Method for identifying performance degradation state of rolling linear guide rail pair under variable working conditions

Abstract

The invention discloses a rolling linear guide pair performance degradation state identification method under variable working conditions, and relates to the technical field of rolling linear guide pair performance state monitoring. Vibration signals at different operation speeds and different positions in the whole-life periodic performance degradation process of the rolling linear guide rail pair are collected; the vibration signals are preprocessed by using a combined denoising method based on CEEMDAN and a wavelet adaptive threshold value; extracting time domain, frequency domain and multi-scale fuzzy entropy features from the denoised vibration signals to construct a feature set; constructing a field adaptive degradation state recognition model based on a one-dimensional convolutional neural network; and carrying out degradation state identification tests of multiple groups of variable working conditions, randomly selecting data of at least two groups of working conditions as labeled source domain training data, taking data of other working conditions as unlabeled target domain test data, and identifying the performance degradation state of the rolling linear guide rail pair by utilizing a trained model. The method is suitable for linear guide pair performance degradation state recognition under various working conditions.

Description

Technical Field

[0001] The present invention relates to the technical field of performance status monitoring of rolling linear guide pairs, and more particularly to a method for identifying the performance degradation status of rolling linear guide pairs under variable working conditions. Background Art

[0002] As a core functional component of the transmission system of high-end CNC machine tools, linear rolling guides employ interference fit between the balls and raceways to ensure optimal performance. This contact and compression between the balls and raceways generates a preload, eliminating clearance between the two. As the CNC machine tool continues to operate, friction and wear between the balls and raceways of the linear rolling guide lead to a degradation of the preload. This degradation in preload leads to a degradation in the guide's accuracy retention and other performance characteristics, further reducing the machine's feed and machining accuracy. Therefore, accurately identifying the degradation of linear rolling guide performance is crucial to improving the reliability of CNC machine tools.

[0003] The linear guides in CNC machine tools are not easy to disassemble directly to measure their current performance indicators, and traditional methods are unable to monitor the performance degradation of the guides. Currently, obtaining the operating status of mechanical equipment through vibration signals is the most commonly used state detection method. Research on state monitoring and fault diagnosis based on vibration signals focuses on bearings and gears, and research on the performance degradation state identification of linear guides is relatively scarce. The complex service environment of rolling linear guides requires consideration of factors such as noise interference and changes in working conditions that guides face in industrial scenarios, further increasing the difficulty of identifying the performance degradation state of linear guides. It is necessary to design a linear guide performance degradation state identification method with high accuracy, strong generalization, and high robustness.

[0004] Therefore, it is an urgent problem for those skilled in the art to propose a method for identifying the performance degradation state of a rolling linear guide pair under variable working conditions to solve the difficulties existing in the prior art. Summary of the Invention

[0005] In view of this, the present invention provides a method for identifying the performance degradation state of a rolling linear guide pair under variable working conditions, which is suitable for identifying the performance degradation state of a linear guide pair under various working conditions and provides a new solution to the technical field of rolling linear guide pair state monitoring.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] A method for identifying the performance degradation state of a rolling linear guide pair under variable working conditions comprises the following steps:

[0008] Collect vibration signals at different operating speeds and positions during the performance degradation of the rolling linear guide pair throughout its life cycle;

[0009] The vibration signal is preprocessed using a joint denoising method based on CEEMDAN and wavelet adaptive threshold;

[0010] The time domain, frequency domain and multi-scale fuzzy entropy features are extracted from the denoised vibration signal to construct a feature set;

[0011] Construct a domain-adaptive degradation state recognition model based on one-dimensional convolutional neural network;

[0012] Conduct multiple sets of degradation state identification tests under varying working conditions. Randomly select at least two sets of working condition data as labeled source domain training data, and the remaining working condition data as unlabeled target domain test data. Use the trained model to identify the performance degradation state of the rolling linear guide pair.

[0013] Optionally, vibration signals at different operating speeds and positions during the performance degradation process of the rolling linear guide pair throughout its life cycle are collected. Specifically, the rolling linear guide pair is installed on a loaded running-in test bench, and the running-in mileage is set. After a single running-in is completed, the guide rail to be tested is removed from the loaded running-in test bench, and a three-axis vibration sensor is attached to the top surface of the slider and a single-axis vibration sensor is attached to one end of the guide rail. The vibration signal acquisition system collects vibration signals of four round trips at the specified operating speed and sampling frequency.

[0014] Optionally, the vibration signal is preprocessed using a joint denoising method based on CEEMDAN and wavelet adaptive threshold, which specifically includes the following steps: decomposing the original vibration signal into different intrinsic mode components through an experience-based signal decomposition method, screening out the intrinsic mode components containing degradation information from the decomposed intrinsic mode components, and eliminating irrelevant false components; using wavelet adaptive threshold decomposition to suppress the residual noise components in the selected intrinsic mode components, and adding the processed intrinsic mode components to obtain a denoised reconstructed signal.

