Method of detecting defect of at least one gear in aircraft turbomachine based on vibration signal of gear
By combining the whitening vibration spectrum and Bayesian method, the modulation module of the gear is quasi-blindly estimated, which solves the problem of difficult detection of gear defects in the prior art, and realizes accurate detection without knowing the modulation period, which improves the Robustness and reliability of detection.
Patent Information
- Application Number
- CN202380036228.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2022-05-09
- Filing Date
- 2023-05-09
- Publication Date
- 2024-12-06
AI Technical Summary
The prior art is difficult to accurately and reliably detect defects in aircraft turbine gears without knowing the gear modulation cycle, and the existing phenomenological model determination is complex and unreliable.
Through the whitening spectrum of the vibration signal, multiple modulation modules with modulation index are estimated, and quasi-blind estimation is used to use carrier coefficients and symmetric modulation coefficients to automatically estimate the target signal, combine with Bayesian method to determine health indicators and compare them with the database to detect defect.
It realizes accurate detection of gear defects without knowing the gear modulation period, improves detection robustness and reliability, and can effectively distinguish spectrum peaks from noise.
Smart Images

Figure CN119096120A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the field of detecting defects in gears of aircraft turbomachines by means of vibration signals of the gears. Background Art
[0002] It is known to measure the vibration signal of a turbine and to analyze the signal by spectrum analysis to detect the signal characteristics when a defect is present.
[0003] To be effective, this type of spectrum analysis requires knowledge of the turbine kinematics in order to be able to monitor changes in the signal characteristics. This monitoring method is only possible if the signal characteristics are known "a priori", i.e. the characteristic frequencies in the spectrum are known. In practice, a priori knowledge of the turbine kinematics cannot be guaranteed.
[0004] In order to eliminate this drawback, it has been proposed to model the modulation phenomenon of the gear by means of a phenomenological model which makes it possible to determine the theoretical characteristic frequencies of each type of gear defect in the frequency spectrum. In practice, such a phenomenological model is complex to determine and is also unreliable.
[0005] There is also a need to determine the gear modulation without knowing the gear modulation period, i.e. by a blind method, so that gear defects can be detected accurately and reliably at any time.
[0006] In the prior art, an article numbered XP93002238A in MDP1 2021 entitled “Application of Structured Prior Models in Bayesian Compressed Sensing of Modulated Signals” proposed a method in which the input is a signal sampled in time. Summary of the invention
[0007] The present invention aims to at least partially eliminate these disadvantages.
[0008] The invention relates to a method for detecting defects of at least one gear in an aircraft turbomachine based on a vibration signal of the gear, each gear having a determined meshing frequency corresponding to a carrier frequency, the method comprising:
[0009] - a step of determining a vibration spectrum from the vibration signal,
[0010] - a step of whitening the vibration spectrum to obtain a whitened vibration spectrum,
[0011] - a step of estimating a target signal comprising a set of characteristic spectral lines based on the whitened vibration spectrum,
[0012] - a step of determining at least one health indicator based on at least one spectral line feature of the target signal,
[0013] - a step of comparing at least one health indicator with a database of indicators to detect defects,
[0014] The present invention is notable in that the step of estimating the target signal comprises:
[0015] - a step of determining a plurality of modulation modules having a modulation index in the whitened vibration spectrum,
[0016] -Each modulation module includes:
[0017] o a carrier coefficient, which is a function of a predetermined carrier frequency and a modulation index, and
[0018] ○Multiple modulation coefficients that are symmetrical with the carrier coefficients.
[0019] Thanks to the invention, the modulation coefficient can be estimated in a quasi-blind manner, exploiting its symmetrical structure, while being robust to measurement noise and non-modulation periodicity. In this way, the relevant target signal is determined, improving the detection of gear defects. Advantageously, the whitening step makes it possible to reveal the symmetry assumption of the original signal and thus exploit it.
[0020] As taught in the 2021 MDPI paper XP93002238A entitled “Application of Structured Prior Models in Bayesian Compressed Sensing of Modulated Signals,” this symmetry assumption cannot be exploited in undersampled input signals.
