Remaining service life prediction method of drilling rig spindle rolling bearing based on CVA and RNN
Through the CVA and RNN-based method, typical variable analysis and recurrent neural network construction are used to use acceleration sensor data to solve the problem of accurate prediction of the spindle rolling bearing of the drilling rig in the bearing degradation stage, improve the accuracy of the residual service life prediction, reduce the risk of failure and Maintenance cost.
Patent Information
- Application Number
- CN202111533045.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-12-15
- Publication Date
- 2024-12-31
- Estimated Expiration
- 2041-12-15
AI Technical Summary
The existing method of predicting the remaining service life of the spindle rolling bearing of the drilling rig has shortcomings in the accurate prediction of the bearing degradation stage, and it is difficult to effectively predict the time of bearing failure.
Using CVA and RNN-based methods, the vibration acceleration signals collected by the acceleration sensor are obtained, the frequency domain characteristics are extracted, and the model is constructed using typical variable analysis and recurrent neural networks, and the state transition matrix and the output residual service life prediction are calculated.
Improve the accuracy of bearing residual service life prediction, reduce economic and safety risks caused by failures, and reduce maintenance costs.
Smart Images
Figure CN114254454B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of drilling rig fault prediction and relates to a method for predicting the remaining service life of a drilling rig main shaft rolling bearing based on CVA and RNN. Background Art
[0002] As one of the most basic components in rotating machinery, rolling bearings are widely used in various mechanical equipment. The stability of their operating status has the most direct impact on the safety and production of the equipment. For a drilling rig, if a bearing fails, not only will the drill rod be bent, but the drill and the borehole will not be in a straight line, which seriously affects the stability of the drilling rig. Drilling or drilling explosion may also occur, causing greater economic losses and even personal injury. Therefore, in order to better reduce safety hazards and reduce maintenance costs, it is necessary to predict and manage the faults of drilling rigs, among which the prediction of the remaining useful life of bearings is particularly important. By predicting the remaining useful life of bearings, new bearings can be replaced when the bearings are about to fail, thereby avoiding the adverse effects of failures. The existing remaining useful life prediction methods are divided into three categories: failure mechanism analysis methods, data-driven methods, and fusion methods. At present, the data-driven method based on machine learning is the mainstream method for remaining useful life prediction. Summary of the invention
[0003] Purpose of the invention: The main technical problem solved by the present invention is to accurately predict the remaining service life of a bearing in the degradation stage of the bearing.
[0004] Technical solution: The above problem can be solved by the following technical solution: A method for predicting the remaining service life of a drilling rig spindle rolling bearing based on CVA and RNN, the method comprising the following steps:
[0005] (1) Obtaining data: Using an acceleration sensor radially installed on the outer wall of the main shaft bearing of the drilling rig to collect the vibration acceleration signal of the bearing throughout its life cycle, and dividing it into a training set and a test set;
[0006] (2) Data processing: extracting the frequency domain features of the vibration acceleration data of the training set and the test set, and normalizing the labels of the training set and the test set;
[0007] (3) Feature extraction: for the training set, let the time lag order constant be N, then starting from the Nth sample point, a pair of past data vectors and future data vectors are divided one by one, and the total number of pairs of divided vectors is M, where M=Z-N+1, and Z is the total number of sample points; the above-mentioned past data vector is composed of the frequency domain features of the N sample points before the current sample point after data processing in sequence, and the future data vector is composed of the label values of the current sample point and the N-1 sample points after the current sample point after normalization in sequence; the CVA method is used to extract features from the past data vectors and the future data vectors, and the state estimation matrix of the past data is calculated. The data matrix composed of the past data vectors is multiplied by the state estimation matrix to obtain the typical variable matrix of the past data, that is, the data matrix with the maximum correlation with the corresponding remaining useful life data matrix;
[0008] (4) Construct RNN deep recurrent neural network;
[0009] (5) For the training set, the past data matrix and the future data matrix are both known, while the future data matrix in the test set is unknown. The state transition matrix of CVA is calculated through the training set, and the data matrix after the state transition matrix conversion of the training set is used as input and the normalized remaining useful life label is used as output to train the recurrent neural network. The performance of the CVA-RNN model is verified through the test set.
[0010] Preferably, in step (1), the vibration acceleration signal of the bearing throughout its life cycle is obtained by sampling on an acceleration sensor, wherein the sampling frequency is 2.56 kHz, the sampling interval is 10 s, and each sample point has 2560 values.
