Offshore wind turbine generator fault diagnosis method and system based on multi-source sensor data fusion

Abstract

The invention provides an offshore wind turbine generator fault diagnosis method and system based on multi-source sensor data fusion. The method comprises the steps that a vibration signal from at least one component of a wind turbine generator is acquired through a vibration sensor; performing time-frequency conversion on the vibration signal by applying synchronous compression wavelet transform to obtain time-frequency representation of the vibration signal; when the reconstruction error exceeds a preset threshold value, it is judged that an abnormal event exists in the vibration signal; obtaining the position of a part corresponding to the abnormal event; starting an image sensor and an acoustic sensor according to the position of the component, and acquiring an image signal and a sound signal of the component according to the image sensor and the acoustic sensor; according to the DS evidence theory, the vibration signal, the image signal and the sound signal, obtaining the confidence of the fault type; the fault type of the component is judged according to the maximum confidence allocation principle, high-resolution time-frequency analysis can be achieved through synchronous compression wavelet transform (SST), and the fault feature identification degree is improved in combination with the self-encoding neural network and the D-S evidence theory.

Description

Technical Field

[0001] This application relates to the field of wind power equipment monitoring technology, and in particular to a fault diagnosis method and system for offshore wind turbines based on multi-source sensor data fusion. Background Technology

[0002] As the global energy structure transitions towards clean and low-carbon energy, offshore wind power, as an important component of renewable energy, is showing a trend of rapid development. However, offshore wind turbines operate in complex environments with high humidity, high salinity, strong winds, and strong corrosion for extended periods. Key components such as gearboxes, main shafts, generators, blades, and connecting bolts are highly susceptible to fatigue, corrosion, vibration, and shock, leading to failures such as wear, cracks, loosening, and overheating.

[0003] Currently, the industry uses vibration sensors, thermocouples, cameras, and infrared thermal imagers for condition monitoring, but most of these methods remain at the level of independent monitoring based on a single data source. The lack of data fusion mechanisms and intelligent analysis methods results in many potential faults not being effectively identified in their early stages. Summary of the Invention

[0004] In view of this, it is necessary to provide a fault diagnosis method and system for offshore wind turbines based on multi-source sensor data fusion, which can at least overcome one of the above-mentioned defects.

[0005] In a first aspect, embodiments of this application provide a fault diagnosis method for offshore wind turbines based on multi-source sensor data fusion, the method comprising:

[0006] Vibration signals from at least one component of the wind turbine are acquired using vibration sensors.

[0007] The vibration signal is converted to time and frequency using synchronous compressed wavelet transform to obtain a time-frequency representation of the vibration signal;

[0008] The time-frequency representation is input into a pre-trained autoencoder neural network;

[0009] The time-frequency representation is processed by feature compression and reconstruction to obtain the reconstruction error;

[0010] When the reconstruction error exceeds a preset threshold, it is determined that an abnormal event exists in the vibration signal;

[0011] Obtain the location of the component corresponding to the abnormal event;

[0012] The image sensor and acoustic sensor are activated according to the position of the component, and the image signal and sound signal of the component are acquired according to the image sensor and the acoustic sensor.

[0013] The confidence level of the fault type is obtained based on the DS evidence theory, the vibration signal, the image signal, and the sound signal;

[0014] The fault type of the component is determined according to the maximum confidence allocation principle.

[0015] In one embodiment, applying synchronous compressed wavelet transform to perform time-frequency conversion on the vibration signal to obtain a time-frequency representation of the vibration signal includes:

[0016] The Morlet mother wavelet is used as the analysis basis function;

[0017] Set the scale set and frequency set;

[0018] The vibration signal is decomposed using the scale set and the frequency set to obtain the scale coefficients.

[0019] Calculate the instantaneous frequency based on the scale coefficient;

[0020] The continuous wavelet transform is redistributed to the trajectory of the instantaneous frequency using the Dirac function to generate the time-frequency representation.

[0021] In one embodiment, calculating the instantaneous frequency based on the scale coefficient includes:

[0022] The partial derivative of the scaling coefficient is calculated using the following formula:

[0023]

[0024] Among them, W x (a j b) represents the scale factor indicating the vibration signal at scale a. j and local features at time b; a j φ is the scale parameter, inversely proportional to frequency; b is the translation parameter, representing the time position of the signal; φ(a j b) represents the phase of the scaling factor;

[0025] The phase derivative is obtained by taking the imaginary part of the partial derivative and normalizing it.

[0026] The instantaneous frequency is obtained based on the phase derivative.

[0027] In one embodiment, the step of redistributing the continuous wavelet transform to the trajectory of the instantaneous frequency using the Dirac function includes:

[0028] The energy of the continuous wavelet transform coefficients is redistributed to the trajectory of the instantaneous frequency using the Dirac function to generate the high-frequency time-frequency representation. The generation formula is as follows:

[0029]

[0030] Among them, T x (b,ω) represents the time-frequency representation, Δa j For the scale step size, Where is the normalization factor, J is the total scale number, δ(·) is the Dirac function, and W x (a j b) represents the instantaneous frequency.

[0031] In one embodiment, obtaining the confidence level of the fault type based on DS evidence theory, the reconstruction error, the image signal, and the sound signal includes:

[0032] Define trust assignment functions for the vibration signal, the image signal, and the sound signal respectively;

[0033] Calculate the degree of conflict among the vibration signal, the image signal, and the sound signal;

[0034] The trust assignment functions for the vibration signal, the image signal, and the sound signal are sequentially fused pairwise to generate a comprehensive confidence score. The generation function for the comprehensive confidence score is as follows:

[0035]

[0036] Wherein, m(F) j F represents the overall confidence level. j The fault type is represented by A and B, which are intermediate fusion results. is the normalization factor, and m1 and m2 are the basic probability assignment functions.

[0037] In one embodiment, determining the fault type of the component based on the maximum confidence allocation principle includes:

[0038] Obtain at least two of the aforementioned combined confidence levels;

[0039] Compare the difference between the two comprehensive confidence levels. If the difference is less than or equal to a preset threshold, a manual review is triggered.

[0040] If the difference is greater than a preset threshold, the fault type corresponding to the highest overall confidence level is selected as the fault type of the component.

[0041] In one embodiment, the method further includes performing image enhancement processing on the image signal, the image enhancement processing including:

[0042] When the wind turbine is in a backlit scene, the contrast of the image signal is enhanced according to the local histogram equalization algorithm to adjust the exposure of the image signal.

[0043] When the wind turbine is in a low-light scene, the image signal is processed according to a non-local mean algorithm to reduce environmental noise.

[0044] In one embodiment, the method further includes:

[0045] Set the initial threshold based on empirical coefficients;

[0046] The preset threshold is dynamically reconstructed based on the peak threshold theory to update the preset threshold.

[0047] If the preset threshold is higher than the initial threshold, then the preset threshold is skipped.

