Wind driven generator fault diagnosis method and system based on Mamba-ResNet

Abstract

The invention 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 the following steps: carrying out feature extraction and feature fusion by utilizing preprocessed data, namely constructing adaptive window short-time Fourier transform (AW-STFT) to carry out dynamic time-frequency resolution analysis, carrying out parallel feature extraction and constructing a multi-dimensional heterogeneous feature vector, and carrying out a cross-modal adaptive gating fusion mechanism based on a bidirectional cross gating unit; the method comprises the following steps: constructing a Mamba-ResNet hybrid deep network model architecture; performing model training on the constructed network model architecture; and performing fault diagnosis on the wind driven generator by using the trained model architecture. A tedious manual feature design process in a traditional method is avoided, and the automation level and adaptability of a diagnosis system are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology, and in particular to a method and system for fault diagnosis of wind turbines based on Mamba-ResNet. Background Technology

[0002] As wind turbines evolve towards larger, more intelligent systems and are deployed in complex environments such as deep seas and high altitudes, the reliability of their core drivetrain faces unprecedented challenges. Bearings, as critical rotating components connecting blades, gearboxes, and generators, are subjected to alternating loads, complex stresses, and environmental erosion over extended periods. If the initiation and spread of microscopic damage in bearings are not detected in time, they can easily trigger a chain reaction of failures, ultimately leading to unplanned downtime of the entire turbine, resulting in significant direct economic losses and secondary safety risks. Therefore, the health of the bearings directly determines the safe, stable operation and economic efficiency of the wind turbine throughout its entire lifecycle.

[0003] Currently, mainstream bearing fault diagnosis methods in the industry still heavily rely on the traditional paradigm of "signal processing + expert experience + shallow models." Specifically, engineers or algorithms first manually design and extract time-domain statistical features (such as mean, variance, peak value, kurtosis, impulse factor, waveform factor, etc.) or frequency-domain features (such as characteristic frequencies, sidebands, harmonic components, spectral kurtosis, etc. obtained through Fast Fourier Transform) of vibration signals based on physical priors. These manually extracted features are then input into classic machine learning classifiers such as Support Vector Machines (SVM), Random Forests, and k-Nearest Neighbors (k-NN) for pattern recognition and state determination. While this paradigm has proven effective under certain conditions, its inherent technical limitations are becoming increasingly apparent when addressing the demands of modern and complex wind power operation and maintenance. (1) Feature extraction heavily relies on prior knowledge from experts, resulting in insufficient adaptability and generalization ability. The effectiveness of traditional methods largely depends on the quality of feature engineering, which requires diagnostic experts to have a deep understanding of the failure mechanism, vibration dynamics, and signal characteristics of bearings. This "manual design-fixed rules" model has inherent defects: on the one hand, it cannot adaptively and automatically extract the most discriminative deep-level abstract features from massive, high-dimensional, and non-stationary original vibration signals; on the other hand, when faced with novel failure modes (such as fatigue fracture of new materials) or complex failures involving multiple coupled failures that have not appeared in the training data, the predefined feature set often fails, leading to a sharp decline in the model's generalization performance and making it difficult to achieve intelligent diagnosis that is "learned in one place and applicable in many places".

[0004] (2) Weak ability to model the temporal dynamic characteristics of vibration signals. In recent years, deep learning methods based on convolutional neural networks (CNNs) have been introduced, which can automatically learn spatial features from time-frequency images (such as wavelet scale maps and short-time Fourier transform spectra). However, these methods mainly focus on extracting spatial patterns from local receptive fields, which is essentially a static analysis of signal "snapshots". The development of bearing faults is a typical dynamic process, and its vibration signals contain long-sequence temporal dependencies and evolutionary laws closely related to the rotation period, fault impact period, and modulation effect. Existing CNN methods are limited by their structural characteristics and cannot effectively capture this long-range temporal context information spanning thousands of sampling points. Therefore, they cannot comprehensively and coherently characterize the dynamic evolution process of faults from generation to development to deterioration, which limits the detection capability of early weak faults and the accurate assessment of fault severity.

[0005] (3) The fusion mechanism of multi-source heterogeneous information is simple and inefficient, failing to achieve information complementarity and synergistic enhancement. Bearing vibration signal is a natural multimodal information carrier: the time domain signal contains impact intensity and statistical distribution information, the frequency domain signal reveals the characteristic frequency components of the fault, and the time-frequency domain signal can characterize the change of 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 housing also provide complementary spatial vibration information. However, most current advanced methods still rely on simple vector splicing or early fusion of features from different sources. This "patchwork" approach ignores the inherent and 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, the information utilization rate is low, and noise may even be introduced due to differences in the dimensions of features or information redundancy, failing to achieve the synergistic diagnostic effect of "1+1>2".

[0006] (4) Lack of robustness to variable operating conditions in actual operation, the model is prone to performance drift. The actual operating conditions of wind turbine generators are complex and variable. Fluctuations in wind speed directly lead to changes in rotational speed, while grid dispatching demands cause load changes. Changes in rotational speed and load directly cause significant shifts in the amplitude, frequency, and even statistical distribution of bearing vibration signals. This phenomenon is called "domain shift" or "operating condition shift". Traditional diagnostic models usually assume that training data and test data follow the same independent and identically distributed distribution. 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, resulting in a significant decrease in diagnostic accuracy and a serious lack of long-term deployment robustness and stability in the variable environment of actual wind farms. How to enable the diagnostic model to have the ability to "learn by analogy" and adapt across operating conditions is the core problem that must be overcome to promote it from the laboratory to engineering applications.

