A double-gaussian bias fault diagnosis method for strong noise wind power bearing

Abstract

The application provides a double Gaussian bias fault diagnosis method for strong noise wind power bearings, and relates to the technical field of wind power generation operation and maintenance.The application constructs a bearing health state monitoring model to monitor the health state of a target wind turbine generator bearing, including extracting a feature sequence from input signals at different times based on a multi-physical sensing wavelet front end module;aggregating the feature sequences at different times into aggregated features based on an adaptive impact aggregation module;performing fusion processing on the aggregated features based on a mixed confidence degree gating mechanism module to extract a high confidence degree feature vector; and predicting the high confidence degree feature vector based on a double Gaussian induction bias module to obtain a predicted wind power bearing fault type.The application reduces noise based on a double Gaussian induction bias mechanism and constructs a high-dimensional decision boundary to suppress complex background interference in a wind power field and accurately capture weak early fault features at an extremely low signal-to-noise ratio.

Description

Technical Field

[0001] This invention belongs to the field of wind power operation and maintenance technology, specifically relating to a dual Gaussian bias fault diagnosis method for noisy wind turbine bearings. Background Technology

[0002] With the transformation of the global energy structure, wind power has become a major force in renewable energy. Wind turbines are typically deployed in areas such as deserts, mountains, or offshore, operating under complex conditions of variable speed, variable load, and strong gusts. As core components of the transmission chain, the main bearings and gearbox bearings are subjected to alternating stress and impact loads over long periods, making them highly susceptible to pitting, spalling, or breakage. Failure to detect early, minor faults can lead to unplanned downtime and huge maintenance costs, or even catastrophic accidents such as collapse. Therefore, vibration signal-based monitoring and diagnostic technology is crucial for ensuring the safe operation of wind turbines.

[0003] Traditional bearing fault diagnosis methods mainly rely on signal processing techniques, such as envelope demodulation, wavelet transform (WT), and empirical mode decomposition (EMD). While these methods are effective to some extent, they heavily depend on expert prior knowledge for frequency band selection or threshold setting, making it difficult to adapt to the non-stationary and non-Gaussian signal characteristics of wind power sites, and they cannot achieve end-to-end automated diagnosis.

[0004] In recent years, deep learning algorithms, represented by Convolutional Neural Networks (CNNs) and Transformers, have achieved remarkable results in the field of fault diagnosis. However, when faced with the high-noise environment of wind power sites, these data-driven models have revealed significant shortcomings. First, vibration signals collected at wind power sites are often overwhelmed by high-intensity background noise (such as aerodynamic noise and gear meshing frequency) and electromagnetic interference, resulting in extremely low signal-to-noise ratios (SNR). Existing deep learning models typically lack inductive biases for noise distribution, making them prone to overfitting to noise features rather than fault impact features during training. This leads to a sharp decline in the model's generalization performance across operating conditions or on high-noise test sets, resulting in domain shift problems. Second, traditional deep neural networks suffer from severe black-box properties. Their feature extraction process lacks physically meaningful mappings, making it impossible for maintenance personnel to determine whether the model's decisions are based on actual fault impact frequencies or spurious correlations in the data. This severely limits their application in high-reliability industrial scenarios.

[0005] The introduction of Kolmogorov-Arnold networks (KANs) and their shortcomings in signal processing: To address the problem of poor model interpretability, KAN networks, built based on the Kolmogorov-Arnold representation theorem, have recently received widespread attention. Unlike traditional multilayer perceptrons (MLPs), KANs place learnable activation functions on the edges, theoretically possessing better function fitting capabilities and parameter efficiency. However, directly applying standard KAN networks to bearing fault diagnosis under strong noise conditions still faces significant technical bottlenecks. On the one hand, standard KANs primarily use B-splines as basis functions. While B-splines offer good smoothness, they struggle to capture high-frequency, abrupt fault characteristics when processing bearing vibration signals containing transient impact components, and their computational complexity increases exponentially with grid density, hindering real-time monitoring. On the other hand, most existing KAN architectures are purely mathematically driven, failing to consider the physical and statistical properties of the signal itself. In the context of strong noise, the signal amplitude distribution often presents a complex mixed state. KAN networks that lack physical constraints and specific statistical priors are difficult to effectively separate noise and fault components and cannot establish robust fault feature mapping relationships under low signal-to-noise ratio.

[0006] In summary, the existing technology lacks a diagnostic method that can both constrain the network learning process using physical mechanisms and resist strong noise interference through specific statistical inductive biases. Summary of the Invention

[0007] To address the shortcomings of existing technologies in diagnosing strong noise in wind turbine bearings, such as weak feature extraction capabilities, poor noise resistance, and lack of physical interpretability, this invention aims to provide a method for diagnosing strong noise faults in wind turbine bearings based on a dual Gaussian inductive bias. By integrating physical mechanism constraints with specific statistical distribution priors, this invention effectively overcomes the limitations of traditional deep learning and standard KAN networks under low signal-to-noise ratios and complex operating conditions, achieving high-precision and robust bearing fault diagnosis.

