signal.
[0045] More specifically, each sensor 4A, 4B and 4C is configured to measure the physical quantity of system 2 or to receive a time signal from system 2 and generate a signal representative of the physical quantity.
[0046] The representative signal is usually an analog signal.
[0047] Thus, each sensor 4A, 4B and 4C corresponds, for example, to an antenna or to a position, pressure, speed or temperature sensor.
[0048] Typically, sensors 4A, 4B and 4C are probes, event counters, radars, sonars, or electromagnetic spectrum analyzers.
[0049] Each 4A, 4B and 4C sensor is typically used with a 5-channel processing circuit.
[0050] The processing circuit 5 is positioned downstream of sensors 4A, 4B and 4C.
[0051] For example, the processing circuit 5 includes a set of signal conditioners 6A, 6B and 6C and an analog-to-digital converter 8.
[0052] The signal shaping units 6A, 6B and 6C, the number of which depends on the number of sensors 4A, 4B and 4C, are connected to the latter and allow, for example, the amplification and / or shaping of the analog signals from sensors 4A, 4B and 4C.
[0053] Advantageously, a shaping unit 6A, 6B and 6C is present for each sensor 4A, 4B, 4C and each shaping unit 6A, 6B and 6C is connected to a single sensor 4A, 4B and 4C.
[0054] The analog-to-digital converter 8 is connected to the signal conditioning units 6A, 6B, and 6C and converts the analog signals from the conditioners into digital signals, each digital signal being composed of a plurality of measurement points. Such a converter is often referred to by the abbreviation "ADC" (from the English "Analog to Digital Converter").
[0055] The anomaly detection device 10 is designed to implement a detection process in acquired time series.
[0056] More specifically, the anomaly detection device 10 is configured to receive the plurality of measurement points, obtain a time series from them and deduce a normal or abnormal classification associated with said time series.
[0057] For example, a time series is a vector of predetermined size P and composed of P measurement points.
[0058] The anomaly detection device 10 is configured to obtain a time series each time P measurement points are received and to deduce a normal or abnormal classification from said time series obtained.
[0059] A normal classification for a time series is defined as a series corresponding to an expected behavior of the system 2.
[0060] An abnormal classification for a time series is defined as a series corresponding to an unexpected behavior of the system 2.
[0061] The classification of the time series is then transmitted to an external electronic medium 12.
[0062] For example, external electronic medium 12 is a computer, a smartphone or any electronic device capable of transferring information to an operator.
[0063] In one embodiment, an operator can monitor the classifications received in real time and an alert is sent to an operator in case of an abnormal classification.
[0064] For example, the alert transmitted takes the form of a notification on the external electronic medium 12 or an audible alert.
[0065] In the example in Figure 1, the anomaly detection device 10 includes a computer 15 comprising, for example, a processor 17 and a memory 19 associated with the processor 17.
[0066] Calculator 15 is configured to process digital signals from sensors 4A, 4B and 4C.
[0067] The calculator 15 is an electronic circuit designed to manipulate and / or transform data represented by electronic or physical quantities in the calculator's registers and / or memories into other similar data corresponding to physical data in the register memories or other types of display, transmission, or storage devices. As specific examples, the calculator 15 is implemented as a programmable logic component, such as an FPGA (Field Programmable Gate Array), or as an integrated circuit, such as an ASIC (Application Specific Integrated Circuit).
[0068] Alternatively, when the process is implemented as one or more software programs, that is, as a computer program, also called a computer program product, it is also capable of being stored on a computer-readable medium, not shown here. A computer-readable medium is, for example, a medium capable of storing electronic instructions and being connected to a bus of a computer system. Examples of such a readable medium include an optical disc, a magneto-optical disc, ROM, RAM, any type of non-volatile memory (e.g., FLASH or NVRAM), or a magnetic card. A computer program containing software instructions is then stored on this readable medium.