[0007] Therefore, there is an urgent need for a next-generation intelligent fault diagnosis system for wind turbine bearings that can adaptively extract deep features, accurately model temporal dynamics, intelligently integrate multi-source information, and robustly adapt to changes in operating conditions. Summary of the Invention

[0008] To address the shortcomings of existing wind turbine bearing fault diagnosis methods, such as reliance on manual feature extraction, insufficient temporal dynamic modeling, inadequate multi-source information fusion, poor adaptability to varying operating conditions, and weak model interpretability, this invention provides a wind turbine fault diagnosis method and system based on Mamba-ResNet. Through key technologies such as dual-channel vibration signal input, automatic multi-scale feature extraction, deep interaction of temporal and spatial features, dynamic gating fusion, and adaptive compensation based on operating conditions, high-precision, robust, and interpretable bearing fault diagnosis is achieved.

[0009] In a first aspect, the present invention provides a wind turbine fault diagnosis method based on Mamba-ResNet, which adopts the following technical solution: A fault diagnosis method for wind turbines based on Mamba-ResNet includes: Acquire dual-channel vibration signals from wind turbine generators; Data preprocessing is performed based on the acquired dual-channel vibration signals; Feature extraction and feature fusion are performed 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 multi-dimensional heterogeneous feature vectors, and a cross-modal adaptive gating fusion mechanism based on bidirectional cross-gating units. Construct a hybrid deep network model architecture of Mamba-ResNet; Train the model on the constructed network model architecture; The trained model architecture is used for wind turbine fault diagnosis.

[0010] Secondly, a wind turbine fault diagnosis system based on Mamba-ResNet includes: The data acquisition module is configured to acquire dual-channel vibration signals from the wind turbine. The preprocessing module is configured to perform data preprocessing based on the acquired dual-channel vibration signals; The feature module is configured to perform 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 multi-dimensional heterogeneous feature vectors, and a cross-modal adaptive gating fusion mechanism based on bidirectional cross-gating units. The model building module is configured to build a Mamba-ResNet hybrid deep network model architecture. The model training module is configured to train the constructed network model architecture. The diagnostic module is configured to perform fault diagnosis of wind turbines using the trained model architecture.

[0011] Thirdly, the present invention provides a computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the aforementioned method for diagnosing wind turbine faults based on Mamba-ResNet.

[0012] Fourthly, the present invention provides a terminal device, including a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, the instructions being adapted to be loaded and executed by the processor to provide the described method for fault diagnosis of wind turbine generators based on Mamba-ResNet.

[0013] In summary, the present invention has the following beneficial technical effects: (1) Achieve end-to-end intelligent feature learning and reduce reliance on expert experience: By using deep neural networks to automatically learn multi-scale features with strong discriminative power from the original vibration signals, the tedious manual feature design process in traditional methods is avoided, and the automation level and adaptability of the diagnostic system are significantly improved.

[0014] (2) Deep integration of information to enhance feature expression ability: The time domain, frequency domain, and time-frequency domain features are innovatively integrated with the spatial and temporal features of deep learning, making full use of the multidimensional information in the vibration signal, which significantly enhances the discriminative ability and robustness of the features.

[0015] (3) Efficient time-series modeling capability to capture the dynamic evolution of faults: The introduction of the Mamba model to process long-sequence dependencies in vibration signals overcomes the difficulties of traditional RNN training and the limitations of CNN receptive field, and can effectively capture the dynamic evolution process of fault features.

[0016] (4) Strong robust adaptability to changing working conditions, enhancing practical application value: Through working condition perception embedding and feature distribution alignment technology, the influence of working condition changes on feature distribution is explicitly modeled, so that the diagnostic model can maintain stable performance under changing working conditions, thus improving its application value in actual wind fields.

[0017] (5) Enhanced diagnostic interpretability and improved system credibility: Provides multi-level visualization tools and attribution analysis methods, enabling maintenance personnel to intuitively understand the diagnostic basis of the model, enhancing the credibility and acceptability of the diagnostic results, and providing strong support for predictive maintenance decisions.

[0018] (6) Modular design, easy to expand and maintain: The present invention adopts a modular design, with each functional module being relatively independent, which makes it easy to adjust and optimize according to actual needs, and also makes it easy to maintain and upgrade the system.

[0019] (7) Highly efficient online diagnostic capabilities to meet real-time monitoring needs: Through model optimization and deployment strategies, millisecond-level fault diagnosis response is achieved, meeting the needs of real-time monitoring and early warning of wind farms and providing a time window for timely maintenance. Attached Figure Description

[0020] Figure 1 This is a flowchart of a wind turbine fault diagnosis method based on Mamba-ResNet, an implementation of the present invention; Figure 2 This is a model diagram of a wind turbine fault diagnosis method based on Mamba-ResNet according to Embodiment 1 of the present invention; Figure 3 This is a comparative test result of the accuracy of a wind turbine fault diagnosis method based on Mamba-ResNet implemented in this invention; Figure 4 This is a comparative test result of the F1 score of a wind turbine fault diagnosis method based on Mamba-ResNet implemented in this invention; Figure 5 This is the model training and testing loss of a wind turbine fault diagnosis method based on Mamba-ResNet implemented in this invention; Figure 6 This is a model accuracy transformation curve of a wind turbine fault diagnosis method based on Mamba-ResNet implemented in this invention. Detailed Implementation

[0021] The present invention will be further described in detail below with reference to the accompanying drawings.