[0021] Preferably, a Bayesian approach is used to estimate the target signal. This approach enables, on the one hand, the automatic estimation of point estimators and, on the other hand, the determination of complete a posteriori probabilities. This enables the calculation of uncertainty deviations.
[0022] Preferably, the modulation modules are different from each other, in particular disjoint. This makes it possible to add additional conditions to facilitate the estimation of the modulation coefficients.
[0023] Preferably, the symmetrical modulation coefficients are estimated jointly, which can speed up the estimation by emphasizing the symmetry.
[0024] Preferably, for each modulation block, each symmetrical modulation coefficient is estimated at the same position relative to the carrier coefficient. In this way, we can use the symmetry within and between blocks to determine the position of the modulation coefficient.
[0025] Preferably, each symmetrical modulation coefficient is considered to be a realization of a multivariate heavy-tailed distribution. This allows the spectral peaks to be distinguished from the noise. This is because the spectral peaks have a heavy-tailed distribution (a distribution with very low standard deviation). The noise is more evenly distributed throughout the spectrum. Therefore, this distribution can be used to distinguish the modulation peaks in the spectrum from random noise.
[0026] Preferably, the whitened vibration spectrum γ is defined by the following formula:
[0027] γ=Φ -1 x+w
[0028] -in:
[0029] -Φ is the Fourier transform is the conjugate transpose of the matrix Φ, I N is the identity matrix of size N;
[0030] -x∈C N is the vector of unknown Fourier coefficients of the target signal x(n);
[0031] -W is additive noise.
[0032] Preferably, the modulation coefficient includes a real part and an imaginary part, the real part and the imaginary part follow a mean of zero and a covariance matrix of τγ q Σ is a Gaussian distribution, where
[0033] γ q ≥0 is the mixing parameter according to a certain mixing distribution
[0034] · in is the Kronecker product, I2 is the identity matrix of size 2×2 and
[0035] Global variable v b Potential parsimony in tuning modulation modules: small amounts of v b The appropriate value may result in an over-contraction of the activity coefficients, especially those with smaller amplitudes.
[0036] Preferably, the modulation module is in the form of the following hierarchical model
[0037] v b ~IG(α b , β b )
[0038]
[0039] - where IG(α b ,β b ) is a positive constant parameter α b and β b The inverse gamma distribution, C + (0,e) is a half-Cauchy distribution with site 0 and scale e.
[0040] It is recommended to use semi-Cauchy priorities for both local and global parameters to improve shrinkage. An inverse gamma distribution is also suitable given that its density vanishes at the origin, which artificially makes v bThe value of is far from zero. On the other hand, it is the conjugate density of the Gaussian distribution, which makes the calculation easier.
[0041] Preferably, the vector of carrier coefficients is determined by a Gaussian prior, such as x0~CN(0,τγ0Diag(v)), where γ0>0.
[0042] If activity group b proves to be very parsimonious, v can be avoided b The value of is zero. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] The invention will be better understood on reading the following description, given by way of example, and with reference to the following drawings, given by way of non-limiting example, in which like references indicate similar objects.
[0044] Figure 1 It is a schematic diagram of a method for detecting defects of a gear according to a specific embodiment of the present invention.
[0045] Figure 2 Schematic diagram of the vibration spectrum represented in logarithmic scale and the whitened vibration spectrum.
[0046] Figure 3 A close-up of the whitened vibrational spectrum on a linear scale.
[0047] Figure 4 is a schematic diagram of the estimation steps of the modulation module.
[0048] Figure 5 It is a schematic diagram of the coefficients of the modulation module.
[0049] Figure 6 is a schematic diagram of the whitened vibration spectrum with the estimated modulation module in the initial experiment.
[0050] Figure 7 is Figure 6 Close-up diagram of the whitened vibration spectrum at the second modulation module in .
[0051] Figure 8 is a schematic diagram of the whitened vibration spectrum with the estimated modulation module in the second experiment.
[0052] Fig. 9 yes Figure 8 A close-up diagram of the whitened vibration spectrum of
[0053] Fig.10 is a schematic diagram of the whitening vibration spectrum measured on day 1 of the third experimental period.
