Fault Diagnosis Method for Wind Turbine Drive Train Based on Expectation Maximization in Time-Frequency Plane

Abstract

The present invention proposes a fault diagnosis method for the drive train of a wind turbine based on time-frequency plane expectation maximization, belonging to the technical field of fault diagnosis; the method includes the following steps: configuring a cylindrical MEMS acceleration sensor on the drive train of the wind turbine, and collecting the vibration signal of the fan drive train through the cylindrical MEMS acceleration sensor; adopting an unsupervised classification method based on the zero points of the spectrogram, collecting the vibration signal and the rotational speed signal during the operation of the wind turbine, generating a time-frequency representation of the vibration signal by using the short-time Fourier transform, extracting the zero points of the spectrogram and performing unsupervised classification to realize the noise reduction processing of the vibration signal; based on the multi-component signal estimation method under time-frequency plane expectation maximization, accurately estimating the instantaneous frequency and the instantaneous amplitude of the multi-component signal on the time domain plane of the denoised signal; performing order spectrum analysis to identify the vibration signal characteristics of the drive train of the wind turbine and realizing the fault diagnosis of the drive train of the wind turbine under variable rotational speed.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and particularly to a fault diagnosis method for a wind turbine drive train based on expectation maximization in the time-frequency plane. Background Art

[0002] Wind turbines are important devices in modern renewable energy utilization and are widely used in wind farms around the world. With the increasing global demand for clean energy, the number and scale of wind turbines are also growing rapidly. Wind turbines are usually installed in high-altitude and harsh environments, which makes the reliability and lifespan of the key components of the wind turbine drive train particularly important. Under complex operating conditions such as variable speed, the blades of wind turbines are subjected to short and frequent impacts, causing the drive train to be subjected to alternating loads, which greatly affects its working lifespan. Therefore, effectively monitoring the operating state of the wind turbine drive train can improve operating reliability and prevent major safety accidents. The wind turbine drive train mainly consists of rotating components such as the main shaft and support bearings (direct drive type), speed-increasing gearboxes (doubly-fed type), etc. Currently, there are two prominent problems in the fault diagnosis of wind turbine drive trains: 1) Due to continuous changes such as speed, the fault characteristics have typical time-varying and non-stationary properties; 2) Due to the relatively low main shaft speed, the fault characteristic frequencies of the key components on the drive train are in a relatively low frequency band, and at the same time, weak fault signals are easily submerged by strong noise, making it difficult to identify early weak faults and extract variable speed and weak fault characteristics.

[0003] Therefore, it is very necessary to provide a fault diagnosis method for a wind turbine drive train based on expectation maximization in the time-frequency plane, collect the vibration signals of the wind turbine drive train under variable speed conditions, accurately characterize the service state of the key components of the wind turbine drive train, accurately obtain the instantaneous amplitude and instantaneous frequency of different component signals, and achieve early detection and fault diagnosis of wind turbine drive train faults. Summary of the Invention

[0004] In view of this, the present invention proposes a fault diagnosis method for a wind turbine drive train based on expectation maximization in the time-frequency plane, which uses a cylindrical MEMS sensor to collect the vibration signals of the wind turbine drive train under variable speed conditions, extracts the zero points of the spectrogram and performs unsupervised classification to achieve denoising preprocessing of the signals, and accurately obtains the instantaneous amplitude and instantaneous frequency of different component signals.

[0005] The present invention provides a fault diagnosis method for a wind turbine drive train based on expectation maximization in the time-frequency plane, including the following steps:

[0006] S1; Configure a cylindrical MEMS acceleration sensor on the drive train of the wind turbine, and collect the vibration signals of the wind turbine drive train through the cylindrical MEMS acceleration sensor;

[0007] S2: Adopt an unsupervised classification method based on the zero points of the spectrogram. Collect the vibration signal and rotational speed signal during the operation of the wind turbine. Use the short-time Fourier transform to generate the time-frequency representation of the vibration signal, extract the zero points of the spectrogram and perform unsupervised classification to achieve noise reduction processing of the vibration signal;

[0008] S3: Based on the multi-component signal estimation method under the expectation maximization in the time-frequency plane, accurately estimate the instantaneous frequency and instantaneous amplitude of the multi-component signal on the time domain plane of the denoised vibration signal;

[0009] S4: Conduct order spectrum analysis, identify the vibration signal characteristics of the wind turbine drive train, and achieve fault diagnosis of the wind turbine drive train under variable rotational speed.