[0008] This invention provides a dual Gaussian bias fault diagnosis method for high-noise wind turbine bearings, comprising the following steps: A dataset is constructed based on vibration signals collected by sensors installed on wind turbine units; A bearing health status monitoring model is constructed, which includes a multi-physical sensing wavelet front-end module, an adaptive impact aggregation module, a hybrid confidence gating mechanism module, and a dual Gaussian induction bias module set in sequence. The bearing health status monitoring model is trained and validated based on the dataset to obtain a trained bearing health status monitoring model; the health status of the target wind turbine bearing is monitored based on the trained status monitoring model. The multi-physical sensing wavelet front-end module is used to perform preliminary noise reduction on the input signals of the bearing health status monitoring model at different times to obtain feature sequences at different times; The adaptive impact aggregation module is used to aggregate feature sequences at different times to obtain aggregated features; The hybrid confidence gating mechanism module is used to fuse aggregated features to obtain high-confidence feature vectors; The dual Gaussian inductive bias module is used to predict high-confidence feature vectors to obtain the predicted wind turbine bearing fault types.

[0009] Furthermore, the specific method for constructing the dataset is as follows: The vibration signal is normalized and segmented to obtain multiple time series samples and construct a time series sample set; the time series sample set includes time series samples of various fault types and different noise levels, and the fault types include at least inner ring, outer ring, rolling element faults and normal state. Preprocessing of time series samples includes outlier handling and data normalization; the outlier handling includes setting an adaptive sliding window and performing median filtering on the time series samples within the adaptive sliding window. Training and test sets are constructed based on the preprocessed time series samples.

[0010] Furthermore, the specific method for preliminary noise reduction of the input signals at different times of the input bearing health status monitoring model is as follows: The convolution kernel is parameterized as a complex Morlet wavelet kernel function that includes trainable parameters. The input signal at time t is convolved to obtain... The analytic signal at time; the complex Morlet wavelet kernel function includes two parallel convolution channels, one for the real part and one for the imaginary part; Based on modulus activation pairs The modulus of the analytic signal at time t is taken to obtain The characteristic sequence of time.

[0011] Furthermore, the specific method for aggregating the feature sequences at different times is as follows: based on The probability density of the input signal at time t is used to aggregate the feature sequences at different times to obtain aggregated features; the probability density is a trainable parameter.

[0012] Furthermore, the hybrid confidence gating mechanism module includes a physical branch, a metacognitive branch, and a data-driven branch; The physics branch is used to construct a probabilistic physical mask based on bearing geometry and estimated rotational speed; the metacognitive branch is used to calculate a confidence scalar based on aggregated features; and the data-driven branch is used to extract high-confidence feature vectors based on aggregated features and the confidence scalar. The fusion processing of aggregated features includes: Obtain the bearing geometry and estimated rotational speed, and construct a probabilistic physical mask based on the physical branch; The aggregated features are input into the metacognitive branch, and the confidence scalar is calculated based on the metacognitive branch. The aggregated features are input into the data-driven branch, and high-confidence feature vectors are extracted based on the data-driven branch.

[0013] Furthermore, the specific method for constructing a probabilistic physical mask based on physical branches is as follows: Theoretical fault frequencies are calculated based on bearing geometry and estimated rotational speed. A Gaussian mixture model (GMM) is used to model the frequency distribution of theoretical fault characteristics, and a probabilistic physical mask is obtained.

[0014] Furthermore, the specific method for calculating the confidence scalar based on metacognitive branches is as follows: The aggregated features are input into a global average pooling layer to obtain a global feature vector; the global feature vector is then input into a nonlinear gated layer that includes a learnable Gaussian radial basis function, and finally outputs a confidence scalar through a Sigmoid function layer.

[0015] Furthermore, the specific method for extracting high-confidence feature vectors based on data-driven branching is as follows: The aggregated features are input into the data attention generation network. By learning the correlation of feature channels in different frequency bands, the data attention weights are obtained. The probabilistic physical mask and the data attention weights are weighted and summed based on the confidence scalar to obtain the hybrid attention weight matrix. The hybrid attention weight matrix and the aggregated features are multiplied element-wise, and then the feature is aggregated with the aggregated features through residual connections to obtain a high-confidence feature vector.

[0016] Furthermore, the specific method for predicting high-confidence feature vectors is as follows: A parameterized Gaussian radial basis function is used as a learnable activation function to perform nonlinear mapping and impact aggregation on high-confidence feature vectors, resulting in high-fidelity manifold features. The high-fidelity manifold features are input into a fully connected mapping layer and a Softmax activation function to calculate the predicted probability distribution of the input signal belonging to various bearing health states. The category corresponding to the maximum probability is taken to obtain the predicted wind turbine bearing fault type.

[0017] The beneficial effects of adopting the above technical solution are as follows: This invention provides a dual Gaussian bias fault diagnosis method for noisy wind turbine bearings. The proposed dual Gaussian inductive bias mechanism utilizes complex wavelet-locked resonance bands for noise reduction at the front end and employs Gaussian radial basis functions to construct high-dimensional decision boundaries at the back end, thereby effectively suppressing complex background interference at wind power sites and accurately capturing weak early fault features even with extremely low signal-to-noise ratios. By introducing a hybrid confidence gating mechanism and adaptive generalized mean pooling, the model can dynamically evaluate the quality of the input signal and adaptively adjust the fusion weights of the physical prior mask and data features. This process breaks the "black box" limitation of traditional deep learning, giving diagnostic decisions clear physical meaning and statistical basis. Combining a parameterized front end and an optimized Kolmogorov-Arnold network (FastKAN), this invention effectively overcomes the computational redundancy problem caused by standard spline networks. This architecture achieves high-resolution approximation of the feature space while significantly reducing the number of invalid parameters, improving the overall computational efficiency and deployment convenience of model feature extraction. Attached Figure Description