[0015] Optional, the specific steps of CEEMDAN are as follows:

[0016] The original vibration signal x(t) is repeatedly superimposed with i groups of Gaussian white noise m with specific amplitudes. i , generate the signal to be decomposed x i (t):

[0017] x i (t) = x(t) + m i ;

[0018] Treat the decomposed signal x i (t) Perform EMD decomposition, that is, identify the signal to be decomposed x i All the maximum and minimum points of (t) are fitted with the upper and lower envelopes respectively using cubic spline interpolation to construct the upper and lower envelopes e imax(t) and e imin (t), calculate the mean envelope m i (t):

[0019] m i (t)=[e imax (t)+e imin (t)] / 2;

[0020] Use h i (t) = x i (t)-m i (t), extract the candidate intrinsic mode components, judge whether the number of its extreme points and zero-crossing points is equal or differs by 1 and whether the mean envelope is zero. If the eigenmode component conditions are met, the first eigenmode component IMF1 is obtained; if not, let x i (t) = h i (t), repeat the above steps until the eigenmode component conditions are met;

[0021] The method for selecting the eigenmode components that are sensitive to performance degradation includes calculating the correlation coefficient between the eigenmode components and the original signal, which is expressed as follows:

[0022]

[0023] in, and are the means of the original signal and the eigenmode components, respectively. The threshold of the correlation coefficient is set, and the eigenmode components with correlation coefficients higher than the specified threshold are selected;

[0024] For the selected intrinsic mode components, wavelet threshold is used to reduce the noise components in the selected intrinsic mode components; for a set of noisy intrinsic mode components x(t), the wavelet threshold denoising process is as follows:

[0025] Determine the appropriate wavelet basis function and decomposition layer number, perform N-layer wavelet decomposition on x(t), and obtain a series of wavelet coefficients ωj ,k ;

[0026] Determine the appropriate threshold T, quantize the detail coefficients of the first layer N layer through the threshold function, and obtain the estimated value of the detail coefficient ω j,k ;

[0027] Perform inverse wavelet transform on the approximate coefficients of the Nth layer and the detail coefficients after quantization to obtain the denoised signal x(t);

[0028] The denoised intrinsic mode components are added together to obtain the denoised reconstructed signal.

[0029] Optionally, extracting time domain, frequency domain, and multi-scale fuzzy entropy features from the denoised vibration signal to construct a feature set specifically includes the following steps:

[0030] A model for identifying the performance degradation of rolling linear guideways under varying operating conditions uses a one-dimensional convolutional neural network as a feature extraction and classification model. A domain adaptation layer is introduced to minimize the distribution differences of domain features in this layer, enabling identification of guideway performance degradation under varying operating conditions.

[0031] The grid structure includes an input layer, a feature extraction layer, a domain adaptation layer, a fully connected layer, and an output layer. The loss function includes source domain classification loss and domain adaptation loss.

[0032] The feature extraction layer includes single-scale convolution layer, multi-scale convolution layer and other feature extraction modules;

[0033] The classification layer includes the Softmax function;

[0034] Gradient descent calculation optimizers include Adam optimizer;

[0035] The domain adaptation layer aims to achieve domain distribution adaptation on the fully connected layer, including calculating the MMD distance of inter-domain features on the adaptation layer and adding MMD to the loss function;

[0036] The loss function includes source domain classification loss and domain adaptation loss. The optimization goal is to minimize the loss after the two. The source domain classification loss uses the cross entropy loss function. The total loss expression is as follows:

[0037] L=L s +δ m L m ;

[0038] Among them, L s , L m are the cross entropy loss and the MMD distance of the domain adaptation layer, δ m >0 is L m The penalty coefficient of

[0039] During network training, the parameter update expression of each network layer is expressed as:

[0040]

[0041] Among them, α is the network learning rate, ω l and b l are the weights and bias parameters of the network layers during training.

[0042] Optionally, multiple sets of degradation state recognition tests with varying working conditions are conducted. Data from at least two working conditions are randomly selected as labeled source domain training data, and data from the remaining working conditions are used as unlabeled target domain test data. The trained model is used to identify the performance degradation state of the rolling linear guide pair. Specifically, the following steps are included:

[0043] Select at least two sets of working condition data as labeled source domain training data, and the remaining working condition data as unlabeled target domain test data;

[0044] Initialize the network parameters and set the initial learning rate of the degradation state recognition network. The learning rate decay method includes decaying with the training process. The expression is:

[0045] δ=αδ0;

[0046] Among them, δ0 is the initial learning rate, α is the attenuation coefficient corresponding to different attenuation methods;

[0047] The source domain data and target domain data are fed into the degradation state recognition network. The training labels of the source domain data and the feature distributions of the source and target domains on the fully connected layer are obtained through forward propagation. The cross entropy loss and the MMD distance on the domain adaptation layer are calculated and combined to form the total loss function.

[0048] Calculate the propagation gradient of network weights and biases based on the total loss function, use error backpropagation and optimization algorithm to train model parameters according to the maximum number of iterations to complete network training;

[0049] The remaining unlabeled target domain test data sets are respectively fed into the trained degradation state recognition network to obtain prediction results, thereby realizing performance degradation identification under the target working conditions.

