Bearing current damage identification method and device based on multi-mode manifold maintenance

Abstract

The invention relates to the technical field of wind driven generator fault diagnosis, in particular to a bearing current damage identification method and device based on multi-mode manifold maintenance, and the method comprises the steps: collecting fault operation data of damage and fault-free normal bearing operation data; performing noise reduction on the acquired data, then decomposing the acquired data into a plurality of modal signals in a manifold space, calculating a high-dimensional sample entropy of each manifold subspace modal signal, and obtaining multi-modal manifold sample entropies of normal, fault and current-damaged bearings; on the basis of the sample entropy corresponding signals, state exclusive analytic signals of all bearings are constructed through Hilbert transformation, instantaneous amplitude, phase, frequency and sample entropy statistical features are extracted, and an initial feature set is constructed; and carrying out dimensionality reduction by adopting a locality preserving projection method to obtain a low-dimensional mainstream feature set, inputting the low-dimensional mainstream feature set into a classifier of a neural network model, and comparing current damage with other state feature differences to realize accurate recognition. According to the invention, the accuracy of bearing current damage identification is improved.

Description

Technical Field

[0001] This invention relates to the field of wind turbine fault diagnosis technology, and in particular to a bearing current damage identification method and device based on multimodal manifold maintenance. Background Technology

[0002] In the field of wind power generation, the main bearing is a key component of the wind turbine's transmission system, bearing complex loads, including directional wind loads and impact loads. As wind turbines become larger and higher, the main bearing not only faces the risk of wear from mechanical loads but may also experience shaft currents due to factors such as electromagnetic field coupling from the generator. When these shaft currents flow through the rolling elements and contact surfaces of the main bearing, they can trigger electrolytic erosion, leading to pitting or metal melting on the surface, which can then develop into spalling and cracks. Ultimately, this can cause the main bearing to fail, resulting in high replacement costs and downtime losses, severely impacting the economic benefits of wind farms.

[0003] Currently, the industry mainly relies on vibration monitoring technology, which collects vibration signals from the main bearing housing or adjacent components and uses methods such as Fourier transform and envelope analysis to extract fault characteristic frequencies to determine damage. However, traditional vibration monitoring methods face three main challenges: First, shaft current damage signals are weak and coupled with the mechanical fault spectrum, making them easily masked by background vibration; second, the envelope spectrum characteristics of shaft current damage and outer ring pitting corrosion overlap, and traditional envelope analysis cannot distinguish between the two, potentially leading to misjudgment; finally, wind speed fluctuations cause changes in the main bearing speed, resulting in fault characteristic frequency drift, which traditional methods struggle to adapt to, leading to poor identification stability.

[0004] Therefore, existing vibration monitoring methods are insufficient to meet the requirements for early and accurate identification of shaft current damage in main bearings. There is an urgent need to develop new methods that can overcome technical bottlenecks to improve identification accuracy and ensure the reliable operation of wind turbine units. Summary of the Invention

[0005] This invention provides a bearing current damage identification method and device based on multimodal manifold preservation, which solves the problems of difficulty in extracting shaft current damage signals under complex working conditions of wind turbine main bearings, difficulty in distinguishing them from mechanical faults such as pitting on the outer ring, unstable identification due to speed fluctuations, and insufficient utilization of full life data.

[0006] The objective of this invention can be achieved through the following technical solutions: The first aspect of this invention is to provide a bearing current damage identification method based on multimodal manifold retention, comprising: Step S1: Initiate damage to the same type of bearing and collect corresponding data. Simultaneously collect the full life data of the current damage of this type of bearing, and at the same time collect the fault operation data of the damaged bearing and the normal operation data of the fault-free bearing. Step S2: Denoise the collected data. After denoising, the collected data is decomposed into multiple modal signals in the manifold space. Calculate the high-dimensional sample entropy of the modal signals in each manifold subspace to obtain the multimodal manifold sample entropy of normal, faulty, and current-damaged bearings. Step S3: Based on the signal corresponding to the sample entropy, construct the specific analytical signal for each bearing state through Hilbert transform, extract the instantaneous amplitude, phase, frequency and sample entropy statistical features, and construct the initial feature set; Step S4: Use the local preservation projection method to reduce the dimensionality to obtain a low-dimensional mainstream feature set, input it into the classifier of the neural network model, and compare the differences between current damage and other state features to achieve accurate identification.

[0007] Further, step S1 specifically includes: Laser electro-etching pen was used to create faults in the inner ring, ball, and outer ring of the same type of experimental bearing, and the full life data of current damage was acquired simultaneously. Acceleration sensors are installed in the X, Y, and Z directions of the bearing bracket to collect data on the corresponding faults and normal bearing operation.

[0008] Furthermore, in step S2, the collected data is denoised using wavelet threshold denoising. The wavelet basis is db4, the decomposition level is 3-5 levels, and the threshold is the VisuShrink default threshold. The manifold subspace dimension d = k / 2 - 2k / 3, where k is the total number of modal functions and d is an integer, ensuring that each subspace corresponds to 1-2 modal functions to avoid aliasing.

[0009] Furthermore, the parameters for calculating the sample entropy include: The embedding dimension m=2, the similarity tolerance r=0.18-0.22 times the standard deviation of the modal signal, and the data length N=1000-2000 points; the statistical features of the sample entropy include mean, variance, peak factor, kurtosis, skewness, and impulse factor; the impulse factor is used to distinguish between bearing current damage and mechanical failure.

[0010] Furthermore, in step 4, the classifier is a support vector machine, random forest, or convolutional neural network. When SVM is selected, the kernel function is a radial basis function, the penalty parameter C = 1-10, and the kernel function parameters are... =0.1-1.0.

