Comparison of fuzzy expert system based strategies of offline and online estimation of flank wear in hard milling process

Asif Iqbala, , , Ning Hea, Naeem Ullah Darb and Liang Lia

aCollege of Mechanical and Electrical Engineering, Nanjing University of Aeronautics and Astronautics, 29, Yu Dao Street, Nanjing, Jiangsu 210016, PR China
bDepartment of Mechanical Engineering, University of Engineering and Technology, Taxila, Pakistan


Available online 2 May 2006.

 


Abstract
Accurate estimation of flank wear during any in-progress machining process is highly important for the purpose of controlling product quality and the production rate. Hard milling is among few of the recently popularized technologies of metal cutting domain and is found under intense research for the purpose of estimation and control of tool wear. In the presented paper two fuzzy rules based strategies are explained and compared for accurate estimation of tool’s flank wear in hard milling process. The offline strategy uses length of cut (LoC) as major input besides tool helix angle and workpiece material hardness, while for the online strategy LoC is replaced with the cutting force. Series of hard milling experiments were performed in order to obtain data for the development of two fuzzy expert systems as well as for testing of both of the strategies. ANOVA showed LoC and cutting force were more significant than other input parameters for the estimation of flank wear, and the design of fuzzy sets for input parameters was based upon this analysis. Two expert systems were tested using experimental data and the results showed that online strategy was 67.9% more accurate than offline one in estimating flank wear.

Keywords: Wear estimation; Hard milling; Fuzzy sets; Cutting force; Expert system
 

Article Outline
1. Introduction
2. Experimental work
2.1. Experimental results
3. The fuzzy expert system
3.1. Fuzzy rules
4. Testing and comparison
5. Conclusions
References

 
1. Introduction
Hard milling is the name given to the process of milling—employing the cutting speeds in excess of 150 m/min—of steels hardened to the values of more than 40HRc. Hard milling possesses numerous advantages over conventional milling process, like reduction of cutting forces, removal of process heat along with chips, improved dimensional accuracy, better surface quality, compressive residual stresses in the workpiece, minimal alterations in micro-hardness of workpiece, and machining in the range not subject to critical vibrations (Axinte and Dewes, 2001 and Schulz, 2004). Unfortunately, the exceedingly high amount of heat generated in hard milling process, as compared to that in conventional milling, makes the cutting tool vulnerable to rapid tool wear (Zeren & Ozel, 2002). Flank wear is one of three forms of tool wear that occurs when freshly cut workpiece slides pass the flank face of the tool. Increase in flank wear reduces the surface quality of machined workpiece significantly; besides, it also causes the increase in power utilization of the machine tool. So it becomes necessary to replace the tool with a new one as the flank wear reaches a specific value. If it is not done so, increased flank wear may cause fracture of tool that will lead to loss in productivity besides possible loss of workpiece itself. So, it is important to know the values of tool’s flank wear at any stage of the process.

It is almost impossible to measure flank wear when milling process is in progress. The only remedy to this problem is to have estimations of flank wear using different possible strategies. Quite a many research papers can be found explaining different techniques for estimation of tool wear for different forms of machining processes. Fang (1995) developed knowledge-intensive fuzzy feature-state relationship matrices for the diagnosis of tool wear state in finish turning process, using force signals and tool vibrations as input parameters. Obikawa, Kaseda, Matsumura, Gong, and Shirakashi (1996) made use of artificial neural network to estimate tool wear states, onset of chatter, and chip tangling by using force signals as input parameters. Morikawi and Mori (1993) used neural network based multi-sensor integration technique to determine tool wear states in turning process. Du et al. (2001) interpreted input parameters: cutting speed, feed, width of cut, machining time, spindle motor current, and feed motor current, using transition fuzzy probability to estimate wear conditions of tool in a boring process. Luo, Osypiw, and Irle (2001) made use of Gaussian Wavelet Algorithm to transform the spindle vibration signals for estimation of tool wear states in machining of medium density fiber-board. Aforementioned papers are few of those explaining utilization of artificial intelligence in the estimation of tool wear. From literature survey it can be found that most of the research has targeted the turning process and the limited amount of research has been carried in estimation of tool wear related to the milling process.