[0048] In one embodiment, the components of the offshore wind turbine include blades, a gearbox, a main bearing, and a generator;

[0049] The method further includes: initiating a drone inspection when the abnormal event occurs, wherein the drone is equipped with the image sensor, the acoustic sensor, the lidar, and the infrared thermal imager;

[0050] If the component corresponding to the abnormal event is a blade, then the image sensor and the lidar of the UAV are activated to collect visible light images and three-dimensional point cloud data of the surface cracks, deformations and loose bolts of the blade;

[0051] If the component corresponding to the abnormal event is the gearbox, then the infrared thermal imager of the UAV is activated to collect the temperature distribution data of the gearbox shell;

[0052] If the component corresponding to the abnormal event is the main bearing, then the infrared thermal imager of the UAV is activated to monitor the temperature gradient change of the bearing housing of the main bearing, and the acoustic sensor is activated to obtain the abnormal acoustic signature characteristics caused by the friction of the main bearing, and the three-dimensional point cloud data of the bearing housing is scanned by the lidar.

[0053] If the component corresponding to the abnormal event is a generator, then the infrared thermal imager of the UAV is activated to collect the winding temperature distribution data of the generator.

[0054] Secondly, embodiments of this application provide a fault diagnosis system for offshore wind turbines based on multi-source sensor data fusion, the system comprising:

[0055] A vibration sensor unit is used to acquire vibration signals from at least one component of the wind turbine.

[0056] The signal conversion unit is used to apply synchronous compressed wavelet transform to perform time-frequency conversion on the vibration signal to obtain a time-frequency representation of the vibration signal;

[0057] An autoencoder neural network is used to accept the time-frequency representation and process the time-frequency representation through feature compression and reconstruction to obtain reconstruction error;

[0058] An abnormal event determination module is used to determine that an abnormal event exists in the vibration signal when the reconstruction error exceeds a preset threshold.

[0059] An abnormal event confirmation module is used to obtain the location of the component corresponding to the abnormal event; activate the image sensor and acoustic sensor according to the location of the component, and obtain the image signal and sound signal of the component according to the image sensor and the acoustic sensor; obtain the confidence level of the fault type according to the DS evidence theory, the reconstruction error, the image signal and the sound signal; and determine the fault type of the component according to the maximum confidence allocation principle.

[0060] This application provides a fault diagnosis method and system for offshore wind turbines based on multi-source sensor data fusion. It can achieve high-resolution time-frequency analysis through synchronous compressed wavelet transform (SST), and dynamically fuse vibration, image, acoustic and infrared multimodal data by combining autoencoder neural network and DS evidence theory, which significantly improves the fault feature identification and early anomaly detection sensitivity. Attached Figure Description

[0061] Figure 1 This is a flowchart illustrating a method for diagnosing offshore wind turbine faults based on multi-source sensor data fusion, provided as an embodiment of this application.

[0062] Figure 2 This is a schematic diagram of a fault diagnosis system module for offshore wind turbines provided in an embodiment of this application.

[0063] Figure 3 A schematic diagram of an electronic device provided in an embodiment of this application.

[0064] Explanation of main component symbols

[0065] Offshore wind turbine fault diagnosis system 10

[0066] Task Management Module 100

[0067] Unmanned Aerial Vehicle (UAV) Airport Array 200

[0068] Scheduling optimization module 300

[0069] Route planning module 400

[0070] Rescheduling module 500

[0071] Electronic devices 20

[0072] Processor 21

[0073] Memory 22

[0074] Method steps S100-900 Detailed Implementation

[0075] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0076] It should be noted that, in the embodiments of this application, "at least one" refers to one or more, and "more than one" refers to two or more. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the specification of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application.

[0077] It should be noted that in the embodiments of this application, the terms "first," "second," etc., are used only for descriptive purposes and should not be construed as indicating or implying relative importance, nor as indicating or implying order. Features specified as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of the embodiments of this application, words such as "exemplary" or "for example" are used to indicate examples, illustrations, or explanations. Any embodiment or design scheme described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design schemes. Specifically, the use of words such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0078] Based on the embodiments described in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0079] As the global energy structure transitions towards clean and low-carbon energy, offshore wind power, as an important component of renewable energy, is showing a trend of rapid development. However, offshore wind turbines operate in complex environments with high humidity, high salinity, strong winds, and strong corrosion for extended periods. Key components such as gearboxes, main shafts, generators, blades, and connecting bolts are highly susceptible to fatigue, corrosion, vibration, and shock, leading to failures such as wear, cracks, loosening, and overheating.

[0080] Currently, the industry uses vibration sensors, thermocouples, cameras, and infrared thermal imagers for condition monitoring, but most still operate at the level of independent monitoring based on a single data source. For example, traditional vibration signal analysis relies on empirical rules, is susceptible to noise interference, and struggles to capture transient features; image inspection performs poorly in low-light, backlight, or rotating environments; and infrared imaging is greatly affected by background temperature fluctuations, making it difficult to accurately identify local anomalies. The lack of data fusion mechanisms and intelligent analysis methods results in many potential faults not being effectively identified in their early stages.

[0081] This application provides a fault diagnosis method and system for offshore wind turbines based on multi-source sensor data fusion. It achieves high-resolution time-frequency analysis through Synchronous Compressed Wavelet Transform (SST), and dynamically fuses vibration, image, acoustic, and infrared multimodal data using an autoencoder neural network and DS evidence theory, significantly improving fault feature identification and early anomaly detection sensitivity. A dynamic threshold mechanism adapts to sudden wind loads and high-salt-spray environments at sea, reducing the false alarm rate to below 5%. Unmanned aerial vehicles (UAVs) equipped with multiple sensors collaboratively collect data on key defects such as blade cracks and bearing wear, reducing invalid inspections by more than 30%. The system supports tiered early warning and expert review, effectively extending component lifespan and improving the operational reliability of offshore wind farms.

[0082] Figure 1 This is a schematic flowchart of a fault diagnosis method for offshore wind turbines based on multi-source sensor data fusion provided in an embodiment of this application, as shown below. Figure 1 The offshore wind turbine fault diagnosis method based on multi-source sensor data fusion, as shown, includes at least the following steps: S100: acquiring vibration signals from at least one component of the wind turbine through vibration sensors; S200: applying synchronous compressed wavelet transform to perform time-frequency conversion on the vibration signals to obtain a time-frequency representation of the vibration signals; S300: inputting the time-frequency representation into a pre-trained autoencoder neural network; S400: processing the time-frequency representation through feature compression and reconstruction to obtain a reconstruction error; S500: determining that an abnormal event exists in the vibration signal when the reconstruction error exceeds a preset threshold; S600: obtaining the location of the component corresponding to the abnormal event; S700: activating image sensors and acoustic sensors according to the location of the component, and acquiring image and sound signals of the component based on the image and acoustic sensors; S800: obtaining the confidence level of the fault type based on DS evidence theory, vibration signals, image signals, and sound signals; S900: determining the fault type of the component according to the maximum confidence allocation principle.

[0083] S100: Obtain vibration signals from at least one component of the wind turbine through a vibration sensor.

[0084] Specifically, vibration sensors are deployed on key components of the wind turbine, including the blade roots, gearbox housing, main bearing housing, and generator base, to collect multi-axis vibration data (such as X / Y / Z acceleration signals) in real time during equipment operation. The sensors employ high-sensitivity piezoelectric ceramic elements, with a sampling rate covering high-frequency vibration components (≥10 kHz), and transmit the raw data to the edge computing unit via a wireless communication module. In the high-salt-spray and high-wind-load environment at sea, the sensor housing features an anti-corrosion coating and a waterproof seal design to ensure long-term stable operation. The collected vibration signals include mechanodynamic characteristics, such as blade aerodynamic load fluctuations, gear meshing impacts, and bearing rolling element friction, providing fundamental data for subsequent fault analysis.

