Method for monitoring engine speed of aircraft

Through vibration sensor measurement and spectrum diagram processing, the problems of harmonic interference and noise interference in aircraft engine speed monitoring are solved, and high-accurate speed monitoring is achieved, which is suitable for aircraft engines.

CN118974530APending Publication Date: 2024-11-15INST NAT DES SCI APPLIQUEES DE LYON +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202380031932.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Priority Date
2022-03-28
Filing Date
2023-03-27
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

The prior art is difficult to effectively monitor the rotation speed of aircraft engines, especially in high speed and non-stationary operating conditions. Due to harmonic interference and noise interference, the signal-to-noise ratio is low, it is difficult to distinguish the frequency of interest, and it is difficult to install the sensor.

Method used

Vibration signals are measured by vibration sensors, the original spectrum diagram is constructed, and the denoising process is performed to obtain an equalized spectrum diagram, separate the frequency of interest, reduce harmonic interference, and determine the engine speed. This method performs denoising and source separation in the frequency domain, and uses the bottom of the spectrum to reduce the noise impact, which is suitable for aircraft engines.

Benefits of technology

It realizes the separation of harmonic content of interest in high noise environments, reduces the sensitivity of harmonic interference, improves the accuracy and robustness of engine speed monitoring, and is suitable for aircraft engines and reduces the difficulty of sensor installation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118974530A_ABST
    Figure CN118974530A_ABST
Patent Text Reader

Abstract

One aspect of the invention relates to a method for monitoring the engine speed of a rotary machine, comprising the steps of:-constructing a spectrogram from vibration signals measured by a vibration sensor; -de-noising the spectrogram in order to obtain an equalized spectrogram; -separating the sources in the equilibrium spectrogram by determining an estimate of the frequency of interest; and-determining an instantaneous rotational speed of the engine rotational speed from the estimate of the frequency of interest.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The technical field of the invention is the monitoring of rotating machinery, in particular aircraft engines, by analyzing vibration signals and engine power transmission.

[0002] The present invention relates to a method for monitoring the engine speed of a rotating machine. Background Art

[0003] There are many methods for monitoring rotating machinery, in particular by measuring the vibrations caused by the operation of the machinery and drive elements. However, the methods known from the prior art in this field are difficult to apply to aviation machinery, in particular aircraft and helicopter engines. In fact, monitoring the operation of aircraft engines is very difficult, on the one hand, due to the accessibility and overall size of the shaft system and the rotating engine in a critical environment, which hinders the installation of dedicated monitoring equipment and systems, and, on the other hand, due to the very high engine speeds of these rotating machines, which leads to the appearance of additional harmonics in the vibration signal, while the vibration signal spectrum of the engine in a low-speed operating state has fewer harmonics. These additional frequencies may interfere and mask the frequencies of interest for evaluating the engine operating state. This is called harmonic interference.

[0004] Furthermore, since aircraft engines are designed to reach very high rotational speeds, they experience very high acceleration and deceleration phases. Therefore, the vibration signal obtained during these phases is non-stationary and has a strong variability. Moreover, different parts of the engine do not operate in the same way during acceleration or deceleration phases. Therefore, the harmonic disturbances depend on the operating mode of the engine.

[0005] Furthermore, an aircraft engine can be arranged in several parts that do not necessarily have synchronous operating modes. In general, a turbine engine comprises two rotating shafts, each with a different rotational speed. In this case, these are the rotational speeds N1 and N2 of the shafts. It is also possible that the different parts, in particular the shafts at speeds N1 and N2, have asynchronous operating modes during stationary phases of the engine speed and during acceleration and deceleration phases. These asynchronous operating modes also generate additional harmonic contents that are detrimental to resolving in the signal spectrum the multiple frequencies associated with the operation of the different rotating elements of the engine. In fact, some harmonic components overlap or mask each other, leading to a poor estimation of the engine operating state. The harmonic interference also depends on the operating modes of the different parts of the engine, which may not be synchronized and whose rotational speeds may also vary greatly over time.

[0006] Finally, when an aircraft engine is running, there are many sources of noise, either solid, airborne or thermal. The main sources of noise are the movement of the aircraft due to the combustion of fuel, the aerodynamic flow of fluids in and around the engine, and the resonance of different elements in the engine and aircraft due to the vibrations generated by the engine. These multiple noise sources, most of which are broadband, mean that the signal-to-noise ratio of the acquired vibration signal is very low, making it difficult to distinguish the frequency content of interest, which represents the operating state of the engine, from the noise.

[0007] The vast majority of state-of-the-art monitoring methods rely on measuring a reference vibration signal using a speed or angular position signal (e.g., patents and applications FR3087265A1, WO2019207242A1, FR3102858A1, FR3096133B1, FR3079977B1). Therefore, for these methods, the rotational speed must be measured, which also forces the addition of hardware instrumentation to the monitoring system. However, in the aviation field, the number of sensors that can be installed is very limited. In addition, due to the overall size and compactness requirements of the engine system, it is very difficult or even impossible to install a position sensor, as the resolution of the position sensor is usually too low relative to the measured signal. In addition, commonly used speed or angular position sensors are not very suitable for use in aircraft engines, as they are not very robust to the critical conditions of engine operation (very strong vibrations, high temperature and pressure gradients, high speeds, etc.).

[0008] Furthermore, methods of single frequency analysis of vibration signals are also known, by which the instantaneous rotational speed is estimated. This speed is determined by applying a bandpass filter around a single component of the signal. However, these methods are not compatible with aviation applications due to the above-mentioned harmonic interference, the multiplicity of the harmonics of interest present in such vibration signals, and the high noise levels measured.

[0009] There are also known estimation methods based on comparison with a reference vibration signal, or applied to simple structures where the harmonic content is easily identifiable, because the operating conditions are low speed, the changes are slow or step-like. These methods are not compatible with aviation applications, because the rotating elements of an aircraft are subject to sudden transient conditions, each with its own specific transients, and to high operating conditions, which generate very rich harmonic content. In addition, in non-steady-state conditions, aircraft engines generate considerable spectral leakage in the measured vibration signal spectrum, making it impossible to identify peaks by comparison with the reference signal. In the spectrum, spectral leakage leads in particular to the diffusion of frequency content and the appearance of parasitic harmonics.

[0010] Finally, there are some known methods for the temporal analysis of vibration signals. The disadvantage of these methods is that in the time domain, the noise is highly correlated with the signal of interest, which means that all harmonics must be taken into account during the analysis. In the case of aviation signals, it is not possible to know all the harmonic content of the signal due to the complexity of the machinery and the variety of engine operating modes. Moreover, harmonics may lose amplitude locally and be masked by high noise levels or harmonic interference. Moreover, in these methods, the noise is modeled with a Gaussian behavior, which is unrealistic in the aviation case where the noise is of an impulsive nature.