In the following paper, two strategies, for estimation of tool’s flank wear at different stages of in-progress hard milling process, are presented and compared for accuracy. The offline strategy involves length of cut (LoC) as input parameter besides tool’s helix angle and workpiece material hardness. The online strategy replaces LoC with the Fxy (resultant of peak values of two components of cutting force acting in x- and y-directions). The online strategy has been named so because the Fxy values can be known only when the process is in progress and tool is in contact with the workpiece. The peak values of two components of force signals (Fx and Fy) can be measured and fed into a computer in order to calculate instantaneous values of Fxy, and thus decision can be taken online by following a certain rule base. Both the strategies make use of expert systems, each of them utilizing fuzzy logic as reasoning mechanism. The design of fuzzy sets and combination of rules is based upon experimental data obtained. The detail has been provided in the upcoming sections.

2. Experimental work
Table 1 shows the design of experiments for the purpose of development of rule bases. Two levels for helix angle, 3 levels for workpiece material hardness, and 6 levels each for LoC and Fxy were tested for maximum width of flank wear land (VB). This technique gives 36 (= 2 × 3 × 6) data points for each strategy.

 Table 1.

Levels of input parameters used in experiments Level Helix angle (°) Hardness (HRc) LoC (m) Fxy (N)
1 30 52 1 400
2 50 57 2.5 500
3 – 62 5 700
4 – – 7.5 900
5 – – 10 1100
6 – – 12.5 1300

 


All the experiments were performed on Micron UCP 710, 5-axis, vertical milling center having maximum power of 16 kW. The workpiece material used was AISI D2 having dimensions 100 mm × 33 mm × 33 mm, and hardened to 3 different values as shown in Table 1. The cutting tools used were flat end solid K30 carbide cutters with PVD coated mono layer of TiAlN, having diameter (D) of 10 mm, corner radius (R) of 1.5 mm, rake angle (γ) of 5°, flank angle (α) of 6° (primary) and 10° (secondary), and number of flutes (z) equal to 4. The measurement of forces were performed using Kistler piezoelectric dynamometer 9265B, utilizing force plate 9443B, having measuring range of 0–15 kN in x- and y-directions and range of 0–30 kN in z-direction. The dynamometer was connected to four channel Gould Classic oscilloscope using charge amplifiers. Flank wear was measured using 10× tool maker’s microscope.

In the entire hard end milling experiments fixed cutting parameters were used: cutting speed (Vc) = 250 m/min (7957.7 rpm); feed rate (fz) = 0.1 mm/tooth (Vf = 3.183 m/min); radial depth of cut (ae) = 0.4 mm; and axial depth of cut (ap) = 5 mm. Down-milling was employed as milling orientation and no coolant was used. Milling was performed in straight line with length of cut for a single pass equal to 100 mm.

For the purpose of testing of expert systems of both of strategies, further hard end milling experiments were done with different values of input parameters and resulting VB values were measured. Detail has been provided in upcoming sections.

2.1. Experimental results
Experimental results are shown in graphical form in Fig. 1, Fig. 2, Fig. 3 and Fig. 4. The maximum width of flank (VB) wear seems to be more or less in direct relationship with LoC as well as with Fxy. It can also be observed from the figures that the workpiece material of higher hardness causes rapid increase in VB of tool as the process progresses. These obtained sets of data were analyzed used ANOVA (analysis of variance) technique. ANOVA revealed that effect of LoC upon VB is 2.8 and 1.45 times more significant as that of helix angle and workpiece material hardness, respectively. For the other set of data, ANOVA showed that effect of Fxy upon VB was 8.7 and 4.4 times more significant than that of helix angle and workpiece material hardness, respectively. Fuzzy sets for input parameters were designed based upon these experimental results.