[0069] The operation of the anomaly detection device 10 is now described with reference to Figure 2, which illustrates an example of the implementation of an anomaly detection process, and to Figures 3 and 4, which schematically represent the steps of certain phases of the anomaly detection process.
[0070] The anomaly detection process includes a training phase 20 of the anomaly detection device 10, a measurement point acquisition phase 30, a time series acquisition phase 40 and an anomaly detection phase 50 in the time series obtained.
[0071] During the training phase 20, the anomaly detection device 10 is trained to be able to recognize in an acquired time series whether the time series is normal or abnormal.
[0072] Training phase 20 is implemented on a plurality of training time series.
[0073] In the example described, a training time series is a vector of size T defined by a classification predetermined by the operator among the group consisting of: training normal, calibration normal, anomalous.
[0074] A normal training time series is characterized by a consistent evolution of the associated physical quantity and is intended for training the anomaly detection device 10.
[0075] A normal calibration time series is characterized by a consistent evolution of the associated physical quantity and is intended for the calibration of the anomaly detection device 10. An abnormal time series is characterized by an associated physical quantity following a dangerous behavior or not following any usual behavior.
[0076] Training time series are, for example, derived from data from previous studies, simulations, or measurements of real-world cases.
[0077] According to the example described, as seen in Figure 3, the training phase 20 includes a first step of applying a statistical technique 22, a second step of applying the statistical technique 24, a calculation step 26 and a training step 28.
[0078] During the first step of applying a statistical technique 22, the calculator 15 applies a statistical technique to each of the normal training time series.
[0079] Such a first step in applying a statistical technique 22 allows us to obtain a plurality of training representation parameters.
[0080] A representation parameter is a vector representation of a time series in a predetermined basis.
[0081] The plurality of representation parameters includes at least one representation parameter representing an amplitude and one representation parameter representing a phase.
[0082] In this context, amplitude refers to the magnitude or height of the values taken by a time series at each given point.
[0083] Furthermore, phase corresponds to the horizontal position of the characteristics of the curve or signal, that is, their position in time or space. Thus, a phase difference between two time series means that the characteristic points (such as peaks or troughs) occur at different times.
[0084] A training representation parameter is a representation parameter associated with a normal training time series.
[0085] For example, the statistical technique applied is an elastic alignment algorithm.
[0086] The elastic alignment algorithm uses the Fisher-Rao metric to calculate the distance between two curves while taking into account their phase and amplitude transformations.
[0087] Typically, such transformations are induced by acquisition errors between the different sensors 4A, 4B, 4C.
[0088] For example, the elastic alignment algorithm is a square root velocity function. For example, to obtain the representative phase parameter, calculator 15 applies the following equation: sign
[0089] Or :
[0090] • f: time series,
[0091] • g. : Karcher's mean of all normal training time series,
[0092] • y: any function belonging to the set of increasing bijective functions from [0,1] to [0,1], and
[0093] • Yf :representative parameter of the phase associated with the time series which also belongs to the set of increasing bijective functions from [0,1] to [0,1],
[0094] For example, to obtain the representation parameter q f As a representative of the amplitude, calculator 15 applies the following equation:
[0095] The examples above are given for illustrative purposes only; the representation parameters can be representative of any quantity in any basis and calculated in different ways.
[0096] In the second step of applying statistical technique 24, the calculator 15 applies the statistical technique to each of the normal, calibration, and abnormal time series. This second step of applying statistical technique 24 allows for obtaining a plurality of calibration representation parameters.
[0097] A calibration representation parameter is a representation parameter associated with a normal calibration time series or an abnormal time series.
[0098] During calculation step 26, calculator 15 calculates a list of training distances for each calibration representation parameter.
[0099] A distance list is defined as a list of distances between a representation parameter and a plurality of other representation parameters.
[0100] The training distance list is a list of distances between a calibration representation parameter and each of the training representation parameters.