[0010] Based on the above technical solutions, preferably, the cylindrical MEMS acceleration sensor includes a cylindrical base, a mass block and a plurality of electrode plates. The surface of the base is provided with a sliding groove, and the sliding groove extends along the axial direction of the base; the mass block is embedded in the sliding groove and is slidably connected to the inner surface of the sliding groove; both ends of the mass block are respectively provided with anchor points, one end of the anchor point is fixedly connected to the mass block, and the other end of the anchor point extends away from the mass block, and the anchor point fits on the surface of the base and is slidably connected to the base; a plurality of comb teeth are symmetrically arranged at non-end positions on both sides of the mass block, and the plurality of comb teeth are arranged at intervals along the axial direction of the base. One end of the plurality of comb teeth is fixedly connected to the mass block, and the other end of the plurality of comb teeth extends away from the mass block, and the comb teeth are slidably connected to the surface of the base. There are intervals between adjacent comb teeth and between the comb teeth and the anchor points; a plurality of electrode plates are fixedly arranged on the surface of the base between adjacent comb teeth, and the position of the electrode plates relative to the base remains unchanged; when the wind turbine drive train generates acceleration, the mass block drives the anchor points and a plurality of comb teeth to slide along the axis of the base, changing the distance between the comb teeth and the adjacent electrode plates, thereby generating a current signal, sampling the magnitude of the current signal and converting it into the vibration signal of the wind turbine drive train under variable rotational speed conditions.

[0011] Preferably, the specific content of the noise reduction processing of the vibration signal in step S2 includes:

[0012] Convert the vibration signal of the wind turbine drive train into a time-frequency representation, use a Gaussian window function to segment the vibration signal, the window slides on the vibration signal, obtain the spectral information corresponding to each time window, and combine the spectral information corresponding to each window to obtain the time-frequency spectrogram of the entire vibration signal;

[0013] Find the zero points in the time-frequency spectrogram, and classify the zero points into SS class, SN class and NN class according to the properties of the zero points;

[0014] After completing the zero - point classification, in the time - frequency plane, select the SS - type and SN - type zero points for Delaunay triangulation, obtain the estimation of the signal domain in the time - frequency plane, and use the inverse short - time Fourier transform to reconstruct the vibration signal within the signal domain estimation to obtain the estimated value of the vibration signal without noise.

[0015] Further preferably, find the zero points in the time - frequency spectrogram, classify the zero points according to their properties. Set a threshold γ, and consider the points whose spectrogram values in the time - frequency spectrogram are lower than the threshold as zero points. The determination method of the threshold γ is: γ = μ X -k 0 σ X , where μ X is the mean value of the spectrogram values, σ X is the standard deviation of the spectrogram values, and k 0 is an adjustment factor; according to the properties of the zero points, the zero points are divided into three categories: 1) SS - type: signal - signal zero points; 2) SN - type: signal - noise zero points; 3) NN - type: noise - noise zero points; Assume that N noise simulations are performed to obtain N time - frequency spectrograms. Define the position of zero point i in different noise simulations as (t Ki , f Ki ), K = 1, 2,..., N. For each zero point i, calculate the position change amount d i of the zero point in different noise simulations: where collectively represent the average position of zero point i in the N - th noise simulation. Take δ SS and δ NN as the thresholds of the position change amount d i . δ SS is the right offset value covering the first peak of the position change amount of the signal - signal zero points, and δ NN is the left offset value covering the second peak of the position change amount of the noise - noise zero points. The zero - point classification rule is: SS - type = {i|d i ≤δ SS}, NN - type = {i|d i ≥δ NN}, and SN - type = {i|δ SS ≤d i ≤δ NN}.

[0016] Even more preferably, the value of the adjustment factor k 0 is 2.5.

[0017] More preferably, for the estimation of the signal domain in the time-frequency plane, the inverse short-time Fourier transform is used to reconstruct the vibration signal in the signal domain estimation to obtain an estimated value of the noise-free vibration signal. Specifically, after completing the zero-point classification, on the time-frequency plane, the SS class and SN class zero points are selected for Delaunay triangulation. The distribution of the zero points and the set characteristics of the corresponding triangles are used to identify the signal domain. The side length l of the triangulation ij The calculation formula is: i and j represent two different zero points, (t i , f i ) and (t j , f j ) are the positions of the two zero points respectively; a weight factor ω ij is introduced to make the triangulation consider both the density and geometric distance of the zero points. where ρ(t, f) is the local density of the zero points on the time-frequency plane, and the weighted side length l′ of each side of the triangulation is obtained through the weight factor ij , l′ ij = ω ij ·l ij , dynamically adjusting the effective length of each side; for each triangle of the triangulation, if the weighted side length of any side is greater than the preset side length threshold then the triangle is considered to belong to the signal domain; the triangles with side lengths greater than the side length threshold are clustered together to obtain the estimation of the signal domain in the time-frequency plane: Based on the obtained estimation of the signal domain in the time-frequency plane, the inverse short-time Fourier transform is used to reconstruct the signal in the signal domain, and the estimated value of the noise-free signal is calculated

[0018] Even more preferably, the content of step S3 is:

[0019] S31. Expectation step: Calculate the expected value of the vibration signal. The noise-free signal estimation value with multiple components is expressed as: k = 1, 2,..., m, A k (t) is the instantaneous amplitude of the k-th component, φ k (t) is the instantaneous phase, and n(t) is the noise; then the Monte Carlo sampling is used to estimate the posterior distribution, and M samples θ (i) are generated from the prior distribution. Here, the superscript i represents different moments. Assuming that the observed signal is independent at each time point t, for each sample θ (i) , calculate the likelihood function value of the estimated value of the noise-free signal t = 1, 2, 3,..., T, and then calculate the weight of each sample Then, the weighted average method is used to estimate the expected value of the posterior distribution:

