Você está na página 1de 17

Advanced Engineering Informatics 25 (2011) 259275

Contents lists available at ScienceDirect

Advanced Engineering Informatics


journal homepage: www.elsevier.com/locate/aei

Stochastic model of near-periodic vertical loads due to humans walking


Vitomir Racic *, James Mark William Brownjohn
Department of Civil and Structural Engineering, University of Shefeld, Sir Frederick Mappin Building, Shefeld S1 3JD, UK

a r t i c l e

i n f o

a b s t r a c t
A mathematical model has been developed to generate realistic synthetic vertical force signals induced by people walking. This model is both stochastic and narrow-band, which are the two essential features of walking loading not addressed adequately in the existing design guidelines for pedestrian structures, such as footbridges, long-span oors and staircases. The key reasons for this are (1) the lack of a comprehensive database of walking forces in the form of continuously recorded time series that can be used for development of statistically reliable characterisation of these forces for application in the civil engineering context, and (2) the lack of an adequate modelling strategy which can account for their narrow-band nature. This paper addresses both issues by establishing a large database of measured walking time series recorded by an instrumented treadmill, while the modelling strategy was motivated by the existing numerical generator of electrocardiogram (ECG) signals and speech recognition techniques. Hence, the new approach presented in this paper can serve as a framework for a more thorough and realistic treatment of vertical forces induced by people walking that could be adopted in the design practice. 2010 Elsevier Ltd. All rights reserved.

Article history: Received 28 May 2010 Received in revised form 13 July 2010 Accepted 16 July 2010 Available online 17 August 2010 Keywords: Vibration serviceability Human-structure dynamic interaction Biomechanics Walking Modelling Forces

1. Introduction Predicting vibration performance of civil structures due to loads generated during walking is an increasingly critical aspect of the design process for structures such as footbridges, long-span oors, staircases, operating theatres and vibration-sensitive test and manufacturing facilities [1]. Structures for which this procedure has been inadequately or incorrectly applied may, when constructed, be unt for purpose leading to very expensive remedial measures. For example, rectifying the lateral sway problem of the London Millennium Bridge under pedestrian loading increased the cost of the structure by 30%. However, remedial measures may, sometimes, be even impossible to implement. A typical example is vibration sensitive semiconductor facilities where suspended oors cannot house vibration sensitive equipment, primarily due to failure to cater for walking excitation [1,2]. These and hundreds of similar examples reported in the last decade have increased research interests in vibration serviceability design of the above mentioned ve types of structures which are, by their very nature, occupied and dynamically excited by people walking. However, due to the sheer complexity of the problem, there is still a rm requirement for better understanding of walking loads and developing of realistic mathematical models which can be used in more reliable vibration performance assessment.

Measured continuous force time histories are invariably nearperiodic (Fig. 1a), indicating their narrow-band nature (Fig. 1b and c). However, to simplify analysis and utilisation of the measured forcing time histories in design, they have usually been assumed to be perfectly periodic. This means that actual forces due to continuous walking can be re-created or synthesised, by adding a sequence of identical single footfall traces, temporally displaced by integer multiples of an exact footfall interval (Fig. 2). Based on this assumption, walking force time histories are presentable via Fourier series sinusoids:
m X n1

F t a0

an sin 2ptnfw un

* Corresponding author. Tel.: +44 0 114 222 5727; fax: +44 0 114 222 5700. E-mail addresses: v.racic@shefeld.ac.uk (V. Racic), james.brownjohn@shefeld.ac.uk (J.M.W. Brownjohn). 1474-0346/$ - see front matter 2010 Elsevier Ltd. All rights reserved. doi:10.1016/j.aei.2010.07.004

Here, F(t) is the synthetic forcing function, fw is pacing rate, a0 is a mean value (equivalent to body weight of a pedestrian), an are harmonic amplitudes and un are the corresponding phase angles. The value of m determines the number of harmonics considered in a model. The simplest such model involving a single sinusoidal function can be found in the current UK [3] and Canadian [4] design codes for footbridges, while vibration design guidelines primarily for oors [58] provide for up to four harmonics. Some models even use up to the sixth harmonic [9]. In a time domain analysis procedure that uses such synthetic walking time series, the structure is assumed to be linearly elastic and, using modal decomposition, the response of each vibration mode can be analysed separately using harmonic loads modulated by the appropriate mode shape to account for the moving load [10].

