Adaptive simple mode decomposition method for wind turbine gearbox fault diagnosis

Abstract

This invention belongs to the field of fault diagnosis technology. It proposes an adaptive simplified mode decomposition method for fault diagnosis of wind turbine gearboxes. The method acquires the original vibration signal of the gearbox and constructs a multi-channel input matrix. Initial modal components are extracted using an initial multi-channel finite impulse response filter bank. The filter coefficients are updated at the parameter level using reweighted kurtosis as an indicator to enhance the impact characteristics. Then, a two-stage structural optimization is implemented: first, redundant channels are merged based on the consistency of the time-domain periodic structure; then, secondary simplification is performed based on the consistency of the frequency-domain modulation characteristics. Finally, the target filter bank and modal components are obtained. This invention effectively suppresses modal splitting and automatically extracts high signal-to-noise ratio fault features through a two-layer evolution mechanism of parameter updating and structural merging, significantly improving the accuracy and efficiency of early gearbox fault diagnosis.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology, and specifically to an adaptive simplified mode decomposition method for fault diagnosis of wind turbine gearboxes. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] As a typical large-scale rotating machinery system, the operating status of wind turbines directly affects the safety and economy of wind farms. The gearbox, as the core component in a wind turbine that realizes speed transformation and torque transmission, operates under conditions of variable speed, heavy load, and complex environments, making it highly susceptible to early failures. To ensure reliable equipment operation, fault diagnosis techniques based on vibration signals are widely used. In recent years, mode decomposition methods have shown great potential in mechanical fault feature extraction due to their ability to adaptively decompose complex non-stationary signals into several physically meaningful components. Among these, decomposition strategies based on finite impulse response filter banks, through the construction of a multi-channel analysis framework and the optimization of filter parameters by combining impulsive or periodic indices, have gradually become a research hotspot.

[0004] However, existing mode decomposition methods generally suffer from the problem of dispersed periodic impact characteristics when processing vibration signals from wind turbine gearboxes. Because the filter structure remains fixed during decomposition, periodic impact components from the same fault source are often split into multiple modes, leading to dilution of characteristic energy and decreased diagnostic sensitivity. Although some methods introduce reweighting indices to optimize filter coefficients, they lack a dynamic adjustment mechanism for the filter channel structure and cannot automatically merge redundant channels based on the inherent consistency of the signal during iteration. This makes it difficult to eradicate mode splitting, severely affecting the concentrated extraction and accurate identification of fault features. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an adaptive simplified mode decomposition method for wind turbine gearbox fault diagnosis. This method constructs a multi-channel finite impulse response (FIR) filter bank to perform parallel filtering of the input signal, and updates the filters in two stages: the first stage iteratively updates the filter parameters based on reweighted kurtosis; the second stage optimizes, merges, and simplifies the filter structure by utilizing the characteristic consistency between the outputs of each channel. Through this two-stage filter update process, the filter bank gradually evolves into an optimal structure that matches the periodic impulse characteristics of the signal during the iteration process, ultimately obtaining signal components containing the main periodic impulse information, thereby improving the accuracy of wind turbine gearbox fault diagnosis.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] In a first aspect, the present invention provides an adaptive simplified mode decomposition method for fault diagnosis of wind turbine gearboxes.

[0008] An adaptive simplified mode decomposition method for fault diagnosis of wind turbine gearboxes includes the following steps:

[0009] The original vibration signal of the wind turbine gearbox is acquired and constructed into a multi-channel input signal matrix.

[0010] Based on the multi-channel input signal matrix, the initial mode components corresponding to each channel are obtained by filtering using an initial multi-channel finite impulse response filter bank.

[0011] Using the reweighted kurtosis of each initial modal component as the evaluation index, the filter coefficients of the initial multi-channel finite impulse response filter bank are updated at the parameter level to obtain the first updated filter bank and the corresponding first updated modal component.

[0012] Based on the first updated mode component, the consistency of the output signal of each channel in the time domain periodic structure is analyzed. The first structure-level merging is performed on the filter channels that meet the time domain periodic structure consistency condition to obtain the second updated filter group and the corresponding second updated mode component.

[0013] Based on the second updated mode component, the consistency of the output signal of each channel in the frequency domain modulation characteristics is analyzed. The filter channels that meet the frequency domain modulation characteristic consistency condition are subjected to a second structural level merging to obtain the target filter group and the corresponding target mode component.

[0014] Fault characteristics of wind turbine gearboxes are determined using target modal components.

[0015] In one implementation of the first aspect of the present invention, based on a multi-channel input signal matrix, filtering is performed using an initial multi-channel finite impulse response filter bank to obtain the initial mode components corresponding to each channel, including:

[0016] The original vibration signal is converted into a column vector form, and the initial number of filter channels is set.

[0017] The normalized frequency range is evenly divided into sub-bands with the same number of initial filter channels;

[0018] For each sub-band, the initial finite impulse response (FIR) filter coefficients are designed using a window function, and the initial FIR filter coefficients are then subjected to energy normalization to obtain the initial multi-channel FIR filter bank.

[0019] The original vibration signal is copied multiple times to construct a multi-channel input signal matrix. The multi-channel input signal matrix is ​​then convolved and filtered using an initial multi-channel finite impulse response filter bank to obtain the initial modal components corresponding to each channel.

[0020] In one implementation of the first aspect of the present invention, the filter coefficients of the initial multi-channel finite impulse response filter bank are updated at the parameter level using the reweighted kurtosis of each initial modal component as the evaluation index, to obtain a first updated filter bank and the corresponding first updated modal component, including:

[0021] Each initial modal component is divided into multiple segments, and the kurtosis value of the signal in each segment is calculated.