[0050] It can be seen from the above technical solution that, compared with the prior art, the present invention provides a method for identifying the performance degradation state of a rolling linear guide pair under variable working conditions, which has the following beneficial effects:

[0051] 1. The present invention adopts CEEMDAN and wavelet adaptive threshold combined denoising to decompose and reconstruct the original vibration signal, selects the effective IMF component for wavelet adaptive threshold denoising, and the reconstructed signal suppresses the original noise component while retaining important information characterizing the degradation state.

[0052] 2. The trend of fuzzy entropy change with scale in the present invention can characterize the performance degradation trend of the guide pair. The domain adaptive network model based on one-dimensional convolutional neural network (CNN) can accurately identify the performance degradation state of the rolling linear guide pair under variable working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0054] Figure 1 A flow chart of a method for identifying the performance degradation state of a rolling linear guide pair under variable working conditions provided by the present invention;

[0055] Figure 2 Flowchart of the CEEMDAN and wavelet adaptive threshold combined denoising method provided by the present invention;

[0056] Figure 3a Schematic diagram showing the comparison of separability between state 2 and state 1 of the performance degradation of the original signal of the rolling linear guide pair before and after noise reduction of the original signal of the rolling linear guide pair according to the present invention;

[0057] Figure 3b Schematic diagram comparing the separability of the rolling linear guide pair before and after noise reduction of the original signal of the present invention in state 3 and state 1;

[0058] Figure 3c Schematic diagram showing the comparison of separability between state 4 and state 1 of the performance degradation of the original signal of the rolling linear guide pair before and after noise reduction of the present invention;

[0059] Figure 4 A schematic diagram of the domain-adaptive network structure based on a one-dimensional convolutional neural network (CNN) provided by the present invention;

[0060] Figure 5 A flowchart for identifying the performance degradation state of a rolling linear guide pair using a domain adaptive model based on a one-dimensional convolutional neural network (CNN) provided by the present invention;

[0061] Figure 6a This is a schematic diagram of the confusion matrix for identifying the degradation state of a rolling linear guide pair for migration task 1 of the present invention;

[0062] Figure 6b This is a schematic diagram of the confusion matrix for identifying the degradation state of a rolling linear guide pair for migration task 2 of the present invention;

[0063] Figure 6c This is a schematic diagram of the confusion matrix for identifying the degradation state of the rolling linear guide pair for migration task 3 of the present invention. DETAILED DESCRIPTION

[0064] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0065] Example 1

[0066] An embodiment of the present invention discloses a method for identifying the performance degradation state of a rolling linear guide pair under variable working conditions, comprising the following steps:

[0067] S1, collect the vibration signals at different running speeds and different positions during the performance degradation of the rolling linear guide pair throughout its life cycle. The running speeds are set to v1, v2, ..., v n , the corresponding working conditions are working condition 1, working condition 2, ..., working condition n;

[0068] S2, using the joint denoising method based on CEEMDAN and wavelet adaptive threshold to preprocess the original vibration signal to reduce noise interference;

[0069] S3, extracting time domain, frequency domain and multi-scale fuzzy entropy features from the denoised vibration signal to construct a feature set;

[0070] S4. Construct a domain-adaptive degradation state recognition model based on a one-dimensional convolutional neural network (CNN);

[0071] S5. Conduct degradation state identification tests for multiple groups of variable working conditions. Randomly select data from at least two groups of working conditions as labeled source domain training data, and the data from the remaining working conditions as unlabeled target domain test data. Use the trained model to identify the performance degradation state of the rolling linear guide pair.

[0072] Furthermore, S1 specifically includes:

[0073] The rolling linear guide pair is installed on the loading running-in test bench, and the running-in mileage is set. After a single running-in is completed, the guide rail to be tested is removed from the loading running-in test bench. The three-axis vibration sensor is attached to the top surface of the slider and the single-axis vibration sensor is attached to one end of the guide rail. The vibration signal acquisition system is used to collect four round-trip vibration signals at the specified running speed and sampling frequency.

[0074] Furthermore, S2 specifically includes:

[0075] The original vibration signal x(t) is repeatedly superimposed with i (i = 1, 2, ..., m) groups of Gaussian white noise m with specific amplitudes. i , generate the signal to be decomposed x i (t):

[0076] x i (t) = x(t) + m i ;

[0077] Treat the decomposed signal x i (t) Perform EMD decomposition, that is, identify the signal to be decomposed x i All the maximum and minimum points of (t) are fitted with the upper envelope (connecting the maximum points) and the lower envelope (connecting the minimum points) using cubic spline interpolation to construct the upper and lower envelopes e imax (t) and e imin (t), and then calculate the mean envelope m i (t):

[0078] m i (t)=[e imax (t)+e imin (t)] / 2;

[0079] Use h i (t) = x i (t)-m i (t), extract the candidate intrinsic mode component (IMF), and judge whether the number of its extreme points and zero-crossing points is equal or differs by 1 and whether the mean envelope is zero. If the IMF condition is met, the first intrinsic mode component IMF1 is obtained; if not, let x i (t) = h i (t), repeat the above steps until the IMF conditions are met.

