Method for predicting residual life of wind power rolling bearing based on transfer learning

Abstract

The invention discloses a method for predicting the residual life of a wind power rolling bearing based on transfer learning. The method comprises the following steps: firstly, processing a vibration acceleration signal by using an envelope demodulation algorithm, and screening demodulation characteristics with good monotonicity to construct a bearing health index; thirdly, through a minimum distance sample pairing algorithm, pairing a target domain degradation sample with a source domain full-life sample; and then, constructing a deep convolutional migration neural network, and learning feature extraction and mapping by using paired samples to realize bearing health state estimation. And finally, fitting a state estimation result and combining a failure threshold value to predict the residual life. The method has the beneficial effects that an effective health index construction method is provided, an accurate sample pairing algorithm is designed, a high-performance state estimation model is established, and the prediction precision of the residual life of the wind power bearing is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of power generation equipment state monitoring and remaining life prediction, and in particular to a method for predicting the remaining life of a wind power rolling bearing based on transfer learning. Background Art

[0002] Rolling bearings, as a supporting component of rotating machinery, are widely used in the transmission chain and other rotating parts of wind turbines. The operating performance of rolling bearings has an important impact on the safe operation of wind turbines. If the failure of rolling bearings cannot be predicted in time and proper maintenance is not performed, it will lead to unexpected shutdown of wind turbines or even catastrophic accidents. Therefore, it is of great significance to conduct health status assessment and remaining life prediction of rolling bearings in wind turbines. However, existing research methods have limitations in migration prediction. Migration sample pairing does not take sample similarity into consideration and does not incorporate cross-domain sample similarity into migration learning, which is not conducive to accurate prediction of remaining life. Summary of the invention

[0003] Based on the above problems and the defects of the prior art, the present invention proposes a method for predicting the remaining life of wind power rolling bearings based on transfer learning.

[0004] The present invention comprises the following steps:

[0005] Step 1: Divide several filter frequency bands based on the wavelet packet algorithm, demodulate the demodulation characteristics of the vibration signal of each filter frequency band based on the envelope algorithm, and obtain demodulation characteristics of different scales.

[0006] Step 2: Select demodulation features with good monotonicity as monotonic features, and use monotonicity measurement to quantify the monotonic nature of the demodulation features. The calculation expression is:

[0007]

[0008] Where Mon(n,α) is the monotonicity measure of the envelope spectrum energy characteristics at sampling time n and scale α, and F Y (m,α) is the envelope spectrum energy characteristic at sampling time m and scale α, and ε(.) is a piecewise function.

[0009] Step 3: The degradation increments of monotonic features are fused and normalized to construct a bearing health index with good monotonicity. A degradation increment accumulation algorithm is established. The calculation expression is:

[0010]

[0011] Where D i (n) is the degenerate increment of the i-th monotonic feature at sampling time n, F Z(n,i) is the envelope spectrum energy of the i-th monotonic feature at sampling time n, D(n) is the degradation increment of all monotonic features at sampling time n, FDS is the degradation starting point, and I is the number of monotonic features.

[0012] Step 4: Normalize the fused degradation increments to construct a health indicator with good monotonicity.

[0013] Step 5: Calculate the cross-domain Euclidean distance between each degradation sample in the target domain and the full-life sample in the source domain. The calculation expression is:

[0014]

[0015] Where K is the number of sample dimensions, (x1,x2,…,x K ) and (y1,y2,…,y K ) are the elements of the sample matrices of the target domain and the source domain, respectively.

[0016] Step 6: Search for the source domain sample with the smallest distance to each degraded sample in the target domain to form a sample pair.

[0017] Step 7: Build a deep transfer neural network, use paired samples for training, learn feature extraction and mapping, and realize bearing health status estimation.

[0018] Step 8: Fit the state estimation results and combine them with the failure threshold to predict the remaining life. The calculation expression is:

[0019] RUL(t n )=inf{t:HI(t n +t)≥FT|t n}

[0020] In the formula, RUL(t n ) is the sampling time t n The remaining life, t n is the current moment, HI(t n +t) is the sampling time t n +t is the health indicator, FT is the failure threshold.

[0021] The beneficial effects of the present invention are: a method for constructing a bearing health index with good monotonicity is proposed, a sample pairing algorithm based on minimum distance is designed, a state estimation method based on a deep convolutional transfer neural network is established, and the remaining life prediction accuracy of wind turbine bearings is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 A flowchart showing a method for predicting the remaining life of a wind turbine rolling bearing based on transfer learning.