[0022] The calculated kurtosis value sequence is sorted and a weight sequence is calculated. Based on the weight sequence and the sorted kurtosis value sequence, the reweighted kurtosis value under different number of segments is calculated, and the number of segments with the largest reweighted kurtosis value is selected as the optimal number of segments.

[0023] The input signal is segmented according to the optimal number of segments, and the Toplitz matrix of the input signal for each sub-segment is constructed.

[0024] Based on the partial derivative condition of the kurtosis maximization objective function with respect to the filter coefficients, the optimal filter coefficients for each segment are solved using the Toplitz matrix and the output signal vector, and then normalized.

[0025] The optimal filter coefficients corresponding to each segment are weighted and fused using the weight sequence to obtain the updated filter coefficients and normalize them to form the first updated filter group.

[0026] The first updated mode component is obtained by filtering the multi-channel input signal matrix using the first updated filter bank.

[0027] As a further limitation of the first aspect of the present invention, calculating the kurtosis value of each sub-segment signal includes: for any sub-segment signal, calculating the ratio of the fourth power sum to the square of the second power sum of all sampling points in the sub-segment signal, and using the ratio as the kurtosis value of the sub-segment signal;

[0028] The reweighted kurtosis value is calculated based on the weight sequence and the sorted kurtosis value sequence for different number of segments, including: multiplying the elements in the sorted kurtosis value sequence with the corresponding elements in the sorted weight sequence, summing all the product results, and using the summation result as the reweighted kurtosis value for that number of segments.

[0029] As a further limitation of the first aspect of the present invention, based on the partial derivative condition of the kurtosis maximization objective function with respect to the filter coefficients, the optimal filter coefficients corresponding to each segment are solved using the Topletz matrix and the output signal vector, including:

[0030] ;

[0031] in, Representing the The filter channel, the first The optimal filter coefficient vector corresponding to each sub-segment; The filtered output signal vector represents the nth sub-segment; Representative to The sequence obtained by squaring each element; Representative to The sequence after cubeing each element in the sequence; Representative to The sequence obtained by raising each element to the fourth power; This represents the Toplitz matrix constructed from the input signal of the nth sub-segment; Represents the Toplitz matrix The transpose of .

[0032] In one implementation of the first aspect of the present invention, based on the first updated modal component, the consistency of the output signals of each channel in the time-domain periodic structure is analyzed, and a first structural-level merging is performed on the filter channels that satisfy the time-domain periodic structure consistency condition to obtain a second updated filter bank and the corresponding second updated modal component, including:

[0033] Calculate the Pearson correlation coefficient between each first updated modal component and the original vibration signal, and sort the modes according to the magnitude of the Pearson correlation coefficient;

[0034] Starting from the first sorted filter channel, calculate the normalized cross-correlation spectrum of the output signals of any two filter channels;

[0035] Peak detection was performed on the normalized cross-correlation spectrum to extract the set of significant peaks and the peak position sequence;

[0036] If the number of significant peak sets satisfies the minimum number constraint, the maximum peak amplitude exceeds the preset threshold, and the peak position interval sequence passes the statistical consistency test, then the corresponding two filter channels are determined to be consistent in the time-domain periodic structure representation.

[0037] The filter coefficients and modal components corresponding to the two filter channels that are determined to be consistent are added and fused to obtain the second updated filter group and the corresponding second updated modal components.

[0038] In one implementation of the first aspect of the present invention, calculating the normalized cross-correlation spectrum of the output signals of any two filter channels includes: calculating the cross-correlation function of the output signals of the two filter channels; calculating the mean and standard deviation terms of the cross-correlation function under different time delays; and obtaining the normalized cross-correlation spectrum by subtracting the mean term from the cross-correlation function and dividing it by the standard deviation term.

[0039] In one implementation of the first aspect of the present invention, based on the second updated mode component, the consistency of the output signals of each channel in the frequency domain modulation characteristics is analyzed, and a second structural-level merging is performed on the filter channels that meet the frequency domain modulation characteristic consistency condition to obtain the target filter bank and the corresponding target mode component, including:

[0040] Hilbert transform is performed on each of the second updated mode components after the first structural-level merging to obtain the analytic signal and extract the envelope signal;

[0041] Calculate the power spectral density of each envelope signal and normalize the power spectral density to obtain the normalized envelope power spectrum.

[0042] Calculate the correlation coefficient between the normalized envelope power spectra of any two filter channels;

[0043] If the correlation coefficient is greater than or equal to the preset correlation coefficient threshold, then the two corresponding filter channels are determined to be consistent at the modulation frequency structure level.

[0044] The filter coefficients and modal components corresponding to the two filter channels that are determined to be consistent are added and fused to obtain the target filter group and the corresponding target modal components.

[0045] In one implementation of the first aspect of the present invention, calculating the power spectral density of each envelope signal includes:

[0046] ;

[0047] Where q represents the filter channel index; k represents the iteration number index; n represents the discrete time index; and f represents the frequency variable. This represents the filtered output time-domain signal of the q-th channel at the k-th iteration; j represents the imaginary unit. Represents the Hilbert transform; This represents the analytic signal of the filtered output signal of the q-th channel at the k-th iteration. This represents the envelope signal of the filtered output signal of the q-th channel at the k-th iteration. Represents power spectral density; This represents the power spectral density of the envelope signal of the q-th channel at the k-th iteration.

[0048] Secondly, the present invention provides an adaptive simplified mode decomposition system for fault diagnosis of wind turbine gearboxes.