[0022] Example 1 Reference Figure 1 This embodiment of a wind turbine fault diagnosis method based on Mamba-ResNet includes: Its overall process is as follows Figure 1 As shown, it includes the following steps: S1: Enhanced Simultaneous Acquisition and Adaptive Preprocessing of Dual-Channel Vibration Signals S1.1 Multi-directional Synchronous Acquisition and High-Fidelity Digitization To ensure comprehensive capture of the three-dimensional spatial vibration pattern of wind turbine bearings under complex loads, this method deploys high dynamic range, low-noise IEPE piezoelectric accelerometers in two mechanically orthogonal directions: the horizontal radial direction (X-axis) and the vertical axial direction (Y-axis) of the bearing housing. The two sensors achieve microsecond-level synchronous acquisition through hardware triggering, recording the vibration acceleration time-domain signals separately. and .

[0023] sampling frequency The Hz frequency is set to 25,600 Hz. This parameter is determined based on the following principles: First, it must satisfy the Nyquist sampling theorem to acquire the fault characteristic frequencies and their higher harmonics (typically more than 10 times the bearing's operating frequency) of each bearing component (inner ring, outer ring, rolling elements, and cage) without distortion. Second, a high sampling rate helps to accurately capture the details of the instantaneous impact waveform caused by defects such as local spalling and cracks, providing high-resolution input for subsequent time-frequency analysis and deep learning models.

[0024] S1.2 Adaptive Normalization Based on Sliding Window Given the time-varying nature of wind turbine load and rotational speed, the overall energy distribution of the vibration signal will fluctuate accordingly. To eliminate the impact of differences in signal amplitude scale under different operating conditions on model training and to preserve the relative structural information within the signal, a sliding window local normalization is used instead of global normalization. For each signal channel... At any moment Calculate its length before and after. The mean of the signal within the window with standard deviation And normalize it: , in, This is a numerical stability constant introduced to prevent division by zero. This method can adaptively eliminate baseline drift caused by slowly varying operating conditions, allowing the model to focus more on the fault characteristics inherent in the local fluctuation patterns of the signal.

[0025] S1.3 Dynamic Data Augmentation Strategies Oriented Towards Physical Mechanisms To improve the model's robustness to disturbances in real-world environments and its generalization ability to unknown operating conditions, the normalized signal is processed during the training phase. 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 expanding the data.

[0026] (1) Adaptive Signal-to-Noise Ratio Noise Injection: To simulate actual interference such as sensor electronic noise and background mechanical noise, and to avoid noise masking the real fault impact, a noise injection method based on adaptive signal local energy is adopted. Specifically, the noise intensity is correlated with the signal energy (variance) within a local time window: , in For random intensity factor, For The variance of the signal within the sliding window centered on the signal.

[0027] (2) Elastic time warp: Nonlinear scaling of signal timing to simulate speed fluctuations: , Wherein is the distortion amplitude. This is the signal length.

[0028] S2: Adaptive Extraction and Fusion of Multi-Scale Time-Frequency Features S2.1 Constructing an Adaptive Windowed Short-Time Fourier Transform (AW-STFT) to achieve dynamic time-frequency resolution analysis. To overcome the inherent limitation of traditional short-time Fourier transform (SFT) in simultaneously achieving high time resolution and high frequency resolution within a fixed window, this invention constructs an adaptive window SFT. Its core lies in utilizing a time-varying Gaussian window function with a standard deviation... The system can be dynamically adjusted based on 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 caused by local defects in the bearing (such as pitting or spalling) (characterized by high frequency and short duration), the algorithm automatically adopts a narrower time window, sacrificing some frequency resolution to accurately locate the moment of impact and obtain a time-spectrum with high temporal resolution; conversely, when analyzing the steady-state modulation or harmonic components caused by bearing wear or misalignment, it automatically switches to a wider frequency window to clearly separate dense spectral lines and obtain high frequency resolution. This "adaptive windowing" strategy ensures that the generated two-dimensional time-frequency map can simultaneously depict the transient impact timing and steady-state spectral structure of the fault signal at optimal resolution, providing a low-level representation with maximized information density for subsequent visual feature extraction.

[0029] , in The time-varying Gaussian window function has a standard deviation. Adaptive adjustment based on the instantaneous frequency change rate of the signal: a narrow window is used to improve time resolution for high-frequency transient components, and a wide window is used to improve frequency resolution for stationary components.

[0030] During the long-term operation of wind turbine bearings, the energy distribution patterns of different fault types in the time-frequency domain exhibit significant differences. Typical localized damage faults (such as pitting and spalling) usually manifest as short-duration, high-frequency impact signals, with their energy highly concentrated in several localized areas in the time-frequency plane. In contrast, progressive faults such as wear and increased clearance tend to exhibit wide-duration, low-contrast energy diffusion characteristics. Relying solely on traditional global time-frequency energy or statistical characteristics is insufficient to effectively distinguish these different fault patterns.

[0031] Therefore, this invention introduces a time-frequency impact focusing index in the multi-scale time-frequency feature extraction stage. This index is used to quantitatively characterize the local concentration of vibration signal energy in the time-frequency plane, and its definition is shown in the following formula. This index compares the significant impact areas... The proportion of energy within the fault and the overall time-frequency energy distribution can characterize whether the fault energy exhibits structural characteristics of "high focus" or "dispersed expansion".