[0018] Figure 1 Flowchart of the dual Gaussian bias fault diagnosis method for noisy wind turbine bearings provided in Embodiment 1 of the present invention; Figure 2 A schematic diagram of the multi-physics sensing wavelet front-end module provided in Embodiment 1 of the present invention; Figure 3 A schematic diagram of the adaptive impact aggregation module provided in Embodiment 1 of the present invention; Figure 4 A schematic diagram of the hybrid confidence gating mechanism module provided in Embodiment 1 of the present invention; Figure 5 A schematic diagram of the dual Gaussian induction bias module provided in Embodiment 1 of the present invention; Figure 6 The experimental results of the bearing health status monitoring model provided in Embodiment 1 of this invention under different signal-to-noise ratios; Figure 7 The present invention provides confusion matrices based on different models at a signal-to-noise ratio of -6 dB in Embodiment 1, wherein (a) is the confusion matrix based on PI-WKAN at a signal-to-noise ratio of -6 dB, (b) is the confusion matrix based on WDCNN at a signal-to-noise ratio of -6 dB, (c) is the confusion matrix based on ResNet 18 at a signal-to-noise ratio of -6 dB, (d) is the confusion matrix based on CWCNN at a signal-to-noise ratio of -6 dB, (e) is the confusion matrix based on DRSN at a signal-to-noise ratio of -6 dB, and (f) is the confusion matrix based on ViT-1D at a signal-to-noise ratio of -6 dB. Figure 8 The present invention provides t-SNE visualization results based on feature representations learned by different models at a signal-to-noise ratio of -6 dB in Embodiment 1. Among them, (a) is the t-SNE visualization result based on PI-WKAN at a signal-to-noise ratio of -6 dB, (b) is the t-SNE visualization result based on WDCNN at a signal-to-noise ratio of -6 dB, (c) is the t-SNE visualization result based on ResNet 18 at a signal-to-noise ratio of -6 dB, (d) is the t-SNE visualization result based on CWCNN at a signal-to-noise ratio of -6 dB, (e) is the t-SNE visualization result based on DRSN at a signal-to-noise ratio of -6 dB, and (f) is the t-SNE visualization result based on ViT-1D at a signal-to-noise ratio of -6 dB. Figure 9 The present invention provides an analysis of the impact of FastKAN grid size (G) on diagnostic performance and computational complexity at a signal-to-noise ratio of -6 dB, wherein (a) is the accuracy under different grids, (b) is the recall under different grids, (c) is the number of parameters under different grids, and (d) is the inference time under different grids. Figure 10 A schematic diagram of the evolution trajectory of probability density p during the training process provided in Embodiment 1 of the present invention. Detailed Implementation

[0019] The specific implementation methods of this application will be further described in detail below with reference to the accompanying drawings and embodiments.

[0020] Example 1: A dual Gaussian bias fault diagnosis method for high-noise wind turbine bearings, such as Figure 1 As shown, it includes the following steps: Step 1: Collect vibration signals, normalize and segment the vibration signals to construct a time series sample set; Vibration signals are collected using sensors installed at key nodes of the wind turbine drivetrain. Considering the cost and difficulty of acquiring vibration signals during actual monitoring, appropriate sensors need to be selected. A large number of vibration signals will significantly reduce the model training speed. Therefore, this embodiment uses only a single type of vibration signal for condition monitoring. Commonly used signal acquisition sensors include temperature sensors, acceleration sensors, and acoustic emission sensors. This embodiment selects an acceleration sensor as the single signal acquisition device for monitoring the health status of wind turbine bearings. From a cost perspective, accelerometers are relatively inexpensive, offering a significant price advantage compared to composite sensors that measure multiple physical quantities such as vibration, temperature, and acoustic emission, thus greatly reducing the deployment cost of monitoring systems. Regarding signal effectiveness, accelerometers can directly capture high-frequency vibration signals of 1-10kHz caused by bearing failures, accurately covering the characteristic frequencies of inner ring, outer ring, and rolling element failures. For example, monitoring the characteristic frequency of outer ring failures can help determine the extent of outer ring damage. Considering ease of installation, accelerometers can be quickly fixed to the bearing housing end face via magnetic attraction or bolts, eliminating the need for complex installation processes and additional coupling agents, effectively reducing installation difficulty and time costs. Furthermore, accelerometers provide stable output signals and strong anti-interference capabilities, ensuring data quality even in complex electromagnetic environments, providing a reliable basis for subsequent condition monitoring.

[0021] The accelerometer is fixed to the spindle bearing housing in both horizontal and vertical directions, preferably near the end face of the bearing's outer ring. This position minimizes vibration transmission path loss and directly acquires the vibration signal generated by the bearing's operation. The accelerometer is installed using either magnetic attachment or bolts. Magnetic attachment facilitates quick installation and removal, suitable for temporary monitoring or equipment debugging; bolt fixing provides a more secure connection, suitable for long-term online monitoring scenarios, ensuring tight coupling between the accelerometer and the bearing housing and preventing signal distortion due to loosening.

[0022] The sampling frequency was set to 2.56kHz, which not only fully captures the fault characteristics in the vibration signal but also effectively reduces the amount of data collected, lowering storage and transmission pressure. The sampling interval was set to 5 minutes, with each 1-second vibration signal sample collected as one data set. By covering multiple operating states, the collected data ensures a comprehensive reflection of the bearing's vibration characteristics under different working conditions, improving the adaptability and generalization ability of subsequent monitoring models.

[0023] Based on the variable speed and strong noise characteristics of wind power sites, the vibration signal is normalized and segmented to obtain multiple time series samples and construct a time series sample set, including time series samples of various fault types and different noise levels. A corresponding operating condition label is added to each time series sample. The fault types include at least inner ring, outer ring, rolling element faults and normal conditions.

