A method for predicting the residual life of a wind power rolling bearing by using incremental mapping and recursive estimation

Abstract

The application discloses a method for predicting the residual life of wind power rolling bearings by using incremental mapping and recursive estimation, and relates to renewable energy equipment state monitoring. First, a multi-band envelope demodulation two-dimensional spectrum diagram is obtained by transforming a vibration acceleration signal, a spectrum diagram variation is calculated and normalized, a health index and a normalized degradation increment are obtained. Second, a channel attention module is used to extract features from two dimensions of the two-dimensional spectrum diagram, and a two-dimensional convolutional neural network is designed to learn the nonlinear mapping of the spectrum diagram and the degradation increment. Third, a recursive deep learning model is established, model weight Gaussian sampling is combined, the posterior probability of future health indexes is calculated, and rolling prediction is realized. Finally, the residual life is predicted according to a failure threshold. The application has the advantages that a health index based on a multi-band envelope demodulation two-dimensional spectrum diagram variation increment is constructed, a two-dimensional convolutional neural network containing a channel attention module is designed, recursive deep learning and weight Gaussian sampling are fused, and the prediction accuracy of the residual life of wind power bearings is improved.

Description

Technical Field

[0001] This invention relates to the field of power generation equipment condition monitoring and remaining life prediction, and in particular to a method for predicting the remaining life of wind turbine rolling bearings using incremental mapping and recursive estimation. Background Technology

[0002] Rolling bearings are widely used in the main shaft, gearbox, and generator systems of wind turbine transmission systems. Operating in harsh environments and subjected to alternating loads and random excitations, they are prone to failure and continuous degradation, leading to a gradual decline in performance. Failure to predict bearing failure in a timely manner and implement appropriate maintenance can result in unexpected wind turbine shutdowns or even catastrophic accidents. Deep learning algorithms, with their powerful feature extraction and nonlinear expression capabilities, are widely used in the prediction of the remaining life of bearings and other electromechanical equipment. However, existing research has limitations in incremental mapping and recursive estimation. Most normalized health indicators do not reflect the dynamic degradation process of bearings, and traditional recursive deep learning cannot calculate the posterior probability of the degradation state, affecting the accuracy and robustness of wind turbine bearing remaining life prediction. Summary of the Invention

[0003] Based on the above problems and the shortcomings of existing technologies, this invention proposes a method for predicting the remaining life of wind turbine rolling bearings using incremental mapping and recursive estimation, which mainly predicts the remaining life by estimating the development trend of health indicators.

[0004] This invention includes the following steps:

[0005] Step 1: Based on the wavelet packet algorithm, divide the vibration signal into several filter bands with different center frequencies and bandwidths. Perform narrow-band envelope demodulation on the vibration signal for different frequency bands to obtain the envelope spectrum. Then, stitch the spectrum together in order of increasing center frequency of each filter to form a multi-band envelope demodulation two-dimensional spectrum for degradation information extraction.

[0006] Step 2: Construct the dynamic degradation increment using the changes in the two-dimensional spectrum at different time points. The calculation expression is as follows:

[0007] X = {x1, x2, ..., x} K}

[0008]

[0009] In the formula, X represents the spectrum of the bearing's full-life degradation process, K is the number of sampling times for the bearing's full-life degradation process, and Y... (a) The expression represents the absolute increment, |.| represents taking the absolute value, and sum(.) represents summing all elements of the matrix.

[0010] Step 3: By subtracting the average value of the healthy phase from the absolute increment, the impact of the inherent characteristics of each bearing on the construction of the health index is reduced. The calculation expression is as follows:

[0011]

[0012] In the formula, FDT represents the degradation initiation point, Y (f) This represents the baseline increment, and mean(.) represents calculating the average value.

[0013] Step 4: Normalize the health indicators for the deterioration stage, and construct normalized health indicators and corresponding normalized increments.

[0014] Step 5: The convolutional neural network first uses the channel attention module to perform compression and excitation attention operations on the two dimensions of the two-dimensional spectrum, namely the band and frequency, to obtain weight vectors. These weight vectors are then multiplied with the original spectrum to obtain a weighted spectrum. Convolution is then used to extract deep degradation features and learn the nonlinear mapping relationship between features and degradation increments to achieve increment mapping estimation.

[0015] Step 6: The recurrent neural network first trains the network weights using the training set data, and then calculates the likelihood estimate of the health indicators at future times by Gaussian sampling of the network weights, thus obtaining the posterior probability of the health indicators.