[0020] S32. Maximization step: After obtaining the expected value that conforms to the characteristics of the actual vibration signal , the estimated values of the instantaneous frequency and amplitude need to be updated; first, substitute the expected value of the vibration signal characteristics obtained above, maximize the expected log-likelihood function, and find the optimal instantaneous frequency and amplitude: where θ (t+1) represents the parameter estimation after the (t + 1)-th iteration. Using the result after maximizing the expectation, update the parameters of the instantaneous frequency and amplitude:

[0021]

[0022] where is the instantaneous phase after the (t + 1)-th iteration, is the instantaneous amplitude of the k components after the t-th iteration, and η is the learning rate;

[0023] S33. To ensure the stability and accuracy of the parameter estimation, the expectation step S31 and the maximization step S32 need to be repeated until the parameter estimation converges. The convergence condition of the parameter estimation is |θ (t+1) -θ (t) | < ε, where θ (t) is the parameter estimation after the t-th iteration, and ε is the convergence threshold; finally, extract the instantaneous frequency and the instantaneous amplitude of each signal component as the input conditions for the next-order spectral analysis.

[0024] Further preferably, the content of step S4 is: After completing the extraction of the instantaneous frequency and instantaneous amplitude of the signal components of the vibration signal, obtain the instantaneous frequencies f(t) and f F (t) of the signal feature classification and the rotational speed feature classification, and use the instantaneous frequencies f(t) and f F (t) to extract the features of the variable rotational speed signal. According to the result of the variable rotational speed signal feature extraction, realize the fault diagnosis of the wind turbine drive train.

[0025] Even further preferably, using the instantaneous frequencies f(t) and f F (t) to extract the features of the variable rotational speed signal adopts the order ratio analysis method. Define the order ratio L as a multiple of the rotational speed, and characterize different fault types and fault locations of the wind turbine drive train according to the magnitude of the order ratio.

[0026] Preferably, the gap between adjacent comb teeth is greater than the sliding distance of the mass block in the chute.

[0027] The fault diagnosis method for the drive train of a wind turbine based on expectation maximization in the time-frequency plane provided by the present invention has the following beneficial effects compared with the prior art:

[0028] (1) By collecting the vibration signals of the drive train of the wind turbine under variable speed conditions, the present invention accurately characterizes the service states of the key components of the wind turbine drive train, providing effective data support for subsequent time-frequency analysis. Secondly, a method for unsupervised classification based on the zero points of the spectrogram is proposed. By collecting the vibration and speed signals during the operation of the wind turbine, a time-frequency representation is generated using the short-time Fourier transform, the zero points of the spectrogram are extracted and subjected to unsupervised classification to achieve denoising preprocessing of the signals. Then, a method for estimating multi-component signals under the time-frequency plane expectation maximization algorithm is proposed to accurately obtain the instantaneous amplitude and instantaneous frequency of different component signals, realizing a high-resolution time-frequency representation of the time-frequency characteristics of multi-component vibration signals under non-stationary working conditions. Finally, the time-frequency ridges of the rotation frequency and fault characteristic components are extracted from the highly readable time-frequency plane obtained from the above time-frequency analysis method, and the means of order spectrum analysis is used to realize the early detection and fault diagnosis of the faults of the wind turbine drive train;

[0029] (2) A structure of a cylindrical MEMS sensor is provided. Compared with the planar structure of the existing sensor, it can limit the degrees of freedom of the mass block so that it can only move along the axial direction of the cylinder, restricting the movement of the mass block in other directions, thereby eliminating measurement interference. Multiple circular ring-shaped electrodes and circular ring-shaped comb teeth attached to the surface of the cylinder, and the comb teeth move up and down on the cylinder surface following the mass block, ensuring that the capacitance changes only due to the change in the pole distance, while avoiding the influence of the change in the covered area on the measurement, ensuring the accuracy and stability of the measurement. Description of the Drawings

[0030] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention, and those of ordinary skill in the art can also obtain other drawings based on these drawings without creative efforts.

[0031] Figure 1 It is the flow chart of the steps of the fault diagnosis method for the drive train of a wind turbine based on expectation maximization in the time-frequency plane of the present invention;

[0032] Figure 2 It is the three-dimensional view of the cylindrical MEMS acceleration sensor of the fault diagnosis method for the drive train of a wind turbine based on expectation maximization in the time-frequency plane of the present invention;

[0033] Figure 3Stereogram of the base of the cylindrical MEMS acceleration sensor for the fault diagnosis method of the wind turbine drive train based on time-frequency plane expectation maximization of the present invention;

[0034] Figure 4 Stereogram of the combined state of the mass block, anchor points and comb teeth of the cylindrical MEMS acceleration sensor for the fault diagnosis method of the wind turbine drive train based on time-frequency plane expectation maximization of the present invention;

[0035] Figure 5 Stereogram of the electrode plate of the cylindrical MEMS acceleration sensor for the fault diagnosis method of the wind turbine drive train based on time-frequency plane expectation maximization of the present invention. Detailed implementation manners

[0036] To make the objectives, technical solutions and advantages of the embodiments of the present disclosure clearer, the technical solutions of the embodiments of the present disclosure will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present disclosure. Obviously, the described embodiments are some but not all of the embodiments of the present disclosure. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the scope of protection of the present invention.

