A wind turbine fault diagnosis method and system based on Mamba-ResNet

Abstract

The present application relates to the technical field of fault diagnosis, in particular to a wind driven generator fault diagnosis method and system based on Mamba-ResNet. The method comprises feature extraction and feature fusion using preprocessed data, including constructing an adaptive window short-time Fourier transform (AW-STFT) for dynamic time-frequency resolution analysis, parallel feature extraction and construction of a multi-dimensional heterogeneous feature vector, and a cross-modal adaptive gating fusion mechanism based on a bidirectional cross-gating unit; a Mamba-ResNet hybrid deep network model architecture is constructed; the constructed network model architecture is subjected to model training; and the trained model architecture is used for wind driven generator fault diagnosis. The method avoids the cumbersome manual feature design process in traditional methods, and significantly improves the automation level and adaptability of the diagnosis system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fault diagnosis, in particular to a wind generator fault diagnosis method and system based on Mamba-ResNet. BACKGROUND

[0002] With the trend of large-scale, intelligent, deep-sea and high-altitude complex environment site development of wind turbine generators, the reliability of its core transmission chain is facing unprecedented challenges. Bearings, as the key rotating components connecting blades, gearboxes and generators, have long been subjected to alternating loads, complex stresses and environmental erosion. If the initiation and propagation of micro-damage cannot be detected in time, it will easily cause a chain failure, ultimately leading to unplanned shutdown of the whole machine, causing huge direct economic losses and secondary safety risks. Therefore, the health status of bearings directly determines the safety and stability of the wind turbine generator during the whole life cycle and the economy.

[0003] Currently, the mainstream bearing fault diagnosis method in the industry still highly depends on the traditional paradigm of "signal processing + expert experience + shallow model". Specifically, first, engineers or algorithms manually design and extract time-domain statistical features (such as mean, variance, peak value, kurtosis, pulse factor, waveform factor, etc.) or frequency-domain features (such as characteristic frequency, sideband, harmonic component, spectral kurtosis, etc.) of the vibration signal according to physical priori. Then, these hand-crafted features are input into classic machine learning classifiers such as support vector machine (SVM), random forest (Random Forest), k-nearest neighbor (k-NN) for pattern recognition and state determination. Although this paradigm has been proven effective under certain conditions, its inherent technical limitations are increasingly evident in the face of modern and complex wind power operation and maintenance needs:

[0004] (1) Feature extraction is highly dependent on expert prior knowledge, and has insufficient self-adaptation and generalization ability. The effectiveness of traditional methods largely depends on the quality of feature engineering, which requires diagnosis experts to have a deep understanding of the fault mechanism, vibration mechanics and signal characteristics of bearings. This "manual design-fixed rule" mode has essential defects: on the one hand, it cannot automatically mine the most discriminative deep abstract features from massive, high-dimensional, non-stationary raw vibration signals; on the other hand, when faced with new fault patterns (such as fatigue fracture of new materials) or multiple fault coupling composite faults that have not appeared in the training data, the pre-defined feature set often fails, leading to a sharp decline in model generalization performance, making it difficult to achieve intelligent diagnosis of "one place learning, multiple applications".

[0005] (2) Weakness in modeling the time-series dynamic characteristics of vibration signals. In recent years, deep learning methods based on convolutional neural networks (CNN) have been introduced, which can automatically learn spatial features from time-frequency images (such as wavelet scale maps, short-time Fourier transform spectrograms). However, these methods mainly focus on extracting spatial patterns from local receptive fields, and their essence is a static analysis of signal "snapshots". The development of bearing faults is a typical dynamic process, and the vibration signals contain long-range time-series dependencies and evolution laws related to the rotation period, fault impact period, and modulation effects. Existing CNN methods are limited by their structural characteristics and cannot effectively capture long-range time-series context information spanning thousands of samples, making it difficult to comprehensively and coherently represent the dynamic evolution process from fault generation, development, to deterioration, limiting the detection of early weak faults and the accurate assessment of fault severity.

[0006] (3) Simple and inefficient fusion mechanism for multi-source heterogeneous information, failing to achieve information complementarity and synergistic enhancement. Bearing vibration signals are a natural multi-modal information carrier: time-domain signals contain impact strength and statistical distribution information, frequency-domain signals reveal the characteristic frequency components of faults, and time-frequency domain signals can depict the changes in frequency components over time (i.e., non-stationary characteristics). In addition, synchronous signals from different directions (such as horizontal and vertical) of the same bearing seat provide complementary spatial vibration information. However, most current advanced methods still rely on simple vector concatenation or early fusion of features from different sources. This "cassette" processing approach ignores the inherent, complex correlations and differences between different modal features, failing to establish an adaptive, interactive, and interpretable multi-scale cross-modal feature fusion mechanism. As a result, information utilization is low, and even noise may be introduced due to dimensional differences or information redundancy between features, making it impossible to achieve a "1+1>2" synergistic diagnostic effect.

[0007] (4) Lack of robustness to variable operating conditions in actual operation, and the model is prone to performance drift. The actual operating conditions of wind turbine generators are complex and variable, with fluctuations in wind speed directly causing changes in rotational speed, and changes in load due to grid scheduling requirements. Changes in rotational speed and load will directly cause significant shifts in bearing vibration signal amplitude, frequency, and even statistical distribution, a phenomenon known as "field shift" or "operating condition shift". Traditional diagnostic models usually assume that training data and test data follow the same independent and identically distributed assumption. Under severe variable operating conditions, this assumption no longer holds. The decision boundary learned by the model under fixed operating conditions often fails when faced with new operating condition data with different distributions, leading to a significant decrease in diagnostic accuracy and a lack of long-term deployment robustness and stability in actual wind field variable environments. How to make the diagnostic model have the ability to "draw a conclusion from one case and apply it to another" is a core problem that must be overcome to move from the laboratory to engineering applications.

[0008] Therefore, there is an urgent need for a next-generation wind turbine bearing intelligent fault diagnosis system that can adaptively extract deep features, accurately model timing dynamics, intelligently fuse multi-source information, and robustly adapt to changing working conditions. SUMMARY

[0009] To solve the problems of existing wind turbine bearing fault diagnosis methods, such as feature extraction relying on manual, insufficient timing dynamic modeling, insufficient multi-source information fusion, poor variable working condition adaptability, and weak model interpretability, the present application provides a wind turbine fault diagnosis method and system based on Mamba-ResNet. Through key technologies such as dual-channel vibration signal input, multi-scale feature automatic extraction, time-space feature deep interaction, dynamic gated fusion, and working condition adaptive compensation, high-precision, robust, and interpretable bearing fault diagnosis is achieved.

[0010] In the first aspect, the present application provides a wind turbine fault diagnosis method based on Mamba-ResNet, which adopts the following technical solution:

[0011] A wind turbine fault diagnosis method based on Mamba-ResNet, comprising:

[0012] Obtain dual-channel vibration signals of the wind turbine;

[0013] Perform data preprocessing based on the obtained dual-channel vibration signals;

[0014] Perform feature extraction and feature fusion using the preprocessed data, including constructing an adaptive window short-time Fourier transform (AW-STFT) for dynamic time-frequency resolution analysis, parallel feature extraction, and constructing a multi-dimensional heterogeneous feature vector, and a cross-modal adaptive gating fusion mechanism based on a bidirectional cross-gating unit;

[0015] Construct a Mamba-ResNet hybrid deep network model architecture;

[0016] Perform model training on the constructed network model architecture;

[0017] Use the trained model architecture to perform wind turbine fault diagnosis.