[0080] After the original signal is decomposed by CEEMDAN, a series of IMFs containing information from different frequency bands are obtained. The correlation coefficient is used to select the IMF components that are sensitive to performance degradation information and remove irrelevant false components. The correlation coefficient is expressed as follows:

[0081]

[0082] in, and are the means of the original signal and the intrinsic mode component respectively, the threshold of the correlation coefficient is set to 0.3, and the IMF components with a correlation coefficient greater than 0.3 are selected.

[0083] For the selected IMF components containing important information, wavelet thresholding (WT) is used to reduce the noise components in the selected IMF. For a set of noisy IMF components x(t), the wavelet threshold denoising process is as follows:

[0084] Determine the appropriate wavelet basis function and decomposition layer number, perform N-layer wavelet decomposition on x(t), and obtain a series of wavelet coefficients ω j,k ;

[0085] Determine the appropriate threshold T, quantize the detail coefficients of the first layer N layer through the threshold function, and obtain the estimated value of the detail coefficient ω j,k ;

[0086] Perform inverse wavelet transform on the approximate coefficients of the Nth layer and the detail coefficients after quantization to obtain the denoised signal x(t).

[0087] The denoised IMF components are added together to obtain the denoised reconstructed signal.

[0088] For further information, see Figure 3a 、 Figure 3b 、 Figure 3c , S3 specifically includes:

[0089] The denoised vibration signal is sampled non-overlappingly, and the length of each small sample is 1024. 14 time domain statistical features are extracted from the expanded sample, including maximum value, minimum value, mean, absolute average amplitude, peak value, variance, standard deviation, kurtosis, skewness, root mean square value, form factor, peak factor, pulse factor and margin factor; 12 frequency domain statistical features include mean, variance, third-order central moment of variance, fourth-order central moment of variance, total mean, standard deviation, C index, D index, E index, G index, third-order central moment and fourth-order central moment.

[0090] In addition, multi-scale fuzzy entropy (MFE) is selected to characterize the degradation state of the linear guide pair. The multi-scale fuzzy entropy is used for the original vibration signal {x1,x2,x3,......,x N}, introduce the scaling factor τ = 1, 2, 3, ..., n, perform coarse-graining on the original signal, observe the trend change of the fuzzy entropy of the signal at different time scales, and then characterize the degradation state of the linear guide pair. The new coarse-grained vector is expressed as follows:

[0091]

[0092] Furthermore, S4 specifically includes:

[0093] A one-dimensional convolutional neural network is used as the feature extraction and classification model, and a domain adaptation layer is introduced into it. By minimizing the distribution difference of domain features in this layer, the performance degradation state of the guide rail pair under different working conditions can be identified.

[0094] The network structure consists of an input layer, two convolutional layers, two batch normalization (BN) layers, a fully connected layer, and an output layer. The two convolutional layers are configured with 32 5×1 kernels. Relu is used as the activation function for the convolutional layers. Finally, Softmax is selected as the classification layer with 4 output nodes. The Adam optimizer is used for gradient descent. Domain distribution adaptation is implemented on the fully connected layers by calculating the MMD distance between domain features on the adaptation layer and incorporating the MMD into the loss function. The loss function includes the source domain classification loss and the domain adaptation loss. The optimization goal is to minimize the difference between the two. The source domain classification loss uses the cross-entropy loss function. The total loss is expressed as follows:

[0095] L=L s +δ m L m ;

[0096] Among them, L s 、L m are the cross entropy loss and the MMD distance of the domain adaptation layer, δ m >0 is L m The penalty coefficient of .

[0097] During network training, the parameter update expression of each network layer is expressed as:

[0098]

[0099] Furthermore, S5 specifically includes the following steps:

[0100] S51, select the data of working conditions 1 and 5 as source domain training data, and the data of working conditions 2, 3 and 4 as unlabeled target domain test data;

[0101] S52. Initialize the network parameters and set the initial learning rate of the deep migration network to 0.001. The learning rate decays as the training process progresses. The expression is:

[0102]

[0103] Among them, δ0 is the initial learning rate, m is the current number of iterations, p is the number of training rounds, which is set to 30, l is the number of samples of source domain data, and batchsize is the batch size of training data, which is set to 100;

[0104] S53, the source domain data and the target domain data are transmitted to the recognition network model, and the training labels of the source domain data and the feature distributions of the source domain and the target domain on the fully connected layer are obtained through forward propagation. The cross entropy loss and the MMD distance on the domain adaptation layer are calculated, and the two are combined to form the total loss function;

[0105] S54. Calculate the propagation gradient of the network weights and biases according to the total loss function, use the error back propagation and Adam optimization algorithm to train the model parameters according to the maximum number of iterations, and complete the network training;

[0106] S55. The three groups of target domain test data sets are respectively transmitted to the trained recognition network model to obtain prediction results, thereby realizing the recognition of performance degradation status under target working conditions.