[0011] Therefore, a method for monitoring the operating status of a rotating machine through a vibration signal is needed to distinguish the harmonic content of interest from the noise generated by the operation of the machine and reduce the sensitivity of the vibration signal to harmonic interference. Summary of the invention

[0012] The present invention provides a solution to the above-discussed problems, namely, monitoring rotating machinery by measuring vibration signals, thereby distinguishing the harmonic content of interest from the noise generated by the operation of the machinery, while reducing the sensitivity of the vibration signal to harmonic interference.

[0013] A first aspect of the invention relates to a method for monitoring the engine speed of a rotating machine, the method comprising the following steps:

[0014] - constructing a raw spectrum graph based on the vibration signal measured by the vibration sensor;

[0015] - De-noise the original spectrum to obtain an equalized spectrum. The denoising steps include:

[0016] o A first sub-step of determining the bottom of the original spectrogram, the first sub-step comprising:

[0017] A first sub-sub-step of constructing a regression function that is robust to the peaks of the original spectrogram;

[0018] A second sub-sub-step of determining the bottom of the spectrum according to the regression function;

[0019] o a second sub-step of determining the equalized spectrum graph based on the bottom of the spectrum graph;

[0020] -separating sources in the equalized spectrogram by determining estimates of the frequencies of interest;

[0021] - Determining the instantaneous speed of the engine speed based on the estimate of the frequency of interest.

[0022] A "rotating machine" is an engine machine whose engine converts energy supplied to the engine into a rotational motion, for example via a shaft system. In the context of the present invention, this refers in particular to aircraft such as airplanes or helicopters, but may also refer to wind turbine engines, engines of wheeled vehicles, etc.

[0023] "Engine speed monitoring" means monitoring the operating state of an engine that rotates a rotating machine, in particular by estimating the instantaneous speed of the rotating machine. The monitoring then makes it possible to predict and / or detect faults in the operation of the rotating machine. For example, by estimating the instantaneous speed, it is possible to check against reference data whether the engine speed of the rotating machine is as expected, in order to perform maintenance on the rotating machine or detect operating anomalies. The instantaneous speed can be defined as the estimated speed within a time interval whose duration is shorter than the duration of the acquisition of the vibration signal.

[0024] A "spectrogram" refers to the frequency content of one or more signals as a function of time. Furthermore, the terms "frequency content", "harmonic content", "harmonic components" and "harmonics" all refer to a set of frequencies (or harmonics) contained in the spectrum of a signal, which is obtained by applying a Fourier transform from the time domain to the frequency domain. Thus, a "harmonic content of interest" is a frequency content representative of the engine speed of a rotating machine, in particular, its instantaneous speed can be determined, the frequency of interest being linearly related to the instantaneous speed.

[0025] A "vibration sensor" is a sensor used to measure the vibrations of a structure. For example, it can be an accelerometer-type sensor based on the piezoelectric effect, a laser vibrometer, a capacitive or eddy current displacement sensor, etc.

[0026] "Harmonic interference" is the phenomenon that two frequency contents coming from different sources and measured simultaneously by the same sensor can partially or completely interfere with each other. This interference can partially or completely weaken the frequency content of the signal, in particular by superimposing (in phase) or masking (anti-phase) the harmonic content.

[0027] According to the present invention, the engine speed of a rotating machine can be monitored by a high noise vibration signal (whose harmonic content of interest is damaged by harmonic interference) and its instantaneous speed can be determined. Therefore, the method is particularly suitable for high-speed applications, such as helicopter and aircraft engines in aviation. The method can also be applied to other types of rotating machines or combustion and explosion machines.

[0028] Unlike applications in the time domain, implementing monitoring methods in the frequency domain can speed up the signal denoising and source separation process. In addition, in the frequency domain, the assumption that noise is decorrelated with the signal is more realistic than in the time domain, so denoising is easier and less biased.

[0029] Thus, the step of denoising the original spectrogram of the vibration signal makes it possible to obtain a balanced spectrogram, since the bottom of the spectrogram representing the broadband noise is used, thus having a uniform noise level at all points on the time axis and the frequency axis. Furthermore, the spectrogram bottom makes the spectrogram robust to the presence of peaks in its spectrum.

[0030] Then, the step of determining an estimate of the frequency of interest enables reducing the effect of harmonic interference in the equalized spectrum by separating the sources in the vibration signal so as to better distinguish the frequencies of interest to determine the instantaneous rotational speed of the rotating machine.

[0031] Finally, the method is easy to implement within the scope of aeronautical applications, since it requires at least one vibration sensor whose size is compatible with the overall size and compactness issues of aeronautical systems. Therefore, the method according to the invention requires only a small amount of equipment to be installed to operate. Moreover, since vibration sensors are more robust and more sensitive to high frequencies than speed or position sensors in critical environments, the method is suitable for aeronautical applications.

[0032] In addition to the features discussed in the previous paragraph, the method according to the first aspect of the invention may have one or more of the following additional features, which may be considered alone or in any technically feasible combination.

[0033] According to an alternative embodiment, the method further comprises the step of detecting an operational anomaly of the engine of the rotary machine based on the instantaneous rotational speed.

[0034] By means of this alternative, abnormal operation of one or more rotating elements in the engine shafting of a rotating machine can be detected. Abnormal operation means that the operating state of the engine speed does not meet the expectation of the relevant engine speed compared with the rated or reference operating state.

[0035] According to an alternative embodiment, the raw spectrogram is constructed by applying a Fourier transform of the vibration signal over a plurality of consecutive time windows of short duration.

[0036] With this alternative, the spectrogram can be constructed in time slices such that the time interval over which the Fourier transform is applied is small enough to make the assumption of a stationary signal valid.

[0037] According to an alternative embodiment, the source separation step is repeated until a convergence criterion is met, the source separation step comprising a first sub-step of determining the frequency of interest based on the amplitude of the frequency of interest, and a second sub-step of estimating the amplitude of the frequency of interest based on the frequency of interest, the first sub-step and the second sub-step being alternately performed in each iteration of the source separation step, and further determining the estimate of the frequency of interest based on the frequency of interest and the amplitude of the frequency of interest.

[0038] By this alternative, an estimate of the frequency of interest can be determined when the amplitude of the frequency of interest is not initially known. By an iterative search that alternates the estimation of the frequency of interest and the associated amplitude, the frequency of interest content of the harmonic disturbance can be corrected. This correction is also performed adaptively by alternating the steps of estimating the frequency of interest and estimating the associated amplitude. The adaptive nature of this alternative means that the method is robust to changes in the speed of the aircraft engine. Therefore, the method is suitable for estimating the instantaneous speed even if the vibration signal is non-stationary.

[0039] According to an alternative embodiment, the source separation step is implemented by an expectation-maximization algorithm, wherein the first sub-step of determining the frequencies of interest is an E-step of the expectation-maximization algorithm, and wherein the second sub-step of estimating the amplitudes of the frequencies of interest is an M-step of the expectation-maximization algorithm.