[0037] The existing methods for detecting the wind turbine drive train have the following disadvantages: 1. When using a planar MEMS accelerometer to measure vibration signals, it is easily interfered by gravity and centrifugal force when installed on rotating equipment, resulting in a decrease in measurement accuracy; in the face of complex vibration modes, it is unable to effectively distinguish signals from different vibration sources, especially when the frequencies of vibration signals are close or overlapping, which is prone to misjudgment. 2. In signal denoising processing, the wavelet denoising algorithm based on threshold is one of the currently widely used methods, but the wavelet denoising algorithm based on threshold has the following disadvantages: the wavelet transform itself depends on the pre-selected wavelet basis, and these wavelet bases do not always perfectly match the characteristics of the signal, resulting in the loss of signal details or the generation of artifacts. 3. Synchronous compression transform rearranges the time-frequency map obtained by the traditional continuous wavelet transform or short-time Fourier transform, and concentrates the signal energy at the position of the instantaneous frequency, but in essence, it belongs to a post-processing method of time-frequency analysis, and it is sensitive to noise. Especially when the frequency components of the signal are dense or overlapping, the noise will cause great interference to the rearrangement process, resulting in an increase in estimation error. Especially when dealing with large-scale data or real-time signals, it is not suitable for industrial applications. In view of this, as Figure 1 shown, the present invention provides a fault diagnosis method for the wind turbine drive train based on time-frequency plane expectation maximization, which specifically includes the following steps:

[0038] S1; Configure a cylindrical MEMS acceleration sensor on the drive train of the wind turbine, and collect the vibration signals of the fan drive train through the cylindrical MEMS acceleration sensor.

[0039] As shown Figures 2 - 5 in the figure, the cylindrical MEMS acceleration sensor includes a cylindrical base 1, a mass block 2 and a plurality of electrode plates 3. A chute is provided on the surface of the base 1, and the chute extends along the axial direction of the base 1; the mass block 2 is embedded in the chute and is slidably connected to the inner surface of the chute; anchor points 2.1 are respectively arranged at both ends of the mass block 2, one end of the anchor point 2.1 is fixedly connected to the mass block 2, and the other end of the anchor point 2.1 extends away from the mass block 2, and the anchor point 2.1 is attached to the surface of the base 1 and is slidably connected to the base 1; a plurality of comb teeth 2.2 are symmetrically arranged at non-end positions on both sides of the mass block 2, and the plurality of comb teeth 2.2 are arranged at intervals along the axial direction of the base 1. One end of the plurality of comb teeth 2.2 is fixedly connected to the mass block 2, and the other end of the plurality of comb teeth 2.2 extends away from the mass block 2, and the comb teeth 2.2 are slidably connected to the surface of the base 1. There are intervals between adjacent comb teeth 2.2 and between the comb teeth 2.2 and the anchor point 2.1; a plurality of electrode plates 3 are fixedly arranged on the surface of the base 1 between adjacent comb teeth 2.2, and the position of the electrode plate 3 relative to the base 1 remains unchanged; when the wind turbine drive chain generates acceleration, the mass block 2 drives the anchor points 2.1 and a plurality of comb teeth 2.2 to slide along the axis of the base 1, changing the distance between the comb teeth 2.2 and the adjacent electrode plates 3, thereby generating a current signal. The magnitude of the sampled current signal is sampled and converted into a vibration signal of the wind turbine drive chain under variable speed conditions. The main function of the electrode plate 3 is to cooperate with the comb teeth on the mass block to facilitate the sensing of the displacement of the mass block. Since the distance between the comb teeth and the adjacent electrode plate 3 can change, the capacitance between the electrodes in the output also changes accordingly, thereby generating a detection signal corresponding to the acceleration generated by the vibration. In order to prevent the comb teeth 2.2 from overlapping with the electrode plate 3 in position, the gap between adjacent comb teeth 2.2 can be set to be greater than the sliding distance of the mass block 2 in the chute.

[0040] In the drive chain of the wind turbine of this embodiment, the bearing is a key part for bearing load and vibration, and the gearbox is the core component for transmitting power. Therefore, the sensor is installed above the outer shell of the gearbox bearing seat to ensure that the axial direction of the sensor is aligned with the expected vibration direction, so as to accurately measure the vibration in the up and down directions. This can effectively monitor the vibration state of the wind turbine drive chain and identify potential faults. As shown Figures 2 - 5 in the attached figure, both the comb teeth 2.2 and the electrode plate 3 are arc-shaped structures attached to the base 1. The chute has a wedge-shaped cross-sectional shape, which restricts the movement of the mass block in other directions, thereby reducing the measurement error caused by lateral movement. The sensing area can be greatly expanded through the comb tooth structure, the measurement accuracy can be improved, and the difficulty of signal processing can be reduced.