260

V. Racic, J.M.W. Brownjohn / Advanced Engineering Informatics 25 (2011) 259275

Nomenclature Ti cycle intervals normalised cycle intervals ASD of si Fourier amplitudes of si Gaussian heights (weights) Gaussian centres bj, dj, bj Iw,i DIw,i q0, q1 Gaussian widths weight normalised impulses disturbance term coefcients of linear regression angular frequency of rotation shapes of successive cycles

si
Ss(f) As Wj, Aj cj, tj, hj

xi
Zk(t)

Phase angles have been found to vary considerably between various measurements [11,12], indicating that their random nature should be considered when generating synthetic walking loads. However, little utilization of this important nding has been made to date. The values of normalised coefcients an = an/a0 are often reported as dynamic load factors (DLFs) and depend on walking speed and individual characteristics of a human test subject [1].

The DLFs for walking, running, and jumping are presented by Rainer et al. [13] for three subjects, while a more comprehensive study of walking has been undertaken by Kerr [14]. His study found wide variability of DLFs among subjects participating in the tests (intersubject variability) and also among different tests with the same subject (intra-subject variability). Based on Kerrs data, Young [15] described this variability by frequency-dependent mean and

1200

a)

total force

right footfalls

left footfalls

Force amplitudes [N]

1000 800 600 400 200 0 0 2 4 6 8 10 Time [s] 12 14 16 18 20

Fourier amplitudes/ body weight [-]

0.25 b) 0.20 0.15 0.10 0.05 0 0 0.5fw fw 1.5fw 2fw 3fw 2.5fw 3.5fw Frequency/walking rate fw [-] 4fw 4.5fw 5fw

4 3 c)

Fourier phases [rad]

2 1 0 -1 -2 -3 -4 1 20 3 4 5 6 Frequency [Hz] 7 8 9 10

Fig. 1. (a) An example of continuously measured walking force signal and (b and c) the corresponding Fourier amplitudes and phases derived from a window comprising 64 successive footfalls.

V. Racic, J.M.W. Brownjohn / Advanced Engineering Informatics 25 (2011) 259275

261

1200

a) 1000
Force [N]

total force

right footfalls

left footfalls

800 600 400 200


0 0 0.2
0.4 0.6 0.8 1 Time [s] 1.2 1.4 1.6 1.8

Fourier amplitudes/ body weight [-]

0.25 b) 0.20 0.15 0.10 0.05 0

0
4

fw

2fw

3fw Frequency/walking rate fw [-]

4fw

5fw

c)

Fourier phases [rad]

2 0 -2 -4 0 2 4 Frequency [Hz] 6 8 10

Fig. 2. (a) Reconstruction of walking force time history by replicating a single footfall, and (b and c) the corresponding Fourier amplitudes and phases derived from a window comprising 64 successive footfalls.

coefcient of variation for DLFs. The mean DLFs dened in this way gave rise to the current state-of-the-art mathematical models featuring deterministic approach to describe walking loads (i.e. there is a uniform force model for all individuals within the global human population). Numerous studies have shown the Fourier approach to be inappropriate for real situations of single and multiple occupants because it leads to signicant loss of information and introduction of inaccuracies during the data reduction process [16,17,12]. Brownjohn et al. [16] showed that there were signicant differences between the resonant responses due to the imperfect real forces and the equivalent perfectly periodic simulation. These differences are most signicant (1) for higher Fourier harmonics where the simulated vibration response is regularly overestimated and (2) for resonant vibrations of lightly damped structures where the simulated response is often underestimated. In the former case the differences are caused by unnaturally high DLF values of periodic force representation, whereas the latter case is a direct result of neglecting energy of the actual walking excitation around dominant harmonics (Fig. 1b). However, excessive structural vibrations due to walking are not necessarily resonant. For instance, in the case of high-frequency oors (i.e. when the natural frequency of the fundamental mode of vibration is far above pacing rate), strong vibration response

due to people walking typically shows a transient character [7]. Such response is induced by high-frequency components of pedestrian dynamic loading generated when the heel strikes the walking surface [1]. As each following heel strike is made, another transient decay is generated, but no build-up of vibrations can be developed over several oscillations between footfalls due to the damping [2]. The difference in the nature of resonant and transient vibration responses has led to a requirement for two conceptually distinct walking force models used in design guidelines of low and highfrequency structures. For oors, the most up to date guidance is available in Appendix G [7] of the Concrete Society Technical Report 43 (CSTR43). It provides design values of DLFs for single pedestrians for low-frequency oors (i.e. if natural frequency of the fundamental mode is below 10 Hz) and equivalent impulse values for high frequency oors (i.e. if the fundamental natural frequency is above 10 Hz). However, this articial division proves unreliable when a oor has strong responses in both the low and high-frequency region, which clearly indicates the need for a uniform model capable of taking into account the complete frequency content of walking forces [2]. In an attempt to overcome this problem, Zivanovic and Pavic [18] merged together the Appendix G impulses and their previously published low-frequency force model based on the Fourier approach [11]. The new model takes into account the differences in the walking force induced by different