[0054] Fig.11 is a schematic diagram of the albino vibration spectrum measured on the 12th day of the third experimental period.
[0055] Fig.12 yes Fig.10 A close-up diagram of .
[0056] Fig.13 yes Fig.11 A close-up diagram of .
[0057] Fig.14 It is a schematic diagram of the evolution of the index over time.
[0058] It should be noted that the accompanying drawings illustrate the invention in detail in order to implement the invention and, if necessary, said drawings can of course be used to better define the invention. DETAILED DESCRIPTION
[0059] The invention relates to a method for detecting defects in at least one gear of an aircraft turbomachine based on a vibration signal of the gear.
[0060] The vibration signal can be measured by a vibration or force sensor installed in the turbine. The measurement signal can be a pure vibration or a vibroacoustic signal.
[0061] The invention will be presented in the context of a single gear, but it is more generally applicable to multiple gears, and in particular to the gear train of a turbine accessory gearbox.
[0062] According to the present invention, reference Figure 1 , the method comprises the following steps:
[0063] - a step E2a of determining the vibration spectrum z(n) from the vibration signal s(n),
[0064] - a step E2b of whitening said vibration spectrum z(n) to obtain a whitened vibration spectrum γ(n),
[0065] - a step E3 of estimating a target signal x(n) comprising a set of characteristic spectral lines by whitening the vibration spectrum γ(n),
[0066] - a step E4 of determining at least one health indicator IND by means of at least one spectral feature of the target signal x(n), and
[0067] - A step E5 of comparing at least one health indicator IND with a database of indicators DB(IND) in order to detect defects DEF.
[0068] In this example, the vibration signal s(n) is sampled at a sampling frequency fs to obtain n samples (n varies from 1 to N).
[0069] Using the method described in the present invention, a health indicator IND related to the detection of a gear defect DEF can be accurately determined.
[0070] It is complicated to estimate the target signal x(n) including a set of characteristic spectral lines from the whitened vibration spectrum y(n) in a quasi-blind method. In the present invention, only the meshing frequency fc (hereinafter referred to as carrier frequency fc) peculiar to the gears is known.
[0071] The individual steps in this method are described below.
[0072] Step E1 to recalibrate the vibration signal
[0073] In this example, reference Figure 1 , the method includes an optional preliminary step, which includes: if the rotation speed N of the gear rot In this case, if the speed v of the gear is not fixed, i.e., non-constant, the vibration signal s(n) acquired in the time domain is resampled to obtain a vibration signal s recalibrated in angle. ang (n). This step requires knowing the rotation frequency N of the gear in advance. rot The latter can be estimated by means of a position sensor, for example, located on the gear shaft. If the speed is constant (v=vcons), the vibration signal s(n) can be directly converted into the vibration spectrum z(n).
[0074] Step E2a of determining the vibration spectrum from the vibration signal
[0075] Using known methods, such as Fourier transform, the vibration signal s(n) (or the recalibrated vibration signal s ang (n)) determines the vibration spectrum z(n) as follows.
[0076] Step E2b of whitening the vibration spectrum z(n) to obtain a whitened vibration spectrum y(n)
[0077] When the vibration signal s(n) is measured, it is usually filtered by a mechanical transfer function associated with the transmission between the excitation source and the sensor. This transfer function affects the shape of the vibration spectrum z(n) and may change various characteristics of the control modulation that we wish to exploit in the present invention.
[0078] The method also includes the step of whitening the vibration spectrum z(n) to weaken the influence of the transfer function associated with the transmission. In order to whiten the vibration spectrum z(n), an inverse transfer function is determined and applied to the vibration spectrum z(n) to compensate for the transfer function associated with the transmission. In other words, the influence of the transmission on the vibration spectrum z(n) is eliminated, so that the modulation can be highlighted, as will be shown later.
[0079] A number of methods are known for determining the inverse transfer function. For example, the method known as “equalization correction” includes determining the baseline of the spectrum V(n), for example by using a median filter or a parametric method. Figure 2Next, the vibration spectrum z(n) is normalized to obtain a whitened spectrum y(n)=s(n) / V(n), also known as a “flat spectrum y(n)”.
