An Adaptive Wind Turbine Bearing Fault Diagnosis Method Based on Modal Reliability Feedback and Agent Backward Search

By constructing a multi-dimensional modal credibility evaluation system and optimizing VMD parameters using a proxy reverse search strategy, the problems of low modal screening accuracy and high computational cost in wind turbine bearing fault diagnosis were solved, and efficient fault feature extraction and identification under complex noise conditions were achieved.

CN123115984APending Publication Date: 2026-11-17SHANDONG UNIV OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611329730.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-31
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies for wind turbine bearing fault diagnosis suffer from low modal screening accuracy, poor adaptability to operating conditions, and high computational costs, especially in complex noise conditions where it is difficult to effectively extract fault features.

Method used

An adaptive bearing fault diagnosis method based on modal credibility feedback and surrogate reverse search VMD is adopted. By constructing multi-dimensional indicators to quantify IMF modal credibility, and combining surrogate models and reverse search strategies to optimize VMD parameters, closed-loop parameter optimization and modal screening are achieved, thereby improving the accuracy of fault feature extraction and diagnostic stability.

Benefits of technology

It effectively improves the accuracy of fault feature extraction and diagnostic stability under complex noise conditions, reduces computing costs, adapts to signal characteristics of multiple operating conditions and multiple noise intensities, and achieves efficient and accurate fault identification.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115984A_ABST
    Figure CN123115984A_ABST
Patent Text Reader

Abstract

This invention discloses a surrogate reverse search adaptive wind turbine bearing fault diagnosis method based on modal reliability feedback. The method introduces a surrogate model to predict the fitness values ​​of different VMD parameter combinations, and generates surrogate reverse candidate parameters by combining the current optimal parameters, parameter search boundaries, and MCS feedback results. The surrogate model is used to pre-screen the candidate parameters, and the screened parameters undergo true VMD decomposition and MCS evaluation. The bearing vibration signal is decomposed using the optimal parameters, and reliable IMF components are screened based on MCS. After reconstructing the fault feature signal, it is input into the diagnostic model, and the wind turbine bearing fault diagnosis result is output. This invention overcomes the shortcomings of traditional VMD parameters relying on manual experience, improves the inadequacy of using only a single fitness index for parameter optimization, reduces computational overhead, enhances fault feature extraction and anti-interference capabilities, and further improves the accuracy of wind turbine bearing fault diagnosis.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention patent relates to the fields of mechanical equipment fault diagnosis, vibration signal processing and intelligent operation and maintenance technology, specifically to a proxy reverse search VMD adaptive bearing fault diagnosis method based on modal credibility feedback, which is particularly suitable for early fault identification and condition monitoring of wind turbine bearings. Background Technology

[0002] Wind turbine rolling bearings operate under complex and harsh conditions of variable loads, strong vibrations, and intense environmental noise for extended periods, making them highly susceptible to typical failures of the inner ring, outer ring, and rolling elements. As a core load-bearing rotating component, if latent early-stage failures in bearings are not accurately identified, they will gradually deteriorate, exacerbating equipment wear, leading to turbine shutdowns, significantly increasing maintenance costs, and in severe cases, causing turbine safety accidents.

[0003] Raw bearing vibration signals collected in industrial settings are often mixed with a large amount of environmental noise and irrelevant interference components. The fault impact characteristics are weak and have low identifiability, making direct use for fault diagnosis difficult to achieve effective feature extraction. Current mainstream solutions rely on intelligent algorithms such as PSO, GWO, and GA to optimize VMD parameters. While this avoids the limitations of manual parameter tuning, the overall optimization architecture has significant flaws: these methods rely solely on single indicators such as envelope entropy and reconstruction error to construct fitness functions, failing to comprehensively measure the fault feature carrying capacity of each IMF mode; they also lack a refined evaluation mechanism for the reliability of individual mode components, easily including noisy and redundant modes in the feature set, and the fixed weight and fixed threshold selection method cannot adapt to signal characteristics with multiple operating conditions and varying noise intensities; furthermore, the iterative execution of the complete VMD decomposition operation suffers from high computational redundancy and low optimization efficiency, making it difficult to meet the high-precision and high-efficiency fault diagnosis requirements of industrial settings.