[0085] Understandably, collecting vibration data throughout the entire lifecycle using high-precision vibration sensors can capture subtle anomalies in the early stages of equipment development (such as local resonance in the early stages of blade crack propagation or sideband energy changes in the early stages of gearbox wear), significantly improving the sensitivity of fault detection. The distributed deployment strategy of the sensors covers key components, and combined with anti-interference design, adapts to the complex marine environment, effectively reducing the impact of noise interference on data quality and laying a reliable foundation for subsequent signal processing.

[0086] S200: Apply synchronous compressed wavelet transform to perform time-frequency conversion on the vibration signal to obtain the time-frequency representation of the vibration signal.

[0087] Specifically, the vibration signal is processed using Synchronous Compressed Wavelet Transform (SST). First, the Morlet mother wavelet is selected as the basis function, possessing excellent time-frequency locality. Then, the vibration signal is decomposed into wavelet coefficients at different scales based on a preset scale and frequency set. Further, the instantaneous frequency is calculated based on the phase derivative of the wavelet coefficients to determine the true trajectory of the signal energy. Finally, the energy of the continuous wavelet transform is redistributed to the instantaneous frequency trajectory using the Dirac function, generating a high-resolution time-frequency representation. This process effectively suppresses the energy divergence problem in traditional wavelet transforms, resulting in clear time-frequency ridges for fault characteristic frequencies on the time-frequency graph, such as low-frequency periodic impacts caused by blade cracks or high-frequency energy surges caused by broken gear teeth.

[0088] Understandably, the synchronous compressed wavelet transform (SST) significantly improves time-frequency resolution through an energy redistribution mechanism, enabling the accurate capture of fault characteristics under non-stationary operating conditions (such as frequency drift caused by sudden wind speed changes). This method can dynamically track the fault frequency evolution path without relying on a speed sensor, solving the technical challenge of monitoring variable operating conditions in offshore wind turbines. Furthermore, the high-resolution time-frequency map generated by SST can clearly present the modulation effects of fault characteristics (such as the amplitude modulation / frequency modulation relationship between bearing fault frequency and rotational frequency), providing high-quality input data for subsequent anomaly detection.

[0089] In this embodiment, synchronous compressed wavelet transform is applied to perform time-frequency conversion on the vibration signal to obtain a time-frequency representation of the vibration signal. Specifically, this includes: using the Morlet mother wavelet as the analysis basis function; setting a scale set and a frequency set; performing wavelet decomposition on the vibration signal according to the scale set and frequency set to obtain scale coefficients; calculating the instantaneous frequency based on the scale coefficients; and redistributing the continuous wavelet transform to the trajectory of the instantaneous frequency using the Dirac function to generate a time-frequency representation.

[0090] Specifically, in the application of synchronous compressed wavelet transform, the Morlet mother wavelet is first used as the analysis basis function. The Morlet wavelet, modulated by complex exponential and Gaussian functions, possesses both good time-domain localization characteristics and frequency-domain resolution, making it particularly suitable for extracting non-stationary impact components from offshore wind turbine vibration signals. Subsequently, scale sets and frequency sets are set according to the signal frequency band characteristics. The scale parameters are divided according to a logarithmic distribution, covering the complete vibration spectrum from low to high frequencies, ensuring that different fault characteristics (such as blade aerodynamic load fluctuations and gear meshing impacts) can be effectively captured. Based on the set scales, wavelet decomposition is performed on the vibration signal to generate a multi-scale wavelet coefficient matrix, which characterizes the energy distribution of the signal at different time points and scales. Further, the instantaneous frequency is obtained by calculating the phase derivatives of the wavelet coefficients: after taking the partial derivatives of the complex wavelet coefficients, the phase change rate is separated and normalized to a physical frequency value, thereby determining the true trajectory of the signal energy in the time-frequency plane. Finally, the energy of the continuous wavelet transform is redistributed to the instantaneous frequency trajectory using the Dirac function, generating a high-resolution time-frequency representation. This process, through mathematical operations, concentrates the energy that was originally dispersed due to scale expansion onto the instantaneous frequency ridge, significantly improving the focus of the time-frequency plot. This allows fault features (such as harmonics of bearing fault frequencies or gearbox sidebands) to exhibit a clear peak structure in the time-frequency domain, facilitating subsequent anomaly detection and feature extraction.

[0091] Understandably, the synchronous compressed wavelet transform (SST) addresses the energy divergence problem of traditional wavelet transform by introducing phase derivatives and energy redistribution mechanisms, making it particularly suitable for fault feature extraction under complex operating conditions of offshore wind turbines. In high-salt-spray and high-wind-load environments at sea, this method can accurately capture weak impact signals caused by early faults (such as local resonance in the early stages of blade crack propagation) and visually demonstrate the evolution of fault frequencies through high-resolution time-frequency plots. Furthermore, SST can dynamically track frequency changes of non-stationary signals without relying on speed sensors, adapting to spectral drift caused by sudden wind speed changes. This characteristic significantly improves the robustness of fault diagnosis in offshore wind power scenarios, avoiding misjudgments caused by environmental disturbances, and provides high-quality time-frequency input for subsequent feature compression and anomaly detection of autoencoder neural networks, further enhancing the reliability of multi-source data fusion.

[0092] In this embodiment of the application, the mother wavelet is defined as:

[0093]

[0094] Where ω0 is the center frequency (ranging from 8 to 20), and j is the imaginary unit. This mother wavelet possesses both good time-domain localization characteristics and frequency-domain resolution capabilities, making it suitable for non-stationary analysis of vibration signals from offshore wind turbines.

[0095] Define discrete scale set Where a j Let be the scale parameter (inversely proportional to frequency), and J be the total number of scales (e.g., 64–128). The corresponding frequency set {ω} j} is obtained by transformation from the scale parameter, such as ω j =ω c / (a j ·s), ω c denoted as the center angular frequency of the mother wavelet, and s as the sampling rate.

[0096] Performing a continuous wavelet transform (CWT) on the vibration signal x(t) yields the scaling coefficient W. x (a j b):

[0097]

[0098] Among them, a j Ψ is the scale parameter, b is the translation parameter (time axis coordinate), and Ψ is the translation parameter (time axis coordinate). * This is the conjugate function of the mother wavelet. The scaling coefficient matrix reflects the energy distribution of the signal at different scales and time points.

[0099] In this embodiment, calculating the instantaneous frequency based on the scaling coefficients specifically includes: taking the partial derivatives of the scaling coefficients; taking the imaginary part of the partial derivatives and normalizing them to obtain the phase derivative; and obtaining the instantaneous frequency based on the phase derivative. This process involves calculating the wavelet coefficients in complex form.

[0100] The formula for finding partial derivatives is:

[0101]

[0102] Among them, W x (a j b) represents the scale factor indicating the vibration signal at scale a. j and local features at time b; a j φ is the scale parameter, inversely proportional to frequency; b is the translation parameter, representing the time position of the signal; φ(a j b) represents the phase of the scaling factor.