[0024] Step 2: Preprocess the time series sample set constructed in Step 1, and divide the preprocessed time series sample set into a training set and a test set; Preprocessing includes handling outlier data and data normalization; In this embodiment, abnormal data processing is implemented using a median filtering method based on an adaptive sliding window. An adaptive sliding window is set, and median filtering is performed on the time series samples within the adaptive sliding window. Abnormal mutation data in the time series samples are detected within the adaptive sliding window. After filtering, the abnormal mutation data is replaced with normal data, while other normal signal data remains unaffected. Data normalization is performed on the dataset after processing outlier data. In order to eliminate the influence of units of measurement, accelerate model convergence and improve training stability, this embodiment uses Z-Score Normalization to process the time series samples after processing outlier data.

[0025] Step 3: Construct the bearing health status monitoring model PI-WKAN based on the dual Gaussian inductive bias, which includes the multi-physical sensing wavelet front-end module, the adaptive impact aggregation module, the hybrid confidence gating mechanism module, and the dual Gaussian inductive bias module set in sequence. The multi-physics sensing wavelet front-end module is used to perform preliminary noise reduction on input signals at different times and extract feature sequences; The adaptive impact aggregation module is used to aggregate sequence features from different time periods to obtain aggregated features; The hybrid confidence gating mechanism module is used to process aggregated features and extract high-confidence feature vectors; The dual Gaussian inductive bias module is used to predict high-confidence feature vectors to obtain the predicted wind turbine bearing fault types.

[0026] Step 3.1: Perform preliminary noise reduction on the input signals at different times based on the multi-physics sensing wavelet front-end module to obtain the feature sequence; In this embodiment, the preprocessed time series sample set in the training set is used as the input signal to the bearing health status monitoring model. The multi-physical sensing wavelet front-end module is used to perform physical layer envelope demodulation and preliminary noise reduction, extract high signal-to-noise ratio fault envelope features, and obtain feature sequences. In the multi-physics sensing wavelet front-end module, the convolution kernel is parameterized as a complex Morlet wavelet kernel function, as shown in the following formula: ; in, For bandwidth, For the center frequency, This is the normalization constant.

[0027] Complex Morlet wavelet kernel function The convolution is decomposed into two parallel channels: a real part and an imaginary part, as shown in the following formula: ; in, These are the in-phase components, i.e., the real part convolution channels; These are orthogonal components, i.e., the imaginary part of the convolution channel; Unlike wavelet transforms with fixed parameters, the bandwidth of this embodiment... and center frequency These are trainable parameters. During initialization, a Mel-scale distribution is used to cover the entire frequency domain, simulating the human ear's frequency perception characteristics. During backpropagation, the center frequency is driven by gradient descent. It automatically finds the resonant frequency band where the energy is most concentrated in the signal, while the bandwidth... It will automatically adjust to match the duration of the impact. This mechanism allows the network to automatically search for the optimal demodulation frequency band. To ensure the numerical stability of the training, this embodiment employs a Logit spatial parameterization strategy to ensure the center frequency. Always within the Nyquist frequency range, and bandwidth It is always a positive value.

[0028] To achieve physical envelope demodulation, modulus activation is used to extract feature sequences. The core theoretical basis of this design lies in the analytic property of the Morlet wavelet. Since the spectral energy of the Morlet wavelet is mainly concentrated on the positive half-axis, for… Input signal at time The convolution operation with complex Morlet kernels essentially constructs the frequency band within that band. Analytic Signal at Time: ; in, for Analyzing the signal at any given moment; imaginary unit ,express The extracted components are orthogonal components; Based on this, the activation pair based on the modulus value Analytical signal at time Modulus extraction is performed to extract high signal-to-noise ratio fault envelope features, resulting in a feature sequence as shown in the following formula: ; in, for The feature sequence of time; It is the numerical stability constant; In signal processing, the standard method for extracting the envelope signal is to construct an analytic signal. .in This is the Hilbert transform. This embodiment implements a parameterized form of this process using complex Morlet convolution.

[0029] When the center frequency of the filter When aligning the resonant frequencies of the system, the complex convolution operation actually outputs the analytic signal within that frequency band. Taking the modulus of the analytic signal is mathematically rigorously equivalent to Hilbert demodulation. In this embodiment, the high-frequency resonant band characteristics are losslessly demodulated to the baseband through an explicit physical layer, generating an envelope signal containing the fault frequency. This avoids the information loss and resonance paradox caused by downsampling, providing neural networks with more discriminative low-frequency envelope features. The overall structure is as follows: Figure 2 As shown.

[0030] Step 3.2: Input the feature sequences at different times into the adaptive shock aggregation module, and retain the key transient shock components while compressing the dimensions through learnable generalized mean pooling; From the perspective of statistical signal processing, Input signal at time The probability density function follows a generalized Gaussian distribution (GGD), as shown in the following formula: ; in, It is a shape parameter. When When, probability density The bearing is in a healthy state if the distribution is Gaussian; when When, probability density The bearing is in a fault impact state due to the tail section distribution.

[0031] Building upon this, this embodiment introduces adaptive generalized mean pooling (GeM) to aggregate sequence features at different time points, resulting in aggregated features. As shown in the formula below: ; in, The length is the time dimension. Let be the probability density.

