A bearing degradation trend prediction method based on multi-resolution feature extraction and Bi-LSTM network

Abstract

The present invention provides a bearing degradation trend prediction method based on multi-resolution feature extraction and Bi-LSTM network. It aims to solve the problem that the bearing vibration signal contains a large amount of random noise in complex working conditions and environments, which affects the bearing degradation trend prediction accuracy. It fully considers the impact of historical vibration state on current vibration. The main steps are as follows: applying the full life cycle vibration signal of the accelerated bearing degradation test, performing time domain analysis on it to extract multi-angle eigenvalues ​​to form a feature set, screening sensitive features according to the trend and correlation of the features, using the multi-resolution singular value decomposition method to decompose and reconstruct the sensitive features, inputting the normalized sensitive features into the constructed Bi-LSTM network for training, and predicting the bearing degradation trend. Its purpose is to provide a prediction method that fully considers the impact of the historical vibration state of the bearing on the current moment, eliminates random interference during operation, retains important fault information in the bearing degradation process, and improves the accuracy of bearing degradation trend prediction.

Description

Technical Field

[0001] The present invention belongs to the field of fault prediction and health management, and specifically relates to a bearing degradation trend prediction method. Background Art

[0002] Bearings play an irreplaceable role in the industrial applications of rotating machinery. However, actual operating conditions are often extremely harsh, making bearings highly susceptible to performance degradation. Mechanical system management based on predictive maintenance is extremely important. Reliable and accurate prediction of bearing degradation processes is of great practical significance for the safe, stable, and efficient operation of rotating machinery.

[0003] Traditional bearing degradation trend prediction methods mainly focus on the impact of the current vibration state on the future vibration state. However, in actual work, the historical vibration state of the bearing is also closely related to the future vibration state, and both have a certain indicative effect on the current vibration state. The existing recurrent neural network used to predict bearing degradation trends, due to the special structure of the hidden layer, although it connects the time series information of the samples in time, it will produce gradient instability problems (gradient disappearance or explosion) when learning long-term information, and does not fully consider the impact of subsequent vibration anomalies on future states. In addition, the existing health indicators used for degradation trend prediction are often simple linear types, which do not fully consider the fault vibration characteristics of the bearing, resulting in low accuracy in bearing degradation trend prediction. Summary of the Invention

[0004] The purpose of the present invention is to provide a prediction method that fully considers the correlation between the historical and future vibration states of a bearing as a reference for current vibration, eliminates random interference from irrelevant noise during operation, retains important fault information during bearing degradation, and effectively and accurately predicts bearing degradation trends. To achieve the above objectives, the technical solutions adopted by the present invention are as follows:

[0005] A bearing degradation trend prediction method based on multi-resolution feature extraction and Bi-LSTM network is characterized by comprising the following steps:

[0006] S1: Obtain the vibration acceleration signal from the beginning to complete failure of the bearing accelerated life test;

[0007] S2: Perform time domain analysis on the vibration signal, extract multi-angle eigenvalues, and preliminarily construct a feature set that comprehensively describes the operating status of the bearing;

[0008] S3: Analyze the trend and correlation of the feature set and select appropriate features;

[0009] S4: Input the feature into the multi-resolution singular value decomposition method, use monotonicity to determine the appropriate number of decomposition layers, and construct sensitive features;

[0010] S5: After normalizing the sensitive features to [0,1], the starting prediction point is determined by the μ+3σ criterion, and the training set and prediction set are divided;

[0011] S6: Build a Bi-LSTM network framework and use the normalized sensitive features as input to obtain health indicators that comprehensively represent the operating status of the bearing.

[0012] S7: training the Bi-LSTM network using the training set and verifying the prediction set;

[0013] S8: The trained and predicted health indicators are denormalized and converted into predicted bearing degradation trends, and model evaluation indicators are constructed to evaluate the accuracy of the predicted degradation trends.

[0014] Preferably, in step S2, the time domain features included in the time domain analysis include root mean square value, variance, peak value, mean value, absolute mean value, and kurtosis.

[0015] Preferably, in step S3, the trend indicator is defined as The relevance index is defined as Finally, select the features with better physical properties (such as root mean square value).

[0016] Preferably, in step S4, multi-resolution singular value decomposition is performed on the selected image, and the multi-resolution singular value decomposition result is selected according to the principle that the monotonicity of the approximate component of a certain layer (such as the fourth layer) is the best, and the decomposition is stopped to obtain the approximate component after removing the random noise as the sensitive feature. The monotonicity index is defined as

[0017] Preferably, in step S5, the normalization processing model is Fit the linear regression model to the root mean square value RMS = wt + b where The coefficient w represents the slope of the RMS value relative to time t, and the absolute value of w is used to establish the threshold. The threshold is determined by the mean μ and standard deviation σ of the bearing's healthy phase data. The threshold is determined by calculating μ + 3σ. The starting prediction point is determined when the newly calculated w exceeds the threshold for n consecutive times (for example, n = 5). The data before the starting prediction point is divided into the training set, and the data after the starting prediction point is divided into the test set.

