Prediction Method for Remaining Service Life of Ball Screw Pair Based on Raceway Surface Profile

Abstract

The present application provides a method for predicting the remaining service life of a ball screw pair based on the raceway surface profile, which mainly includes the following steps: collecting the surface profile curves of the screw raceway of the ball screw pair at three positions by a Taylor Hobson profilometer, and performing shape removal and Gaussian filtering on the profile; extracting the main features of the raceway surface profile curve, including roughness, maximum peak-valley height, root mean square, recurrence law, fractal dimension, and multifractal spectrum width, to construct a mixed feature set; then performing dimensionality reduction processing, evaluating the importance of each feature in the feature set by the random forest method, and selecting the features containing 95% of the information to construct a new feature set; establishing a bidirectional long short-term memory neural network model; and importing the mixed feature set at a preset position into the established bidirectional long short-term memory neural network model for training. The method provided by the present application is convenient, has high accuracy, and small error.

Description

Technical Field

[0001] This application relates to the technical field of ball screw pairs, and particularly to a method for predicting the remaining service life of a ball screw pair based on the raceway surface profile. Background Art

[0002] Ball screw pairs are widely used in numerical control machine tools due to their good positioning accuracy and load-bearing performance. As the main transmission mechanism of the machine tool system, the wear of the ball screw will cause the performance of the ball screw pair to degenerate continuously, and even cause damage to the machine tool and endanger people's lives. Therefore, it is particularly important to be able to accurately predict the remaining service life of the ball screw pair.

[0003] At present, the methods for establishing the remaining service life model of ball screw pairs are mainly divided into methods based on physical models and methods based on data-driven. Due to the complexity of the ball screw pair system, it is very difficult to establish an accurate physical model to predict the remaining service life of the ball screw pair. With the continuous development of computer technology, data-driven methods are applied more and more widely. At present, the prediction of the remaining service life of ball screw pairs is achieved by collecting vibration signals during the operation of the ball screw pair and extracting their features. However, since this method is easily interfered by noise and the like, the prediction result is poor.

[0004] The wear condition during the operation of the ball screw pair can be reflected by the change of the raceway surface profile. Therefore, the present invention proposes a method for predicting the remaining service life of a ball screw pair based on the raceway surface profile. This method not only overcomes the shortcomings of signal means, but also has the advantages of practicality, convenience, high accuracy, small error, etc., and provides a new method for predicting the remaining service life of the ball screw pair. Summary of the Invention

[0005] This application provides a method for predicting the remaining service life of a ball screw pair based on the raceway surface profile, which can be used to solve the technical problem that the prediction of the remaining service life of the ball screw pair is not accurate enough.

[0006] This application provides a method for predicting the remaining service life of a ball screw pair based on the raceway surface profile, and the method includes: This application provides a method for predicting the remaining service life of a ball screw pair based on the raceway surface profile, and the method includes:

[0007] Collect the contour curves of the screw raceway surface of the ball screw pair at multiple preset positions within a preset area, and perform shape removal and Gaussian filtering processing on the contour curves;

[0008] Use comprehensive statistical analysis, recursive analysis, and fractal analysis methods to extract features from the contour curves of the raceway surface and construct a mixed feature set;

[0009] Perform dimensionality reduction processing. Evaluate the importance of each feature in the mixed feature set by the random forest method, sort each feature in the mixed feature set from high to low according to the importance value, and select the features that accumulate more than 95% of the information to construct a new feature set;

[0010] Input the new mixed feature set at the preset position into the bidirectional long short-term memory neural network model at the target accuracy to obtain the prediction result of the remaining service life of the ball screw pair;

[0011] Among them, the bidirectional long short-term memory neural network model at the target accuracy is determined by the following method:

[0012] Step 1, collect the contour curves of the screw raceway surface of the ball screw pair at multiple preset positions in the preset area through a Taylor Hobson profilometer, and perform shape removal and Gaussian filtering processing on the contour curves;

[0013] Step 2, use comprehensive statistical analysis, recursive analysis, and fractal analysis methods to extract features from the contour curves of the raceway surface and construct a mixed feature set; the features include roughness, maximum peak-to-valley height, root mean square, recurrence law, fractal dimension, and multifractal spectrum width;

[0014] Step 3, perform dimensionality reduction processing. Evaluate the importance of each feature in the mixed feature set by the random forest method, sort each feature in the mixed feature set from high to low according to the importance value, and select the features that accumulate more than 95% of the information to construct a new feature set;

[0015] Step 4, establish a bidirectional long short-term memory neural network model, and set the number of network layers and the number of neurons in the hidden layer according to the experiment;

[0016] Step 5, import the new mixed feature set at the preset position into the established bidirectional long short-term memory neural network model for training until the bidirectional long short-term memory neural network model at the target accuracy is obtained.