[0016] Step 7: Obtain posterior estimates of health indicators for future timeframes through recursive rolling prediction, and combine this with a failure threshold to predict remaining lifespan. The calculation expression is as follows:

[0017]

[0018] In the formula, t now Indicates the current sampling time, FT represents the failure threshold, and x t This represents the system state at sampling time t.

[0019] The beneficial effects of this invention are as follows: a method for constructing health indicators based on the incremental change of two-dimensional spectra in multi-band envelope demodulation is proposed; a two-dimensional convolutional neural network containing a channel attention module is designed to learn the nonlinear mapping relationship of features; a recursive deep learning model and weighted Gaussian sampling are integrated to enhance the rolling prediction capability of the model and improve the prediction accuracy of the remaining life of wind turbine bearings. Attached Figure Description

[0020] Figure 1 The flowchart shows a method for predicting the remaining life of wind turbine rolling bearings based on incremental mapping and recursive estimation.

[0021] Figure 2 This represents the degenerative dynamic normalized health index constructed by this method and the corresponding degenerative increment.

[0022] Figure 3 This represents the attention convolutional neural network structure constructed using this method.

[0023] Figure 4 This is a graph showing the predicted remaining lifetime of the instance. Detailed Implementation

[0024] A preferred embodiment of the present invention will now be described in detail with reference to 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 providing a clearer and more explicit definition of the scope of protection of the present invention.

[0025] like Figure 1 The diagram shows a flowchart of a method for predicting the remaining life of wind turbine rolling bearings based on incremental mapping and recursive estimation. The data in this example comes from vibration acceleration signals from high-speed shaft measuring points of a 2.0MW wind turbine generator set in a wind farm. The data sampling frequency is 25600Hz, the sampling duration is 1.28s, the sampling interval is 1 day, and the recorded duration from healthy to degradation to failure is 187 days.

[0026] According to the embodiments, the implementation process of the present invention includes the following steps:

[0027] S1: Divide the vibration signal into several filter bands with different center frequencies and bandwidths. Perform narrow-band envelope demodulation on the vibration signal for each frequency band to obtain the envelope spectrum. Then, stitch the spectrum together in order of increasing center frequency of each filter to form a two-dimensional spectrum of multi-band envelope demodulation.

[0028] S2: Utilize the changes in the two-dimensional spectrum at different times to construct dynamic degradation increments, and accumulate them to obtain dynamic health indicators.

[0029] S3: Normalize health indicators for the deterioration stage, construct normalized health indicators and corresponding normalized increments, such as... Figure 2 As shown.

[0030] S4: Design a two-dimensional convolutional neural network to extract features from a two-dimensional spectral map and learn the nonlinear mapping relationship between features and degradation increments. The network structure is as follows: Figure 3 As shown.

[0031] S5: Build a recursive deep learning model and train the network weights using the training set data.

[0032] S6: Integrate Gaussian sampling of network weights to calculate the posterior probability of health indicators at future moments, thereby enabling rolling prediction of degradation states.

[0033] S7: Combine the recursive posterior estimation results of the degradation state at future time points with the preset failure threshold to predict the remaining lifetime.

[0034] The detailed explanations of steps S1-S7 above are as follows:

[0035] Step 1: Construct a two-dimensional spectrum for multi-band envelope demodulation:

[0036] First, based on the wavelet packet algorithm, several filter frequency bands with different center frequencies and bandwidths are divided. These frequency bands should cover the range of possible fault frequencies in the rolling bearing.

[0037] Narrowband envelope demodulation: For each filtered frequency band, the vibration signal is demodulated using a narrowband envelope to obtain the envelope spectrum. The envelope spectrum can reflect the vibration characteristics of the rolling bearing within the corresponding frequency band.

[0038] Two-dimensional spectrum splicing: The resulting envelope spectra are spliced ​​together in ascending order of the center frequencies of each filter to form a multi-band envelope demodulation two-dimensional spectrum. This two-dimensional spectrum can comprehensively reflect the vibration characteristics of the rolling bearing, providing a basis for subsequent degradation increment calculations.

[0039] Step 2: Construct dynamic health indicators:

[0040] Calculate the degradation increment: Construct a dynamic degradation increment using the changes in the two-dimensional spectrum at consecutive time points. Specifically, calculate the difference between the two-dimensional spectra at adjacent time points, and sum the absolute values ​​to obtain the absolute increment.

[0041] Cumulative health index: The absolute increments are accumulated to obtain a dynamic health index. This index reflects the degree of degradation of the rolling bearing and gradually increases as the degradation process progresses.

