Prediction method of part surface roughness and tool wear based on multi-task learning

Abstract

A prediction method of part surface roughness and tool wear based on multi-task learning belong to the file of machining technology. Firstly, the vibration signals in the machining process are collected; next, the part surface roughness and tool wear are measured, and the measured results are corresponding to the vibration signals respectively; secondly, the samples are expanded, the features are extracted and normalized; then, a multi-task prediction model based on deep belief networks (DBN) is constructed, and the part surface roughness and tool wear are taken as the output of the model, and the features are extracted as the input to establish the multi-task DBN prediction model; finally, the vibration signals are input into the multi-task prediction model to predict the surface roughness and tool wear.

Claims

1. A prediction method of part surface roughness and tool wear condition based on multi-task learning, comprising the following steps: collecting vibration signal during machining by fixing tri-axial acceleration sensor at a position of a spindle, wherein the tri-axial acceleration sensor closest to tool holder to collect a tri-vibration signal of the spindle during a machining process of a CNC machine tool, and the vibration signal of a machine tool load is intercepted from a tri-axial vibration signal; using measurement equipment to measure a part surface roughness and tool wear; when measuring the part surface roughness, dividing a surface of the part into equal intervals according to a sampling length, and measuring a roughness value in each interval, wherein whether surface quality of the part is qualified or not is divided according to the roughness; tool wear is measured every fixed cutting distance, and the tool is categorized into three conditions being normal, worn and damaged according to whether there is obvious wear and damage of the tool; finally, dividing vibration data collected in the machining process by a length of vibration data corresponding to each sampling length, wherein a surface roughness label and a tool wear condition label are respectively corresponding to the vibration data after division; according to equation (1), adding M degree of Gaussian white noise to the collected vibration signal; $\begin{array}{cc}{V}_i^s=\left(k^{2}i+k\right)\frac{\sigma }{M}{V}_{\mathrm{st}}^s& \left(1\right)\end{array}$ where: V.sub.st.sup.s is an original vibration signal, s is a sample index, k is a different intensity parameter, i=1, 2, . . . , M is used to control a different noise level, σ is a sample standard variance; after adding noise, each sample is expanded to M+1 group sample {V.sub.st.sup.s, V.sub.1.sup.s, . . . , V.sub.M.sup.s}, the sample expansion is as follows:
Sa=V.sub.st.sup.s+V.sub.i.sup.s  (2) normalizing the variance, root mean square, kurtosis, impulse factor and skewness features extracted from equation (3)˜(7); $\begin{array}{cc}\mathrm{Va}=\frac{1}{N-1}\underset{t=1}{\overset{N}{.Math.}}{\left(V_i^s\left(t\right)-\overline{V}\right)}^2& \left(3\right)\\ \mathrm{RMS}={\left(\frac{1}{N}\underset{t=1}{\overset{N}{.Math.}}{V_i^s\left(t\right)}^2\right)}^{1/2}& \left(4\right)\\ \mathrm{KUR}=\left[\frac{1}{N}\underset{t=1}{\overset{N}{.Math.}}{\left(V_i^s\left(t\right)-\overline{V}\right)}^4\right]/{\left[\frac{1}{N}\underset{t=1}{\overset{N}{.Math.}}{\left(V_i^s\left(t\right)-\overline{V}\right)}^2\right]}^2& \left(5\right)\\ \mathrm{IF}=\left[{V_i^s\left(t\right)}_{\max }-{V_i^s\left(t\right)}_{\min }\right]/\left[\frac{1}{N}\underset{t=1}{\overset{N}{.Math.}}.Math.V_i^s\left(t\right).Math.\right]& \left(6\right)\\ \mathrm{Ske}=\left[\frac{1}{N}\underset{t=1}{\overset{N}{.Math.}}{\left(V_i^s\left(t\right)-\overline{V}\right)}^3\right]/{\left[\frac{1}{N}\underset{t=1}{\overset{N}{.Math.