[0101] Thus, the list of training distances associated with a calibration representation parameter takes the form of a vector whose size is less than or equal to the number of training representation parameters. For example, to calculate the list of training distances associated with a representation parameter representing amplitude, calculator 15 applies the following equation: Or :
[0102] • d a (f,g) is the distance between the representation parameter representing the amplitude of the calibration time series f and the representation parameter representing the amplitude of the training time series g,
[0103] • is the representative parameter of the amplitude of the calibration time series f, and
[0104] • q g is the representative representation parameter of the amplitude of the training time series g.
[0105] For example, to calculate the list of training distances associated with a representative phase parameter, calculator 15 applies the following equation: Or :
[0106] • d p (f, g): distance between the parameter representing the phase of the calibration time series f and the parameter representing the phase of the training time series g,
[0107] • Yf : representative representation parameter of the phase of the calibration time series f, and
[0108] • Yg : representative representation parameter of the amplitude of the training time series g.
[0109] During step 28 of the training stage, the calculator 15 determines an abnormality threshold.
[0110] Typically, the abnormality threshold is determined based on the consistency rate associated with each of the normal calibration and abnormal time series.
[0111] The consistency rate is determined by applying a non-parametric method to the lists of training distances associated with said time series.
[0112] A non-parametric method is a statistical technique that does not rely on specific assumptions about the underlying distribution of the data.
[0113] For example, the non-parametric method is a k-nearest neighbors technique. The k-nearest neighbors (k-NN) algorithm is a non-parametric classification and regression algorithm that predicts the class or value of a data point (called an observation or test point) based on the k nearest points in the training set.
[0114] For example, the anomaly threshold is determined by the operator. In such an example, the operator uses a two-dimensional representation of the points obtained to determine said anomaly threshold, called k in the rest of the description.
[0115] In other words, during the detection phase 50, a time series will be considered anomalous if the nearest neighbors in amplitude are k times (or more) different from the nearest neighbors in phase.
[0116] Typically, the abnormality threshold k is less than or equal to 10.
[0117] During the acquisition phase 30 of measurement points of a physical quantity, the anomaly detection device 10 obtains measurement points from system 2 acquired by sensors 4A, 4B and 4C.
[0118] For example, the anomaly detection device 10 stores the measurement points in memory 19.
[0119] During the time series acquisition phase 40, the anomaly detection device 10 forms an acquired time series (denoted X hereafter) from a plurality of previously stored measurement points.
[0120] For example, the time series X thus obtained is a vector of size E, E being the number of measurement points forming said time series obtained.
[0121] During the detection phase 50, the anomaly detection device 10 detects anomalies in the obtained time series X.
[0122] According to the example described, as seen in Figure 4, the detection phase 50 includes an application step 52, a calculation step 54 and a classification step 56.
[0123] During application step 52, the calculator 15 applies the statistical technique to the time series obtained to obtain a plurality of representation parameters.
[0124] During calculation step 54, calculator 15 calculates a list of training distances for each representation parameter. The distance list associated with a representation parameter is a list of distances between that representation parameter and each of the training representation parameters.
[0125] Typically, calculator 15 applies the same equation as in calculation step 26 (applied to the representation parameters of the obtained time series).
[0126] During classification step 56, the calculator 15 determines whether the resulting time series X is normal or abnormal. More specifically, the calculator 15 classifies the resulting time series X as abnormal if the consistency rate between the lists of distances associated with each of the representation parameters is less than the abnormality threshold.
[0127] For example, calculator 15 implements the following inequality:
[0128] {{ / i'” . ''} n 0 '' . r 0} < fc}
[0129] Or :
[0130] • ■ “-’ denotes the training function corresponding to the i-th distance associated with the amplitude representation of the time series f ,
[0131] • denotes the training function corresponding to the i-th distance associated with the phase representation of the time series f,
[0132] • denotes the degree of consistency between the amplitude representation and the phase representation (classified as such
[0133] • k denotes the abnormality threshold, a time series being classified as abnormal if the inequality is not validated by said time series.