[0032] , in, This represents the time-frequency amplitude matrix obtained through adaptive short-time Fourier transform or continuous wavelet transform; This indicates the impact-dominant region automatically identified through energy mutation detection or threshold discrimination methods. By accumulating the squared amplitude values, the sensitivity to high-energy impact components can be enhanced, and interference from weak noise in the index calculation can be avoided.

[0033] When a bearing is in the early stage of localized damage, the impact energy is concentrated in a limited time-frequency region, resulting in a high degree of focus. Higher values ​​are achieved; however, during the wear-accelerated or complex fault stages, energy tends to disperse in the time-frequency plane. The value decreases accordingly. Therefore, this indicator can effectively reflect the type characteristics and evolution stage of bearing failure.

[0034] In the model implementation, the time-frequency impact focus index can be used as an explicit auxiliary feature input to the ResNet spatial feature extraction branch, or used to modulate the weights of multi-scale convolutional feature channels, thereby guiding the network to pay more attention to the time-frequency region with high impact and high discriminative power, and improving the model's ability to perceive weak fault features and distinguish different fault modes.

[0035] S2.2 Constructing multi-dimensional heterogeneous feature vectors to achieve feature synergy between physical statistics and deep learning. From the above time-frequency representation and the original signal, three types of complementary feature vectors are extracted in parallel to construct a feature pool that integrates prior knowledge and data-driven approaches. Time-domain statistical eigenvectors: directly derived from the original vibration signal and its first difference Second-order difference In this study, a set of 24 classical statistics is calculated. These features include, but are not limited to: the mean describing the center of the signal distribution, the standard deviation describing the amplitude of the fluctuation, the skewness describing the asymmetry of the distribution, the kurtosis describing the sharpness of the distribution, as well as the root mean square (RMS), peak factor, impulse factor, and waveform factor. These features have clear physical meaning and can robustly characterize the amplitude distribution, impulse intensity, and waveform shape of the signal, providing a validated and interpretable physical insight basis for the model.

[0036] , Frequency band energy entropy eigenvector: To quantify the frequency domain distribution pattern of fault energy, the power spectral density of the signal is divided into... This involves dividing the bearing into non-uniform frequency bands with clearly defined physical orientations. The division is based on a combination of bearing geometry parameters and theoretical fault characteristic frequencies, such as: a low-frequency band related to bearing throughput frequencies, a mid-frequency band related to rolling element fault characteristic frequencies, and a high-frequency band related to structural resonance high-frequency modulation. The calculation of each frequency band... Energy percentage Furthermore, its Shannon entropy is calculated. This entropy value effectively measures the degree of concentration or dispersion of 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), while a high entropy value indicates that the energy is dispersed (such as a complex fault or severe wear). This feature provides a key frequency domain distribution fingerprint for distinguishing different fault modes.

[0037] , Different fault types often exhibit different energy distribution characteristics in different frequency bands.

[0038] Singular value eigenvectors of the time-frequency matrix: Treating the time-frequency graph as a two-dimensional matrix Singular value decomposition (SVD) is then performed on it. Singular value sequence This represents the projected intensity of the time-frequency graph energy in each orthogonal direction, and its distribution characteristics reflect the stability and complexity of the time-frequency structure. (Take the previous value) The feature vector is composed of several major singular values. This feature highly compresses the global structural information of the time-frequency plot and can effectively characterize the "trajectory" pattern formed by the fault impact on the time-frequency plane, such as whether it is a discrete point trajectory (corresponding to a single impact) or a continuous band trajectory (corresponding to modulation phenomenon), thus providing the model with a macroscopic and structured time-frequency morphological description.

[0039] S3: Constructing a Mamba-ResNet hybrid deep network architecture S3.1 Multi-scale Convolutional Attention Module (MS-ConvFormer): Enables spatial multi-granularity feature focusing. To address the different spatial scales (such as subtle short-frequency ripples and broad harmonic bands) that fault features may exhibit in time-frequency images, this module designs a parallel multi-branch convolutional structure. This structure uses convolutional kernels of three different sizes (3×3, 5×5, and 7×7) to perform parallel convolution operations on the input time-frequency image, simultaneously capturing local fine-grained textures, medium-range patterns, and global contextual 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 weights of each channel through a two-layer fully connected network containing a ReLU activation function, and finally... The function generates attention coefficients between 0 and 1. This process enables the network to automatically suppress irrelevant or noisy channels and enhance feature channels that are crucial to the current fault diagnosis task, achieving adaptive focusing and fusion of multi-scale spatial features.

[0040] , Introducing a channel attention mechanism to automatically learn the weights of features at different scales: , GAP represents global average pooling. It is the ReLU activation function. for Function. S3.2 Improved ResNet Spatial Feature Extractor: Constructing Hierarchical Deep Spatial Representations Building upon the primitive features extracted by MS-ConvFormer, a deep residual network (ResNet) is introduced as the backbone extractor for spatial features. We made two key improvements to the standard ResNet to adapt it to fault diagnosis tasks: First, we constructed a three-stage progressive network, with each stage containing multiple improved residual blocks. This downsamples the spatial size of the feature maps layer by layer, while expanding the channel dimension from 64 to 128 and 256, thereby achieving a hierarchical abstraction from low-level edge textures to mid-level component patterns and then to high-level fault semantics. Second, we integrated [the following] within each residual block... The attention module explicitly models the interdependencies between channels and dynamically adjusts the feature responses of each channel, thereby enhancing information interaction and collaboration between feature channels and enabling the network to learn more discriminative channel combinations. Ultimately, the network outputs multi-layered feature maps. These features together constitute a complete and robust deep representation of the fault in the time-frequency space.