[0107] Example 2

[0108] Reference Figure 1 As shown, the present invention discloses a method for identifying the performance degradation state of a rolling linear guide pair under variable working conditions, comprising the following steps:

[0109] S1. Collect vibration signals at different operating speeds and positions during the performance degradation of the rolling linear guide pair throughout its life cycle;

[0110] S2, using the joint denoising method based on CEEMDAN and wavelet adaptive threshold to preprocess the original vibration signal to reduce noise interference;

[0111] S3, extracting time domain, frequency domain and multi-scale fuzzy entropy features from the denoised vibration signal to construct a feature set;

[0112] S4. Construct a domain-adaptive degradation state recognition model based on a one-dimensional convolutional neural network (CNN);

[0113] S5. Carry out three sets of degradation state identification tests under variable working conditions. Select the data of working condition 1 (7m / min) and working condition 5 (21 / min) as the source domain training data, and the data of working condition 2 (10.5m / min), working condition 3 (14m / min) and working condition 4 (17.5m / min) as the unlabeled target domain test data. Use the trained model to identify the performance degradation state of the rolling linear guide pair.

[0114] Furthermore, S1 obtains the vibration signal data of the rolling linear guide pair, installs the rolling linear guide pair on the loading running-in test bench, sets the running-in mileage, and after a single running-in, removes the guide rail to be tested from the loading running-in test bench. The three-axis vibration sensor is attached to the top surface of the slider, and the single-axis vibration sensor is attached to one end of the guide rail. The sampling length is set to 60% of the guide rail length, the sampling frequency is 5000HZ, and the slider running speed is set to 7m / min, 10.5m / min, 14m / min, 17.5m / min and 21m / min. The vibration signal acquisition system collects four round trip vibration signals.

[0115] Furthermore, in S2, the collected vibration signal of the rolling linear guide pair is preprocessed to suppress the influence of noise in the signal.

[0116] Specifically, the complete ensemble empirical mode decomposition based on adaptive noise (CEEMDAN) and wavelet adaptive threshold joint denoising method are used to optimize the quality of the original vibration signal and suppress the influence of noise. The denoising method flow chart is as follows: Figure 2 As shown, the specific steps include:

[0117] Add i (i=1,2,......,m) groups of Gaussian white noise m to the original signal x(t) i , get the signal x to be decomposed i (t):

[0118] x i (t) = x(t) + m i ;

[0119] Treat the decomposed signal x i (t) Perform EMD decomposition, that is, identify the signal to be decomposed x i All the maximum and minimum points of (t) are fitted with the upper envelope (connecting the maximum points) and the lower envelope (connecting the minimum points) using cubic spline interpolation to construct the upper and lower envelopes e imax (t) and e imin (t), and then calculate the mean envelope m i (t):

[0120] m i (t)=[e imax (t)+e imin (t)] / 2;

[0121] Use h i (t) = x i (t)-m i (t), extract the candidate intrinsic mode component (IMF), and judge whether the number of its extreme points and zero-crossing points is equal or differs by 1 and whether the mean envelope is zero. If the IMF condition is met, the first intrinsic mode component IMF1 is obtained; if not, let x i (t) = h i (t), repeat the above steps until the IMF conditions are met.

[0122] Calculate the average value of the first IMF component to obtain the first-order modal component of CEEMDAN decomposition

[0123]

[0124] Calculate the first-order modal component residual r1(t):

[0125]

[0126] Add white noise to the first-order modal component residual r1(t) to form a new signal to be decomposed r1(t)+ε1E1(m i ), calculate the second modal component

[0127]

[0128] The same steps as the first-order modal component residual are used to calculate the second-order modal component residual r2(t):

[0129]

[0130] Repeat the above steps until the number of extreme points of the residual signal is less than 3, then stop decomposition and wait until the K-order modal component and the residual r k (t), the decomposition result is expressed as:

[0131]

[0132] After the original signal is decomposed by CEEMDAN, a series of IMFs containing information from different frequency bands are obtained. The correlation coefficient is used to select the IMF components that are sensitive to performance degradation information and remove irrelevant false components. The correlation coefficient is expressed as follows:

[0133]

[0134] in, and are the means of the original signal and the intrinsic mode component respectively, the threshold of the correlation coefficient is set to 0.3, and the IMF components with a correlation coefficient greater than 0.3 are selected.

[0135] For the selected IMF components containing important information, wavelet thresholding (WT) is used to reduce the noise components in the selected IMF. For a set of noisy IMF components x(t), the wavelet threshold denoising process is as follows:

[0136] Determine the appropriate wavelet basis function and decomposition layer number, perform N-layer wavelet decomposition on x(t), and obtain a series of wavelet coefficients ω j,k ;

[0137] Determine the appropriate threshold T, quantize the detail coefficients of the first layer N layer through the threshold function, and obtain the estimated value of the detail coefficient ω j,k ;

[0138] Perform inverse wavelet transform on the approximate coefficients of the Nth layer and the detail coefficients after quantization to obtain the denoised signal x(t).

[0139] This method uses DB4 wavelet to perform three-layer wavelet decomposition on the noisy IMF component, and uses hard threshold or soft threshold function to process the detail coefficients obtained by decomposition. The judgment criteria are as follows:

[0140]

[0141] Among them, sign() is the sign function, and T is the set threshold.