[0042] Step 3: Normalized Health Indicators (NHI):

[0043] Determine the degradation initiation point: In the degradation process of rolling bearings, a degradation initiation point is determined, which is usually selected at the moment when the performance of the rolling bearing begins to decline significantly.

[0044] Calculate the baseline increment: Before the point of degradation, calculate the average absolute increment of the healthy phase as the baseline increment.

[0045] Normalization: The dynamic health index is subtracted from the baseline increment and then normalized to obtain the normalized health index. Simultaneously, the normalized increment is calculated to reflect the degradation rate of the rolling bearing.

[0046] Step 4: Design a two-dimensional convolutional neural network:

[0047] Feature extraction: A two-dimensional convolutional neural network is used to extract features from the multi-band envelope demodulated two-dimensional spectrum. The convolutional layers in the network can learn deep degradation features in the spectrum.

[0048] Incremental mapping estimation: Design a network structure that can learn the nonlinear mapping relationship between features and degradation increments. By training the network, incremental mapping estimation is achieved, that is, predicting the normalized increment based on the extracted features.

[0049] Step 5: Build a recursive deep learning model:

[0050] Constructing recursive networks: Establish recursive deep learning models, such as Long Short-Term Memory (LSTM) networks or Gated Recurrent Units (GRUs), to capture the temporal dependencies in the degradation process of rolling bearings.

[0051] Training network weights: Using the training set data, the network weights are trained through the backpropagation algorithm, enabling the recursive network to accurately predict health indicators at future moments.

[0052] Step 6: Implement rolling prediction of degradation state:

[0053] Fusion Gaussian Sampling: During the training of a recurrent network, Gaussian sampling of the network weights is fused to simulate the network output under different parameter combinations. This helps to enhance the robustness and generalization ability of the model.

[0054] Calculating posterior probabilities: Using Gaussian sampling of network weights, the posterior probability distribution of health indicators at future times is calculated. This reflects the uncertainty of the prediction results and provides a reliable basis for rolling predictions.

[0055] Rolling forecasting: Based on current health indicators and predicted future health indicators, rolling forecasts are achieved. During the rolling forecasting process, the forecasting model is continuously updated using new observational data to improve forecast accuracy.

[0056] Step 7: Predict remaining lifespan:

[0057] Setting a failure threshold: Based on the failure mechanism and performance degradation law of rolling bearings, a failure threshold is set. When the predicted health indicators fall below this threshold, the rolling bearing is considered to have failed.

[0058] Calculating remaining lifetime: The remaining lifetime of the rolling bearing is calculated by combining the recursive posterior estimation results of the degradation state at future time points with a preset failure threshold. Specifically, the remaining lifetime is determined based on the time difference between the predicted trend of health indicators and the failure threshold.

[0059] The following provides a detailed explanation of each step:

[0060] S1: Divide the filter frequency band and generate a two-dimensional spectrum.

[0061] First, it is necessary to divide the signal into several filter bands with different center frequencies and bandwidths based on the wavelet packet algorithm. In this embodiment, the vibration signal is divided into 8 different frequency bands, each with a different center frequency and bandwidth. The setting of these frequency bands is determined based on the analysis of the characteristic frequencies of rolling bearing faults and prior knowledge.

[0062] Next, narrowband envelope demodulation is performed on the vibration signals for different frequency bands. Narrowband envelope demodulation is a signal processing technique used to extract the envelope component from the rolling bearing vibration signal, which contains fault information of the rolling bearing. By performing narrowband envelope demodulation on the vibration signal of each frequency band, the corresponding envelope spectrum can be obtained.

[0063] Finally, the envelope spectra of each frequency band are stitched together in ascending order of the filter center frequencies to form a multi-band envelope demodulation two-dimensional spectrum. This two-dimensional spectrum contains rich information about the rolling bearing's degradation process throughout its entire lifespan and can be used for subsequent degradation analysis and remaining life prediction.

[0064] S2: Constructing dynamic degradation increments and health indicators

[0065] After obtaining the two-dimensional spectrum, the changes in the spectrum between consecutive time points need to be used to construct the dynamic degradation increment. Specifically, the absolute increment between the two-dimensional spectra at two adjacent time points can be calculated, which is the sum of the absolute values ​​of the differences between corresponding elements in the two-dimensional spectra at the two time points. This absolute increment reflects the degree of degradation of the rolling bearing at the current time point relative to the previous time point.

[0066] The dynamic degradation increment is constructed by using the changes in the two-dimensional spectrum at different time points. The calculation expression is as follows:

[0067] X = {x1, x2, ..., x} K}

[0068]

[0069] In the formula, X represents the spectrum of the bearing's full-life degradation process, K is the number of sampling times for the bearing's full-life degradation process, and Y... (a) The expression represents the absolute increment, |.| represents taking the absolute value, and sum(.) represents summing all elements of the matrix.