[0032] The key to this embodiment lies in the probability density. It is a learnable parameter, not a fixed hyperparameter. This is achieved by using the probability density... Setting these as trainable parameters enables the adaptive impact aggregation module to possess physical sensing capabilities, i.e., automatically adjusting the aggregation strategy based on the state of the feature sequence: in a healthy state, the bearing health status monitoring model tends to learn smaller... value( At this point, the bearing health condition monitoring model focuses on overall energy fluctuations to avoid misjudgments caused by individual random noises. During fault conditions, as training progresses, the bearing health condition monitoring model will automatically learn larger... The value. This evolution process shows that the bearing health condition monitoring model identifies fault characteristics as sparse, narrow pulses. By adaptively approximating the statistical order of the kurtosis, the model successfully isolates weak periodic impacts from the strong noise background. The adaptive impact aggregation module structure is as follows: Figure 3 As shown.

[0033] Step 3.3: Aggregate features The input hybrid confidence gating mechanism module dynamically calculates the weights of the physical mask and data attention based on the signal characteristics, realizing the adaptive fusion of physical priors and data features to obtain a high-confidence feature vector; The hybrid confidence gating mechanism module includes a physical branch, a metacognitive branch, and a data-driven branch. The physical branch is used to construct a probabilistic physical mask based on the bearing geometry and estimated rotational speed. The metacognitive branch is used to calculate the confidence scalar based on aggregated features. The data-driven branch is used to calculate the data-driven attention weights based on aggregated features. The confidence scalar is used as the weight coefficient, and the probabilistic physical mask and the data-driven attention weights are weighted and fused to obtain the hybrid attention weight matrix output by the hybrid confidence gating mechanism module. Relying solely on physical models carries the risk of parameter drift, while relying solely on data-driven models lacks robustness under strong noise. To overcome this duality, this embodiment introduces a hybrid confidence gating mechanism. By quantifying the uncertainty of the model regarding physical priors, it achieves a dynamic interplay and optimal fusion of physical mechanisms and data features.

[0034] First, a soft constraint based on physical mechanisms is constructed, rather than a rigid hard threshold mask. The theoretical fault frequency is then calculated based on the bearing geometry and estimated rotational speed. And construct a probabilistic physical mask. To address speed fluctuations in scenarios without a tachometer or under unsteady operating conditions, a Gaussian mixture model (GMM) is used to model the frequency distribution of theoretical fault characteristics, resulting in a probabilistic physical mask. As shown in the formula below: ; in, For frequency variables, For the first The standard deviation of the Gaussian window at the first harmonic. The total order of the fault characteristic harmonics; the mask consists of a series of fundamental frequencies located at the fault frequency and their... The Gaussian window is composed of the subharmonics.

[0035] This embodiment does not blindly trust probabilistic physical masks, but instead designs a gated sub-network within the metacognitive branch. Specifically, the gated sub-network first aggregates features... The input is fed into a global average pooling layer, where aggregation is performed over the time dimension to obtain a global feature vector representing the overall energy distribution and signal-to-noise ratio of the signal. Subsequently, this global feature vector is directly input into a nonlinear gated layer composed of learnable Gaussian radial basis functions. Finally, a Sigmoid function layer restricts the output to the [0,1] interval, thereby calculating and generating a dynamic confidence scalar. ; The aggregated features extracted by the multi-physics sensing wavelet front-end and adaptive impact aggregation module As a data-driven branch, the data attention generation network learns the correlation between feature channels in different frequency bands to obtain data attention weights. ; Based on confidence scalar For probabilistic physical masks and data attention weights By performing a weighted summation, we obtain the hybrid attention weight matrix. As shown in the formula below: ; Furthermore, the calculated hybrid attention weight matrix and aggregation features Perform element-wise multiplication and aggregate features through residual connections. Feature aggregation is performed to generate a high-confidence feature vector that combines physical constraints with data feature adaptability. This mechanism endows the bearing health condition monitoring model with metacognitive capabilities, enabling it to assess the reliability of physical priors. This is achieved by scaling the confidence scalar... The analysis reveals the dynamic decision-making logic of the model under different signal-to-noise ratios (SNR): When the input signal-to-noise ratio is higher than the preset threshold and the rotational speed is stable within the preset range, the fault characteristics match the theoretical frequency well, and the bearing health condition monitoring model will automatically adjust the confidence scalar. Push to 1. At this point, the probabilistic physical mask, as a strong prior, acts like a bandpass filter to filter out irrelevant noise and pinpoint fault characteristics.

[0036] When encountering complex faults, strong nonlinear disturbances, or drastic speed fluctuations that cause the physical model to fail, the network will perceive the mismatch between the physical mask and the actual features, thereby reducing the confidence scalar. The value approaches 0. At this point, it automatically switches to data-driven mode, relying on the data-driven attention weights learned by the neural network. To conduct a diagnosis.

[0037] This dynamic weighting mechanism not only resolves the conflict between rigid physical models and flexible data models, but also achieves this through an explicit confidence scalar. This achieves interpretability of the model's decision-making process and strikes an optimal balance between robustness and adaptability. The hybrid confidence gating mechanism module structure is as follows: Figure 4 As shown.