[0040] According to an alternative embodiment, the sub-step of determining the frequency of interest from its amplitude comprises a first sub-step of determining the posterior probability of the frequency of interest and a second sub-step of determining an expectation of the log-likelihood of the amplitude of the frequency of interest.

[0041] According to an alternative embodiment, the sub-step of estimating the amplitude of the frequency of interest based on the frequency of interest includes a first sub-step of determining the posterior probability of the amplitude of the frequency of interest, a second sub-step of determining the posterior probability of the amplitude of the frequency of interest, and a third sub-step of determining the variance of the smoothed generalized Gaussian distribution of the equalized frequency spectrum.

[0042] According to an alternative embodiment, when the stopping criterion is met, an estimate of the frequency of interest is determined based on the posterior probability of the frequency of interest.

[0043] According to an alternative embodiment, the method further comprises the step of interpolating the estimates of the frequencies of interest.

[0044] According to an alternative embodiment, the method further comprises a step of smoothing the interpolation of the estimate of the frequency of interest.

[0045] A second aspect of the invention relates to a computer program product comprising instructions which, when the program is executed on a computer, cause the computer to implement the steps of the method according to the first aspect of the invention.

[0046] A third aspect of the present invention relates to a computer-readable recording medium comprising instructions which, when executed by a computer, cause it to implement the steps of the method according to the first aspect of the present invention.

[0047] The invention and its various applications will be better understood by reading the following description and by viewing the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The drawings are for illustration purposes only and are not intended to limit the present invention in any way.

[0049] - Figure 1 is a block diagram illustrating a sequence of steps of a method according to the present invention.

[0050] - Figure 2 Two examples of spectrograms constructed and processed by the method according to the invention are shown for signals acquired from an operating aircraft engine and a helicopter engine.

[0051] - Figure 3 shows the original spectrogram, i.e. the spectrogram not processed using the method according to the invention, Figure 2 The processed spectrograms in are generated from these raw spectrograms.

[0052] - Figure 4 A denoising method for a vibration signal spectrum according to the present invention is shown.

[0053] - Figure 5 It is shown that after the denoising step according to the method of the present invention, Figure 3 The equalized spectrum diagram is obtained after equalizing the original spectrum diagram.

[0054] - Figure 6 Shown for Figure 5 Graph of an equalized spectrum, examples of the a posteriori probabilities of frequencies of interest determined based on the source separation step of the method according to the invention. DETAILED DESCRIPTION

[0055] Unless otherwise stated, identical elements appearing in different drawings have a single reference number.

[0056] A first aspect of the invention relates to a method for determining the engine speed of a rotating machine. The method comprises four main sequential steps.

[0057] A rotating machine is, for example, an aircraft engine, which comprises one or more shaft systems that enable the energy consumed by the engine to be converted into rotational motion. The shaft system may comprise several elements, such as wheels and blades, to enable the aircraft to move.

[0058] Figure 1 A schematic representation of the main steps of a method 100 according to the invention is shown.

[0059] The method 100 according to the invention comprises a first step 110 of constructing a spectrogram. A spectrogram is a time-frequency representation of the harmonic content of an acquired signal. Preferably, this is a vibration signal.

[0060] The vibration signal is collected by a vibration sensor (such as a piezoelectric accelerometer). In addition, the sensor is placed on or close to the shaft system of the rotating machine, for example.

[0061] Figure 2 Two examples of spectrograms of signals collected from a running aircraft engine and a helicopter engine, respectively, are provided in .

[0062] Figure 2 The x-axis of the spectrogram represents time, the y-axis represents frequency, and the grayscale represents the amplitude associated with each frequency.

[0063] The spectrogram represents the time evolution of the frequency content of the measured vibration signal. In addition, the harmonic content of the spectrogram represents the state of the engine speed. In fact, depending on the operating state of the engine, the harmonic content differs due to the natural modes of the elements in the shaft system, which are caused by the excitation and resonance of these elements. It is known that the frequency content of the vibration signal is linearly related to the engine speed of the rotating machine due to the periodicity of the engine rotation. Moreover, this periodicity is found in the harmonic content of the spectrum. Therefore, the evolution of the harmonics and their amplitudes in the vibration signal over time provides direct information about the engine speed of the aircraft engine. In this case, the term "trajectory in the time-frequency plane" is used to refer to the evolution of one or more frequencies that are linearly related to the engine speed. Specifically, these frequencies depend on the instantaneous speed of the engine. For the sake of commonality, the instantaneous speed of the engine is also called the instantaneous frequency, because they are linearly related to each other.

[0064] exist Figure 2 13, spectrograms 1300a and 1300b have been processed by the method 100 according to the invention in accordance with the steps to be described later, in order to better illustrate the points related to the trajectory on the time-frequency plane and to support the significance of the invention compared with the prior art. It can be seen from the processed spectrograms 1300a and 1300b that the evolution of the trajectory 1310 of the highest amplitude harmonic on the time-frequency plane is linearly related to the change in the instantaneous speed of the aircraft engine. It is worth noting that in the first stationary phase RM1 of the engine speed, the trajectory 1310 is initially stationary, and then in the second longer acceleration phase RM2 of the engine speed, the trajectory 1310 changes and its frequency increases until it becomes stationary again when the third stationary phase RM3 of the engine speed is reached.

[0065] In the original spectrogram, ie the spectrogram not processed using the method 100 according to the present invention, it is difficult to track the trajectory 1310 due to the high noise level and harmonic interference. Figure 3 This problem is illustrated in the figure where the original spectrograms 1100a and 1100b are derived from the vibration signal mentioned above and compared with Figure 21300a and 1300b in the processed spectrograms 1300a and 1300b. On these original spectrograms 1100a and 1100b, the harmonic content of interest is masked by broadband noise, the level of which is similar to the level of the frequency of interest. In addition, it is difficult to distinguish the frequency or frequencies representing the engine speed (usually a few peaks related to the rotation of the shaft system) because the harmonic interference cannot distinguish the frequency of interest from the frequencies related to the noise (solid or air source of the engine that is not directly a rotating machine) by amplitude. In addition, depending on the engine speed, the noise may also contain harmonic content, which comes from the resonance of different components that make up the aircraft due to the excitation of their natural modes.

[0066] In addition, from Figure 3 It is obvious from the spectrograms in Figure 1100 that the harmonic interference phenomenon depends on the engine speed. In fact, the harmonic content of the original spectrograms 1100a and 1100b varies greatly between engine speeds RM1 to RM3. Figure 2 and Figure 3 The advantages of the method 100 according to the present invention can be understood by looking at the processed spectrograms 1300a and 1300b and the original spectrograms 1100a and 1100b.