[0142] Regarding the choice of threshold function, considering that |ω j,k |≥T k There is a deviation between the wavelet coefficients processed by soft thresholding and the original wavelet coefficients, which leads to a reconstruction error between the denoised IMF components and the original IMF components. Therefore, this method uses a hard threshold function to process the wavelet detail coefficients.

[0143] The following is the process of determining the threshold T. VisuShrink is a fixed universal threshold method, expressed as follows:

[0144]

[0145] Among them, λ k ω j,k The empirical coefficient, N is the length of the wavelet coefficient, σ k is the variance of the estimated noise, which is expressed as:

[0146]

[0147] Threshold selection and empirical coefficient λ k Closely related, coefficient λ k As the noisy signal changes, this method introduces a dimensionless numerical kurtosis to describe the peak sharpness in the signal, which is expressed as:

[0148]

[0149] This method takes the empirical coefficient λ k is the inverse of the IMF component kurtosis, and the threshold expression is rewritten as:

[0150]

[0151] The signal containing more performance degradation information has a higher kurtosis value. A lower threshold is used to denoise the IMF with high kurtosis, while a higher threshold is used to denoise the IMF with low kurtosis. The denoised IMF components are summed to obtain the reconstructed denoised signal.

[0152] Furthermore, in S3, the sample set of the denoised signal is expanded, and then time domain, frequency domain and fuzzy entropy analysis are performed to extract time domain, frequency domain and multi-scale fuzzy entropy features, and construct a feature set for characterizing the performance degradation state of the rolling linear guide pair.

[0153] Specifically, the vibration signal after noise reduction is sampled non-overlappingly, and the length of each small sample is 1024. 14 time domain statistical features are extracted from the expanded sample, including maximum value, minimum value, mean, absolute average amplitude, peak value, variance, standard deviation, kurtosis, skewness, root mean square value, form factor, peak factor, pulse factor and margin factor; 12 frequency domain statistical features include mean, variance, third-order central moment of variance, fourth-order central moment of variance, total mean, standard deviation, C index, D index, E index, G index, third-order central moment and fourth-order central moment.

[0154] In addition, multi-scale fuzzy entropy (MFE) is selected to characterize the degradation state of the linear guide pair. The multi-scale fuzzy entropy is used for the original vibration signal {x1,x2,x3,......,x N}, introduce the scaling factor τ = 1, 2, 3, ..., n, perform coarse-graining on the original signal, observe the trend change of the fuzzy entropy of the signal at different time scales, and then characterize the degradation state of the linear guide pair. The new coarse-grained vector is expressed as follows:

[0155]

[0156] Furthermore, in S4, a model for identifying the degradation state of rolling linear guide pair under variable working conditions is constructed. The model network structure is as follows: Figure 4 The specific contents are as follows:

[0157] A one-dimensional convolutional neural network is used as the feature extraction and classification model, and a domain adaptation layer is introduced into it. By minimizing the distribution difference of domain features in this layer, the performance degradation state of the guide rail pair under different working conditions can be identified.

[0158] The network structure consists of an input layer, two convolutional layers, two batch normalization (BN) layers, a fully connected layer, and an output layer. The two convolutional layers are configured with 32 5×1 kernels. Relu is used as the activation function for the convolutional layers. Finally, Softmax is selected as the classification layer with 4 output nodes. The Adam optimizer is used for gradient descent. Domain distribution adaptation is implemented on the fully connected layers by calculating the MMD distance between domain features on the adaptation layer and incorporating the MMD into the loss function. The loss function includes the source domain classification loss and the domain adaptation loss. The optimization goal is to minimize the difference between the two. The source domain classification loss uses the cross-entropy loss function. The total loss is expressed as follows:

[0159] L=L s+δ m L m ;

[0160] Among them, L s , L m are the cross entropy loss and the MMD distance of the domain adaptation layer, δ m >0 is L m The penalty term coefficient. During network training, the parameter update expression of each network layer is expressed as:

[0161]

[0162] Furthermore, in S5, the data is divided into a training data set and a test data set according to the working conditions, and the data set is sent to the rolling linear guide pair performance degradation state recognition model in S4 for training, and the performance degradation state recognition is performed using the trained model. The recognition method process is as follows: Figure 5 As shown, the specific steps are:

[0163] S51, select the data of working conditions 1 and 5 as source domain training data, and the data of working conditions 2, 3 and 4 as unlabeled target domain test data;

[0164] S52. Initialize the network parameters and set the initial learning rate of the deep migration network to 0.001. The learning rate decays as the training process progresses. The expression is:

[0165]

[0166] Among them, δ0 is the initial learning rate, m is the current number of iterations, p is the number of training rounds, which is set to 30, l is the number of samples of source domain data, and batchsize is the batch size of training data, which is set to 100;

[0167] S53, the source domain data and the target domain data are transmitted to the recognition network model, and the training labels of the source domain data and the feature distributions of the source domain and the target domain on the fully connected layer are obtained through forward propagation. The cross entropy loss and the MMD distance on the domain adaptation layer are calculated, and the two are combined to form the total loss function;

[0168] S54. Calculate the propagation gradient of the network weights and biases according to the total loss function, use the error back propagation and Adam optimization algorithm to train the model parameters according to the maximum number of iterations, and complete the network training;

[0169] S55. The three groups of target domain test data sets are respectively transmitted to the trained recognition network model to obtain prediction results, thereby realizing performance degradation identification under target working conditions.

