AN EFFECTIVE IMAGE BASED SURFACE ... - IEEE Xplore

0 downloads 0 Views 4MB Size Report
ALI AKBAR AKBARI, DEPARTMENT OF MECHANICAL ENGINEERING ... AMIR GOODARZVAND CHEGINI, DEPARTMENT OF MECHANICAL ENGINEERING.
Copyright - World Automation Congress (WAC) 2006, July 24-26, Budapest, Hungary

AN EFFECTIVE IMAGE BASED SURFACE ROUGHNESS ESTIMATION APPROACH USING NEURAL NETWORK ALI AKBAR AKBARI, DEPARTMENT OF MECHANICAL ENGINEERING FERDOWSI UNIVERSITY, MASHHAD, IRAN, [email protected] AMIN MILANI FARD, DEPARTMENT OF COMPUTER ENGINEERING FERDOWSI UNIVERSITY, MASHHAD, IRAN, [email protected] AMIR GOODARZVAND CHEGINI, DEPARTMENT OF MECHANICAL ENGINEERING SHARIF UNIVERSITY OF TECHNOLOGY, TEHRAN, IRAN, [email protected]

ABSTRACT

The accurate measurement of surface roughness is essential in ensuring the desired quality of machined parts. The most common method of measuring the surface roughness of machined parts is using a surface profile-meter with a contact stylus, which can provide direct measurements of surface profiles. This method has its own disadvantageous such as workpiece surface damage due to mechanical contact between the stylus and the surface. In this paper we proposed a contactless method using image processing and artificial neural network as a pattern classifier. Having trained the network for any specific workpiece with 10 sample patterns, the system would learn how to approximate the actual surface roughness with 3D texture features of the surface image. The input parameters of a training model are RaArea and RqArea, defined parameters for gray level of surface image, arithmetic mean value, and standard deviation of gray levels from the surface image, without involving cutting parameters (cutting speed, feed rate, and depth of cut). Experimental results show effectiveness of this estimation method.

KEYWORDS: Image Processing, ANN, Non-Destructive Surface Roughness Measurement 1. INTRODUCTION

In manufacturing engineering, the surface quality of a machined component is fundamental importance to ensure its integrity. In recent years, as the demand for higher quality has increased, surface roughness monitoring has become increasingly important in ensuring the quality of parts and products. The most common method of measuring the surface roughness of machined parts is applying a surface profile-meter with a contact stylus, which can provide direct measurements of surface profiles. It traverses across the test surface and traces its micro-profile. Although this technique has been developed to a very sophisticated level, the stylus instrument has a number of

drawbacks and limitations including the followings: [1, 2] * On surfaces with deep valleys, the stylus measuring probe due to its tip size may not be able to penetrate fully to the bottom. * When a spherical stylus passes over a sharp peak, the point of contact moves across the stylus, from one side to another. This causes the stylus to follow a path that is more rounded than the peak. * Surface of the workpiece may be damaged during the mechanical contact between the stylus and the surface. This is very important in the case of mirror-finished surfaces. * Surface scratches due to high contact pressure and the wear of the stylus tip, long measuring times, and the difficulty of on-line measurement [3]. The transducer and stylus tips are often fragile, hence the instrument must be employed in a fairly vibration free environment. For these reasons, various other potential techniques for inprocess measurement have been studied. Optical techniques which are contactless and non-

destructive appears to be suitable alternative for carrying out measurement of surface quality including surface roughness. [4]

2. RELATED WORKS

A remarkable number of researches have been done to measure the surface roughness using optical methods [4]. Quantitative analysis of surface topology using electron microscopy, development of a surface roughness measurement system using reflected laser beam, that a SEMcomputer system could be considered as an adequate instrument of surface analysis with capabilities both for morphology and quantitative roughness measurements, as well as surface roughness evaluation [5]. On the other hand, an optical method relying on the reflected beam peak power intensity variation may not be reliable for fine surfaces, since this method uses diffusive reflection, which plays an important role only for considerably rough surfaces. Optical methods for roughness measurement include laser profilometry, which in general utilizes expensive and large equipment and poses difficulties in the implementation. Other optical methods for surface roughness measurement include astigmatic focus, laser speckle and light scattering techniques. The laser speckle pattern method is based on Fraunhoffer diffraction model in which a collimated light is directed on a test surface and the resulting image at the back focal plane of the measuring lens which constitutes the power spectral density (PSD) of the test surface is recorded. This method, however, can be used only for isotropic random surfaces with Rq (root mean square of surface profile) less than the laser wavelength and is time consuming. In particular, the test surfaces should be statistically isotropic and homogeneous [1]. Hence, these methods are incapable of measuring surface roughness where specimens have preferred orientations resulting from machining processes such as turning, grinding or planning.

