International Journal of Machine Tools & Manufacture 71 (2013) 26–40 Contents lists available at SciVerse ScienceDirect International Journal of Machine Tools & Manufacture journal homepage: www.elsevier.com/locate/ijmactool Estimation of CNC machine–tool dynamic parameters based on random cutting excitation through operational modal analysis Bin Li a,b, Hui Cai b, Xinyong Mao b,n, Junbin Huang b, Bo Luo b a b State Key Laboratory of Digital Manufacturing Equipment and Technology, Huazhong University of Science and Technology, Wuhan 430074, China National NC System Engineering Research Center, Huazhong University of Science and Technology, Wuhan 430074, China art ic l e i nf o a b s t r a c t Article history: Received 10 January 2013 Received in revised form 7 April 2013 Accepted 9 April 2013 Available online 17 April 2013 Dynamic properties of the whole machine tool structure including tool, spindle, and machine tool frame contribute greatly to the reliability of the machine tool in service and machining quality. However, they will change during operation compared with the results from static frequency response function measurements of classic experimental modal analysis. Therefore, an accurate estimation of the dynamic modal parameters of the whole structure is of great value in real time monitoring, active maintenance, and precise prediction of a stability lobes diagram. Operational modal analysis (OMA) developed from civil engineering works quite efficiently in modal parameters estimation of structure in operation under an intrinsic assumption of white noise excitation. This paper proposes a new methodology for applying this technique in the case of computer numerically controlled (CNC) machine tools during machining operations. A novel random excitation technique based on cutting is presented to meet the white noise excitation requirement. This technique is realized by interrupted cutting of a narrow workpiece step while spindle rotating randomly. The spindle rotation speed is automatically controlled by G-code part program, which contains a series of random speed values produced by MATLAB software following uniform distribution. The resulting cutting produces random pulses and excites the structure in all three directions. The effect of cutting parameters on the excitation frequency and energy was analyzed and simulated. The proposed technique was experimentally validated with two different OMA methods: the Stochastic Subspace Identification (SSI) method and the poly-reference least square complex frequency domain (pLSCF or PolyMAX) method, both of which came up with similar results. It was shown that the proposed excitation technique combined successfully with OMA methods to extract dynamic modal parameters of the machine tool structure. & 2013 Elsevier Ltd. All rights reserved. Keywords: CNC Machine tools Dynamics Random cutting technique Operational modal analysis 1. Introduction The objective of high performance cutting is to machine the parts in the shortest time, while respecting the physical constraints of the process, such as torque, power, vibrations, tool wear and failure, surface quality, and tolerances [1].This puts high demands on good dynamic behavior of the machine tool during machining when the power of the spindle is settled. Since Tobias et al. [2] recognized that the tool or spindle should not be treated in isolation from the whole mechanical system which includes the machine– tool structure and the foundation the characteristics of which partly determine its dynamic behavior. Kolar et al. [3] studied the impact of the machine frame on the dynamic properties at the tool end n Correspondence to: National NC System Engineering Research Center, School of Mechanical Science and Engineering, Huazhong University of Science and Technology (HUST), 1037 Luoyu Road, Hongshan District, Wuhan 430074, Hubei Province, PR China. Tel.: +86 27 87542613 8428, mobile: +86 15007120546; fax: +86 27 87540024. E-mail addresses: maoxyhust@hust.edu.cn, maoxyhust@163.com (X. Mao). 0890-6955/$ - see front matter & 2013 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.ijmachtools.2013.04.001 point experimentally with a representative stiff face milling cutter and with a slim compliant shank cutter. It was recognized that the machine tool frame causes new critical compliances in the low frequency range in the former case as well as a shift of existing eigenfrequencies and an increased level of damping in both cases. Thus, an accurate estimation of the dynamic modal parameters of the whole mechanical system, including the tool, the spindle and the frame, is of great value in online/real time monitoring, active maintenance and precise prediction of stability lobes diagrams in order to achieve high performance cutting. This is generally done by experimental modal analysis (EMA) approaches, an impact test or a shaker test, which calculate frequency transfer functions (FRF) from measurements of both input excitations and corresponding responses at the tool tip when the machine is under rest [4]. However, impact testing usually cannot be repeated unless a special device such as a swing for the impact hammer is used. The frequency bandwidth of the excitation depends on what hammer tip is chosen, and its energy depends more on the operator's force. All these put high demands on the operator's experience. Also, some parts of the machine tools, for B. Li et al. / International Journal of Machine Tools & Manufacture 71 (2013) 26–40 example, the high speed spindle, are too delicate to be hammered heavily upon while weak input causes poor data quality. In addition, the slight nonlinearities of the machine structure appear regularly in impact testing resulting in less reliable results of the linear modal analysis. Moreover, significant changes in modal parameters are expected to occur due to spindle rotation and changes of machine–tool– workpiece boundary conditions between the inactive state and the machining operation state [5,6]. However, impact testing at the tool tip is not feasible during high-speed machining because the operation presents high risk of injury both to machine tools and to the hammer, and thus violates health and safety regulations. Shaker testing is also impossible because the shaker stinger has to be in contact with the rotating tool. All these lead to a failure of applying EMA to identify the dynamic modal parameters of machine tools during machining operation. Fortunately, the cutting force resulting from machining has long been a natural source to excite the structure. Kwiatkowski and Al-Samarai [7] determined the dynamics of a milling machine measuring its structural responses to the random component of the cutting force during normal milling. However, the random component of cutting force is too weak compared with the harmonic ones making it quite difficult to distinguish between natural frequencies and tooth passing frequency and its harmonics. A lot of methods have been developed to create strong and broadband excitation. Bonzanigo and Tsudi [8] used an unequally spaced milling cutter to generate a uniform, broadband cutting force that excited all the modes of a milling machine in the bandwidth of interest. However, their results are valid only for a particular direction of the cutting force. Opitz and Weck [9] excited the machine tool's structure by the random cutting force generated from the continuous cutting of a “random” workpiece. They derived the transfer functions of the structure in a limited