[0004] Therefore, there is an urgent need for a bearing fault diagnosis method that can comprehensively evaluate the reliability of IMF modes, dynamically screen high-reliability fault modes, and use the modal reliability feedback results to guide the adaptive optimization of VMD parameters in a closed loop. Summary of the Invention

[0005] This invention addresses the shortcomings of existing VMD parameter optimization and bearing fault diagnosis technologies, such as simplistic architecture, low modal screening accuracy, poor adaptability to operating conditions, and high computational costs. It proposes a surrogate-based reverse search VMD adaptive bearing fault diagnosis method based on modal credibility feedback. This invention aims to construct a closed-loop parameter optimization and modal screening system. It quantifies IMF modal credibility through multi-dimensional indicators and optimizes core VMD parameters by combining a surrogate model and a reverse search strategy. This solves the problems of traditional methods, such as one-sided evaluation indicators, severe interference from invalid modalities, and low parameter optimization efficiency, effectively improving the accuracy of bearing fault feature extraction and diagnostic stability under complex noise conditions.

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

[0007] An adaptive wind turbine bearing fault diagnosis method based on modal reliability feedback and proxy reverse search includes the following steps:

[0008] S1: Acquire bearing vibration signals and perform segmentation, mean removal, and normalization on the signals;

[0009] S2: Define the parameter vector to be optimized for Variational Mode Decomposition (VMD), including the number of mode decomposition layers K and the penalty factor. and determine K and Search scope;

[0010] S3: Generate multiple sets of candidate parameter combinations within the search range using a proxy reverse search strategy:

[0011] Let the parameter vector be:

[0012]

[0013] Randomly generate a set of candidate parameters within the parameter search range:

[0014]

[0015] Where s is the number of candidate parameters.

[0016] For each candidate parameter Construct its inverse candidate parameters:

[0017]

[0018] Where L represents the lower bound of the search space and U represents the upper bound of the search space.

[0019] The expanded candidate pool is formed by combining the initial candidate parameters and the reverse candidate parameters:

[0020]

[0021] S4: Perform VMD decomposition on the signal to be decomposed using each candidate parameter combination to obtain several Intrinsic Mode Components (IMFs). Based on the indices of each IMF, determine the weights of each index using the entropy weight method, and calculate the Modal Credibility Score (MCS) for each IMF. The specific calculation process is as follows:

[0022] The evaluation indicators of each IMF are combined into an indicator matrix:

[0023]

[0024] Where K represents the number of IMFs and q represents the number of rating indicators.

[0025] Normalize the positive indicators:

[0026]

[0027] Normalize negative indicators:

[0028]

[0029] Among them, the redundancy index is a negative index.

[0030] Calculate the percentage of indicators:

[0031]

[0032] Calculate the information entropy of the j-th indicator:

[0033]

[0034] Calculate the weight of the j-th indicator:

[0035]

[0036] Calculate the modal reliability index for the i-th IMF modal component:

[0037]

[0038] in This represents the modal confidence level of the i-th IMF component; Indicates kurtosis index; Indicates modal redundancy; Indicates a shock indicator; This indicates the frequency band energy ratio. This is an index representing the correlation between the IMF component and the original vibration signal. For the corresponding entropy weight method weights, and satisfying .

[0039] S5: Construct a fitness function based on the modal credibility score (MCS), evaluate the candidate parameter combinations, and feed the evaluation results back to the agent reverse search process to update the next round of candidate parameter combinations;

[0040] First, the training set is composed of the parameters that have already been evaluated and their fitness values:

[0041]

[0042] Training the agent model:

[0043] Where g represents the proxy model. Indicates the proxy model parameters. This represents the fitness value predicted by the surrogate model.

[0044] Then, a large number of parameters to be evaluated are generated in the candidate pool, and their fitness is predicted using a surrogate model. Only a small number of parameters with high scores are selected to perform real VMD decomposition and feedback fitness calculation.

[0045] At the same time, for the current optimal parameters Execute elite local reverse perturbation:

[0046]

[0047] Where η is the perturbation coefficient and ϵ is the random perturbation term.

[0048] S6. Repeat the parameter search, VMD decomposition, modal reliability evaluation, and feedback update until the iteration termination condition is met to obtain the optimal VMD parameter combination;

[0049] S7: The bearing vibration signal is decomposed using the optimal VMD parameter combination, and the IMF with high modal reliability is selected by adaptive thresholding for signal reconstruction to obtain the fault feature enhancement signal.

[0050]

[0051] Where λ is the threshold adjustment coefficient, which is taken here. .

[0052] When a certain IMF satisfies:

[0053]

[0054] It is then classified as a highly reliable IMF.