[0018] In the second aspect, a wind turbine fault diagnosis system based on Mamba-ResNet is provided, comprising:

[0019] A data acquisition module configured to acquire dual-channel vibration signals of the wind turbine;

[0020] A preprocessing module configured to perform data preprocessing based on the acquired dual-channel vibration signals;

[0021] The feature module is configured to perform feature extraction and feature fusion by using the preprocessed data, and includes the following steps: constructing an adaptive window short-time Fourier transform (AW-STFT) to perform dynamic time-frequency resolution analysis, performing parallel feature extraction and constructing a multi-dimensional heterogeneous feature vector, and performing cross-modal adaptive gating fusion based on a bidirectional cross-gating unit.

[0022] The model construction module is configured to construct a Mamba-ResNet hybrid deep network model architecture.

[0023] The model training module is configured to perform model training on the constructed network model architecture.

[0024] The diagnosis module is configured to perform wind turbine fault diagnosis by using the trained model architecture.

[0025] In a third aspect, the present application provides a computer-readable storage medium, wherein a plurality of instructions are stored in the computer-readable storage medium, and the instructions are adapted to be loaded and executed by a processor of a terminal device to implement the Mamba-ResNet-based wind turbine fault diagnosis method.

[0026] In a fourth aspect, the present application provides a terminal device, comprising a processor and a computer-readable storage medium, wherein the processor is used to implement instructions, and the computer-readable storage medium is used to store a plurality of instructions, and the instructions are adapted to be loaded and executed by the processor to implement the Mamba-ResNet-based wind turbine fault diagnosis method.

[0027] In summary, the present application has the following beneficial technical effects:

[0028] (1) End-to-end intelligent feature learning is realized, and the dependence on expert experience is reduced: multi-scale features with strong discriminability are automatically learned from original vibration signals by a deep neural network, avoiding the cumbersome manual feature design process in traditional methods, and significantly improving the automation level and adaptability of the diagnosis system.

[0029] (2) Deep information fusion improves feature expression ability: time domain, frequency domain, time-frequency domain features and spatial, time sequence features of deep learning are organically fused, multi-dimensional information in the vibration signal is fully utilized, and the discriminability and robustness of the features are significantly enhanced.

[0030] (3) Efficient time sequence modeling capability to capture the dynamic evolution law of the fault: the Mamba model is introduced to handle the long sequence dependence relationship in the vibration signal, overcoming the training difficulty of traditional RNN and the limited receptive field of CNN, and effectively capturing the dynamic evolution process of the fault features.

[0031] (4) Strong robustness of variable working condition adaptability, improve the practical application value: through the working condition perception embedding and feature distribution alignment technology, the influence of working condition change on feature distribution is modeled explicitly, so that the diagnostic model still maintains stable performance in variable working condition environment, and the application value in actual wind field is improved.

[0032] (5) Enhanced diagnostic explainability, improve system credibility: provide multi-level visualization tools and attribution analysis methods, so that the operation and maintenance personnel can intuitively understand the diagnostic basis of the model, enhance the credibility and acceptability of the diagnostic results, and provide strong support for predictive maintenance decision.

[0033] (6) Modular design, easy to extend and maintain: the present application adopts modular design, and each functional module is relatively independent, which is convenient for adjustment and optimization according to actual demand, and is also convenient for system maintenance and upgrading.

[0034] (7) Efficient online diagnosis capability, meet the real-time monitoring demand: through model optimization and deployment strategy, millisecond level fault diagnosis response is realized, which meets the demand of real-time monitoring and early warning of wind farm, and provides time window for timely maintenance. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 is a flow chart of a wind turbine fault diagnosis method based on Mamba-ResNet implemented by the present application;

[0036] Figure 2 is a model diagram of a wind turbine fault diagnosis method based on Mamba-ResNet of the present application embodiment 1;

[0037] Figure 3 is a comparison test accuracy comparison result of a wind turbine fault diagnosis method based on Mamba-ResNet implemented by the present application;

[0038] Figure 4 is a comparison test F1 score comparison result of a wind turbine fault diagnosis method based on Mamba-ResNet implemented by the present application;

[0039] Figure 5 is the model training and test loss of a wind turbine fault diagnosis method based on Mamba-ResNet implemented by the present application;

[0040] Figure 6 is the model accuracy transformation curve of a wind turbine fault diagnosis method based on Mamba-ResNet implemented by the present application. DETAILED DESCRIPTION

[0041] The present application will be further described in detail below in combination with the drawings.

[0042] Embodiment 1

[0043] Reference Figure 1 The wind turbine fault diagnosis method based on Mamba-ResNet in the embodiment comprises:

[0044] The overall flow thereof is as shown in Figure 1 , and comprises the following steps:

[0045] S1: Synchronous acquisition and adaptive preprocessing enhancement of double-channel vibration signals

[0046] S1.1 Multi-directional synchronous acquisition and high-fidelity digitization To ensure comprehensive capture of the three-dimensional spatial vibration form of the bearing of the wind turbine under complex load, high-dynamic-range and low-noise IEPE type piezoelectric acceleration sensors are respectively arranged in the horizontal radial (X-axis) and vertical axial (Y-axis) two mechanically orthogonal directions of the bearing seat. The two sensors are microsecond-level synchronous acquisition through hardware triggering, and record vibration acceleration time domain signals and .

[0047] The sampling frequency is set to 25,600 Hz. This parameter is determined according to the following principles: first, the Nyquist sampling theorem needs to be met, and the fault characteristic frequency of each element (inner ring, outer ring, rolling body, retainer) of the bearing and its high-order harmonics (usually more than 10 times the bearing pass frequency) can be acquired without distortion; second, high sampling rate helps to accurately capture the details of the instantaneous impact waveform caused by local peeling, cracks and other defects, providing high-resolution input for subsequent time-frequency analysis and deep learning models.

[0048] S1.2 Adaptive normalization based on sliding window

[0049] Due to the time-varying nature of the load and speed of the wind turbine, the overall energy distribution of the vibration signal will fluctuate. In order to eliminate the influence of the difference in signal amplitude scale under different working conditions on model training and retain the relative structural information inside the signal, sliding window local normalization is adopted instead of global normalization. For each signal channel , at time , the mean and standard deviation of the signal in the window with length before and after it are calculated, and normalized:

[0050] ,

[0051] wherein, is a numerical stability constant introduced to prevent division by zero. This method can adaptively eliminate the baseline drift caused by slow-changing working conditions, making the model more focused on the fault characteristics contained in the local fluctuations of the signal.

[0052] S1.3 Dynamic data augmentation strategy oriented to physical mechanism

[0053] To improve the robustness of the model to interference in the real environment and the generalization ability to unknown working conditions, the normalized signal Implement dynamic data augmentation based on physical constraints. This strategy introduces random perturbations in the time domain that conform to actual physical laws, rather than simply data augmentation.

[0054] (1) Adaptive signal-to-noise ratio noise injection: In order to simulate actual interference such as sensor electronic noise and background mechanical noise, and to avoid noise masking real fault impacts, an adaptive noise injection method based on local signal energy is used. Specifically, the noise intensity is related to the energy (variance) of the signal in the local time window:

[0055] ,

[0056] where is a random intensity factor, is the signal variance in the sliding window centered at .

[0057] (2) Elastic time warping: nonlinear stretching of signal timing to simulate speed fluctuations:

[0058] ,

[0059] where is the warping amplitude, is the signal length.