[0011] A second aspect of the present invention is to provide a bearing current damage identification device based on multimodal manifold retention, comprising: Data acquisition module: used to artificially create damage to the same type of bearing and collect corresponding data, simultaneously collect the full life data of the current damage of this type of bearing, and at the same time collect the fault operation data of the damaged bearing and the normal operation data of the fault-free bearing; Signal processing and sample entropy calculation module: used to denoise the acquired data. After denoising, the acquired data is decomposed into multiple modal signals in the manifold space. The high-dimensional sample entropy of the modal signals in each manifold subspace is calculated to obtain the multimodal manifold sample entropy of normal, fault and current-damaged bearings. The analytical signal construction and feature extraction module is used to construct the specific analytical signal for each bearing state based on the signal corresponding to the sample entropy through Hilbert transform, extract the instantaneous amplitude, phase, frequency and sample entropy statistical features, and construct the initial feature set. Dimensionality Reduction and Recognition Module: This module uses a local preservation projection method to reduce the dimensionality of the current to obtain a low-dimensional mainstream feature set, which is then input into the classifier of the neural network model. By comparing the differences between current damage and other state features, accurate recognition can be achieved.

[0012] Furthermore, the data acquisition module includes: The sensor unit is configured to mount acceleration sensors in the X, Y, and Z directions of the bearing bracket for collecting operational data. Data storage unit, used to classify and store various types of collected data.

[0013] Furthermore, the signal processing and sample entropy calculation module includes: The signal noise reduction unit is configured to use db4 wavelet, 3-5 level decomposition, and VisuShrink's default wavelet threshold to perform noise reduction on the acquired data. The sample entropy calculation unit is configured to calculate sample entropy based on embedding dimension m=2, similarity tolerance r=0.18-0.22 times the standard deviation of the modal signal, and data length N=1000-2000 points; The modal decomposition unit is configured to set the manifold subspace dimension to d = k / 2 - 2k / 3, where k is the total number of modal functions and d is an integer to avoid modal aliasing.

[0014] Furthermore, the analytical signal construction and feature extraction module includes: Analytic signal construction unit: configured as a Hilbert transform unit, used to parse signals and construct them; The feature extraction unit is used to extract statistical features of the sample entropy, including mean, variance, peak factor, kurtosis, skewness, and impulse factor.

[0015] Furthermore, the dimensionality reduction and recognition module includes: Dimensionality Reduction Unit: Configured to construct LPP optimization function - similarity matrix - diagonal matrix - solve mapping function, used for dimensionality reduction of high-dimensional data; Classification and recognition unit: Built-in support vector machine, random forest, and convolutional neural network for classifying and recognizing bearing current damage types.

[0016] Compared with existing technologies, the beneficial effects of this invention are as follows: It artificially creates damage to bearings of the same type and collects corresponding data, simultaneously collecting full-lifetime data of current damage to this type of bearing, as well as fault operation data and normal bearing operation data without faults; by artificially creating damage and simultaneously collecting full-lifetime data, a complete dataset covering normal, fault, and current-damaged bearings is constructed, providing a realistic and comprehensive data foundation for model training; the collected data is denoised, and after denoising, it is decomposed into multiple modal signals in the manifold space, calculating the high-dimensional sample entropy of the modal signals in each manifold subspace to obtain the multimodal manifold sample entropy of normal, fault, and current-damaged bearings; through manifold space decomposition and high-dimensional sample entropy calculation, the differences in bearing states in nonlinear dynamic characteristics are effectively extracted, enhancing the discriminability and noise resistance of state features; based on the sample entropy corresponding signals, through Hil... The BERT transform is used to construct analytical signals specific to each bearing state, extracting instantaneous amplitude, phase, frequency, and sample entropy statistical features to build an initial feature set. The Hilbert transform is then used to construct analytical signals and extract multi-dimensional instantaneous features, enabling in-depth information mining of non-stationary bearing signals and forming a high-dimensional initial feature set. A local preservation projection method is used to reduce dimensionality to obtain a low-dimensional mainstream feature set, which is input into the classifier of a neural network model. The differences between current damage and other state features are compared to achieve accurate identification. Manifold learning dimensionality reduction and neural network classification are employed to achieve accurate classification of high-dimensional features while preserving the local structure of the data, ultimately achieving automated and high-precision identification of current damage states. This solves the problems of difficulty in extracting shaft current damage signals under complex operating conditions of wind turbine main bearings, difficulty in distinguishing them from mechanical faults such as pitting corrosion on the outer ring, unstable identification due to speed fluctuations, and insufficient utilization of full-lifetime data. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This invention provides a schematic flowchart of the steps for identifying bearing current damage based on multimodal manifold maintenance. Figure 2 This invention provides a schematic diagram of the module flow of a bearing current damage identification device based on multimodal manifold maintenance. Detailed Implementation

[0019] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0020] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0021] To address the problems existing in the background technology, a bearing current damage identification method and device based on multimodal manifold maintenance has been designed, which has important practical significance.

[0022] like Figure 1 As shown, the first aspect of the present invention is to provide a bearing current damage identification method based on multimodal manifold retention, comprising the following steps: Step S1: Initiate damage to the same type of bearing and collect corresponding data. Simultaneously collect the full life data of the current damage of this type of bearing, as well as the fault operation data of the damaged bearing and the normal operation data of the bearing without faults.

[0023] Among them, laser electro-etching pens are used to create faults in the inner ring, balls, and outer ring of the same type of experimental bearings, and the full life data of current damage is acquired simultaneously; acceleration sensors are installed in the X, Y, and Z directions of the bearing bracket to collect corresponding fault and normal bearing operation data.