262

V. Racic, J.M.W. Brownjohn / Advanced Engineering Informatics 25 (2011) 259275

people and, as a result, can estimate the probability distribution of vibration responses generated by a pedestrian population. However, Middleton showed recently that inter-subject variations do not affect the dynamic response as much as variations in an individuals pace rate for successive steps [19]. In the case of high-frequency oors, a lack of these variations can overestimate response up to 40%. A number of authors [20,16] showed that the actual narrow-band effect seen for real loads was more suited to the techniques of stochastic analysis in frequency domain, where the measured forces are presented as auto-spectral densities (ASDs). The resulting response prediction (typically acceleration) is necessarily given as a root mean square (RMS) value, and this is consistent with commonly used vibration acceptance criteria [7]. However, this single number contains no information about expected performance of the structure in real time while human actions are taking place (e.g. where the peak responses are and how often they happen, as it can be described by peak and crest factors). For such advanced analysis, a reliable time domain model of walking forces is clearly the way forward. The most comprehensive dataset on walking forcing functions available worldwide currently is based on data from approximately 1000 individual time series generated by about 40 individuals collected by Kerr [14] as they passed over a force plate and made a single footfall. However, apart from the limited number of test subjects, the key problem with this data set is that it cannot represent the variability of real walking between subsequent steps of the same pedestrian during one walking exercise. Gait, or manner of walking, is the subject of considerable research in the biomechanics community [1]. However, no database of walking forces in the form of continuously recorded time series has yet been found that can be used for development of statistically reliable characterisation of these forces for application in a civil engineering context. By combining multidisciplinary experimental and analytical approaches, this paper advances the eld by bringing together: (1) an extensive database of continuously measured vertical walking forces generated by individuals, and (2) a unique stochastic model of these forces which can simulate accurately what is being measured. This includes modelling of near-periodic features of actual forces, such as variations in timing, shape and amplitudes, as well as the complete frequency spectrum. An instrumented treadmill for experimental identication of the forces was transferred and adapted from the biomechanical eld of human locomotion [21], while the modelling strategy was motivated by existing models for replicating human electrocardiogram (ECG) signals [22] and speech recognition [23]. Therefore, the paper presents a new generation of synthetic walking loads which is a radical departure from the conventional models used in the current design guidelines for pedestrian structures and which, more importantly, can simulate reality better. 2. Experimental data acquisition A database comprising many high-quality walking force records is an essential component for the development of a stochastic model of individual walking loads. Section 2.1 aims to justify the choice of equipment used in this study to build such a database. This is followed by a brief description of the experimental setup in Section 2.2 and details on test protocol in Section 2.3, so that it is clear how the data was collected. 2.1. Instrumentation for measuring contact forces during continuous walking Dynamic forces generated by people walking on a structure are ideally determined by direct measurement of the contact forces

between the feet and the structure itself, hence they are generally known as ground reaction forces (GRFs). For this purpose Rainer et al. [13] used a continuously measured reaction of a oor strip having known dynamic properties, whereas Ebrahimpour et al. [24] used a short length of walkway instrumented with several force plates. However, such an experimental setup can provide force records for a limited number of successive footfalls and requires considerable laboratory space. One solution to overcome these drawbacks is to use an instrumented force measuring treadmill (IFMT). It requires less laboratory space and enables quick collection of GRFs during a large number of successive cycles and over a wide range of steady-state gait speeds [25]. Although there is no theoretical difference in the physics of treadmill and overground locomotion, some early experimental studies based on very rudimentary treadmill designs reported signicant discrepancies between overground and by treadmill measured GRFs [26], as well as with temporal and spatial parameters of gait such as step length and step frequency [27]. The reason for this was thought to be imposed constant speed of movement, controlled by rotation of the treadmill belt, as opposed to the overground locomotion when the gait speed has no similar control [28]. More recently, using their advanced treadmill design [29,30] performed a series of experiments to make more accurate comparisons between treadmill versus overground walking. Although it was possible to detect subtle differences in kinematics and kinetics of human gait, the magnitudes of these differences were within a normal range of variability of the gait parameters in the biomedical domain. Interestingly, van de Putte et al. [31] reported that differences between treadmill and overground walking vanished after only ten minutes of practice of treadmill walking in the case of inexperienced treadmill users. This was concluded based on comparison between three-dimensional knee kinematics and step lengths recorded on ten healthy test subjects as they walked on a treadmill and free-eld. These studies demonstrated the essential equivalence of treadmill and overground gait in biomedical application, such as design of prosthetic legs and articial hip joints. Hence, they justied treadmill-based force measurements used in this paper for design of less delicate civil engineering structures. 2.2. Experimental setup All walking tests were carried out in the Light Structures Laboratory in the University of Shefeld. Continuously measured force records were collected of one person walking at a time on the state-of-the-art instrumented treadmill ADAL3D-F [32]. The treadmill is installed in a recessed pit and ush with the laboratory oor, as illustrated in Fig. 3. To measure independently the left and right footfalls, ADAL3D-F design includes two identical IFMTs, one for the left and one for the right side of the walking area (Fig. 3). All components (including a motor, belt and secondary elements) of each treadmill are mounted on a single metal frame and mechanically connected to the supporting ground only through a pair of Kistler 9077B three-axial piezoelectric force sensors [33]. Therefore, assuming the rigidity of the above mentioned ensemble, the entire treadmill is mechanically isolated. This means that the forces due to belt friction and belt rotation can be considered as internal forces and are not detected by the sensors [21]. The stiffness constants for the force sensors are large enough that natural frequencies of the supported treadmill frame are very high compared to the bandwidth of signicant walking forces. Thus, the sensors measure only external forces, i.e. actual three-axial forces exerted by the feet on the treadmill belts. Only the vertical time series are considered here (Fig. 1a), while lateral and longitudinal components are beyond the scope of this paper. Each treadmill belt is driven by a brushless servomotor equipped with internal velocity controllers to maintain the speed