 
 (24K)

Fig. 1. Progress of maximum width of flank wear land along length of cut for 30° helix angle tool.

 


 (23K)
Fig. 2. Progress of maximum width of flank wear land along length of cut for 50° helix angle tool.

 


 (22K)
Fig. 3. Maximum width of flank wear land versus Fxy for 30° helix angle tool.

 


 (22K)
Fig. 4. Maximum width of flank wear land versus Fxy for 50° helix angle tool.

 

3. The fuzzy expert system
After obtaining the data from experiments, the next step is to divide the experimented range of input and output variables into fuzzy sets. Fuzzy sets and logic is a discipline that has proved itself successful in automated reasoning of expert systems (Konar, 2000). It deals with the theory of vague reasoning in order to model human-like reasoning problems of real life. In recent past, fuzzy logic has found high degree of applicability in development of expert systems and the same has been selected as the reasoning mechanism in development of presented rule bases.

In this research work, the experimented range of more significant variables is divided into more number of fuzzy sets as compared to that of lesser significant variable. Ten fuzzy sets were designed for the output variable (VB) applicable to both of the strategies. For all of the variables, equally distributed triangular shaped fuzzy sets were utilized. Fig. 5, Fig. 6, Fig. 7 and Fig. 8 show the detail, while Table 2 and Table 3 provide the abbreviation details of these fuzzy sets.

 
 (18K)

Fig. 5. Fuzzy sets for helix angle and workpiece material hardness.

 


 (16K)
Fig. 6. Fuzzy sets for length of cut.

 


 (18K)
Fig. 7. Fuzzy sets for Fxy.

 


 (27K)
Fig. 8. Fuzzy sets for maximum width of flank wear land.

 

Table 2.
Expressions used in fuzzy sets of helix angle, hardness, and length of cut Helix angle abbreviation Expression Hardness abbreviation Expression Length of cut abbreviation Expression
S Small LH Less hard VSH Very short
L Large MH Medium hard SH Short
  HH Highly hard AV Average
    LN Long
    VLN Very long

 


Table 3.
Expression used in fuzzy sets of Fxy and VB Fxy abbreviation Expression VB abbreviation Expression
VS Very small VEL Very extremely low
SM Small EL Extremely low
ME Medium VL Very low
LA Large LO Low
VL Very large SL Slightly low
  SH Slightly high
  HI High
  VH Very high
  EH Extremely high
  VEH Very extremely high

 


3.1. Fuzzy rules
The relationship between inputs and output in a fuzzy system is characterized by set of linguistic statements which are called fuzzy rules (Hashmi, Graham, & Mills, 2003). The number of fuzzy rules in a fuzzy system is related to the number of fuzzy sets for each input variable. In this research work, there are 30 (=2 × 3 × 5) possible rules for each of two strategies. An important question arises here, “which VB fuzzy sets to be assigned to 30 possible combinations of input fuzzy sets, for each of two strategies”? For a simple 2-inputs 1-output fuzzy model, the designer has to select the most optimum set of fuzzy rules from more than 10,000 combinations (Wong & Hamouda, 2002). For each strategy in this study, there are 30 fuzzy rules with 10 possibilities each. Thus the total number of possible fuzzy rules combination will be 1030.

For the most optimal possible combination of rules, simulated annealing algorithm was utilized and the best possible combination of rules for both strategies—giving minimum values of estimation error—are provided in Table 4. The max–min inference was employed for aggregation of the rules. The detail of this methodology can be read from Cox (1994).

 Table 4.