[0067] During the first step 110, a raw spectrogram 1100 can be obtained by applying a Fourier transform (FT) to the vibration signal within a time window of short duration. This is called a short-time Fourier transform (STFT) or a sliding window Fourier transform (SWFT). The raw spectrogram 1100 is obtained by connecting the spectra determined by the STFT for a plurality of consecutive time windows along the time axis. The time interval between two time windows is a predefined duration. For example, the time interval can be shorter than the width of the time window to allow the vibration signal to overlap at least partially with two or more consecutive windows. The advantage of using such a FT is to have a sufficiently short time window so that the vibration signal and therefore its spectral content can be considered to be stationary in this window, even if the engine state is in a transient or non-stationary phase.

[0068] From a mathematical point of view, during the first step 110, constructing the original spectrogram 1100 from the vibration signal can be represented as follows:

[0069] Consider the discrete vibroacoustic signal x(t) {t=1,…,K} , where t represents the time sample and K is the signal length, as complement:

[0070] x(t)=s(t)+v(t) . (Formula 1)

[0071] s(t) can be considered as a sum of sinusoids characterized by a power spectral density (PSD) with finite spectral quality and discrete frequencies. v(t) can be considered as a residual signal consisting of random components. For example, in the case of a vibration signal, s(t) represents the periodic component generated by the rotating machinery and v(t) represents the effect of noise, where the source(s) may be fluid flow around and inside the aircraft, turbulence, or transient events related to the operation of the rotating machinery engine.

[0072] Therefore, the signal s(t) can be considered to contain the harmonic content of interest for estimating the instantaneous rotational speed, and the harmonic series of non-interest can be considered to be contained in the residual signal v(t). This leads to the following analytical expression:

[0073]

[0074] P is the harmonic order of the frequency content,

[0075] is the instantaneous phase associated with the instantaneous frequency f(t) to be estimated. j is an imaginary number here.

[0076] A p (t) is defined as the complex Fourier coefficient of each harmonic of index p. The last term can also be written as:

[0077]

[0078] where c p and θ p is the p-times amplitude and p-times phase of the p-times amplitude.

[0079] The STFT can then be obtained by multiplying the vibration signal x(t) with a sliding window w along the time axis to obtain a time segment of the vibration signal, and then applying the FT to the time segment. The window w used can be any type of function commonly used in signal processing. For example, it can be a Gaussian function, a Hamming function, a Hanning function, etc. The method 100 according to the present invention has no restrictions on the window applied. Preferably, the window applied is selected to minimize its effect on the harmonic content, for example to avoid artificially adding frequencies to the spectrogram and / or to avoid distorting the harmonic content of interest.

[0080] Then, the STFT of the discrete vibration signal x(t) to the discrete spectrum X(n,k) is expressed as:

[0081]

[0082] R is the overlap size between two temporally adjacent windows w(l), where l is the index of a temporal segment. Δ is the frequency step size and L is the segment size. f is the frequency. n and k are the indices of the time and frequency samples in the spectrogram, respectively.

[0083] After substituting formula (Formula 2) into (Formula 4), the expression of X(n,k) becomes:

[0084]

[0085] V(n,k) is the spectrum of the noise and frequencies of non-interest contained in the signal v(t).

[0086] When constructing the original spectrogram 1100, in the first step 110, the duration of the time segment can be considered to be small. Therefore, the instantaneous frequency can be considered to be almost constant within each time segment. This means that f(l+nR)≈f(nR). The expression φ(l+nR) can then be approximated by:

[0087]

[0088] where F s is the sampling frequency. Approximately It is also effective.

[0089] Similarly, for time segments of short duration, the Fourier coefficients, i.e. the amplitude of the frequency content, can be considered constant over each segment. Thus: A p (l+nR)≈A p (nR).

[0090] For simplicity, in the following, the term R will be omitted from the notation so that A p,n ,φ n and f n A p (nR), φ(nR) and f(nR). The following expressions are obtained:

[0091]

[0092] The expression can be defined as

[0093]

[0094] as the Fourier transform of the window w. This quantity is used to introduce the frequency resolution of the STFT, or equivalently, the filter bandwidth of the STFT. Therefore, the following expression can be derived from the previous expression:

[0095]

[0096] It can also be considered that Xn =[X(0,n),…,X(F-1,n)] T , V n =[V(0,n),…,V(F-1,n)] T And A n =[A n ,…,A Pn ] T It can also be considered is H(f n ) kp =e jpφ(n) W(kΔf-pf n ). Therefore, the following expression is obtained:

[0097]

[0098] in The number of time slices.

[0099] Therefore, the original spectrum diagram 1100 constructed in step 110 of the method 100 according to the present invention can be constructed according to the multi-harmonic model defined in the formula (Formula 10), and the multi-harmonic model can be expressed as X=[X0, ..., X n ,...,X N-1 ] T .

[0100] The method 100 according to the invention comprises, after a first step 110 of constructing an original spectrogram 1100, a second step 120 of denoising the original spectrogram 1100. This step 120 makes it possible to obtain a denoised spectrogram that can be used without the risk of causing a deviation in the determination of the engine speed due to amplification of certain frequency regions by broadband noise, which can cause a deviation of one or more traces and locally amplify the amplitudes in the original spectrogram 1100.

[0101] Preferably, the second, denoising step 120 enables to obtain an equalized spectrogram. An equalized spectrogram is an original spectrogram 1100 whose noise has been normalized over all time segments along the time and frequency dimensions of the spectrogram. This step makes it possible to improve the appearance of the harmonic content of interest, which characterizes the shafting gear with respect to the noise. As a reminder, denoising in the Fourier domain is facilitated, since, unlike the time domain, the noise can be considered independent of the harmonic content of interest.

[0102] The step 120 of denoising the original spectrogram 1100 can be based on estimating the bottom of the spectrum. The "bottom of the spectrum" refers to all random harmonic components contained in the original spectrogram 1100. Therefore, this definition includes all effects of broadband noise, such as aerodynamic noise, combustion noise and resonance noise. From a graphical point of view, the bottom of the spectrum is a line that smoothly passes through the broadband components, as if there were no peaks. Therefore, the denoising step 120 can include a first sub-step 121 for estimating the bottom of the spectrum of the original spectrogram 1100. After the first sub-step 121 of estimating the bottom of the spectrum of the original spectrogram 1100, the denoising step 120 can also include a second sub-step 122 for determining an equalized spectrogram. In the second sub-step 122, the equalized spectrogram can be determined by dividing the original spectrogram 1100 by the bottom of the spectrum of the original spectrogram 1100. This can be, for example, a weighted division according to the frequency considered, or a simple division.

[0103] Furthermore, in the second step 120 of the method 100, denoising can be performed by estimating the spectral floor for each time segment defined during the calculation of the STFT. The estimation of the spectral floor in the first sub-step 121 can then be an estimation of the spectral floor for each time segment.

[0104] Furthermore, denoising may be performed by dividing the amplitude in the original spectrogram 1100 of each time segment by the estimated spectral floor. Then, the equalized spectrogram for each time segment determined in the second sub-step 122 may be determined by dividing the original spectrogram 1100 of the time segment by the spectral floor of the time segment estimated in the first sub-step 121. The division may be a simple division or a weighted division, for example as a function of the frequency of the spectrogram.