[0041] S3.3 Mamba Temporal Dynamic Modeling Module: Implements Long-Range Selective Temporal Dependency Capture To model the long-range dynamic evolution of vibration signals spanning thousands of sampling points, this module introduces the latest structured state-space sequence model (Mamba). Its core idea is to transform one-dimensional time-series vibration signals... The observations are treated as a continuous-time system and modeled using the following selective state-space equations: , in It is a diagonal state matrix. and This is the projection matrix. The continuous system is transformed into a discrete form using a discretization method, enabling efficient computation.

[0042] To enhance the model's ability to perceive the impact characteristics of early bearing failures and to avoid excessive memorization of irrelevant stationary signals during long-term time-series modeling, this invention introduces a selective memory modulation mechanism based on the intensity of the failure impact in the Mamba time-series modeling module.

[0043] Specifically, firstly, statistical analysis is performed on the input vibration signal within a local time window to calculate the signal at the current moment. Relative to its local mean The degree of deviation, combined with local standard deviation. The deviation is normalized. The fault impact sensing coefficient is defined by equation (1). This is used to quantify the significance of potential fault impacts in the vibration signal at the current moment: , Based on this, as shown in equation (2), the impact sensing coefficient is introduced into the Mamba state update process to dynamically modulate the fusion ratio between historical states and current candidate states. Specifically, Used to control the strength of the model's memory of information at the current moment, and This is used to control the degree of retention of historical states, thereby enabling the key memory of fault-related critical moments and the rapid forgetting of smooth and redundant information.

[0044] Through the above mechanism, the model can explicitly introduce physical mechanism constraints of bearing failure while maintaining the advantages of Mamba long-range time series modeling. This makes the time series modeling process pay more attention to impact events closely related to the failure evolution, effectively improve the detection capability of weak faults and early anomalies, and enhance the robustness and stability of the model under complex working conditions.

[0045] , Mamba's key innovation lies in its selectivity mechanism and efficient hardware-aware design. Firstly, its system parameters... (Input matrix) (Output matrix) and the most important step size parameter It is no longer fixed, but depends on 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 shocks, 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 shocks and the evolution of modulation phenomena.

[0046] , 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.

[0047] S4: Cross-modal adaptive gating fusion mechanism S4.1 Bidirectional Cross-Gated Unit: Realizing Dynamic Dialogue and Collaborative Enhancement Based on Spatiotemporal Features 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.

[0048] 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.

[0049] 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 yields a new temporal feature enhanced by spatial information. This process simulates a deep "dialogue" between two modal features—for example, a persistent high-energy region in a frequency band of the time-frequency plot (spatial feature) can suggest that the Mamba module should pay more attention to periodic impulses in the signal during the corresponding time period (temporal feature enhancement); conversely, a strong transient impulse sequence in the signal (temporal feature) can suggest that ResNet should focus on local patterns at the corresponding time and frequency in the time-frequency plot (spatial feature enhancement). Finally, these two interactively enhanced features are fused to provide a complementary and internally consistent unified representation for subsequent decision-making.

[0050] , in The degree to which spatial features enhance temporal features is controlled. The degree to which temporal features enhance spatial features.

[0051] To further constrain the physical consistency of temporal and spatial features across time scales, a temporal-spatial consistency energy constraint is introduced. , which serves as a regularization term in the fusion phase.

[0052] , in: This indicates that Mamba is at time [time]. Output timing characteristics. These are the spatial features extracted by ResNet in the corresponding time-frequency region. For space-time aligned projection operators (such as time pooling + linear mapping) This energy term measures the degree of consistency between the temporal dynamic model and the spatial time-frequency model on the same time scale. When a bearing actually fails, the time-domain impact sequence and the time-frequency energy accumulation region should be highly consistent in time, and this formula explicitly constrains this physical consistency.

[0053] S4.2 Multi-Expert Dynamic Fusion Network: Implementing a Task-Oriented Adaptive Feature Synthesis Strategy Considering that different fault types (such as pitting, wear, and combined faults) or different operating conditions may lead to different effective diagnostic logic and dependent feature combinations, we constructed a multi-expert dynamic fusion network to simulate such diverse decision-making paths. This network includes... One (for example, Parallel "expert" subnetworks.

[0054] , Each expert is an independent small neural network, trained to learn and excel at a specific, underlying fault-feature mapping relationship. For example, expert A might excel at diagnosing early pitting by analyzing high-frequency transient features; expert B might excel at diagnosing wear by analyzing low-frequency modulation features; and expert C might excel at integrating various features to handle complex faults. These experts are differentiated and complementary.

[0055] The key innovation lies in a lightweight routing network. This network takes the aforementioned fused features as input, analyzes the feature attributes of the current sample in real time, and dynamically calculates a weight vector assigned to each expert. This weight vector is obtained through... Function normalization represents the degree to which each expert's output should be trusted or emphasized in the current diagnostic decision. Ultimately, the network's output is a weighted sum of all expert network outputs. This mechanism allows the model to adaptively select the most suitable feature synthesis and inference path based on the characteristics of the input signal itself, rather than forcing the same fixed set of fusion rules to be used for all samples. This greatly improves the model's flexibility and its ability to diagnose different fault modes.