[0080] Figure 1 The original spectrum s(n) and the baseline spectrum V(n) are shown. Figure 2 The whitened spectrum y(n) is shown.
[0081] Step E3 of estimating a target signal x(n) comprising a set of characteristic spectral lines according to the whitened vibration spectrum y(n)
[0082] The determination of the characteristic lines by whitening the vibration spectrum y(n) is complex, especially in a blind manner. In this case, even in a blind determination, certain properties of the characteristic lines are still known.
[0083] In fact, the spectral lines can be identified from the noise in the whitened vibration spectrum y(n) because their values are usually greater than the noise. Therefore, this whitened vibration spectrum can be approximated as a set of carrier frequency families (base frequency and its multiples) and a set of modulation modules (carrier and its modulation).
[0084] like Figure 4 As shown, step E3 of estimating the target signal x(n) includes: determining a plurality of modulation modules BMb with a modulation index b in the whitened vibration spectrum y(n), the modulation index being between 1 and B.
[0085] refer to Figure 5 , each modulation module BM b include:
[0086] -Carrier coefficient x 0,b (x 0,b = b*fc), which is the predetermined carrier frequency fc and the modulation module BM b is a function of the exponential b, and
[0087] -Multiple with carrier coefficient x 0,b Symmetrical modulation coefficient.
[0088] Through the modulation module BM bModeling makes sense because gears are the mechanical elements that exhibit the most modulation phenomena. According to the Fourier series decomposition, the contact between the gear teeth is represented by the meshing frequency fc and its harmonics. Gears can have various defects, such as machining defects, uneven spacing defects between teeth, or irregularities in the contact surface of the teeth. During the operation of the gear, other defects may appear, such as defects where material is detached from the teeth. If the defect is related to an uneven hardness distribution along the tooth surface, a periodic amplitude modulation will appear. If the defect is related to a change in the meshing period (which may be related to a change in the tooth spacing), a phase modulation will appear. It is worth noting that amplitude modulation produces a completely symmetrical spectrum, which is different from phase modulation.
[0089] In general, due to the advantage of amplitude modulation, the set of lines is globally symmetric. Therefore, the definition includes multiple carrier coefficients x 0,b Modulation module BM with symmetrical modulation coefficient b is suitable.
[0090] Since the carrier coefficient x 0,b and modulation coefficients are represented by lines different from each other, so the modulation module BM b is considered "parsimonious". This allows avoiding results that are not constrained by structure, noise, etc.
[0091] Figure 2 The whitened vibration spectrum y(n) on a logarithmic scale is shown. Figure 3 yes Figure 2 Close-up view of the whitened vibration spectrum y(n) on a linear scale to show the modulation of the meshing frequency fc of the gears due to the defect frequency.
[0092] In the following, the whitened vibration spectrum y(n) is defined by the following formula:
[0093] y=Φ -1 x+W
[0094] in
[0095] -y is the whitened vibration spectrum
[0096] -Φ is the normalized Fourier transform and
[0097] - is the conjugate transpose of the matrix Φ, I N is the identity matrix of size N;
[0098] -x∈C N is the vector of coefficients of the “target signal x”;
[0099] -w is additive noise.
[0100] The Fourier coefficient x is defined according to the following formula: x = ΦS, where {s(n)} {n=1,...N} It is a periodically modulated vibration signal.
[0101] As mentioned before, the target signal x(n) comprises harmonics and modulation of the meshing frequency fc. The target signal x(n) represents the modulation in the spectral domain. Preferably, a Bayesian approach is used to estimate the target signal x(n). On the one hand, the approach enables the automatic estimation of point estimators and, on the other hand, the determination of the complete a posteriori probability. This enables the calculation of uncertainty deviations, as will be shown later.
[0102] In the following, the additive noise coefficients w are assumed to be independent and identically distributed Gaussian variables (iid) with unknown variance τ > 0. The distribution density of the observed values is given by the following formula:
[0103]
[0104] Then, the variance
[0105] [Mathematical formula 3]
[0106] τ
[0107] and the target signal x will be estimated together. Then, the variance
[0108] [Mathematical formula 3]
[0109] τ
[0110] and the target signal x through the non-informative prior Jeffrey's law
[0111] [Mathematical formula 3]
[0112] τ
[0113] [Mathematical formula 3]
[0114] p(τ)∝τ -1 Preferably, as will be shown later, the modulation module BM b In the form of a hierarchical model, it provides great flexibility in considering the structure of the parsimony coefficients with the help of latent variables.