[0105] Mathematically speaking, the implementation of the second step 120 for denoising the original spectrogram 1100 according to the estimate of the spectral floor of each time segment can be transformed as follows.

[0106] First, for the vibration signal X n For each time segment of the spectrogram, the first sub-step 121 estimates the spectrum The bottom of

[0107] When x(t) contains only random components, the bottom of the spectrum can be defined as However, in addition to the random component v(t), the vibration signal x(t) also includes a deterministic part s(t), which contains the frequencies of interest of the vibration signal. In this case, there is no spectrum for the deterministic part of the signal The exact definition of the bottom. However, it can be based on X nThe trend is used to estimate the bottom of the spectrum of the deterministic part of the signal, even if X n This estimate of the deterministic portion of the vibration signal can be considered equivalent to the original spectrogram 1100X n The regression is performed so as to be robust with respect to peak values.

[0108] Thus, a first sub-step 121 of estimating the bottom of the spectrum may include constructing a raw spectrogram 1100X n The regression function may be constructed in time segments and may furthermore be robust with respect to the peaks contained in the original spectrogram 1100 .

[0109] The regression at sub-step 121a can be written as

[0110] X n =g n +∈(k).(Formula 11)

[0111] g n (.) is the regression function, ∈(.) is the approximate error of the regression.

[0112] The regression function g of the first sub-sub-step 121a n It can be constructed according to a trend estimation method. Preferably, the trend estimation method used to construct the regression function is time efficient and simple so as to control the curvature of the regression.

[0113] Furthermore, preferably, the regression function is modeled using a heavy-tailed distribution of the error in order to take into account the impulsive nature of the noise in the original spectrogram 1100. This takes into account the fact that the aviation vibration signal is affected by multiple components. The use of a heavy-tailed error distribution can also improve the robustness of the regression to interfering harmonics. In fact, it is important to remember that these vibration signals usually include at least two harmonic series, associated with two synchronous and asynchronous states N1 and N2, respectively. If the goal is to determine the instantaneous speed of N1, then N2 is also considered as a noise source in addition to the harmonics of no interest associated with N1.

[0114] The regression function may be a spline regression function, with a flexible choice of error distribution ∈. The basis of the spline may be chosen to be numerically stable. In this case, it may be a B-spline basis, which then in a first sub-substep 121a comprises constructing a B-spline basis of order o and number of knots O. The order and number of knots parameters are predetermined so as to be able to control the original spectrogram 1100X n The regularity of the regression function and reduce its edge effects.

[0115] To make this regression robust, the regression function can also be applied to the original spectrogram 1100X nInstead of applying the logarithm of the original spectrum directly to 1100X n In fact, the logarithmic transformation makes the spectrum amplitude uniform and robust to peaks. In fact, X n The logarithm of the distribution ratio X n The distribution is more symmetrical.

[0116] Therefore, the regression of the original spectrogram 1100 in sub-substep 121a can be written as follows:

[0117]

[0118] B is a B-spline basis, and {c j ; j = 1, ..., O} are the coefficients of the B-spline basis B. The B-spline basis can be constructed by conventional methods in the literature.

[0119] The distribution of the regression error ∈(k) makes the regression more robust to peaks. Here, the distribution of the error ∈(k) can be re-tailed to convert X n The presence of a peak in the error is taken into account. For example, such a distribution of errors ∈(k) may follow Student's law. The maximum likelihood may then be determined based on the Student's law model using, for example, an iteratively reweighted least squares (IRLS) algorithm. The algorithm estimates the error weights by optimizing an objective function, where the error itself depends on the weights of the objective function. In each iteration of the IRLS algorithm, sample values ​​whose errors are greater than a predefined threshold may be truncated. This provides greater robustness with respect to the peak. The predefined threshold may be, for example, a constant multiplied by the median of the absolute value of the error. In addition, the IRLS algorithm makes it possible to obtain the coefficients c of the B-spline basis j .

[0120] The first sub-step 121 of estimating the bottom of the frequency spectrum may also comprise a second sub-step 121b of determining the bottom of the frequency spectrum according to a regression function.

[0121] Given an estimated value {c min,j ; j = 1, ..., o}, the spectrum in the second sub-sub-step 121b The bottom of can be determined by:

[0122]

[0123] Figure 4An example application of the first sub-step 121 is shown, namely estimating the bottom 1220 of the spectrum 1210 of the signal. The signal used here is a noise vibration signal, representative of the vibration characteristics of an aircraft engine and containing multiple peaks. After the sub-sub-steps 121a and 121b of regressing the spectrum 1210 and determining the bottom 1220 of the spectrum 1210, a regression function that is robust to the peaks of the logarithm of the spectrum 1210 is first constructed, and then the bottom 1220 of the spectrum 1210 is determined.

[0124] It can thus be observed that the estimate of the bottom 1220 of the spectrum is in fact robust with respect to the peaks: it follows the trend of the spectrum 1210 and is not affected by nearby peaks. It can also be noted that estimating the bottom 1220 of the spectrum can be likened to determining the spectral density of the noise in the spectrum 1210.

[0125] Then, in a second sub-step 122, the spectrum X n Divided into spectrum The function at the bottom determines the equalized spectrum of each time segment 1200Y n , that is, the spectrum of the noise is uniform along the time axis and frequency axis, so that:

[0126]

[0127] To illustrate the effect of denoising by equalizing the spectrum, Figure 5 Two equalized spectrum diagrams 1200a and 1200b are shown. The equalized spectrum diagrams 1200a and 1200b are respectively Figure 3 The original spectrograms 1100a and 1100b are obtained by denoising according to the second step 120 of the method 100. It should be noted that the spectrogram 1100a is generated by the vibration signal collected by the aircraft engine, while the spectrogram 1100b is generated by the vibration signal collected by the helicopter engine. As can be seen from these two examples, the equalized spectrograms 1200a and 1200b have been denoised from the broadband noise contained in the original spectrograms 1100a and 1100b. In addition, the frequencies of interest emerge more clearly from the background noise. Therefore, the denoising step 120 is very robust to the peaks contained in the original spectrograms 1100a and 1100b; in addition, considering the large number of spectral peaks generated by different rotating parts in the aircraft vibration signal, this robustness is crucial for implementing the method 100 according to the present invention.

[0128] Advantageously, the denoising step 120 by equalizing the spectrum according to the provided time segments can not only attenuate the background noise, but also attenuate the quasi-stationary components of the time-frequency plane when the sought component is highly non-stationary (speed increase, speed decrease), and conversely, attenuate the non-stationary components when the sought component is quasi-stationary.

