Jtam.dvi JOURNAL OF THEORETICAL AND APPLIED MECHANICS 46, 3, pp. 521-530, Warsaw 2008 QUANTITATIVE ESTIMATION OF THE TOOL WEAR EFFECTS IN A RIPPING HEAD BY RECURRENCE PLOTS Grzegorz Litak Technical University of Lublin, Department of Applied Mechanics, Lublin, Poland e-mail: g.litak@pollub.pl Jakub Gajewski Technical University of Lublin, Department of Machine Construction, Lublin, Poland Arkadiusz Syta Technical University of Lublin, Department of Applied Mathematics, Lublin, Poland Józef Jonak Technical University of Lublin, Department of Machine Construction, Lublin, Poland We investigate the time series of a torque applied to the ripping head in the process of a cutting concrete rockwith sharp andblunt tools. By ap- plying nonlinear embeddingmethods and the recurrence plots technique to the corresponding time series, we indicate the changes in nonlinear dynamics lying behind the ripping process, andwe propose a test of the ripping machine efficiency and a way of monitoring the state of tools. Key words: ripping head, nonlinear vibrations, recurrence plot 1. Introduction The dynamics of a cutting process, in the case of metals, was described by Merchant several decades ago (Merchant, 1945). His proposals, involving such phenomena as friction, impacts, and chips breaking, are still inspiring in the current research of metal cutting with higher speeds (Litak, 2002; Sen et al., 2007;Warmiński et al., 2003). Theywere also extended to the problem of cut- ting of brittle materials including cutting of rocks. In this paper, we discuss experimental results obtained using the technology ofmulti-tool rippingheads (Eichbaum, 1980; Fowell, 1993; Gajewski and Jonak, 2006; Jonak andGajew- ski, 2008).Wewill show time histories of the ripping head torque applied to a 522 G. Litak et al. standard rock and later examine thembymeans of embedding and recurrence theories. The aim of the present paper is to show how recurence plots (RP) and recurence quantification analysis (RQA) can help in the assessment of states of the ripping head tools. 2. Considered problem Multi-tool ripping torque time histories T(i) versus the sampling index i for sharp and blunt tools are presented in Figs. 1a and 1b, respectively. All other working conditions besides the states of tools were the same. Note, that the times series of sharp tools is characterized by slightly larger peak valu- es. On the other hand, blunt time series shows a characteristic time interval i ∈ [3700,7100] of a fairly smaller amplitude, which can indicate on possible intermittency in the examined system. Fig. 1. Multi-tool ripping torque time histories T(i) versus the sampling index i for sharp (a) and blunt (b) tools. The sampling time was 0.005s while the rotational velocity was 42rev/min To go further with the standardmethods of nonlinear time series analysis (Fig.1), one should investigate the embedding properties. This approach is based on properties of a phase space which can be reconstructed by supple- mentingmissing coordinates with time delayed ones (Takens, 1981). Thus, the Quantitative estimation of the tool wear effect... 523 examined torque T i on the ripping head at the sampling time given by i can be expressed as the following vector T i = [Ti,Ti−∆i,Ti−2∆i, . . . ,Ti−(M−1)∆i] (2.1) where ∆i is the time delay in sampling time units, while M is the embedding dimension. Now we explain how to obtain proper values of ∆i and M by examining the average mutual information (AMI) (Fraser and Swinney, 1986; Hegger et al., 1999;Kantz andSchreiber, 1997) and the false nearest neighbour fraction (FNNF) (Abarbanel, 1996; Hegger et al., 1999; Kennel and Brown, 1992; Sen et al., 2007). AMI is defined via conditional probabilities of the sequence events AMI(δi) =− ∑ kl pkl(δi)ln pkl(δi) pkpl (2.2) where, for some partition (16 equal parts) of the ripping head, torque values are in an interval T ∈ [Tmin,Tmax], pk is the probability of finding a time series value in the k-th interval, and pkl is the joint probability that an ob- servation falls later into the l-th element, and the observation time is given by δi. The optimal time delay ∆i= δi is to be determined by the first AMI minimum for which the examined events are independent enough to define a new coordinate. Note, that AMI is positively defined and its smallest value (theoretically AMI=0) can be reached when pkl can be factorized to indivi- dual probabilities pk and pl (pkl ≈ pkpl) for any k and l far from each other by δi. On the other hand, to get FNNF one has to choose a point indicated by T i and calculate the distance to its nearest neighbour point Tj in the m-dimensional space. For an Euclidean distance, which is typically used here, it is |T i−T j|m. By iterating both points along the time series, we compute the control parameter Qi,m defined as Qi,m = |T i−Tj|m+1 |T i−Tj|m (2.3) Bycomparing theabovevalue to achosen threshold Qc,we calculate a fraction of cases for which Qi,m exceeds the threshold value Qc. The FNNF can be then estimated from the following expression FNNF(m)= 1 N ∑ i Θ(Qi,m−Qc) (2.4) 524 G. Litak et al. where N is the number of vector elements in the vector time series, Θ(x) is the Heaviside step function. This, so called, fraction analysis is repeated by choosing different values of the dimension m. The optimal value M = m is defined when the fraction of false nearest neighbours tends to zero (note, in some cases, depending on Qc in respect to the standard square deviation of examined time series T(i), some points are omitted and FNNF reaches the minimum for the optimal dimension m=M) lim m→M FNNF(m)→ 0 (2.5) 3. Results of numerical analysis Using the above definitions for AMI and FNNF, we have estimated the em- bedding for the time series of sharp tools. InFigs.2a,b one can easily find that the optimal values are ∆i = 3 and M = 7 representing the time delay and the embedding dimension. Fig. 2. AMI versus time delay δi (a), FNNF versus embedding dimension m (b) for sharp tools These results would be a frame for the RP and RQA techniques. Here we assumed that these quantities were slowly evaluating during cutting of rock. To simplify the procedure of monitoring of the tools state, we fixed the embedding in the initial stage (for sharp tools). This can be additionally justified by the existence of important embedding invariants including the Shannon information entropy (Thiel et al., 2004). The recurrence plot is usually defined by the following matrix form Rm,ǫ with the corresponding elements R m,ǫ ij (Casdagli, 1997; Eckmann et al., 1987; Quantitative estimation of the tool wear effect... 525 Litak et al., 2007; Marwan, 2003, 2006; Marwan et al., 2007; Thiel etal, 2004; Webber and Zbilut, 1994; Wendeker et al., 2004) R m,ǫ ij =Θ(ǫ−|T i−T j|) (3.1) having 0 and 1 elements to be translated into the recurrence plot as an empty place and a black dot, respectively. In this method (RP), patterns showing the diagonal and vertical or horizontal structure of lines are examined.Having obtained such a structure, one can easily classify the dynamics of a studied system (Marwan et al., 2007). In Fig. 3a and 3b, wemapped the correspondingmatrix elements Rij into recurrence plot graphs. One can see that the lines in Fig.3a are more dense than those in Fig.3b. As the space between lines corresponds directly to the characteristic period of torque variations, we can conclude that sharp tools variations possess a higher frequency component in their spectrum. Fig. 3. Recurrence plots for sharp (a) and blunt (b) tools The other possibility is stronger appearance of random oscillations which usually fill the space between lines uniformly (see themost dense region in the middle of Fig.3a). On the other hand, a finite length in some of the vertical lines visible in Figs. 3a and 3b invoke the intermittent character of torque changes in both cases (Litak et al., 2007; Marwan et al., 2007; Wendeker et al., 2004). For quantitative analysis, we define the recurrence rate RR RR= 1 N2 N∑ i,j=1 R m,ǫ ij for |i− j| ­w (3.2) 526 G. Litak et al. which determines the black dots fraction in the RP graph. w denotes the Theiler window used to exclude identical and neighbour points (in our case w=1) from the above summation, Eq.