[0060] S2: Adaptive extraction and fusion of multi-scale time-frequency features

[0061] S2.1 Construct adaptive window short-time Fourier transform (AW-STFT) to realize dynamic time-frequency resolution analysis

[0062] In order to overcome the inherent defect of traditional short-time Fourier transform that cannot simultaneously consider high time resolution and high frequency resolution under fixed window, the invention constructs an adaptive window short-time Fourier transform. The core is to use a time-varying Gaussian window function, whose standard deviation The dynamic adjustment can be made according to the instantaneous physical characteristics of the vibration signal. Specifically, the system calculates the instantaneous frequency change rate of the signal in real time: when analyzing the transient impact component in the signal caused by local defects (such as pitting and spalling) of the bearing (which is manifested as high frequency and short time width), the algorithm automatically adopts a narrower time window to sacrifice part of the frequency resolution, accurately locates the time of impact occurrence, and obtains a time-frequency spectrum with high time resolution; on the contrary, when analyzing the smooth modulation component or harmonic component caused by bearing wear and misalignment, the algorithm automatically switches to a wider frequency window to clearly separate the dense spectral lines and obtain high frequency resolution. This strategy of 'windowing according to the material' ensures that the generated two-dimensional time-frequency diagram can simultaneously depict the transient impact timing and the steady-state spectral structure of the fault signal with optimal resolution, providing an information density maximized underlying representation for subsequent visual feature extraction.

[0063] ,

[0064] wherein is a time-varying Gaussian window function, and the standard deviation is adaptively adjusted according to the instantaneous frequency change rate of the signal: a narrow window is used for high-frequency transient components to improve time resolution, and a wide window is used for smooth components to improve frequency resolution.

[0065] During the long-term operation of the wind turbine bearing, different fault types have significant differences in the energy distribution patterns in the time-frequency domain. Typical local damage faults (such as pitting and spalling) usually exhibit short-time high-frequency impact signals, and their energy is highly concentrated in several local areas in the time-frequency plane; while progressive faults such as wear and gap enlargement tend to exhibit energy diffusion characteristics with wide time periods and low contrast. It is difficult to effectively distinguish the above different fault patterns by relying only on traditional global time-frequency energy or statistical features.

[0066] Therefore, in the multi-scale time-frequency feature extraction stage, the present application introduces a time-frequency impact focusing degree index , which is used to quantitatively represent the local concentration degree of the energy of the vibration signal in the time-frequency plane, and its definition is shown in the following formula. This index compares the energy proportion in the impact significant area with the overall time-frequency energy distribution to depict whether the fault energy presents a 'highly focused' or 'dispersed and expanded' structural feature.

[0067] ,

[0068] wherein, represents a time-frequency amplitude matrix obtained by adaptive short-time Fourier transform or continuous wavelet transform; Impulse dominant region automatically identified by energy mutation detection or threshold discrimination method. By accumulating the square amplitude, the sensitivity to high-energy impact components can be enhanced, and the interference of weak noise on index calculation can be avoided.

[0069] When the bearing is in the early stage of local damage, the impact energy is concentrated in a limited time-frequency region, making the focusing degree index Take higher values; while in the stage of accelerated wear or compound failure, the energy tends to be dispersed in the time-frequency plane, The value of the index decreases accordingly. Therefore, this index can effectively reflect the type characteristics and evolution stage of bearing failure.

[0070] In model implementation, the time-frequency impact focusing degree index can be input as an explicit auxiliary feature into the ResNet spatial feature extraction branch, or used to modulate the weight of the multi-scale convolution feature channel, so as to guide the network to pay more attention to the time-frequency region with high impact and high discriminability, and improve the perception ability of the model to weak fault features and the discrimination ability to different fault modes.

[0071] S2.2 Build a multi-dimensional heterogeneous feature vector to realize the feature cooperation of physical statistics and deep learning

[0072] From the above time-frequency representation and original signal, three types of complementary feature vectors are extracted in parallel to build a feature pool that combines prior knowledge and data-driven:

[0073] Time domain statistical feature vector: directly from the original vibration signal And its first-order difference Second-order difference A set of 24-dimensional classical statistics are calculated. These features include but are not limited to: mean value describing the distribution center of the signal, standard deviation describing the fluctuation intensity, skewness describing the asymmetry of the distribution, kurtosis describing the sharpness of the distribution, and root mean square (RMS), peak factor, pulse factor, waveform factor, etc. These features have clear physical meaning and can stably represent the amplitude distribution, impact intensity and waveform shape of the signal, providing verified and interpretable physical insight basis for the model.

[0074] ,

[0075] Frequency band energy entropy feature vector: to quantify the frequency domain distribution pattern of fault energy, the power spectral density of the signal is divided into Non-uniform frequency bands with clear physical direction. The division is based on the combination of the geometric parameters of the bearing and the theoretical fault characteristic frequency, such as: low frequency band related to bearing pass frequency, medium frequency band related to rolling element fault characteristic frequency, high frequency band related to structural resonance high frequency modulation, etc. The Energy proportion of each frequency band The energy proportion of each frequency band is calculated, and the Shannon entropy of the energy proportion is further calculated. The entropy value can effectively measure the concentration or dispersion degree of the fault energy in the frequency spectrum: a low entropy value indicates that the fault energy is concentrated in a few frequency bands (such as a single impact feature), and a high entropy value indicates that the energy is dispersed (such as a composite fault or severe wear). This feature provides a key frequency domain distribution fingerprint for distinguishing different fault modes.

[0076] ,

[0077] Different fault types often exhibit different energy distribution characteristics in different frequency bands.

[0078] Singular value feature vector of time-frequency matrix: the time-frequency map is regarded as a two-dimensional matrix , and singular value decomposition (SVD) is performed on it. The singular value sequence represents the projection intensity of the time-frequency map energy in each orthogonal direction, and its distribution characteristics reflect the stability and complexity of the time-frequency structure. The first principal singular values are taken to form the feature vector. This feature highly compresses the global structural information of the time-frequency map and can effectively represent the "track" pattern formed by the fault impact on the time-frequency plane, such as discrete point-like tracks (corresponding to single impact) or continuous band-like tracks (corresponding to modulation phenomena), thereby providing a macroscopic and structured time-frequency pattern description for the model.

[0079] S3: Constructing Mamba-ResNet hybrid deep network architecture

[0080] S3.1 Multi-scale convolution attention module (MS-ConvFormer): realizing spatial multi-granularity feature focusing

[0081] In view of the different spatial scales that the fault features in the time-frequency image may exhibit (such as fine short-time-frequency patterns and wide harmonic bands), this module designs a parallel multi-branch convolution structure. This structure uses 3×3, 5×5, and 7×7 convolution kernels of different sizes to perform parallel convolution operations on the input time-frequency map, so as to simultaneously capture local fine-grained textures, medium-range patterns, and global context information. Subsequently, an adaptive recalibration of the fused multi-scale features is performed through a channel attention gating mechanism. This mechanism first aggregates the global information of each feature channel through global average pooling (GAP), then learns the importance weight of each channel through a two-layer fully connected network containing a ReLU activation function, and finally generates attention coefficients between 0 and 1 through function. This process enables the network to automatically suppress irrelevant or noisy channels and strengthen the feature channels that are crucial to the current fault diagnosis task, realizing adaptive focusing and fusion of multi-scale spatial features.