frequency range of 40–80 Hz using spectral density measurements of both the force and the displacement signals. Minis [10] developed an improved technique, similar to that of Opitz and Weck, which provides a strong, broadband excitation by interrupted cutting of a specially designed workpiece, where the surface is modulated with pseudorandomly distributed teeth and channels. They identified the dynamics of machine tool structure from input–output measurements. This also proved to be an effective method to excite a 5-axis machining center in Budak's case [11]. However, the interrupted cutting of such a workpiece generated a pseudorandom periodic force signal resulting in a lower limit of the valid excitation frequency band. What's more, this lower limit could be much higher in high-speed machining which will cut off the properties in lower frequency range due to machine tool frame, for example, in Kolar's case. Besides, new workpieces have to be designed and machined if the frequency bandwidth of interest changes according to different machine tools. Also, the machining of the complicated workpiece is time and effort consuming resulting in high cost. Tounsi and Otho [12] developed a pulse-like cutting force excitation as a result of the interrupted cutting of a narrow workpiece width through single tooth milling operations. In the above cases, all the cutting forces and output responses must be measured by an expensive dynamometer in order to identify the dynamics of the structure. This results in high cost to do real time monitoring in a factory background and is sometimes even impossible. Operational modal analysis (OMA), also known as output-only modal analysis, works quite efficiently in modal parameters estimation of a structure on duty under an intrinsic assumption of white noise excitation. The response is the only information to identify the modal parameters of the structure making it quite simple and relatively inexpensive. The theoretical assumption of white noise excitation turns out to be too strict in practical applications. Fortunately, as long as (unknown) input spectral is reasonably flat, OMA methods will work fine [13]. Some of the popular methods are the 27 numerically robust Stochastic Subspace Identification (SSI) method [14–19], the user friendly Frequency Domain Decomposition (FDD) method [20], the industry standard, Least Square Complex Exponential (LSCE) method [5,21,22] and the poly-reference Least Square Complex Frequency domain (pLSCF or PolyMAX) method [13,23,24]. As the model order increases, the SSI method requires a large expenditure of memory and is not suitable for cases that need to handle large amounts of data or require high computational efficiency [25], such as real time monitoring. Also, poles arising from noise and redundant model orders can be so scattered as to mess up the physical poles. FDD is under the assumption that the structure is lightly damped and the modal shapes of close modes are geometrically orthogonal [20]. This may not be the case for machine tools whose damping ratios can be much higher because of many joint interfaces between different parts, and the modal shapes can have arbitrary directions. The PolyMAX method which is a polyreference version of the LSCF method can also be applied to operational data (referred to as Op. PolyMAX) when appropriate preprocessing and post-processing is applied [13]. It proceeds similarly and as fast as the polyreference LSCE method. Moreover, it can identify closely spaced poles quite well, and produce extremely clear stabilization diagrams making the automatic parameter identification process rather straightforward. This enables continuous monitoring of the dynamic properties of machine tools. So far, the only complete methodology to apply OMA under machining operations was detailed by Zaghbani and Songmene [5]. They tried to estimate modal parameters through SSI and the autoregressive moving average (ARMA) method during normal milling operations. However, it is quite difficult to distinguish between natural frequencies and tooth-passing frequencies and their harmonics. Although some criteria were presented to eliminate these harmonic frequencies, the methods were complex and rather experiencedependent. The reason for this problem is that the cutting force generated by normal machining is rather periodic, which does not fulfill the assumption of white noise excitation for OMA. In the present study, the main goals are: (1) to propose a new random excitation technique in accord with the excitation requirement of OMA, (2) to develop a new complete method to apply OMA during machining operations based on the proposed technique, (3) to verify the effectiveness of the excitation through spectral analysis, and (4) to estimate the dynamic modal parameters of the machine tool structure during machining through OMA methods and compare them with the results of conventional impact test and normal cutting tests. The employed OMA methods were SSI and Op.PolyMAX methods. This paper is organized as follows: Section 2 presents the theoretical background of OMA and describes the proposed excitation technique originating from random impulses signal. This signal is modeled and simulated and then the technique to realize this signal with cutting force is proposed. After an analysis of the effect of cutting parameters on excitation bandwidth and energy, the steps to estimate CNC machine-tool dynamic parameters through OMA are summed up. Section 3 presents the experimental verification of the method. The characteristics and feasibility of the proposed technique are discussed and compared with the impact testing. The work is summarized and future works are presented in the conclusions. 2. Operational modal analysis based on random cutting 2.1. Background of OMA The relationship between the inputs x(t) and the responses y(t) [26] can be expressed as Gyy ðωÞ ¼ HðωÞGxx ðωÞHðωÞH ð1Þ 28 B. Li et al. / International Journal of Machine Tools & Manufacture 71 (2013) 26–40 where Gxx(ω) is the (n n) power spectral density (PSD) matrix of the input, Gyy(ω) is the (r r) PSD matrix of the responses, H(ω) is the (r n) frequency response function (FRF) matrix and superscript H denotes the hermitian of a matrix. The FRF matrix can be written in the following well-known partial fraction, i.e. pole/ residue form: N Q r Ψ r Ψ Tr Q nr Ψ nr Ψ H r ð2Þ þ HðωÞ ¼ ∑ n jω−λr jω−λr r¼1 where λr, Ψr and Qr are respectively, the pole, the mode shape vector and a scalar constant of mode r. Superscripts T and n denote the transpose and complex conjugate of a matrix, respectively. The poles occur in complex-conjugated pairs and are related to the natural frequencies ωr and damping ratios ξr as follows: qffiffiffiffiffiffiffiffiffiffiffi ð3Þ λr ; λnr ¼ −ξr ωr 7jωr 1−ξ2r Operational modal analysis (OMA) is developed under an intrinsic assumption of white noise excitation, i.e. the Gxx (ω)¼ const., at least in the frequency band of interest, then Eq. (1) becomes n N ar Ψ r Ψ Tr anr Ψ nr Ψ H br Ψ r Ψ Tr br Ψ nr Ψ H r r þ þ þ ð4Þ Gyy ðωÞ ¼ ∑ jω−λr −jω−λr jω−λnr jω−λnr r¼1 n where ar, anr , br and br all are scalar constant coefficients. The goal of operational modal analysis is to identify the right hand side four terms of Eq. (4) based on measured output data pre-processed into output spectral. However, it is obvious that this PSD model of outputs has four poles (λr,−λr, λnr , and −λnr ) for each mode r, which means its order is twice the order of the FRF model as shown in Eq. (2). Fortunately, it