[0023] Figure 2 Indicates the health indicator construction result.

[0024] Figure 3 Represents the deep transfer neural network structure.

[0025] Figure 4 A graph showing the remaining life prediction results of an instance. DETAILED DESCRIPTION

[0026] A preferred embodiment of the present invention is described in detail below in conjunction with the accompanying drawings so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby making a clearer and more explicit definition of the protection scope of the present invention.

[0027] As shown in the figure, it is a flow chart of a method for predicting the remaining life of a wind turbine rolling bearing based on transfer learning. The data in the embodiment comes from the vibration acceleration signal of the high-speed shaft measurement point of a 2.0MW wind turbine in a wind farm. The data sampling frequency is 25600Hz, the sampling time is 1.28s, the sampling interval is 1 day, and the duration from health to degradation and then to failure is 212 days.

[0028] According to the embodiment, the implementation process of the present invention includes the following steps:

[0029] S1: Based on the wavelet packet algorithm and envelope algorithm, the vibration signal is demodulated to obtain demodulation features of different scales.

[0030] S2: Screen demodulation features with good monotonicity as monotonic features.

[0031] S3: The degraded increments of monotonic features are fused and normalized to construct a bearing health index with good monotonicity, such as Figure 2 shown.

[0032] S4: Calculate the cross-domain Euclidean distance between each degraded sample in the target domain and the full-lifetime sample in the source domain.

[0033] S5: Search for the source domain sample with the smallest distance to each degraded sample in the target domain to form a sample pair.

[0034] S6: Constructing a deep transfer neural network, such as Figure 3 As shown, paired samples are used for training to learn feature extraction and mapping to achieve bearing health status estimation.

[0035] S7: Fit the state estimation results and predict the remaining life in combination with the failure threshold, such as Figure 4 shown.

[0036] The following is a detailed description of each step:

[0037] S1: Based on the wavelet packet algorithm and envelope algorithm, the vibration signal is demodulated to obtain demodulation features of different scales.

[0038] In wind turbines, the vibration signal of rolling bearings contains rich state information, and the extraction of this information is crucial for predicting the remaining service life of bearings. In order to effectively extract this information from the vibration signal, a demodulation method combining wavelet packet algorithm and envelope algorithm can be used.

[0039] 1.1 Wavelet packet algorithm.

[0040] The wavelet packet algorithm is an extension of the wavelet transform, which can analyze the frequency components of the signal in more detail. It divides the frequency band into multiple levels, so as to more accurately capture the characteristics of different frequency bands. In the vibration signal analysis of rolling bearings, the wavelet packet algorithm can divide multiple filter frequency bands, and the signal in each frequency band reflects the vibration characteristics of the bearing under different conditions.

[0041] The wavelet packet algorithm first selects a suitable wavelet basis function, and then decomposes the signal into multiple sub-bands by performing multiple convolution and downsampling operations on the original signal. The signal in each sub-band contains information within a specific frequency range, which is of great significance for subsequent feature extraction and state monitoring.

[0042] 1.2 Envelope algorithm.

[0043] The envelope algorithm is a commonly used signal demodulation technology, which can reflect the amplitude change of the signal by extracting the envelope of the signal. In the vibration signal of the rolling bearing, the envelope signal can more intuitively reflect the vibration state of the bearing.

[0044] In order to obtain the envelope signal, it is usually necessary to perform a Hilbert transform on the original signal. The Hilbert transform is a linear transform that can convert a real signal into an analytical signal. The real part of the analytical signal is the original signal, and the imaginary part is the Hilbert transform result of the original signal. By calculating the modulus of the analytical signal, the envelope of the original signal can be obtained.

[0045] In the vibration signal analysis of rolling bearings, by performing envelope demodulation on the signal in each filter frequency band, demodulation features of different scales can be obtained. These demodulation features reflect the vibration characteristics of the bearing under different conditions, providing a basis for the subsequent construction of health indicators and remaining life prediction.

[0046] S2: Screen demodulation features with good monotonicity as monotonic features.

[0047] After obtaining the demodulated features of different scales, these features need to be screened to find out the monotonic features that can reflect the bearing degradation process. Monotonic features refer to features whose values ​​show monotonic changes (such as monotonic increase or monotonic decrease) as the bearing degradation process proceeds.

[0048] 2.1 Monotonicity measures.

[0049] In order to quantify the monotonic nature of the demodulated features, a monotonicity measure can be introduced. A monotonicity measure is a mathematical expression that calculates the degree of monotonicity of the feature value over time. The monotonicity measure value of the feature value at each sampling time can be calculated, and then these values ​​can be statistically analyzed to evaluate the monotonic nature of the feature.