[0082] ,

[0083] Channel attention mechanism is introduced to automatically learn the weights of different scale features:

[0084] ,

[0085] where GAP denotes global average pooling, is the ReLU activation function, is function. S3.2 Improved ResNet spatial feature extractor: Constructing hierarchical deep spatial representation

[0086] On the basis of the primitive features extracted by MS-ConvFormer, a deep residual network (ResNet) is introduced as the backbone extractor of spatial features. We made two key improvements to the standard ResNet to adapt to the fault diagnosis task: First, a three-stage progressive network is constructed, each stage containing multiple improved residual blocks, which gradually down-samples the spatial size of the feature map and expands the channel dimension from 64 to 128, 256, thereby realizing the hierarchical abstraction from low-level edge texture to middle-level component pattern to high-level fault semantics. Second, an attention module is integrated inside each residual block, which dynamically adjusts the feature response of each channel by explicitly modeling the interdependence between channels, thereby enhancing the information interaction and collaboration between feature channels, allowing the network to learn more discriminative channel combinations. Finally, the network outputs multi-level feature maps , which together constitute a complete and robust deep representation of faults in the time-frequency space.

[0087] S3.3 Mamba temporal dynamic modeling module: Realize long-range selective temporal dependence capture

[0088] To model the long-range dynamic evolution rule spanning thousands of sampling points in the vibration signal, this module introduces the latest structured state-space sequence model (Mamba). The core idea is to regard the one-dimensional time series vibration signal as an observation of a continuous-time system, and model it through the following selective state-space equation: ,

[0089] where is a diagonal state matrix, and are projection matrices. Through the discretization method, the continuous system is converted into a discrete form to realize efficient calculation.

[0090] ​To enhance the model's ability to perceive the early fault impact characteristics of the bearing and avoid excessive memory of irrelevant stationary signals during long-term modeling, the application introduces a selective memory modulation mechanism based on fault impact intensity in the Mamba time series modeling module.

[0091] Specifically, first, statistical analysis is performed on the input vibration signal within a local time window, and the deviation of the current time signal from its local mean is calculated, and the deviation is normalized by the local standard deviation . The fault impact perception coefficient is defined by equation (1) to quantify the significance of potential fault impact in the current vibration signal:

[0092] ,

[0093] On this basis, as shown in equation (2), the impact perception coefficient is introduced into the state update process of Mamba to dynamically modulate the fusion ratio between the historical state and the current candidate state. Specifically, is used to control the memory strength of the model for the current time information, and is used to control the retention degree of the historical state, thereby achieving key memory of fault-related key moments and rapid forgetting of stationary redundant information.

[0094] Through the above mechanism, the model can maintain the advantages of Mamba long-term time series modeling while explicitly introducing the bearing fault physical mechanism constraint, so that the time series modeling process pays more attention to the impact events closely related to fault evolution, effectively improves the detection ability of weak faults and early abnormalities, and enhances the robustness and stability of the model under complex working conditions.

[0095] ,

[0096] The key innovation of Mamba lies in its selective mechanism and efficient hardware perception design. First, its system parameters (input matrix), (output matrix), and most importantly, the step size parameter are no longer fixed, but are adjusted according to the current input Dynamic generation via linear projection. This means the model can "selectively" remember or ignore historical information: for input segments containing critical fault impacts, the system extends the memory (increasing the effective state window); for stationary or noisy segments, it quickly forgets. This input-dependent selectivity allows it to precisely focus on sparse but critical fault events in the signal. Secondly, through ingenious discretization and parallel scanning algorithms, Mamba achieves training parallel efficiency and inference complexity comparable to Transformer with linear sequence lengths while maintaining theoretically unlimited memory capacity. This enables it to efficiently process high-frequency vibration sequences of wind turbine bearings that last for several seconds and have tens of thousands of sampling points, accurately characterizing long-range time-series patterns such as the periodicity of fault impacts and the evolution of modulation phenomena.

[0097] ,

[0098] This input-dependent parameterization enables the model to focus on key temporal patterns, significantly improving its ability to model long-term dependencies in vibration signals.

[0099] S4: Cross-modal adaptive gating fusion mechanism

[0100] S4.1 Bidirectional Cross-Gated Unit: Realizing Dynamic Dialogue and Collaborative Enhancement Based on Spatiotemporal Features

[0101] This step aims to break down the modal barriers between spatial features (from ResNet, representing the static morphology in time-frequency images) and temporal features (from Mamba, representing the dynamic evolution of vibration signals), rather than simply splicing them together. To this end, we designed a bidirectional cross-gated unit, the core idea of ​​which is to allow the features of the two modes to guide and enhance each other.

[0102] Specifically, the unit receives spatial feature vectors from ResNet. and temporal feature vectors from Mamba The system runs two gating networks in parallel: the first gating network uses temporal features as context to compute a gating weight applied to spatial features. The second gating network uses spatial features as context to compute a gating weight applied to temporal features. These two weight vectors are obtained through... The function is compressed to the [0,1] interval, and its value reflects the strength of the "enhancement proposal" of another modality feature on the current feature channel.

[0103] Subsequently, a bidirectional enhancement operation is performed: spatial features and their corresponding gating weights. Element-wise multiplication yields a new spatial feature enhanced by temporal information; simultaneously, the temporal feature is coupled with its corresponding gating weight. Element-wise multiplication, resulting in a new temporal feature that is guided by the spatial information for enhancement. This process simulates a deep "dialogue" between the two modalities - for example, a persistent high-energy region in a certain frequency band of the spectrogram (spatial feature) can hint that the Mamba module should pay more attention to the periodic impulses in the signal during the corresponding time period (temporal feature enhancement); conversely, a strong sequence of transient impulses in the signal (temporal feature) can hint that the ResNet should focus on the local patterns in the spectrogram at the corresponding time and frequency (spatial feature enhancement). Ultimately, these two interactively enhanced features will be fused to provide a unified representation that is complementary in information and internally consistent for subsequent decision-making.

[0104] ,

[0105] where the degree of enhancement of the spatial features over the temporal features, controls the degree of enhancement of the temporal feature on the spatial feature.

[0106] To further constrain the physical consistency of the temporal feature and the spatial feature in the time scale, introduce the temporal-spatial consistency energy constraint as a regularization term in the fusion stage.

[0107] ,

[0108] where: denotes the temporal feature output by Mamba at time . is the spatial feature extracted by ResNet in the corresponding time-frequency region. is the spatial-temporal alignment projection operator (such as time pooling + linear mapping)

[0109] This energy term is used to measure the degree of consistency between the temporal dynamic pattern and the spatial time-frequency pattern in the same time scale. When the bearing actually fails, the time-domain impulse sequence and the time-frequency energy accumulation region should be highly consistent in time. This formula explicitly constrains this physical consistency.

[0110] S4.2 Multi-expert dynamic fusion network: implementation of task-oriented adaptive feature synthesis strategy

[0111] Considering different fault types (such as pitting, wear, and composite faults) or different working condition stages, their effective diagnostic logic and dependent feature combinations may be different. In order to simulate this diversified decision path, we construct a multi-expert dynamic fusion network. This network contains parallel "expert" sub-networks (for example, ).

[0112] ,

[0113] Each expert is a small individual neural network trained to learn and excel at a specific, latent fault-feature mapping relationship. For example, Expert A might excel at diagnosing early pitting corrosion by analyzing high-frequency transient features; Expert B might excel at diagnosing wear by analyzing low-frequency modulated features; and Expert C might excel at handling complex faults by synthesizing various features. These experts are differentiated and complementary.

[0114] The key innovation lies in a lightweight routing network. This network takes the aforementioned fused features as input, analyzes the feature properties of the current sample in real-time, and dynamically computes a weight vector assigned to each expert . This weight vector is normalized by a function, representing the degree to which each expert’s output should be trusted or emphasized in the current diagnostic decision. Ultimately, the network’s output is the weighted sum of all expert network outputs. This mechanism enables the model to adaptively select the most suitable feature synthesis and reasoning path according to the characteristics of the input signal itself, rather than forcing the use of the same fixed fusion rules for all samples, greatly enhancing the model’s flexibility and targeted diagnostic ability for different fault modes.