}}{\left(V_i^s\left(t\right)-\overline{V}\right)}^2\right]}^{3/2}& \left(7\right)\end{array}$ where: V.sub.i.sup.s(t) is the vibration data of a sample, t is an index of vibration data, N is a number of data contained in each sample, and V is an average value of each sample; constructing a deep neural network A based on deep belief networks, wherein the deep neural network A is composed of three layers of restricted Boltzmann machine and a BP network; among them, the restricted Boltzmann machine is an energy-based model, and its joint probability distribution is specified by an energy function, and a definition of the energy function is shown in equation (8), $\begin{array}{cc}E\left(v,h\text{;}\theta \right)=-\underset{\alpha =1}{\overset{V}{.Math.}}\underset{\beta =1}{\overset{H}{.Math.}}{w}_{\alpha \beta }{v}_{\alpha }{h}_{\beta }-\left(\underset{\alpha =1}{\overset{V}{.Math.}}{b}_{\alpha }{v}_{\alpha }+\underset{\beta =1}{\overset{H}{.Math.}}{a}_{\beta }{h}_{\beta }\right)& \left(8\right)\end{array}$ where v.sub.α and h.sub.β are the binary states of visible and hidden units, θ={w,b,a} are the parameters of the model; w.sub.αβ are the weight between visible units α and hidden units β, b.sub.α and a.sub.β are offsets of visible units α and hidden units β, respectively, V and H is the number of visible and hidden units; the energy of hidden neurons is as follows: $\begin{array}{cc}{E}_{\beta }=\underset{\alpha =1}{\overset{V}{.Math.}}{w}_{\alpha \beta }{v}_{\alpha }+b_{\alpha }& \left(9\right)\end{array}$ similarly, the energy of visible neurons is as follows: $\begin{array}{cc}{E}_{\alpha }=\underset{\beta =1}{\overset{H}{.Math.}}{w}_{\alpha \beta }{h}_{\beta }+a_{\beta }& \left(10\right)\end{array}$ marking the input layer to the first layer restricted Boltzmann machine as B1, and from the second layer restricted Boltzmann machine to the output layer as B2, copying B2 and marking it as B3, wherein B3 is connected with the first layer restricted Boltzmann machine of deep neural network A, at this time, arranging the network structure B2 and B3 side by side and connecting with B1; according to the task requirements, setting the number of output layer nodes of B2 and B3 to complete the improvement of deep belief network, wherein the improved deep belief network B is composed of network structure B1, B2 and B3, which can realize the function of multi-task learning; using a greedy algorithm to pre-train the restricted Boltzmann machine of each layer in the deep neural network A, and assigning the weight of the trained deep neural network A to the improved deep belief networks B, wherein based on the back-propagation algorithm, the supervised fine-tuning of the improved deep belief network B is carried out, and the loss function as equation (11) is adjusted to the optimal value to obtain multi-task predicting model; $\begin{array}{cc}L\left(\theta \right)=-\underset{j=1}{\overset{W}{.Math.}}\frac{{\lambda }_j}{m}\left[\underset{k=1}{\overset{m}{.Math.}}{\hat{y}}_k\log y_k\right]& \left(11\right)\end{array}$ where, W is the total number of tasks in the multi-task model, λ.sub.j is the weight of each task, m is the total number of fine-tuning samples, ŷ.sub.k is the actual label, y.sub.k is the prediction label; in the actual machining process, pre-processing the vibration data collected and inputting the pre-processed vibration data into the multi-task prediction model after training; the multi-task prediction model outputs the part surface roughness and tool wear condition; as task 1, outputting 1 when the part surface roughness is qualified, and outputting 0 when the part surface roughness is not qualified; and as task 2, outputting 0 when the tool is not worn, outputting 1 when it is worn, and outputting 2 when it is damaged.

Description

DESCRIPTION OF THE DRAWINGS

(1) FIG. 1: Training flow chart of multi-task DBN prediction model.

(2) FIG. 2: Network structure of multi-task DBN prediction model.

(3) FIG. 3: Confusion matrix of tool wear prediction accuracy.

(4) FIG. 4: Confusion matrix of prediction accuracy of part roughness.

DETAILED DESCRIPTION

(5) In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be described in detail below with reference to the accompanying drawings.