[0115] Therefore, given that the carrier frequency fc is known, the target signal x can be modeled as a set of modulation modules BM whose positions are partially known b In fact, the modulation module BM b The center position of is known, which is equal to the carrier frequency fc multiplied by the modulation index b. Modulation module BM b The position of the modulation coefficient in is unknown.
[0116] Preferably, only some modulation coefficients around the carrier frequency fc and its integer multiple harmonics are non-zero, while other modulation coefficients can be regarded as zero. In addition, the positions of non-zero modulation coefficients around different harmonics of the carrier frequency fc remain almost the same.
[0117] Hereinafter, B is the number of harmonics in the carrier, i.e. the modulation module BM b Each modulation module BM b As mentioned above, it is assumed that the modulation module is symmetrical, that is, each modulation module BM b The left and right sides of the carrier coefficient centered at harmonic b*fc contain coefficients of the same number Q, such as Figure 5 shown.
[0118] Preferably, it is also assumed that the modulation module BM b are spaced from each other, i.e. do not overlap. This is the case with gears, since the carrier frequency fc is very large compared to the fundamental modulation frequency.
[0119] After this, A is all modulation modules BM b This means that we assume at most B(2Q+1) non-zero coefficients (the B harmonics of the carrier (carrier coefficients) and the 2Q non-zero coefficients around each carrier (modulation coefficients)).
[0120] As mentioned above, the number of non-zero modulation coefficients around each carrier is actually much less than 2Q. is the set of interference modulation coefficients. x k = 0. In the following, for all b∈{1,..,B}, we use X 0,b Indicates frequency bf c The carrier coefficients on the , for all q∈{1,...,Q], we use x -q,b and x qb represents the qth modulation coefficient located on the left and right of b*fc, such as Figure 5 shown.
[0121] Preferably, each modulation coefficient x q,b can be obtained by using an appropriate sparse matrix P q,b It is easy to obtain from the target signal x, that is, X q,b =P q,b X, the sparse matrix P q,b A permutation matrix containing the appropriate rows.
[0122] As mentioned above, the target signal x is modulated around the carrier frequency and its harmonics. In order to utilize each modulation module BM bThe method comprises the following steps: jointly estimating the symmetric modulation coefficient X -q,b and X q,b , that is, the modulation coefficient of the frequency group that is symmetrical to the bth harmonic of the carrier frequency b*fc. To this end, we define the vector X q,b ∈C 2 ,in,
[0123] X q,b =[x -q,b , x q,b ].
[0124] In order to fully utilize the modulation module BM b The similarity between them determines the modulation coefficient at the same position of the harmonics relative to the carrier.
[0125] To this end, all modulation modules BM at the same position b The coefficients of and their symmetries with respect to the harmonics of the carrier b*fc are stacked to construct a multi-block coefficient vector of size 2B.
[0126]
[0127] in
[0128] [Formula 4]
[0129]
[0130] is the transpose of a vector
[0131] [Formula 4]
[0132] X q,1
[0133] Also known as
[0134] [Formula 4]
[0135] X0
[0136] The vector of size B contains the coefficients of the harmonics of the carrier B (carrier coefficients). The parsimony of the spectrum and the dependence between the Fourier coefficients are obtained by assuming
[0137] [Formula 4]
[0138] (X q ) 1≤q≤Q
[0139] is a realization of a multivariate heavy-tailed distribution, which is considered to be a class of scaled mixtures of multivariate Gaussians (SMG).
[0140] -[Mathematical formula 5] xq ~CN(0,γ q Σ)
[0141] -in
[0142] [Formula 5]
[0143] CN
[0144] Here it refers to the multivariate complex normal distribution. The covariance matrix is real-valued.
[0145] Modulation coefficient X q It includes real and imaginary parts, which follow the mean of zero and the covariance matrix of τγ q Gaussian distribution of ∑.