[0017] Optionally, collecting the contour curves of the screw raceway surface of the ball screw pair at multiple preset positions in the preset area through a Taylor Hobson profilometer, and performing shape removal and Gaussian filtering processing on the contour curves includes:

[0018] Select three raceway positions of the screw in the uniform motion area, perform notch marking on them, place the ball screw pair on the wear test bench. When the ball screw pair runs 300,000 revolutions before operation, stop the test bench every 30,000 revolutions and remove the screw, and collect the surface contour curves of its three raceway positions through a Taylor Hobson profilometer; among them, the preset area is the area where the screw runs at a uniform speed during the operation of the ball screw pair;

[0019] After the ball screw pair runs 300,000 revolutions, stop the test bench every 60,000 revolutions and remove the screw to conduct a profile acquisition test once.

[0020] Smooth it through Gaussian filtering, and adopt the quintic polynomial method to remove the shape of the smoothed curve, and finally export the required raceway surface profile curve.

[0021] Optionally, the mixed feature set is determined by the following method:

[0022] Step 201: Solve the roughness, maximum peak-to-valley height, and root mean square feature of the extracted raceway profile curve through statistical analysis methods. The used formulas are:

[0023]

[0024] Rz = z max -z min

[0025]

[0026] Among them, Ra is the roughness, Rz is the maximum peak-to-valley height, Rms is the root mean square, z i is the profile height, z min and z max are the minimum profile height and the maximum profile height respectively, is the mean height of the profile, and n is the sampling rate;

[0027] Step 202: Solve the recurrence law of the extracted raceway profile curve by using the recurrence analysis method. The formula is:

[0028] r ij = |z i -z j |

[0029] ε = 0.5σ

[0030] R ij (ε) = θ(ε - r ij )

[0031]

[0032]

[0033] Among them, RR is the recurrence law, z i is one of the profile heights, z j is the other profile height, i, j = 1, 2,..., n, r ij represents the distance between any two points, R ijis an element of the matrix, σ is the standard deviation, ε is the threshold, and θ(x) is the Heaviside function;

[0034] Step 203: The W-M function is used to characterize the non-linear raceway surface profile, and the formula is:

[0035]

[0036] where 1 < D < 2, γ > 1, D is the fractal dimension, G is the height scale coefficient, γ n is the spectrum of the rough surface, n is the sampling rate, and n1 is the minimum sampling rate. L is the sampling length, γ = 1.5, z(x) is the height of the random profile, and x is the position coordinate of the profile;

[0037] The power spectrum function of the above formula is expressed as the formula:

[0038]

[0039] Define the incremental variance of z(x) as the structure function, as shown in the following formula:

[0040]

[0041] where τ = nΔL and ΔL is the sampling interval;

[0042] Combining Equation and Equation gives Equation:

[0043]

[0044] where C = Γ(2D - 3)sin((D - 1.5)π) / (4 - 2D)lnγ, and Γ(*) is the gamma function.

[0045] Taking the logarithm of both sides of Equation gives Equation:

[0046] lgS(τ) = (4 - 2D)lgτ + lgC + 2(D - 1)lgG

[0047] According to the above formula, the fractal dimension D is:

[0048]

[0049] where k is the slope of the straight line;

[0050] Step 204: Use the box-counting method to calculate the multifractal spectrum of the raceway profile. Extract the minimum value of the collected raceway profile data as the lower limit zero to ensure that all amplitudes are positive. Cover the raceway profile with multiple small boxes of size ε (0 < ε < 1). The total profile height S iLet \(S_i(\varepsilon)\) denote the sum of all contour amplitudes within the \(i\)-th small box when the box size is \(\varepsilon\), then the probability measure \(P(\varepsilon)\) i is defined by the formula:

[0051]

[0052] where \(\sum S(\varepsilon)\) i is the sum of the amplitudes of all raceway contour data;

[0053] In the scale-free interval, \(P(\varepsilon)\) i is expressed in exponential form, such as the formula:

[0054] \(P(\varepsilon)\) i \(\sim \varepsilon^{-\alpha}\) α

[0055] where \(\alpha\) is the singularity exponent, reflecting the singularity strength of \(P(\varepsilon)\); i

[0056] Suppose the number of boxes with the same singularity exponent \(\alpha\) is \(N(\varepsilon)\), then in the scale-free interval \(N(\varepsilon)\) α is expressed in exponential form: α

[0057] \(N(\varepsilon)\) α \(\sim \varepsilon^{-f(\alpha)}\) -f(α)

[0058] where \(f(\alpha)\) represents the fractal dimension corresponding to the singularity exponent \(\alpha\), and the smaller \(f(\alpha)\) is, the smaller \(N(\varepsilon)\) is; α

[0059] Define the partition function of multifractals as \(\chi(\varepsilon)\), and the formula is: q

[0060] \(\chi(\varepsilon)=\sum P(\varepsilon)^q\) q i \(=\varepsilon^{-\tau(q)}\) q τ(q)