[0038] Step 3.4: Use the high-confidence feature vector as the input of the dual Gaussian inductive bias module, and use the Gaussian radial basis function to construct a high-dimensional decision boundary to accurately predict the fault type of wind turbine bearings; This embodiment is based on the FastKAN architecture and high-confidence feature vectors. Before entering the FastKAN classification layer, the data is first processed by Layer Norm (layer normalization); subsequently, the feature stream is absorbed and decomposed by parallel activation layers. Unlike traditional multilayer perceptrons (MLPs) that rely on fixed nonlinear activation functions, this module uses a parameterized Gaussian radial basis function (RBF) as a learnable activation function. The basic components, mathematically expressed as: ; in, These are learnable weight coefficients. For the first Each RBF can learn combination coefficients For the total number of RBF, For the first The center position of each Gaussian radial basis function As a smoothing factor, These are high-confidence feature vectors. This is a layer normalization operation. Through weighted summation using a self-learning Gaussian kernel, the nonlinear mapping of features in the latent manifold space is completed. The core of this process lies in its physical and mathematical isomorphism: the Gaussian RBF basis function used in this classifier, together with the Gaussian envelope Morlet wavelet used in the front end of step 3.1, constitutes an equivalent and symmetrical double Gaussian inductive bias in the entire network architecture. This Gaussian isomorphic design, echoing each other at the beginning and end, not only maximizes the transmission and reconstruction efficiency of transient signal impact energy in the latent manifold feature space, enabling the model to generate an S-shaped high-dimensional decision boundary with extremely high curvature, but also completely decouples the weak early fault features severely entangled under strong background noise, effectively eliminating the spectral bias defect of high-frequency feature loss in traditional deep learning models when processing low signal-to-noise ratio impact signals. Unlike MLP, this embodiment uses the FastKAN architecture, parameterizing the activation function as a learnable nonlinear function on the edges.

[0039] To address the issue of similar faults becoming entangled in the feature space, this embodiment increases the grid density of the radial basis function (RBF) to G = 32, expanding the coverage to [-3, 3]. This high-resolution design allows the model to approximate complex decision boundaries with extremely high curvature on the feature manifold. More importantly, the Morlet wavelet (Gaussian window) at the front end and the RBF (Gaussian window) at the back end mathematically constitute a double Gaussian inductive bias at both ends. This symmetry maximizes the transmission efficiency and representational consistency of signal features throughout the network.

[0040] After multiple nonlinear transformations by the FastKAN network, the extracted features are input into a fully connected mapping layer and a Softmax activation function to calculate the predicted probability distribution of the input signal belonging to various bearing health states. The category corresponding to the maximum probability is taken to obtain the final predicted wind turbine bearing fault type. The overall structure is shown in Figure 5.

[0041] Step 4: Train the bearing health status monitoring model using the training set and test it using the test set to obtain the trained bearing health status monitoring model. The purpose of training the bearing health condition monitoring model is to optimize the network parameters of the bearing health condition monitoring model using the training dataset, so that it can achieve better performance on the training set, and at the same time, make the network generalize to other data outside the training set as much as possible.

[0042] This embodiment acquires datasets for experiments under six typical variable operating conditions. Detailed dataset information is shown in Table 1. Based on high-frequency sampling at 64 kHz, weak early features in the broadband spectrum are preserved.

[0043] Table 1. Detailed information about the dataset

[0044] Labeled wind turbine bearing samples were used to train the PI-WKAN model end-to-end. During the testing phase, a noisy vibration signal under unknown operating conditions was input, and the network output the final assessment results of the fault category and degradation degree.

[0045] To comprehensively evaluate the superiority of the bearing health condition monitoring model, it was compared with several mainstream deep learning-based fault diagnosis methods. Experiments were conducted on these methods under noise levels of 6dB, 4dB, 2dB, 0dB, -2dB, -4dB, and -6dB. The results are as follows: Figure 6 As shown.

[0046] To deeply analyze the model's decision-making behavior and feature learning capabilities under harsh conditions, such as Figure 7 and Figure 8 As shown, Figure 7 (a) is the confusion matrix based on PI-WKAN at a signal-to-noise ratio of -6 dB. Figure 7 (b) is the confusion matrix based on WDCNN at a signal-to-noise ratio of -6 dB. Figure 7 (c) is the confusion matrix based on ResNet 18 at a signal-to-noise ratio of -6 dB. Figure 7 (d) is the confusion matrix based on CWCNN at a signal-to-noise ratio of -6 dB. Figure 7 (e) is the confusion matrix based on DRSN at a signal-to-noise ratio of -6 dB. Figure 7 (f) is the confusion matrix based on ViT-1D at a signal-to-noise ratio of -6 dB; Figure 8 (a) Visualization of t-SNE results based on PI-WKAN at a signal-to-noise ratio of -6 dB. Figure 8 (b) Visualization results of t-SNE based on WDCNN at a signal-to-noise ratio of -6 dB. Figure 8 (c) Visualization results of t-SNE based on ResNet 18 at a signal-to-noise ratio of -6 dB. Figure 8 (d) shows the t-SNE visualization results based on CWCNN at a signal-to-noise ratio of -6 dB. Figure 8 (e) shows the t-SNE visualization results based on DRSN at a signal-to-noise ratio of -6 dB. Figure 8(f) shows the t-SNE visualization results based on ViT-1D at a signal-to-noise ratio of -6 dB. This embodiment jointly analyzes the confusion matrix and t-SNE feature visualization under a strong noise environment of -6 dB. The results show that the bearing health status monitoring model exhibits excellent diagonal dominance and manifold separability. Even for early faults with weak features (KA30 / H3), it still maintains a high recall rate of 90.66%. The extracted features form a highly dense cluster with clear boundaries and compact intra-class features in two-dimensional space.

[0047] The grid size G in the FastKAN classifier determines the network's approximation resolution in the feature space. To investigate the impact of G on the ability to separate complex manifolds under strong noise, this embodiment tested the performance of different G values ​​on a dataset with a signal-to-noise ratio of -6 dB, and the results are shown in Table 2.