3. SURFACE ROUGHNESS PARAMETERS

Surface roughness parameters in ASME and ISO standards have been defined as follows: Ra: Average Roughness, Also known as Arithmetic Average (AA), Center Line Average (CLA), Arithmetical Mean Deviation of the Profile. The average roughness is the area between the roughness profile and its mean line, or the integral of the absolute value of the roughness profile height over the evaluation length Ra IJ-L r(x) dx .R . R, does not tell the whole story about a osnttl h hl tr bu LO0

surface. If we want to distinguish between surfaces that differ in shape or spacing, we need to calculate other parameters for a surface that measure peaks and valleys and profile shape and spacing. Rq: Root Mean Square Roughness, The RMS average roughness of a surface is calculated from another integral of the roughness profile: Rq = f r2(x)dx . For a pure sine wave of any LO0

wavelength and amplitude Rq is proportional to Ra; it's about 1.1 1 times larger. Older instruments made use of this approximation by calculating Rq with analog electronics (which is easier than calculating with analog electronics) and then multiplying by 1.1 1 to report Rq [6]. However, real profiles are not simple sine waves, and the approximation often fails miserably. R.: is the maximum peak or lowest valley vertical distance within a single sample length. Rz: (ISO) is the sum of the height of the highest peak plus the lowest valley depth within a sampling length. S111: is the mean spacing between peaks, now with a peak defined relative to the mean line.

4. IMAGE PROCESSING BASED METHOD

If the variation of the gray level of an image along a line be into a pure sine wave with amplitude 1 and wavelength 27c unit, and the surface roughness parameters has been defined along this line, the parameters would be obtained as table 1.

Sm RV 1.0000 1.0000 6.2832 Table 1 Surface roughness parameters for a sin wave Ra (unit) 0.6366

RZ Din

R

0.7071

The result of the above table shows that they are coinciding mathematically and structurally to stylus result. In the next stage we evaluated the image of machined surface. Incipience estimate the variation of the gray level along the perpendicular against lay of surface. The image of surface (SEM result) is shown in Figure 1. Dimension of picture is 1000 X 1000 pixels with magnification of 1200 times, whereas the real size is 83.33 ptm by 83.33 rim. 3 2 4 Evaluation of the variation of the gray level has been done along the perpendicular against lay of surface. We obtained the vertical angle by a simple image processing algorithm finding the maximum amount [7]. Figure 2 shows the difference between gray level and average of gray level to along a line of pixel of image. As it 1 shows from 0-100 to 900-1000, the average gray level decreased, that depend to slope of sample and it will be obviation by Image Processing algorithms and mapping them to specific range. In this stage we can calculate surface roughness parameters from their gray level texture on formula 23 base. For example, the texture of Figure 2 of these parameters Figure 1. Evaluated image for is shown in Table 2 and by comparing these data with the surface roughness estimation and result of measurement by stylus instrument, we can define variation of estimation angle to some experimental coefficients as in Table 2. achieve perpendicular angle 200