[0055] The selected high-confidence IMFs are then reconstructed:

[0056]

[0057] Where S represents the set of highly reliable IMFs, This indicates a high-confidence reconstructed signal.

[0058] S8: Input the enhanced fault feature signal into a convolutional neural network (CNN) for training and recognition, and output the bearing fault diagnosis result.

[0059] Compared with the prior art, the present invention has the following beneficial effects:

[0060] 1. This invention constructs a closed-loop optimization architecture for VMD parameters driven by modal credibility feedback, breaking the single-layer optimization mode of the traditional algorithm of "parameter iteration - single evaluation - direct update". It uses the contribution of real modal fault features to constrain the direction of parameter optimization, so that VMD parameter iterative optimization can fully adapt to the needs of bearing fault feature extraction, and fundamentally solves the problems of blind parameter optimization and poor adaptability of decomposition effect.

[0061] 2. A multi-dimensional IMF modal credibility evaluation system is constructed, which integrates five core indicators: signal impact characteristics, characteristic energy concentration, signal correlation, and modal redundancy. It can accurately identify effective modes, irregular noise modes, and information redundancy modes that carry fault characteristics. Compared with the single indicator evaluation method, it significantly improves the accuracy and reliability of fault mode screening.

[0062] 3. The entropy weight method is introduced to adaptively allocate the weights of each evaluation index, abandoning the traditional fixed weight evaluation mode. The index weights can be dynamically adjusted according to the signal differences of different fault types, operating speeds, and environmental noise intensities, effectively adapting to complex and ever-changing industrial conditions and improving the generalization ability and robustness of the algorithm. Attached Figure Description

[0063] Figure 1 This is a flowchart of an adaptive wind turbine bearing fault diagnosis method based on modal credibility feedback and proxy reverse search.

[0064] Figure 2 This is a schematic diagram of VMD decomposition and MCS screening;

[0065] Figure 3 This is a bar chart comparing the results of VMD parameter optimization;

[0066] Figure 4 This is a bar chart comparing the diagnostic rates of different models under different signal-to-noise ratios;

[0067] Figure 5 This is a schematic diagram of high-confidence modal component screening. Detailed Implementation

[0068] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments of the present invention will now be described in conjunction with the accompanying drawings.

[0069] This embodiment provides an adaptive wind turbine bearing fault diagnosis method based on modal reliability feedback and proxy reverse search, which specifically includes the following steps:

[0070] S1. Signal Acquisition and Preprocessing.

[0071] An acceleration sensor is installed at the bearing housing or motor end cover to collect vibration acceleration signals of the rolling bearing under different operating conditions. These operating conditions include normal operation, inner ring fault condition, outer ring fault condition, and rolling element fault condition. The sampling frequency is set to 12kHz (a common sampling frequency for wind turbine bearing fault detection, which can fully cover the fault characteristic frequency range).

[0072] The acquired raw vibration signals were processed by removing the mean, normalizing, and segmenting. Specifically, the continuous vibration signal was divided into multiple sample segments of fixed length, with each sample segment having a length of 1024 points. A 50% overlap rate was set between adjacent sample segments to increase the number of samples and maintain the continuity of fault characteristics. Subsequently, each sample segment was divided into a training set, a validation set, and a test set according to a preset ratio of 7:2:1.

[0073] S2, Parameter range setting.

[0074] By fixing the VMD parameter search range, the modal decomposition layer number K ranges from [3,10], and the penalty factor α ranges from [500,5000], a two-dimensional parameter optimization space is constructed to provide boundary constraints for subsequent iterative optimization.

[0075] S3. Constructing the candidate set of reverse search parameters for the proxy.

[0076] An initial set of candidate parameters is randomly generated within a preset parameter range. Inverse parameters are generated through global inverse mapping and fused to obtain an expanded candidate pool to broaden the search range. The fitness of parameters in the candidate pool is pre-predicted based on the random forest surrogate model to quickly screen out high-potential parameter combinations, eliminate invalid parameters, and reduce subsequent redundant calculations. At the same time, elite local inverse perturbation is implemented on the optimal parameters in each iteration to improve the accuracy of local optimization and avoid the algorithm from getting stuck in local optima.

[0077] S4, VMD decomposition and modal credibility calculation.