[0129] The method 100 according to the invention further comprises a third step 130 of separating the sources. Source separation refers to distinguishing between two or more signal sources present in the same signal (here a vibration signal). Preferably, source separation is used to distinguish between the harmonic content of interest masked by the harmonic content of non-interest caused by harmonic interference and various noise sources that may coexist, intersect, be very close to or even merge in the same frequency band. Thus, this step reduces the sensitivity of the equalized spectrum diagram 1200 to harmonic interference. Thus, one or more trajectories 1310 and their evolution can be tracked. It is to be remembered that the different mechanical elements in the shaft system are sources of the harmonics of interest, which can be identified in the spectrum diagram as trajectories that vary with the instantaneous speed. The engine speed, in particular the instantaneous rotational speed, is determined based on these trajectories.

[0130] The third step 130 is preferably a step of determining an estimate of the frequency of interest. The estimate of the frequency of interest represents the harmonic content of interest of the equalized spectrogram 1200. The estimate of the frequency of interest is determined, for example, based on the restrictions imposed on the trajectories 1310 observed in the equalized spectrogram 1200. In addition, this can be a restriction related to the frequency of one or more trajectories 1310, and / or a restriction on the amplitude of these frequencies. In practice, the engine speed, in particular its instantaneous speed, can be determined by following the maximum amplitude in a given frequency band. Therefore, it is necessary to ensure that the maximum amplitude is associated with the instantaneous speed to be determined, and there is a risk that it may deviate from the correct trajectory and therefore provide an erroneous estimate of the instantaneous frequency.

[0131] The restrictions on the trajectory 1310 may also be related to the regularity of the trajectory. The regularity may, for example, define the evolution of one or more trajectories 1310 and / or the associated amplitudes. The regularity of the trajectory 1310 may also be defined as a function of one or more predefined frequency bands. These may be frequency bands that are, for example, a function of the shafting size.

[0132] Using such a regularity constraint may also avoid jumps and discontinuities in the trajectory 1310 , for example due to isolated peaks of residual noise present after the denoising step 120 .

[0133] Furthermore, an estimate of the frequency of interest may be determined based on one or more frequencies of interest and associated amplitudes in the equalized spectrum graph 1200. Preferably, the estimate of the frequency of interest is determined based on the distribution of the harmonic content of interest in the equalized spectrum graph 1200. The estimate of the frequency of interest may be determined, for example, by maximizing the distribution of the frequency of interest in the equalized spectrum graph 1200.

[0134] As mentioned previously, multiple trajectories 1310 can be tracked simultaneously. The advantage of considering multiple harmonics instead of just one harmonic is that, for example, different trajectories and their harmonics can be weighted in relation to each other. Thus, more weight can be given to trajectories that are considered to be close to the desired instantaneous frequency and the sensitivity of the harmonic content of interest to harmonic interferences, such as the sensitivity to coupling moments between different components, can be reduced. For example, the harmonics monitored are predetermined depending on the size of the shafting. They may also be the most energetic harmonics, a function of the size of the rotating elements of the engine, which makes it possible to avoid deviations due to one or more energy interference components.

[0135] This multi-component approach is possible because the vibration signal is processed in the spectral domain. The assumption that the harmonics of interest and non-interest are uncorrelated is actually robust enough in the spectral domain to allow using only some of the harmonics of the vibration signal instead of all of them as in the time domain. This decorrelation assumption is advantageous because aviation vibration signals are very rich in harmonic content and considering all of them would be intractable.

[0136] When considering the multiharmonic model proposed in formula (Equation 10), only the original spectrum of the vibroacoustic signal X is known. n 1100. In particular, the amplitude A of the frequency content of interest n However, in order to estimate the frequency of interest f n , to determine the instantaneous engine speed, the amplitude A must be known n In practice, the instantaneous rotation speed can be identified based on the amplitudes of the different tracks, preferably based on the track with the highest amplitude. Therefore, the third source separation step 130 is to estimate the amplitude A of the harmonic content of interest by estimating the frequency of interest. n Identify the frequency of interest f in the harmonic content in the estimate of n Therefore, this estimate of the frequency of interest enables the frequency of interest f n Distinguish from the effects of harmonic interference.

[0137] An estimator of the frequency of interest can be determined using an iterative search algorithm. For example, this can be an algorithm based on minimization of a cost function, such as a multi-objective gradient method or a genetic algorithm. Methods based on Bayesian theory can also be used.

[0138] The iterative search algorithm is preferably an expectation maximization (EM) algorithm. Since it uses a Bayesian approach, this algorithm allows the introduction of regularity and continuity constraints in the estimation process. This reduces the risk of bias in the trajectory 1310 estimate. The EM algorithm is used to obtain the posterior probability distribution of the frequency. The estimate of the frequency of interest can be further determined according to the posterior probability law of the frequency.

[0139] The EM algorithm includes two steps of alternating iterations: an E step or expectation step, and an M step or maximization step. Therefore, preferably, the third source separation step 130 includes two sub-steps 131 and 132, corresponding to the E step and the M step of the EM algorithm, which are repeated alternately until a predefined convergence criterion is met. Therefore, step 130 is an alternating iteration of sub-steps 131 and 132.

[0140] The first sub-step 131 or E-step is to calculate the amplitude A of the harmonic content of interest. n Estimate the frequency of interest f n The second sub-step 132 or M step is based on the frequency value f determined in the previously executed sub-step 131. n Estimated Amplitude n steps.

[0141] The predefined convergence criterion is preferably a criterion defined such that the frequency of interest f n and the amplitude A of the frequency of interest n The final error of the estimate of is small enough, for example, below a predefined error threshold. The convergence criterion may be, for example, a predefined error threshold to be reached or a maximum number of iterations of step 130 to be performed.

[0142] It should be noted that the spectral content of the noise V n Modeled by a heavy-tailed distribution to account for the multi-harmonic nature of non-interesting harmonic content. Preferably, the spectral content V(n,f), n∈{1,…,N} and k∈{0,…,F-1} are considered independent and identically distributed according to a smoothed generalized Gaussian (SGG) distribution. The SGG distribution can be as follows:

[0143]

[0144] where β>0 is the shape parameter, σ>0 is the shape factor (also called the variance of the SGG distribution), δ>0 is a constant, and C(σ,δ,β) is a normalizing constant.

[0145] The advantage of using SGG over the generalized Gaussian (GG) distribution is that the constant δ that does not exist in GG usually takes a smaller value, such as δ = 10 -4 , which makes SGG have Gaussian-like behavior near 0 and preserves robust statistics. In addition, the parameter δ allows the probability density to be differentiable at the origin, allowing for more efficient use of gradient-based methods.

[0146] The shape parameter β represents the impulse nature of the spectral noise. The shape parameter β is preferably chosen to have a value between 0 and 2 to preserve the mathematical properties of the Gaussian distribution. To avoid any divergence due to convexity, the shape parameter β should not be chosen too small. Preferably, the parameter is greater than 0.1.