[0056] , S5: Adaptive Feature Calibration and Robust Classification for Operating Conditions S5.1 Condition-Aware Embedded Learning: Encoding environmental context as part of features Wind turbine generator speed With load These are the core operating condition parameters that affect the vibration signal performance. To eliminate signal distribution shifts caused by changes in operating conditions, we use these continuous operating condition parameters as explicit inputs to the model. Through a dedicated embedding layer, rotational speed and load are mapped into a low-dimensional, dense, continuous vector condition. This operating condition embedding vector encodes the contextual information of the current operating environment.

[0057] , The operating condition embedding not only serves as additional information input to the network, but also utilizes an operating condition-sensitive gating function. Dynamic modulation of the feature channels.

[0058] , in, These are the characteristics after fusion. For speed-load condition embedding, Represented as the characteristic rate of change over time, Dynamic modulation coefficient S5.2 Feature Distribution Alignment: Driving the Model to Learn Essential Fault Features That Remain Unchanged Under Operating Conditions Simply adding operating condition information as input is insufficient. To ensure the model truly focuses on fault features independent of operating conditions, we introduce a domain-adaptive loss during training—Maximum Mean Difference (MMD). Its core objective is to force the distributions of samples from different operating conditions (e.g., high RPM versus low RPM) but of the same fault type to be as close as possible in the feature space. MMD minimizes the distance between the feature distributions under training and testing conditions: , in For feature mapping function, For the regenerated nucleus Hilbert space.

[0059] MMD achieves this goal by calculating the distance between the mean features of samples under different operating conditions in the regenerative kernel Hilbert space (RKHS). By minimizing this distance, we are essentially constraining the feature extractor to ignore changes caused by operating conditions, while preserving and amplifying changes caused by the fault itself. This makes the feature representation learned by the model invariant to changes in operating conditions, thus maintaining stable diagnostic performance when faced with new operating conditions not encountered during training.

[0060] S5.3 Multi-task Collaborative Optimization: Driving Main Task Performance Improvement with Auxiliary Tasks To further stabilize training and improve the discriminative power of features, we designed a multi-task learning framework. In addition to the main task (fault classification), the model executes an auxiliary task (operating condition identification) in parallel, simultaneously predicting the current rotational speed and load condition.

[0061] 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 fault classification loss, condition recognition loss, and the aforementioned MMD distribution alignment loss (MMD), the model is guided to learn a rich feature representation that is both highly sensitive to fault categories and implicitly reflects condition information. The auxiliary task acts as a powerful regularizer, preventing the network from overfitting to specific combinations of conditions in the training data on the fault classification task and encouraging the feature extractor to build more comprehensive and generalizable internal representations, ultimately feeding back into and improving the classification robustness of the main task.

[0062] , Furthermore, we employed label smoothing in the classification loss of the main task. This technique replaces the original "hard" labels (e.g., category [0, 1, 0, 0]) with "soft" labels (e.g., [0.025, 0.925, 0.025, 0.025]). This introduces a slight "uncertainty" into model training, effectively preventing the model from overconfident in its prediction probabilities of training samples, thus alleviating overfitting and improving the model's calibration and generalization ability under noise interference.

[0063] , in This is the smoothed label distribution.

[0064] S6: Model Training Strategies and Optimization Methods S6.1 Progressive Training Strategy: Unlocking the network in stages to achieve stable optimization. Given the complexity of the model structure and the large number of parameters, direct end-to-end training can easily lead to gradient instability or getting trapped in local optima. Therefore, we adopt a three-stage progressive training strategy: Feature Extractor Freezing Phase: First, the weights of core feature extraction backbone networks such as ResNet and Mamba are fixed, making them pre-trained feature extractors. Only the newly added classification head (fully connected layer) at the end of the model is trained. This phase allows the classification head to quickly learn how to use existing features for initial classification, while stabilizing the initial stage of the entire training process.

[0065] Intermediate layer fine-tuning stage: After the classification head converges, the deeper layers near the output of the backbone network are unfrozen. These layers typically contain more abstract, task-relevant features. During this stage, the unfrozen layers are fine-tuned together with the classification head to adaptively adjust the feature representations to the current diagnostic task.

[0066] Global End-to-End Fine-Tuning Phase: Finally, all network layers are unfrozen, and global end-to-end joint training is performed. Due to the large number of model parameters, we employ gradient accumulation: the gradients from multiple mini-batches are summed before performing a weight update. This is equivalent to simulating the training effect of large batches with limited GPU memory, which helps to obtain more stable gradient estimates and a better final convergence point.

[0067] S6.2 Adaptive Learning Rate Scheduling: Dynamically adjusting and optimizing the step size to guide the model to converge smoothly. 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: , in .

[0068] 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.

[0069] 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.

[0070] 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.

[0071] 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.

[0072] 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.

[0073] , in and These are the first and second moment estimates of the gradient, respectively. This is the weight decay coefficient.

[0074] Furthermore, to address the gradient explosion problem that can occur in deep neural networks, we implemented gradient clipping before each parameter update. Specifically, we calculated the norm of the gradient vectors for all trainable parameters of the entire model. If this norm exceeded a preset threshold (clip_norm), all gradients were scaled proportionally to make their norm equal to the threshold. This operation acts like a "stabilizer" for the training process, effectively preventing drastic gradient fluctuations caused by anomalous data in individual batches or deep backpropagation within the network, thus ensuring the numerical stability of the training process. This is especially crucial for processing long sequences in the Mamba module.