[0103] The formula for obtaining the phase derivative by taking the imaginary part and normalizing is:

[0104]

[0105] Finally, the instantaneous frequency ω is extracted. x (a j b)

[0106]

[0107] ω x It represents the instantaneous frequency, reflecting the true frequency trajectory of the signal energy.

[0108] In this embodiment, redistributing the continuous wavelet transform to the trajectory of the instantaneous frequency using the Dirac function includes: redistributing the energy of the continuous wavelet transform coefficients to the trajectory of the instantaneous frequency using the Dirac function to generate a high-frequency time-frequency representation, the generation formula being:

[0109]

[0110] Among them, T x (b,ω) represents the time-frequency representation, Δa j For the scale step size, ω is the normalization factor, J is the total scale number, δ(·) is the Dirac function, and ω x (a j b) represents the instantaneous frequency.

[0111] Specifically, this method first employs the Morlet mother wavelet as the analysis basis function. Its complex exponential and Gaussian modulation characteristics balance time-domain localization and frequency-domain resolution, making it suitable for the non-stationary analysis of vibration signals from offshore wind turbines. By defining a discrete scale set and corresponding frequency set, a continuous wavelet transform (CWT) is performed on the vibration signal to generate a scale coefficient matrix, characterizing the energy distribution of the signal at different scales and time points. Subsequently, the instantaneous frequency is calculated based on the complex wavelet coefficients: by obtaining the phase derivative of the scale coefficients and combining it with imaginary part normalization, the instantaneous frequency is finally obtained, accurately locating the true frequency trajectory of the signal energy. Finally, the energy of the continuous wavelet transform is redistributed to the instantaneous frequency trajectory using the Dirac function, generating a high-resolution time-frequency representation.

[0112] Understandably, the phase derivative and energy redistribution mechanism of Synchronous Compressed Wavelet Transform (SST) solves the energy divergence problem of traditional wavelet transform, improving the identification of fault characteristic frequencies (such as the periodic impact of blade cracks and gearbox sidebands) on the time-frequency graph by more than 3 times, supporting early identification of blade cracks at the 0.1mm level and temperature anomalies below 1℃. SST can dynamically track the frequency drift problem of offshore wind turbines caused by sudden changes in wind speed without relying on speed sensors, such as the aerodynamic load fluctuation of blades or the modulation effect of gearbox meshing frequency, which is significantly better than traditional FFT or Short Time Fourier Transform (STFT). The normalization factor suppresses the energy attenuation caused by scale expansion, ensuring the comparability of energy distribution at different scales; the Dirac function concentrates energy to the instantaneous frequency trajectory, effectively reducing the impact of sea wind turbulence noise (≥80dB) on fault characteristics, and improving the signal-to-noise ratio by more than 5dB.

[0113] S 300: Input the time-frequency representation into a pre-trained autoencoder neural network.

[0114] Specifically, the autoencoder neural network in this embodiment adopts a convolutional-deconvolutional architecture, comprising an encoder, a bottleneck layer, and a decoder. The encoder consists of multiple layers of convolutional kernels and pooling operations, applying 32, 64, and 128 filters sequentially (kernel size 3×3, stride 1), extracting local features of the time-frequency map through the ReLU activation function; the bottleneck layer compresses the feature vector to a 128-dimensional latent space, preserving the core pattern of the signal; the decoder gradually restores the feature map size through deconvolution operations, ultimately generating the reconstructed time-frequency map. During the training phase, the network uses a vibration signal dataset under normal operating conditions (such as time-frequency maps with stable wind speed and no cracks / wear on the components), optimizing parameters through a weighted loss function that minimizes reconstruction error and sparsity constraints.

[0115] Understandably, autoencoder neural networks, through unsupervised learning, construct a feature library of normal operating conditions, enabling them to automatically identify abnormal events where signals deviate from normal patterns (such as local resonance caused by blade crack propagation or sideband energy enhancement due to gearbox wear). The convolutional-deconvolutional structure effectively extracts the spatial correlation of time-frequency maps, avoiding the limitations of manually designed features. Sparsity constraints (KL divergence) enhance the model's robustness to noise, ensuring stable operation in complex marine environments. This method exhibits high sensitivity to early faults (e.g., a 15% increase in reconstruction error recognition rate when blade crack length is <1mm), while dynamically adjusting the loss function weights λ to adapt to the vibration characteristics of different components (e.g., frequency band differences between blades and gearboxes), significantly improving the versatility and reliability of the diagnosis.

[0116] S 400: Time-frequency representation is processed by feature compression and reconstruction to obtain reconstruction error.

[0117] Specifically, the encoder part of the autoencoder network compresses the time-frequency map into a low-dimensional feature vector through multiple convolutional operations. The decoder then gradually recovers the feature map through deconvolution operations, ultimately generating the reconstructed time-frequency map. The reconstruction error is the L2 norm difference between the original time-frequency map and the reconstructed map.

[0118] Understandably, the feature compression and reconstruction process of the autoencoder network quantifies the degree to which the signal deviates from the normal pattern through the L2 loss function. The reconstruction error is highly sensitive to early faults (such as minor bearing spalling or early gearbox tooth breakage) and can capture weak anomalies in the signal (such as a sudden increase in reconstruction error ≥15%). The dynamic threshold mechanism, combined with environmental parameter compensation (wind speed, temperature), avoids the misjudgment problem of fixed thresholds under complex operating conditions. For example, the threshold is relaxed at high wind speeds (to suppress turbulence interference) and tightened at high temperatures (to enhance sensitivity to material fatigue). This method achieves high-precision detection of abnormal events in offshore wind power scenarios, reducing the false alarm rate to below 5%, significantly outperforming traditional threshold methods.

[0119] S 500: When the reconstruction error exceeds the preset threshold, it is determined that there is an abnormal event in the vibration signal.

[0120] Specifically, when the reconstruction error generated by the autoencoder network exceeds a dynamic threshold, the system determines that an abnormal event has occurred in the vibration signal. This threshold combines an initial threshold (based on historical data statistics μ+3σ) with real-time environmental parameter compensation. Furthermore, a safety factor is adjusted to ensure suppression of turbulence interference under high wind speeds and enhance sensitivity to material fatigue under high-temperature environments. If the dynamically calculated threshold is higher than the initial threshold, the initial value is retained to avoid erroneous threshold increases caused by extreme environments.

[0121] In this embodiment, the method further includes: setting an initial threshold based on an empirical coefficient; dynamically calculating the reconstruction error distribution based on a preset threshold according to peak threshold theory to update the preset threshold; and skipping the preset threshold if the preset threshold is higher than the initial threshold.

[0122] Specifically, based on historical reconstruction error data under normal operating conditions (such as the reconstruction error distribution of vibration signal time-frequency diagrams over three consecutive months), the mean and standard deviation are calculated, and an initial threshold is set according to empirical coefficients. Safety factors are set based on component type and operating condition differences (e.g., blade α = 2.5, gearbox α = 3.0) to balance the fault sensitivities of different components. The Peak Overthreshold Theory (POT) is dynamically updated: the threshold is updated every time window (1 hour to 24 hours) using a sliding window mean and weighted standard deviation.