[0078] The optimal parameter combination obtained in step S3 is used to perform VMD decomposition one by one, resulting in K intrinsic mode components IMF1, IMF2, ..., IMF1. K For each IMF component, five indicators are calculated: kurtosis, original signal correlation coefficient, fault band energy ratio, envelope spectrum impact consistency, and modal redundancy. Positive and negative indicators are distinguished and normalized. The weights of each indicator are adaptively solved using the entropy weight method, and the MCS credibility score of each IMF group is obtained by weighted calculation, thus quantifying the fault characteristic carrying capacity of each mode.

[0079] Table 1. Reliability Indicators of Some IMF Component Modes

[0080] Quantity <![CDATA[IMF1]]> <![CDATA[IMF2]]> <![CDATA[IMF3]]> <![CDATA[IMF4]]> <![CDATA[IMF5]]> <![CDATA[IMF6]]> MCS 0.762 0.743 0.582 0.597 0.773 0.658

[0081] S5. Fitness evaluation and parameter feedback update.

[0082] The mean MCS of all IMFs is used as the core fitness index to evaluate the decomposition and adaptation effect of VMD parameters for each group, and the evaluation results are fed back into the algorithm of this model. The random forest surrogate model is trained and updated using the iterated samples to continuously optimize the parameter prediction accuracy and guide the next round of candidate parameter iteration, forming a closed-loop optimization mechanism.

[0083] S6. Iteration termination determination.

[0084] The process of parameter optimization, signal decomposition, modal evaluation, and parameter update is executed in a loop. Iteration is stopped strictly according to the preset termination condition: a maximum of 30 iterations or 5 consecutive iterations with no fluctuation in fitness. Finally, the optimal VMD parameter combination that fits the current bearing signal is output.

[0085] S7, High-Confidence Modal Screening and Signal Reconstruction.

[0086] VMD decomposition was performed based on optimal parameters. The mean MCS of all IMFs was calculated, and an adaptive screening threshold was generated using an adjustment coefficient λ=0.2. High-confidence fault modes with MCS scores higher than the threshold were selected, while noisy and redundant modes were removed. The effective modes were then superimposed and reconstructed to obtain a reconstructed signal with thorough noise reduction and significantly enhanced fault characteristics. This reconstructed vibration signal retains the main fault impact characteristics while reducing the impact of environmental noise and irrelevant frequency band components on subsequent classification.

[0087] S8. Train and test the fault diagnosis model.

[0088] The reconstructed fault feature enhancement signal is input into the CNN diagnostic model. The model is iteratively trained using the cross-entropy loss function and the Adam optimizer. The model parameters are fine-tuned based on the validation set results. After training, the diagnostic performance is verified through the test set. Finally, the bearing operating status and fault type identification results are output.

[0089] This invention solves the problems of blind parameter optimization and low computational efficiency in traditional VMD by using proxy reverse search and proxy model pre-screening mechanism. Relying on multi-index modal credibility evaluation and adaptive screening strategy, it maximizes the preservation of weak fault impact characteristics and suppresses interference from complex environmental noise. Combined with CNN model, it achieves accurate end-to-end fault identification and is suitable for actual operation and maintenance scenarios of wind turbine bearings with changing operating conditions and strong noise. It has strong engineering practicality.

[0090] To verify the parameter optimization performance, computational efficiency, and noise robustness of the method of this invention, multiple sets of comparative experiments were conducted with traditional PSO-VMD, GWO-VMD, and GA-VMD fault diagnosis methods. All experiments maintained consistent hardware environments, datasets, and model hyperparameters. Experimental results are as follows: Figure 3 As shown.

[0091] The performance comparison results of parameter optimization show that, under the experimental conditions and signal sample conditions presented in this paper, the optimal mode decomposition layer K of the four comparative algorithms is 2. The optimal penalty factor α of the method in this invention is 1831, the optimal fitness value is -0.5044, and the number of iterations required for convergence is only 3. Compared with the 5-6 iterations required for convergence of PSO, GWO, and GA algorithms, the optimization convergence speed of this invention is significantly improved, and the optimal parameter fitting accuracy is higher and the iteration stability is stronger, effectively solving the problems of redundancy and slow convergence in traditional intelligent algorithms.

[0092] Furthermore, to verify the algorithm's anti-interference capability, noise immunity experiments were conducted under different signal-to-noise ratio (SNR) conditions of 10dB, 5dB, 0dB, and -5dB. Figure 4 It can be seen that under normal low-noise conditions, the diagnostic accuracy of each algorithm is basically the same; under severe conditions of strong noise of 0dB and -5dB, the diagnostic accuracy of traditional algorithms is greatly reduced, while the method of this invention can still maintain a high diagnostic accuracy of 91.29% and 80.61%, respectively. Its noise resistance and fault feature extraction capabilities are significantly better than those of traditional algorithms.