[0049] An adaptive simplified mode decomposition system for fault diagnosis of wind turbine gearboxes includes:

[0050] The signal construction unit is configured to: acquire the original vibration signal of the wind turbine gearbox and construct the original vibration signal into a multi-channel input signal matrix;

[0051] The initial filtering unit is configured to perform filtering based on the multi-channel input signal matrix using an initial multi-channel finite impulse response filter bank to obtain the initial mode components corresponding to each channel.

[0052] The parameter update unit is configured to update the filter coefficients of the initial multi-channel finite impulse response filter bank at the parameter level, using the reweighted kurtosis of each initial modal component as the evaluation index, to obtain the first updated filter bank and the corresponding first updated modal component.

[0053] The time-domain merging unit is configured to: analyze the consistency of the output signals of each channel in the time-domain periodic structure based on the first updated modal component, perform the first structure-level merging on the filter channels that meet the time-domain periodic structure consistency condition, and obtain the second updated filter group and the corresponding second updated modal component;

[0054] The frequency domain merging unit is configured to: analyze the consistency of the frequency domain modulation characteristics of the output signals of each channel based on the second updated mode component, and perform a second structural-level merging on the filter channels that meet the frequency domain modulation characteristic consistency condition to obtain the target filter group and the corresponding target mode component;

[0055] The fault diagnosis unit is configured to determine the fault characteristics of the wind turbine gearbox using the target modal components.

[0056] Compared with the prior art, the beneficial effects of the present invention are:

[0057] This invention fundamentally changes the passive situation of traditional mode decomposition (MDD) where filter structures are fixed or rely on manual presets by constructing a two-layer adaptive evolution mechanism of parameter-level fine-tuning and structure-level dynamic merging. In the decomposition of strong non-stationary signals such as wind turbine gearboxes, this method can automatically drive a multi-channel finite impulse response (FIR) filter bank to reconstruct the topology in real time based on the inherent periodic impulse characteristics of the signal. This inherent adaptive capability allows the algorithm to quickly lock onto key frequency bands in the early stages of iteration without external intervention, and efficiently converge the dispersed periodic impulse energy into at least a few or even a single dominant mode component in subsequent processes. This not only fundamentally avoids mode aliasing and mode splitting, but also significantly reduces the number of iterations required to reach convergence while ensuring high-fidelity extraction of fault features. It achieves a dual breakthrough in decomposition accuracy and computational efficiency, providing solid theoretical support and technical pathways for real-time fault diagnosis under complex operating conditions.

[0058] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0059] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0060] Figure 1 A flowchart illustrating an adaptive simplified mode decomposition method for fault diagnosis of wind turbine gearboxes, provided as an exemplary embodiment of the present invention.

[0061] Figure 2 The measured vibration signal of a wind turbine gearbox in a wind farm, provided as an exemplary embodiment of the present invention, wherein... Figure 2 (a) in the diagram is the time-domain plot of the original signal. Figure 2 (b) in the diagram is the original signal frequency domain diagram;

[0062] Figure 3 The MRMD decomposition signal component map (iteration 4 times) is provided as an exemplary embodiment of the present invention. Figure 3 (a) in the figure is the time-domain plot of the dominant periodic mode. Figure 3 (b) in the diagram is the frequency domain plot of the dominant periodic mode. Figure 3 (c) in the diagram is the time-domain plot of the residual modes. Figure 3 (d) in the diagram represents the frequency domain plot of the residual modes;

[0063] Figure 4 The RMD decomposition signal component map (iterated 14 times) provided as an exemplary embodiment of the present invention, wherein, Figure 4In the diagram, (a) and (b) are the time-domain and frequency-domain plots of the first component after decomposition. Figure 4 In the diagram, (c) and (d) are the time-domain and frequency-domain plots of the second component after decomposition;

[0064] Figure 5 This is a schematic diagram of an adaptive simplified mode decomposition system for fault diagnosis of wind turbine gearboxes, provided as an exemplary embodiment of the present invention. Detailed Implementation

[0065] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0066] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0067] Mode decomposition based on vibration signals has been widely applied in fault feature extraction due to its ability to decompose complex signals into components with specific frequency bands or physical meanings. Traditional empirical mode decomposition and its improved methods have improved the processing capability of non-stationary signals to some extent, but still suffer from problems such as mode aliasing, endpoint effects, and insufficient noise resistance. Subsequent variational mode decomposition methods achieve frequency band separation by constructing an optimization model, theoretically improving decomposition stability, but their decomposition effect is highly sensitive to parameter settings and has high computational complexity. Building on this, eigenmode decomposition (FMD) and its derivatives introduce adaptive finite impulse response filter banks and combine impulsiveness and periodicity indices to model and extract periodic impact features in mechanical fault signals, showing promising application prospects in rotating machinery fault diagnosis. Furthermore, reduced mode decomposition (RMD) is proposed within this framework, driving filter iteration updates with indices such as reweighted kurtosis and reducing similar modes to gradually concentrate periodic impact information, providing a new approach for periodic impact signal analysis.

[0068] In the complex operating environment of wind turbine gearboxes, methods such as Ensemble Empirical Mode Decomposition (EEMD), Variational Mode Decomposition (VMD), and Singular Value Decomposition (SVD) have been widely studied and applied. EEMD suppresses mode aliasing by introducing noise, but its decomposition results are sensitive to noise amplitude and decomposition order, and the periodic impact component may still be weakened under strong noise conditions. VMD requires pre-setting the number of modes and penalty factors, and improper parameter selection can easily lead to frequency band division deviations, causing periodic impact information to be scattered across multiple modes. SVD mainly decomposes based on energy characteristics, lacking direct constraints on periodic structure and modulation characteristics, and has limited adaptability to non-stationary and weak impact signals. Therefore, the above methods are difficult to stably and centrally extract periodic impact features under complex operating conditions.