[0115] ,

[0116] S5: Condition-adaptive feature calibration and robust classification

[0117] S5.1 Condition-aware embedding learning: Encoding environmental context as part of features

[0118] The rotational speed of the wind turbine generator set and the load are the core operating condition parameters that affect the performance of vibration signals. In order to eliminate the signal distribution shift caused by changes in operating conditions, we take these continuous operating condition parameters as explicit inputs to the model. Through a dedicated embedding layer, the rotational speed and load are mapped to a low-dimensional, dense continuous vector condition. This operating condition embedding vector encodes the context information of the current operating environment.

[0119] ,

[0120] The operating condition embedding not only serves as additional information input into the network, but also dynamically modulates the feature channels through a condition-sensitive gating function .

[0121] ,

[0122] where is the fused feature, To embed the speed-load operating condition, denotes the characteristic time rate of change, Dynamic modulation coefficient

[0123] S5.2 Feature distribution alignment: driving model to learn operating condition invariant intrinsic fault features

[0124] It is not enough to simply append operating condition information as input. To make the model truly focus on the operating condition independent, pure fault features, we introduce a domain adaptation loss, Maximum Mean Discrepancy (MMD), in the training process. The core goal is to force the sample distributions from different operating conditions (e.g., high speed vs. low speed) but the same fault type to be as close as possible in the feature space. Maximum Mean Discrepancy (MMD) minimizes the distance between the feature distributions under training and testing conditions:

[0125] ,

[0126] where is the feature mapping function, is the Reproducing Kernel Hilbert Space.

[0127] MMD achieves this goal by calculating the distance between the mean values of sample features under different operating conditions in the Reproducing Kernel Hilbert Space (RKHS). By minimizing this distance, we are actually constraining the feature extractor to ignore the changes caused by operating conditions and retain and strengthen the changes caused by the fault itself. This makes the feature representation learned by the model invariant to operating condition changes, so that it can maintain stable diagnostic performance when facing new operating conditions that have not appeared in training.

[0128] S5.3 Multi-task collaborative optimization: driving main task performance improvement with auxiliary task To further stabilize training and improve the discriminability of features, we design a multi-task learning framework. The model performs an auxiliary task (operating condition identification) in parallel with the main task (fault classification), i.e., simultaneously predicting the current speed and load operating condition.

[0129] The two tasks share most of the feature extraction layers but have independent final classification heads. By jointly optimizing a joint loss function that combines the fault classification loss (classification), the operating condition identification loss (condition), and the aforementioned MMD distribution alignment loss (MMD), the model is guided to learn a rich feature representation that is highly sensitive to fault categories and implicitly reflects operating condition information. The auxiliary task acts as a powerful regularizer that prevents the network from overfitting to specific operating condition combinations in the training data on the fault classification task, and encourages the feature extractor to build a more comprehensive and more generalizable internal representation, ultimately benefiting and improving the classification robustness of the main task.

[0130] ,

[0131] In addition, we adopted label smoothing technique in the classification loss of the main task. This technique replaces the original "hard" labels (such as class [0, 1, 0, 0]) with "soft" labels (such as [0.025, 0.925, 0.025, 0.025]). This practice introduces a slight "uncertainty" into model training, effectively preventing the model from being overly confident in its predictions of training samples, thereby alleviating overfitting and improving the model's calibration and generalization ability under noise interference.

[0132] ,

[0133] where is the smoothed label distribution.

[0134] S6: Model training strategy and optimization method

[0135] S6.1 Progressive training strategy: unlock network in stages to achieve stable optimization

[0136] Given the complexity of the model structure and the large number of parameters, direct end-to-end training can easily lead to unstable gradients or get stuck in local optima. Therefore, we adopted a three-stage progressive training strategy:

[0137] Feature extractor freezing stage: First, fix the weights of core feature extraction backbone networks such as ResNet and Mamba, making them serve as pre-trained feature extractors. Only the newly added classification head (fully connected layer) at the end of the model is trained. This stage allows the classification head to quickly learn how to use existing features for preliminary classification, while stabilizing the early stages of the entire training process.

[0138] Intermediate layer fine-tuning stage: After the classification head converges, unfreeze the deep layers of the backbone network close to the output end. These layers usually contain more abstract and task-related features. In this stage, the unfrozen layers are fine-tuned together with the classification head to adapt the feature representation to the current diagnostic task.

[0139] Global end-to-end fine-tuning stage: Finally, unfreeze all network layers and perform global end-to-end joint training. Due to the large number of model parameters, we use gradient accumulation technique: accumulate the gradients of multiple small batches and perform weight update again. This is equivalent to simulating large batch training under limited GPU memory, which helps to obtain more stable gradient estimates and better final convergence points.

[0140] S6.2 Adaptive learning rate scheduling: dynamically adjust the optimization step size to guide the model to smooth convergence

[0141] The learning rate is one of the most critical hyperparameters for training deep models. We employ a pre-warming cosine annealing strategy for fine-tuning:

[0142] ,

[0143] in .

[0144] Warm-up period: the initial stage of training Within a certain number of steps, the learning rate starts from a very small initial value. And linearly increasing to the preset maximum value This "warm-up" process avoids the model from experiencing severe oscillations in the early stages due to the high randomness of initial parameters and unstable gradients, thus laying a stable foundation for subsequent training.

[0145] Cosine annealing period: After warm-up, the learning rate follows the trajectory of the cosine function from... Smoothly decay to near The advantage of cosine annealing lies in its smooth decay process, which avoids a step-like abrupt change in the learning rate. This allows the model to finely search around the optimal solution with a very small step size in the later stages of training, thereby achieving better performance and generalization ability.

[0146] This strategy of "warming up" first and then "smoothing down" effectively guides the huge model parameter space, composed of Mamba, ResNet and complex fusion modules, to converge towards the global optimum or near-optimal solution region along a stable and efficient path.

[0147] To ensure that the large parameter space consisting of Mamba, ResNet and complex fusion modules can converge to the optimal solution efficiently and stably, and to avoid overfitting, this method integrates a set of cutting-edge gradient optimization and regularization techniques.

[0148] We employ the AdamW optimizer as the core optimization algorithm. AdamW is an improved version of the classic Adam optimizer, with the key difference being the proper decoupling of weight decay (L2 regularization) from gradient updates. In traditional Adam, the weight decay term is scaled by the adaptive learning rate, which may weaken its regularization effect. AdamW, however, treats weight decay as a separate penalty term applied directly after the weight update step. Its update rule aims to effectively utilize the first moment estimate of the gradient (…). (similar to momentum) and second-order moment estimation ( (used to adaptively adjust the learning rate of each parameter), while using independent decay coefficients. By continuously penalizing large weights, the model is encouraged to learn simpler, more generalizable solutions.

[0149] ,

[0150] where and are the first and second order moment estimates of the gradient, respectively, is the weight decay coefficient.

[0151] In addition, to address the problem of gradient explosion that may occur in deep neural networks, we implemented gradient clipping before each parameter update. Specifically, we calculated the norm of the gradient vector of all trainable parameters in the entire model, and if the norm exceeded a pre-set threshold (clip_norm), all gradients were scaled proportionally so that their norm was equal to the threshold. This operation is like installing a "stabilizer" for the training process, which can effectively prevent the dramatic fluctuations of individual batches of abnormal data or network deep back propagation, ensuring the numerical stability of the training process, especially for the Mamba module that handles long sequences.