[0041] S2: Adopt an unsupervised classification method based on spectrogram zeros. Collect the vibration signal and rotational speed signal during the operation of the wind turbine. Use the short-time Fourier transform to generate the time-frequency representation of the vibration signal, extract the spectrogram zeros and perform unsupervised classification to achieve noise reduction processing of the vibration signal.

[0042] The specific content of step S2 includes:

[0043] Convert the vibration signal of the wind turbine drive train into a time-frequency representation. Use a Gaussian window function to segment the vibration signal. The window slides on the vibration signal to obtain the spectral information corresponding to each time window. Combine the spectral information corresponding to each window to obtain the time-frequency spectrogram of the entire vibration signal. This step uses the short-time Fourier transform (STFT) to generate the time-frequency (TF) representation of the vibration signal x(t) measured by the cylindrical MEMS accelerometer, that is, X(t, f) = STFT(x(t)). Through this method, a one-dimensional time signal can be converted into a two-dimensional time-frequency spectrogram, so as to observe the variation law of the frequency of the measured vibration signal with time. Next, use a Gaussian window function to segment the signal. In the short-time Fourier transform, the window function slides on the signal, and only the signal within a short time window is processed each time, and then the Fourier transform is performed on these short time segments. In this way, the spectral information corresponding to each time window can be obtained, and combining them forms the time-frequency spectrogram of the entire signal.

[0044] Find the zeros in the time-frequency spectrogram. According to the properties of the zeros, classify the zeros into SS class, SN class, and NN class. Specifically, set a threshold γ. Any point in the time-frequency spectrogram whose spectrogram value is lower than the threshold is regarded as a zero. The determination method of the threshold γ is: γ = μ X -k 0 σ X , where μ X is the mean of the spectrogram values, σ X is the standard deviation of the spectrogram values, and k 0 is an adjustment factor. According to the properties of the zeros, the zeros are divided into three categories: 1) SS class: signal-signal zero; generated by the interference between signal components, usually the most stable. 2) SN class: signal-noise zero; generated by the interference between the signal and noise, and its stability is between the SS class and the NN class. 3) NN class: noise-noise zero; only generated by noise, the most unstable.

[0045] Assume that N noise simulations are performed to obtain N time-frequency spectrograms. Define the position of zero i in different noise simulations as (t Ki , f Ki ), K = 1, 2,..., N. For each zero i, calculate the position change amount d of the zero in different noise simulations i: where together represent the average position of the zero point i in the Nth noise simulation, and δ SS and δ NN are taken as the thresholds of the position change amount d i , δ SS is the right offset value of the first peak covering the position change amount of most signal-signal zero points, δ NN is the left offset value of the second peak covering the position change amount of most noise-noise zero points. The rule for zero point classification is: SS class = {i|d i ≤δ SS}, NN class = {i|d i ≥δ NN}, and SN class = {i|δ SS ≤d i ≤δ NN}.

[0046] Through cross-validation, this embodiment provides that the value of the adjustment factor k 0 is 2.5.

[0047] After completing zero point classification, on the time-frequency plane, select the SS class and SN class zero points for Delaunay triangulation to obtain an estimate of the signal domain in the time-frequency plane, and use the inverse short-time Fourier transform to reconstruct the vibration signal within the signal domain estimate to obtain an estimated value of the vibration signal without noise. Specifically, after completing zero point classification, on the time-frequency plane, select the SS class and SN class zero points for Delaunay triangulation, and ignore the NN class zero points to reduce the influence of noise. The distribution of zero points and the set characteristics of corresponding triangles are used to identify the signal domain. The calculation formula for the side length l ij of the triangulation is: i and j represent two different zero points, (t i , f i ) and (t j , f j ) are the positions of the two zero points respectively; introduce a weight factor ω ij to make the triangulation consider both the density and geometric distance of zero points, where ρ(t, f) is the local density of zero points on the time-frequency plane. The weighted side length l′ ij of each side of the triangulation is obtained through the weight factor, and l′ ij =ω ij ·l ij , dynamically adjusting the effective length of each side; for each triangle of the triangulation, if the weighted side length of any side is greater than the preset side length threshold then the triangle is considered to belong to the signal domain; the side length greater than the side length threshold Triangles are aggregated to obtain an estimate of the signal domain in the time-frequency plane: Based on the estimated signal domain in the obtained time-frequency plane, the signal within the signal domain is reconstructed using the inverse short-time Fourier transform (ISTFT) to calculate the estimated value of the noise-free signal

[0048] Traditional methods may not be sufficient to effectively distinguish signals from noise in some cases, especially when signal components are close or the noise is strong. By introducing a weight factor, the present invention can more accurately reflect the actual importance and distribution characteristics of zeros, thereby improving the denoising effect. Introducing the weight factor makes the triangulation not only depend on the geometric distance but also take into account the density of zeros. Since regions with high density may contain more noise zeros, while regions with low density are more likely to be the signal domain. In traditional methods, regions with large noise are easily misjudged as the signal domain. By considering the local density of zeros and introducing the weight factor, such misjudgments can be effectively reduced, improving the overall denoising effect. This improvement makes the algorithm more robust when dealing with dense noise regions and reduces misjudgments. This method effectively utilizes the stability characteristics of zeros to accurately distinguish signals from noise and achieves the purpose of signal denoising.