[0123] Understandably, this dynamic threshold mechanism significantly improves the adaptability of offshore wind power under complex operating conditions. By combining the initial threshold with dynamic updates, the system can flexibly cope with non-stationary signal interference. In high salt spray environments, vibration sensors may experience baseline drift due to corrosion. The dynamic threshold update mechanism (POT sliding window) automatically adapts to such slow changes, avoiding frequent manual calibration while retaining the ability to detect real faults (e.g., triggering an early warning when the reconstruction error slowly increases in the early stages of blade crack propagation). Furthermore, the "preserve initial value" strategy of the threshold determination logic effectively solves the limitations of the traditional fixed threshold method in extreme environments. For example, during typhoon season when wind speed fluctuates frequently, if the dynamic threshold is higher than the initial value due to short-term noise interference, the system still uses the initial threshold to avoid missed detections; while under stable operating conditions, the dynamic threshold can be lower than the initial value, improving the ability to detect weak anomalies (e.g., an 8% increase in reconstruction error when the blade crack length is <0.5mm can trigger an early warning).

[0124] S 600: Get the location of the component corresponding to the abnormal event.

[0125] Specifically, the system uses GPS positioning and matches it with a wind turbine coordinate database to determine the component corresponding to the abnormal event (such as the blade root, gearbox housing, main bearing housing, or generator base). The coordinate database pre-defines the mapping relationship between each sensor number and its physical location (e.g., the coordinates of sensor number BL-01 at the blade root correspond to x=0m, y=50m, z=80m). If the abnormality is in the blade, the UAV activates the path planning module and sets a hovering point (e.g., 10 meters from the blade tip); if the abnormality is in the main bearing, the UAV hovers in the area 15 meters above the tower top and activates the infrared thermal imager and directional acoustic array; if the abnormality is in the gearbox or generator, the fixed infrared thermal imager and vibration sensor array are activated.

[0126] Understandably, this positioning and linkage mechanism, through a closed-loop design of GPS and a coordinate database, ensures component positioning accuracy ≤0.1 meters and response time <2 seconds. The drone's hovering path planning, combined with component location and wind speed data, avoids turbulence interference and improves the efficiency of multi-sensor data acquisition. For example, after a blade anomaly is triggered, the drone can complete path planning and hover to the target area within 5 seconds, with high-definition cameras and LiDAR simultaneously collecting data, significantly reducing manual intervention and improving diagnostic timeliness.

[0127] S 700: Activate the image sensor and acoustic sensor according to the position of the component, and acquire the image signal and sound signal of the component according to the image sensor and acoustic sensor.

[0128] Specifically, the corresponding sensor modules are activated based on the component type. When a blade malfunctions, the UAV's high-definition camera (0.05mm / pixel resolution) and lidar (range accuracy ±1cm) work together, covering an angle from 0° to 180°, while an ultrasonic sensor (center frequency 40kHz) captures high-frequency acoustic emission signals. When a gearbox malfunctions, a fixed infrared thermal imager (temperature range -20℃ to 300℃, resolution 640×512) collects the gearbox housing temperature distribution, and a vibration sensor array (sampling rate 10kHz) captures meshing frequency sideband data. When the main bearing malfunctions, the UAV hovers 15 meters above the tower, activating the infrared thermal imager to monitor the bearing housing temperature gradient (>5℃ triggers an alert), and a directional microphone array (5cm spacing, beamforming towards the bearing housing) captures friction noise (20kHz to 50kHz). When the generator malfunctions, the infrared thermal imager monitors the winding temperature gradient (>10℃ triggers an alert), and an electromagnetic vibration sensor collects 100Hz harmonic energy.

[0129] In this embodiment, the components of the offshore wind turbine include blades, gearboxes, main bearings, and generators. The method further includes: in the event of an abnormal event, activating a drone for inspection. The drone is equipped with an image sensor, an acoustic sensor, a lidar, and an infrared thermal imager. If the abnormal event corresponds to a blade, the drone's image sensor and lidar are activated to acquire visible light images and 3D point cloud data of surface cracks, deformation, and loose bolts on the blade. If the abnormal event corresponds to a gearbox, the drone's infrared thermal imager is activated to acquire temperature distribution data of the gearbox housing. If the abnormal event corresponds to a main bearing, the drone's infrared thermal imager is activated to monitor the temperature gradient change of the main bearing housing, while the acoustic sensor acquires abnormal acoustic signature characteristics caused by friction in the main bearing, and the lidar scans the 3D point cloud data of the bearing housing. If the abnormal event corresponds to a generator, the drone's infrared thermal imager is activated to acquire winding temperature distribution data of the generator.

[0130] Understandably, the multi-sensor linkage mechanism overcomes the limitations of traditional fixed monitoring systems, making it particularly suitable for detecting high-altitude blades and complex structural components. For example, a drone hovering and collecting high-frequency acoustic emission signals (40 kHz to 100 kHz) within a 10-meter range of the blade tip, combined with lidar point cloud data (accuracy 0.1 mm), can identify cracks as small as 0.1 mm. The joint analysis of main bearing friction noise (20 kHz to 50 kHz) and temperature gradient (>5℃) increases the accuracy of bearing spalling fault identification to 95%. Furthermore, the synergistic application of infrared thermography and vibration spectrum (such as the joint determination of a gearbox local overheating temperature difference >10℃ and sideband energy enhancement) significantly reduces the false alarm rate (<5%) and improves diagnostic reliability.

[0131] S 800: The confidence level of the fault type is obtained based on DS evidence theory, vibration signals, image signals and sound signals.

[0132] Specifically, the system performs fusion analysis on multi-source sensor data based on DS evidence theory to obtain the comprehensive confidence level for each fault type. First, a basic probability assignment function (BPA) is defined for vibration signals, image signals, and sound signals to quantify the degree of confidence each sensor type has in different fault hypotheses. Then, the system sequentially fuses the confidence assignment functions of the three sensors pairwise. First, the vibration signal and image signal are calculated using DS combination rules. Next, the intermediate results are fused a second time with the confidence assignment function of the acoustic sensor to finally generate the comprehensive confidence level.

[0133] In this embodiment, the confidence level of the fault type is obtained based on DS evidence theory, reconstruction error, image signal, and sound signal, including: defining trust assignment functions for vibration signal, image signal, and sound signal respectively; calculating the conflict degree between vibration signal, image signal, and sound signal; and sequentially fusing the trust assignment functions of vibration signal, image signal, and sound signal pairwise to finally generate a comprehensive confidence level. The generation function of the comprehensive confidence level is:

[0134]

[0135] Wherein, m(F) j F represents the overall confidence level. j The fault type is represented by A and B, which are intermediate fusion results. The normalization factor ensures that the sum of the confidence levels for all fault types is 1, and the basic probability assignment function is m1 and m2.