[0152] S7: Fault diagnosis decision and explainability analysis

[0153] S7.1 Multi-dimensional confidence evaluation: from "single prediction" to "trusted decision"

[0154] To convert the "black box" output of the model into a decision basis that can be trusted by operation and maintenance personnel, we went beyond the single maximum probability value and built a multi-dimensional confidence evaluation system. This system integrates information from three levels to calculate the final confidence score:

[0155] Prediction probability level: the maximum probability value of the model Softmax output is the basis for confidence.

[0156] Feature space level: calculate the separation index. This index quantifies the distance between the current sample's feature vector and the prototype feature center of the fault class , as well as the distance from other class centers. If the current sample features are close to the class center they belong to and far from other class centers, the separation degree is high and the confidence is high; if they fall into the fuzzy overlap region of the feature space, the separation degree is low and the confidence is low.

[0157] Time series level: introduce historical consistency consideration. By comparing the consistency degree of the current diagnosis result with the recent historical diagnosis results (such as predictions within the sliding window), the reliability of the current diagnosis is evaluated. A stable and continuous fault warning is more credible than an isolated and jumping abnormal point.

[0158] where the separation index is defined as:

[0159] ,

[0160] By weighted and fused scores from these three dimensions, the system outputs not just a fault label, but an intelligent diagnostic report with a quantified confidence score. Low-confidence diagnoses automatically trigger a review process (such as requiring manual intervention or initiating a more refined secondary analysis), thereby significantly improving the reliability and usability of the entire monitoring system.

[0161] S7.2 Explainability Visualization Tools: A Series of Keys to Unlocking the "AI Black Box"

[0162] To build user trust and facilitate in-depth analysis, we provide a multi-layered set of interpretable visualization tools:

[0163] Gated weight heatmap: This visually displays features from different sources, including the time domain, frequency domain, spatial ResNet branch, and temporal Mamba branch, in a specific diagnostic case, as well as the contribution weights of each expert in the multi-expert fusion network to the final decision. Operations personnel can clearly see at a glance which information the model primarily "referenced" for this decision.

[0164] Attention distribution map:

[0165] Timing Attention: This demonstrates the intensity of attention the Mamba module pays to different time steps when processing vibration signals, and can pinpoint the impact moments or modulation periods that the model determines to be critical.

[0166] Spatial / Frequency Domain Attention: This section demonstrates the regions of focus for the attention modules in MS-ConvFormer and ResNet on time-frequency images, revealing the key frequency bands and time segments that the model considers to be related to faults.

[0167] Feature space dimensionality reduction visualization: Using t-SNE or UMAP algorithms, the high-dimensional fused features learned by the model are projected onto a two-dimensional plane. Points of different colors / shapes represent different fault states. This image can be used to macroscopically evaluate model performance: the denser the clustering of each category and the clearer the boundaries between categories, the stronger the model's feature learning and separation capabilities; it can also be used to analyze specific misclassified samples to observe whether they fall into the wrong category cluster or are in a region with ambiguous boundaries.

[0168] Decision attribution analysis (such as Grad-CAM): For time-frequency plot input, Grad-CAM can generate a heatmap, highlighting the regions in the time-frequency plot that have the greatest influence on the model's current fault decision. This can directly "map" the fault diagnosis results back to the original physical signal representation, achieving "pixel-level" attribution and greatly assisting in fault root cause analysis.

[0169] Experimental verification and performance evaluation

[0170] To verify the effectiveness of the method, we conducted comprehensive experiments on the data set collected from an actual wind farm. The data set contains 4 types of bearing faults (cage fault, inner ring fault, outer ring fault, and inner and outer ring composite fault) under 3 different working conditions (35Hz / 12kN, 37.5Hz / 11kN, 40Hz / 10kN), with a total of 15 bearings' complete life cycle monitoring records.

[0171] Comparative method

[0172] To verify the effectiveness of the Mamba-ResNet fusion model proposed in this paper in the fault diagnosis of wind turbine gearboxes, three representative deep learning models are selected as benchmark comparison methods:

[0173] (1) CAAN model (Cross-domain Attention Adaptive Network): Using cross-domain attention and adaptive learning, the problem of insufficient generalization ability of fault diagnosis model under variable working conditions is solved.

[0174] (2) EMD-1DCNN model: Through empirical mode decomposition, the vibration signal is automatically decomposed into multiple channel intrinsic mode functions, and then 1D convolutional neural network is used for end-to-end fault classification, effectively dealing with the diagnosis challenge under non-stationary working conditions.

[0175] (3) LMD-MSE-SVM model: Through adaptive signal decomposition, multi-scale complexity features are extracted, and support vector machine is used to realize high-precision fault classification.

[0176] (4) VibrMamba: An efficient long sequence modeling framework based on state space model (such as Mamba) specially designed for vibration signal analysis.

[0177] Evaluation index: Accuracy, F1 score and other indicators are used to evaluate the performance of the model.

[0178] The experimental results are shown in Figure 3 and Figure 4 :

[0179] To verify the effectiveness and superiority of the Mamba-ResNet model proposed in this paper in the fault diagnosis task of wind turbine gearbox, four representative comparative methods are selected for performance comparison experiments, including Emd-1dcnn based on empirical mode decomposition and one-dimensional convolutional neural network, Virmamba based on state space sequence modeling, LMD-MSE-SVM combining local mean decomposition and support vector machine, and multi-channel convolutional attention network CANN. Each model is trained and tested under the same experimental data set and working conditions, and the evaluation indicators are classification accuracy (Accuracy) and F1 score (F1-score). As shown in FIG. 3, Figure 4 Specifically, in terms of accuracy, the recognition rates of Emd-1dcnn, Virmamba, LMD-MSE-SVM and CANN models are 82.75%, 87.60%, 90.00% and 85.70% respectively, while the model proposed in this paper reaches 99.73%, which is about 14.03 percentage points higher than the optimal comparative model. This shows that the hybrid architecture of the proposed model that combines structured state space and residual network has significantly improved feature expression and classification discrimination ability.

[0180] In terms of F1 score, as shown in FIG. 3, the model proposed in this paper also achieves the best performance: the F1 scores of Emd-1dcnn, Virmamba, LMD-MSE-SVM and CANN are 71.59%, 88.41%, 87.67% and 93.10% respectively, while the model proposed in this paper reaches 99.73%. Compared with other methods, the improvement of F1 score shows that the model is more robust in handling the classification balance and recall rate of different categories of samples, especially in the case of uneven sample distribution, it can still maintain high diagnostic accuracy. Figure 5 and Figure 6 The results show that the improved Mamba-ResNet model significantly improves the accuracy of fault diagnosis when dealing with sample data imbalance. Compared with existing methods, the accuracy is improved by an average of 14.03 percentage points, fully verifying the effectiveness and superiority of the proposed method.

[0181] Overall, the Mamba-ResNet model proposed in this paper realizes the multi-modal adaptive fusion of time series, time-frequency and statistical features by introducing the Cross-Gated Fusion mechanism, significantly enhancing the feature interaction and fault pattern recognition ability. In complex scenarios with multiple working conditions, this model not only maintains high accuracy, but also performs outstandingly in feature balance and generalization performance, fully verifying its effectiveness and superiority in the fault diagnosis task of wind turbine gearbox.

[0182] A computer readable storage medium, wherein a plurality of instructions are stored, the instructions being adapted to be loaded by a processor of a terminal device and to execute the Mamba-ResNet-based wind power generator fault diagnosis method.