Optimal combination of fuzzy rules for offline and online strategies Rule No. Helix angle Hardness Offline strategy Online strategy
   LoC VB (output) Fxy VB (output)
1, 31 S LH VSH VEL VS EL
2, 32 S LH SH VL SM SH
3, 33 S LH AV SH ME VH
4, 34 S LH LN VH LA EH
5, 35 S LH VLN VEH VL VEH
6, 36 S MH VSH EL VS EL
7, 37 S MH SH SH SM SL
8, 38 S MH AV VH ME HI
9, 39 S MH LN EH LA EH
10, 40 S MH VLN VEH VL VEH
11, 41 S HH VSH EL VS VEL
12, 42 S HH SH SH SM SL
13, 43 S HH AV VEH ME SH
14, 44 S HH LN VEH LA VH
15, 45 S HH VLN VEH VL VEH
16, 46 L LH VSH VEL VS EL
17, 47 L LH SH LO SM LO
18, 48 L LH AV SL ME SH
19, 49 L LH LN SH LA EH
20, 50 L LH VLN EH VL VEH
21, 51 L MH VSH EL VS VEL
22, 52 L MH SH SL SM SL
23, 53 L MH AV HI ME HI
24, 54 L MH LN VH LA EH
25, 55 L MH VLN EH VL VEH
26, 56 L HH VSH VL VS VEL
27, 57 L HH SH SH SM LO
28, 58 L HH AV VH ME SH
29, 59 L HH LN EH LA VH
30, 60 L HH VLN VEH VL VEH

 


4. Testing and comparison
For the purpose of testing and comparison of expert systems related to offline and online strategies, further hard milling experiments were done upon workpieces having hardness values of 54 and 60HRc, using milling tools of 40° and 45° helix angles. The VB values were measured at different levels of LoC and Fxy, as shown in Table 5. For each and every combination of input parameters, the values of VB were estimated from the expert systems and compared with the actual VB values obtained from the experiments. Table 5 shows the detail. The term estimation error can be defined as follows:

 (1)


In above equation l, m, and n stand for number of levels of input variables used. For our case of testing data: l = 2, m = 2, and n = 5, thus giving total of 20 data points for each strategy. Table 5 shows that estimation capability of both of strategies is very good but still the online strategy outperforms the offline one, as the former gives 67.9% better estimation of VB as compared to later one.
 Table 5.

Comparison of offline and online strategies S/No. Helix Hardness Offline strategy Online strategy
   LoC VB (actual) VB (expert) Fxy VB (actual) VB (expert)
1 45 54 1 0.041 0.043 403 0.048 0.046
2 45 54 3 0.091 0.09 588 0.085 0.0804
3 45 54 5 0.123 0.119 806 0.128 0.1253
4 45 54 7 0.148 0.1434 899 0.147 0.1443
5 45 54 10 0.178 0.1674 1100 0.187 0.19
6 45 60 1 0.069 0.0617 426 0.039 0.041
7 45 60 3 0.119 0.1035 600 0.074 0.076
8 45 60 5 0.151 0.1499 802 0.114 0.1188
9 45 60 7 0.176 0.173 896 0.133 0.137
10 45 60 10 0.206 0.187 1107 0.175 0.1788
11 40 54 1 0.045 0.439 399 0.049 0.046
12 40 54 3 0.095 0.0902 603 0.09 0.0873
13 40 54 5 0.127 0.12 806 0.13 0.131
14 40 54 7 0.152 0.147 897 0.149 0.149
15 40 54 10 0.182 0.1695 1099 0.189 0.1891
16 40 60 1 0.073 0.06 420 0.04 0.044
17 40 60 3 0.123 0.1027 604 0.076 0.0769
18 40 60 5 0.155 0.1551 801 0.116 0.118
19 40 60 7 0.18 0.1762 895 0.135 0.1368
20 40 60 10 0.21 0.187 1106 0.177 0.179
  Estimation error (offline) 0.007966 Estimation error (online) 0.00256

 


The reason for the better performance of online strategy is that it deals with real time input data, which possesses better information of condition of in-progress process. On the other hand, offline strategy makes use of length of cut as primary input, which has not the ability to account for real process uncertainties and anomalies. For instance, if the tool, unluckily, undergoes some rapid wear due to possible local hardening of workpiece or due to tool unbalance, such abnormal increase in wear would be reported by increase in cutting force, while, on the other hand, the input parameter LoC remains indifferent to this accelerated amount of wear. This proves that fuzzy expert system technique incorporating cutting forces as major input can successfully diagnose tool wear states.