[0170] Example 3

[0171] The method of the present invention is described below by way of an embodiment.

[0172] In this embodiment, the test sample selected is the LGAS35AN guide rail produced by Guangdong Kate Precision Machinery Co., Ltd., and the specific parameters of the guide rail pair are shown in Table 1.

[0173] Table 1

[0174]

[0175]

[0176] A running-in test was conducted on the rolling linear guideway, collecting vibration data under four different performance degradation states. Five slider speeds were set for each performance degradation state, and five sets of vibration data were collected. Five sets of vibration data were collected for each performance degradation state, for a total of 20 sets of vibration data.

[0177] Furthermore, the vibration data is decomposed using the complete ensemble empirical mode decomposition method with adaptive noise, resulting in a series of intrinsic mode components (IMFs) containing information from different frequency bands. The correlation coefficient between the IMFs and the original vibration signal is then calculated, and IMFs with a correlation coefficient greater than 0.3 are selected. Wavelet adaptive thresholding is then applied to these selected IMFs to remove noise components. Finally, the denoised IMFs are summed to obtain the denoised reconstructed signal.

[0178] Furthermore, the reconstructed signal was truncated to 1024 data samples using non-overlapping sampling. The number of samples for each degradation state under various operating conditions is shown in Table 2. For each sample, 14 time-domain features, 12 frequency-domain features, and 16 fuzzy entropy features were extracted. These features were combined into a 42 × 1 data sample. Signal features from three position channels, namely the slider top surface, vertical side surface, and guide rail, were superimposed to obtain a 126 × 1 input sample, completing the fusion of sensor information from different positions at the feature set level.

[0179] Furthermore, the datasets of the source and target domains are fed into a domain adaptation network model based on a one-dimensional convolutional neural network. The accuracy of the constructed model is trained and verified using three sets of migration tasks. The prediction is repeated 10 times for each migration task, and the average of the 10 prediction results is taken as the recognition result of the proposed method. The prediction results are shown in Table 3. Figure 6a 、 Figure 6b 、 Figure 6cThis is the confusion matrix of the recognition results of one of the training processes in the three groups of migration tasks. The confusion matrix shows that the recognition model can accurately distinguish between degradation states 3 and 4. The vibration signal features of the first two degradation states partially overlap, so there are some recognition errors, but the recognition accuracy of each degradation state reaches more than 90%, indicating that the method proposed in the present invention can realize the accurate identification of the performance degradation state of the linear guide pair under variable working conditions.

[0180] Table 2

[0181]

[0182]

[0183] Table 3

[0184] Migration tasks Task 1 Task 2 Task 3 Recognition results (%) 94.14±0.88 95.17±0.65 95.35±0.34

[0185] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0186] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for identifying the performance degradation state of a rolling linear guide pair under variable working conditions, characterized in that: The following steps are involved: Collect vibration signals at different operating speeds and positions during the performance degradation of the rolling linear guide pair throughout its life cycle; The vibration signal is preprocessed using a joint denoising method based on CEEMDAN and wavelet adaptive threshold; The time domain, frequency domain and multi-scale fuzzy entropy features are extracted from the denoised vibration signal to construct a feature set; Construct a domain-adaptive degradation state recognition model based on one-dimensional convolutional neural network; Conduct multiple sets of degradation state identification tests under varying working conditions. Randomly select at least two sets of working condition data as labeled source domain training data, and the remaining working condition data as unlabeled target domain test data. Use the trained model to identify the performance degradation state of the rolling linear guide pair.

2. The method for identifying the performance degradation state of a rolling linear guide pair under variable working conditions according to claim 1, characterized in that: The vibration signals of the rolling linear guide pair at different operating speeds and different positions during the performance degradation process of the entire life cycle are collected. Specifically, the rolling linear guide pair is installed on the loaded running-in test bench, and the running-in mileage is set. After a single running-in is completed, the guide to be tested is removed from the loaded running-in test bench. The three-axis vibration sensor is attached to the top surface of the slider, and the single-axis vibration sensor is attached to one end of the guide rail. The vibration signals of four round trips are collected through the vibration signal acquisition system at the specified operating speed and sampling frequency.

3. The method for identifying the performance degradation state of a rolling linear guide pair under variable working conditions according to claim 1, characterized in that: The vibration signal is preprocessed using a joint denoising method based on CEEMDAN and wavelet adaptive threshold. The method specifically includes the following steps: the original vibration signal is decomposed into different intrinsic mode components through an empirical signal decomposition method, the eigenmode components containing degradation information are screened out from the decomposed eigenmode components, and irrelevant false components are eliminated; the residual noise components in the selected eigenmode components are suppressed using wavelet adaptive threshold decomposition, and the processed eigenmode components are added together to obtain the denoised reconstructed signal.