[0134] To illustrate this process more concretely, its application to the detection of anomalies in a drone is now presented.
[0135] In this case, system 2 is a drone.
[0136] For example, drone 2 is configured to transmit its altitude at regular time intervals.
[0137] Thus, drone 2 includes a radio communication module equipped with an antenna.
[0138] Typically, a drone emits a signal with a frequency between 900 MHz and 6 GHz.
[0139] Thus, the information is received by at least one of the sensors 4A, 4B, 40 which is a receiving antenna.
[0140] A processing circuit downstream of at least one antenna converts the received analog signal into a digital signal.
[0141] The anomaly detection device 10 is configured to form a set of time series representing the evolution of the altitude of drone 2 during a given time interval.
[0142] A normal classification implies that the drone 2 behaves as expected, following, for example, a predefined flight plan or flight commands transmitted by the operator. Conversely, an abnormal classification implies that the flight plan was not followed or that a flight command was not obeyed.
[0143] For example, in the event of an engine failure, the trajectory is abnormal because the propeller's thrust is lower than its nominal value. In such a case, the flight instructions are not followed, and the system exhibits an anomaly compared to the drone's autopilot's commands.
[0144] The causes of this abnormal classification could be an engine or electronic failure which may involve a destination error, a collision or damage to drone 2.
[0145] The operator could then monitor the evolution of the classification of signals transmitted by drone 2 to ensure flight safety.
[0146] Thus, such a process makes it possible to estimate an error rate of a time series without any expert knowledge being required.
[0147] Furthermore, the amount of data required for training is extremely low depending on the use case.
[0148] Advantageously, inference using such a process is very fast and inexpensive in terms of computing resources.
[0149] Finally, the presence of relatively rare normal behaviors does not necessarily generate many false positives.
[0150] Other ways of implementing the process just described are conceivable.
[0151] In one embodiment, during the application steps, the statistical technique is also applied to a weighted average of normal training series.
[0152] For example, calculator 15 applies the following inequality:
[0153] • y: any function belonging to the set of increasing bijective functions from [0,1] to [0,1],
[0154] • y f : a parameter representing the phase associated with the time series, and
[0155] • Z: Karcher mean of the training time series.
[0156] For example, to determine the average Karcher pressure, calculator 15 applies the following equation:
[0157] WHERE: • f: training time series,
[0158] • Z: average pressure washer pressure, and
[0159] • : geodesic distances between . and each training time series .
[0160] The use of Karcher's average improves the time complexity when calculating amplitude and phase distances via joint alignment with respect to Z.
[0161] In another embodiment, the detection phase 50 further includes a step of determining a calculation index lower than the number of measurement points of the plurality of measurement points, the detection phase then being applied to a time series obtained limited to a restricted number of measurement points, the restricted number being equal to the calculation index.
[0162] In one embodiment, the detection process 50 includes a preprocessing phase implemented after the time series acquisition phase 40 and in which a predetermined preprocessing is applied to said time series.
[0163] For example, preprocessing may include a data smoothing, filtering, segmentation or derivation step.
[0164] Subsequently, the acquisition and detection phases are applied to the time series thus processed.
[0165] In another embodiment, during the acquisition phase 40, a configurable time window is applied to the acquired measurement points by discarding the measurement points outside of said window to form the time series.
[0166] For example, applying such a window makes it possible to exclude points from acquisitions that are distant in time for a rapid reaction of the device in case of error.
[0167] In yet another embodiment, the external electronic support 12 is a human-machine interface from which the operator can perform one or more of the following actions: moderate time series classified as abnormal (for example, moderating a time series could mean ignoring the series), alert an external agent, or send a command to system 2; access the memory where previously analyzed and classified time series are stored; initiate a test phase to evaluate the quality of the anomaly detection device 10; and initiate an alert phase to notify the operator if a time series from the anomaly detection device is classified as abnormal. Advantageously, such embodiments allow for numerous preprocessing options and enable the anomaly detection device to be adapted to specific problems.