is sufficient to compute the so-called half spectral, Gþ yy ðωÞ, which only consists of the first two terms in Eq. (4) [24]: N ar Ψ r Ψ Tr anr Ψ nr Ψ H r Gþ þ ð5Þ yy ðωÞ ¼ ∑ n jω−λr jω−λr r¼1 This expression of Gþ yy ðωÞ is almost the same as the expression of H(ω) identified by Eq. (2) except for the scalar constant Qr and ar. Each element of matrix Gþ yy ðωÞ is a spectral density function. The diagonal elements of the matrix are the so-called auto-power spectral density (Auto PSD) functions which are the magnitudes of the spectral densities between a response and itself. The offdiagonal elements are the cross-power spectral densities (CSDs) between different responses. The Auto PSDs are all real-valued elements while the CSDs take complex values, carrying the phase information between the measured and the reference degree of freedom. The matrix is symmetric with complex conjugate elements around the diagonal, namely a Hermitian. Any column or row of the matrix carries enough information to extract the modal parameters like the H(ω) matrix. Then the natural frequency ωr, damping ratio ξr and unscaled mode shape Ψr can be estimated based on Gþ yy ðωÞ with classical frequency domain identification methods based on FRF in EMA. Of course, there are some time domain identification methods of OMA based on the correlation function model similar to the impulse response function (IRF) in EMA. It should be noted that when OMA is applied in the estimation of machine–tool dynamics, there are two critical requirements for excitation from the analysis above. First, it needs white noise excitation intrinsically, namely the PSD of the excitation should be reasonable flat over the frequency bandwidth of interests. Second, the corresponding frequency range and energy of the excitation should be adjustable according to different machine tools and actual situations so that all the structure modes in the frequency range of interest are excited. The following section presents the proposed random excitation technique based on cutting that meets the needs mentioned above. 2.2. Structure excitation with random cutting force 2.2.1. The random impulses excitation In EMA, random and impulse signals are two of the most popular excitations used by the shaker test and the impact test respectively. The force signal for random excitation is an ergodic, stationary random signal containing all frequencies within the frequency range. It is nondeterministic which cannot be described by a mathematical function but rather by its statistical characteristics. The white noise excitation needed in OMA is actually a pure random signal whose energy is equally distributed through the whole frequency range (−∞∼+∞); namely, the PSD of the signal is a flat line over the entire frequency band. However, ideal white noise excitation cannot be obtained in reality while the most common one has a reasonable flat PSD in a limited frequency range. Fig. 1a shows the PSD of such a typical random signal generated by MATLAB which has the uniform distribution over [0,100]. Of course, whatever distribution having a flat PSD is fine. Random excitation has the tendency to linearize the behavior of a structure from the measurement data even though it behaves nonlinearly. However, the fact that neither the force signal nor the response is periodic with infinite time history gives rise to an error called leakage. The ideal impulse signal is a Dirac function δ(t) and its spectral is a straight line over the entire frequency range. The practical impulse signal for impact testing is deterministic with limited amplitude and duration of time which can be described by a mathematical function as ( A0 ; t∈½0; τ f si ðt Þ ¼ ð6Þ 0; t∉½0; τ where fsi(t) is the time function of duration of the impulse, A0 is its spectrum Fsi(f) ¼(A0e−iπft sin πft)/πf, signal processing book. Then the Gss(f) of fsi(t) is defined as [21] 2 sin πτf Gss ðf Þ ¼ F si ðf ÞF nsi ðf Þ ¼ A20 πf single impulse input, τ is the amplitude, and its frequency which can be found in any auto-power spectral density ð7Þ Its zeros are f¼ k/τ (k ¼ 71, 72 …). It is clear that the frequency bandwidth of the first power spectral lobe (k¼ 1) increases as τ decreases, as shown in Fig. 1b. Though impact excitation is convenient to use and very portable for field and laboratory tests, it is difficult to control either the force level or the frequency range of the impact. This could affect the signal-to-noise ratio in the measurement, thus resulting in poor quality data. Fortunately, distributing many impulses randomly in time domain gives a new random excitation fri(t): ( Ai ; t∈½t i ; t i þ τ f ri ðt Þ ¼ ð8Þ 0; t∉½t i ; t i þ τ where n is the number of impulses the excitation contained, Ai (i¼1, 2, …, n) is the amplitude of each impulse, and ti (ti≥0, i¼1, 2, …, n) is a random variable representing the start moment of the ith impulse. The spectral of fri(t) is Z F si ðf Þ ¼ ∞ −∞ n f ri ðtÞe−j2πf t dt ¼ ∑ n ∑ Ai e i¼1 i¼1 −j2πf ðτ=2þt i Þ Z t i þτ ti Ai e−j2πf t dt ¼ sinπτf πf ð9Þ B. Li et al. / International Journal of Machine Tools & Manufacture 71 (2013) 26–40 29 Fig. 1. (a) Random signal with reasonable flat PSD in a limited frequency band and (b) Time signal and PSD of impulses with different durations. yielded: The PSD of the signal is Grr ðf Þ ¼ F ri ðf ÞF nri ðf Þ ¼ 2 sin πτf πf 2 Grr ðf Þ ¼ nðAτÞ2 3 6 n 7 n n 6 7 6 ∑ A2i þ ∑ ∑ Ai Aj cos 2πf ðt i −t j Þ7 4i ¼ 1 5 i¼1 j¼1 i≠j ð10Þ j≠i where tj (j¼ 1, 2,…, n; j≠i) following the same distribution as ti. It is clear that Eq. (10) contains three independent parameters, namely τ, Ai (or Aj), and ti (or tj), which originate from the excitation signal fri(t) and has the same zeros as Eq. (7). Then the bandwidth of the first spectral lobe (referred to as BW1st) is the inverse of the pulse duration: BW 1st ¼ 1 τ ð11Þ tj (or tj; i, j ¼1, 2,…, n; j≠i) is a very important variable characterizing the statistical properties of the random impulses excitation. It determines the moment of when each impulse occurs and the number of impulses in unit time (referred to as density ρ). However, the variable ρ, which is analyzed here, is much more controllable than the moment, which is rather random. Because ti and tj are both random variables obeying the same distribution, the term Ai Aj cos ω(ti−tj) in Eq. (10) becomes zero after many times of averaging. If the amplitude of each impulse is almost the same, namely Ai ¼A (i¼ 1, 2, …, n), the averaged PSD of this signal is sin πτf πτf 2 ð12Þ Grr ðf Þ represents the distribution of the signal energy, referred to as Ee, as a function of frequency. Eq. (11) indicates that the frequency range BW1st of the excitation is inversely proportional to the duration of each impulse τ. Eq. (12) indicates that the excitation energy Ee is proportional to the number of impulses contained in one sample, which is determined by ρ, and the square of the impulse amplitude A. Fig. 2a presents three different random impulse signals of which the amplitude A is 1 N, the duration of each impulse τ is 2 ms and the total length of time (referred to as T) is 1 s while the density ρ is 1, 8, and 32 impulses per second respectively. Fig. 2b presents the PSD of the three signals. It is clear that BW1st does not change while the amplitude of the PSD increases almost eightfold (9 dB) as ρ octuples and increases almost fourfold (6 dB) as ρ quadruples. Fig. 2c presents the PSD of signal2 in Fig. 2a, while A takes up 1 N, 4 N and 16 N. The PSD of the three signals have the same BW1st while