[0147] It should also be noted that the frequency of interest f n is a discrete variable. They can take values from a set of possible frequencies of base B . For example, can be the entire frequency axis of the STFT, in which case B = F, or just a subset of the frequency axis, in which case B < F, if the minimum and maximum bounds of the frequency of interest f n are known. Thus, the log-likelihood l(.) can be written, where the frequency of interest f n is marginalized from the joint distribution:

[0148]

[0149] where v k represents the values that the frequency f n can take in the frequency set , is the Kronecker delta.

[0150] In equation (Equation 16), the logarithm is applied to the sum of the densities. Thus, the sum in the logarithm prevents the logarithm from being directly applied to the joint probability p(Y n , f n = ν k |A n ). This leads to a complex expression for the solution of the posterior maximum. However, for a given frequency of interest f n , given the negative log-likelihood function of {Y n , f n |A n} given by -logp(Y n , f n}, the estimator of the magnitude A n of the frequency of interest f can be estimated from the posterior distribution n of the magnitude after obtaining the observed value Y n . This is equivalent to minimizing the negative logarithm of the posterior probability given by:

[0151]

[0152] where G(A n |Y n , f n ) is the posterior distribution of the magnitude after obtaining the observed value Y nThe posterior probability of the latter consideration. p(A n ). The posterior probability represents the amplitude A n The log-likelihood l(A n ; Y n ) represents the observed value Y n Can be combined with parameter A n Related possibilities.

[0153] When the frequency f is observed n When , minimizing (16) is straightforward. However, only the observations Y n is known, because f n In the case of the vibration signal of an aircraft engine, it is not observable. Instead, the frequency distribution f n The a priori law p k (n) are observable at each time step n: p1(n),…,p B (n). It gives the frequency f n exist where ν1,…,ν B The prior probability of .

[0154] This approach to solving the problem of unobservable parameters can be achieved by using an algorithm such as the EM algorithm. Therefore, the algorithm aims to obtain an a posteriori maximum (APM) estimator from the observations. In the EM algorithm, the frequency of interest f n is a hidden variable, i.e., used to estimate the amplitude A n This choice is motivated by the ease of computing the expected log-likelihood of discrete variables. In practice, the frequencies of interest, f n has discrete values, and the amplitude A n Has continuous values.

[0155] Mathematically speaking, the EM algorithm can be reflected as follows.

[0156] The expectation step or E-step, which is also a sub-step 131 of the method 100 according to the present invention, may include a first sub-sub-step 131a for determining a frequency of interest f n The posterior probability z k (n) (i+1) This first sub-step 131a then makes it possible to determine in a second sub-step 131b the log-likelihood l(A) according to the formula (Formula 16) n ; Y n ) In addition, the frequency of interest f n The conditional distribution of Given the observed data Y n and the current estimate of the magnitude given To determine the expected log-likelihood The index i represents the current iteration number of the EM algorithm.

[0157] In the first sub-substep 131a, the posterior probability z k (n) (i+1) It can be defined as:

[0158]

[0159] Using the Bayesian formula, the posterior probability z of formula (Formula 18) is k (n) (i+1) can be written as:

[0160]

[0161] A priori law of frequency distribution p k (n) can be initialized as a function of the known values ​​of the frequencies of interest in the spectrogram. These values ​​may be known, for example, from a dimensional analysis of the shafting of an aircraft engine. The purpose is to provide information about f n Prior information about the frequencies of interest. These values ​​can be the minimum value f min and the maximum value f max , or values ​​around harmonics ([kf min ,kf max ]).

[0162] Then, by considering the frequency of interest f n The previous estimate of p k (n) is updated from time n to time n+1. In addition, the frequency of interest f n The previous estimate of is integrated into the joint distribution p(f n ,…,f n-M ) to obtain the marginal distribution.

[0163] The joint distribution p(f n ). Assuming M = 1, the joint distribution becomes where q(.|f n-1 =v k ; ε 2 ) is the mean value ν k and variance ε 2 This is equivalent to replacing the previous conditional density p(f n-1 |A n-1 ,Y n-1 ) is convolved with a Gaussian kernel. It should be noted that v k Indicates frequency f n The values ​​that can be taken in the frequency set F.

[0164] Figure 6 The posterior probability distribution z of the frequency is listed k These posterior probability laws z k is from Figure 6 The a posteriori probability law 1310a is derived from the equalization spectrum diagram 1200a for an aircraft engine application. The a posteriori probability law 1310b is derived from the equalization spectrum diagram 1200b for a helicopter engine application. This figure emphasizes that, depending on the choice of shape parameter β, the a priori law p k (n) The continuity of the frequency curve and harmonic order between time points is guaranteed, and the trajectory of interest can become the trajectory with the highest energy in the spectrum diagram.

[0165] According to the posterior probability distribution z k , one or more point estimators can be determined to estimate the instantaneous frequency f. For example, the instantaneous frequency The point estimator of can be the one that maximizes the posterior probability z k (n) (i+1) The value of .

[0166] In sub-step 131, the negative log-likelihood -log(f n |Y n ,A n ) can be e.g. relative to the posterior probability Write out. Negative log-likelihood -log(f n |Y n ,A n ) can be given by the following formula:

[0167]

[0168] Then, the posterior probability z k (n) (i+1) and the expected log-likelihood Can be used in the M step to determine the amplitude A of the frequency of interest n .

[0169] The maximization step or M-step, which is also a sub-step 132 of the method 100 according to the invention, is used to determine the amplitude A n The value of .

[0170] During the maximization step, the amplitude A can be determined by finding the argument that maximizes its joint density or equivalently minimizes its negative logarithm n :

[0171]

[0172] During a first sub-substep 132a, the a posteriori probability G of the amplitude is determined (i+1) (A n |Y n ). For example, the posterior probability G (i +1) (A n |Y n ) is determined by the following formula:

[0173]

[0174] Unknown Amplitude A n The prior distribution p(A n ) can follow a predefined model. Preferably, the amplitude A n The evolution of follows an autoregressive model taking into account old measurements, which makes it possible to limit the amplitude of the trajectory A n For example, the predefined model can be a linear model such that A n =αA n-1 +∈ n , where ∈ n Following a complex circular Gaussian distribution, α is a known constant such that 1 ≥ α > 0. Alternatively, the predefined model may be a nonlinear model.

[0175] In step 132, the amplitude A is determined n The solution formula (Formula 21) can be further included in the second sub-sub-step 132b. The second sub-sub-step 132b allows, for example, determining the posterior probability G by an iterative algorithm. (i+1) (A n |Y n ). In addition, since the posterior probability G (i+1) (A n |Y n ) is differentiable and convex for β ≥ 1, so an iterative algorithm based on gradient estimation can be used. For example, the mainization-minimization (MM) algorithm can be used.