[0070] Then, these absolute increments are accumulated to obtain a dynamic health index. This dynamic health index reflects the cumulative degree of degradation of the rolling bearing from a healthy state to the current moment. It is important to note that during the accumulation process, the absolute increments need to be normalized to eliminate dimensional and numerical range differences between different frequency bands.

[0071] S3: Normalized Health Indicators and Increments

[0072] After obtaining the dynamic health indicators, they need to be normalized for the degradation stage. The purpose of normalization is to map the value range of the health indicators to a fixed range for subsequent prediction and analysis. In this embodiment, the value range of the health indicators is mapped to the range [0,1].

[0073] By subtracting the average value of the healthy phase from the absolute increment, the impact of the inherent characteristics of each bearing on the construction of the health index is reduced. The calculation expression is as follows:

[0074]

[0075] In the formula, FDT represents the degradation initiation point, Y (f) This represents the baseline increment, and mean(.) represents calculating the average value.

[0076] Specifically, the average health index value of the rolling bearing during the healthy phase can be calculated first as a baseline value. Then, the health index value at each moment is compared with the baseline value, and the relative value is calculated. Finally, these relative values ​​are normalized to obtain the normalized health index.

[0077] Additionally, the increment of the normalized health index can be calculated, which is the difference between the normalized health index at two adjacent time points. This increment reflects the rate of degradation of the rolling bearing at the current time point relative to the previous time point.

[0078] S4: Design a 2D Convolutional Neural Network

[0079] To learn the nonlinear mapping between two-dimensional spectral features and degradation increments, a two-dimensional convolutional neural network (CNN) needs to be designed. This network can automatically extract deep features from the two-dimensional spectral image and learn the mapping between these features and degradation increments.

[0080] In this embodiment, the designed two-dimensional CNN structure is as follows: Figure 3 As shown, the network consists of multiple convolutional layers, pooling layers, fully connected layers, and an output layer. The convolutional layers extract local features from the two-dimensional spectral map; the pooling layers reduce the dimensionality of the feature map and decrease computational cost; the fully connected layers map the extracted features to the degradation increment output space; and the output layer outputs the prediction result.

[0081] Furthermore, a channel attention module is introduced to enhance the network's focus on important features in the two-dimensional spectrogram. The channel attention module performs compression and excitation attention operations on both the bandwidth and frequency dimensions of the two-dimensional spectrogram, obtaining weight vectors, which are then multiplied sequentially with the original spectrogram to obtain a weighted spectrogram. This allows the network to pay closer attention to features that significantly influence the prediction of degradation increments.

[0082] S5: Building a recursive deep learning model

[0083] After obtaining the two-dimensional CNN, a recurrent deep learning model needs to be built to predict health indicators at future moments. A recurrent neural network (RNN) is a neural network structure suitable for processing sequential data, capable of capturing the temporal dependencies within the sequence.

[0084] In this embodiment, a Long Short-Term Memory (LSTM) network is chosen as the basic structure of the recursive deep learning model. LSTM is a special type of RNN structure that overcomes the shortcomings of traditional RNNs in dealing with long-term dependencies by introducing gating mechanisms and memory units.

[0085] The weights of the LSTM network are trained using the training set data. During training, features extracted by a 2D CNN are used as input to the LSTM network, and normalized health indicators are used as output. By optimizing the network weights, the LSTM network can learn the mapping relationship between features and health indicators.

[0086] To improve the generalization ability and prediction accuracy of the LSTM network, a Gaussian sampling method for weights is employed. Specifically, during training, the weights of the LSTM network are sampled using a Gaussian distribution to simulate the network output under different parameter combinations. This method yields multiple prediction results, and the mean and variance of these results are calculated to obtain the posterior probability distribution of the health indicator at future time points.

[0087] S6: Enable rolling prediction of degradation state

[0088] After obtaining the posterior probability distribution of health indicators at future times, rolling predictions of the degenerative state can be achieved. Specifically, the rolling prediction error can be calculated based on the current health indicators and the predicted future health indicators, and the prediction model can be adjusted according to the error.

[0089] The rolling forecasting process is iterative. In each iteration, the forecasting model is updated using the latest observational data and forecast results, and the predicted values ​​of health indicators for the next time step are calculated. In this way, the actual degradation process can be gradually approximated, improving the accuracy and reliability of the forecast.