the PSD amplitude increases almost 16 times (12 dB) as the impulse amplitude quadruples. Fig. 2d presents the PSD of signal2 while τ takes up 2 ms, 4 ms and 8 ms and the corresponding BW1st of each signal is 500 Hz, 250 Hz and 125 Hz respectively. The PSD spectral is reasonably flat over almost the entire first lobe with 15 dB roll-off which is acceptable [27]. Moreover, as τ doubles the amplitude of the first lobe increases almost fourfold (6 dB). This is 30 B. Li et al. / International Journal of Machine Tools & Manufacture 71 (2013) 26–40 Fig. 2. (a) Three random impulse time signals with different ρ (1, 8, and 32 impulses per second); (b) PSD of the three signals; (c) PSD of signal2 while A changes to 1 N, 4 N, and 16 N; (d) PSD of signal2 while τ changes to 2 ms, 4 ms, and 8 ms. because when τ is quite small, sin πτf≈πτf and then Grr ðf Þ∝τ2 according to Eq. (12). To sum up, the frequency bandwidth and energy of this random excitation can be adjusted easily by its parameters, namely the duration of impulse τ, the density of the impulses ρ, and the amplitude of the impulses A. BW1st of the excitation is inversely proportional to τ, while the energy Ee is proportional to ρ and the square of both A and τ. Its PSD is reasonably flat over half the first lobe BW1st/2 with 10 dB variations which is acceptable. So it is a fairly good input for OMA. 2.2.2. Realization of random impulses excitation with cutting force Minis [10] developed an improved method to provide a strong, broadband excitation by interrupted cutting of a specially designed workpiece the surface of which modulated with pseudorandomly distributed teeth and channels. The interrupted cutting of that workpiece generated a pseudorandom impulses force signal. This also proved to be an effective method in Budak's case [11] to excite a 5-axis machining center. However, new workpieces have to be designed and machined if the frequency bandwidth of interest changes according to different machine tools. The machining of the complicated workpiece are time and effort consuming resulting in high cost. Moreover, the limited dimension of the workpiece results in a pseudorandom impulses force signal instead of a pure random one. In order to generate random impulses excitation with cutting force and to simplify the machining of the workpiece at the same time, a new excitation technique specifically suitable for widely used CNC machine tools is proposed. Thanks to technological progress, both the main motion and feed motion of CNC machine tools can be controlled automatically and accurately. It is common that the main motion and feed motion are stable in a single pass. However, when the speed of the main motion is random in a single pass, random cutting force will be generated. One step further, if what is being machined is a narrow tooth of a workpiece, cutting force like random impulses finally occurs. As the impulse duration decreases, the frequency bandwidth of the first spectral lobe increases; this is analyzed in the above section. In order to make sure that the random impulses cutting force excites all the structure modes in the bandwidth of interest, the frequency range of the excitation must be wide enough and the impulse energy should be powerful enough. Therefore, the duration of the impulses should be short enough, namely mill cuts engagement should not last much time, but not too short, which will result in too week impulses. On the other hand, it is known that the interaction between the machine tool and the cutting process can be represented by a closed loop system that incorporates two fundamental blocks [10]. One is the machine tool block GMT describes the relationship between the dynamic cutting force, which is applied on the tool, and the response of the structure. The other is the cutting process block GCP which describes the effect of the relative displacement of the tool and workpiece on the variation of the cutting force. In order to estimate the B. Li et al. / International Journal of Machine Tools & Manufacture 71 (2013) 26–40 dynamics of the machine-tool structure GMT during machining, GCP should be uncoupled from the structure dynamics. Fortunately, the cutting process block is valid only in a period of time τ when the cutting is performed. As this tends to zero, the cutting force becomes more pulse-like and then is uncoupled from the structure dynamics. Thus, a narrow radial depth was chosen to reduce the time of engagement and eliminate the effect of the cutting process GCP [12]. Fig. 3 shows the schematization of the proposed technique applied in a milling machine. In the figure, a workpiece with a single tooth is machined while the face milling cutter rotates randomly (the revolution speed N or n(t) is a random variable). The cutting forces are assumed to be proportional to the uncut chip area Ac [28]. In the above case, Ac consists of two faces, the flank and the button, as Fig. 6 shows. So Ac is calculated as Ac ¼ aw ad þ aw T w sin α ð13Þ where ad is the axial depth, aw is the width of the tooth, Tw is the width of the blade, and α is the angle of tool cutting edge. The cutting speed v (mm/s) is v¼ πnD 60 ð14Þ where n is the mean of revolution speed N, and D is the diameter of the cutter. Then the duration τ of each pulse is aw 60aw τ¼ ¼ v πnD ð15Þ Substituting Eq. (11) into (15) gives BW 1st ¼ πnD 60aw ð16Þ Because the PSD spectral is reasonably flat over half the first lobe, which has been discussed in the above section, BW1st/2 can be set as the upper limit fh of the effective bandwidth of the excitation. While the lower frequency limit of the excitation fl given by Minis [10] is fl ¼ 15 Δf π ð17Þ 31 where Δf is the inverse of the period of revolution T (¼60/n, and n is the constant spindle speed). So Eq. (17) becomes fl ¼ 15 n n ¼ π 60 4π ð18Þ There exists a lower frequency limit in Minis' method because both force components of orthogonal cutting are pseudorandom periodic signals. In contrast, the proposed excitation in this paper is truly random resulting in no lower limit. From the above two sections, the relationships between BW1st (and Ee) of the excitation and impulse parameters and the relationships between impulse parameters and cutting parameters are summed up in Table 1. It can be seen from Table 1 that BW1st is only related to one impulse parameter τ while Ee is far more adjustable through all three impulse parameters. The task of realizing random impulses excitation with cutting force is how to choose the cutting parameters according to actual needs. What is known before the work is just the frequency range of interest (referred to as BW1st/2). Before the milling operation, the tools and workpieces at hand are surely known, so the diameter D of the cutter and the width aw of the tooth for excitation are actually determinate. In practice, aw should be chosen carefully in order to provide strong impulses as well as to eliminate the effect of the cutting process. Then Eq. (16) can be rewritten as n¼ 60BW 1st aw πD ð19Þ and the mean n of the revolution speed N is obtained. The random variable N is the key to fulfill white noise excitation. Taking into account the rotational inertia and response speed of the motor, the spindle speed cannot change immediately and dramatically. A parameter Δn is defined to present the maximum difference between adjacent speed value ni and ni+1 (i¼ 1, 2, …, k) where k is the desired number of the speed values. Define a continuous random Table 1 Summary of the relationships between cutting parameters and excitation signal. Energy Ee Bandwidth BW1st Impulse parameters Cutting parameters BW1st ¼1/τ A∝aw, ad τ∝aw, ∝−n ρ∝n, af aw, ad a w, n n, af 2 Ee∝A Ee∝τ2 Ee∝ρ where ∝ denotes proportional relationship and ∝− denotes inverse relationship. Fig. 3. Schematization of the proposed excitation technique. 