[0176] Sub-step 132 may then include sub-sub-step 132c for updating the shape factor σ or variance of the SGG distribution of the noise. The update may be performed by estimating the maximum likelihood of the shape factor σ of the noise distribution according to formula (Formula 17). For example, the maximum likelihood estimate may be expressed as

[0177]

[0178] In the q+1 iteration of the MM algorithm and the i+1 iteration of the EM algorithm, once the posterior probability gradient is estimated It can give an estimated value

[0179]

[0180] The amplitude A of the SGG distribution can then be used in further iterations of the EM algorithm n and variance To determine the posterior probability z of the frequency of interest k (n).

[0181] At the end of step 130, after the convergence criterion of the EM algorithm has been satisfied, the estimated value of the frequency of interest determined is, for example, the posterior probability z for all time instants n determined in the last iteration of substep 131. k (n).

[0182] Finally, the method 100 according to the invention comprises a fourth step 140 for determining the instantaneous rotational speed of the aircraft engine. The instantaneous rotational speed is preferably estimated from an estimate of the frequency of interest. Advantageously, the instantaneous rotational speed is estimated using a regression function whose errors follow a heavy-tailed distribution, which makes it possible to demodulate the instantaneous rotational speed from a plurality of harmonics simultaneously. Thus, the robustness of the estimate is greatly improved with respect to deviations caused by other harmonics that are not of interest.

[0183] The instantaneous speed can be determined from a portion of the estimate at the frequency of interest. For example, the estimate at the frequency of interest is the a posteriori probability z at all times n. k In the case of (n), the instantaneous frequency can be determined based on the maximum value of the estimate of the frequency of interest. The maximum value of the estimate of the frequency of interest is particularly suitable for the case where there are multiple local maxima in the frequency of interest that need to be distinguished.

[0184] Alternatively, its mean can be used rather than the maximum of the estimator, giving a smoother estimate of the frequency of interest.

[0185] Instantaneous frequency It can be determined at each moment n by the following formula:

[0186]

[0187] Due to the instantaneous frequency is determined for the instant n, so the method 100 according to the present invention may include a fifth step 150, for the instantaneous frequency The interpolation makes it possible to obtain a signal of the same length as the discrete vibration signal x(n). The interpolation can be performed, for example, by spline-type interpolation.

[0188] The method 100 according to the present invention may further comprise a sixth step 160 of determining the instantaneous frequency This smoothing process can reduce the step-like effect caused by frequency resolution.

[0189] The corrected spectrum graph 1300 may then be plotted, such as Figure 2 Spectrograms 1300a and 1300b of FIG. 1300a and FIG. 1300b show a trajectory 1310, where the trajectory 1310 is clearly visible. Therefore, the trajectory 1310 with the largest amplitude is the trajectory of the instantaneous speed of the aircraft engine.

[0190] The method 100 according to the invention may also comprise a seventh step 170 of detecting an operational anomaly of the aircraft engine. Preferably, this detection is performed based on the instantaneous rotation speed or the associated trajectory 1310 determined in one of the steps 140 to 160 .

[0191] For example, the operating anomaly is detected by comparing the instantaneous rotational speed or the associated trajectory 1310 with the reference rotational speed or data related to the reference trajectory 1310. In addition, the detection can inform the existence of the anomaly by issuing an alarm. For example, when it is determined that there is a significant deviation between the determined instantaneous rotational speed and the reference instantaneous rotational speed, the detection can be issued. For example, when the deviation is greater than a predefined normal operating threshold, the deviation is significant, and the threshold can be known by performing a dimensional analysis of the shafting of the aircraft engine.

Claims

1. A method (100) for monitoring engine speed of a rotating machine, the method (100) comprising the following steps: - constructing (110) an original spectrogram (1100) based on the vibration signal measured by the vibration sensor; - De-noising (120) the original spectrogram (1100) to obtain an equalized spectrogram (1200), the de-noising step (120) comprising: o A first sub-step (121) of determining the bottom of the original spectrum graph (1100), the bottom of the spectrum being a set of random harmonic components contained in the original spectrum graph (1100), the first sub-step (121) comprising: - a first sub-sub-step (121a) of constructing a regression function that is robust to the peaks of the original spectrogram (1100), the regression function being applied to the logarithm of the original spectrogram (1100) or to the original spectrogram (1100); A second sub-sub-step (121b) of determining the bottom of the spectrum according to the regression function; o a second sub-step (122) of determining the equalized spectrum graph (1200) based on the bottom of the spectrum graph; - separating (130) sources in the equalized spectrum graph (1200) by determining estimates of frequencies of interest; - Determining (140) an instantaneous engine speed based on the estimate of the frequency of interest.

2. The method (100) for monitoring the engine speed of a rotary machine according to the preceding claim, the method (100) further comprising a step (170) of detecting an operational abnormality of the engine of the rotary machine based on the instantaneous speed.

3. A method (100) for monitoring the engine speed of a rotating machine according to the preceding claim, wherein the original spectrogram (1100) is constructed by applying a Fourier transform to the vibration signal over a plurality of consecutive time windows of short duration.

4. A method (100) according to any of the preceding claims, wherein the source separation step (130) is repeated until a convergence criterion is met, the source separation step (130) comprising a first sub-step (131) of determining a frequency of interest based on the amplitude of the frequency of interest, and a second sub-step (132) of estimating the amplitude of the frequency of interest based on the frequency of interest, the first sub-step and the second sub-step (131, 132) being performed alternately in each iteration of the source separation step (130), and further determining an estimate of the frequency of interest based on the frequency of interest and the amplitude of the frequency of interest.

5. A method (100) according to the preceding claim, wherein the source separation step (130) is implemented by an expectation-maximization algorithm, wherein the first sub-step (131) of determining the frequency of interest is an E-step of the expectation-maximization algorithm, and wherein the second sub-step (132) of estimating the amplitude of the frequency of interest is an M-step of the expectation-maximization algorithm.

6. A method (100) according to one of claims 4 and 5, wherein the sub-step (131) of determining the frequency of interest based on the amplitude of the frequency of interest includes a first sub-step (131a) of determining the posterior probability of the frequency of interest, and a second sub-step (131b) of determining the expectation of the log-likelihood of the amplitude of the frequency of interest.

7. A method (100) according to any one of claims 4 to 6, wherein the sub-step (132) of estimating the amplitude of the frequency of interest based on the frequency of interest includes a first sub-step (132a) of determining the posterior probability of the amplitude of the frequency of interest, a second sub-step (132b) of determining the posterior probability of the amplitude of the frequency of interest, and a third sub-step (132c) of determining the variance of the smoothed generalized Gaussian distribution of the equalized spectrum diagram (1200).

8. The method (100) according to one of the preceding claims, wherein when a stopping criterion is met, an estimate of the frequency of interest is determined based on the posterior probability of the frequency of interest.

9. The method (100) according to any of the preceding claims, further comprising the step of interpolating (150) the estimate of the frequency of interest.

10. The method (100) according to the preceding claim, further comprising a step of smoothing (160) the interpolated value of the estimate of the frequency of interest.