[0018] Preferably, in step S6, the Bi-LSTM network (bidirectional long short-term memory network) includes an input layer, multiple hidden layers, and an output layer. Backpropagation and gradient descent algorithms are used, the Adam optimization algorithm is used for training, the root mean square error is used as the loss function, the tanh function is used as the output layer activation function, and other parameters are set to default values. Training is performed for 300 rounds.

[0019] Preferably, in step S8, the denormalization model is x_scaled=x_std×(x max -x min )+x min The present invention uses root mean square error (RMSE), mean absolute error (MAE), correlation coefficient (R 2 ) and correlation (Con) to evaluate the prediction accuracy and compare it with the LSTM network (Long Short-Term Memory Network).

[0020] Compared with the prior art, the present invention has the following beneficial effects:

[0021] 1) The present invention adopts a multi-resolution singular value decomposition method to identify irrelevant vibrations in bearing operation, eliminate random abnormal fluctuations, and construct a root mean square value that is more consistent with the degradation trend in terms of trend and physical properties as a sensitive feature for subsequent degradation trend prediction. Compared with existing neural network predictions, the prediction accuracy is improved.

[0022] 2) The present invention fully considers the correlation between the historical vibration state and the future vibration state of the bearing for predicting the vibration trend at the current moment in the long term. It retains more important fault information when processing bearing degradation sequence data and has broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below with reference to the accompanying drawings, in which:

[0024] Figure 1 This is a flow chart of feature extraction based on multi-resolution singular value decomposition;

[0025] Figure 2 A flowchart of a bearing degradation trend prediction method based on a Bi-LSTM network with multi-resolution singular value decomposition provided in a preferred embodiment of the present invention;

[0026] Figure 3 for Figure 2 The training results and prediction results of the Bi-LSTM network;

[0027] Figure 4The prediction results of bearing degradation trend are compared using Bi-LSTM network based on multi-resolution singular value decomposition and existing LSTM network. DETAILED DESCRIPTION

[0028] The following describes embodiments of the present invention with reference to schematic diagrams and preferred specific examples. Those skilled in the art will readily understand the other advantages and features of the present invention from the disclosure herein. The present invention may also be implemented or applied through various other specific embodiments, and various details may be modified or altered based on different viewpoints and applications without departing from the spirit of the present invention. Therefore, the following description should be understood as a general guide for those skilled in the art and is not intended to limit the present invention.

[0029] See Figures 1 to 4 The RMS value of a bearing vibration signal can, to a certain extent, reflect the state of the vibration signal and the bearing degradation trend. The RMS value is widely used to evaluate vibration signals and, as a result, effectively reflects bearing degradation trends. Therefore, this paper selects the RMS value of the bearing vibration signal as a sensitive feature for subsequent degradation trend prediction.

[0030] like Figure 1 、 Figure 2 As shown, a bearing degradation trend prediction method based on multi-resolution feature extraction and Bi-LSTM network includes steps S1 to S9, which are as follows:

[0031] S1: Obtain the vibration acceleration signal from the beginning to the complete failure of the bearing in the accelerated life test;

[0032] S2: Perform time domain analysis on the vibration signal and extract six time domain features, including root mean square value, variance, peak value, mean value, absolute mean value, and kurtosis;

[0033] S3: Analyze the trend and monotonicity of each time domain feature. The trend is measured by the correlation between each time domain feature and the original vibration signal. The correlation model is The relevance index is defined as Finally, the root mean square value is selected as the sensitive feature;

[0034] S4: Perform multi-resolution singular value decomposition on the feature, use monotonicity to determine the appropriate number of decomposition layers, and construct sensitive features. The monotonicity model is

[0035] S5: Normalize the sensitive features. The normalized model is: After normalizing the sensitive eigenvalue data, its value is between 0 and 1, where x is the eigenvalue to be normalized, and x max 、x minare the maximum and minimum values ​​in the original data, respectively, x norm The data is normalized; the starting prediction point is determined by the μ+3σ criterion, and the training set and prediction set are divided; the linear regression model RMS=wt+b is fitted within the root mean square value range, where The coefficient w represents the slope of the RMS value relative to time t, and the absolute value of w is used to establish the threshold. The threshold is determined by the mean μ and standard deviation σ of the bearing health stage data. The threshold is determined by calculating μ + 3σ. The starting prediction point is determined when the newly calculated w exceeds the threshold for n consecutive times (for example, n = 5). The data before the starting prediction point is divided into the training set, and the data after the starting prediction point is divided into the test set.