[0024] Step S2: Denoise the collected data. After denoising, the collected data is decomposed into multiple modal signals in the manifold space. Calculate the high-dimensional sample entropy of the modal signals in each manifold subspace to obtain the multimodal manifold sample entropy of the bearings in normal, faulty, and current-damaged conditions.

[0025] The denoising method is wavelet thresholding, with the wavelet basis being db4, the decomposition level being 3-5 levels, and the threshold being the VisuShrink default threshold. The manifold subspace dimension is d=k / 2-2k / 3, where k is the total number of modal functions and d is an integer, ensuring that each subspace corresponds to 1-2 modal functions to avoid aliasing.

[0026] Step S3: Based on the signal corresponding to the sample entropy, construct the specific analytical signal for each bearing state through Hilbert transform, extract the instantaneous amplitude, phase, frequency and sample entropy statistical features, and construct the initial feature set.

[0027] The sample entropy calculation parameters include: embedding dimension m=2, similarity tolerance r=0.18-0.22 times the standard deviation of the modal signal, and data length N=1000-2000 points; the sample entropy statistical features include mean, variance, peak factor, kurtosis, skewness, and impulse factor; the impulse factor is used to distinguish between bearing current damage and mechanical failure.

[0028] Step S4: Use the local preservation projection method to reduce the dimensionality to obtain a low-dimensional mainstream feature set, input it into the classifier of the neural network model, and compare the differences between current damage and other state features to achieve accurate identification.

[0029] The classifier can be a support vector machine, random forest, or convolutional neural network. When SVM is selected, the kernel function is a radial basis function, the penalty parameter C = 1-10, and the kernel function parameters are... =0.1-1.0.

[0030] The data examples from the specific experimental process of this method are shown below: Step 1: Experimental data collection and classification labeling (constructing a full-lifecycle, structured dataset to address the problem of insufficient utilization of full-lifecycle data); 1.1 Fault simulation and data acquisition (high sample realism, covering the entire damage cycle); Three sets of bearings of the same model 6319 were selected: one set of normal bearings (sample A, as the baseline) and two sets of faulty bearings (samples B and C) – to avoid characteristic deviations caused by differences in bearing models; Sample B manufacturing defects: Three 0.2mm×0.1mm pits (X1) were created on the inner ring using a laser etching pen, two 0.15mm×0.1mm pits (X2) were created on one ball bearing, and four 0.2mm×0.1mm pits (X3) were created on the outer ring, completely simulating common mechanical defects of wind turbine main bearings; Sample C collects full-lifetime data: 100V AC voltage is applied to the bearing housing to simulate shaft current (matching the actual amplitude of shaft current in the wind farm), and it is continuously operated for 30 days (covering the damage initiation period: 1-5 days, the development period: 6-20 days, and the failure period: 21-30 days). Sampling is performed 3 times a day (avoiding peak grid loads to ensure high data stability), with each sampling lasting 10 seconds (ensuring that 1500 complete vibration cycles are included). Sensor installation: The sensor is installed on the X-axis (axial), Y-axis (radial), and Z-axis (vertical end face) of the test bench support, consistent with the installation method of the wind farm main bearing seat - to ensure the correlation between experimental data and field data and avoid the problem of "effective in the laboratory but ineffective in the field".

[0031] 1.2 Data classification and labeling (structured storage to improve subsequent processing efficiency); Normal data: The X / Y / Z data of sample A are denoted as A1, A2, and A3, with each group containing 200,000 data points (10s × 20kHz). Mechanical failure data: The failure data of the inner ring / ball / outer ring of sample B are recorded as C1-C5, D1-D5, and B1-B5 respectively (each group is repeated 5 times to avoid random errors). Slight misalignment failures are recorded as X4 (covering more failure scenarios). Axis current damage data: The full lifespan data of sample C is recorded by day as S1-S30 and stored in MATLAB .mat format (path " / Data / CurrentDamage / "). It includes labels such as sampling time, rotation speed, and temperature, which facilitates subsequent analysis of characteristic changes according to damage stages and solves the limitation of existing technology that "only focuses on a single damage state".

[0032] Step 2: Multimodal manifold signal processing and sample entropy calculation (solving the problem of weak signal extraction and quantifying fault characteristics). 2.1 Signal denoising (efficiently filtering out background noise while preserving subtle damage characteristics); Denoising was performed using MATLAB's wden function: wavelet basis was set to 'db4', decomposition level was 4, and threshold rule was 'visushrink'. Effect verification (taking the Z-axis signal of sample C on day 10 as an example): Before noise reduction, the signal-to-noise ratio (SNR) was 12.3dB (tower swaying at 1-5Hz and gearbox interference at 50-100Hz masked the damage signal). After noise reduction, the SNR was 28.5dB (SNR increased by 131.7%), the background noise amplitude decreased from 0.8g to 0.15g, and the characteristic frequency band (200-500Hz) of shaft current damage was clearly presented—verifying the effectiveness of the "db4 wavelet noise reduction" of this invention and solving the pain point of "weak signals being masked" in traditional methods.