[0049] S3: Based on the multi-component signal estimation method under the expectation maximization of the time-frequency plane, accurately estimate the instantaneous frequency and instantaneous amplitude of the multi-component signal on the time domain plane of the denoised vibration signal.

[0050] After the above denoising preprocessing, the noise in the wind turbine vibration signal is effectively suppressed. However, in order to further accurately extract the instantaneous frequency and instantaneous amplitude of different signal components (such as rotational frequency and fault characteristic frequency) in the vibration signal of the wind turbine drive train under variable speed, further processing is required, that is, to extract the time-frequency ridges corresponding to the rotational frequency and fault characteristic frequency in the two-dimensional time-frequency plane. This method uses the method based on the expectation maximization algorithm to perform deeper processing on the denoised signal to accurately estimate the instantaneous frequency and instantaneous amplitude of the multi-component signal, thus laying a foundation for the next-order spectrum analysis under variable speed.

[0051] The specific content of step S3 is as follows:

[0052] S31. Expectation step: Calculate the expected value of the vibration signal, and represent the estimated value of the noise-free signal with multiple components as: k = 1, 2,..., m, A k (t) is the instantaneous amplitude of the k-th component, φ k (t) is the instantaneous phase, and n(t) is the noise; then use Monte Carlo sampling to estimate the posterior distribution and generate M samples θ from the prior distribution (i), where the superscript i represents different moments. It is assumed that the observed signal is independent at each time point t. For each sample θ (i) , calculate the estimated value of the noise-free signal of the likelihood function value t = 1, 2, 3,..., T, and then calculate the weight of each sample Then use the method of weighted average to estimate the expected value of the posterior distribution:

[0053] S32, Maximization step: After obtaining the expected value that conforms to the characteristics of the actual vibration signal, it is necessary to update the estimated values of the instantaneous frequency and amplitude. First, substitute the expected value of the vibration signal characteristics obtained above, maximize the expected log-likelihood function, and find the optimal instantaneous frequency and amplitude: where θ (t+1) represents the parameter estimation after the (t + 1)-th iteration. Using the result after maximizing the expectation, update the parameters of the instantaneous frequency and amplitude:

[0054]

[0055] where is the instantaneous phase after the (t + 1)-th iteration, is the instantaneous amplitude of the k components after the t-th iteration, η is the learning rate;

[0056] S33. To ensure the stability and accuracy of the parameter estimation, it is necessary to repeat the expectation step S31 and the maximization step S32 until the parameter estimation converges. The convergence condition of the parameter estimation is |θ (t+1) - θ (t) | < ε, θ (t) is the parameter estimation after the t-th iteration, and ε is the convergence threshold; finally, extract the instantaneous frequency and the instantaneous amplitude of each signal component as the input conditions for the next order spectrum analysis.

[0057] S4: Conduct order spectrum analysis to identify the vibration signal characteristics of the wind turbine drive train and achieve fault diagnosis of the wind turbine drive train under variable speed.

[0058] The content of step S4 is: After completing the extraction of the instantaneous frequency and instantaneous amplitude of the signal components of the vibration signal, obtain the instantaneous frequencies f(t) and f F (t) of the signal feature classification and the rotational speed feature classification, and use the instantaneous frequencies f(t) and f F(t)Extract the characteristics of the variable-speed signal, and realize the fault diagnosis of the wind turbine drive train according to the results of the variable-speed signal feature extraction. If the speed of the device is not uniform or fluctuates greatly, various characteristic signals will not appear at equal intervals, and the traditional spectrum analysis method based on Fourier transform is no longer applicable. For this situation, the order ratio analysis method shows greater advantages. The order ratio L is defined as a multiple of the rotational speed, that is According to the magnitude of the order ratio, different fault types and fault positions of the wind turbine drive train are characterized. The order ratio L = 1 means that the signal occurrence frequency is the same as the rotational frequency; the order ratio L = 2 means that the signal occurrence frequency is twice the rotational frequency. It can be seen that the order ratio is actually a method of frequency representation, equivalent to a multiple of the rotational frequency. For rotating machinery such as wind turbines, its vibration signal is closely related to the rotational speed. Using the order ratio can better represent this corresponding relationship, and the influence of the rotational speed change can be eliminated by using the order ratio. For example, by calculating that a specific order ratio appears, such as an integer multiple of 0.6×Z corresponding to an inner ring fault and an integer multiple of 0.4×Z corresponding to an outer ring fault, where Z represents the number of rolling elements of the rolling bearing, it can be used to characterize different fault types and fault positions, so as to realize the fault diagnosis of the wind turbine under variable speed.