[0061] where \(q\) is the weight factor and \(\tau(q)\) is the mass exponent; in this application, the value of \(q\) is taken as \([-20, 30]\) with an interval of 1;

[0062] When \(\varepsilon \to 0\), \(\tau(q)\) is expressed as:

[0063]

[0064] Relate the three parameters \(\alpha\), \(f(\alpha)\), and \(\tau(\alpha)\) and obtain the formula according to the Legendre transform relationship existing among the three:

[0065]

[0066] f(α) = q·α(q) - τ(q)

[0067] α and f(a) form a multifractal spectrum diagram; the multifractal spectrum width Δα is defined as:

[0068] Δα = α max -α min

[0069] where α min and α max are the minimum singularity index and the maximum singularity index respectively, Δα represents the inhomogeneity of the probability measure of the sequence, corresponding to the fluctuation range of the surface profile height. The larger the fluctuation range, the larger Δα;

[0070] Step 205, Combine the roughness, maximum peak-valley height, root mean square, recurrence law, fractal dimension, and multifractal spectrum width of the surface profiles at the three raceway positions through the above method to construct a hybrid feature set.

[0071] Optionally, the objective function of the dimensionality reduction processing is defined as:

[0072]

[0073] where f is the feature according to which the splitting is performed, D p and D j are the parent node and the jth child node, I is the impurity content, N p is the number of samples of the parent node, N j is the number of samples of the child node, and m is the number of child nodes owned by each parent node; The Gini impurity (I G ) is selected as the impurity metric and is defined as:

[0074]

[0075] where p(i|t) represents the probability that a certain node t belongs to the c-class sample.

[0076] Select 25 decision trees to form a random forest, use the Gini impurity as the judgment criterion, import the feature mixture set in step 3 into Python, extract the features containing 95% of the original information to construct a new hybrid feature set, and thus achieve dimensionality reduction.

[0077] Optionally, the bidirectional long short-term memory neural network model includes an input layer, a Bi-LSTM layer, a fully connected layer, and an output layer;

[0078] The number of neurons in the input layer is set to 6, the number of neurons in the fully connected layer and the output layer is set to 1. According to the trial method, two Bi-LSTM layers are set, and the number of neurons is 8 and 12 respectively. The learning rate is taken as 0.01, the maximum number of iterations is set to 500, the mini-batch gradient descent method is adopted, and the batch size is set to 15. The neural network model is optimized by the Adam algorithm.

[0079] Optionally, import the mixed feature set at the preset position into the established bidirectional long short-term memory neural network model for training until a bidirectional long short-term memory neural network model with the target accuracy is obtained, including:

[0080] Preprocess the mixed feature sets at three preset positions to be between [-1, 1];

[0081] Import the mixed feature set constructed by the raceway surface profiles at position one and position two into the bidirectional long short-term neural network model for training; then import the mixed feature set of the raceway at position three into the trained neural network model to realize the prediction of the remaining service life of the ball screw pair, compare it with the true value, and calculate the root mean square error rmse. The formula is:

[0082]

[0083] Among them, is the network predicted value, y i is the true value;

[0084] When the mean square error reaches the expectation, a bidirectional long short-term memory neural network model with the target accuracy is obtained.

[0085] The method provided by this application can calculate the remaining service life of the ball screw pair more accurately and concisely. It overcomes the disadvantage that the extraction of features from vibration signals in the traditional method has large noise interference, resulting in inaccurate prediction of the remaining service life. Moreover, the bidirectional long short-term memory neural network model established by the method of the present invention can well realize the prediction of the remaining service life of the ball screw pair, with small errors, and the method provided by the embodiments of this application has high accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] Figure 1 is a schematic structural diagram of the acquisition of the raceway surface profile of the ball screw pair provided by the embodiment of this application;

[0087] Figure 2 is a schematic structural diagram of the ball screw pair wear test bench provided by the embodiment of this application;

[0088] Figure 3 is a schematic flow diagram of the prediction of the remaining service life of the ball screw pair based on the raceway surface profile provided by the embodiment of this application;

[0089] Figure 4Schematic diagram of fatigue spalling of ball screw pair provided by the embodiment of the present application;

[0090] Figure 5 Schematic diagram of dimensionality reduction scoring based on random forest evaluation provided by the embodiment of the present application;

[0091] Figure 6 Schematic diagram of training of remaining service life prediction model of ball screw pair based on raceway surface profile provided by the embodiment of the present application;

[0092] Figure 7 Schematic diagram of prediction of remaining service life prediction model of ball screw pair based on raceway surface profile provided by the present application implementation;

[0093] Figure 8 Schematic diagram of model effects before and after dimensionality reduction provided by the embodiment of the present application. Detailed implementation manners

[0094] To make the objectives, technical solutions and advantages of the present application clearer, the following will further describe the implementation manners of the present application in detail with reference to the accompanying drawings.