[0033] 2.2 Modal decomposition and sample entropy calculation (avoiding aliasing and accurately quantifying fault differences). EMD decomposition: The denoised signal is decomposed into 6 IMF components (IMF1: 500-1000Hz, IMF2: 300-500Hz, IMF3: 200-300Hz, IMF4: 100-200Hz, IMF5: 50-100Hz, IMF6: 20-50Hz), covering all characteristic frequency bands of shaft current damage; Manifold subspace partitioning: Divide into 3 subspaces with d=3 (subspace 1: IMF1-IMF2 high frequency band, subspace 2: IMF3-IMF4 mid frequency band, subspace 3: IMF5-IMF6 low frequency band), each subspace corresponds to 2 IMFs, and the modal aliasing rate is <1% (traditional EMD aliasing rate is >15%). Sample entropy calculation: The `sample_entropy` function is called, and the results are shown in the table below (each data set represents the mean ± standard deviation of 5 replicates):

[0034] The sample entropy of shaft current damage was significantly higher than that of normal bearings (+54.2%) and mechanical faults (+16.7%), and the difference was most obvious in subspace 3 (low frequency band) - indicating that multimodal manifold decomposition can accurately separate the features of different faults, provide quantitative basis for subsequent identification, and solve the problem of "feature ambiguity" in traditional methods.

[0035] Step 3: Signal construction and feature extraction (solving fault aliasing problems and capturing dynamic features); 3.1 Analytical signal construction (adapting to speed fluctuations and highlighting instantaneous characteristics); The MATLAB function Hilbert is used to perform a Hilbert transform on the signals corresponding to the entropy of each subspace sample, yielding the analytical signals: z(t) = s(t) + j·s^(t) (where s(t) is the original modal signal and s^(t) is the result of the Hilbert transform). Extract instantaneous parameters: instantaneous amplitude a(t)=|z(t)| (reflects the impact intensity of damage), instantaneous phase φ(t)=arg(z(t)), instantaneous frequency f(t)=(1 / (2π))·dφ(t) / dt (dynamically tracks the characteristic frequency under rotational speed fluctuations, solving the problem of poor adaptation of traditional fixed frequency algorithms).

[0036] 3.2 Feature Extraction (Introducing exclusive distinguishing features to solve the aliasing problem); Six statistical features are extracted: mean, variance, kurtosis, skewness, and impulse factor (as defined in claim 6). An example is taken comparing a normal bearing (A1), shaft current damage (S10), and outer ring failure (B1).

[0037] The pulse factor of shaft current damage (2.4±0.15) is significantly lower than that of outer ring fault (3.8±0.2), with a difference rate of 36.8%—while the BPFO frequency (characteristic frequency of outer ring fault) of both is 185Hz in traditional envelope analysis, making them indistinguishable. This invention completely solves the core problem of "overlapping envelope spectra of shaft current damage and outer ring fault" by using the unique feature of pulse factor, avoiding inappropriate maintenance strategies due to misjudgment (such as using the wrong outer ring pitting repair scheme to treat shaft current damage, thus saving maintenance costs).

[0038] Step 4: LPP manifold dimensionality reduction and fault identification (improving efficiency and accuracy, adapting to field conditions). 4.1 Feature dimensionality reduction (reduce redundancy and improve recognition efficiency). Initial feature set: Each sample contains "3 subspaces × 6 categories of features = 18-dimensional features", for a total of 100 sets of samples (20 sets of normal samples, 30 sets of mechanical fault samples, and 50 sets of shaft current damage samples). LPP dimensionality reduction process: ① Construct the optimization function: ; ( (These are elements of a similarity matrix, characterizing the correlation between samples). ② Construct a similarity matrix S using k-nearest neighbors (k=5), and the diagonal elements of the diagonal matrix X. ; ③ Under constraints The following solution is obtained, and the eigenvectors corresponding to the first three eigenvalues ​​(λ1=5.2, λ2=3.8, λ3=2.5) are used to construct the mapping function W; Dimensionality reduction effect: The feature dimension was reduced from 18 dimensions to 3 dimensions, the data redundancy was reduced by 83.3%, and the subsequent SVM operation time was shortened from 120s to 15s, which significantly improved the efficiency of engineering applications, while preserving the local structural features of the samples (feature retention rate > 95%).

[0039] 4.2 Fault Identification (High Precision, High Adaptability, Verification of Solution Effectiveness) The classifier is built using MATLAB's fitcsvm function with kernel 'RBF', parameters C=5 and γ=0.5. Comparison of experimental results (with traditional envelope analysis methods):

[0040] The overall recognition accuracy of this invention reaches 98.5%, with only one group of shaft current damage samples being misjudged (due to weak characteristics in the early stage), which is far higher than traditional methods; The misjudgment rate of shaft current and outer ring fault is only 1.2%, completely solving the "aliasing misjudgment" problem of traditional methods; The accuracy rate remains at 97.8% when the rotational speed fluctuates by ±5% (simulating a wind speed change scenario in a wind farm), while the traditional method drops to 72.3%—verifying the working condition adaptability of the "instantaneous frequency tracking + LPP dimensionality reduction" method of this invention and solving the problem of unstable identification caused by rotational speed fluctuations.

[0041] Implementation examples verify conclusions (quantifying technical advantages and demonstrating engineering value); Significantly enhanced weak signal extraction capability: Through "db4 wavelet denoising + multimodal manifold decomposition", the SNR of axial current damage signal was increased from 12.3dB to 28.5dB, and the extraction rate was increased by 131.7%, solving the pain point of "features being masked by noise" in traditional methods, and enabling early identification of damage in the bud stage (days 1-5). Industry-leading fault differentiation accuracy: The pulse factor reduces the false diagnosis rate of shaft current-outer ring faults from 28.5% to 1.2%, avoiding improper maintenance caused by misdiagnosis (such as a loss of 250,000 yuan / set due to incorrect bearing replacement), and reducing the cost of a single operation and maintenance by 320,000 yuan (including downtime losses). The adaptability to operating conditions meets the needs of the field: the accuracy rate is still over 97% when the speed fluctuation is ±5%, and it adapts to the main bearing speed fluctuation caused by wind speed changes in wind farms (actual fluctuation range ±10%), solving the problem of "poor adaptability of fixed frequency algorithm" in traditional methods. The project is highly feasible: all experimental equipment are commercially available products (total procurement cost < 150,000 yuan), the algorithm can be integrated into the existing SCADA system of the wind farm (no hardware modification required), and the data processing latency is < 5 minutes, meeting the wind farm's "real-time monitoring and rapid response" operation and maintenance needs.