[0069] Eigenmode Decomposition (FMD) enhances the extraction capability of periodic impulse features to some extent by constructing a multi-channel finite impulse response filter bank and introducing impulse or periodicity indices to iteratively update the filter parameters. However, FMD typically fixes the number of filter channels and the frequency band division structure in the initial decomposition stage, and subsequent iterations mainly focus on numerical optimization of the filter coefficients. When the signal composition is complex or the noise is strong, the impulse information of the same period may still be distributed in the output of multiple filter channels, resulting in mode splitting. This often requires subsequent screening or demodulation analysis to achieve feature aggregation.

[0070] Reduced Mode Decomposition (RMD), building upon Free Mode Decomposition (FMD), reduces redundant modes by introducing indices such as reweighted kurtosis, gradually concentrating periodic impulse information and improving the simplification of some decomposition results. However, the simplification of modes by the RMD method mainly occurs at the result level; the number and structure of filter channels remain unchanged during the decomposition process, lacking a mechanism for dynamic merging and evolution as signal characteristics change. Furthermore, some improved schemes introduce intelligent optimization algorithms to search for decomposition-related parameters, which improves the decomposition effect to some extent, but adds an extra optimization iteration process, resulting in high computational complexity. In complex operating conditions or real-time application scenarios, its ability to rapidly aggregate periodic impulse features remains limited.

[0071] In view of the problems existing in the existing schemes, this implementation proposes an adaptive simplified mode decomposition method for periodic impact signals, such as... Figure 1 As shown, the process includes the following:

[0072] Step S1: Construct the initial multichannel analysis framework for the multichannel reweighted mode decomposition method (MRMD) to provide stable initial values ​​for adaptive optimization.

[0073] The original vibration signal (i.e., the input signal) Initialize as a column vector, based on a preset initial segmentation number. The frequency band is uniformly divided, and an initial finite impulse response filter bank is designed.

[0074] Step S101: Collect the discrete-time raw vibration signal Convert to column vector form:

[0075] (1);

[0076] in, The signal length is and the sampling frequency is . ; represents the 1st, 2nd, ..., Mth discrete-time vibration signal sampling point; T represents transpose.

[0077] Step S102: Set the initial number of filter channels (i.e., the initial number of sub-bands) to... Normalize the frequency range Evenly divided into Sub-band, its first The boundary frequency corresponding to each sub-band for:

[0078] (2);

[0079] Step S103: For The construction length is initial Filter coefficient vector :

[0080] (3);

[0081] in, These represent the coefficients of each order of the initial filter.

[0082] Compared to the Hanning window, the Blackman window has a lower sidelobe level, which can effectively suppress the interference of adjacent frequency band leakage on the initial mode, thereby improving the stability of filter convergence and the reliability of mode decomposition results in the subsequent reweighted kurtosis optimization process. This invention uses the Blackman window function to design the initial filter:

[0083] (4);

[0084] in, This represents the value of the Blackman window function at point n; n represents the discrete-time index of the window function.

[0085] Blackman window function In the design process of the bandpass filter, the impulse response of the ideal bandpass filter is windowed to obtain the initial FIR filter coefficients corresponding to the q-th sub-band. To eliminate the impact of amplitude scale differences between different channels on the reweighted kurtosis (RK) index, the filter coefficients are energy normalized.

[0086] (5);

[0087] in, It is a very small positive number, used to avoid numerical instability caused by a denominator of zero; Represents the filter coefficients after energy normalization; ∥ ∥2 represents the second norm.

[0088] Step S104: Construct a multi-channel input signal matrix :

[0089] (6);

[0090] in, It represents the space of M rows and Q columns of real numbers.

[0091] Original signal copy Each column of signals serves as a separate unit. The input signal of an initial filter.

[0092] Step S2: First-stage filter update operation.

[0093] For each channel's filter, the reweighted kurtosis (RK) of its filtered output signal is used as the evaluation metric. A filter coefficient solution model based on the Toeplitz matrix is ​​constructed, and the filter coefficients of each channel are updated to gradually evolve the filter bank towards a direction that matches the periodic impulse characteristics of the signal, completing the first stage of the filter update process. Step S2 specifically includes:

[0094] Step S201: Filtering operation.

[0095] Using the current number The filter coefficients of the next iteration For input signal Perform convolution filtering:

[0096] (7);

[0097] in, This represents the filtered output at the k-th iteration, the q-th channel, and time n. Represents the original signal in n The sampled value at time l+1; Representing the kth The filter coefficient of the l-th channel in the 1st iteration.

[0098] Step S202: Calculate the reweighted kurtosis (RK) and determine the optimal number of segments. .

[0099] To overcome the limitation of traditional kurtosis being sensitive to occasional impulses, a reweighted kurtosis RK index is introduced. Divided into equal parts Each segment The length of each segment is: Calculate the kurtosis of each segment of the signal. :

[0100] (8);

[0101] in, The kurtosis value of the nth sub-segment signal is represented by h; h represents the index of the sampling point within the sub-segment; y n (h) represents the signal value at point h in the nth sub-segment.

[0102] For the calculated kurtosis sequence Sort in ascending order to get Calculate weights :

[0103] (9);

[0104] For the calculated weight sequence Sort in descending order to get Calculate the reweighted kurtosis value for this number of segments. :

[0105] (10);

[0106] Selecting makes The maximum number of segments is taken as the optimal number of segments in this iteration. .

[0107] Step S203: Construct the observation matrix based on the optimal segmentation.

[0108] according to The input signal is segmented. For the first segment... Each segment ( Construct the Toeplitz matrix of the input signal. :

[0109] (11);

[0110] Where, x n (·) represents the sampled value of the nth sub-segment; C represents the number of sampling points in the sub-segment.