4. The method for identifying the performance degradation state of a rolling linear guide pair under variable working conditions according to claim 1, characterized in that: The specific steps of CEEMDAN are as follows: The original vibration signal x(t) is repeatedly superimposed with i groups of Gaussian white noise m with specific amplitudes. i , generate the signal to be decomposed x i (t): x i (t)=x(t)+m i ; Treat the decomposed signal x i (t) Perform EMD decomposition, that is, identify the signal to be decomposed x i All the maximum and minimum points of (t) are fitted with the upper and lower envelopes respectively using cubic spline interpolation to construct the upper and lower envelopes e imax (t) and e imin (t), calculate the mean envelope m i (t): m i (t)=[e imax (t)+e imin (t)] / 2; Use h i (t) = x i (t)-m i (t), extract the candidate intrinsic mode components, judge whether the number of its extreme points and zero-crossing points is equal or differs by 1 and whether the mean envelope is zero. If the eigenmode component conditions are met, the first eigenmode component IMF1 is obtained; if not, let x i (t) = h i (t), repeat the above steps until the eigenmode component conditions are met; The method for selecting the eigenmode components that are sensitive to performance degradation includes calculating the correlation coefficient between the eigenmode components and the original signal, which is expressed as follows: in, and are the means of the original signal and the eigenmode components, respectively. The threshold of the correlation coefficient is set, and the eigenmode components with correlation coefficients higher than the specified threshold are selected; For the selected intrinsic mode components, wavelet threshold is used to reduce the noise components in the selected intrinsic mode components; for a set of noisy intrinsic mode components x(t), the wavelet threshold denoising process is as follows: Determine the appropriate wavelet basis function and decomposition layer number, perform N-layer wavelet decomposition on x(t), and obtain a series of wavelet coefficients ω j,k ; Determine the appropriate threshold T, quantize the detail coefficients of the first layer N layer through the threshold function, and obtain the estimated value of the detail coefficient ω j,k ; Perform inverse wavelet transform on the approximate coefficients of the Nth layer and the detail coefficients after quantization to obtain the denoised signal x(t); The denoised intrinsic mode components are added together to obtain the denoised reconstructed signal.

5. The method for identifying the performance degradation state of a rolling linear guide pair under variable working conditions according to claim 1, characterized in that: The time domain, frequency domain and multi-scale fuzzy entropy features are extracted from the denoised vibration signal to construct a feature set, which specifically includes the following steps: A model for identifying the performance degradation of rolling linear guideways under varying operating conditions uses a one-dimensional convolutional neural network as a feature extraction and classification model. A domain adaptation layer is introduced to minimize the distribution differences of domain features in this layer, enabling identification of guideway performance degradation under varying operating conditions. The grid structure includes an input layer, a feature extraction layer, a domain adaptation layer, a fully connected layer, and an output layer. The loss function includes source domain classification loss and domain adaptation loss. The feature extraction layer includes single-scale convolution layer, multi-scale convolution layer and other feature extraction modules; The classification layer includes the Softmax function; Gradient descent calculation optimizers include Adam optimizer; The domain adaptation layer aims to achieve domain distribution adaptation on the fully connected layer, including calculating the MMD distance of inter-domain features on the adaptation layer and adding MMD to the loss function; The loss function includes source domain classification loss and domain adaptation loss. The optimization goal is to minimize the loss after the two. The source domain classification loss uses the cross entropy loss function. The total loss expression is as follows: L=L s +δ m L m ; Among them, L s , L m are the cross entropy loss and the MMD distance of the domain adaptation layer, δ m >0 is L m The penalty coefficient of During network training, the parameter update expression of each network layer is expressed as: Among them, α is the network learning rate, ω l and b l are the weights and bias parameters of the network layers during training.

6. The method for identifying the performance degradation state of a rolling linear guide pair under variable working conditions according to claim 1, characterized in that: Conduct multiple sets of degradation state recognition tests under varying working conditions. Randomly select at least two sets of working condition data as labeled source domain training data, and the remaining working condition data as unlabeled target domain test data. Use the trained model to identify the performance degradation state of the rolling linear guide pair. Specifically, the following steps are included: Select at least two sets of working condition data as labeled source domain training data, and the remaining working condition data as unlabeled target domain test data; Initialize the network parameters and set the initial learning rate of the degradation state recognition network. The learning rate decay method includes decaying with the training process. The expression is: δ=αδ0; Among them, δ0 is the initial learning rate, α is the attenuation coefficient corresponding to different attenuation methods; The source domain data and target domain data are fed into the degradation state recognition network. The training labels of the source domain data and the feature distributions of the source and target domains on the fully connected layer are obtained through forward propagation. The cross entropy loss and the MMD distance on the domain adaptation layer are calculated and combined to form the total loss function. Calculate the propagation gradient of network weights and biases based on the total loss function, use error backpropagation and optimization algorithm to train model parameters according to the maximum number of iterations to complete network training; The remaining unlabeled target domain test data sets are respectively fed into the trained degradation state recognition network to obtain prediction results, thereby realizing performance degradation identification under the target working conditions.

Images