[0136] Understandably, this application utilizes DS evidence theory to achieve collaborative decision-making from heterogeneous data sources—vibration, image, and acoustic—effectively improving the robustness and accuracy of fault identification in offshore wind turbines. This method dynamically assesses conflicts between sensors and adaptively adjusts their weights, preventing overall diagnostic failure due to misjudgment by a single sensor. For example, when salt spray interference causes image blurring, the system automatically reduces image weight and increases the contribution ratio of vibration and acoustic data, ensuring diagnostic stability. Simultaneously, through a step-by-step iterative fusion strategy, the diagnostic process possesses good interpretability and flexibility, applicable to various typical fault scenarios (such as blade cracks, gear wear, and bearing spalling), and supports expert system intervention to handle multi-hypothesis conflicts. Real-world data shows that in complex marine environments, this method improves the confidence level of multi-source fusion by an average of 15%–20%, significantly outperforming single-sensor diagnostic results, greatly reducing false alarms and missed alarms, and improving operational efficiency and equipment safety.

[0137] S 900: Determine the fault type of the component based on the maximum confidence allocation principle.

[0138] Specifically, in this embodiment, the system uses DS evidence theory to fuse vibration, image, and acoustic sensor data to generate a comprehensive confidence level for multiple fault types, and determines the fault type based on the maximum confidence principle. When the system identifies at least two fault hypotheses with high confidence, it calculates the difference between them and compares it with a preset threshold (e.g., 0.1). If the difference is less than or equal to the preset threshold, it indicates that the confidence levels of multiple fault types are close, and the diagnostic result is uncertain. In this case, a manual review mechanism is triggered, and multimodal data (e.g., time-frequency maps, high-definition images, thermal images, acoustic spectrograms) are pushed to the operation and maintenance platform for further analysis by professionals combining historical data and expert experience. If the difference is greater than the preset threshold, the fault type with the highest confidence is selected as the final diagnostic result and output to the operation and maintenance scheduling system, triggering the corresponding early warning or maintenance process.

[0139] In this embodiment, determining the fault type of a component based on the maximum confidence allocation principle includes: obtaining at least two comprehensive confidence levels; comparing the difference between the two comprehensive confidence levels; if the difference is less than or equal to a preset threshold, triggering manual review; if the difference is greater than the preset threshold, selecting the fault type corresponding to the maximum comprehensive confidence level as the fault type of the component.

[0140] Understandably, this fault type determination method combines automatic diagnosis with manual intervention, effectively improving the accuracy and reliability of fault identification for offshore wind turbines. Due to the complex marine environment, a single sensor may be biased by salt spray interference, airflow noise, or lighting effects, and the confidence level after multi-source data fusion may still contain multiple similar fault hypotheses. Introducing a dual mechanism of "maximum confidence allocation + difference judgment" not only improves the algorithm's response speed to high-confidence events but also avoids the risk of misjudgment caused by blind decision-making in low-discrimination situations, thus ensuring the scientific and practical nature of the diagnostic results. Furthermore, the manual review mechanism provides maintenance personnel with ample data support (such as high-definition images, infrared thermal images, vibration time-frequency diagrams, and acoustic spectrograms collected by drones), enabling them to make quick and accurate judgments, reducing the number of on-site inspections, and improving overall maintenance efficiency. Actual test data shows that when the confidence difference is greater than 0.1, the system can autonomously complete more than 90% of the fault determination tasks with an accuracy rate of over 95%. In scenarios with smaller differences, manual review further improves the diagnostic accuracy rate to over 98%, which is significantly better than traditional single-source diagnostic methods.

[0141] In this embodiment of the application, the method further includes image enhancement processing of the image signal, which includes: when the wind turbine is in a backlit scene, enhancing the contrast of the image signal according to a local histogram equalization algorithm to adjust the exposure of the image signal.

[0142] Specifically, the Local Histogram Equalization (CLAHE) algorithm is used to enhance the image contrast. The image is divided into 8×8 local windows, the histogram distribution of each window is calculated, and a contrast limiting factor (typically set to 2 to 4) is applied to prevent artifacts caused by over-enhancement. Gamma correction is then performed on the enhanced image to adjust the overall brightness distribution, making details in backlit areas (such as the shaded side of leaves) clearly visible.

[0143] When the wind turbine is in a low-light environment, the image signal is processed according to the non-local mean algorithm to reduce environmental noise.

[0144] Specifically, a nonlocal means (NLM) denoising algorithm is applied to suppress noise in low-light environments. This algorithm denoises each pixel by calculating a weighted average of similar regions in the image. The specific steps include: defining the sizes of the search window and the similarity window (e.g., 21×21 and 7×7); and calculating the similarity weights.

[0145]

[0146] Where w(i,j) is the similarity weight, Z is the normalization factor, and G is the similarity weight. σ Here, is the Gaussian kernel function, and h is the attenuation coefficient (ranging from 10 to 20). and Let represent the local neighborhood grayscale distribution of pixels i and j. A weighted average is then applied to similar regions based on their weights to generate the denoised image.

[0147] Understandably, for contrast optimization in backlit scenes, Local Histogram Equalization (CLAHE) avoids the oversaturation problem of global equalization through block processing, making details such as loose bolts at the blade root and cracks in the gearbox housing clearly visible under backlight conditions (contrast improvement ≥30%). Gamma correction further adjusts the brightness distribution to ensure that defects in backlit areas (such as blade leading edge corrosion) are not lost in subsequent feature extraction.

[0148] Understandably, for noise suppression in low-light scenes, Nonlocal Means Denoising (NLM) effectively suppresses sensor noise (such as CMOS dark current noise) by utilizing redundant information in the image (such as the repetitiveness of leaf surface texture), thereby improving the signal-to-noise ratio (PSNR) of images in low-light environments (PSNR improvement ≥ 3 dB). By dynamically adjusting the search window and attenuation coefficient, the NLM algorithm can smooth random noise while preserving edge details, making it suitable for image acquisition by drones during nighttime inspections or in rainy weather.

[0149] The following describes an overall exemplary process for a fault diagnosis method for offshore wind turbines based on multi-source sensor data fusion.

[0150] First, vibration signals from key components of the wind turbine are collected using vibration sensors, and the signals are then converted to time-frequency signals using Synchronous Compressed Wavelet Transform (SST). This process involves decomposing the signal using the Morlet mother wavelet, calculating the instantaneous frequency using the phase derivative, and redistributing the energy to the true frequency trajectory using the Dirac function to generate a high-resolution time-frequency map. This method significantly improves the energy focusing capability of traditional wavelet transform, and is particularly suitable for extracting fault features under non-stationary operating conditions of offshore wind turbines, such as identifying periodic impacts caused by blade cracks or the sidebands of gearbox meshing frequencies.

[0151] Subsequently, the time-frequency graph is input into a pre-trained autoencoder neural network. This network compresses and reconstructs signal features using a convolutional-deconvolutional structure, generating a reconstruction error as a preliminary basis for judging abnormal events. When the reconstruction error exceeds a dynamic threshold, the system determines that an abnormal event exists in the vibration signal and triggers a multi-source sensor collaborative diagnostic process. This dynamic threshold is automatically adjusted based on historical data statistics and environmental parameters (such as wind speed and temperature), effectively reducing the probability of misjudgment caused by sudden changes in operating conditions. For example, the threshold is increased under high wind speeds to suppress turbulent noise interference, and the threshold is decreased under high temperature environments to enhance sensitivity to material fatigue.