[0042] In summary, this embodiment fully verifies the effectiveness, stability, and engineering applicability of the present invention in identifying shaft current damage in wind turbine main bearings. It fully complies with the technical solution defined in the claims and is significantly superior to the prior art, possessing value for large-scale promotion and application. This concludes the embodiment.

[0043] like Figure 2 As shown, a second aspect of the present invention is to provide a bearing current damage identification device based on multimodal manifold retention, comprising: Data acquisition module 101: used to artificially create damage to the same type of bearing and collect corresponding data, simultaneously collect the full life data of the current damage of this type of bearing, and at the same time collect the fault operation data of the damaged bearing and the normal operation data of the bearing without fault.

[0044] The data acquisition module includes a sensor unit and a data storage unit. The sensor unit is configured to mount acceleration sensors in the X, Y, and Z directions of the bearing bracket for collecting operational data. Data storage unit, used to classify and store various types of collected data.

[0045] Signal processing and sample entropy calculation module 102: used to denoise the acquired data. After denoising, the acquired data is decomposed into multiple modal signals in the manifold space. The high-dimensional sample entropy of the modal signals in each manifold subspace is calculated to obtain the multimodal manifold sample entropy of normal, fault and current-damaged bearings.

[0046] The signal processing and sample entropy calculation module includes a signal denoising unit, a sample entropy calculation unit, and a mode decomposition unit. The signal noise reduction unit is configured to use db4 wavelet, 3-5 level decomposition, and VisuShrink's default wavelet threshold to perform noise reduction on the acquired data. The sample entropy calculation unit is configured to calculate sample entropy based on embedding dimension m=2, similarity tolerance r=0.18-0.22 times the standard deviation of the modal signal, and data length N=1000-2000 points; The modal decomposition unit is configured to set the manifold subspace dimension to d = k / 2 - 2k / 3, where k is the total number of modal functions and d is an integer to avoid modal aliasing.

[0047] Analytical signal construction and feature extraction module 103: Based on the sample entropy corresponding signal, it constructs the specific analytical signal of each bearing state through Hilbert transformation, extracts the instantaneous amplitude, phase, frequency and sample entropy statistical features, and constructs the initial feature set.

[0048] The analytical signal construction and feature extraction module includes an analytical signal construction unit and a feature extraction unit. Analytic signal construction unit: configured as a Hilbert transform unit, used to parse signals and construct them; The feature extraction unit is used to extract statistical features of the sample entropy, including mean, variance, peak factor, kurtosis, skewness, and impulse factor.

[0049] Dimensionality reduction and recognition module 104: Used to reduce the dimensionality of the current damage to obtain a low-dimensional mainstream feature set by using the local preservation projection method, input it into the classifier of the neural network model, and compare the differences between current damage and other state features to achieve accurate recognition.

[0050] The dimensionality reduction and recognition module includes a dimensionality reduction unit and a classification and recognition unit. Dimensionality Reduction Unit: Configured to construct LPP optimization function - similarity matrix - diagonal matrix - solve mapping function, used for dimensionality reduction of high-dimensional data; Classification and recognition unit: Built-in support vector machine, random forest, and convolutional neural network for classifying and recognizing bearing current damage types.

[0051] The specific implementation of this device is shown below: This embodiment of the device addresses the shaft current damage identification requirement of the main bearing of a 1.5MW wind turbine (model 6319 deep groove ball bearing, design speed 1500 r / min, rated load 28 kN). Relying on the collaborative work of four modules—experimental data acquisition, multimodal signal processing, signal feature analysis and construction, and LPP dimensionality reduction and classification—it achieves early and accurate identification of shaft current damage under complex operating conditions (including background noise and speed fluctuations of ±5%). This device effectively solves the core pain points of traditional devices—weak signal extraction, difficulty in fault differentiation, and poor adaptability to operating conditions. Actual verification shows that a single maintenance operation can help wind farms reduce costs by 320,000 yuan.

[0052] The device consists of an experimental data acquisition module, a multimodal signal processing module, an analytical signal feature construction module, and an LPP dimensionality reduction and classification module. All modules are connected via a USB 3.0 data bus (with a transmission rate of up to 5Gbps, ensuring real-time data interaction without delay). The core hardware uses commercially available, mature equipment, and the software is developed based on MATLAB R2022b. It can be directly integrated into wind farm SCADA systems without requiring additional modifications to existing operation and maintenance systems, making it highly practical for engineering implementation.

[0053] Specific implementation details for each module; Experimental data acquisition module – laying the foundation for data based on “real working conditions + full life cycle”; This module focuses on data source quality, providing full-cycle data relevant to the actual situation through fault simulation, multi-directional sensing acquisition, and structured storage. It comprises three functional units: Fault simulation unit: equipped with DE-200 laser electro-etching pen (power 50W, spot diameter 0.1mm) and JZ-10 bearing test bench (speed adjustment range ±5%). On the one hand, a laser electro-etching pen was used to create real faults in the 6319 experimental bearing: inner ring faults (0.2mm diameter, 0.1mm depth, 3 evenly distributed along the inner ring raceway), ball faults (2 0.15mm×0.1mm pits on the surface of a single ball), and outer ring faults (4 0.2mm×0.1mm pits on the outer ring raceway), simulating common mechanical faults of main bearings in the field. On the other hand, a 100V AC voltage was applied to the bearing housing to simulate the shaft current environment, and the bearing was run continuously for 30 days to cover the entire damage cycle (1-5 days in the budding stage, 6-20 days in the development stage, and 21-30 days in the failure stage). At the same time, the rotational speed was controlled at 1425-1575r / min on the test bench to match the main bearing rotational speed fluctuation caused by wind speed changes in the wind farm, ensuring the data condition adaptability.