V. Racic, J.M.W. Brownjohn / Advanced Engineering Informatics 25 (2011) 259275

263

minutes long warming up, which included walking on the treadmill while the walking speed was varied randomly and controlled by the speed of rotation of the treadmill belts. As already explained in the previous section, this interval was reported as long enough for pre-test habituation to treadmill walking, when differences between treadmill and overground walking vanish even for inexperienced test subjects [31]. The acquisition of walking forces started at a speed of 2 km/h and continued in increments of 0.5 km/h up to the maximum comfortable walking speed, which was determined by each test subject individually. This speed differs between people and depends primarily on the length of a test subjects legs and therefore their overall height [34]. Pacing rate was not prompted by any stimuli such as a metronome, and it was determined only from subsequent analysis of the generated force signals. Each test followed the same pattern whereby a participant was rst asked to stand still on the treadmill, and then: (1) was informed when the treadmill belts would start rotating at a particular constant speed, (2) was asked to walk until at least 64 successive footfalls were recorded, (3) was informed when the treadmill will be stopped, (4) was asked to keep walking until the treadmill came to a halt, and (5) was asked to leave the treadmill and rest. In total, 80 volunteers (55 males and 25 females, body mass 76.1 15.3 kg, height 174.8 8.1 cm, age 29.7 9.3 years) were drawn from students, academics and technical staff of the University of Shefeld. On average, forces corresponding to ten different walking speeds were collected for each test subject depending on their maximum comfortable walking speed. All together they generated 824 vertical walking force time series of kind illustrated in Fig. 1a. Each walking time series was sampled at 200 Hz.

Fig. 3. Experimental setup.

as constant as possible. Velocity of treadmill belts in the range 0.1 10 km/h can be controlled and monitored remotely either with a control panel or with ADAL3D-F software, Adisoft2000 [32], run from the data acquisition PC. Similar to tness treadmills, the remote control panel and the treadmill itself are equipped with a safety stop switch. 2.3. Test protocol Prior to the measurements, a test protocol approved by the Research Ethics Committee of the University of Shefeld required each participant complete a Physical Activity Readiness Questionnaire and pass a preliminary tness test (by satisfying predened criteria for blood pressure and resting heart rate) to check whether they were suited for physical activity required during the measurements. Also, measurements of the body mass, age and height were taken for each test subject who passed the preliminary test. All participants wore comfortable at sole shoes. Participants who had no experience with treadmill walking were given a brief training prior to the measurements supervised by a qualied instructor. All test subjects did approximately ten

3. Modelling individual force signals This section describes the concept of the new model development, from analysis of continuous walking parameters of a single force record to its synthetic counterpart, so the reader can follow the rationale for the remaining parts of the paper. In Section 4 it will be applied to each force record in the database presented in Section 2 yielding a mathematical generator of stochastic nearperiodic walking time series.

1.6 total force 1.4 1.2


Force/body weight [-]

right footfalls

left footfalls

1 0.8 0.6 0.4 0.2 0 0 0.5 1 1.5 Time [s] 2 2.5 3 footfall Ti area = Iw,i Ti+1 walking cycle = stride

Fig. 4. A portion of a continuously measured GRF due to walking. The total duration of the signal is 40 s.