[0095] The following will first combine Figure 1 to introduce the raceway surface profile acquisition test of the ball screw pair.

[0096] Please refer to Figure 1 , the raceway surface profile acquisition test of the ball screw pair mainly includes Taylor Hobson profilometer, support unit, PC and V-blocks, etc. The Taylor Hobson profilometer is used to extract the surface profile of the screw raceway, and the ball screw pair is placed on two V-blocks for support. The profilometer needs to be leveled before measurement, and its sampling length is set to 1 mm. The measurement position is the 50th raceway, and it is marked with a notch to record the measurement position. The same operation is also performed on the 68th raceway and the 98th raceway to collect the surface profile data. The three selected raceway positions are all located in the uniform motion area. Here, they are respectively denoted as Position 1 to Position 3.

[0097] Then, the wear test of the ball screw pair is carried out. Refer to Figure 2, the ball screw pair wear test bench mainly consists of an eddy current brake, a loaded ball screw pair, linear guides, a workbench, and a stepping motor. By setting the magnitude of the eddy current brake current, the axial load applied to the ball screw pair can be adjusted. In this test, the eddy current brake current is set to 2.5 A, and the corresponding axial load is 25 KN. The ball screw pair to be measured is installed on this wear test bench. Before the ball screw pair runs 300,000 revolutions, the test bench is stopped every 30,000 revolutions and the screw is removed. The surface profile curves of its three raceway positions are collected by a Taylor Hobson profilometer. After the ball screw pair runs 300,000 revolutions, the test bench is stopped every 60,000 revolutions and the screw is removed for a profile collection test. Then, it is smoothed by Gaussian filtering, and the shape of the smoothed curve is removed by the fifth-order polynomial method, and finally the required raceway surface profile curve is exported.

[0098] The method provided by this application includes the following steps:

[0099] Collect the profile curves of the screw raceways of the ball screw pair at multiple preset positions within a preset area, and perform shape removal and Gaussian filtering on the profile curves;

[0100] In the embodiment of this application, there are three operating regions of uniform acceleration, uniform speed, and uniform deceleration in the screw region during the operation of the ball screw pair. Since the change of the raceway surface profile in the uniform speed operating region is more uniform and stable, the three raceway positions are all taken from the uniform speed operating region.

[0101] Use comprehensive statistical analysis, recursive analysis, and fractal analysis methods to extract features from the profile curves of the raceway surface and construct a mixed feature set;

[0102] Perform dimensionality reduction processing. Evaluate the importance of each feature in the mixed feature set by the random forest method, sort each feature in the mixed feature set from high to low according to the importance value, and select the features that accumulate more than 95% of the information to construct a new feature set;

[0103] Input the new mixed feature set at the preset position into the bidirectional long short-term memory neural network model under the target accuracy to obtain the prediction result of the remaining service life of the ball screw pair;

[0104] Among them, the bidirectional long short-term memory neural network model under the target accuracy is determined by the following method:

[0105] Step 1, collect the profile curves of the screw raceways of the ball screw pair at multiple preset positions within a preset area by a Taylor Hobson profilometer, and perform shape removal and Gaussian filtering on the profile curves.

[0106] Specifically, Step 1 includes the following steps:

[0107] Select three raceway positions of the lead screw in the uniform motion area, make indentation marks on them, place the ball screw pair on the wear test bench. Before the ball screw pair runs 300,000 revolutions, stop the test bench every 30,000 revolutions and remove the lead screw, and collect the surface profile curves of its three raceway positions through a Taylor Hobson profilometer; among them, the preset area is the area where the lead screw runs at a uniform speed during the operation of the ball screw pair;

[0108] After the ball screw pair runs 300,000 revolutions, stop the test bench every 60,000 revolutions and remove the lead screw for a profile acquisition test;

[0109] Perform smoothing processing on it through Gaussian filtering, adopt the quintic polynomial method to remove the shape of the smoothed curve, and finally export the required raceway surface profile curve.

[0110] Step 2, use comprehensive statistical analysis, recursive analysis, and fractal analysis methods to extract the characteristics of the profile curve of the raceway surface and construct a mixed feature set; the features include roughness, maximum peak-to-valley height, root mean square, recurrence law, fractal dimension, and multifractal spectrum width;

[0111] The mixed feature set is determined by the following method. Further, extract the characteristic values of the collected raceway surface profile curve through statistical analysis method, recursive analysis method, and fractal analysis method. The specific steps are as follows:

[0112] Step 201: First, solve the roughness, maximum peak-to-valley height, and root mean square characteristics of the extracted raceway profile curve through the statistical analysis method. The characteristics extracted by this method can reflect the most primitive characteristics of the raceway profile and are more sensitive to wear changes. The larger the three, the rougher and more uneven the raceway profile. The formula used in this method is:

[0113]

[0114] Rz = z max -z min

[0115]

[0116] Among them, Ra is the roughness, Rz is the maximum peak-to-valley height, Rms is the root mean square, z i is the profile height, z min and z max are the minimum profile height and the maximum profile height respectively, is the mean height of the profile, and n is the sampling rate.