[0054] Sensor Unit: Three PCB352C33 accelerometers (range ±50g, sensitivity 10mV / g) are deployed, installed on the bearing bracket in the X-axis (along the bearing axis), Y-axis (radial), and Z-axis (perpendicular to the bearing end face), respectively, in the same manner as the sensors installed on the main bearing housing in the field. The sensors synchronously collect three-dimensional vibration signals, with a sampling frequency set to 20kHz (preliminary verification shows that this frequency can completely capture the characteristic frequency band of shaft current damage 50-1000Hz). Each sampling lasts for 10 seconds, and data is collected once a day at 8:00, 14:00, and 20:00 to ensure data timeliness and continuity.

[0055] Data classification and storage unit: Based on a Core i7-12700H computer (32GB RAM), paired with an NIcDAQ-9178 data acquisition card (sampling rate 0-256kHz). The acquisition card can simultaneously receive signals from three sensors, avoiding time delay deviations in multi-channel data. The computer classifies and stores the acquired data according to "state-type" in MATLAB.mat format: normal data (denoted as A1-A3, corresponding to X / Y / Z axes), mechanical fault data (inner ring faults C1-C5, ball bearing faults D1-D5, outer ring faults B1-B5, other minor mechanical faults X4), and shaft current damage lifespan data (denoted as S1-S30, one set per day), with the storage path set to " / Data / CurrentDamage / ". Structured storage reduces subsequent module data preprocessing time by 30%, allowing for direct signal analysis.

[0056] Module advantages: It solves the problem of "heavy artificial traces in fault samples and lack of full-cycle data" in traditional devices. The micron-level faults generated by laser electro-erosion are highly consistent with the morphology of on-site shaft current electro-erosion damage. The 30-day full-cycle data covers the complete process of damage from its inception to failure, providing high-quality data support for subsequent feature extraction and identification.

[0057] Multimodal signal processing module – overcoming the pain point of “difficulty in extracting weak signals”; This module addresses the background noise interference problem in wind farms by using layered noise reduction, mode decomposition, and entropy quantization to separate and enhance the characteristics of shaft current damage from complex signals. It comprises three core processing units: Signal denoising unit: Wavelet threshold denoising is implemented based on the MATLAB wden function, with parameters set to db4 wavelet basis, 4 decomposition layers (within the 3-5 layer optimization range), and the default VisuShrink threshold. This scheme can specifically filter out typical background noise in wind farms: tower sway (1-5Hz) and gearbox vibration interference (50-100Hz). Taking the shaft current damage sample (Z-direction signal of sample C on day 10) as an example, the signal-to-noise ratio was only 12.3dB before denoising, and improved to 28.5dB after denoising. The extraction rate of weak damage signals increased by 131.7%, which is far better than traditional low-pass filtering (only 40% improvement), effectively preserving damage characteristics without signal distortion.

[0058] Modal decomposition unit: The noise-reduced signal is decomposed using the EMD algorithm, resulting in 6 mode functions (IMFs). The frequency distribution of each IMF is as follows: IMF1 (500-1000Hz), IMF2 (300-500Hz), IMF3 (150-300Hz), IMF4 (80-150Hz), IMF5 (50-80Hz), and IMF6 (<50Hz). Based on the manifold subspace dimension formula d=k / 2-2k / 3 (where k is the total number of IMFs), we calculate d=3-4. Taking the integer d=3, we divide the 6 IMFs into 3 manifold subspaces: subspace 1 (IMF1-IMF2, high frequency band), subspace 2 (IMF3-IMF4, mid frequency band), and subspace 3 (IMF5-IMF6, low frequency band). Each subspace corresponds to 2 IMFs, completely avoiding the modal aliasing problem of traditional EMD (the aliasing rate of traditional EMD reaches 35%, while the aliasing rate of this module is reduced to 0).

[0059] Sample entropy calculation unit: The sample entropy of the modal signals of each manifold subspace is calculated using the MATLAB function `sample_entropy`, with parameters set as follows: embedding dimension m=2, similarity tolerance r=0.2 times the standard deviation of the modal signal (within the optimized range of 0.18-0.22 times), and data length N=1500 points (within the reasonable range of 1000-2000 points). The calculation results show that the average sample entropy of the shaft current damage sample (S10) is 1.82±0.07, significantly higher than that of normal bearings (1.18±0.04) and bearings with inner ring failures (1.56±0.06), achieving quantitative differentiation of damage characteristics. Furthermore, after 10 sets of repeated experiments, this parameter combination demonstrates the highest feature discrimination, improving discrimination by 25% compared to traditional sample entropy parameters (m=3, r=0.3 times the standard deviation).

[0060] Module advantages: By combining the techniques of "db4 wavelet denoising + multimodal manifold decomposition", the module breaks through the technical bottleneck of "weak axis current signals being masked by background noise" in traditional devices, providing a clear and quantifiable signal basis for subsequent feature extraction.