[0011] Preferably, the frequency domain feature extraction method in step (2) is: performing a fast Fourier transform on the vibration signal to obtain the frequency domain amplitude feature, and then calculating its mean.
[0012] Preferably, in step (2), the method of normalizing the labels of the training set is as shown in formula (1):
[0013]
[0014] Where n is the number of the sample point, T is the full life cycle of the bearing, and T n The label for the nth sample point at which the bearing data record begins, i.e., the remaining useful life, RUL n is the normalized value of the label of the nth sample point, where T n The calculation formula is as follows:
[0015] T n =10(n-1)(2).
[0016] Preferably, the CVA in step (3) is a canonical variable analysis, which maximizes the correlation between past data and future data. The specific solution steps are:
[0017] Construct the historical vibration acceleration data vector and the future remaining service life data vector respectively:
[0018] Assume that at a certain time t, the historical data vector and the future data vector are as shown in equation (3) and equation (4):
[0019] p(t)=[x T (t-1),…,x T (tN)] T ∈R N×p (3)
[0020] f(t)=[y T (t),y T (t+1)…,y T (t+N)] T ∈R N×f (4)
[0021] Where: N is the time lag order constant, x(t)∈R p is the frequency domain feature of the p-dimensional vibration acceleration signal of the sample point sampled at time t, y(t)∈R f The f-dimensional remaining useful life data vector of the sample point sampled at time t;
[0022] The two data matrices are normalized to zero mean and unit variance along each variable at each moment. The normalization method is shown in equations (5) and (6):
[0023]
[0024]
[0025] Where: is the mean vector of the sample, σ x(t) ,σ y(t) is the sample variance;
[0026] Construct the past output matrix X p and the future output matrix X f ;
[0027] The past output matrix X consists of M past output variables and future output variables p and the future output matrix X f :
[0028]
[0029]
[0030] Where: M is the total number of past or future data vectors divided, is the normalized data vector, where i = t+1, t+2, ..., t+M;
[0031] The goal of CVA is to find a pair of projection directions J,L such that the matrix X p and the future data matrix X f The projections in this direction have the greatest correlation, and the projected data is the typical variable matrix JX p With LX f , by solving the optimization problem:
[0032]
[0033] Obtain the projection matrix. Where: Σ pf Output matrix X for the past p and the future output matrix data matrix X f The cross covariance matrix between them, J and L are projection matrices;
[0034] The method to solve the above optimization problem is: first solve the covariance matrix Σ of the past output matrix pp and the covariance matrix of the future output matrix Σ ff , and the cross-covariance matrix Σ between the past output matrix and the future output matrix data matrix pf , the calculation formula is shown in formula (10), (11), (12):
[0035]
[0036]
[0037]
[0038] Secondly, the matrix Do singular value decomposition:
[0039]
[0040] Where: U is the left singular matrix, Σ is the singular value matrix, and V is the right singular matrix. The projection matrix calculation formula for the past and future data matrices is:
[0041]
[0042]
[0043] Therefore, the state transition matrix of the past data matrix is:
[0044]
[0045] Where: U k are the first k columns of the matrix U, k is the dimension of the data after CVA processing, and the value of k is 1 in this method.