[0048] Table 2. Comparison of model performance at different mesh sizes G with a signal-to-noise ratio of -6 dB.

[0049] Figure 9 This paper analyzes the impact of FastKAN grid size (G) on diagnostic performance and computational complexity at a signal-to-noise ratio of -6 dB, where (a) represents accuracy under different grid sizes, (b) represents recall under different grid sizes, (c) represents the number of parameters under different grid sizes, and (d) represents inference time under different grid sizes. Figure 9 As shown, when G is small, the model is heavily influenced by spectral bias and tends to learn low-frequency smoothing functions, resulting in an inability to effectively separate severely entangled early fault clusters (such as KA30 and KI04) in the feature space. As G increases to 32, the network obtains sufficient high-frequency basis functions to construct high-curvature decision boundaries, significantly improving accuracy by 12.41%. Although the number of parameters increases slightly, thanks to FastKAN's matrix computation optimization, the inference latency only increases by 1.7ms, meeting industrial real-time requirements. This confirms that G=32 is a key threshold for achieving high-fidelity approximation of the back-end manifold in double Gaussian inductive bias.

[0050] Learnable parameters introduced by the adaptive GeM pooling layer It acted as a statistical filter. Figure 10 illustrates the training process. The evolution trajectory of the value. Experimental results show that the adaptive pooling parameter p always converges within the range of 3 ± 0.2. From the perspective of statistical signal processing, this value is consistent with the inherent kurtosis characteristics of bearing fault signals. Compared with the feature dilution caused by global average pooling (GAP, p=1) and the sensitivity of max pooling (GMP, p→∞) to noise, p≈3 provides a cubic weighted nonlinear mechanism that can enhance the sparse fault impact component while suppressing Gaussian background noise. This adaptive convergence behavior indicates that PI-WKAN can autonomously learn that the sparsity in the signal is a key statistical invariant for fault diagnosis.

[0051] Step 5: Use the trained condition monitoring model to test the health status of the bearing to be monitored; Step 6: Visualize and make decisions based on the predicted bearing failure conditions; Due to the harsh operating environment of wind turbines, the diagnostic results of a single time window may be affected by sudden strong impacts, resulting in fluctuations. To further improve the reliability of the diagnostic system, this embodiment introduces an interpretable visualization module at the output of the bearing health condition monitoring model, including the following: (1) Dynamic decision confidence display: Real-time output and visualization of confidence scalar The trajectory of change intuitively reflects whether the model's current decision-making tends to rely on physical prior masks or purely data-driven features; (2) Physical statistical parameter evolution tracking: learnable probability density in adaptive generalized mean pooling layers Dynamic tracking was performed to verify whether the model accurately captured the sparsity and kurtosis characteristics of bearing failure impact; (3) Display of high-dimensional feature manifold distribution: Using feature dimensionality reduction algorithms such as t-SNE, the high-dimensional features extracted by the FastKAN backend network are mapped to a two-dimensional visualization plane, which intuitively presents the clustering effect and manifold boundary of different fault categories in the feature space under strong noise environment; (4) Comprehensive confusion matrix evaluation: The diagnostic confusion matrix is ​​generated by combining the prediction results of multiple consecutive time windows to smooth the risk of misjudgment caused by a single burst of noise, and to provide on-site operation and maintenance personnel with transparent and reliable multi-dimensional diagnostic decision-making basis.

[0052] Example 2: This embodiment proposes an electronic device, including: one or more processors, and a memory for storing instructions. When the instructions are executed by the one or more processors, the one or more processors cause the one or more processors to perform the dual Gaussian bias fault diagnosis method for noisy wind turbine bearings.

[0053] The electronic device may be a mobile phone, computer, or tablet computer, etc., and includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, implements the dual Gaussian bias fault diagnosis method for noisy wind turbine bearings as described in the embodiments. It is understood that the electronic device may also include input / output (I / O) interfaces and communication components.

[0054] The processor is used to execute all or part of the steps in the dual Gaussian bias fault diagnosis method for noisy wind turbine bearings as described in the above embodiments. The memory is used to store various types of data, which may include, for example, instructions for any application or method in the electronic device, as well as application-related data.

[0055] The processor can be implemented as an Application Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), controller, microcontroller, microprocessor, or other electronic components, and is used to execute the dual Gaussian bias fault diagnosis method for noisy wind turbine bearings described in the above embodiments.

[0056] Example 3: This embodiment proposes a computer-readable storage medium that stores executable instructions. When these instructions are executed, if they are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium.

[0057] The computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the dual Gaussian bias fault diagnosis method for noisy wind turbine bearings described in the various embodiments of this application.

[0058] The aforementioned storage media include: flash memory, hard disks, multimedia cards, card-type memory (e.g., SD (Secure Digital Memory Card) or DX (Memory Data Register, MDR) memory), random access memory (RAM), static random-access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic storage, disks, optical discs, servers, APP (Application) application stores, and other media capable of storing program verification codes. These media store computer programs, which, when executed by a processor, can implement the various steps of the aforementioned dual Gaussian bias fault diagnosis method for noisy wind turbine bearings.

[0059] Example 4: This embodiment proposes a computer program product, including a computer program or instructions, which, when executed by a processor, implements the aforementioned dual Gaussian bias fault diagnosis method for noisy wind turbine bearings.

[0060] Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a computer program product.

[0061] The various embodiments in this application are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0062] The scope of protection of this application is not limited to the embodiments described above. Obviously, those skilled in the art can make various modifications and variations to this disclosure without departing from the scope and spirit of this disclosure. If such modifications and variations fall within the scope of this disclosure and its equivalents, then the intent of this disclosure also includes these modifications and variations.