[0117] Step 202: Use the recursive analysis method to solve the recurrence law of the extracted raceway contour curve. The recurrence law can reflect the volatility and stability of the raceway contour. The larger the recurrence law, the smoother the corresponding contour and the smaller the volatility. The formula used in this method is:

[0118] r ij =|z i -z j |

[0119] ε=0.5σ

[0120] R ij (ε)=θ(ε - r ij )

[0121]

[0122]

[0123] Among them, RR is the recurrence law, z i is one of the contour heights, z j is the other contour height, i, j = 1, 2,..., n, r ij represents the distance between any two points, R ij is an element of the matrix, σ is the standard deviation, ε is the threshold, and θ(x) is the Heaviside function.

[0124] Both the statistical analysis method and the recursive analysis method are scale-dependent analysis methods and are related to the sampling length of the device. Since the fractal dimension has self-similarity and scale invariance, which are scale-independent characteristics, the fractal dimension is introduced. There are many methods for solving the fractal dimension, and the structure function method is relatively accurate. The present invention solves the fractal dimension according to the structure function method. The W-M function is suitable for characterizing the non-linear raceway surface contour, and the formula is:

[0125]

[0126] Among them, 1 < D < 2, γ > 1, D is the fractal dimension, G is the height scale coefficient, γ n is the spectrum of the rough surface, n is the sampling rate, n1 is the minimum sampling rate, L is the sampling length, γ = 1.5, z(x) is the height of the random contour, and x is the position coordinate of the contour,

[0127] The power spectrum function of the above formula can be expressed as the formula:

[0128]

[0129] Define the incremental variance of z(x) as the structure function, as shown in the following formula:

[0130]

[0131] Among them, τ = nΔL, where ΔL is the sampling interval.

[0132] By combining the equations and performing summation, the equation can be obtained.

[0133]

[0134] Among them, C = Γ(2D - 3)sin((D - 1.5)π) / (4 - 2D)lnγ, where Γ(*) is the gamma function.

[0135] Taking the logarithm of both sides of the equation gives the equation:

[0136] lgS(τ) = (4 - 2D)lgτ + lgC + 2(D - 1)lgG

[0137] According to the above equation, the fractal dimension D is:

[0138]

[0139] Among them, k is the slope of the straight line. The fractal dimension can reflect the complexity and irregularity of the raceway profile. The larger the fractal dimension, the more complex and irregular the raceway profile.

[0140] Step 204: Since the fractal dimension only describes the raceway profile from a single measure, and the multifractal spectrum can describe the raceway profile from multiple measures, the multifractal spectrum is introduced. The box-counting method is used to calculate the multifractal spectrum of the raceway profile. The minimum value of the collected raceway profile data is extracted as the lower limit zero to ensure that all amplitudes are positive. The raceway profile is covered with many small boxes of size ε (0 < ε < 1). The total profile height S i (ε) represents the sum of all profile amplitudes within the i-th small box when the box size is ε. Then the probability measure P i (ε) can be defined by the formula:

[0141]

[0142] Among them, ∑S i (ε) is the sum of all raceway profile data amplitudes.

[0143] In the scale-free interval, P i (ε) can also be written in exponential form, such as the formula:

[0144] P i (ε) ~ ε α

[0145] Among them, α is the singularity exponent, which can reflect P i(ε) singular strength.

[0146] Assume the number of boxes with the same singularity index α is N α (ε), then within the scale-free interval, N α (ε) can be written in exponential form:

[0147] N α (ε) ~ ε -f(α)

[0148] where f(α) represents the fractal dimension corresponding to the singularity index α, and the smaller f(α) is, the smaller N α (ε) is. Since it is difficult to directly calculate the number of boxes with the singularity index, the partition function is introduced for calculation. Define the partition function of multifractal as χ q (ε), and its formula is:

[0149] χ q (ε) = ∑P i (ε) q = ε τ(q)

[0150] where q is the weight factor and τ(q) is the mass exponent. The value range of q should be [-∞, +∞], however, in actual calculation, the value of q cannot take infinity, so in this invention, the value of q is [-20, 30] with an interval of 1.

[0151] When ε → 0, τ(q) can be written as the formula:

[0152]

[0153] Relate the three parameters α, f(α) and τ(α) and obtain the formula according to the Legendre transform relationship existing among the three:

[0154]

[0155] f(α) = q·α(q) - τ(q)

[0156] α and f(a) form the multifractal spectrum diagram. Define the multifractal spectrum width Δα as:

[0157] Δα = α max -α min

[0158] where α min and α max are the minimum singularity index and the maximum singularity index respectively. Δα represents the inhomogeneity of the probability measure of the sequence, corresponding to the fluctuation range of the surface profile height. The larger the fluctuation range is, the larger Δα is.