[0075] S7: Fault Diagnosis Decision Making and Interpretability Analysis S7.1 Multi-dimensional Confidence Assessment: From "Single Prediction" to "Credible Decision Making" To transform the model's "black box" output into a reliable decision-making basis for operations and maintenance personnel, we went beyond a single maximum probability value and constructed a multi-dimensional confidence assessment system. This system integrates information from three levels to calculate the final confidence score: Predictive probability level: i.e., the maximum probability value output by the model's Softmax function. This is the basis of confidence level.

[0076] Feature space level: Calculate the separation index. This index is calculated by comparing the feature vectors of the current sample. With fault category prototype feature center The separation is quantified by the distance between the current sample feature and its class center, as well as its distance from other class centers. If the current sample feature is close to its class center and far from other class centers, the separation is high and the confidence is high; if it falls into a fuzzy overlapping region of the feature space, the separation is low and the confidence is low.

[0077] At the time series level: Historical consistency is considered. The reliability of the current diagnosis is assessed by comparing the consistency between the current diagnostic results and recent historical diagnostic results (such as predictions within a sliding window). A stable and consistent fault warning is more reliable than an isolated, abrupt anomaly.

[0078] The separation index is defined as follows: , 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.

[0079] S7.2 Explainability Visualization Tools: A Series of Keys to Unlocking the "AI Black Box" To build user trust and facilitate in-depth analysis, we provide a multi-layered set of interpretable visualization tools: 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.

[0080] Attention distribution map: 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.

[0081] 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.

[0082] 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.

[0083] 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.

[0084] Experimental verification and performance evaluation To verify the effectiveness of the method of this invention, we conducted a comprehensive experiment on a dataset collected from an actual wind farm. The dataset contains data on four types of bearing faults (cage fault, inner ring fault, outer ring fault, and combined inner and outer ring fault) under three different operating conditions (35Hz / 12kN, 37.5Hz / 11kN, and 40Hz / 10kN), totaling complete life cycle monitoring records for 15 bearings.

[0085] Comparison Methods To verify the effectiveness of the proposed Mamba-ResNet fusion model in wind turbine gearbox fault diagnosis, this paper selects three representative deep learning models as benchmark comparison methods: (1) CAAN model (Cross-domain Attention Adaptive Network): It uses cross-domain attention and adaptive learning to solve the problem of insufficient generalization ability of fault diagnosis models under varying working conditions.

[0086] (2) EMD-1DCNN model: The vibration signal is automatically decomposed into multi-channel intrinsic mode functions through empirical mode decomposition, and then a 1D convolutional neural network is used for end-to-end fault classification, which effectively addresses the diagnostic challenges under non-stationary working conditions.

[0087] (3) LMD-MSE-SVM model: Extracts multi-scale complexity features through adaptive signal decomposition and uses support vector machine to achieve high-precision fault classification.

[0088] (4) VibrMamba: An efficient long sequence modeling framework based on state-space models (such as Mamba) specifically designed for vibration signal analysis.

[0089] Evaluation metrics: The model performance is comprehensively evaluated using multiple metrics such as accuracy and F1 score.

[0090] Experimental results are as follows Figure 3 and Figure 4 As shown: To verify the effectiveness and superiority of the proposed Mamba-ResNet model in wind turbine gearbox fault diagnosis, four representative comparative methods were selected for performance comparison experiments: Emd-1dCNN based on Empirical Mode Decomposition and a one-dimensional convolutional neural network; Virmamba based on state-space sequence modeling; LMD-MSE-SVM combining Local Mean Decomposition and Support Vector Machine; and CANN multi-channel convolutional attention network. Each model was trained and tested under the same experimental dataset and operating conditions. The evaluation metrics were classification accuracy and F1 score. (See Figure 3.) Figure 4As shown, the proposed model demonstrates significant advantages in both metrics. 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 our model achieves 99.73%, an improvement of approximately 14.03 percentage points compared to the best-performing comparative model. This indicates that the proposed hybrid architecture integrating structured state space and residual networks significantly improves feature representation and classification capabilities.

[0091] Regarding the F1 score, as shown in the figure, our model also achieved the best performance: the F1 scores of Emd-1dcnn, Virmamba, LMD-MSE-SVM, and CANN were 71.59%, 88.41%, 87.67%, and 93.10%, respectively, while our model reached 99.73%. Compared to other methods, the improved F1 score indicates that the model is more robust in handling classification balance and recall of different classes of samples, especially maintaining high diagnostic accuracy even under imbalanced sample distribution. Figure 5 and Figure 6 The results show that the improved Mamba-ResNet model significantly improves the accuracy of fault diagnosis when dealing with imbalanced sample data. Compared with existing methods, the accuracy is improved by an average of 14.03 percentage points, fully validating the effectiveness and superiority of the proposed method.

[0092] Overall, the Mamba-ResNet model proposed in this paper achieves multimodal adaptive fusion of time-series, time-frequency, and statistical features by introducing a dynamic gated fusion mechanism (Cross-Gated Fusion), significantly enhancing feature interaction and fault mode recognition capabilities. Under complex multi-condition scenarios, the model not only maintains high accuracy but also demonstrates outstanding performance in feature balancing and generalization, fully validating its effectiveness and superiority in wind turbine gearbox fault diagnosis.