(6) Cutting experiments were carried out on a 3-axis vertical machining center with a vertical milling cutter. Among them, the basic information of the 3-axis vertical machining center is: the maximum stroke of X axis, Y axis and Z axis is 710 mm, 500 mm and 350 mm, the maximum feed speed is 32 m/min, 32 m/min and 30 m/min, and the maximum spindle speed is 15000 r/min. The basic information of the tool is: the type of the tool is vertical milling cutter; the tool material is cemented carbide; the diameter of the tool is 10 mm; and the number of cutting edges is 4. The basic information of the workpiece to be cut is as follows: the workpiece material is 45# steel; the workpiece shape is 200 mm×100 mm×10 mm. The cutting process parameters are as follows: the cutting depth is 2 mm; the feed rate is 80 mm/min; the spindle speed is 6000 r/min.

(7) The multi-task model training process is shown in FIG. 1, and the specific implementation mode is as follows:

(8) Step One: Collect Vibration Signals During Machining

(9) The cutting experiments is carried out in the 3-axis vertical machining center. The tri-axial acceleration sensor is installed at the position of the spindle near the toolholder, and the vibration signals of the spindle in X, Y and Z directions are collected and saved at the sampling frequency of 1000 Hz.

(10) Step Two: Measure the Part Surface Roughness and Tool Wear

(11) The part surface roughness and tool wear are measured by special measuring equipment. When measuring the surface roughness, the surface roughness of the part is divided into equal intervals according to the sampling length (4 mm), and a roughness value Ra is measured in each interval. According to the roughness size, the surface quality of the parts is divided into qualified or not with the threshold value of 0.8 μm. The tool wear condition is detected every 100 mm cutting distance, and the tool is divided into normal, worn and damaged condition according to whether the tool has obvious wear and damage. Finally, taking the vibration data length corresponding to each sampling length as the standard, 119 vibration data segments were obtained by dividing the collected vibration data into equal intervals, and 119 sets of sample datasets were obtained by matching the surface roughness label and tool wear condition label with the vibration data after division.

(12) Step Three: Sample Expansion and Feature Extraction

(13) According to equation (1), four different degrees of white Gaussian noise are added to the segmented vibration signal. After expansion, the total number of samples was 595 groups. Then, according to equation (3)˜(7), the extracted features include kurtosis, margin factor, root mean square, variance and skewness. Since the vibration signal is divided into three directions, each sample contains 15 features.

(14) Step Four: The Construction and Training of Multi-Task Prediction Model Based on Improved DBN

(15) A deep neural network A is constructed based on deep belief network. The network consists of three layers of restricted Boltzmann machine and a BP network. The number of neuron nodes in the input layer is 15, the number of neuron nodes in each layer is 200, 80 and 60, and the output layer is 2. Copy the restricted Boltzmann machine from the second layer to the output layer, connect the network B3 with the restricted Boltzmann machine in the second layer, and set the number of nodes in B3 output layer to 3 according to the requirements of task 2, that is, to complete the improvement of the deep belief network. The network structure of multi-task prediction model B based on improved DBN is 15-200-80 (80)-60 (60)-2 (3), which can realize the function of multi-task prediction. The data set is randomly divided into training set test set according to the ratio of 4:1. Firstly, the weight of neural network a is adjusted unsupervised by using the training set data, and the network weight is initially determined. At this time, the learning rate is 0.05 and the number of iterations is 1000. Next, the weight of the trained neural network A is assigned to the prediction model B as the initial parameter of the supervised fine-tuning multi task prediction model B Value. Then, the multi-task prediction model B is trained alternately by using the surface roughness label, tool wear condition label and the corresponding vibration signal characteristics. The weight of the prediction model is adjusted to the minimum and the loss function (equation 8) is adjusted to the minimum. Finally, the multi-task prediction model based on the improved DBN is obtained. Then test the multi-task model with test dataset. The results show that the accuracy of the proposed multi task prediction model for tool wear prediction is 99%, and the prediction accuracy of surface roughness of parts is 93%. The multi-task prediction model can be used to predict tool wear and part surface roughness.

(16) Step Five: Predict the Surface Roughness and Tool Wear Condition.

(17) The real-time vibration collected in the actual machining process is input into the multi-task prediction model based on improved DBN after data preprocessing, and the corresponding surface roughness and tool wear condition are obtained.