[0183] A terminal device, comprising a processor and a computer readable storage medium, the processor being configured to implement instructions, and the computer readable storage medium being configured to store a plurality of instructions, the instructions being adapted to be loaded by the processor and to execute the Mamba-ResNet-based wind power generator fault diagnosis method.

[0184] The above are preferred embodiments of the present application, and are not intended to limit the protection scope of the present application, and therefore: any equivalent changes made according to the structure, shape, principle of the present application should be covered within the protection scope of the present application.

Claims

1. A wind turbine fault diagnosis method based on Mamba-ResNet, characterized in that, The method comprises the following steps: acquiring double-channel vibration signals of a wind turbine; performing data preprocessing based on the acquired double-channel vibration signals; performing feature extraction and feature fusion using the preprocessed data, including constructing an adaptive window short-time Fourier transform (AW-STFT) for dynamic time-frequency resolution analysis, parallel feature extraction, and construction of a multi-dimensional heterogeneous feature vector, and a cross-modal adaptive gating fusion mechanism based on a bidirectional cross-gating unit; constructing a Mamba-ResNet hybrid deep network model architecture; training the constructed network model architecture; performing wind turbine fault diagnosis using the trained model architecture. The parallel feature extraction and construction of the multi-dimensional heterogeneous feature vector include parallel extraction of three types of complementary feature vectors from time-frequency representation and original signals, construction of a feature pool that fuses prior knowledge and data-driven features, including time-domain statistical feature vectors, frequency band energy entropy feature vectors and singular value feature vectors of time-frequency matrix ​ The cross-modal adaptive gating fusion mechanism based on the bidirectional cross-gating unit includes receiving a spatial feature vector from a ResNet and a timing feature vector from a Mamba , compressed to the interval [0, 1] by a function, followed by a bidirectional enhancement operation, to obtain new spatial features enhanced by timing information guidance; meanwhile, the timing features are multiplied by the corresponding gating weights, to obtain new timing features enhanced by spatial information guidance, and finally the two features enhanced by interaction will be fused; The constructed Mamba-ResNet hybrid deep network model architecture includes a parallel multi-branch convolution structure MS-ConvFormer designed for different spatial scales of fault feature presentation in time-frequency images, parallel convolution operations on input time-frequency images are performed using convolution kernels of different sizes to simultaneously capture local fine-grained texture, medium-range patterns and global context information, then the fused multi-scale features are adaptively recalibrated through a channel attention gating mechanism, the global information of each feature channel is aggregated through global average pooling GAP, the importance weight of each channel is learned through a two-layer fully connected network containing a ReLU activation function, and finally the output of the network is a multi-level feature map The function generates attention coefficients between 0 and 1, then introduces a channel attention mechanism to automatically learn the weights of different scale features, and finally introduces a deep residual network ResNet as a backbone extractor of spatial features based on the primitive features extracted by MS-ConvFormer, wherein a three-stage progressive network is first constructed, the spatial size of the feature map is down-sampled layer by layer, and the channel dimension is expanded from 64 to 128, 256, realizing hierarchical abstraction from low-level edge texture to middle-level component pattern to high-level fault semantics, secondly, an SE attention module is integrated in each residual block, and finally, the network outputs a multi-level feature map which together constitute a complete and robust deep representation of the fault in the time-frequency space. Also included is introducing a structured state-space sequential model Mamba, taking one-dimensional time-series vibration signals x ( t ) as observations of a continuous-time system and modeling them through selective state-space equations; then introducing a selective memory modulation mechanism based on fault shock intensity in the Mamba time-series modeling module, statistically analyzing the input vibration signals within a local time window, calculating the deviation of the current time signal from its local mean , and combining with the local standard deviation to normalize the deviation and define a fault shock awareness coefficient to quantify the significant degree of potential fault shock in the current time vibration signal, and then introducing the shock awareness coefficient into the state update process of Mamba to dynamically modulate the fusion ratio between the historical state and the current candidate state.

2. The Mamba-ResNet-based wind turbine fault diagnosis method according to claim 1, characterized in that, The data preprocessing based on the acquired raw vibration signals includes, firstly, processing each signal channel... At any moment Calculate the mean of the signal within a window of length 𝑊 before and after. with standard deviation And after normalization, it is expressed as: Then the normalized signal Physically constrained dynamic data augmentation is implemented by employing a noise injection method that adapts to the local energy of the signal to introduce random perturbations in the time domain that conform to actual physical laws. The noise intensity is correlated with the signal energy within a local time window. ,in For random intensity factor, For The signal variance within the centered sliding window is used; and the signal timing is nonlinearly scaled to simulate rotational speed fluctuations. ,in For the amplitude of distortion, This is the signal length.

3. The Mamba-ResNet-based wind turbine fault diagnosis method according to claim 2, characterized in that, The adaptive window short-time Fourier transform (AW-STFT) is constructed for dynamic time-frequency resolution analysis, including constructing an adaptive window short-time Fourier transform, using a time-varying Gaussian window function with a standard deviation According to the instantaneous physical characteristics of the vibration signal, the instantaneous frequency change rate of the signal is calculated in real time, a narrow time window is used when analyzing the transient impact component generated by the local defect of the bearing, a time-frequency spectrum with high time resolution is obtained, and the generated two-dimensional time-frequency diagram simultaneously depicts the transient impact timing and the steady-state spectrum structure of the fault signal with optimal resolution, and is expressed as: , wherein is a time-varying Gaussian window function, whose standard deviation According to the adaptive adjustment of the signal instantaneous frequency rate of change, the high-frequency transient component uses a narrow window to improve the time resolution, and the smooth component uses a wide window to improve the frequency resolution; Then, aiming at the fact that the global statistical features cannot distinguish different fault modes, in the multi-scale time-frequency feature extraction stage, the time-frequency impact focusing degree index is introduced to quantitatively represent the local concentration degree of the vibration signal energy in the time-frequency plane, and by comparing the energy proportion in the impact significant area with the overall time-frequency energy distribution, the structural characteristics of the fault energy are described, and are expressed as: wherein, represents the time-frequency amplitude matrix obtained by adaptive short-time Fourier transform; represents the impact dominant area identified by energy mutation detection.

4. The Mamba-ResNet-based wind turbine fault diagnosis method according to claim 3, characterized in that, The parallel feature extraction and construction of multi-dimensional heterogeneous feature vectors include parallel extraction of three types of complementary feature vectors from time-frequency representation and original signals, and construction of a feature pool that fuses prior knowledge and data-driven features, including time-domain statistical feature vectors , a set of 24-dimensional statistics are calculated directly from the original vibration signals and their first-order differences and second-order differences ; a frequency band energy entropy feature vector is obtained by dividing the power spectral density of the signal into K=4 non-uniform frequency bands with clear physical direction, quantifying the frequency domain distribution pattern of the fault energy, then calculating the energy proportion of each frequency band , and further calculating the Shannon entropy, which measures the concentration and dispersion of fault energy in the frequency spectrum, and is expressed as: ; a time-frequency matrix singular value feature vector is obtained by regarding the time-frequency diagram as a two-dimensional matrix and performing singular value decomposition SVD, using the distribution characteristics of the singular value sequence to reflect the stability and complexity of the time-frequency structure, and taking the first main singular values to construct the feature vector, which represents the trajectory pattern formed by the fault impact on the time-frequency plane.