[0146] γ q ≥0 is the mixing parameter according to a certain mixing distribution, and in represents the Kronecker product. I2 is the identity matrix of size 2×2, and
[0147] The vector of carrier coefficients is also a Gaussian prior, ie x0~CN(0,τγ0Diag(v)), where γ0>0. It should be noted that although the diagonal covariance matrix has been considered, X q The coefficients of are uncorrelated, but still dependent.
[0148] Variable v b acts as a global indicator of the total amplitude of the harmonics in the active module b, while the local variable γ q Controls parsimony within a module. If γ q If γ is too small, the spectral coefficients at position q and the spectral coefficients at symmetrical positions relative to the harmonics of the carrier on different activity groups B approach zero. q is very high, then in the case of non-zero v b Any modulation module BM b In , all coefficients at positions q and -q around the carrier harmonics are also non-zero, so the probability of being modulated is high. It can be seen that if almost all γ q are close to zero, then the modulation module BM b Simplicity within.
[0149] It should be noted that this model favours the structure of the spectrum by introducing a dependency between the coefficient amplitudes, which are related to their position relative to the harmonics of the carrier. This dependency can be explained as follows: coefficients located in the same or symmetrical position relative to their carrier frequency contribute almost equally to the total energy of the modulation module BMb in which they are located.
[0150] The definition of global and local parameters in a minimalistic modular model is non-trivial. In this example, we propose the following hierarchical model for these parameters:
[0151] v b ~IG(α b , β b )
[0152]
[0153] Among them, IG(α b , β b ) is a positive constant parameter α b and β b The inverse gamma distribution, C + (0,e) is a half-Cauchy distribution with site 0 and scale e.
[0154] This hierarchical model has many advantages. The global variable v b Potential parsimony in tuning modulation modules: small v b The appropriate value v may lead to over-shrunk activity coefficients, especially those with small magnitudes. Therefore, we prefer to define its prior to avoid spurious values v in case the activity group b is very sparse. b .
[0155] This is because we assume that the number of carrier harmonics is known, i.e., in each modulation block BM b There is at least one non-zero modulation coefficient in .
[0156] It is recommended to use half-Cauchy priorities for local and global parameters to improve shrinkage. However, during experiments it was found that half-Cauchy priorities for global parameters can make reconstruction of small lines worse compared to the inverse gamma distribution. The inverse gamma distribution is suitable given that its density vanishes at the origin, which artificially makes v b The value of is far from zero. On the other hand, it is the conjugate density of the Gaussian distribution, which will make subsequent calculations easier.
[0157] On the other hand, we prefer local variables v to global scales. b has imaginary values in order to properly model the ultra-minimalist signal. This is achieved by using the semi-Cauchy priority, which, in addition to having a tail similar to the Cauchy, has the advantage of having a smaller value in v compared to the inverse gamma. b = 0 is non-zero. This makes x q The priority has an infinitely high peak at the origin, providing good shrinkage properties.
[0158] Hyperparameter Φ q For mixed qPriorities provide another degree of freedom.
[0159] Finally, compared to the inverse gamma using the parameter expansion trick, the semi-Cauchy prior does not introduce any additional difficulties for posterior inference.
[0160] Using the detection method described in the present invention, the characteristic lines forming the target signal x(n) can be accurately estimated.
[0161] In fact, those skilled in the art know the kinematics of gears and can predetermine the signal sources that contribute to the measured vibration signal. For example, for planetary gears, it is necessary to determine the known frequencies of the various rotating components of the gears (sun gear, planetary gear, ring gear) and / or the (unknown) linear combinations of these known frequencies in the spectrum around the meshing frequency. These frequencies are referred to as "expected frequencies" hereinafter.
[0162] To define the size of the modulation module, it is advantageous to define an upper limit to these expected frequencies. This allows the size of the modulation module to be defined around the carrier frequency. After estimating the spectrum of the target signal, to monitor a specific rotating component (e.g., a planetary gear), it is only necessary to extract the energy of all sidebands related to the expected frequency of that specific rotating component.
[0163] A step (E4) of determining at least one health indicator (IND) from at least one spectral line feature of the target signal (x).