5. Conclusions
This paper describes the application of artificial intelligence to the domain of hard milling process using two strategies. The offline strategy utilizes length of cut as major input while offline strategy uses force signals for the purpose of estimation of tool’s flank wear. Following conclusions can be drawn from the contents of this paper:

1. Fuzzy expert system approach can be utilized for accurate estimation of tool wear.

2. Online strategy, using cutting force signals as primary input, gives better estimation of tool wear as compared to offline strategy that utilizes length of cut as primary input. Aforementioned input parameters are considered as primary ones because their effects upon flank wear are exceedingly more significant as compared to those of other inputs, namely helix angle and workpiece material hardness.

3. Cutting force signals give better real time information about condition of in-progress machining process, whereas, length of cut does not possess such ability.


 
References
Axinte and Dewes, 2001 Axinte, D. A., & Dewes, R. C. (2001). Tool wear and workpiece surface integrity when high-speed ball nose end milling hardened AISI H13. In Proceedings of 3rd international conference on metal cutting and HSM (pp. 171–179). Metz, France.

Cox, 1994 E. Cox, The fuzzy systems handbook: A practical guide to building, using, and maintaining fuzzy systems, Academic Press, New York (1994).

Du et al., 2001 Du, R., Liu, Y., Xu, Y., Li, X., Wong, Y.S., & Hong, G. S. (2001). Tool condition monitoring using transition fuzzy probability. In Proceedings of 3rd international conference on metal cutting and HSM (pp. 375–392). Metz, France.

Fang, 1995 X.D. Fang, Expert system-supported fuzzy diagnosis of finish-turning process states, International Journal of Machine Tools and Manufacture 35 (1995) (6), pp. 913–924. Abstract | Full Text + Links | PDF (753 K) | Abstract + References in Scopus | Cited By in Scopus

Hashmi et al., 2003 K. Hashmi, I.D. Graham and B. Mills, Data selection for turning carbon steel using a fuzzy logic approach, Journal of Material Processing Technology 135 (2003), pp. 44–58. SummaryPlus | Full Text + Links | PDF (457 K) | Abstract + References in Scopus | Cited By in Scopus

Konar, 2000 A. Konar, Artificial intelligence and soft computing, CRC Press LLC (2000).

Luo et al., 2001 Luo, G., Osypiw, D.,& Irle, M. (2001). Tool wear monitoring by online vibration analysis with wavelet algorithm. In Proceedings of 3rd international conference on metal cutting and HSM (pp. 393–404). Metz, France.

Morikawi and Mori, 1993 T. Morikawi and Y. Mori, Recognition of cutting state based on neural network sensor fusion, Journal of Japan Society of Precision Engineering 60 (1993), pp. 779–784.

Obikawa et al., 1996 T. Obikawa, C. Kaseda, T. Matsumura, W.G. Gong and T. Shirakashi, Tool wear monitoring for optimizing cutting conditions, Journal of Material Processing Technology 62 (1996), pp. 374–379. Abstract | Full Text + Links | PDF (573 K) | Abstract + References in Scopus | Cited By in Scopus

Schulz, 2004 Schulz, H. (2004). Why high speed cutting. In Proceedings of international conference on high speed machining (pp. 1–20). Nanjing, China.

Wong and Hamouda, 2002 S.V. Wong and A.M.S. Hamouda, Development of genetic algorithm-based fuzzy rules design for metal cutting data selection, Robotics and Computer Integrated Manufacturing 18 (2002), pp. 1–12. SummaryPlus | Full Text + Links | PDF (219 K) | Abstract + References in Scopus | Cited By in Scopus

Zeren and Ozel, 2002 Zeren, E., & Ozel, T. (2002). Hard milling technology. Report No. MARL-02, Department of Industrial and Systems Engineering, The State University of New Jersey.