[0050] The calculation formula of the monotonicity metric can be constructed based on the relative rate of change or the absolute change of the feature value. For example, the difference of the feature values ​​at adjacent sampling moments can be calculated, and the positive and negative values ​​of these differences can be counted to evaluate the monotonicity of the feature. In addition, some smoothing algorithms can be considered to reduce the impact of noise on the monotonicity metric.

[0051] 2.2 Feature screening.

[0052] After obtaining the monotonicity measure of the demodulated features, the monotonic features can be screened according to the value of the monotonicity measure. A threshold can be set to consider features with a monotonicity measure greater than the threshold as monotonic features. These monotonic features can more accurately reflect the degradation process of the bearing and provide a reliable basis for the subsequent construction of health indicators and remaining life prediction.

[0053] It should be noted that feature screening is a complex process that requires consideration of multiple factors. In addition to monotonicity measurement, factors such as feature stability, repeatability, and sensitivity can also be considered to comprehensively evaluate the quality of features. Through comprehensive screening, a set of high-quality monotonic features can be obtained for subsequent health indicator construction and remaining life prediction.

[0054] S3: The degraded increments of monotonic features are fused and normalized to construct a bearing health index with good monotonicity.

[0055] After obtaining the monotonic features, these features need to be further processed to construct a bearing health index with good monotonicity. The health index is a numerical indicator that can reflect the health status of the bearing, and it changes monotonically as the bearing degradation process proceeds.

[0056] 3.1 Degradation increment calculation.

[0057] In order to construct the health index, it is first necessary to calculate the degradation increment of the monotonic feature. The degradation increment refers to the change in the feature value between adjacent sampling moments. By calculating the degradation increment, the slight changes in the feature value over time can be captured, thereby more accurately reflecting the degradation process of the bearing.

[0058] The degradation increment of each monotonic feature at each sampling moment can be calculated, and these degradation increments can be accumulated to obtain the total degradation increment at each sampling moment. The total degradation increment reflects the degree of degradation of the bearing at the current sampling moment and provides a basis for the subsequent construction of health indicators.

[0059] 3.2 Construction of health indicators.

[0060] After the total degradation increment is obtained, it can be normalized to construct a bearing health index with good monotonicity. Normalization is a commonly used data preprocessing technique that can scale the data to a specific range (such as between 0 and 1) to eliminate dimensional differences and numerical differences between different features.

[0061] The total degradation increment can be divided by a reference value (such as the maximum degradation increment or the initial degradation increment) to obtain a normalized health index. This health index changes monotonically as the bearing degradation process proceeds, and can more intuitively reflect the health status of the bearing.

[0062] S4: Calculate the cross-domain Euclidean distance between each degraded sample in the target domain and the full-lifetime sample in the source domain.

[0063] In transfer learning, in order to transfer the knowledge of the source domain to the target domain, it is necessary to find the similarity between the target domain samples and the source domain samples. This similarity can be measured by calculating the distance between the samples. In the present invention, the Euclidean distance is used to calculate the cross-domain distance between each degraded sample in the target domain and the full-life sample in the source domain.

[0064] 4.1 Euclidean distance calculation.

[0065] Euclidean distance is a commonly used distance measurement method, which calculates the straight-line distance between two samples in a multidimensional space. In the present invention, the feature vectors of the target domain sample and the source domain sample are used as coordinate points in the multidimensional space, and then the Euclidean distance between these coordinate points is calculated.

[0066] 4.2 The significance of cross-domain distance.

[0067] The cross-domain distance between the target domain sample and the source domain sample is calculated to find the source domain sample that is most similar to the target domain sample. This similarity measurement is the basis of transfer learning because it allows the knowledge in the source domain to be transferred to the target domain. In the present invention, by calculating the cross-domain distance, the most similar source domain full-life sample can be found for each degraded sample in the target domain, thereby forming a sample pair.

[0068] S5: Search for the source domain sample with the smallest distance to each degraded sample in the target domain to form a sample pair.

[0069] After obtaining the cross-domain distance between the target domain samples and the source domain samples, it is necessary to search for the source domain samples with the smallest distance to each degraded sample in the target domain to form sample pairing. This sample pairing is a key step in transfer learning because it allows the model to be trained using full-life data in the source domain and applied to degraded samples in the target domain.

[0070] 5.1 Search for the minimum distance sample.