[0093] In summary, the proposed surrogate reverse search VMD adaptive bearing fault diagnosis method based on modal credibility feedback innovatively constructs a multi-index modal credibility evaluation system and parameter closed-loop optimization architecture. It significantly reduces computational costs by relying on the surrogate model, strengthens fault features through adaptive modal screening, and achieves high-precision fault identification by combining CNN. It balances efficiency, stability, and robustness, and is suitable for fault diagnosis engineering applications in the complex working conditions of wind turbine bearings.

Claims

1. A proxy-based reverse search adaptive wind turbine bearing fault diagnosis method based on modal reliability feedback, characterized in that, Includes the following steps: S1: Collect bearing vibration signals and preprocess the vibration signals to obtain the signals to be decomposed; S2: Define the parameter vector to be optimized for Variational Mode Decomposition (VMD), including the number of mode decomposition layers K and the penalty factor. and determine K and The search scope; S3: Use a proxy reverse search strategy to generate multiple sets of candidate parameter combinations within the search range; S4: Perform VMD decomposition on the signal to be decomposed using each set of candidate parameters to obtain several intrinsic mode components (IMFs). Based on the kurtosis of each IMF, its correlation with the original signal, the proportion of fault characteristic frequency energy, the envelope spectrum impact consistency index, and the modal redundancy index, determine the weight of each index using the entropy weight method, and calculate the modal credibility score (MCS) of each IMF. S5: Construct a fitness function based on the modal credibility score (MCS), evaluate the candidate parameter combinations, and feed the evaluation results back to the agent reverse search process to update the next round of candidate parameter combinations; S6: Repeat the parameter search, VMD decomposition, modal credibility evaluation and feedback update until the iteration termination condition is met and the optimal VMD parameter combination is obtained; S7: The bearing vibration signal is decomposed using the optimal VMD parameter combination, and the IMF with high modal reliability is screened by adaptive thresholding for signal reconstruction to obtain the fault feature enhancement signal; S8: Input the enhanced fault feature signal into a convolutional neural network (CNN) for training and recognition, and output the bearing fault diagnosis result.

2. The method according to claim 1, characterized in that, In step S3, the proxy reverse search includes a global reverse search and an elite local reverse search: For the parameter vector: Its global inverse parameter is: Where L and U represent the lower and upper bounds of the parameter search range, respectively; its elite local inverse parameter is: Where, θ ∗ Let η represent the current optimal parameters, and let η represent the local search coefficients. This represents a random disturbance term.

3. The adaptive wind turbine bearing fault diagnosis method based on modal reliability feedback and proxy reverse search according to claim 1, characterized in that, In step S4, the entropy weight method determines the weights of each indicator and calculates the Modal Credibility Score (MCS) for each IMF, including: Normalize each evaluation index to obtain a normalized index matrix; Calculate the information entropy of the j-th evaluation index: Where m represents the number of IMF modal components. This represents the normalized proportion of the i-th IMF modal component under the j-th evaluation index; Calculate the weight of the j-th evaluation indicator: Where q represents the number of modal credibility evaluation indicators; Calculate the modal reliability index for the i-th IMF modal component: in This represents the modal confidence level of the i-th IMF component; Indicates kurtosis index; This indicates the proportion of energy at fault characteristic frequencies; Indicates a shock indicator; Indicates the frequency band energy index; This is an index representing the correlation between the IMF component and the original vibration signal. For the corresponding entropy weight method weights, and satisfying .

4. The method according to claim 1, characterized in that, In step S3, the surrogate model is used to fit the nonlinear mapping relationship between the VMD parameter vector and the fitness value, and its inputs are K and The output is a predicted fitness value or a predicted MCS value; the surrogate model is a random forest regression model.

5. The method according to claim 1, characterized in that, In step S7, the adaptive threshold screening reliable mode is: in The adjustment coefficient is used when the i-th IMF mode component satisfies: When this occurs, the IMF mode component is identified as a high-confidence IMF mode component.

6. The method according to claim 1, characterized in that, In step S1, the vibration signal is preprocessed, including segmentation, mean removal, and normalization.

7. The method according to claim 1, characterized in that, The fault diagnosis results include at least inner ring faults, outer ring faults, rolling element faults, and normal conditions.