264

V. Racic, J.M.W. Brownjohn / Advanced Engineering Informatics 25 (2011) 259275

3.1. Basic processing of measured force records Fig. 4 illustrates a portion of a continuously measured force time history generated by a single person walking at 5 km/h and 1.81 Hz pacing rate. The force signal was ltered using a fourth-order low-pass Butterworth digital lter having cut-off frequency at 60 Hz. It is apparent that there is some variation in shape of footfall forces (i.e. pulses), intervals and force amplitudes on a step-by-step basis. IFMTs are normally used in biomechanics of human gait to provide a series of statistics or gait parameters describing footfalls [1]. However, data relevant for structural analysis concerning intra-subject variations in timing of the individual footfalls and their amplitudes are not readily available from the literature. Closer inspection also reveals differences between the left and right footfalls. For instance, the left footfalls have greater rst peak of the characteristic M shape. Sahnaci and Kasperski [35] showed that this asymmetry in walking induces intermediate harmonic load amplitudes (so called sub-harmonics which are illustrated in Fig. 1b) which may lead to larger vibration responses than higher integer harmonics for some test subjects. Therefore, in this study the total vertical force will be modelled on a stride-by-stride basis, where a stride (i.e. walking cycle) consists of two successive footfalls, left and right (Fig. 4). A walking cycle can be dened between any two nominally identical events of the total force signal. Since the amplitudes of the total vertical forces oscillate around body weight of the test subject, the point close to the beginning of the left footfall where amplitude of the total force is equal to body weight was selected as starting (and completing) event (Fig. 4). From the 40 s long force signal illustrated in Fig. 4 yielding about 72 footfalls, a window comprising 32 successive cycles (64 footfalls) was extracted from the middle of the signal for further analysis. Fig. 5 shows values of normalised impulses Iw,i as a function of the cycle intervals Ti. The normalised impulse is dened as the cycle impulse (i.e. the integral of the force over Ti) divided by test subject body weight. The apparent linear trend between these two parameters offers an opportunity to describe weight normalised impulses as a function of the cycle intervals. However, bearing in mind the denition of a cycle impulse, the linear trend does not indicate that the peak force amplitudes of successive cycles depend on the corresponding cycle intervals. This is an important feature for the proposed methodology, as will be shown later in Section 3.5.

3.2. Variability of cycle intervals Variations of intervals Ti (i = 1, . . . ,32) can be represented by a sequence of dimensionless numbers calculated as:

si

T i lT 2

lT lT meanT i

Three methods were tried for modelling si data. Histogram of si amplitudes (Fig. 6) indicated that this data set does not follow any known probability distribution. Small number of data points could be argued, but the same was found for the variations of twice as much footfall intervals. Next, it was assumed that the current si value is a linear combination of previous k values si1, . . . ,sik, which could be described by an autoregressive model [36]. This method proved unreliable since weak correlation was found between the previous and subsequent si values for increasing order of regression from one to ve (k = 1, . . . ,5). Finally, utilisation of the autospectral density (ASD) of a random process resulted in a success. The variance r2 x of a real random process x(t) can be obtained as the total area under its ASD Sx(f) as [37]:

r2 x

Z
0

Sx f df

Hence, the aim is to use ASD of actual si series to generate synthetic s0i series with the same statistical properties, such as standard deviation and structure of variation between successive si values. Under assumption that the variation of si does not change for the given test subject, pacing rate and duration of walking (e.g. caused by fatigue and physical discomfort), this applies regardless of the number of discrete data points in the measured and articial data sets, which will be demonstrated later in this section. The ASD of si can be calculated as:

A2 fm Ss fm s ; 2 Df

fm

m ; 32

m 0; . . . ; 15

where A2 s fm is a single-sided discrete Fourier amplitude spectra having spectral line spacing Df 1=32 (Fig. 7). The ASD ordinates do not depend on the number of discrete data points si used for the calculation of As(fm) via Fast Fourier Transform (FFT) but it is coarse due to the limited number of points

7 1.16 measured data linear fit 6

Number of occurances [-]

1.15

Normalised impulses Iw,i [-]

5 4 3 2 1 0

1.14 1.13 1.12 1.11 1.10 1.09 1.08 1.07 1.07 1.08 1.09 1.10 1.11 1.12 1.13 1.14 Cycle intervals Ti [s] 1.15 1.16

-0.040

-0.036

-0.022

-0.015

0.015

Normalised cycles intervals

Fig. 5. Linear trend between periods Ti and weight normalised impulses Iw,i.

Fig. 6. Histogram of si values.