[0159] Step 205: Combine the roughness, maximum peak-valley height, root mean square, recurrence law, fractal dimension, and multifractal spectrum width of the surface profiles at the three raceway positions through the above method to construct a hybrid feature set.

[0160] Step 3: Perform dimensionality reduction. Evaluate the importance of each feature in the hybrid feature set through the random forest method, sort each feature in the hybrid feature set from high to low according to the importance value, and select the features that accumulate more than 95% of the information to construct a new feature set.

[0161] Next, perform dimensionality reduction. First, a target function needs to be defined. This target function can maximize the information gain of each split and is defined as:

[0162]

[0163] where f is the feature based on which to split, D p and D j are the parent node and the jth child node, I is the impurity content, N p is the number of samples in the parent node, N j is the number of samples in the child node, and m is the number of child nodes each parent node has. The information gain of the parent node and the child node only differs in terms of the impurity content, that is, the lower the impurity content of the child node, the greater the information gain. The Gini impurity can be understood as a criterion for minimizing the probability of misclassification. Select the Gini impurity degree (I G ) as the impurity metric and define it as:

[0164]

[0165] where p(i|t) represents the probability that a certain node t belongs to the c-class sample.

[0166] Select 25 decision trees to form a random forest. Use the Gini impurity degree as the judgment criterion. Import the feature mixture set in Step 3 into Python, extract the features containing 95% of the original information to construct a new hybrid feature set, and thus achieve dimensionality reduction.

[0167] Step 4: Establish a bidirectional long short-term memory neural network model, and set the number of network layers and the number of neurons in the hidden layer according to the experiment.

[0168] Next, establish a bidirectional long short-term memory neural network model for predicting the remaining service life of the ball screw pair. The flowchart of this model is as Figure 3As shown in the figure, the model mainly consists of an input layer, a Bi-LSTM layer, a fully connected layer, and an output layer. Since there are a total of six features, the number of neurons in the input layer is set to 6. Since only the remaining service life of the ball screw pair needs to be output, the number of neurons in the fully connected layer and the output layer is set to 1. According to the experimental method, two layers of Bi-LSTM layers are set, and the number of neurons is 8 and 12 respectively. The learning rate is set to 0.01, the maximum number of iterations is set to 500, the mini-batch gradient descent method is adopted, and the batch size is set to 15. The neural network model is optimized by the Adam algorithm.

[0169] Step 5: Import the new mixed feature set at the preset position into the established bidirectional long short-term memory neural network model for training until the bidirectional long short-term memory neural network model under the target accuracy is obtained.

[0170] Specifically, the mixed feature sets at three preset positions are preprocessed to be between [-1, 1]. Then, the mixed feature set constructed from the raceway surface profiles at position 1 and position 2 is imported into the neural network model for training. Then, the mixed feature set of the raceway at position 3 is imported into the trained neural network model to predict the remaining service life of the ball screw pair. It is compared with the true value, and the root mean square error rmse is calculated. The formula is:

[0171]

[0172] Among them, is the network prediction value, and y i is the true value.

[0173] When the mean square error reaches the expectation, the bidirectional long short-term memory neural network model under the target accuracy is obtained.

[0174] The method provided by the present application will be described below through an embodiment.

[0175] Select the test lead screw as the GD4010 series lead screw produced by Botek Precision Co., Ltd. located in Shandong, China. The main parameters are shown in Table 1.

[0176] Table 1: Ball screw pair parameters

[0177]

[0178] Continuously collect the raceway surface profile data until fatigue spalling appears on the raceway surface of the lead screw. At this time, the ball screw pair has rotated a total of 1.42 million revolutions, as Figure 4 shown. The raceway surface profile is processed by Gaussian filtering and the shape is removed by a fifth-order polynomial and then exported. According to the feature extraction method mentioned above, the changes of six features with the running revolutions of the ball screw pair are extracted. As shown in Table 2, the feature extraction results of the raceway at position 1 are listed.

[0179] Table 2: Changes in Raceway Profile Features (10,000 revolutions)

[0180]

[0181]

[0182] Next, the extracted mixed feature set is imported into the random forest evaluation algorithm for dimensionality reduction, and the model scoring results are as Figure 5 shown. It can be found that the original information contained in the first 5 features has exceeded 95%. Therefore, the multifractal spectrum width is discarded, and a new mixed feature set is constructed with five features: roughness, maximum peak-valley height, root mean square, recurrence law, and fractal dimension.

[0183] After establishing the bidirectional long short-term memory neural network model and completing the corresponding parameter settings, the mixed feature sets at the three raceway positions are preprocessed to be between [-1, 1]. The results of dimensionality reduction and preprocessing of the data in Table 2 are shown in Table 3.