Claims

1. A dual Gaussian bias fault diagnosis method for high-noise wind turbine bearings, characterized in that, Includes the following steps: A dataset is constructed based on vibration signals collected by sensors installed on wind turbine units; A bearing health status monitoring model is constructed, which includes a multi-physical sensing wavelet front-end module, an adaptive impact aggregation module, a hybrid confidence gating mechanism module, and a dual Gaussian induction bias module set in sequence. The bearing health status monitoring model is trained and validated based on the dataset to obtain a trained bearing health status monitoring model; the health status of the target wind turbine bearing is monitored based on the trained status monitoring model. The multi-physical sensing wavelet front-end module is used to perform preliminary noise reduction on the input signals of the bearing health status monitoring model at different times to obtain feature sequences at different times; The adaptive impact aggregation module is used to aggregate feature sequences at different times to obtain aggregated features; The hybrid confidence gating mechanism module is used to fuse aggregated features to obtain high-confidence feature vectors; The dual Gaussian inductive bias module is used to predict high-confidence feature vectors to obtain the predicted wind turbine bearing fault types.

2. The method for diagnosing dual Gaussian bias faults in high-noise wind turbine bearings according to claim 1, characterized in that, The specific method for constructing the dataset is as follows: The vibration signal is normalized and segmented to obtain multiple time series samples and construct a time series sample set; the time series sample set includes time series samples of various fault types and different noise levels, and the fault types include at least inner ring, outer ring, rolling element faults and normal state. Preprocessing of time series samples includes outlier handling and data normalization; the outlier handling includes setting an adaptive sliding window and performing median filtering on the time series samples within the adaptive sliding window. Training and test sets are constructed based on the preprocessed time series samples.

3. The method for diagnosing dual Gaussian bias faults in high-noise wind turbine bearings according to claim 1, characterized in that, The specific method for preliminary noise reduction of the input signals at different times of the input bearing health status monitoring model is as follows: The convolution kernel is parameterized as a complex Morlet wavelet kernel function that includes trainable parameters. The input signal at time t is convolved to obtain... The analytic signal at time; the complex Morlet wavelet kernel function includes two parallel convolution channels, one for the real part and one for the imaginary part; Based on modulus activation pairs The modulus of the analytic signal at time t is taken to obtain The characteristic sequence of time.

4. The method for diagnosing dual Gaussian bias faults in high-noise wind turbine bearings according to claim 1, characterized in that, The specific method for aggregating feature sequences at different times is as follows: based on The probability density of the input signal at time 1 is used to aggregate the feature sequences at different times to obtain aggregated features; The probability density is a trainable parameter.

5. The method for diagnosing dual Gaussian bias faults in high-noise wind turbine bearings according to claim 1, characterized in that, The hybrid confidence gating mechanism module includes a physical branch, a metacognitive branch, and a data-driven branch; The physics branch is used to construct a probabilistic physical mask based on bearing geometry and estimated rotational speed; the metacognitive branch is used to calculate a confidence scalar based on aggregated features; and the data-driven branch is used to extract high-confidence feature vectors based on aggregated features and the confidence scalar. The fusion processing of aggregated features includes: Obtain the bearing geometry and estimated rotational speed, and construct a probabilistic physical mask based on the physical branch; The aggregated features are input into the metacognitive branch, and the confidence scalar is calculated based on the metacognitive branch. The aggregated features are input into the data-driven branch, and high-confidence feature vectors are extracted based on the data-driven branch.

6. The method for diagnosing dual Gaussian bias faults in high-noise wind turbine bearings according to claim 5, characterized in that, The specific method for constructing a probabilistic physical mask based on physical branches is as follows: Theoretical fault frequencies are calculated based on bearing geometry and estimated rotational speed. A Gaussian mixture model (GMM) is used to model the frequency distribution of theoretical fault characteristics, and a probabilistic physical mask is obtained.

7. The method for diagnosing dual Gaussian bias faults in high-noise wind turbine bearings according to claim 5, characterized in that, The specific method for calculating the confidence scalar based on metacognitive branches is as follows: The aggregated features are input into a global average pooling layer to obtain a global feature vector; the global feature vector is then input into a nonlinear gated layer that includes a learnable Gaussian radial basis function, and finally outputs a confidence scalar through a Sigmoid function layer.

8. The method for diagnosing dual Gaussian bias faults in high-noise wind turbine bearings according to claim 5, characterized in that, The specific method for extracting high-confidence feature vectors based on data-driven branching is as follows: The aggregated features are input into the data attention generation network. By learning the correlation of feature channels in different frequency bands, the data attention weights are obtained. The probabilistic physical mask and the data attention weights are weighted and summed based on the confidence scalar to obtain the hybrid attention weight matrix. The hybrid attention weight matrix and the aggregated features are multiplied element-wise, and then the feature is aggregated with the aggregated features through residual connections to obtain a high-confidence feature vector.

9. The method for diagnosing dual Gaussian bias faults in high-noise wind turbine bearings according to claim 1, characterized in that, The specific method for predicting high-confidence feature vectors is as follows: A parameterized Gaussian radial basis function is used as a learnable activation function to perform nonlinear mapping and impact aggregation on high-confidence feature vectors, resulting in high-fidelity manifold features. The high-fidelity manifold features are input into a fully connected mapping layer and a Softmax activation function to calculate the predicted probability distribution of the input signal belonging to various bearing health states. The category corresponding to the maximum probability is taken to obtain the predicted wind turbine bearing fault type.

Images