0.022

0.029

0.036

0.040

V. Racic, J.M.W. Brownjohn / Advanced Engineering Informatics 25 (2011) 259275

265

Cycle intervals Ti [s]

(16 single-sided FFT). More points might reveal a much richer structure but this requires a longer walking force signal. However, concerning ethics in research, the duration was limited to avoid fatigue and discomfort, which can also inuence natural variability of the force records. The ASD Ss(fm) can be analytically represented by a series of Gaussian functions (Fig. 7):

1.15 1.14 1.13 1.12 1.11 1.10 1.09 1.08 measured 1.07 1 5 10 15 20 Cycles [-] 25 30 32 synthetic

S0s f

16 X j 1

f cj W je
2b 2 j

here, parameter Wj is the height of the jth Gaussian peak, cj is the position of the centre of the peak, and bj controls the width (i.e. time duration) of the corresponding Gaussian function. The Gaussian centres cj, j = 1, . . . ,16, are placed in each sample on the quasi-frequency axis in order to t exactly the measured ASD (Fig. 7). For such xed positions of cj and predened widths bj = Df, Gaussian heights Wj (also called weights) can be computed using the non-linear least-square method [38]. Representation of the discrete ASD Ss(fm) in the form of the continuous function S0s f enables generation of synthetic series s0i i 1; . . . ; N of arbitrary length (e.g. N ) 32), which will have statistically the same properties of variations on the sample-bysample basis as the measured set of 32 actual si data points. This is done by calculating a new set of ASD amplitudes for a sequence of discretely spaced frequency points fn = nDf0 using Eq. (5), where n = 0, . . . ,N/2 1 and Df0 = 1/N. The new set of ASD amplitudes S0s fn and Eq. (4) are then used to generate a new set of Fourier q amplitudes A0s fn 2Df 0 S0s fn . Finally, assuming random distribution of phase angles in the range [p, p], these amplitudes are used in inverse FFT algorithm to generate a synthetic set of N variations s0i . Different realisations of the random phases may be specied by varying the seed of the random number generator, hence many different series s0i may be generated with the same spectral properties (Fig. 8). According to Eq. (2), scaling s0i by lT and adding the offset value lT calculated from the measured Ti data, results in a series of synthetic intervals T 0i , as would be generated by the test subject during nominally identical walking exercises. Empirical evidence for this is presented in Section 3.5.

Fig. 8. Measured si and an example of synthetic

s0i series.

3.3. Variability of normalised impulses The measured weight normalised impulses Iw,i can be expressed as a function of the cycle intervals Ti (Fig. 9) using the following linear regression model [39]:

Iw;i q1 T i q0 DIw;i

Here, q1 = 1.05 and q0 = 0.05 are regression coefcients and DIw,i is the subsequent error (also known as a disturbance term), which is a random variable. Given the regression coefcients q1 and q0, DIw,i, can be calculated as:

DIw;i Iw;i q1 T i q0

It is common to model DIw,i as a random noise having a Gaussian distribution [39]. Given a set of synthetic cycle intervals T 0i generated as explained in Section 3.2, a series of the corresponding synthetic normalised impulses I0w;i can be calculated using Eq. (6). These are used in Section 3.5 to scale amplitudes of successive strides, thus make a smooth transition of walking energy between successive cycles.

x 10 12 10 8
i [-]

-4

0.5 measured fit 0.4 0.3

Normalised force [-]


c1 0 c4 0.1 c8 0.2 0.3 Quasi-frequency [-] c12 0.4 c16 0.5

0.2 0.1 0 -0.1 -0.2

6 4 2 0

ASD of

-0.3 -0.4 0 0.2 0.4 0.6 Tme [s] 0.8 1 1.2

Fig. 7. Single-sided ASD of si values.

Fig. 9. Normalised walking cycles.

266

V. Racic, J.M.W. Brownjohn / Advanced Engineering Informatics 25 (2011) 259275

3.4. Template waveform of walking force signals The 32 normalised walking cycles have been detrended (body weight has been removed), resampled to the maximum length of all the cycles and aligned with respect to the common time axis, as shown in Fig. 9. Visual inspection suggests that there is a common waveform of the walking force which distorts along time and amplitude axes on a cycle-by-cycle basis. The key challenge here is to extract this waveform and provide its reliable mathematical representation which can be used to generate synthetic time series. The average cycle is considered unreliable representative due to misalignments and shifts of the common events, such as positions of the local extreme values. To correct for these slight variations, a procedure called dynamic time warping (DTW) is applied to the measured cycles before their average is calculated. DTW was originally developed for speech recognition [23], but in recent years it has found application in a ngerprint verication [40], chromatography [41] and in gene expression studies [42]. The procedure nonlinearly warps the two trajectories in such a way that similar events are aligned and the summed squared differences are minimised [23]. This process is illustrated schematically in Fig. 10a. Similarly, wavelet analysis linearly shifts and stretches the original (or mother) wavelet. However, due to the apparent nonlinear misalignments of the common events in the force data, linear shifting and stretching is unsuitable for determining the template cycle. Since the DTW method warps two cycles at one time, the next task is to choose the reference cycle and then warp all the other cycles onto that one. Such a cycle minimises the sum of Euclidiean distances between each data point to the average of all the cycles [43]. When the warping has been completed and the common events in the cycles aligned, the template cycle is formed as average of the warped cycles, as illustrated in Fig. 10b. For fs = 200 Hz sampling rate and interval of 1.15 s, the underlying shape of the template cycle (Fig. 10b) can be modelled mathematically as a sum of 115 Gaussian functions Z(t):
115 X j