[0111] Step S204: For the first obtained in step S203 Based on the condition of the partial derivative of the kurtosis maximization objective function with respect to the filter coefficients, by setting the partial derivative of the objective function with respect to the filter coefficients to zero, we can obtain the nth sub-segment signal. The optimal filter coefficients for each sub-segment :

[0112] (12);

[0113] in, For the first The Topulitz matrix constructed from the input signals of each segment. This indicates that the elements of the output signal vector are cubed, y n The vector representing the filtered output signal of the nth segment is solved. Then normalization is performed:

[0114] (13);

[0115] Step S205: Weighted fusion update filter. Utilize the sorting weight vector obtained in step S202. , to press The sub-filters arranged in ascending order are weighted and fused to obtain the updated filter:

[0116] (14);

[0117] The normalized updated filter:

[0118] (15);

[0119] Step S206: If the current inner iteration count is... If the preset maximum value is not reached, return to step S201, recalculate the reweighted kurtosis based on the new output signal and update the filter; otherwise, end the iteration process of this channel and output the modal components corresponding to the current channel.

[0120] Step S3: Second stage filter update operation.

[0121] This process analyzes the consistency relationship between the outputs of different channels in the time and frequency domain feature spaces, and merges, simplifies, and reorganizes the filters, so that the structure of the filter bank gradually evolves into a simplified representation that matches the inherent periodic impulse structure of the signal. That is, the first structural merging is based on the consistency of the time-domain periodic structure, and the second structural merging is based on the consistency of the frequency-domain modulation characteristics.

[0122] Step S301: First structural merging based on temporal periodic structural consistency.

[0123] Calculate the modes output in step S2 With the original signal Pearson correlation coefficient:

[0124] (16);

[0125] in, This represents the Pearson correlation coefficient between the q-th mode in the k-th iteration and the original signal; This represents the output of the q-th channel in the k-th iteration; This represents the average output value of the q-th channel; represents the mean of the original signal; n represents the time index.

[0126] according to The modes are sorted in descending order of their magnitude, prioritizing modes with strong correlation to the original signal. Starting from the first filter channel after sorting, any two filter channels are then sorted... and Construct the normalized cross-correlation spectrum of its output signal. :

[0127] (17);

[0128] in, It is a cross-correlation function. and These represent the mean and standard deviation terms under the corresponding time delay, respectively, and are used to eliminate the influence of amplitude scale differences.

[0129] right Peak detection is performed to extract the set of significant peaks and the peak position sequence. If the following conditions are met simultaneously: 1) the number of significant peaks meets the minimum constraint; 2) the maximum peak amplitude exceeds a preset threshold; 3) the peak position interval sequence passes the statistical consistency test, indicating that the two channel outputs have a stable and consistent periodic structure, then the filter channel is determined. and They exhibit consistency in the representation of time-domain periodic structures, and a first structural-level merging is performed on them:

[0130] (18);

[0131] in, and These are the filter coefficients for channels q and r, respectively; and These are the output signals for channels q and r, respectively. and These are the filter coefficients and output signal for the first update of channels q and r, respectively.

[0132] Step S302: Second structural-level update based on frequency domain modulation feature consistency.

[0133] For the mode set fused in step S302, calculate its envelope power spectral density (PSD):

[0134] (19);

[0135] Normalize it:

[0136] (20);

[0137] For any two filter channels and Calculate the correlation coefficient between their normalized envelope power spectra:

[0138] (twenty one);

[0139] Set threshold It is assumed that the corresponding filter channels have a high degree of consistency at the modulation frequency structure level, and a second structure-level merging is performed on them:

[0140] (twenty two);

[0141] in, and These are the filter coefficients for the first update of channels q and r, respectively; and These are the output signals for the first update of channels q and r, respectively. and These are the filter coefficients and output signals for the second update of channels q and r, respectively.

[0142] Step S303: Use the filter bank and its channel outputs after two structural level updates as the first... The input for the next outer iteration. If the number of outer iterations has not reached the preset maximum value, and the current number of filter channels is still greater than the target number of channels, then return to step S2 to continue performing parameter-level and structure-level updates; otherwise, output the current filter bank. and its corresponding channel output This serves as the final simplified mode decomposition result.

[0143] More specifically, this implementation also provides the following calculation example:

[0144] This embodiment takes a 2MW wind turbine gearbox as the research object. The gearbox adopts a two-stage planetary gear plus a single-stage parallel shaft structure, with the vibration sensor installed at the high-speed shaft bearing housing. The sampling frequency is 25600Hz, the sampling duration is 1s, and the acquired signal includes gear meshing impact components and random wind load interference noise. The MRMD method constructs a multi-channel finite impulse response filter bank and introduces a reweighted kurtosis-driven parameter update mechanism and a modal structure simplification mechanism during the iteration process. This allows the filter bank to gradually evolve towards representing the main periodic impact characteristics of the signal, thereby achieving efficient and simplified modal decomposition.

[0145] The vibration acceleration signal of a rotating machinery transmission system was collected, and the sampling frequency was set to [value missing]. =25600 Hz, the amplitude of the signal is in units of Set the initial number of filter channels (i.e., the initial number of sub-bands) to... Set the filter length to Calling MATLAB functions use Window design initial bandpass filter bank The original signal is copied 15 times to form a multi-channel input matrix. .

[0146] Perform the first stage filter parameter update. For each channel input signal... The following iterations are performed:

[0147] (1) Perform filtering operation Initial stage ;

[0148] (2) Calculate the current output signal reweighted kurtosis Determine the optimal number of segments (Pick The range is 2 to 10.

[0149] (3) According to , input signal Segmentation, using MATLAB functions Construct the Toplitz matrix for each sub-segment .