[0059] The above is only the preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A wind turbine transmission chain fault diagnosis method based on time-frequency plane expectation maximization, characterized in that: The steps include: S1; configuring a cylindrical MEMS acceleration sensor in the transmission chain of the wind turbine, and collecting vibration signals of the wind turbine transmission chain through the cylindrical MEMS acceleration sensor; S2: The unsupervised classification method based on the zero point of the spectrum is used to collect the vibration signal and speed signal of the wind turbine during operation. The short-time Fourier transform is used to generate a time-frequency representation of the vibration signal, and the zero point of the spectrum is extracted and unsupervised classification is performed to achieve noise reduction of the vibration signal. S3: Based on the multi-component signal estimation method under the expectation maximization of the time-frequency plane, the instantaneous frequency and instantaneous amplitude of the multi-component signal are accurately estimated on the time domain plane of the denoised signal; S4: Perform order spectrum analysis to identify the vibration signal characteristics of the wind turbine transmission chain and realize fault diagnosis of the wind turbine transmission chain under variable speed.

2. The wind turbine transmission chain fault diagnosis method based on time-frequency plane expectation maximization according to claim 1 is characterized in that: The cylindrical MEMS acceleration sensor comprises a cylindrical base (1), a mass block (2) and a plurality of pole plates (3); a slide groove is provided on the surface of the base (1), and the slide groove is extended along the axial direction of the base (1); the mass block (2) is embedded in the slide groove and is slidably connected to the inner surface of the slide groove; anchor points (2.1) are respectively provided at both ends of the mass block (2); one end of the anchor point (2.1) is fixedly connected to the mass block (2), and the other end of the anchor point (2.1) extends in a direction away from the mass block (2); the anchor point (2.1) is attached to the surface of the base (1) and is slidably connected to the base (1); a plurality of comb teeth (2.2) are symmetrically provided at non-end positions on both sides of the mass block (2); the plurality of comb teeth (2.2) are spaced apart along the axial direction of the base (1), and the plurality of comb teeth (2.2) are spaced apart from each other. One end is fixedly connected to the mass block (2), the other ends of the plurality of comb teeth (2.2) extend in a direction away from the mass block (2), the comb teeth (2.2) are slidably connected to the surface of the base (1), and adjacent comb teeth (2.2) and comb teeth (2.2) and anchor points (2.1) are arranged at intervals; a plurality of pole plates (3) are fixedly arranged on the surface of the base (1) between adjacent comb teeth (2.2), and the position of the pole plates (3) relative to the base (1) remains unchanged; when the fan drive chain generates acceleration, the mass block (2) drives the anchor points (2.1) and the plurality of comb teeth (2.2) to slide along the axial direction of the base (1), changing the spacing between the comb teeth (2.2) and adjacent pole plates (3), thereby generating a current signal, sampling the magnitude of the current signal and converting it into a vibration signal of the fan drive chain under variable speed conditions.

3. The wind turbine transmission chain fault diagnosis method based on time-frequency plane expectation maximization according to claim 2 is characterized in that: The noise reduction process of the vibration signal in step S2 specifically includes: The vibration signal of the fan transmission chain is converted into a time-frequency representation, and the vibration signal is processed in segments using a Gaussian window function. The window slides on the vibration signal to obtain the spectrum information corresponding to each time window, and the spectrum information corresponding to each window is combined to obtain the time-frequency spectrogram of the entire vibration signal; Find the zero points in the time-frequency spectrum and classify them into SS, SN and NN types according to their properties; After completing the zero point classification, SS and SN zero points are selected on the time-frequency plane for Delaunay triangulation to obtain an estimate of the signal domain in the time-frequency plane. The vibration signal in the signal domain estimate is reconstructed using the inverse short-time Fourier transform to obtain an estimate of the noise-free vibration signal.

4. The wind turbine transmission chain fault diagnosis method based on time-frequency plane expectation maximization according to claim 3 is characterized in that: Find the zero point in the time-frequency spectrum and classify it according to its properties. The threshold γ is set. Any point whose spectrum value of the time-frequency spectrum is lower than the threshold is regarded as a zero point. The threshold γ is determined as follows: γ = μ X -k0σ X , where μ X is the mean of the spectrum values, σ X is the standard deviation of the spectrum value, k0 is the adjustment factor; according to the nature of the zero point, the zero point is divided into three categories: 1) SS category: signal-signal zero point; 2) SN category: signal-noise zero point; 3) NN category: noise-noise zero point; Assuming that N noise simulations are performed and N time-frequency spectra are obtained, the position of zero point i in different noise simulations is defined as (t Ki , f Ki ), K = 1, 2, ..., N, for each zero point i, calculate the position change d of the zero point in different noise simulations i : in Together they represent the average position of zero point i in the Nth noise simulation, taking δ SS and δ NN is the position change d i The threshold value, δ SS is the right offset value of the first peak of the signal-signal zero position change, δ NN To cover the left offset value of the second peak of the noise-noise zero point position change, the zero point classification rule is: SS class = {i|d i ≤δ SS }, NN class={i|d i ≥δ NN } and SN class = {i|δ SS ≤d i ≤δ NN }.