0.4 0.3

Normalised force [-]

0.2 0.1 0 -0.1 -0.2 -0.3 template -0.4 0 0.2 0.4 0.6 Tme [s] 0.8 1 1.2 Gaussians fit

Fig. 11. Fitting the template cycle with a series of Gaussian basis functions.

the normalised force amplitudes, thus to reect completely the corresponding Fourier amplitude spectrum. A less dense distribution of Gaussian functions (e.g. centres are placed in every third or fourth sample) makes the t smoother causing the high-frequency components to vanish. For such xed positions of Gaussian centres and predened widths dj = 2Dt, Gaussian heights Aj can be optimised using non-linear least-square curve t [38]. Scaled in the vertical and horizontal direction, analytical functions Z(t) can be used to generate articial force signals which reect the variations of impulses and intervals for successive walking cycles. This is demonstrated in the next section. 3.5. Dynamic model A repetitive walking pattern could be better visualised when the template cycle is wrapped around the surface of a circular cylinder (Fig. 12). Now the rst and the last sample of the template overlap, thus the corresponding trajectory becomes a closed orbit in three-dimensional (3D) space with coordinates (r, h, z). A synthetic time series is generated by retracing the t Z(t) around the circle of unit radius in (r, h) the plane (Fig. 12). Each revolution on this circle corresponds to one interval of walking cycle. The equations to generate synthetic forces are therefore given by a set of three coupled equations:

t t j
2d2 j

Z t

Aj e

where the parameter Aj is the height of the jth Gaussian peak, tj is the position of the centre of the peak, and dj is the width of the corresponding Gaussian functions in the sum (Fig. 11). The Gaussian centres tj = nDt (n = 1, . . . ,115, Dt = 2/fs) are placed in every second sample on the time axis to t accurately

0.5 a) 0.4 0.3 reference trajectory warping trajectory

0.5 b) 0.4 0.3 template warped signals

Normalised force [-]

0.2 0.1 0 -0.1 -0.2 -0.3 -0.4 0 0.2 0.4 0.6 Tme [s] 0.8 1 1.2

Normalised force [-]

0.2 0.1 0 -0.1 -0.2 -0.3 -0.4 0 0.2 0.4 0.6 Tme [s] 0.8 1 1.2

Fig. 10. (a) Schematic illustration of dynamic time warping. (b) Template cycle.

V. Racic, J.M.W. Brownjohn / Advanced Engineering Informatics 25 (2011) 259275

267

r
j

Fig. 12. Trajectory of the template cycle in a three-dimensional (3D) space.

r t 1 ht xt Z h
115 X j

hhj
2b2 j

Aj e

of unit amplitude to each data point. The resulting signal is an articial, weight normalised walking force time series. Measured continuous walking series and an example of its synthetic counterpart are shown in Fig. 13a and b, respectively. Visual comparison suggests that the uniform scaling of the template cycle is not the way to model inter-cycle variations of the walking force amplitudes. However, no model has been found to describe variations of the Gaussian heights Aj as a function of variations in the cycle intervals. Existence of such a model was already questioned in Section 3.3 due to the strong linear correlation between the cycle intervals and the normalised impulses (Fig. 5). Fig. 14 shows the discrete Fourier amplitude spectra of the measured and synthetic signals. To avoid leakage during the digital signal analysis, the duration of data blocks used to calculate the spectra was xed at exactly 32 cycles. While the dominant harmonics (i.e. integer multiples of the pacing rate fw = 1.81 Hz) are modelled reliably, there is less spread of energy to adjacent spectral lines of synthetic data through lack of inter-cycle variability of force amplitudes. The effect is more prominent for higher harmonics, which can be illustrated better by comparing quasi steadystate structural responses of a single degree of freedom (SDOF) oscillator after the initial transient build up. 3.6. Quantication of structural response and model validation The blue line in Fig. 15 shows response to the measured walking time series for a 1000 kg oscillator with 2% damping and natural frequencies of 1.81, 3.62, 5.43, and 18.1 Hz. These frequencies f correspond exactly to the fundamental and higher harmonics of the walking excitation. The response of the same oscillators to the synthetic excitation (black lines) is given below the response to real walking for comparison. Visually, there is no signicant difference between the two responses when the natural frequency of the oscillator exactly equals the mean pacing rate (f = fw). The comparison also demonstrates that it is possible to build-up to a signicant percentage of resonant response to the fundamental component of the walking force even with real, more variable walking. For the second and third harmonics, the difference is much clearer, particularly for the third harmonic. On average, the SDOF

where x is angular frequency of rotation and bj = xdj in rads. The time positions of the Gaussian centres tj in Eq. (8) now correspond to xed angles hj = xtj around the unit circle, as illustrated in Fig. 12. Variations in the length of successive intervals can be incorporated by varying angular frequency of rotation x. Its values for successive revolutions can be calculated as:

x0i

2p T 0i

10

where T 0i i 1; . . . ; N are synthetic intervals of walking cycles generated as explained in Section 3.2. Moreover, having T 0i series, values of the corresponding synthetic normalised impulses I0w;i can be calculated by using Eq. (6). These are assigned to the synthetic time series generated by Eq. (9) by multiplying amplitudes of successive cycles by appropriate scaling constants and adding an offset