32 B. Li et al. / International Journal of Machine Tools & Manufacture 71 (2013) 26–40 Fig. 4. Random signal and histogram of N and Nu. variable Nu to be the difference between adjacent speeds which has the uniform distribution over [−Δn, Δn] and then the random variable N denoting the spindle speed is defined as n N i −N i−1 ¼ N ui or Ni ¼ N 0 þ ∑ N ui ði ¼ 1; 2; ⋯; k; N 0 ¼ nÞ i¼1 ii. ð20Þ According to the definition of the Wiener process (or Brownian motion) [29], the stochastic process {Ni: i≥0} is a Wiener process if the increments Nui (i¼1, 2,…, k) are normally distributed with mean μ¼ 0 and variance s2 ¼1. However, they are uniformly distributed over [−Δn, Δn] in this case. It can be supposed that {Ni: i≥0} are also uniformly distributed with mean μ¼n which will be checked by Matlab later. It should be noted that the spindle speed N should be present in a limited range [nl, nh] rather than too low or too high during machining (N should be smaller than the maximum speed and larger than zero), and its PSD should be reasonable flat. Such a random variable is easily generated by MATLAB software. Compared with the feed af, the axial depth ad is more effective to adjust Ee, so only changing ad while choosing a constant value for af is a good choice to increase the excitation energy. So far all the cutting parameters are determined. In summary, the complete steps to apply the proposed random impulses excitation technique in a CNC machine tool to do OMA are as follows: i. Design and manufacture the workpiece. Only one tooth is used for excitation and the width aw of this tooth should be chosen iii. iv. v. carefully in order to provide strong impulses as well as to eliminate the effect of the cutting process. Calculate average spindle speed n from Eq. (19) according to the frequency band of interest BW1st/2, the tool diameter D and the width aw; Generate the values of random variable N through MATLAB. A sequence of values of N limited in certain range [nl, nh] can be easily obtained from Eq. (20) through MATLAB; Manually prepare the NC part program. The part program contains a sequence of commands and each command contains a spindle speed value from sequence obtained in step iii. Meanwhile, only change the axial depth ad to adjust the excitation energy while choose a normal value for feed af in the program; Machine the workpiece to excite the structure and pick up vibrations of different points to do OMA. 3. Experimental verification 3.1. Realization of random cutting excitation The proposed excitation technique was conducted on a 3-axis vertical milling center XHK5140 made by Huazhong University of Science & Technology shown in Fig. 5. It is equipped with a spindle, the speed of which is up to 3000 rpm and the maximum B. Li et al. / International Journal of Machine Tools & Manufacture 71 (2013) 26–40 power of which is 11 kW, and with a numerical control system HNC-22M from Wuhan Huazhong Numerical Control Co., Ltd (HNC). The machine structure represents a very common type of machine tool design with a moving spindle box, a worktable and a slide. The primary motion is the spindle rotation and the feed motion is completed by the spindle box moving along the z-axis, the worktable moving along the x-axis and the slide along the y-axis. According to Minis's method, the lower limit fl is 31.8 Hz [see Eq. (18)] if the rotational speed is 400 rpm which is common in milling operations. However, it is not acceptable because the dynamics of lower frequency range (0–50 Hz) corresponding to the spindle speed range (0–3000 rpm) is of great significance to machining behavior. So Minis's method fails while the proposed excitation technique is a good choice. A representative stiff face milling cutter with the diameter of 80 mm and total length of 92 mm was used and only one tool tooth was engaged in the milling operation. The width aw of the workpiece tooth for excitation was 4 mm and the frequency band of interest BW1st/2 was chosen to be 250 Hz. The feed af was kept at 0.1 mm/tooth and the axial depth ad of cut could vary from 0.5 up to 4 mm in order 33 to adjust the excitation energy with a depth of 1.2 mm originally chosen. According to Eq. (19) the average rotation speed n was 477 rpm and it was first made 500 rpm in order to set the speed limits [300,700] rpm. The maximum difference Δn between adjacent spindle speeds is set to be 20 rpm. 800 values of a random variable Nu having the uniform distribution over [−20, 20] is first generated by MATLAB. The mean of the 800 values is 0.365. Then a sequence of values between 300 and 700 rpm for random variable N denoting the spindle speed was generated according to Eq. (20). The final average rotation speed n is 490 rpm resulting in a final BW1st 513.14 Hz. The minimum speed is 311 rpm, the maximum is 676 rpm and the standard deviation is 91.4 rpm. Plotting Nu and N as a function of time in 60 s results in two random signals shown in Fig. 4a and b. A common 45# carbon steel workpiece with two teeth was machined as Fig. 5c shows. One tooth was used here for excitation and the other was used for another experiment. The length of the teeth was 80 mm and the height was 4 mm. Finally a NC part program was manually prepared which contained the spindle speed values obtained above. The theoretical cutting period calculated according to NC part program is 101.489 s. Fig. 5. Experimental setup: (a) XHK5140 milling center and the numbered red squares show the locations of 6 PCB 356A15 accelerometers: 1-Column:L1, 2-Column:L18, 3-Headstock:10, 4-worktable:31, 5-slide:21, 6-base:25; (b) the face cutter used and (c) the workpiece for excitation (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.). 34 B. Li et al. / International Journal of Machine Tools & Manufacture 71 (2013) 26–40 3.2. Experimental setup and measurements Fig. 5 shows the experimental setup. During the machining, the cutting forces in the x, y, and z directions were measured using a three-axis 9253B23 Kistler table dynamometer (Measurement range: 712 kN; Frequency range (−3 dB):0–45 kHz; Resonant frequency: fn(x)≈610 Hz, fn(y)≈570 Hz, fn(z)≈570 Hz). The FRF of the dynamometer was measured when it was mounted on the machine worktable shown in Fig. 8b. It is clear that the frequency band of interest BW1st/2 (250 Hz) is still within the bandwidth of the dynamometer, while the bandwidth BW1st exceeds a little which is acceptable. A steel plate was faced and drilled to adjust the table dynamometer holes to the T-slots of the machine worktable and firmly bolted to the worktable ensuring rigid mounting. Then the dynamometer was bolted to the plate and the workpiece was bolted to the dynamometer. The tooth of the cutter was an indexical carbide insert (SEHT1204) from EGO Machine Tools Co., Ltd. Six three-axis accelerometers of type PCB 356A15 (Measurement range: 750 g; Frequency range (75%): 2–5000 Hz; Resonant Frequency: ≥25 kHz) were mounted to measure the vibrations of the machine structure, the slide, the worktable, the base, the headstock, and the top and bottom of the column, in all three directions due to the cutting force. And a one-axis accelerometer of type PCB 352C34 (Measurement range: 7 50 g; Frequency range: 0.5–10,000 