[0090] Furthermore, the uncertainty of rolling forecasts can be calculated based on the results of rolling forecasts. Uncertainty reflects the reliability and confidence level of the forecast results. By comparing the forecast uncertainties at different times, the stability and robustness of the forecast model can be evaluated.

[0091] S7: Predicting Remaining Lifetime

[0092] Finally, the remaining life of the rolling bearing can be predicted by combining the recursive posterior estimation results of the degradation state at future moments with a preset failure threshold. Specifically, the predicted future health index can be compared with the failure threshold. When the predicted health index is lower than the failure threshold, the rolling bearing is considered to have failed. At this point, the remaining life can be calculated based on the time difference between the current moment and the failure moment.

[0093] It is important to note that the failure threshold is determined based on the failure mechanism and performance degradation characteristics of rolling bearings. In practical applications, an appropriate failure threshold needs to be selected according to the specific type of rolling bearing and the operating environment.

[0094] Furthermore, the failure threshold can be updated using the results of rolling predictions. As the degradation process of rolling bearings continues, their failure mechanisms may change. Therefore, it is necessary to dynamically adjust the failure threshold based on the latest observation data and prediction results to ensure the accuracy and reliability of the prediction results.

[0095] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting the remaining life of wind turbine rolling bearings using incremental mapping and recursive estimation, characterized in that, The method specifically includes the following steps: Step 1: Divide the vibration signal into several filter bands with different center frequencies and bandwidths. Perform narrow-band envelope demodulation on the vibration signal for each frequency band to obtain the envelope spectrum. Then, stitch the spectrum together in order of increasing center frequency of each filter to form a two-dimensional spectrum of multi-band envelope demodulation. Step 2: Construct dynamic degradation increments using the changes in the two-dimensional spectrum at different time points, and accumulate them to obtain dynamic health indicators; Step 3: Normalize health indicators for the deterioration stage, and construct normalized health indicators and corresponding normalized increments; Step 4: Design a two-dimensional convolutional neural network to extract features from the two-dimensional spectral map and learn the nonlinear mapping relationship between features and degradation increment; Step 5: Build a recursive deep learning model and train the network weights using the training set data; Step 6: Combine network weights with Gaussian sampling to calculate the posterior probability of health indicators at future time points, and realize rolling prediction of degradation state; Step 7: Combine the recursive posterior estimation results of the degradation state at future time points with the preset failure threshold to predict the remaining lifetime; in: In step 1, several filter frequency bands with different center frequencies and bandwidths are divided based on the wavelet packet algorithm. Narrow-band envelope demodulation is performed on the vibration signal for different frequency bands to obtain the envelope spectrum. The spectrum is then stitched together in order from low to high filter center frequencies to form a multi-band envelope demodulation two-dimensional spectrum for degradation information extraction. In step 2, the dynamic degradation increment is constructed using the changes in the two-dimensional spectrum at different time points. The calculation expression is as follows: X={x1,x2,...,x K } In the formula, X represents the spectrum of the bearing's full-life degradation process, K is the number of sampling times for the bearing's full-life degradation process, and Y... (a) The expression represents the absolute increment, |.| represents taking the absolute value, and sum(.) represents summing all elements of the matrix. In step 3, by subtracting the average value of the healthy phase from the absolute increment, the impact of the inherent characteristics of each bearing on the construction of the health index is reduced. The calculation expression is as follows: In the formula, FDT represents the degradation initiation point, Y (f) This represents the baseline increment, and mean(.) represents calculating the average value.

2. The method for predicting the remaining life of wind turbine rolling bearings using incremental mapping and recursive estimation as described in claim 1, characterized in that, Health indicators are normalized for the deterioration stage, and normalized health indicators and corresponding normalized increments are constructed.

3. The method for predicting the remaining life of wind turbine rolling bearings using incremental mapping and recursive estimation as described in claim 1, characterized in that, Convolutional neural networks first use channel attention modules to perform compression and excitation attention operations on the two dimensions of the two-dimensional spectrum, namely the band and frequency, to obtain weight vectors. These weight vectors are then multiplied by the original spectrum to obtain a weighted spectrum. Convolution is then used to extract deep degradation features and learn the nonlinear mapping relationship between the features and the degradation increment to achieve increment mapping estimation.

4. The method for predicting the remaining life of wind turbine rolling bearings using incremental mapping and recursive estimation as described in claim 1, characterized in that, Recurrent neural networks first train network weights using training set data, and then calculate the likelihood estimate of health indicators at future times by Gaussian sampling of the network weights, thus obtaining the posterior probability of the health indicators.

Images