[0150] (4) Solve for the optimal filter coefficients for each segment. Then, normalization is performed using a weight vector. The filter coefficients of each sub-segment are weighted and fused to obtain the updated filter. And normalize.

[0151] (5) Repeat the above process until the set number of iterations is completed, and output the current channel mode. Here, it is stipulated that the maximum number of iterations in step 2, i.e., the RK iteration, is 2 fewer than the maximum number of iterations in the MRMD algorithm.

[0152] Perform the second stage filter structure update.

[0153] (1) Consistency merging based on time-domain period: Calculate each mode With the original signal Pearson correlation coefficient Sort the modes in descending order. Starting from the first mode, calculate the normalized cross-correlation spectrum between any two modes. Detect its peak value. If it simultaneously meets the following conditions: number of peaks ≥ 3, maximum peak value > 0.25, and peak interval ≥ 3... For consistency checks, the two modes and their corresponding filters are merged.

[0154] (2) Merging based on frequency domain modulation consistency: calling MATLAB functions Calculate the envelope signal of each mode power spectral density Calculate the correlation coefficient between normalized envelope PSDs. ,like Then the two modes and their corresponding filters are combined.

[0155] (3) Determine whether the preset maximum number of iterations has been reached. If not, use the merged modality set as input and return to step 2; otherwise, output the final feature modality set.

[0156] To verify the effectiveness and engineering adaptability of the method of this invention in extracting early fault impact signals of wind turbine gearboxes, a comparative experiment was conducted between the method of this invention (MRMD) and the prior art method (RMD) under the same experimental conditions and parameter configurations. The method of this invention (MRMD) was set to iterate 4 times, while the prior art method (RMD) output decomposition results at 4 and 14 iterations, respectively, to compare the feature extraction capabilities of the two methods at different decomposition stages.

[0157] Figure 2 This is a time-domain representation of the original signal. Figure 3 This is a signal component image obtained after 4 iterations of MRMD using the method of this invention. Figure 4 The signal component diagram is obtained by iterating RMD 14 times using the existing technology method.

[0158] In the decomposition results of the method (MRMD) of this invention, Mode#1 represents the dominant periodic mode extracted from the original signal, mainly containing periodic impulse components and their harmonic structures; Mode#2 represents the residual mode, mainly composed of broadband random components and high-frequency noise. The method (MRMD) of this invention can achieve effective separation between the dominant periodic impulse components and the residual components in the fourth iteration, and the periodic impulse information can be completely aggregated into a single mode.

[0159] The existing RMD method is still in the early stages of decomposition after four iterations. RC#1 may contain some periodic impact characteristics, but it still contains many non-periodic components. The components from RC#2 to RC#12 mainly exhibit narrow-band or wide-band components with small amplitudes, and some periodic-related frequency information remains. The overall periodic structure is not yet concentrated. Only after 14 iterations does the principal component RC#1 of the existing RMD method exhibit periodic impact characteristics similar to those of the method of this invention after four iterations. Clearly, the intermediate decomposition results of the existing RMD method show obvious modal splitting, meaning that the same periodic impact information is split and distributed across multiple components, requiring multiple iterations to gradually reconstruct the complete mode.

[0160] Experimental results show that the MRMD method of the present invention can effectively extract the principal periodic modes with fewer iterations, significantly reducing the number of iterations required while ensuring decomposition accuracy, thus demonstrating higher computational efficiency and better modal consistency.

[0161] In summary, this invention innovatively embeds the reweighted kurtosis index deep into the parameter update closed loop of the filter coefficients. Instead of blindly performing uniform decomposition across the entire frequency band, this invention utilizes reweighted kurtosis to sensitively evaluate the initial modal components, accurately identifying and strengthening sub-segments containing periodic impulse information while suppressing noise-dominated invalid components. This statistically-driven guiding mechanism forces the filter coefficients to evolve towards the direction with the most significant impulse characteristics in each parameter-level update. This proactively avoids the generation of meaningless modes at the initial stage of the decomposition process, eliminating the need to expend computational resources to generate spurious modes that subsequently need to be eliminated. Instead, it directly outputs high-purity candidate modes, greatly simplifying subsequent processing and ensuring that the final modal components have extremely high signal-to-noise ratios and clear physical meanings, significantly improving the targeting and effectiveness of fault feature extraction.

[0162] To address the common problem of modal splitting in complex mechanical vibration signals, this invention constructs a dual-verification merging mechanism based on the consistency of time-domain periodic structure and frequency-domain modulation characteristics, achieving intelligent pruning and fusion of filter bank structures. In the first stage, the mechanism accurately identifies filter channels with the same periodic repetition pattern in the time-domain waveform by calculating the normalized cross-correlation spectrum and peak statistics, and merges these redundant channels representing the same physical source for the first time. Then, in the second stage, the envelope power spectrum is extracted using Hilbert transform and correlation is calculated to further verify the rationality of the merging from the perspective of frequency-domain modulation, performing secondary structural simplification. This time-to-frequency, progressive logic ensures that only signals truly representing the same fault source are aggregated, while interference or independent fault components from different sources are effectively preserved. This allows the same fault feature, originally dispersed in multiple weak modes, to be completely and compactly recombined into a single strong feature mode, greatly enhancing the identification of fault features and avoiding missed detections or misjudgments caused by feature dispersion.