[0071] In order to find the source domain sample with the shortest distance to the target domain sample, a nearest neighbor search algorithm can be used. This algorithm searches for the nearest neighbor to the target sample in a multidimensional space, that is, the sample with the shortest distance. In the present invention, an existing nearest neighbor search library can be used to implement this step, such as KDTree or BallTree algorithm in scikit-learn.

[0072] 5.2 The significance of sample pairing.

[0073] Through sample pairing, the full life data in the source domain can be combined with the degradation data in the target domain to form a richer dataset. This dataset contains data from the entire process from health to degradation to failure, which can be used to train more accurate prediction models. In addition, sample pairing can also reduce the domain difference problem in transfer learning and improve the generalization ability of the model.

[0074] S6: Build a deep transfer neural network, use paired samples for training, learn feature extraction and mapping, and realize bearing health status estimation.

[0075] After obtaining the paired samples, a deep transfer neural network needs to be built to learn feature extraction and mapping and realize bearing health status estimation. The deep transfer neural network is a neural network model that combines deep learning and transfer learning. It can use the knowledge in the source domain to improve the prediction performance in the target domain.

[0076] 6.1 Deep transfer neural network structure.

[0077] The structure of a deep transfer neural network usually includes a feature extraction layer, a transfer layer, and a prediction layer. The feature extraction layer is used to extract useful feature representations from the original input data; the transfer layer is used to transfer knowledge from the source domain to the target domain; and the prediction layer is used to predict the output in the target domain based on the extracted features and the transferred knowledge.

[0078] In the present invention, the structure of the deep transfer neural network can be designed according to specific tasks and data sets. For example, a convolutional neural network (CNN) can be used as a feature extraction layer to extract time-frequency features in vibration signals; a fully connected network can be used as a transfer layer to transfer knowledge in the source domain; and a regression network can be used as a prediction layer to predict the health status or remaining service life of a bearing.

[0079] 6.2 Model training and optimization.

[0080] After building a deep transfer neural network, it is necessary to use paired samples for training. During the training process, the back propagation algorithm can be used to optimize the network's weights and bias parameters to minimize the prediction error. In addition, some regularization techniques can be used to prevent overfitting problems, such as L1 / L2 regularization, dropout, etc.

[0081] By training and optimizing the deep transfer neural network, a model that can accurately estimate the health status of the bearing can be obtained. This model can be used to monitor the operating status of wind turbine rolling bearings in real time and detect potential faults in a timely manner.

[0082] S7: Fit the state estimation results and predict the remaining life in combination with the failure threshold.

[0083] After obtaining the estimated result of the bearing health status, it is necessary to fit it and combine it with the failure threshold to predict the remaining service life. The remaining service life refers to the time or number of cycles that the bearing will experience from the current state to the failure state. By predicting the remaining service life, it can provide an important reference for maintenance decisions of wind power rolling bearings.

[0084] 7.1 Fitting of state estimation results.

[0085] In order to fit the state estimation results, some commonly used fitting methods can be used, such as linear regression, polynomial regression, exponential regression, etc. These methods can build a fitting model according to the changing trend of the health state estimation value over time. In the present invention, the fitting method that best suits the data characteristics and prediction requirements can be selected.

[0086] The purpose of the fitting model is to smooth the estimated health status and reduce the impact of noise and random fluctuations, so as to more accurately reflect the degradation trend of the bearing. Through fitting, a continuous health status curve can be obtained, which describes the whole process of the bearing from health to degradation and then to failure.

[0087] 7.2 Determination of failure threshold.

[0088] The failure threshold refers to the threshold at which the bearing is considered to be about to fail or require maintenance when the bearing health status reaches a certain critical value. This threshold can be determined based on actual engineering needs and experience. In the present invention, the failure threshold can be set by analyzing historical data and expert experience.

[0089] Once the failure threshold is determined, it can be combined with the fitted health curve to predict the remaining useful life of the bearing. The point at which the bearing is expected to reach failure can be determined by finding the intersection of the health curve and the failure threshold.

[0090] 7.3 Calculation of remaining life.

[0091] After determining the expected failure time point, the remaining service life of the bearing can be calculated. The calculation formula for the remaining service life can be designed according to actual needs and the characteristics of the prediction model.

[0092] By calculating the remaining service life, an important time reference can be provided for maintenance decisions of wind turbine rolling bearings. For example, when the remaining service life is short, maintenance plans can be arranged in advance to avoid unexpected downtime or accidents caused by bearing failure.