[0184] Table 3: Changes in Raceway Profile Features after Dimensionality Reduction and Preprocessing

[0185]

[0186]

[0187] Next, the reduced mixed feature set constructed from the raceway surface profiles at position 1 and position 2 is imported into the neural network model for training, and the training results are as Figure 6 shown. The root mean square error of training is 0.2834.

[0188] Finally, the mixed feature set of the raceway at position 3 is imported into the trained neural network model to predict the remaining service life of the ball screw pair, and the prediction results are as Figure 7 shown. The root mean square error of testing is 1.7028, and the error is small, so the prediction is accurate.

[0189] To verify the influence of dimensionality reduction on the prediction effect of the established neural network model, the mixed feature sets constructed from the non-dimensionality-reduced data and the dimensionality-reduced data are respectively imported into the bidirectional long short-term memory neural network model, and the results are as Figure 8 shown. It can be observed that after dimensionality reduction, the errors of the training set are almost the same, the errors of the test set decrease relatively, and the prediction effect of the remaining service life of the ball screw pair is improved.

Claims

1. A method for predicting the remaining service life of a ball screw pair based on the raceway surface profile, characterized in that, The method includes: Collect the contour curves of the screw raceway surface of the ball screw pair at multiple preset positions within a preset area, and perform shape removal and Gaussian filtering on the contour curves. Use comprehensive statistical analysis, recursive analysis, and fractal analysis methods to extract features from the contour curves of the raceway surface and construct a mixed feature set. Perform dimensionality reduction processing. Evaluate the importance of each feature in the mixed feature set by the random forest method, sort each feature in the mixed feature set from high to low according to the importance value, and select the features that accumulate more than 95% of the information to construct a new feature set. Input the new mixed feature set at the preset position into the bidirectional long short-term memory neural network model under the target accuracy to obtain the prediction result of the remaining service life of the ball screw pair. Among them, the bidirectional long short-term memory neural network model under the target accuracy is determined by the following method: Step 1: Collect the contour curves of the screw raceway surface of the ball screw pair at multiple preset positions within a preset area by a Taylor Hobson profilometer, and perform shape removal and Gaussian filtering on the contour curves. Step 2: Use comprehensive statistical analysis, recursive analysis, and fractal analysis methods to extract features from the contour curves of the raceway surface and construct a mixed feature set; the features include roughness, maximum peak-to-valley height, root mean square, recurrence law, fractal dimension, and multifractal spectrum width. Step 3: Perform dimensionality reduction processing. Evaluate the importance of each feature in the mixed feature set by the random forest method, sort each feature in the mixed feature set from high to low according to the importance value, and select the features that accumulate more than 95% of the information to construct a new feature set. Step 4: Establish a bidirectional long short-term memory neural network model, and set the number of network layers and the number of neurons in the hidden layer according to the experiment. Step 5: Import the new mixed feature set at the preset position into the established bidirectional long short-term memory neural network model for training until the bidirectional long short-term memory neural network model under the target accuracy is obtained.

2. The method according to claim 1, characterized in that, Collect the contour curves of the screw raceway surface of the ball screw pair at multiple preset positions within a preset area by a Taylor Hobson profilometer, and perform shape removal and Gaussian filtering on the contour curves, including: Select three raceway positions of the screw in the uniform motion area, perform notch marking on them, place the ball screw pair on the wear test bench. When the ball screw pair runs 300,000 revolutions before operation, stop the test bench every 30,000 revolutions and remove the screw, and collect the surface contour curves of its three raceway positions by a Taylor Hobson profilometer; among them, the preset area is the area where the screw runs at a uniform speed during the operation of the ball screw pair. After the ball screw pair runs 300,000 revolutions, stop the test bench every 60,000 revolutions and remove the screw for a contour collection test. Perform smoothing processing on it by Gaussian filtering, and adopt the fifth-order polynomial method to perform shape removal on the smoothed curve, and finally export the required raceway surface contour curve.