(3.2). Furthermore, the RQA can be used to identify vertical or diagonal lines through their lengths up to Lmax, Vmax for diagonal andvertical lines, respec- tively. In its frame, the RQA enables one to perform probability p(l) or p(v) distribution analysis of lines according to their lengths l or v (for diagonal and vertical lines). In practice, they are calculated as p(y)= Pǫ(y) N∑ y=ymin Pǫ(y) (3.3) where y= l or v, depending on the diagonal or vertical structures in a specific recurrence plot. For various collections of diagonal and vertical lines with respect to their lengths distributions, Shannon information entropies (Lentr and Ventr) can be defined via (Marwan, 2003) Lentr =− N∑ l=lmin p(l)lnp(l) Ventr =− N∑ v=vmin p(v)lnp(v) (3.4) Other properties ofRPas the determinism DET and laminarity LAM aswell as the trapping time TT are also based on the probabilities Pǫ(x) DET = N∑ l=lmin lPǫ(l) N∑ i,j=1 R m,ǫ i,j LAM = N∑ v=vmin vPǫ(v) N∑ v=1 vPǫ(v) (3.5) TT = N∑ v=vmin vPǫ(v) N∑ v=vmin Pǫ(v) In the above equations, lmin and vmin (lmin = vmin = 2 in our case) denote minimal lengths of the diagonal and vertical lineswhich should be chosen for a specificdynamical system.Thedeterminismquantity DET is ameasure of the predictability of the examined timeseriesandgives the ratioof recurrentpoints formed in diagonals to all recurrent points. Note that in a periodic system all points would be included in the lines. On the other hand, the laminarity Quantitative estimation of the tool wear effect... 527 LAM is a similar measure which corresponds to points formed in vertical lines. For small point-to-point changes (laminar), consecutive points form a vertical line. These measures tell about the dynamics behind sampling point changes, and are strictly connected to the point fraction spanning the diagonal (DET) andvertical (LAM) patterns, respectively.Thesediagonal andvertical line patterns are skeletons of deterministic features, while any singular point corresponds to randomness in the examined system. Note that for random numbers, the recurence plot is filled uniformly without any pattern. Finally, the trapping time TT refers to the average length of vertical lines measuring the time scale (in terms of sampling intervals) of these small changes in the examined time history. We performed calculations (using the numerical code provided inMarwan (2006)) of all the specified quantities for our time series (Fig.1) and included them into Table 1. Table 1. Summary of recurrence quantification analysis (RQA) for ’sharp’ and ’blunt’ tools, respectively; w=1 (Theiler,s window), M =7,∆i=3 and lmin =2, vmin =2 recurrence threshold ǫ=4 Type RR DET LAM L max V max Lentr Ventr TT ’sharp’ 0.016 0.341 0.323 72 26 0.811 0.710 2.36 ’blunt’ 0.025 0.819 0.849 318 43 1.918 1.564 3.30 One can see that for the same threshold value ǫ=4 (Eq. (3.1)), we have a larger RR for the ’blunt’ data. This could be connected with higher correla- tions of these experimental data. The same effect of correlations is also visible in DET and LAM parameters which tell about the deterministic structure of the data. Note that this structure is much more transparent in the ’blunt’ data.On the other hand, the entropies Lentr and Ventr show that the distribu- tion of length of particular diagonal and vertical lines is complex and broader in the case of the ’blunt’ data. Note also the differences in Lmax and Vmax (Table 1). These quantities are fairly larger for the ’blunt’ data (Lmax =318 for ’blunt’ data while Lmax = 72 for ’sharp’ data and Vmax = 43 for ’blunt’ data while Vmax = 26 for ’sharp’ data) because of the lack of correlations in the ’sharp’ data. 4. Concluding remarks In summary, we carried out recurrence analysis of the dynamical time series of the multi-tool ripping head. By examining the RP results, we found that oscillations in the case of sharp tools have two additional dynamical compo- 528 G. Litak et al. nents in the spectrum in the case of blunt tools (Fig. 3a and 3b). The first one is higher frequency, while the second important factor corresponds to random noise. Comparing the RQA results for both examined cases, one can also notice fundamental differences in the deterministic structure. This structure, expres- sed by both DET and LAM being close to 1 (see in Table 1 DET =0.819, LAM = 0.849), is fairly well defined in the case of blunt tools, while in the case of sharp tools, DET and LAM are much smaller (see in Table 1 DET =0.341,LAM =0.323). This could be an indicator of the effect of wear in tools and used in an automatedmonitoring procedure of condition of tools. In a rock cutting process, more attention is paid to the expected size of loosening rock elements, which is strictly connectedwith energy consumption. For larger pieces, smaller energy is needed. It is obvious that wear in tools directly changing characteristics of friction and impacts could be responsible for different system responses. For instance, the wear effect on the sintered carbides considerably changes their rake angle of the cutting