[0093] A computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device, the aforementioned method for fault diagnosis of a wind turbine based on Mamba-ResNet.

[0094] A terminal device includes a processor and a computer-readable storage medium, the processor being used to implement various instructions; the computer-readable storage medium being used to store multiple instructions, the instructions being adapted to be loaded and executed by the processor as described in the Mamba-ResNet-based wind turbine fault diagnosis method.

[0095] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

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; performing model training on the constructed network model architecture; performing wind turbine fault diagnosis using the trained model architecture.

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 The dynamic data augmentation based on physical constraints is implemented, and the signal local energy adaptive noise injection method is used to introduce random disturbances in the time domain that conform to the actual physical law. The noise intensity is associated with the energy of the signal in the local time window: where is a random intensity factor, is a signal variance within a sliding window centered at and a non-linear stretching of the signal timing to simulate speed fluctuations: wherein is the twist amplitude, is the signal length.

3. The Mamba-ResNet-based wind turbine fault diagnosis method according to claim 1, 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, which is expressed as: ， wherein denotes a time-frequency amplitude matrix obtained by an adaptive short-time Fourier transform; denotes an impact dominant region identified by an energy abruptness detection.

4. The Mamba-ResNet-based wind turbine fault diagnosis method according to claim 1, 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, which are calculated directly from the original vibration signals and first-order differences and second-order differences A total of 24 dimensions of statistics are calculated and represented as: ； The band energy entropy feature vector quantifies the frequency domain distribution pattern of the fault energy by dividing the power spectral density of the signal into K=4 non-uniform frequency bands with clear physical direction, then calculates the energy proportion of each frequency band , and further calculates the Shannon entropy thereof, and the entropy value is used to measure the concentration and dispersion degree of the fault energy on the frequency spectrum, and is expressed as:​ ； singular value eigenvector of time-frequency matrix, by regarding the time-frequency map as a two-dimensional matrix and performing singular value decomposition (SVD), using the distribution characteristics of singular value sequence to reflect the stability and complexity of time-frequency structure, taking the first major singular values to form the eigenvector, representing 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 1, characterized in that, The construction of the Mamba-ResNet hybrid deep network model architecture includes designing a parallel multi-branch convolution structure for different spatial scales of fault feature presentation in the time-frequency image, 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 gate mechanism, the fused multi-scale features are adaptively recalibrated, the global information of each feature channel is aggregated through the global average pooling GAP, and the importance weight of each channel is learned through the two-layer fully connected network containing the ReLU activation function, and finally through The function generates attention coefficients between 0 and 1, expressed 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 Squeeze-and-Excitation attention module is integrated in each residual block, the interdependence 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 1, 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 signal deviation from its local mean , and the deviation is normalized by combining the local standard deviation and defining a fault impact perception coefficient to quantify the significance of potential fault impact in the current 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, for controlling the memory strength of the model to the information at the current time, for controlling the retention degree of the historical state, so as to realize the key memory to the fault-related key time and the fast forgetting to the smooth redundant information, and is expressed as: 。 7. The Mamba-ResNet-based wind turbine fault diagnosis method according to claim 1, characterized in that, The cross-modal adaptive gating fusion mechanism based on bidirectional cross-gating units includes receiving spatial feature vectors from ResNet. and temporal feature vectors from Mamba ,pass The function is compressed to the [0,1] interval, and then a bidirectional enhancement operation is performed to obtain a new spatial feature enhanced by temporal information. Simultaneously, the temporal feature is multiplied element-wise with its corresponding gating weight to obtain a new temporal feature enhanced by spatial information. Finally, the two interactively enhanced features are fused, represented as: ， wherein controlling the degree of enhancement of the spatial features on the temporal features, controlling the degree of enhancement of the temporal features on the spatial features; To constrain the physical consistency of the temporal and spatial features in the time scale, a temporal-spatial consistency energy constraint is introduced which is expressed as a regularization term in the fusion stage ， where: denotes the time instant the output temporal features, is the spatial feature extracted by ResNet in the corresponding time-frequency region, is the spatial-temporal alignment projection operator; finally, a multi-expert dynamic fusion network is constructed: ， Each expert is an independent small neural network, taking fused features as input, real-time analyzing the feature attributes of the current sample, and dynamically calculating the weight vector assigned to each expert Finally, the network output is the weighted sum of all expert network outputs, represented as: 。 8. The Mamba-ResNet-based wind turbine fault diagnosis method according to claim 1, 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 represented as: and being dynamically modulated through a working condition sensitive gating function a feature channel, and being represented as: ， wherein, is the fused feature, is the speed-load condition embedding, is the feature time rate of change, dynamic modulation coefficient; then introduce the maximum mean discrepancy MMD as the domain adaptive loss, use the maximum mean discrepancy MMD to minimize the distance between the feature distribution under the training condition and the test condition, denoted 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 represented as: and introduce uncertainty with 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 on 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 are fine-tuned together with the classification head 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: multiple small batches of gradients are accumulated, and then the weight is updated 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 at the same time, the large weight is continuously punished through an 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 wind turbine fault diagnosis system based on Mamba-ResNet, 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 perform model training on the constructed network model architecture; a diagnosis module configured to perform wind turbine fault diagnosis using the trained model architecture.

Images