3. The method according to claim 1, characterized in that, The mixed feature set is determined by the following method: Step 201: Solve the roughness, maximum peak-valley height, and root mean square characteristics of the extracted raceway profile curve through statistical analysis methods. The formulas used are as follows: Rz = z max -z min Among them, Ra is the roughness, Rz is the maximum peak-to-valley height, Rms is the root mean square, z i is the profile height, z min and z max are the minimum profile height and the maximum profile height respectively, is the mean height of the profile, and n is the sampling rate; Step 202: Solve the recurrence law of the extracted raceway profile curve using the recurrence analysis method. The formula is: r ij = |z i - z j | ε = 0.5σ R ij (ε) = θ(ε - r ij ) where RR is the recursion law, z i is one of the profile heights, z j is the other profile height, i, j = 1, 2,..., n, r ij represents the distance between any two points, R ij is an element of the matrix, σ is the standard deviation, ε is the threshold, and θ(x) is the Heaviside function; Step 203: The W-M function is used to characterize the non-linear raceway surface profile. The formula is: where 1 < D < 2, γ > 1, D is the fractal dimension, G is the height scale coefficient, γ n is the spectrum of the rough surface, n is the sampling rate, n1 is the minimum sampling rate, L is the sampling length, γ = 1.5, z(x) is the height of the random profile, and x is the position coordinate of the profile; The power spectral function of the above formula is expressed as the formula: Define the incremental variance of z(x) as the structure function, as shown in the following formula: where τ = nΔL, and ΔL is the sampling interval; Combining Equation and Equation gives Equation : where C = Γ(2D - 3)sin((D - 1.5)π) / (4 - 2D)lnγ, Γ(*) is the gamma function, Taking the logarithm of both sides of Equation gives Equation : lgS(τ) = (4 - 2D)lgτ + lgC + 2(D - 1)lgG According to the above formula, the fractal dimension D is: where k is the slope of the straight line; Step 204: Calculate the multifractal spectrum of the raceway profile using the box-counting method. Extract the minimum value of the collected raceway profile data as the lower limit zero to ensure that all amplitudes are positive. Cover the raceway profile with multiple small boxes of size ε, where 0 < ε < 1, and the total profile height is S i S(ε) represents the sum of all profile amplitudes in the i-th small box when the box size is ε. Then the probability measure P i (ε) is defined by the formula: Among them, ∑S i (ε) is the sum of the amplitudes of all raceway profile data; In the scale-free interval, P i (ε) is expressed in exponential form, such as the formula: P i (ε)~ε α where α is the singularity index, reflecting the singularity intensity of P i (ε); Suppose the number of boxes with the same singularity index α is N α (ε), then within the scale-free interval, N α (ε) is expressed in exponential form as: N α (ε) to ε f(α) Among them, f(α) represents the fractal dimension corresponding to the singular exponent α, and the smaller f(α) is, the smaller N α (ε) is; Define the partition function of the multifractal as χ q (ε), and the formula is as follows: χ q (ε) = ∑P i (ε) q = ε τ(q) where q is the weight factor and τ(q) is the quality index; the value of q is [-20, 30] with an interval of 1; When ε → 0, τ(q) is expressed as: Relate the three parameters α, f(α), and τ(α) and obtain the formula according to the Legendre transform relationship existing among the three: f(α) = q·α(q) - τ(q) α and f(a) form a multifractal spectrum diagram; define the multifractal spectrum width Δα as: Δα = α max -α min where α min and α max are the minimum singular exponent and the maximum singular exponent respectively, Δα represents the inhomogeneity of the probability measure of the sequence, corresponding to the fluctuation range of the surface profile height. The larger the fluctuation range, the larger Δα is; Step 205: Combine the roughness, maximum peak-valley height, root mean square, recurrence law, fractal dimension, and multifractal spectrum width of the surface profiles at the three raceway positions through the above methods to construct a mixed feature set.

4. The method according to claim 1, characterized in that, The objective function of dimensionality reduction processing is defined as: Among them, f is the feature according to which splitting is performed, D p and D j are the parent node and the j-th child node, I is the impurity content, N p is the number of samples of the parent node, N j is the number of samples of the child node, m is the number of child nodes owned by each parent node; the Gini impurity (I G ) is selected as the impurity metric and is defined as: where p(i|t) represents the probability that a certain node t belongs to the c-class sample; Select 25 decision trees to form a random forest, use the Gini impurity as the judgment criterion, import the feature mixture set in Step 3 into Python, extract the features containing 95% of the original information to construct a new mixed feature set, and thus achieve dimensionality reduction.

5. The method according to claim 1, characterized in that The bidirectional long short-term memory neural network model includes an input layer, a Bi-LSTM layer, a fully connected layer, and an output layer; The number of neurons in the input layer is set to 6, the number of neurons in the fully connected layer and the output layer is set to 1. Set the two Bi-LSTM layers according to the experimental method, and the number of neurons is 8 and 12 respectively. The learning rate is taken as 0.01, the maximum number of iterations is set to 500, and the mini-batch gradient descent method is adopted. The batch size is set to 15, and the neural network model is optimized through the Adam algorithm.

6. The method according to claim 4, characterized in that, Import the mixed feature set at the preset position into the established bidirectional long short-term memory neural network model for training until a bidirectional long short-term memory neural network model with the target accuracy is obtained, including: preprocessing the mixed feature sets at the three preset positions to be between [-1, 1]; Import the mixed feature set constructed from the surface profiles of the raceways at position one and position two into the bidirectional long short-term neural network model for training; then import the mixed feature set of the raceway at position three into the trained neural network model to predict the remaining service life of the ball screw pair, compare it with the true value, and calculate the root mean square error rmse. The formula is: Among them, is the network prediction value, y i is the true value; When the mean square error reaches the expectation, obtain the bidirectional long short-term memory neural network model under the target accuracy.

Images