Claims
DEMANDS 1. A method for detecting anomalies in acquired time series, said method being implemented by an anomaly detection device (10) and comprising the following phases: - training (20) of the anomaly detection device on a plurality of training time series comprising normal training time series, normal calibration time series and anomalous time series, the training phase comprising the steps of: - application of a statistical technique (22) on each of the normal training time series to obtain a plurality of training representation parameters, a representation parameter being a vector representation of a time series, in a predetermined basis, said plurality of representation parameters comprising at least one representation parameter representing an amplitude and one representation parameter representing a phase, - application of the statistical technique (24) on each of the normal and abnormal calibration time series to obtain a plurality of calibration representation parameters, and - calculation (26), for each calibration representation parameter, of a list of training distances between said calibration representation parameter and each of the training representation parameters, - determination of an abnormality threshold (28) based on consistency rates associated with each of the normal calibration and abnormal time series, the consistency rate associated with a time series being determined from the lists of training distances associated with said time series, - acquisition (30) of measurement points of a physical quantity over time by at least one sensor (4A, 4B, 4C), - obtaining (40) a time series acquired from a plurality of acquired measurement points, and - anomaly detection (50) in the obtained time series comprising the following steps: - application of the statistical technique (52) on the time series obtained to obtain a plurality of representation parameters, - calculation (54), for each representation parameter, of a list of distances between said representation parameter and each of the training representation parameters, and - classification of the time series obtained (56) as abnormal if the consistency rate between the lists of distances associated with each of the representation parameters is less than the threshold of abnormality.
2. Method according to claim 1, wherein, during application steps (22, 24, 26), the statistical technique is also applied to a weighted average of normal training series.
3. Method according to claim 1 or 2, wherein, during the training phase (20), the abnormality threshold is determined by an operator.
4. A method according to any one of claims 1 to 3, wherein the statistical technique is an elastic alignment algorithm.
5. A method according to claim 4, wherein the statistical technique is a square root velocity function 6. A method according to any one of claims 1 to 5, wherein the consistency rate associated with a time series is determined by applying a non-parametric method to the training distance lists associated with said time series, the non-parametric method advantageously being a k nearest neighbors technique.
7. A method according to any one of claims 1 to 6, wherein the detection phase (50) further comprises a step of determining a calculation index lower than the number of measurement points of the plurality of measurement points, the detection phase then being applied to a time series obtained limited to a restricted number of measurement points, the restricted number being equal to the calculation index.
8. A method according to the preceding claim, wherein the classification step (56) comprises implementing the following inequality: 18 where: - denotes the training function corresponding to the i-th distance associated with the amplitude representation of the time series f, - denotes the training function corresponding to the i-th distance associated with the phase representation of the time series f, denotes the degree of consistency between the amplitude representation and the phase representation (classified as such and k denotes the abnormality threshold, a time series being classified as abnormal if the inequality is not validated by said time series.
9. A method according to any one of claims 1 to 8, wherein the method further comprises a preprocessing phase after the time series acquisition phase, the time series obtained being preprocessed, for example during a data smoothing, filtering, segmentation or derivation step.
10. A method according to any one of claims 1 to 9, wherein, during the time series acquisition phase (40), a configurable time window is applied to the acquired measurement points by moving the measurement points out of said window to form the time series.
11. A method according to any one of claims 1 to 10, the method further comprising an alert phase in which an external electronic medium (12) alerts the operator if a time series from the anomaly detection device (10) is classified as abnormal.
12. Product computer program comprising a readable information carrier, on which is stored a computer program comprising program instructions, the computer program being loadable onto a data processing unit and adapted to drive the implementation of a method according to any one of claims 1 to 11 when the computer program is implemented on the data processing unit.
13. A readable information medium containing program instructions forming a computer program, the computer program being loadable onto a 19 data processing unit and adapted to drive the implementation of a method according to any one of claims 1 to 11 when the computer program is implemented on the data processing unit.