Hz; Resonant frequency: ≥50 kHz) was used to measure the vibration of the workpiece in the x direction during machining. The locations of all the accelerometers are shown in Fig. 5. In order to make sure that the generated cutting force can be controlled as expected, the axial depth of all the cutting tests were chosen according to previous cutting results so that the cutting Table 2 Summary of the cutting conditions for the experimental tests. Case A is symmetric tooth milling of the workpiece; Case B is symmetric face milling of the workpiece. Test #2, #4 and #6 were later used to extract modal parameters. Case A B Test # Feed (mm/tooth) Axial depth (mm) Spindle speeds (rpm) 1 0.1 1.2 2 0.1 1.5 3 4 5 6 7 8 9 0.05 0.1 0.2 0.1 0.1 0.1 0.1 1 1 1 1 1 0.5 0.7 800 random values over [300,700] 500 random values over [300,700] 400 400 400 600 800 400 400 process is stable. Table 2 summarizes all the cutting tests during the experimental study. During all the cutting tests, the tool cuts symmetrically in end-milling instead of peripheral milling (upmilling or down-milling) while the slide feeding along the +y direction. Of course, the proposed excitation technique can be applied easily in other conditions according to the machine tool considered. The measured boundary condition is that the spindle is rotating, the machine table is moving and the tool is moved at the exact position without contact with the workpiece while machining. Because the face milling cutter is an example of very compact and stiff tool and both the axial depth and radial depth are small, the deflection of the tool during the cutting is negligible. The total 22 signals were collected by the acquisition system LMS SCADAS Mobile SCM05 simultaneously at a sampling rate of 1024 Hz. The cutting forces were only measured to analysis the effectiveness of the excitation. Table 3 Comparison of the machine–tool modal parameters estimated using the Op. PolyMAX method for test #2, #4 and #6 and the PolyMAX method for tap test. The SSI method was also employed for test #2. Modes Test #2 [300,700] rpm SSI 1 ωn (Hz) ζ (%) 2 ωn (Hz) ζ (%) 3 ωn (Hz) ζ (%) 4 ωn (Hz) ζ (%) 5 ωn (Hz) ζ (%) 6 ωn (Hz) ζ (%) 7 ωn (Hz) ζ (%) 8 ωn (Hz) ζ (%) 9 ωn (Hz) ζ (%) Tap test 0 rpm Test #4 400 rpm Test #6 600 rpm Op. PolyMAX 17.611 17.473 17.436 17.073 16.839 6.843 19.593 3.687 19.416 1.078 18.976 1.358 20.037 3.86 20.185 1.83 22.351 1.381 22.844 2.207 21.687 0.55 22.912 0.136 22.594 4.702 69.563 1.55 69.469 3.387 71.066 1.145 73.237 1.937 72.035 4.988 91.677 2.261 91.124 4.819 91.098 0.237 93.396 0.468 93.944 5.024 1.19 120.936 122.266 4.764 122.293 0.169 121.432 0.122 123.397 3.552 0.608 130.46 130.577 3.114 133.027 0.071 134.997 0.274 132.304 10.607 1.937 175.055 175.022 2.548 175.153 0.062 173.32 0.164 173.156 3.819 2.739 243.131 242.788 3.915 239.392 0.083 238.872 0.052 241.675 4.933 0.276 3.323 0.808 Fig. 6. Impact tests with different hammers: (a) Tool tip impact test with HPCB and (b) Column impact test with HDFC. 0.06 B. Li et al. / International Journal of Machine Tools & Manufacture 71 (2013) 26–40 What's more, multiple-input multiple-output (MIMO) impact tests with two different impact hammers were conducted on the machine shown in Fig. 6 under the same experimental setup. One is a light hammer of type PCB-086C04 (referred to HPCB), and the other is a much more powerful hammer of type DFC-1 (referred to as HDFC) which comes from INV (China Orient Institute of Noise & Vibration). The light hammer tapped the cutter and the heavy hammer tapped the top of the column both in x and y directions. 35 The resulting modal parameters acted as a comparative reference and are presented in Table 3 (referred to as EMA). 3.3. Results and discussion 3.3.1. Spectral analysis Fig. 7 presents the cutting force signals in x, y, and z directions and the corresponding acceleration signals of point Headstock: 10 Fig. 7. The cutting force and acceleration signals (point Headstock:10) of test #1 in Table 2: (a) Cutting force and acceleration signal for the whole acquisition time together with an enlarged view of the signal profile for 1 s in x direction; (b) Cutting force and acceleration signal in y direction and (d) Cutting force and acceleration signal in z direction. 36 B. Li et al. / International Journal of Machine Tools & Manufacture 71 (2013) 26–40 of test #1. It is clear that the cutting forces excited the structure in all three directions during the cutting process. A good similarity was observed between the force signals which fulfilled the assumption of random impulses excitation in the time domain. The effective cutting period appeared from 8.125 s to 117.148 s and the average amplitude of the impulses was about 700 N in x, 400 N in y, 380 N in z directions respectively. Fig. 8a presents the PSDs of the cutting forces in cutting tests and the PSDs of impulse force in the impact tests. The spectral estimation was performed using the modified averaged periodogram method (Welch's technique) with an overlap of 50% and a Hanning weighting function. This ensures that all data are equally weighted in the averaging process, minimizing leakage and picket fence effects. It can be seen that the PSDs of the cutting forces in three directions present about 15 dB roll-off cross the first lobe BW1st (¼511.9 Hz, which match the desired 513.14 Hz) with some fluctuations. They are reasonably flat over half the first lobe, which closely meets the excitation requirements of OMA. Besides, the PSD of the x component of the cutting force is higher than the HDFC hammer impact force in the x direction. Yet the other two cutting force components are almost the same level as the HDFC impact force in y direction. In addition, the PSDs of all the cutting force components are at least 10 dB (10 times) higher than the ones of the HPCB hammer impact forces although both of which are inputted at the tool tooth. It means that the designed cutting force excitation is quite powerful which can greatly improve the energy level of the corresponding responses. The real spindle speeds during machining can be estimated from the measured cutting force impulses. Because the tool cut the workpiece once per revolution, the duration Δt between two adjacent force impulses is the time the spindle cost to finish one revolution. The average rotation speed n(Δti) in this revolution can be estimated by n(Δti)¼ 60/Δti rpm. Fig. 8c shows the designed and estimated spindle speed start at the same time 8.125 s. It is clear that the trend of the estimated speed match quite well with the designed speed with some delay in time. The maximum of this measured speed is 682.7 rpm, the minimum is 305.7 rpm, the mean is 492.7 rpm and the standard deviation is 94.3 rpm. Compared with Fig. 4b, the statistic properties and distribution of the estimated speed were almost the same as the ones of designed speed N verifying that the excitation technique was carried out successfully. Fig. 9 presents the PSD of the acceleration signals at the points Headstock:10 and Column:L1 under cutting test #1 and HPCB impact test. Since the vibration modes in all three directions are of interest while the tool tip cannot be tapped in the z direction, only the signals of the x and y directions are presented here to do comparison while all acceleration signals within the effective cutting period would later be used in modal parameters estimation. It can be observed that the PSDs of the acceleration signals in both directions under the cutting excitation move almost parallel to the ones of the hammer impact outputs. The former is at least 10 dB higher than the latter leading to a clearer presentation of the peaks and stronger ant-jam capability in a factory background. What's more, the modes in both directions are excited well in the cutting test compared with impact test. So it can be assumed that the modes in all three directions in the frequency range of interest are excited by the cutting forces. 