[0163] Unlike traditional time-consuming schemes that rely on swarm intelligence optimization algorithms for global search or multi-stage serial processing, this invention employs a deterministic two-stage filter update strategy, significantly reducing the computational complexity of the algorithm and improving its engineering applicability. The entire decomposition process abandons the highly stochastic iterative search phase, instead relying on rigorous mathematical derivations (such as solving the Toplitz matrix and analyzing the signal envelope) to directly calculate the optimal filter coefficients and merging strategy. This deterministic computational path not only eliminates the instability caused by random initialization but also makes the algorithm's runtime controllable and predictable. In practical applications for online monitoring of wind turbines, this method can complete the entire process from acquiring the original vibration signal to outputting the target fault mode within a very short time window, meeting real-time requirements without the need for high-performance computing hardware. It balances the depth and speed of the algorithm, enabling high-precision adaptive mode decomposition technology to truly move out of the laboratory and meet the stringent requirements of low latency and high reliability in industrial settings.

[0164] Figure 5 An adaptive simplified mode decomposition system for wind turbine gearbox fault diagnosis is shown, including:

[0165] The signal construction unit 501 is configured to: acquire the original vibration signal of the wind turbine gearbox and construct the original vibration signal into a multi-channel input signal matrix;

[0166] The initial filtering unit 502 is configured to: perform filtering processing based on the multi-channel input signal matrix using an initial multi-channel finite impulse response filter bank to obtain the initial mode components corresponding to each channel;

[0167] The parameter update unit 503 is configured to update the filter coefficients of the initial multi-channel finite impulse response filter bank at the parameter level using the reweighted kurtosis of each initial modal component as the evaluation index, so as to obtain the first updated filter bank and the corresponding first updated modal component.

[0168] The time-domain merging unit 504 is configured to: analyze the consistency of the output signals of each channel in the time-domain periodic structure based on the first updated modal component, perform the first structure-level merging on the filter channels that meet the time-domain periodic structure consistency condition, and obtain the second updated filter group and the corresponding second updated modal component;

[0169] The frequency domain merging unit 505 is configured to: analyze the consistency of the frequency domain modulation characteristics of the output signals of each channel based on the second updated mode component, and perform a second structural-level merging on the filter channels that meet the frequency domain modulation characteristic consistency conditions to obtain the target filter group and the corresponding target mode component;

[0170] The fault diagnosis unit 506 is configured to determine the fault characteristics of the wind turbine gearbox using the target modal components.

[0171] It is understood that the aforementioned units can be individually or entirely merged into one or more other units, or some of the units can be further divided into multiple functionally smaller units. This achieves the same operation without affecting the technical effects of the embodiments of the present invention. The aforementioned units are based on logical functional division. In practical applications, the function of one unit can be implemented by multiple units, or the function of multiple units can be implemented by one unit. In other embodiments of the present invention, the system may also include other units. In practical applications, these functions can also be implemented with the assistance of other units, and can be implemented collaboratively by multiple units.

[0172] According to another embodiment of the present invention, the system of this embodiment can be constructed by running a computer program (including program code) capable of performing the steps involved in the corresponding method of the present invention on a general-purpose computing device, such as a computer, which includes processing elements and storage elements such as a central processing unit (CPU), random access memory (RAM), and read-only memory (ROM). The computer program can be recorded on, for example, a computer-readable recording medium, loaded into the aforementioned computing device through the computer-readable recording medium, and run therein.

[0173] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An adaptive and reduced modal decomposition method for wind turbine gearbox fault diagnosis, characterized in that, Includes the following processes: The original vibration signal of the wind turbine gearbox is acquired and constructed into a multi-channel input signal matrix. Based on the multi-channel input signal matrix, the initial mode components corresponding to each channel are obtained by filtering using an initial multi-channel finite impulse response filter bank. Using the reweighted kurtosis of each initial modal component as the evaluation metric, the filter coefficients of the initial multi-channel finite impulse response filter bank are updated at the parameter level to obtain the first updated filter bank and the corresponding first updated modal components, including: Each initial modal component is divided into multiple segments, and the kurtosis value of the signal in each segment is calculated. The calculated kurtosis value sequence is sorted and a weight sequence is calculated. Based on the weight sequence and the sorted kurtosis value sequence, the reweighted kurtosis value under different number of segments is calculated, and the number of segments with the largest reweighted kurtosis value is selected as the optimal number of segments. The input signal is segmented according to the optimal number of segments, and the Toplitz matrix of the input signal for each sub-segment is constructed. Based on the partial derivative condition of the kurtosis maximization objective function with respect to the filter coefficients, the optimal filter coefficients for each segment are solved using the Toplitz matrix and the output signal vector, and then normalized. The optimal filter coefficients corresponding to each segment are weighted and fused using the weight sequence to obtain the updated filter coefficients and normalize them to form the first updated filter group. The first updated mode component is obtained by filtering the multi-channel input signal matrix using the first updated filter bank. Based on the first updated mode component, the consistency of the output signal of each channel in the time-domain periodic structure is analyzed. The first structure-level merging is performed on the filter channels that meet the time-domain periodic structure consistency condition to obtain the second updated filter bank and the corresponding second updated mode component, including: Calculate the Pearson correlation coefficient between each first updated modal component and the original vibration signal, and sort the modes according to the magnitude of the Pearson correlation coefficient; Starting from the first sorted filter channel, calculate the normalized cross-correlation spectrum of the output signals of any two filter channels; Peak detection was performed on the normalized cross-correlation spectrum to extract the set of significant peaks and the peak position sequence; If the number of significant peak sets satisfies the minimum number constraint, the maximum peak amplitude exceeds the preset threshold, and the peak position interval sequence passes the statistical consistency test, then the corresponding two filter channels are determined to be consistent in the time-domain periodic structure representation. The filter coefficients and modal components corresponding to the two filter channels that are determined to be consistent are added and fused to obtain the second updated filter group and the corresponding second updated modal components. Based on the second updated mode component, the consistency of the output signals of each channel in the frequency domain modulation characteristics is analyzed. A second structural-level merging is performed on the filter channels that meet the frequency domain modulation characteristic consistency condition to obtain the target filter bank and the corresponding target mode components, including: Hilbert transform is performed on each of the second updated mode components after the first structural-level merging to obtain the analytic signal and extract the envelope signal; Calculate the power spectral density of each envelope signal and normalize the power spectral density to obtain the normalized envelope power spectrum. Calculate the correlation coefficient between the normalized envelope power spectra of any two filter channels; If the correlation coefficient is greater than or equal to the preset correlation coefficient threshold, then the two corresponding filter channels are determined to be consistent at the modulation frequency structure level. The filter coefficients and modal components corresponding to the two filter channels that are determined to be consistent are added and fused to obtain the target filter group and the corresponding target modal components; Fault characteristics of wind turbine gearboxes are determined using target modal components.