[0164] refer to Figure 1 After obtaining the target signal x(n), one or more health indicators IND can be calculated through one or more spectral line features of the target signal x(n). For example, the health indicator IND can be a sideband energy ratio (SER). It is defined as follows:
[0165]
[0166] This health indicator IND is interesting because it is normalized and can show the evolution of harmonics and modulation simultaneously. For example, in the presence of a defect, the energy of the carrier frequency fc may decrease, while the energy of the modulation increases.
[0167] If global monitoring is required (without exploiting kinematics), the maximum size of the modulation module can be set during the estimation phase.
[0168] Then, the global indicator SERG can be calculated based on the obtained signal.
[0169]
[0170] The global metric is noisier because all signal sources are included, but it can still detect defects because the defect increases the total energy of the modulation. However, it cannot be used to locate defects. To locate defects, the frequencies of the various rotating components need to be known.
[0171] Comparing at least one health indicator (IND) with an indicator database (DB(IND)) to determine the defect (DEF) Step (E5).
[0172] Then, the health indicator IND is compared with the indicator database DB(IND) to determine the defect DEF. For SER type health indicators, the indicator database DB(IND) includes health thresholds, etc. It goes without saying that the indicator database DB(IND) may also include reference health indicators.
[0173] Using the method described in the present invention, the target signal x(n) is reliably estimated by a blind method in order to reactively detect gear defects and thus avoid any malfunctions.
[0174] Then some experimental results are given.
[0175] Experiment 1
[0176] In the first experiment, the vibration signal s(n) is a signal generated by simulation and has the following characteristics:
[0177] -N=5,000,
[0178] -B=6.
[0179] -f s =5000, fs is the sampling frequency;
[0180] - 6 modulations,
[0181] -f c =350 / f s , f m =25 / f s .
[0182] The vibration signal s(n) is the product of two periodic signals (s = s c xs m ) is generated and Gaussian noise is added. The whitened vibration spectrum y(n) obtained from the vibration signal s(n) is as follows: Figure 6 shown.
[0183] During the step of estimating the target signal x(n), 6 modulation modules BM1-BM6 are identified (gray vertical lines), each of which consists of a frequency b*fc (carrier coefficient), an amplitude and a phase modulation (modulation coefficient) with different amplitudes. As shown above, this structure is symmetrical in position on the frequency axis, but asymmetrical in amplitude. In this first experiment, the modulation modules BM1-BM6 were detected in the absence of noise with a very low MSE (mean square error), of the order of 0.04. Advantageously, given that it is non-simple and asymmetric in frequency, there is no need to take into account Gaussian noise when reconstructing the modulation. This provides an important gain for the formation of the relevant health indicator IND.
[0184] refer to Figure 7 , which shows a close-up of the whitened vibration spectrum y(n) at the level of the second modulation module BM2, we can see that there are two harmonic noises BR1, BR2 at different frequencies. The blind determination method proposed by the present invention can resist these harmonic noises BR1, BR2. By considering the inter-block symmetry (modulation coefficients within a single modulation module) and the intra-block symmetry (same symmetry between multiple modulation modules), non-modulation lines such as harmonic noises BR1 and BR2 can be excluded.
[0185] Experiment 2
[0186] In the second experiment, reference Figure 8 The vibration signal s(n) is a signal resulting from actual vibration measurements carried out on an aircraft turbine including a planetary gear. This type of gear has a modulation in its healthy state, which is amplified or supplemented by a new modulation in the event of a defect. Figure 8 An estimated modulation model around the 8th order meshing frequency fc is shown.
[0187] Fig. 9 A close-up view of the modulation of the sun gear pinion, planetary gear pinion, and low pressure shaft being estimated correctly. The modulation is well estimated, where the noise is severely attenuated when no modulation estimation is done. The circles are added to show that even though there are peaks within the modulation module, they are not recovered because they do not have a symmetrical structure and are therefore not modulation coefficients.
[0188] Experiment 3
[0189] In the third experiment, the vibration signal s(n) is sampled at a frequency f s = 3 seconds of signal from a spur gear collected at 20kHz. In this example, the carrier frequency of the vibration signal s(n) is f c = 388 Hz. The vibration signal s(n) is measured on different days, especially on the first day ( Fig.10 ) and Day 12 ( Fig.11) is measured.