[0061] Analyzing signal features to build a module—solving the problem of "difficulty in distinguishing fault aliasing"; This module focuses on the overlapping problem of shaft current damage and outer ring pitting corrosion. Through signal analysis and specialized feature extraction, it achieves accurate differentiation of fault types and includes two key units: Analytical signal construction unit: Based on the MATLAB Hilbert function, the Hilbert transform is performed on the sample entropy corresponding signals of each manifold subspace to generate an analytical signal z(t) = s(t) + js^(t) specific to each bearing state (where s(t) is the original modal signal and s^(t) is the Hilbert transform result). The instantaneous amplitude a(t) = |z(t)| and the instantaneous phase φ(t) = arg(z(t)) are extracted from the analytical signal, and the instantaneous frequency f(t) = 1 / (2π) is obtained by differentiating the instantaneous phase. dφ(t) / dt — This instantaneous frequency can dynamically track the characteristic frequency drift under speed fluctuations, solving the problem that traditional fixed-frequency algorithms cannot adapt to speed changes.

[0062] Feature extraction unit: Employs statistical feature algorithms to extract six core features from the analytical signal: mean, variance, peak factor, kurtosis, skewness, and impulse factor. Among these, the impulse factor is a unique distinguishing feature between shaft current damage and outer ring pitting: the impulse factor for the shaft current damage sample (S10) is 2.4±0.15, while the impulse factor for the outer ring fault sample (B1) is 3.8±0.2, with a difference rate of 36.8%. Traditional envelope analysis has a misclassification rate as high as 28.5% for both, but this module reduces the misclassification rate to 1.2% through the impulse factor, completely solving the industry problem of overlapping envelope spectrum features between the two.

[0063] Module advantages: By introducing a pulse factor as a fault-specific distinguishing feature, and by adapting instantaneous frequency to speed fluctuations, it solves the dual pain points of traditional devices, namely "misjudgment of fault type and poor speed adaptation", laying a feature foundation for subsequent accurate identification.

[0064] LPP dimensionality reduction and classification recognition module – achieving “efficient and accurate” recognition; This module improves computational efficiency while maintaining recognition accuracy through high-dimensional feature dimensionality reduction and classifier optimization, and consists of two core units: LPP Dimensionality Reduction Unit: Feature dimensionality reduction is implemented based on the MATLAB LPP toolbox. The specific process is as follows: First, construct the LPP optimization function J(W) (where the optimization function J(W) is the same as above); use the k-nearest neighbor method (k=5) to determine the adjacency relationship between sample points and construct a similarity matrix S; introduce a diagonal matrix X to transform the optimization function; in " Under the constraint of "", the optimization function is solved, and the feature vectors corresponding to the first 3 feature values ​​(λ1=5.2, λ2=3.8, λ3=2.5) are selected to form the mapping function W. The initial feature set is "3 subspaces × 6 types of features = 18 dimensions", which is reduced to 3 dimensions after dimensionality reduction. The data redundancy is reduced by 83.3%, the classifier operation time is shortened from 12s to 2s, and the efficiency is improved by 500%.

[0065] The classification and recognition unit incorporates multiple classifiers (Support Vector Machine (SVM), Random Forest, and Convolutional Neural Network (CNN). This embodiment uses the SVM classifier (using the MATLAB `fitcsvm` function), with parameters set to Radial Basis Function (RBF), penalty parameter C=5 (within the optimal range of 1-10), and kernel parameter γ=0.5 (within a reasonable range of 0.1-1.0). After inputting the low-dimensional feature set into the classifier, the overall recognition accuracy reaches 98.5% (only 1 out of 100 samples with shaft current damage was misclassified as an outer ring fault), with a shaft current damage recognition accuracy of 99.0%. When the test bench speed increases from 1425 r / min (-5%) to 1575 r / min (+5%), the recognition accuracy remains at 97.8%, while the traditional device achieves only 72.3% accuracy under the same speed fluctuation, fully demonstrating the device's adaptability and stability. Furthermore, the classifier supports on-demand switching; if a CNN classifier is selected, the recognition accuracy can be further improved to 99.2%, meeting the requirements for high-precision recognition.

[0066] Module advantages: By using LPP dimensionality reduction, data redundancy is greatly reduced and computing efficiency is improved. Combined with a parameter-optimized classifier, "high efficiency + accuracy" identification is achieved. At the same time, it is adapted to the operating conditions of speed fluctuation, meeting the dual requirements of efficiency and stability for wind farm engineering applications.

[0067] Overall implementation effect and core advantages of the device; Strong signal extraction capability: Relying on the "db4 wavelet denoising + multimodal manifold decomposition" technology, the signal-to-noise ratio of weak signals with shaft current damage is increased from 12.3dB to 28.5dB, and the extraction rate is more than 3 times that of traditional devices, completely solving the problem of "features being masked by noise". High fault differentiation accuracy: With pulse factor as the exclusive differentiation feature, the misjudgment rate of shaft current damage and outer ring pitting is only 1.2%, which is 95.8% lower than that of traditional devices, effectively avoiding inappropriate maintenance strategies due to fault misjudgment; Excellent adaptability to operating conditions: By dynamically tracking speed changes through instantaneous frequency, the recognition accuracy still exceeds 97% under ±5% speed fluctuation, which can cover more than 98% of the operating conditions of wind farms, and its adaptability is far superior to traditional fixed frequency algorithms. The project is highly practical: the core equipment are all commercially available products, with low procurement costs and easy access; the algorithm can be directly integrated into the wind farm's SCADA system without the need to build an additional operation and maintenance platform; according to actual tests at the wind farm, a single operation and maintenance operation can reduce downtime losses and spare parts waste, reducing costs by 320,000 yuan, resulting in significant economic benefits.

[0068] This concludes the embodiment.

[0069] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, optical storage, etc.) containing computer-usable program code.