2. The adaptive simplified mode decomposition method for wind turbine gearbox fault diagnosis as described in claim 1, characterized in that, Based on the multi-channel input signal matrix, filtering is performed using an initial multi-channel finite impulse response filter bank to obtain the initial mode components corresponding to each channel, including: The original vibration signal is converted into a column vector form, and the initial number of filter channels is set. The normalized frequency range is evenly divided into sub-bands with the same number of initial filter channels; For each sub-band, the initial finite impulse response (FIR) filter coefficients are designed using a window function, and the initial FIR filter coefficients are then subjected to energy normalization to obtain the initial multi-channel FIR filter bank. The original vibration signal is copied multiple times to construct a multi-channel input signal matrix. The multi-channel input signal matrix is ​​then convolved and filtered using an initial multi-channel finite impulse response filter bank to obtain the initial modal components corresponding to each channel.

3. The adaptive simplified mode decomposition method for wind turbine gearbox fault diagnosis as described in claim 1, characterized in that, Calculate the kurtosis value of each sub-segment signal, including: for any sub-segment signal, calculate the ratio of the fourth power sum to the square of the second power sum of all sampling points in the sub-segment signal, and use the ratio as the kurtosis value of the sub-segment signal; The reweighted kurtosis value is calculated based on the weight sequence and the sorted kurtosis value sequence for different number of segments, including: multiplying the elements in the sorted kurtosis value sequence with the corresponding elements in the sorted weight sequence, summing all the product results, and using the summation result as the reweighted kurtosis value for that number of segments.

4. The adaptive simplified mode decomposition method for wind turbine gearbox fault diagnosis as described in claim 1, characterized in that, Based on the partial derivative condition of the kurtosis maximization objective function with respect to the filter coefficients, the optimal filter coefficients for each segment are solved using the Topulitz matrix and the output signal vector, including: ； in, Representing the The filter channel, the first The optimal filter coefficient vector corresponding to each sub-segment; The filtered output signal vector represents the nth sub-segment; Representative to The sequence obtained by squaring each element; Representative to The sequence obtained by cubeing each element in the sequence; Representative to The sequence obtained by raising each element to the fourth power; This represents the Toplitz matrix constructed from the input signal of the nth sub-segment; Represents the Toplitz matrix The transpose of .

5. The adaptive simplified mode decomposition method for wind turbine gearbox fault diagnosis as described in claim 1, characterized in that, Calculate the normalized cross-correlation spectrum of the output signals of any two filter channels, including: calculating the cross-correlation function of the output signals of the two filter channels; calculating the mean and standard deviation of the cross-correlation function under different time delays; and obtaining the normalized cross-correlation spectrum by subtracting the mean from the cross-correlation function and dividing by the standard deviation.

6. The adaptive simplified mode decomposition method for wind turbine gearbox fault diagnosis as described in claim 1, characterized in that, Calculate the power spectral density of each envelope signal, including: ； Where q represents the filter channel index; k represents the iteration number index; n represents the discrete time index; and f represents the frequency variable. This represents the filtered output time-domain signal of the q-th channel at the k-th iteration; j represents the imaginary unit. Represents the Hilbert transform; This represents the analytic signal of the filtered output signal of the q-th channel at the k-th iteration. This represents the envelope signal of the filtered output signal of the q-th channel at the k-th iteration. Represents power spectral density; This represents the power spectral density of the envelope signal of the q-th channel at the k-th iteration.

7. An adaptive simplified mode decomposition system for wind turbine gearbox fault diagnosis, utilizing the adaptive simplified mode decomposition method for wind turbine gearbox fault diagnosis as described in any one of claims 1-6, characterized in that, include: The signal construction unit is configured to: acquire the original vibration signal of the wind turbine gearbox and construct the original vibration signal into a multi-channel input signal matrix; The initial filtering unit is configured to perform filtering based on the multi-channel input signal matrix using an initial multi-channel finite impulse response filter bank to obtain the initial mode components corresponding to each channel. The parameter update unit is configured to update the filter coefficients of the initial multi-channel finite impulse response filter bank at the parameter level, using the reweighted kurtosis of each initial modal component as the evaluation index, to obtain the first updated filter bank and the corresponding first updated modal component. The time-domain merging unit is configured to: analyze the consistency of the output signals of each channel in the time-domain periodic structure based on the first updated modal component, perform the first structure-level merging on the filter channels that meet the time-domain periodic structure consistency condition, and obtain the second updated filter group and the corresponding second updated modal component; The frequency domain merging unit is configured to: analyze the consistency of the frequency domain modulation characteristics of the output signals of each channel based on the second updated mode component, and perform a second structural-level merging on the filter channels that meet the frequency domain modulation characteristic consistency condition to obtain the target filter group and the corresponding target mode component; The fault diagnosis unit is configured to determine the fault characteristics of the wind turbine gearbox using the target modal components.

Images