3.3.2. Structure identification All the acceleration signals of 7 points within the effective cutting time period were truncated for operational modal analysis. Each of the signals must contain at least 65,000 data points. Two OMA methods Fig. 8. (a) Power spectral density of different excitations; (b) Driving point FRF (of Point 1) between impact input and dynamometer outputs in x, y and z direction and (c) Designed and estimated spindle speed during the effective cutting period. B. Li et al. / International Journal of Machine Tools & Manufacture 71 (2013) 26–40 (SSI and Op.PolyMAX) were used to estimate the modal parameters of the machine tool, namely the natural frequencies, damping ratios, and unscaled modal shapes. Fig. 10a presents the sum of the PSD of all signals within the mode stabilization diagram for the results of test #2 using the Op.PolyMAX method. It is obvious that 9 modes corresponding to the nine peaks in Fig. 9 fall in the frequency range of interest (0–250 Hz). Another numerical method (SSI) was used to confirm the validity of the results and the efficiency of the Op.PolyMAX method. Fig. 10b presents the stabilization diagram for the results of test #2 using SSI. A tolerance of 1% for frequency stability, 2% for vector stability and 5% for damping stability is used in both diagrams. It turns out that the Op.PolyMAX yields an extremely clear stabilization diagram leading to a quite easy selection of the physical poles. In contrast, the physical poles calculated by the SSI method are messed up by non-physical poles which tend to wander in the stabilization diagram as the model order increases. When it comes to computational efficiency, a lot of research efforts [23,30–33] have investigated. In this work, the Op.PolyMAX method takes about 2 s to calculate and display the stabilization diagram at a model order of 40, while the SSI method takes 4 s on the same PC platform. As the model order increases, the advantage of the Op.PolyMAX goes further. The modal parameters of test #2 (referred to as [300,700] rpm) by the two methods are presented in Table 3. Zaghbani and Songmene's method [5] was used here to extract modal parameters during two representative normal machining tests #4 (referred to as 400 rpm) and #6 (referred to as 600 rpm). All the results are summarized in Table 3. From the table, it can be 37 seen that the results of the Op.PolyMAX method are quite similar to their SSI counterparts, for example, the relative variation of natural frequencies ([ωOp.PolyMAX−ωSSI]/ωSSI) lies between 0.02% (mode 8) and 2.21% (mode 3). Fig. 11 presents the damping ratios and relative variation of the natural frequencies of the tap test (0 rpm), the test at 400 rpm and the test at [300,700] rpm compared with the reference test at 600 rpm. It can be observed that the relative variation of natural frequencies between [300,700] rpm and 600 rpm lies between 0.46% (mode 9) and 3.81% (mode 2) while the variation between 0 rpm and 600 rpm lies between 0.55% (mode 7) and 5.99% (mode 2). The difference of damping ratios between [300,700] rpm and 600 rpm lies between 0.17% (mode 1) and 2.69% (mode 8) while the difference between 0 rpm and 600 rpm lies between 1.45% (mode 3) and 4.87% (mode 9). So the results of test #2 ([300,700] rpm) which employed the proposed random cutting excitation technique are quite close to the results of the normal machining operation (test #6), much closer than the tap test. The same conclusion can be obtained when the reference changes to test #4. Besides, a powerful tool, the modal assurance criterion (MAC), was used to evaluate the quality of the mode shapes. MAC is defined as the squared correlation coefficient between two mode shape vectors [21], which assesses the correlation between these two vectors: MACðΨ r Ψ s Þ ¼ 2 jΨ H r Ψ sj H H ðΨ r Ψ r ÞðΨ s Ψ s Þ Fig. 9. PSD of acceleration signals of Tool holder: 10 and Column:L1 under different excitations. ð21Þ 38 B. Li et al. / International Journal of Machine Tools & Manufacture 71 (2013) 26–40 Fig. 10. Mode stabilization diagram: (a) using operational PolyMAX method (Op. PolyMAX); (b) using the stochastic subspace identification (SSI) method. The treated acceleration signals are from test #2. The symbol ‘o’ indicates a new pole; ‘f’ indicates stable frequency; ‘d’ indicates stable frequency and damping; ‘v’ indicates stable frequency and eigenvector; ‘s’ indicates that all criteria are stable. Fig. 11. Comparison of the machine–tool modal parameters generated by the PolyMAX method at different spindle revolution speeds. The former graph is a relative variation ( ¼ ωn−machining/ωn−600 rpm) of natural frequencies and the latter is an absolute presentation. If two vectors are estimates of the same physical mode shape, the MAC should approach unity (100%), otherwise the MAC should be low. A high quality mode set normally contains diagonal elements which are 100% (by definition) and off-diagonal elements which have a low value (close to 0%). The unscaled mode shapes of test #2 are rather good and the mode shapes of the two OMA methods are very similar, which is evidenced by MAC values represented in Fig. 12a and b. Fig. 12c and d indicated that the mode shapes of test #2 are reasonably similar to the shapes of test #6 compared with the shapes of tap test. Similar results can be observed for test #2 and test #4 according to Fig.12e and f. To sum up, the results obtained from operational modal analysis based on the proposed random cutting technique are extremely close to the results of normal cutting conditions. The estimated modal parameters can characterize the dynamic properties of machine–tool system during machining within the chosen revolution speed range, so they are of great value in on-line monitoring, active maintenance and even precise prediction of the stability lobes diagram which needs further investigation to obtain the modal shape scaling factor. Also, many peaks existing in the lower frequency range of the responses' PSD verifies that the dynamics of lower frequency range determined by machine frame is of great significance to machining behavior. 4. Conclusions The dynamic performance of machine tools will change during machining, so an accurate estimation of the dynamics is of great significance in on-line/real time monitoring, active maintenance, and precise prediction of the stability lobes diagram in order to reach high performance cutting. This paper presents a new technique to generate a strong and broadband excitation to meet the white noise excitation requirements of operational modal B. Li et al. / International Journal of Machine Tools & Manufacture 71 (2013) 26–40 39 Fig.12. MAC values of different mode shapes: (a) Auto-MAC of mode shapes using Op.PolyMAX of test #2; (b) MAC between mode shapes using the Op.PolyMAX and SSI method of test #2; (c) MAC between mode shapes of test #2 and test #6; (d) MAC between mode shapes of tap test and test #6; (e) MAC between mode shapes of test #2 and test #4 and (f) MAC between mode shapes of tap test and test #4. 