[0093] In the present invention, Figure 1 The flowchart shows the overall process of the remaining life prediction method for wind power rolling bearings proposed in the present invention, including key steps such as signal demodulation, feature screening, health indicator construction, sample matching, neural network construction, state estimation and remaining life prediction.

[0094] Figure 2 : Results of constructing health indicators. This figure shows the good monotonicity of the bearing health indicators constructed by fusing the degradation increments of monotonic features and normalizing them. The monotonic change trend of the health indicators over time can intuitively reflect the health status of the bearing and provide a reliable basis for the subsequent remaining life prediction. By observing the changes in health indicators, the degree of degradation and remaining service life of the bearing can be preliminarily judged, providing an important reference for maintenance decisions of wind turbines.

[0095] Figure 3:Deep transfer neural network structure. This figure depicts the structure of the deep transfer neural network, including key components such as feature extraction layer, transfer layer and prediction layer. By showing the hierarchy and connection mode of the neural network, we can deeply understand how the present invention uses deep learning and transfer learning technology to extract features, transfer knowledge and predict status. The design of the neural network structure fully considers the characteristics of the vibration signal of the wind power rolling bearing and the needs of the remaining life prediction, which helps to improve the prediction accuracy and generalization ability.

[0096] Figure 4 :Example remaining life prediction result diagram. This figure shows the result of predicting the remaining life of an actual wind turbine rolling bearing using the method proposed in the present invention. By comparing the prediction result with the actual life, the effectiveness and accuracy of the present invention can be verified. The example prediction result diagram helps to intuitively feel the performance of the present invention in practical applications and enhance trust and understanding of the technical effects of the present invention.

[0097] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for predicting the remaining life of a wind turbine rolling bearing based on transfer learning, characterized in that: The method specifically comprises the following steps: Step 1: Demodulate the vibration signal based on the wavelet packet algorithm and envelope algorithm to obtain demodulation features of different scales; Step 2: Screen demodulation features with good monotonicity as monotonic features; Step 3: Fuse and normalize the degraded increments of monotonic features to construct a bearing health index with good monotonicity; Step 4: Calculate the cross-domain Euclidean distance between each degraded sample in the target domain and the full-life sample in the source domain; Step 5: Search for the source domain sample with the smallest distance to each degraded sample in the target domain to form a sample pair. Step 6: Build a deep transfer neural network, use paired samples for training, learn feature extraction and mapping, and realize bearing health status estimation; Step 7: Fit the state estimation results and combine them with the failure threshold to predict the remaining life.

2. According to the method for predicting the remaining life of a wind power rolling bearing based on transfer learning as described in claim 1, it is characterized in that: The monotonicity metric is used to quantify the monotonicity of the demodulated features. The calculation expression is: Where Mon(n,α) is the monotonicity measure of the envelope spectrum energy characteristics at sampling time n and scale α, and F Y (m,α) is the envelope spectrum energy characteristic at sampling time m and scale α, ε(.) is a piecewise function.

3. According to the method for predicting the remaining life of a wind power rolling bearing based on transfer learning as described in claim 1, it is characterized in that: A degradation increment accumulation algorithm is established, and the calculation expression is: Where D i (n) is the degenerate increment of the i-th monotonic feature at sampling time n, F Z (n,i) is the envelope spectrum energy of the i-th monotonic feature at sampling time n, D(n) is the degradation increment of all monotonic features at sampling time n, FDS is the degradation starting point, and I is the number of monotonic features.

4. According to the method for predicting the remaining life of a wind power rolling bearing based on transfer learning as described in claim 1, it is characterized in that: The fused degradation increments are normalized to construct a health indicator with good monotonicity.

5. According to the method for predicting the remaining life of a wind power rolling bearing based on transfer learning as described in claim 1, it is characterized in that: The Euclidean distance is calculated for each pair of known target domain samples and full-life source domain samples in the cross-domain feature set, and the source domain sample with the smallest distance to each sample in the target domain is searched to obtain sample pairs; the closer the feature distance between two samples, the greater the similarity between the two, and the more suitable they are for cross-domain migration.

6. According to the method for predicting the remaining life of a wind power rolling bearing based on transfer learning as described in claim 1, it is characterized in that: A deep transfer neural network is constructed to learn feature extraction and mapping. The algorithm process is as follows: first, the feature extraction and indicator mapping of the network are trained using the source domain full life data; Secondly, the paired data of the healthy stage of the source domain and the target domain are used to train the hidden layer feature transfer of the network; Then, the paired data of the degradation phase of the source domain and the target domain are used to train the hidden layer feature transfer of the network; Finally, the trained neural network is used for health state estimation.

Images