[0152] After locating the abnormal event, the system activates the corresponding sensors based on the component type: if the abnormality is in the blade, the high-definition camera, lidar, and acoustic sensors on the UAV work together to collect the three-dimensional morphology of the blade surface cracks, aerodynamic noise, and high-frequency acoustic emission signals; if the abnormality is in the gearbox or generator, a fixed infrared thermal imager and vibration sensor array are activated to obtain temperature distribution and spectral characteristics; if the abnormality is in the main bearing, the UAV hovers in the tower top area and activates the infrared thermal imager and directional acoustic array to capture the bearing housing temperature gradient and friction noise. This flexible sensor linkage mechanism breaks through the limitations of traditional fixed monitoring systems, realizing non-contact multi-angle data acquisition, and is especially suitable for the detection of high-altitude blades and complex structural components.

[0153] The acquired image signals undergo dynamic exposure adjustment and nonlocal mean denoising to enhance detail clarity in backlit / low-light scenes. Acoustic signals are analyzed using Mel-frequency cepstral coefficients (MFCC) and high-frequency energy proportion to extract fault-related acoustic signature features. These multimodal data and the time-frequency characteristics of the vibration signals are input into the DS evidence theory framework to define trust assignment functions for the vibration, image, and acoustic sensors, and to calculate the conflict degree between the sensors. If the conflict degree is high (e.g., image blurring causing conflict between crack identification and vibration spectrum), the system automatically adjusts the sensor weights to reduce the contribution of conflict sources; if the conflict degree is low, it is directly fused to generate a comprehensive confidence score. Finally, the fault type corresponding to the highest comprehensive confidence score is selected as the diagnostic result. If the confidence scores of multiple fault types are close (e.g., the difference between the confidence scores of blade cracks and bolt loosening is less than 0.1), the expert system is triggered to intervene, combining rule base matching and additional data acquisition for further judgment.

[0154] This method significantly improves the accuracy and robustness of fault diagnosis through high-resolution time-frequency analysis, multi-source data fusion, and dynamic weight adjustment mechanisms. For example, the multi-sensor system on a UAV can stably identify blade cracks ≥0.1mm in length under salt spray interference, while the fusion strategy based on DS evidence theory improves the diagnostic accuracy of conflicting hypotheses by 15%–20%. Furthermore, the synergistic application of a tiered early warning strategy and an expert system reduces unplanned downtime by 40% and maintenance costs by over 25%, providing reliable technical support for the efficient operation and maintenance of offshore wind farms.

[0155] Figure 2 This is a schematic diagram of the offshore wind turbine fault diagnosis system provided in the embodiments of this application, as shown below. Figure 2 The offshore wind turbine fault diagnosis system 10 shown includes at least the following components: vibration sensor unit 100, signal conversion unit 200, autoencoder neural network 300, abnormal event determination module 400, and abnormal event confirmation module 500.

[0156] In this embodiment, the vibration sensor unit 100 is used to acquire vibration signals from at least one component of the wind turbine. Please refer to [link / reference needed] for details. Figure 1 The details and their corresponding descriptions are not elaborated here.

[0157] In this embodiment, the signal conversion unit 200 is used to apply synchronous compressed wavelet transform to perform time-frequency conversion on the vibration signal to obtain a time-frequency representation of the vibration signal. Please refer to [link / reference needed] for details. Figure 1 The details and their corresponding descriptions are not elaborated here.

[0158] In this embodiment, the autoencoder neural network 300 is used to receive the time-frequency representation and process the time-frequency representation through feature compression and reconstruction to obtain the reconstruction error. Please refer to [link / reference needed] for details. Figure 1 The details and their corresponding descriptions are not elaborated here.

[0159] In this embodiment, the abnormal event determination module 400 is used to determine that an abnormal event exists in the vibration signal when the reconstruction error exceeds a preset threshold. Please refer to [link / reference needed] for details. Figure 1 The details and their corresponding descriptions are not elaborated here.

[0160] In this embodiment, the abnormal event confirmation module 500 is used to obtain the location of the component corresponding to the abnormal event; activate the image sensor and acoustic sensor according to the location of the component, and obtain the image signal and sound signal of the component according to the image sensor and acoustic sensor; obtain the confidence level of the fault type according to DS evidence theory, reconstruction error, image signal and sound signal; and determine the fault type of the component according to the maximum confidence allocation principle. Please refer to [link / reference] for details. Figure 1The details and their corresponding descriptions are not elaborated here.

[0161] Figure 3 This is an electronic device 20 provided in one embodiment of this application. For example... Figure 3 As shown, the electronic device 20 includes at least the following components: a processor 21 and a memory 22.

[0162] In this embodiment, the memory 22 is used to store executable instructions of the processor 21, which, when configured to execute instructions, implement... Figure 1 The method for fault diagnosis of offshore wind turbines based on multi-source sensor data fusion is shown.

[0163] In one embodiment of this application, the program operating in the electronic device 20 may be a program that controls a central processing unit (CPU) to achieve the functions of the above-described embodiments of the present invention (a program that enables the computer to function). Then, the information processed by these devices is temporarily stored in random access memory (RAM) during processing, and subsequently stored in various ROMs such as read-only memory (Flash ROM) and hard disk drives (HDDs), and read, corrected, and written as needed by the CPU.

[0164] It should be noted that a portion of the electronic device 20 described above can also be implemented using a computer. In this case, the program for implementing the control function can be recorded on a computer-readable recording medium, and the program recorded on the recording medium can be read into the computer and executed.

[0165] It should be noted that the term "computer" as used here refers to a computer built into electronic device 20, employing hardware including an operating system and peripheral devices. Furthermore, "computer-readable recording media" refers to removable media such as floppy disks, magneto-optical disks, ROMs, and CD-ROMs, as well as storage devices such as hard drives built into the computer.

[0166] Furthermore, a "computer-readable recording medium" can include: a medium that dynamically stores a program for a short period of time, such as a communication line used when transmitting a program via a network such as the Internet or a communication line such as a telephone line; or a medium that stores a program for a fixed period of time, such as volatile memory inside a computer that serves as a server or client in this case. In addition, the aforementioned program can be a program used to implement the above-mentioned functions, or it can be a program that can implement the above-mentioned functions by combining with programs already recorded in the computer.

[0167] Furthermore, the electronic device 20 in the above embodiments can also be implemented as an assembly (device group) composed of multiple devices. Each device constituting the device group can possess some or all of the functions or functional blocks of the electronic device 20 in the above embodiments. As a device group, it is sufficient to have all the functions or functional blocks of the electronic device 20.

[0168] It is understood that the offshore wind turbine fault diagnosis method and offshore wind turbine fault diagnosis system 10 based on multi-source sensor data fusion provided in this application significantly improve the accuracy, robustness, and environmental adaptability of fault diagnosis by introducing synchronous compressed wavelet transform (SST), autoencoder neural network, DS evidence theory, and dynamic threshold mechanism. Through high-resolution time-frequency analysis, multi-source data fusion, and dynamic threshold mechanism, a complete diagnostic closed loop covering "signal acquisition → anomaly detection → multimodal verification → fault determination" is constructed, significantly improving the accuracy and real-time performance of offshore wind turbine fault identification, and providing an efficient and low-cost technical solution for offshore wind power operation and maintenance.