40 B. Li et al. / International Journal of Machine Tools & Manufacture 71 (2013) 26–40 analysis, which is a powerful tool for dynamic modal parameter estimations during machining. The mathematical model of this excitation signal was first proposed, and the effect of the signal parameters on its energy and effective excitation frequency bandwidth was analyzed. Then the technique to realize this excitation on CNC machine tools with cutting force was detailed. Finally, this technique was experimentally validated and applied successfully using two OMA methods (SSI and Op.PolyMAX) to extract the modal parameters of the whole mechanical system, including the tool, the spindle and the frame under machining conditions. It was demonstrated in the mode stabilization diagram that this technique can generate a quite powerful excitation and avoid the problem of distinguishing between natural frequencies and the harmonics of tooth passing frequency leading to a much easier pole selection. Besides, the modal parameters estimated during the random machining operation are quite close to the estimated parameters during normal machining operations. Thus, they can characterize the dynamic properties of a machine–tool system during machining within the chosen revolution speed range. Because the modal shapes are not mass normalized, a scaling factor (or modal mass) must be estimated to synthesize FRF, which needs further investigation. [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] Acknowledgment [20] This work was funded by the National Natural Science Foundation of China (NSFC) under Grant nos. 51275188 and 51121002, and the Science and Technology Major Special Project of China under Grant no. 2011CB706803. The authors would like to acknowledge Mr. and Mrs. Serody for their help to check and revise the grammatical and spelling errors in the paper. [21] [22] [23] [24] References [25] [1] Y. Altintas, High performance cutting, The International Journal of Advanced Manufacturing Technology 33 (2007) 367. [2] S.A. Tobias, W. Fishwick, The vibrations of radial-drilling machines under test and working conditions, Proceedings of the Institution of Mechanical Engineers 170 (1956) 232–264. [3] P. Kolar, M. Sulitka, M. Janota, Simulation of dynamic properties of a spindle and tool system coupled with a machine tool frame, International Journal of Advanced Manufacturing Technology 54 (2011) 11–20. [4] Y. Altintas, Manufacturing automation: metal cutting mechanics, machine tool vibrations, and CNC design, Cambridge University Press, Cambridge, 2000. [5] I. Zaghbani, V. Songmene, Estimation of machine-tool dynamic parameters during machining operation through operational modal analysis, International Journal of Machine Tools and Manufacture 49 (2009) 947–957. [6] V. Gagnol, T. Le, P. Ray, Modal identification of spindle-tool unit in high-speed machining, Mechanical Systems and Signal Processing 25 (2011) 2388–2398. [7] A.W. Kwiatkowski, H.M. Al-Samarai, Identification of milling machine receptances from random signals during cutting, Annals of the CIRP 16 (1968) 137–144. [8] F. Bonzanigo, J.P. Tsudi, Identification of Milling Machine Structures from the Output Signal Only, Proc., Eurisco-83. [9] H. Opitz, M. Weck, Determination of the transfer function by means of spectral density measurements and its application to the dynamic investigation of [26] [27] [28] [29] [30] [31] [32] [33] machine tools under machining conditions, in: Proceedings of the 10th International MTDR Conference, 1969. I.E. Minis, E.B. Magrab, I.O. Pandelidis, Improved methods for the prediction of chatter in turning, part 1: determination of structural response parameters, Journal of Engineering for Industry 112 (1990) 12–20. O. Özşahin, E. Budak, H.N. Özgüven, Investigating dynamics of machine tool spindles under operational conditions, Advanced Materials Research 223 (2011) 610–621. N. Tounsi, A. Otho, Identification of machine–tool–workpiece system dynamics, International Journal of Machine Tools and Manufacture 40 (2000) 1367–1384. B. Peeters, H. Van der Auweraer, F. Vanhollebeke, P. Guillaume, Operational modal analysis for estimating the dynamic properties of a stadium structure during a football game, Shock and Vibration 14 (2007) 283–303. P. Van Overschee, B. De Moor, Subspace algorithms for the stochastic identification problem, Automatica 29 (1993) 649–660. P.V. Overschee, B. Moor, D.A. Hensher, J.M. Rose, W.H. Greene, K. Train, W. Greene, E. Krause, J. Gere, R. Hibbeler, Subspace Identification for the Linear Systems: Theory–Implementation, Kluwer Academic Publishers, Boston, 1996. B. Peeters, G. De Roeck, Stochastic system identification for operational modal analysis: a review, ASME Journal of Dynamic Systems, Measurement, and Control 123 (2001) 659–667. R. Brincker, P. Andersen, Understanding stochastic subspace identification, in: Proceedings of the 24th IMAC, St. Louis, Missouri, 2006. J. Fan, Z. Zhang, H. Hua, Data processing in subspace identification and modal parameter identification of an arch bridge, Mechanical Systems and Signal Processing 21 (2007) 1674–1689. V. Boonyapinyo, T. Janesupasaeree, Data-driven stochastic subspace identification of flutter derivatives of bridge decks, Journal of Wind Engineering and Industrial Aerodynamics 98 (2010) 784–799. R. Brincker, L. Zhang, P. Andersen, Modal identification of output-only systems using frequency domain decomposition, Smart Materials and Structures 10 (2001) 441–445. W. Heylen, S. Lammens, P. Sas, Modal Analysis Theory and Testing, Katholieke Universteit Leuven, Department of Mechanical Engineering, 1995. R.M. Martinod, G.R. Betancur, L.F.C. Heredia, Identification of the technical state of suspension elements in railway systems, Vehicle System Dynamics 50 (2012) 1121–1135. B. Peeters, H. Van der Auweraer, P. Guillaume, J. Leuridan, The PolyMAX frequency-domain method: a new standard for modal parameter estimation? Shock and Vibration 11 (2004) 395–409. B. Peeters, H. Van der Auweraer, POLYMAX: a revolution in operational modal analysis, in: Proceedings of the 1st International Operational Modal Analysis Conference, Copenhagen, Denmark, 2005. G. Zhang, B. Tang, G. Tang, An improved stochastic subspace identification for operational modal analysis, Measurement 45 (2012) 1246–1256. J.S. Bendat, A.G. Piersol, Random Data: Analysis and Measurement Procedures, Wiley, New York, 2011. P. Avitabile, Modal Space—In Our Own Little World, SEM Experimental Techniques, University of Massachusetts, 2004. M.M. Sadek, Machine Tool Dynamics, Hua Zhong University of Science and Technology Press, Wuhan, 1980. P. Mörters, Y. Peres, Brownian Motion, Cambridge Series in Statistical and Probabilistic Mathematics, Cambridge University Press, Cambridge, 2010. P. Guillaume, P. Verboven, S. Vanlanduit, H. Van Der Auweraer, B. Peeters, A poly-reference implementation of the least-squares complex frequencydomain estimator, the International Modal Analysis Conference, Kissimmee (FL), USA, 2003, pp. 183–192. H. Van der Auweraer, P. Guillaume, P. Verboven, S. Vanlanduit, Application of a fast-stabilizing frequency domain parameter estimation method, ASME Journal of Dynamic Systems, Measurement, and Control 123 (2001) 651–658. P. Verboven, Frequency Domain System Identification for Modal Analysis, Vrije Universiteit Brussel, Belgium250. P. Guillaume, P. Verboven, S. Vanlanduit, Frequency-domain maximum likelihood identification of modal parameters with confidence intervals, in: Proceedings of the 23rd International Seminar on Modal Analysis (ISMA23), Leuven, Belgium, 1998, pp. 359–366.