[0190] When we compare Day 1 ( Fig.10 ) and Day 12 ( Fig.11 ) when we change between Fig.12 (Day 1) and Fig.13 (Day 12) shows a significant amplification and multiplication of the harmonics. When the modulation of each harmonic is about 2f c When the detection method described in the present invention is used, these modulations can be perfectly determined.
[0191] During the 12 days, the health index IND is calculated every day. In this example, the health index IND is based on the energy (frequency) of the modulation. Fig.14 It represents the change of the health index IND over time. Therefore, we can determine that the gear was defect-free until a moderate defect (one tooth was broken) appeared on the 8th day, and a serious defect (one tooth was partially detached) appeared on the 11th day.
[0192] In this way, the determined quality of the target signal x(n) enables the formation of health indicators IND which are accurate and therefore highly relevant for the reactive detection of defects.
Claims
1. A method for detecting a defect (DEF) in at least one gear of an aircraft turbomachine based on a vibration signal (s(n)) of the gear, each gear having a determined meshing frequency (fc) corresponding to a carrier frequency, the method comprising: - a step (E2a) of determining a vibration spectrum (z(n)) from said vibration signal (s(n)), - a step (E2b) of whitening said vibration spectrum (z(n)) to obtain a whitened vibration spectrum (y(n)), - a step (E3) of estimating a target signal (x(n)) comprising a set of characteristic spectral lines based on said whitened vibration spectrum (y(n)), - a step (E4) of determining at least one health indicator (IND) by means of at least one spectral feature of said target signal (x(n)), - a step (E5) of comparing at least one health indicator (IND) with a database of indicators (DB(IND)) in order to detect defects (DEF); -characterized in that the step of estimating the target signal (x(n)) comprises: ○ Determine a plurality of modulation modules (BM) with modulation index (b) in the whitened vibration spectrum (y(n)) b ), ○ Each modulation module (BM b )include: Carrier coefficient (x 0,b ), which is a function of a predetermined carrier frequency (fc) and said modulation index (b), and · Multiple carrier coefficients (x 0,b )Symmetrical modulation coefficient.
2. The method according to claim 1, characterized in that The target signal x(n) is estimated by a Bayesian method.
3. The method according to claim 1 or 2, characterized in that Each modulation module (BM b ) are different from each other.
4. The method according to claim 1, characterized in that The symmetric modulation coefficients are estimated jointly.
5. The method according to claim 1, characterized in that For each modulation module (BM b ), the modulation coefficients of each symmetry are relative to the carrier coefficient (x 0,b )The same position is estimated.
6. The method according to claim 1, characterized in that Each symmetric modulation coefficient is considered to be a realization of a multivariate heavy-tailed distribution.
7. The method according to any one of claims 1 to 6, characterized in that The whitened vibration spectrum (γ(n)) is defined by the following formula: y=Φ -1 x+w -in: -Φ is the Fourier transform is the conjugate transpose of the matrix Φ, I N is the identity matrix of size N; -x∈C N is the vector of unknown Fourier coefficients of the target signal x(n); -w is additive noise.
8. The method according to claim 1, characterized in that The modulation coefficient includes a real part and an imaginary part, the real part and the imaginary part follow a mean of zero and a covariance matrix of τγ q ∑ is a Gaussian distribution, where -γ q ≥0 is the mixing parameter according to a certain mixing distribution; - in is the Kronecker product, I2 is the identity matrix of size 2×2 and v = [v1, ..., v B ] T ∈(R + ) B ..
9. The method according to claim 7 combined with claim 8, characterized in that: The modulation module (BM b ) is in the form of the following hierarchical model - where IG(α b ,β b ) is a positive constant parameter α b and β b The inverse gamma distribution, C + (0,e) is a half-Cauchy distribution with site 0 and scale e.
10. The method according to claim 8 or 9, characterized in that The vector of the carrier coefficients is determined by Gaussian prior, such as x0~CN(0,τγ0Diag(v)), where γ0>0.