[0046] Preferably, the RNN proposed in step (5) is a recurrent neural network with a many-to-many structure, and the recurrent neural network includes an input layer, two hidden layers, an output layer and a fully connected layer. To prevent overfitting, a dropout of size 0.2 is added to each hidden layer. BRIEF DESCRIPTION OF THE DRAWINGS
[0047] Figure 1 It is a flow chart of the remaining service life prediction method of the drilling rig main shaft rolling bearing based on CVA and RNN according to the present invention;
[0048] Figure 2 is a network structure diagram of a recurrent neural network method for predicting the remaining useful life of a bearing in the method of the present invention;
[0049] Figure 3 It is the vibration acceleration signal diagram and frequency domain average amplitude spectrum diagram of the full cycle of the bearing as training data;
[0050] Figure 4 It is the truncated vibration acceleration signal diagram and frequency domain average amplitude spectrum diagram of the bearing used as the test set;
[0051] Figure 5 It is a comparison diagram of the prediction results of the method described in the present invention and that of using only the RNN network. DETAILED DESCRIPTION
[0052] Combination Figures 1 to 5 The present invention proposes a method for predicting the remaining service life of a drilling rig spindle rolling bearing based on CVA and RNN, and the method comprises the following steps:
[0053] (1) Obtaining data: Using an acceleration sensor radially installed on the outer wall of the main shaft bearing of the drilling rig to collect the vibration acceleration signal of the bearing throughout its life cycle, and dividing it into a training set and a test set;
[0054] (2) Data processing: extracting the frequency domain features of the vibration acceleration data of the training set and the test set, and normalizing the labels of the training set and the test set;
[0055] (3) Feature extraction: for the training set, let the time lag order constant be N, then starting from the Nth sample point, a pair of past data vectors and future data vectors are divided one by one, and the total number of pairs of divided vectors is M, where M=Z-N+1, and Z is the total number of sample points; the above-mentioned past data vector is composed of the frequency domain features of the N sample points before the current sample point after data processing in sequence, and the future data vector is composed of the label values of the current sample point and the N-1 sample points after the current sample point after normalization in sequence; the CVA method is used to extract features from the past data vectors and the future data vectors, and the state estimation matrix of the past data is calculated. The data matrix composed of the past data vectors is multiplied by the state estimation matrix to obtain the typical variable matrix of the past data, that is, the data matrix with the maximum correlation with the corresponding remaining useful life data matrix;
[0056] (4) Construct RNN deep recurrent neural network;
[0057] (5) For the training set, the past data matrix and the future data matrix are both known, while the future data matrix in the test set is unknown. The state transition matrix of CVA is calculated through the training set, and the data matrix after the state transition matrix conversion of the training set is used as input and the normalized remaining useful life label is used as output to train the recurrent neural network. The performance of the CVA-RNN model is verified through the test set.
[0058] In step (1), the vibration acceleration signal of the bearing throughout its life cycle is obtained by sampling on an acceleration sensor, wherein the sampling frequency is 2.56 kHz, the sampling interval is 10 s, and each sample point has 2560 values.
[0059] The frequency domain feature extraction method described in step (2) is: perform fast Fourier transform on the vibration signal to obtain the frequency domain amplitude feature, and then calculate its mean.
[0060] In step (2), the method of normalizing the labels of the training set is shown in formula (1):
[0061]
[0062] Where n is the number of the sample point, T is the full life cycle of the bearing, and T n The label for the nth sample point at which the bearing data record begins, i.e., the remaining useful life, RUL n is the normalized value of the label of the nth sample point, where T n The calculation formula is as follows:
[0063] T n =10(n-1) (2).
[0064] The CVA in step (3) is a typical variable analysis. The method of typical variable analysis maximizes the correlation between past data and future data. The specific solution steps are:
[0065] Construct the historical vibration acceleration data vector and the future remaining service life data vector respectively:
[0066] Assume that at a certain time t, the historical data vector and the future data vector are as shown in equation (3) and equation (4):
[0067] p(t)=[x T (t-1),…,x T (tN)] T ∈R N×p (3)
[0068] f(t)=[y T (t),y T (t+1)…,y T (t+N)] T ∈R N×f (4)
[0069] Where: N is the time lag order constant, x(t)∈R p is the frequency domain feature of the p-dimensional vibration acceleration signal of the sample point sampled at time t, y(t)∈R f is the f-dimensional remaining useful life data vector of the sample point sampled at time t; the two data matrices are normalized to zero mean and unit variance along each variable at each time. The normalization method is shown in equations (5) and (6):
[0070]
[0071]
[0072] Where: is the mean vector of the sample, σ x(t) ,σ y(t) is the sample variance;
[0073] Construct the past output matrix X p and the future output matrix X f ;
[0074] The past output matrix X consists of M past output variables and future output variables p and the future output matrix X f :
[0075]
[0076]
[0077] Where: M is the total number of past or future data vectors divided, is the normalized data vector, where i = t+1, t+2, ..., t+M;
[0078] The goal of CVA is to find a pair of projection directions J,L such that the matrix X p and the future data matrix X f The projections in this direction have the greatest correlation, and the projected data is the typical variable matrix JX p With LX f , by solving the optimization problem:
[0079]
[0080] Obtain the projection matrix. Where: Σ pf Output matrix X for the past p and the future output matrix data matrix X f The cross covariance matrix between them, J and L are projection matrices;
[0081] The method to solve the above optimization problem is: first solve the covariance matrix Σ of the past output matrix pp and the covariance matrix of the future output matrix Σ ff , and the cross-covariance matrix Σ between the past output matrix and the future output matrix data matrix pf , the calculation formula is shown in formula (10), (11), (12):
[0082]
[0083]
[0084]
[0085] Secondly, the matrix Do singular value decomposition:
[0086]
[0087] Where: U is the left singular matrix, Σ is the singular value matrix, and V is the right singular matrix.