[0070] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, systems, and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0071] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0072] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0073] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the present invention.

Claims

1. A bearing current damage identification method based on multimodal manifold maintenance, characterized in that, include: Step S1: Initiate damage to the same type of bearing and collect corresponding data. Simultaneously collect the full life data of the current damage of this type of bearing, and at the same time collect the fault operation data of the damaged bearing and the normal operation data of the fault-free bearing. Step S2: Denoise the collected data. After denoising, the collected data is decomposed into multiple modal signals in the manifold space. Calculate the high-dimensional sample entropy of the modal signals in each manifold subspace to obtain the multimodal manifold sample entropy of normal, faulty, and current-damaged bearings. Step S3: Based on the signal corresponding to the sample entropy, construct the specific analytical signal for each bearing state through Hilbert transform, extract the instantaneous amplitude, phase, frequency and sample entropy statistical features, and construct the initial feature set; Step S4: The local preservation projection method is used to reduce the dimensionality to obtain a low-dimensional mainstream feature set, which is then input into the classifier of the neural network model. The difference between current damage and other state features is compared to achieve accurate identification.

2. The bearing current damage identification method based on multimodal manifold maintenance according to claim 1, characterized in that, Step S1 specifically includes: Laser electro-etching pen was used to create faults in the inner ring, ball, and outer ring of the same type of experimental bearing, and the full life data of current damage was acquired simultaneously. Acceleration sensors are installed in the X, Y, and Z directions of the bearing bracket to collect data on the corresponding faults and normal bearing operation.

3. The bearing current damage identification method based on multimodal manifold maintenance according to claim 1, characterized in that, In step S2, the collected data is denoised using wavelet threshold denoising. The wavelet basis is db4, the decomposition level is 3-5 levels, and the threshold is the VisuShrink default threshold. The manifold subspace dimension d = k / 2-2k / 3, where k is the total number of modal functions and d is an integer. This ensures that each subspace corresponds to 1-2 modal functions to avoid aliasing.

4. The bearing current damage identification method based on multimodal manifold maintenance according to claim 1, characterized in that, The parameters for calculating the sample entropy include: The embedding dimension m=2, the similarity tolerance r=0.18-0.22 times the standard deviation of the modal signal, and the data length N=1000-2000 points; the statistical features of the sample entropy include mean, variance, peak factor, kurtosis, skewness, and impulse factor; the impulse factor is used to distinguish between bearing current damage and mechanical failure.

5. The bearing current damage identification method based on multimodal manifold maintenance according to claim 1, characterized in that, In step 4, the classifier can be a support vector machine, random forest, or convolutional neural network. When SVM is selected, the kernel function is a radial basis function, the penalty parameter C = 1-10, and the kernel function parameters are... =0.1-1.

0.

6. A bearing current damage identification device based on multimodal manifold retention, characterized in that, include: Data acquisition module: used to artificially create damage to the same type of bearing and collect corresponding data, simultaneously collect the full life data of the current damage of this type of bearing, and at the same time collect the fault operation data of the damaged bearing and the normal operation data of the fault-free bearing; Signal processing and sample entropy calculation module: used to denoise the acquired data. After denoising, the acquired data is decomposed into multiple modal signals in the manifold space. The high-dimensional sample entropy of the modal signals in each manifold subspace is calculated to obtain the multimodal manifold sample entropy of normal, fault and current-damaged bearings. The analytical signal construction and feature extraction module is used to construct the specific analytical signal for each bearing state based on the signal corresponding to the sample entropy through Hilbert transform, extract the instantaneous amplitude, phase, frequency and sample entropy statistical features, and construct the initial feature set. Dimensionality Reduction and Recognition Module: This module uses a local preservation projection method to reduce the dimensionality of the current to obtain a low-dimensional mainstream feature set, which is then input into the classifier of the neural network model. By comparing the differences between current damage and other state features, accurate recognition can be achieved.

7. The bearing current damage identification device based on multimodal manifold retention according to claim 6, characterized in that, The data acquisition module includes: The sensor unit is configured to mount acceleration sensors in the X, Y, and Z directions of the bearing bracket for collecting operational data. Data storage unit, used to classify and store various types of collected data.

8. The bearing current damage identification device based on multimodal manifold retention according to claim 6, characterized in that, The signal processing and sample entropy calculation module includes: The signal noise reduction unit is configured to use db4 wavelet, 3-5 level decomposition, and VisuShrink's default wavelet threshold to perform noise reduction on the acquired data. The sample entropy calculation unit is configured to calculate sample entropy based on embedding dimension m=2, similarity tolerance r=0.18-0.22 times the standard deviation of the modal signal, and data length N=1000-2000 points; The modal decomposition unit is configured to set the manifold subspace dimension to d = k / 2 - 2k / 3, where k is the total number of modal functions and d is an integer to avoid modal aliasing.

9. A bearing current damage identification device based on multimodal manifold retention according to claim 6, characterized in that, The analytical signal construction and feature extraction module includes: Analytic signal construction unit: configured as a Hilbert transform unit, used to parse signals and construct them; The feature extraction unit is used to extract statistical features of sample entropy, including mean, variance, peak factor, kurtosis, skewness, and impulse factor.

10. A bearing current damage identification device based on multimodal manifold retention according to claim 6, characterized in that, The dimensionality reduction and recognition module includes: Dimensionality Reduction Unit: Configured to construct LPP optimization function - similarity matrix - diagonal matrix - solve mapping function, used for dimensionality reduction of high-dimensional data; Classification and recognition unit: Built-in support vector machine, random forest, and convolutional neural network for classifying and recognizing bearing current damage types.

Images