5. The wind turbine transmission chain fault diagnosis method based on time-frequency plane expectation maximization according to claim 4 is characterized in that: The value of the adjustment factor k0 is 2.

5.

6. The wind turbine transmission chain fault diagnosis method based on time-frequency plane expectation maximization according to claim 4 is characterized in that: The method obtains an estimate of the signal domain in the time-frequency plane, reconstructs the vibration signal in the signal domain estimate using an inverse short-time Fourier transform, and obtains an estimate of the noise-free vibration signal. Specifically, after completing the zero point classification, SS-type and SN-type zero points are selected on the time-frequency plane for Delaunay triangulation. The distribution of the zero points and the collective characteristics of the corresponding triangles are used to identify the signal domain. The side length l of the triangulation ij The calculation formula is: i, j represent two different zero points, (t i , f i ) and (t j , f j ) are the positions of the two zero points respectively; the weight factor ω is introduced ij Make the triangulation take into account both the density and geometric distance of the zero points, Where ρ(t, f) is the local density of zero points in the time-frequency plane, and the weighted length l′ of each edge of the triangulation is obtained by the weight factor ij , l′ ij =ω ij ·l ij , dynamically adjust the effective length of each edge; for each triangulated triangle, if the weighted length of any edge is greater than the preset edge length threshold l max , the triangle is considered to belong to the signal domain; the side length is greater than the side length threshold l max The triangles are gathered together to obtain an estimate of the signal domain in the time-frequency plane: Based on the obtained signal domain estimation in the time-frequency plane, the signal in the signal domain is reconstructed using the inverse short-time Fourier transform to calculate the estimated value of the noise-free signal 7. A wind turbine transmission chain fault diagnosis method based on time-frequency plane expectation maximization according to claim 6, characterized in that: The content of step S3 is: S31, expectation step: calculate the expected value of the vibration signal, and express the estimated value of the noise-free signal with multiple components as: A k (t) is the instantaneous amplitude of the kth component, φ k (t) is the instantaneous phase, n(t) is the noise; Monte Carlo sampling is then used to estimate the posterior distribution and generate M samples θ from the prior distribution. (i) , where the superscript i represents different moments, assuming that the observed signal is independent at each time point t, for each sample θ (i) , calculate an estimate of the noise-free signal The likelihood function value of Then calculate the weight of each sample Then use the weighted average method to estimate the expected value of the posterior distribution: S32, maximization step: obtaining the expected value that meets the actual vibration signal characteristics After that, the estimated values ​​of instantaneous frequency and amplitude need to be updated; first, the expected value of the vibration signal feature obtained above is brought in, the expected log-likelihood function is maximized, and the optimal instantaneous frequency and amplitude are found: where θ (t+1) Represents the parameter estimation after the t+1th iteration. The parameters of the instantaneous frequency and amplitude are updated using the result after maximizing the expectation: in is the instantaneous phase after the t+1th iteration, is the instantaneous amplitude of the k components after the tth iteration, η is the learning rate; S33. In order to ensure the stability and accuracy of parameter estimation, it is necessary to repeat the expectation step S31 and the maximization step S32 until the parameter estimation converges. The convergence condition of the parameter estimation is |θ (t+1) -θ (t) |ε,θ (t) is the parameter estimate after the tth iteration, ε is the convergence threshold; finally, the instantaneous frequency of each signal component is extracted and the instantaneous amplitude As the input condition for the next step of order spectrum analysis.

8. The wind turbine transmission chain fault diagnosis method based on time-frequency plane expectation maximization according to claim 7 is characterized in that: Step S4 includes: after extracting the instantaneous frequency and instantaneous amplitude of the signal component of the vibration signal, obtaining the instantaneous frequency f(t) and f(t) of the signal feature classification and the speed feature classification. F (t), using the instantaneous frequency f(t) and f F (t) Extracting features of the variable speed signal, and realizing fault diagnosis of the wind turbine transmission chain according to the results of the variable speed signal feature extraction.

9. The wind turbine transmission chain fault diagnosis method based on time-frequency plane expectation maximization according to claim 8, characterized in that: Using the instantaneous frequency f(t) and f F (t) To extract the features of the variable speed signal, the order ratio analysis method is used, and the order ratio L is defined as a multiple of the speed. According to the size of the order ratio, different fault types and fault locations of the wind turbine transmission chain are characterized.

10. The wind turbine transmission chain fault diagnosis method based on time-frequency plane expectation maximization according to claim 2, characterized in that: The gap between adjacent comb teeth (2.2) is greater than the sliding distance of the mass block (2) in the slide groove.

Images