5. The Mamba-ResNet-based wind turbine fault diagnosis method according to claim 4, characterized in that, The construction of the Mamba-ResNet hybrid deep network model architecture includes designing a parallel multi-branch convolution structure MS-ConvFormer for different spatial scales of fault feature presentation in time-frequency images, using convolution kernels of different sizes to perform parallel convolution operations on the input time-frequency image to simultaneously capture local fine-grained texture, medium-range patterns and global context information, then through the channel attention gating mechanism to adaptively recalibrate the fused multi-scale features, through the global average pooling GAP to aggregate the global information of each feature channel, and through the two-layer fully connected network containing the ReLU activation function to learn the importance weight of each channel, and finally through the softmax function to generate attention coefficients between 0 and 1, represented as: The function generates attention coefficients between 0 and 1, represented as: ; then introduce the channel attention mechanism, automatically learn the weight of different scale features, represented as: wherein denotes global average pooling, is a ReLU activation function, is function; finally, based on the primitive features extracted by the MS-ConvFormer, a deep residual network ResNet is introduced as the backbone extractor of spatial features, wherein, firstly, a three-stage progressive network is constructed, the spatial size of the feature map is down-sampled layer by layer, and the channel dimension is expanded from 64 to 128, 256, realizing the hierarchical abstraction from low-level edge texture to middle-level part pattern to high-level fault semantics, secondly, the SE attention module is integrated in each residual block, the mutual dependence between channels is modeled explicitly, the feature response of each channel is dynamically adjusted, the information interaction and cooperation between feature channels are enhanced, and the network learns more discriminative channel combinations, finally, the network outputs multi-level feature maps which together constitute a complete and robust deep representation of the fault in the time-frequency space.

6. The Mamba-ResNet-based wind turbine fault diagnosis method according to claim 5, characterized in that, The constructed Mamba-ResNet hybrid deep network model architecture further comprises introducing a structured state space sequence model Mamba for modeling long-range dynamic evolution rules in the vibration signal across thousands of sampling points, regarding the one-dimensional time series vibration signal x ( t ) as an observation of a continuous-time system, and modeling it through a selective state space equation, represented as: wherein is a diagonal state matrix, and is a projection matrix; Then to enhance the model's ability to perceive the early fault impact features of bearings and avoid excessive memory of irrelevant stationary signals in long-term modeling, a selective memory modulation mechanism based on fault impact strength is introduced into the Mamba time series modeling module. The input vibration signal is statistically analyzed in a local time window to calculate the deviation of the current time signal from its local mean , and the deviation is normalized by combining the local standard deviation , and the fault impact perception coefficient is defined to quantify the significance of potential fault impact in the current time vibration signal: Then, the impact-aware coefficient is introduced into the state update procedure of Mamba to dynamically modulate the fusion ratio between the historical state and the current candidate state, specifically, is used to control the memory strength of the model on the current time information, while is used to control the retention degree of the historical state, so as to achieve the key memory of the fault-related key moments and the rapid forgetting of the smooth redundant information, and is expressed as: .

7. The Mamba-ResNet-based wind turbine fault diagnosis method according to claim 5, characterized in that, The cross-modal adaptive gating fusion mechanism based on the bidirectional cross-gating unit includes receiving a spatial feature vector from a ResNet and a timing feature vector from a Mamba , compressed to the [0, 1] interval by a function, and then subjected to a bidirectional enhancement operation to obtain new spatial features enhanced by timing information; meanwhile, the timing features are multiplied by the corresponding gating weights, to obtain new timing features enhanced by spatial information; finally, the two features enhanced by interaction are fused and represented as: wherein controls the enhancement degree of the spatial features on the timing features, controls the enhancement degree of the timing features on the spatial features; To constrain the physical consistency of the temporal-spatial features on the time scale, a temporal-spatial consistency energy constraint is introduced , which is expressed as a regularization term in the fusion stage: , wherein: represents the Mamba temporal feature at time , the output temporal feature, is the spatial feature extracted by the ResNet in the corresponding time-frequency region, is a spatial-temporal alignment projection operator; Finally, a multi-expert dynamic fusion network is constructed Each expert is an independent small neural network, taking the fused features as input, and analyzing the feature attributes of the current sample in real time, and dynamically calculating the weight vector assigned to each expert Finally, the network output is the weighted sum of the outputs of all expert networks, represented as: .

8. The Mamba-ResNet-based wind turbine fault diagnosis method according to claim 7, characterized in that, The cross-modal adaptive gating fusion mechanism based on the bidirectional cross-gating unit further includes taking a continuous working condition parameter as an explicit input of the model to eliminate signal distribution deviation caused by working condition changes, mapping the rotating speed and the load to a low-dimensional and dense continuous vector condition through a dedicated embedding layer, and being expressed as: and being dynamically modulated through a working condition sensitive gating function on the feature channels, and being expressed as: among them, is the feature after fusion, is the speed-load operating condition embedding, is expressed as the feature time rate of change, dynamic modulation coefficient; then introduce the maximum mean difference MMD as the domain adaptive loss, use the maximum mean difference MMD to minimize the distance between the feature distribution under the training condition and the test condition, expressed as: , wherein is a characteristic mapping function, is a reproducing kernel Hilbert space; finally, a multi-task learning framework is designed, and a joint loss function combining the fault classification loss, the working condition identification loss, and the MMD distribution alignment loss is optimized jointly, and is expressed as: and uncertainty is introduced using label smoothing method to prevent the model from being overconfident in its prediction probabilities on the training samples to alleviate overfitting, denoted as: , wherein is the smoothed label distribution.

9. The Mamba-ResNet-based wind turbine fault diagnosis method according to claim 1, characterized in that, The model training of the constructed network model architecture comprises a three-stage progressive training strategy, wherein in the feature extractor freezing stage, the weights of the ResNet and Mamba core feature extractor backbone network are fixed as a pre-trained feature extractor, and only the newly added classification head at the end of the model is trained; in the intermediate layer fine-tuning stage, after the classification head converges, the deep network near the output end of the backbone network is unfrozen, and the unfrozen layers and the classification head are fine-tuned together to adapt the feature representation to the current diagnosis task; in the global end-to-end fine-tuning stage, all network layers are unfrozen, and global end-to-end joint training is performed, and a gradient accumulation method is adopted: the gradients of multiple small batches are accumulated, and then the weight update is performed again, wherein the learning rate is one of the most critical hyperparameters for training deep models, and a preheating cosine annealing strategy is used for fine scheduling, which is represented as: , wherein To ensure the large parameter space composed of Mamba, ResNet and complex fusion module converges to the optimal solution stably, gradient optimization and regularization method is adopted, the weight decay and gradient update are correctly decoupled through AdamW optimizer, the first moment estimation and second moment estimation of gradient are used, and the large weight is continuously punished through independent decay coefficient λ, which is expressed as: , wherein and are the first and second order moment estimates of the gradient, respectively, is the weight decay coefficient.

10. A Mamba-ResNet-based wind turbine fault diagnosis system, which executes a Mamba-ResNet-based wind turbine fault diagnosis method according to claim 1, characterized in that, The method comprises the following steps: a data acquisition module configured to acquire double-channel vibration signals of a wind turbine; a preprocessing module configured to perform data preprocessing based on the acquired double-channel vibration signals; a feature module configured to perform feature extraction and feature fusion using the preprocessed data, including constructing an adaptive window short-time Fourier transform (AW-STFT) for dynamic time-frequency resolution analysis, parallel feature extraction, and construction of a multi-dimensional heterogeneous feature vector, and a cross-modal adaptive gating fusion mechanism based on a bidirectional cross-gating unit; a model construction module configured to construct a Mamba-ResNet hybrid deep network model architecture; a model training module configured to train the constructed network model architecture; a diagnosis module configured to perform wind turbine fault diagnosis using the trained model architecture.

Images