edges. In a con- sequence for blunt tools, the supplied power increases by about 50% in respect to sharp tools (Jonak and Gajewski, 2008). However, to tell if the recurence plots provide a useful criterion, we need to perform more systematic experi- ments, especially with different rotational velocities of the ripping head and feed rates. These results will be reported in next publications. Acknowledgements This research has been partially supported by the PolishMinistry of Science and Higher Education byGrant No. NN501007033. References 1. AbarbanelH.D.I., 1996,Analysis of ObservedChaotic Data, Springer,Berlin 2. Casdagli M.C., 1997, Recurrence plots revisited,Physica D, 108, 12-44 3. Eckmann J.-P., Kamphorst S.O., Ruelle D., 1987, Recurrence plots of dynamical systems,Europhys. Lett., 5, 973-977 4. Eichbaum F., 1980, Schneidend-brechende Gewinnung mit der Schneidsche- ibe. Experimentell und theoretische Untersuchungen, Gluckauf-Betriebsbucher, Band 22, Verlag Gluckauf GMBHEssen 5. Fowell R.J., 1993, The mechanics of rock cutting, Comprehensive Rock En- gineering, 4, 155-175 Quantitative estimation of the tool wear effect... 529 6. Fraser A.M., Swinney H.L., 1986, Independent coordinates for strange at- tractors frommutual information,Phys. Rev. A, 33, 1134-1140 7. Gajewski J., Jonak J., 2006, Utilisation of neural networks to identify the status of the cutting tool point,Tunnelling andUnderground Space Technology, 21, 180-184 8. Hegger R., Kantz H., Schreiber T., 1999, Practical implementation of nonlinear time series methods: The TISEAN package,Chaos, 9, 413-435 9. Jonak J.,Gajewski J., 2008, Identification of ripping tool typeswith the use of characteristic statistical parameters of time graphs, Tunnelling and Under- ground Space Technology, 23, 18-24 10. KantzH., SchreiberT., 1997,Non-linear Time Series Analysis, Cambridge University Press, Cambridge 11. Kennel M.B., Brown R., Abarbanel H.D.I., 1992, Determining embed- dingdimension for phase-space reconstructionusing ageometrical construction, Physical Review A, 45, 3403-3411 12. Litak G., 2002, Chaotic vibrations in a regenerative cutting process, Chaos, Solitons and Fractals, 13, 1531-1535 13. LitakG.,Kamiński T., Czarnigowski J., ŻukowskiD.,WendekerM., 2007, Cycle-to-cycle oscillations of heat release in a spark ignition engine,Mec- canica, 42, 423-433 14. MarwanN., 2003,Encounters with Neighbours: Current Development of Con- cepts Based on Recurrence Plots and their Applications, PhD Thesis, Univer- sität Potsdam, Potsdam 15. Marwan N., 2006, Recurrence Plots Code, http://www.agnld.uni- potsdam.de/∼marwan/6.download/rp.php 16. Marwan N., Romano M.C., Thiel M., Kurths J., 2007, Recurrence plots for the analysis of complex systems,Physics Reports, 438, 237-329 17. MerchantE.M., 1945,Mechanicsof themetal cuttingprocess,J.Appl. Phys., 11, 267-275 18. Sen A.K., Litak G., Syta A., 2007, Cutting process dynamics by nonlinear time series and wavelet analysis,Chaos, 17, 2797-2803 19. Takens F., 1981, Detecting Strange Attractors in Turbulence, Lecture Notes inMathematics, 898, Springer, Heidelberg, 366-381 20. Thiel M., Romano M.C., Read P.L., Kurths J., 2004, Estimation of dy- namical invariants without embedding by recurrence plots,Chaos, 14, 234-243 21. Warmiński J., LitakG., CartmellM.P.,KhaninR.,WiercigrochM., 2003, Approximate analytical solutions for primary chatter in the nonlinear metal cutting model, Journal of Sound and Vibration, 259, 917-933 530 G. Litak et al. 22. Webber C.L. Jr., Zbilut J.P., 1994, Dynamical assessment of physiologi- cal systems and states using recurrence plot strategies, J. App. Physiol., 76, 965-973 23. Wendeker M., Litak G., Czarnigowski J., Szabelski K., 2004, Nonpe- riodic oscillations of pressure in a spark ignition combustion engine, Int. J. Bif. and Chaos, 14, 1801-1806 Szacowanie ilościowe wpływu stępienia narzędzia na pracę głowicy urabiającej za pomocą wykresów rekurencyjnych Streszczenie W artykule przedstawiono wyniki badań nad zastosowaniem nieliniowychmetod zanurzenia oraz wykresów rekurencyjnych do oceny zmian dynamicznych zachodzą- cychwprocesie urabiania. Przebadano przebiegi czasowemomentu urabiania głowicą uzbrojonąw ostre oraz stępione narzędzia górnicze. Proponowanemetodymogą oka- zać się przydatne w kontekście oceny wydajności orazmonitorowania stanu narzędzi głowicy urabiającej. Manuscript received November 11, 2007; accepted for print March 28, 2008