[0088] The projection matrix calculation formula of the past and future data matrices is:
[0089]
[0090]
[0091] Therefore, the state transition matrix of the past data matrix is:
[0092]
[0093] Where: U k are the first k columns of the matrix U, k is the dimension of the data after CVA processing, and the value of k is 1 in this method.
[0094] The RNN proposed in step (5) is a recurrent neural network with a many-to-many structure, which includes an input layer, two hidden layers, an output layer and a fully connected layer. To prevent overfitting, a dropout of size 0.2 is added to each hidden layer.
[0095] Figure 2 To build a deep recurrent neural network: a two-layer RNN.
[0096] Figure 2 The figure shows a schematic diagram of the structure of a double-layer RNN. The structure of the RNN is a recurrent neural network with a "many-to-many" structure. The recurrent neural network includes an input layer, two hidden layers, an output layer, and a fully connected layer. To prevent overfitting, a dropout of size 0.2 is added to each hidden layer. The RNN deep learning network is trained by taking the data matrix of the training set after the state transition matrix conversion as input and the normalized remaining useful life label as output; the performance of the CVA-RNN model is verified by the data matrix of the test set after the state transition matrix conversion. The original data of the training set and the test set are shown in Figure 1. Figure 3 (a), 3(b) and 4(a), 4(b), the data after CVA processing is as follows Figure 3 (c) and 4(c). The initial vibration data is 2-dimensional, which are the vibration signals in the vertical direction and the vibration signals in the horizontal direction. After CVA processing, the dimension is reduced to 1-dimensional.
[0097] The results predicted by the RNN model are as follows Figure 5 As shown, the prediction result after CVA processing and then RNN network is Figure 5 (a) The result predicted by RNN network directly without CVA processing is Figure 5 (b); where the horizontal axis is the number of samples and the vertical axis represents the ratio of the remaining useful life to the total useful life.
[0098] Since we only need to pay attention to the degradation stage of bearing life, we predict the data sets after the time point when the training set and the test set start to degrade, and compare the prediction results. The mean square error (MSE) output result of Figure (a) is: 0.004391. The mean square error (MSE) output result of Figure (b) is: 0.007907. As can be seen from Figure (a), the variance is large at the beginning of the degradation stage and the end of the life. The main reason is that: at the beginning, the data is stable and has a strong correlation under normal working conditions, and the prediction results are closer. At the moment of transition from normal working conditions to the beginning of degradation, the prediction results tend to be a horizontal line, which is more inclined to the normal working state; at the end of the life, there is no future life correlation with it, and the data of the last N samples cannot be processed by CVA, so the prediction results have very large fluctuations.
[0099] In summary, the remaining useful life prediction method of rolling bearings based on CVA and RNN provided in the embodiment of the present invention effectively improves the accuracy of remaining useful life prediction compared with traditional recurrent neural networks.