[0036] S6: Construct the Bi-LSTM network framework, including defining the model structure, selecting the loss function and the optimizer. Compared with the conventional LSTM network, this network contains two more hidden layers. The hidden layer transmits the input information from back to front to each hidden layer node. All information is finally gathered to the output layer, and there is no connection between the hidden layers. The parameter iteration method of the Bi-LSTM network is basically the same as that of the conventional LSTM, with the output y t For example, the update formulas are shown in (1)-(3):

[0037]

[0038]

[0039]

[0040] where h t , u, w, v, f1 represent the state of the forward hidden layer, the weight of the input, the weight of the hidden layer at time t-1, the weight of the output and the activation function; similarly, h t , u, w, v, f2 represent the state of the backward hidden layer, the weight of the input, the weight of the hidden layer at time t+1, the weight of the output and the activation function respectively; y t , f3 represents the output value and the output layer activation function respectively;

[0041] S7: training the Bi-LSTM network using the training set and verifying the prediction set;

[0042] S8: Denormalize the trained and predicted health indicators to convert them into predicted bearing degradation trends, and construct model evaluation indicators to evaluate the prediction accuracy. Denormalization is the inverse operation of step S5. The model is x_scaled = x_std × (x max -x min )+x min , where x_scaled is the result after the denormalization operation, xmax and x min The training result curve of the training set and the test result curve of the test set are as follows: Figure 3 As shown, the horizontal axis is the number of samples sampled every ten minutes, and the vertical axis is the health index after amplitude change. The model evaluation index is constructed using the root mean square error (RMSE), mean absolute error (MAE), correlation coefficient (R 2 ) and correlation (Con) are used to evaluate the prediction accuracy. Figure 3 The calculation result of the root mean square value selected in step S3 is as follows Figure 4 Compared with directly using the LSTM network, the bearing degradation trend prediction method based on multi-resolution feature extraction and Bi-LSTM network proposed in the present invention has improved prediction accuracy in both training and testing stages.

[0043] In summary, in the bearing degradation trend prediction method based on multi-resolution feature extraction and Bi-LSTM network provided in an embodiment of the present invention, it is proposed to use multi-resolution singular value decomposition to extract features and use Bi-LSTM network to predict bearing degradation trend, which effectively improves the prediction accuracy compared with the prediction results using a single LSTM network.

[0044] Finally, it should be noted that the above examples are merely preferred embodiments of the present invention and do not limit the present invention in any way. Although the present invention has been described in detail with reference to the preferred embodiments, any person skilled in the art should understand that any equivalent substitution or modification of the technical solution and technical content of the present invention does not depart from the technical solution of the present invention and is intended to be encompassed by the scope of the claims of the present invention.

Claims

1. A bearing degradation trend prediction method based on multi-resolution feature extraction and Bi-LSTM network, characterized in that: The following steps are involved: S1: Obtain the vibration acceleration signal from the beginning to complete failure of the bearing accelerated life test; S2: Perform time domain analysis on the vibration signal, extract multi-angle eigenvalues, and preliminarily construct a feature set that comprehensively describes the operating status of the bearing; S3: Analyze the trend and correlation of the feature set and select appropriate features; S4: Input the feature into the multi-resolution singular value decomposition method, use monotonicity to determine the appropriate number of decomposition layers, and construct sensitive features; Perform singular value decomposition on the selected features, select the multi-resolution singular value decomposition result based on the principle of the best monotonicity of the approximate component of a certain layer, stop decomposition, and obtain the approximate component without random noise as the sensitive feature; The monotonicity index is defined as: ； S5: After normalizing the sensitive features to [0,1], the starting prediction point is determined by the μ+3σ criterion, and the training set and prediction set are divided; S6: Build a Bi-LSTM network framework and use the normalized sensitive features as input to obtain health indicators that comprehensively represent the operating status of the bearing. S7: training the Bi-LSTM network using the training set and verifying the prediction set; S8: Denormalize the trained and predicted health indicators and convert them into predicted bearing degradation trends. Build model evaluation indicators to evaluate the accuracy of predicted degradation trends. The trend indicator is defined as: The relevance index is defined as: Finally, select the features with better physical properties.

2. A bearing degradation trend prediction method according to claim 1, characterized in that: In step S2, the rolling bearing suitable features included in the time domain analysis include root mean square value, variance, peak value, mean value, absolute mean value, and kurtosis.

3. The method for predicting bearing degradation trend according to claim 1, characterized in that: In step S5, the normalized processing model is A sliding window with a window width of n is established in the RMS value range, and a linear regression fitting operation RMS=wt+b is performed on the n RMS values ​​in the window. Therefore, each time the window slides forward, the slope w and intercept b are re-determined based on the n RMS values ​​contained in the window, where The coefficient w represents the slope of the root mean square value relative to time t, and the absolute value of w is used to establish the threshold. The threshold is determined by the mean μ and standard deviation σ of the bearing health stage data. The threshold is determined by calculating μ + 3σ. The starting prediction point is determined when the newly calculated w is greater than the threshold for n consecutive times. The data before the starting prediction point is divided into the training set, and the data after the starting prediction point is divided into the test set.

4. The method for predicting bearing degradation trend according to claim 1, characterized in that: In step S7, the Bi-LSTM network includes an input layer, multiple hidden layers, and an output layer; the back propagation algorithm and the gradient descent algorithm are used, the training optimization algorithm is Adam, the loss function is the root mean square error, the output layer activation function is the tanh function, other parameters are set to default values, and training is performed for 300 rounds.

5. The method for predicting bearing degradation trend according to claim 1, characterized in that: In step S8, the denormalized model is x_scaled = x_std × (xmax - xmin) + xmin, and the root mean square error, mean absolute error, correlation coefficient and correlation are used to evaluate the accuracy of the prediction and compare it with the prediction results of the LSTM network.

Images