[0169] Those skilled in the art should recognize that the above embodiments are only used to illustrate this application and are not intended to limit this application. Any appropriate changes and variations made to the above embodiments within the essential spirit and scope of this application fall within the scope of protection claimed in this application.

Claims

1. A fault diagnosis method for offshore wind turbines based on multi-source sensor data fusion, characterized in that, The method includes: Vibration signals from at least one component of the wind turbine are acquired using vibration sensors. The vibration signal is converted to time and frequency using synchronous compressed wavelet transform to obtain a time-frequency representation of the vibration signal; The time-frequency representation is input into a pre-trained autoencoder neural network; The time-frequency representation is processed by feature compression and reconstruction to obtain the reconstruction error; When the reconstruction error exceeds a preset threshold, it is determined that an abnormal event exists in the vibration signal; Obtain the location of the component corresponding to the abnormal event; The image sensor and acoustic sensor are activated according to the position of the component, and the image signal and sound signal of the component are acquired according to the image sensor and the acoustic sensor. The confidence level of the fault type is obtained based on the DS evidence theory, the vibration signal, the image signal, and the sound signal; The fault type of the component is determined according to the maximum confidence allocation principle.

2. The method for fault diagnosis of offshore wind turbines based on multi-source sensor data fusion according to claim 1, characterized in that, The application of synchronous compressed wavelet transform to perform time-frequency conversion on the vibration signal to obtain a time-frequency representation of the vibration signal includes: The Morlet mother wavelet is used as the analysis basis function; Set the scale set and frequency set; The vibration signal is decomposed using the scale set and the frequency set to obtain the scale coefficients. Calculate the instantaneous frequency based on the scale coefficient; The continuous wavelet transform is redistributed to the trajectory of the instantaneous frequency using the Dirac function to generate the time-frequency representation.

3. The method for fault diagnosis of offshore wind turbines based on multi-source sensor data fusion according to claim 2, characterized in that, The calculation of instantaneous frequency based on the scale coefficient includes: The partial derivative of the scaling coefficient is calculated using the following formula: Among them, W x (a j b) represents the scale factor indicating the vibration signal at scale a. j and local features at time b; a j φ is the scale parameter, inversely proportional to frequency; b is the translation parameter, representing the time position of the signal; φ(a j b) represents the phase of the scaling factor; The phase derivative is obtained by taking the imaginary part of the partial derivative and normalizing it. The instantaneous frequency is obtained based on the phase derivative.

4. The method for fault diagnosis of offshore wind turbines based on multi-source sensor data fusion according to claim 3, characterized in that, The process of redistributing the continuous wavelet transform to the instantaneous frequency using the Dirac function includes: The energy of the continuous wavelet transform coefficients is redistributed to the trajectory of the instantaneous frequency using the Dirac function to generate the high-frequency time-frequency representation. The generation formula is as follows: Among them, T x (b,ω) represents the time-frequency representation, Δa j For the scale step size, Where is the normalization factor, J is the total scale number, δ(·) is the Dirac function, and W x (a j b) represents the instantaneous frequency.

5. The method for fault diagnosis of offshore wind turbines based on multi-source sensor data fusion according to claim 1, characterized in that, The process of obtaining the confidence level of the fault type based on DS evidence theory, the reconstruction error, the image signal, and the sound signal includes: Define trust assignment functions for the vibration signal, the image signal, and the sound signal respectively; Calculate the degree of conflict among the vibration signal, the image signal, and the sound signal; The trust assignment functions for the vibration signal, the image signal, and the sound signal are sequentially fused pairwise to generate a comprehensive confidence score. The generation function for the comprehensive confidence score is as follows: Wherein, m(F) j F represents the overall confidence level. j The fault type is represented by A and B, which are intermediate fusion results. is the normalization factor, and m1 and m2 are the basic probability assignment functions.

6. The method for fault diagnosis of offshore wind turbines based on multi-source sensor data fusion according to claim 5, characterized in that, The step of determining the fault type of the component based on the maximum confidence allocation principle includes: Obtain at least two of the aforementioned combined confidence levels; Compare the difference between the two comprehensive confidence levels. If the difference is less than or equal to a preset threshold, a manual review is triggered. If the difference is greater than a preset threshold, the fault type corresponding to the highest overall confidence level is selected as the fault type of the component.

7. The method for fault diagnosis of offshore wind turbines based on multi-source sensor data fusion according to claim 1, characterized in that, The method further includes image enhancement processing on the image signal, the image enhancement processing including: When the wind turbine is in a backlit scene, the contrast of the image signal is enhanced according to the local histogram equalization algorithm to adjust the exposure of the image signal. When the wind turbine is in a low-light scene, the image signal is processed according to a non-local mean algorithm to reduce environmental noise.

8. The method for fault diagnosis of offshore wind turbines based on multi-source sensor data fusion according to claim 1, characterized in that, The method further includes: Set the initial threshold based on empirical coefficients; The preset threshold is dynamically reconstructed based on the peak threshold theory to update the preset threshold. If the preset threshold is higher than the initial threshold, then the preset threshold is skipped.

9. The method for fault diagnosis of offshore wind turbines based on multi-source sensor data fusion according to claim 1, characterized in that, The components of the offshore wind turbine include blades, gearbox, main bearing, and generator; The method further includes: initiating a drone inspection when the abnormal event occurs, wherein the drone is equipped with the image sensor, the acoustic sensor, the lidar, and the infrared thermal imager; If the component corresponding to the abnormal event is a blade, then the image sensor and the lidar of the UAV are activated to collect visible light images and three-dimensional point cloud data of the surface cracks, deformations and loose bolts of the blade; If the component corresponding to the abnormal event is the gearbox, then the infrared thermal imager of the UAV is activated to collect the temperature distribution data of the gearbox shell; If the component corresponding to the abnormal event is the main bearing, then the infrared thermal imager of the UAV is activated to monitor the temperature gradient change of the bearing housing of the main bearing, and the acoustic sensor is activated to obtain the abnormal acoustic signature characteristics caused by the friction of the main bearing, and the three-dimensional point cloud data of the bearing housing is scanned by the lidar. If the component corresponding to the abnormal event is a generator, then the infrared thermal imager of the UAV is activated to collect the winding temperature distribution data of the generator.

10. A fault diagnosis system for offshore wind turbines based on multi-source sensor data fusion, characterized in that, The system includes: A vibration sensor unit is used to acquire vibration signals from at least one component of the wind turbine. The signal conversion unit is used to apply synchronous compressed wavelet transform to perform time-frequency conversion on the vibration signal to obtain a time-frequency representation of the vibration signal; An autoencoder neural network is used to accept the time-frequency representation and process the time-frequency representation through feature compression and reconstruction to obtain reconstruction error; An abnormal event determination module is used to determine that an abnormal event exists in the vibration signal when the reconstruction error exceeds a preset threshold. An abnormal event confirmation module is used to obtain the location of the component corresponding to the abnormal event; activate an image sensor and an acoustic sensor according to the location of the component, and obtain image signals and sound signals of the component according to the image sensor and the acoustic sensor; obtain the confidence level of the fault type according to the DS evidence theory, the vibration signal, the image signal and the sound signal; and determine the fault type of the component according to the maximum confidence allocation principle.

Images