Claims
1. A method for predicting the remaining service life of a drilling rig spindle rolling bearing based on CVA and RNN, characterized in that: The method comprises the following steps: (1) Obtaining data: Using an acceleration sensor radially installed on the outer wall of the main shaft bearing of the drilling rig to collect the vibration acceleration signal of the bearing throughout its life cycle, and dividing it into a training set and a test set; (2) Data processing: extracting the frequency domain features of the vibration acceleration data of the training set and the test set, and normalizing the labels of the training set and the test set; (3) Feature extraction: for the training set, let the time lag order constant be N, then starting from the Nth sample point, a pair of past data vectors and future data vectors are divided one by one, and the total number of pairs of divided vectors is M, where M=Z-N+1, and Z is the total number of sample points; the above-mentioned past data vector is composed of the frequency domain features of the N sample points before the current sample point after data processing in sequence, and the future data vector is composed of the label values of the current sample point and the N-1 sample points after the current sample point after normalization in sequence; the CVA method is used to extract features from the past data vectors and the future data vectors, and the state estimation matrix of the past data is calculated. The data matrix composed of the past data vectors is multiplied by the state estimation matrix to obtain the typical variable matrix of the past data, that is, the data matrix with the maximum correlation with the corresponding remaining useful life data matrix; (4) Construct RNN deep recurrent neural network; (5) The past data matrix and future data matrix in the training set are known, while the future data matrix in the test set is unknown. The state transition matrix of CVA is calculated through the training set. The data matrix after the state transition matrix conversion of the training set is used as input, and the normalized remaining service life label is used as output. The recurrent neural network is trained, and the performance of the CVA-RNN model is verified through the test set. The CVA in step (3) is a typical variable analysis. The method of typical variable analysis maximizes the correlation between past data and future data. The specific solution steps are as follows: (1) Construct the historical vibration acceleration data vector and the future remaining service life data vector respectively: Assume that at a certain time t, the historical data vector and the future data vector are as shown in equations (3) and (4): p(t)=[x T (t-1),…,x T (t-N)] T ∈R N×p (3) f(t)=[y T (t),y T (t+1)…,y T (t+N)] T ∈R N×f (4) Where: N is the time lag order constant, x(t)∈R p is the frequency domain feature of the p-dimensional vibration acceleration signal of the sample point sampled at time t, y(t)∈R f The f-dimensional remaining useful life data vector of the sample point sampled at time t; (2) The two data matrices are normalized to zero mean and unit variance along each variable at each moment. The normalization method is shown in equations (5) and (6): Where: is the mean vector of the sample, σ x(t) ,σ y(t) is the sample variance; (3) Construct the past output matrix X p and the future output matrix X f ; The past output matrix X consists of M past output variables and future output variables p and the future output matrix X f : Where M is the total number of past or future data vectors divided, is the normalized data vector, where i=t+1,t+2,...,t+M; (4) The goal of CVA is to find a pair of projection matrices J and L such that the past output matrix X p and the future output matrix X f The projections in this direction have the greatest correlation, and the projected data is the typical variable matrix JX p With LX f , the projection matrix is obtained by solving the following optimization problem: Where: Σ pf Output matrix X for the past p and the future output matrix X f The cross covariance matrix between them, J and L are projection matrices; (5) The method to solve the above optimization problem is: first solve the covariance matrix Σ of the past output matrix pp and the covariance matrix of the future output matrix Σ ff , and the cross-covariance matrix Σ between the past output matrix and the future output matrix pf , the calculation formula is shown in formula (10), (11), (12): Pair Matrix Do singular value decomposition: Where U is the left singular matrix, Σ is the singular value matrix, and V is the right singular matrix; The projection matrix calculation formula of the past and future data matrices is: The state transition matrix of the past data matrix is: Where: U k are the first k columns of the matrix U, k is the dimension of the data after CVA processing, and the value of k is 1.
2. The method for predicting the remaining service life of a drilling rig spindle rolling bearing based on CVA and RNN according to claim 1 is characterized in that: In step (1), the vibration acceleration signal of the bearing throughout its life cycle is obtained by sampling on an acceleration sensor, wherein the sampling frequency is 2.56 kHz, the sampling interval is 10 s, and each sample point has 2560 values.
3. The method for predicting the remaining service life of a drilling rig spindle rolling bearing based on CVA and RNN according to claim 1 is characterized in that: The frequency domain feature extraction method described in step (2) is: perform fast Fourier transform on the vibration signal to obtain the frequency domain amplitude feature, and then calculate its mean.
4. The method for predicting the remaining service life of a drilling rig spindle rolling bearing based on CVA and RNN according to claim 1 is characterized in that: In step (2), the method of normalizing the labels of the training set is shown in formula (1): Where n is the number of the sample point, T is the full life cycle of the bearing, and T n The label for the nth sample point at which the bearing data record begins, i.e., the remaining useful life, RUL n is the normalized value of the label of the nth sample point, where T n The calculation formula is as follows: T n =10(n-1)(2)。 5. The method for predicting the remaining service life of a drilling rig spindle rolling bearing based on CVA and RNN according to claim 1 is characterized in that: The RNN proposed in step (5) is a recurrent neural network with a many-to-many structure, which includes an input layer, two hidden layers, an output layer and a fully connected layer, and a dropout of size 0.2 is added to each hidden layer.