-200

200

200

0

20

I

40

60

80

140

160

180

100

1

i /-200500 100

520

540

560

580

600

-100 600

620

640

660

680

700

720

740

760

780

800

980

1000

0b Wey 7VJ-200 100 00

120

0

100 0

h~~~~~~~~~C

0 0

200

-1

_

io200

_

0220

_

_

240

'V _~ ~ ~ FE 200

Y 0_ ),, tA nj #af

_

_

_

260

_

280

-

300

Jry abV( A

*

-10

700 0

n~~Fgr 2. DIeec betee gra lee an avrg -100 400

420

440

460

Pixel along A Line

480

tA

~~~~~~~~~100*f i

/

500

-100900

A hbi -

raeelogaln 920

940

960

Pil ag A in

RQ

Ragray 26.39

34

RR

RaStylus

Ry

R 106

106

2.78 pm

3.84 pm

20 pm

R 28.6 pm

Racoef. 0.1053

R, coef.

R, coef.

R. coef.

0.113

0.188

0.269

Sm 80

T

Lc pm 117.85 tm =1000pixel

Lc

Sm

76 pm

800 pm

Smcoef.

Lccoef.

0.95

Surface roughness parameters from their gray level texture

Table 2

6.788

5. THE PROPOSED SYSTEM Evaluation along a line of the surface performed by Lee et al [8, 9] using an adaptive neurofuzzy inference system to accurate estimation surface roughness from texture surface image. However here we evaluated the entire surface area for estimation surface roughness parameters using a multi layer perceptron (MLP) artificial neural network, which is described in the next section. To achieve more accurate coefficient for our gray level approach to be close to stylus parameters, we split surface into four partitions and calculate relative error. We proposed sample area (Ac) instead of sample length (Lc). In this case we define all parameters on a surface, like surface roughness parameters along a line. For example the parameters of the first partition are shown in the Table 3.

Size(p) 250 260 270 280 290 300 310 320 330 340 350 360 370

Rqgrat Rygray

Ragray

21.938 22.100 22.272 22.408 22.539 22.676 22.788 22.962 23.149 23.289 23.411 23.482 23.556

30.923 31.073 31.227 31.327 31.408 31.495 31.565 31.693 31.829 31.945 32.036 32.090 32.182

113.795 113.567 113.320 113.048 112.763 112.468 112.265 112.130 111.942 111.783 111.717 111.650 111.545

Table 3

RzDing

113.795 113.567 113.320 113.048 112.763 112.468 112.265 112.130 111.942 111.783 111.717 111.650 111.545

Size(p) 380 390 400 410 420 430 440 450 460 470 480 490 500

Ragray

23.668 23.696 23.701 23.720 23.736 23.731 23.715 23.702 23.695 23.679 23.645 23.648 23.672

Rqgrat Rygray

32.336 32.355 32.357 32.383 32.400 32.377 32.328 32.286 32.263 32.223 32.168 32.149 32.150

The parameters of first partition of image

111.444 111.455 111.437 111.371 111.289 111.245 111.181 111.083 111.009 110.972 110.960 110.935 110.927

RzDingray 111.444 111.455 111.437 111.371 111.289 111.245 111.181 111.083 111.009 110.972 110.960 110.935 110.927

In this case, we define surface roughness parameters such as RaArea and RqArea on the entire surface as below:

RaArea

RqA rea

1

X*y

g*(i x y i: 21 x

1 X*Y 'X *Y i-l x *y

i=4

,y)

(gi (x,

(1) 2

(2)

As we increase the surface evaluation area, the variation of parameters will be decrease. A surface roughness measurement technique, based on area measurement method using a computer vision system, was investigated for in-process inspection of surface quality during a machining process. This method was first employed by Devoe et al [5]. Their proposed roughness measurement system applies the light-scattering theory introduced by Beckman and Spinoza

(1963). The vision system uses a monochrome CCD camera to provide a gray-scale image based on the pattern of light scattered from an area of the machined piece. This gray-scale image is sent to image manipulation software for analysis. The photo-optical measurement method is shown in Figure 9. 1- CCD Camera

2- Micro Computer 3- Feedback to Machining Process 4- Tool 5- Sample 6- Machined Surface

3

Figure 9. The fundamentalprinciple ofthe surface roughness measurement technique In our proposed work, a program was developed using Borland Delphi 6 and MATLAB image processing toolbox. The software provides a user friendly form in order to be used by ordinary workers. Our best result obtained when we set the CCD camera magnification to 50 times. The software is then examines the light scattering pattern in the image, and calculates the roughness parameter of the surface from the image's gray-level histogram. The Ra value for the surface is determined through the use of a correlation curve which uniquely relates a range of the roughness parameter into a range of R values. The resulting Ra value is either displayed on a video monitor for observation or used as feedback to the machining process depending on the system application environment. A relationship was developed between known roughness values and the optical roughness parameter, and this relationship was used to test the repeatability of the optical method by comparing roughness values derived using the optical technique and a stylus device. They were written several macros to determine three image parameters: standard deviation (SD), arithmetical average deviation (AAD), arid RMS values based on the brightness level histogram over a 12.5 mm2 region of each image. These parameters are defined as follows:

1 255 -2 F.I SD= I/OFi(Xi x)

n-1i=0 1 255 2 RMS= | E F. n i =0 in AAD= L Ixi -xl n i-O

(4) (5)

Parameter n is the number of image pixels, x shows gray level of brightness and F is count of pixels in the image at brightness level x. In the next step the results of image processing and MLP neural network were used to classify the obtained surface profile and predict the surface roughness value. The network has 4 inputs: RaArea, RqArea, arithmetic mean value, and standard deviation of gray levels from the surface image, without involving cutting parameters (cutting speed, feed rate, and depth of cut). Determining the number of hidden layer elements has been done experimentally and 3 nodes found to be appropriate for this classification problem. The network is, then, trained and tested to recognize the input vectors with 10 sample pafterns. Table 4 demonstrates the stylus results of machined surface and image processing for four samples. Table 5 shows the input parameters from the image and output parameters from the network.

Raktm)

RQ(tm)

R,Din R(Mtm)

Ra(gary)

R,(gary)

R,(gary) Ry(gary)

Sample 1 2.78 3.84 20 28.6

23.98 31.71 108.64 111.

Sample 2 1.7 2.2 11.5 14.4

15.1 18.24 58.6 51.1

Sample 3

Sample

1.6 2.18 11.2 18.5

4 2.2 3.1 17 21

14 17.1 55.4 65.1

19.93 25.76 83.6 102.1

Table 4: The stylus results of machined surface and image processing for four samples

__

_

Ra(gary)

Rj(gary) R,(gary)

R4gary)

Raktm)

RQ(tm) RzDin

Ry(tm)

Sample A

Sample B

Sample C

Sample

20.7 26.3 107 109.5

D 32.5 40.6 118 121.2

1.96 2.47 17.65 24.34

1.39 1.76 13.45 23.91

2.36 3.19 20.15 25.73

3.74 4.85 22.04 27.93

17.3 20.1 92.4 102.1

12.5 14 68 99.8

Table 5: the input parameters from the image and output parameters from the network

6. CONCLUSIONS

In this paper an image based surface roughness estimation method was introduced using artificial neural network to accurately establish the relationship between actual surface roughness and 3D texture features of the surface image, and consequently effectively estimate surface roughness. The input parameters used for a training model are, RaArea, RqArea defined parameter for gray level of surface image, arithmetic mean value, and standard deviation of gray levels from the surface image, without involving cutting parameters (cutting speed, feed rate, and depth of cut). Experimental results show effectiveness of this estimation method.

7. ACKNOWLEDGEMENTS

Authors would like to thanks Khorasan Axial Parts Co. MFG and IranKhodro - SAPCO for their modest cooperation in providing axial parts microscopic images.

8. REFERENCES

[1] C.J. Tay, S.H. Wang, C.Quan, H.M. Shang, "In situ surface roughness measurement using a laser scattering method" Int. J. Optics Communications, Vol. 218 (2003) 1-10. [2] Z. Yilbas, M.S.J. Hashmi, "An Optical Method and Neural Network for Surface Roughness Measurement" Optics and ,Lasers in Engineering 29 (1998) 1-15. [3] H.Y. Kim, Y.F. Shen, J.H. Ahn, "Development of a surface roughness measurement system using reflected laser beam" J. Of Materials Processing Technology 130-131 (2002) 662667. [4] Z. Yilbas, M.S.J. Hasmi, "Surface roughness measurement using an optical system", Journal of Materials Processing Technology 88 (1999) 10-22. [5] D.DeVoe, L.Knox, G.Zhang, "An Experimental Study of Surface Roughness Assessment Using Image Processing", Technical Reaserch Report, University of Maryland, (1992) NSFD CD 8803012. [6] A.J. Baker, W.J. Giardini. "Developments in Australia's surface roughness measurement system" J.Of Machine Tools & Manufacture 41 (2001) 2087-2093. [7] R.C. Gonzalez, R.E. Woods, S.L. Eddins: Digital Image Processing Using MATLAB, Pearson Prentice-Hall, 2004, ISBN: 0-13-008519-7. [8] K.C. Lee, S.J. Ho, S.Y. Ho, "Accurate estimation of surface roughness from texture features of the surface image using an adaptive neuro-fuzzy inference system" Precision Engineering 29 (2004) 1-15. [9] 5. Damodarasamy, S. Raman. "Texture analysis using computer vision" Computer Industry, 1991 ;16:25-34.