1.5 1.4 1.3 1.2 1.1 1 0.9 0.8 0.7 0.6 1.5 1.4 1.3 1.2 1.1 1 0.9 0.8 0.7 0.6 0

a)

Force/ body weight [-]

b)

Force/ body weight [-]

10

15

20 Time [s]

25

30

35

40

Fig. 13. (a) Measured and (b) an example of synthetic walking time series generated using the template cycle.

268

V. Racic, J.M.W. Brownjohn / Advanced Engineering Informatics 25 (2011) 259275

0.25
Fourier amplitudes/ body weight [-]

a)

0.20 0.15 0.10 0.05 0 0.25 b)

Fourier amplitudes/ body weight [-]

0.20 0.15 0.10 0.05 0 0

8 10 Frequency [Hz]

12

14

16

18

Fig. 14. Discrete Fourier amplitudes of (a) measured and (b) synthetic walking time series. Figs. 13 and 14 represent the same data.

systems have greater response under the excitation of a synthetic walking. However, occasionally the response due to real excitation exceeds its less variable synthetic counterpart. Also note that after the transient build up to a signicant proportion of resonant maximum amplitude, the response fades. A number of authors [44,17] observed this phenomenon with pedestrians exciting very lively footbridges in anti-symmetric vertical modes. When the pacing of a pedestrian becomes out of phase with the vibration of a footbridge structure excited close to resonance, energy is extracted from the structure and the human exciter becomes an active damper. More recently, the authors offered more formal proof of this happening by studying energy ow between active humans and a lively structure due to bouncing and jumping [1]. Similarly to the rst harmonic, there is little to choose between the two responses when the natural frequency of the oscillator is ten times higher than the mean pacing rate (f = 10fw). However, it is worth noticing that occasionally the real response exceeds it synthetic counterpart. Peak vibration amplitude is a commonly used vibration acceptance criteria for oors accommodating sensitive equipment, such as lasers in operating theatres [2]. Further parametric studies can demonstrate that the differences reduce with increase of damping of the SDOF systems studied but are still signicant even for damping as high as 3% found in modern ofce oors with full-height partitions. To get more reliable prediction of vibration response, a logical way forward is development of a more comprehensive model of inter-cycle variations of the force amplitudes which will account for non-uniform scaling. Such a model is presented in the next section.

angular frequencies x0i generate the cycles faster, hence they compress them and vice versa. As shown in Section 3.5, by multiplying amplitudes of each cycle by appropriate scaling constants and adding the offset value, the resulting time series are assigned the corresponding I0w;i values. The similarity between the real measured and synthetic walking forces is illustrated by comparison of Fig. 16a and b. Also, the standard Fourier amplitude spectra are compared in Fig. 17, while vibration responses for the SDOF systems dened in the previous section are shown in Fig. 18. This comparison looks much better than in Sections 3.5 and 3.6 between actual and uniformly scaled amplitudes of synthetic forces. Perfectly identical synthetic signals can be generated only by chance due to the omnipresent randomness of the modelling parameters. However, a family of synthetic forces share several properties inherited from their common measured counterpart: (1) Shapes of the walking cycles are drawn from the same source (i.e. Zk(t) functions), where each shape has the same probability of occurrence. (2) Variations of successive cycle intervals T 0i are the same for all synthetic signals because they have the same ASD. (3) The statistical equivalence between T 0i values reects directly equivalence between the corresponding weight normalised impulses I0w;i according to Eq. (6). This implies that total energies of generated signals are statistically the same. 4. Development of stochastic walking model This section integrates everything presented so far to create a numerical generator of stochastic near-periodic walking force signals. Section 4.1 deals with organisation of the modelling parameters extracted (as explained in Section 3) from the numerous database of the individual force records collected in Section 2. This is followed by a description of the mechanism underlying generation of random walking loads in Section 4.2. 4.1. Utilisation of existing database The 824 force signals measured in Section 2 were classied into 20 categories (clusters) with respect to the actual pacing rate, as

3.7. Variability of amplitudes of walking cycles Each of the 32 signals shown in Fig. 9 is tted to Eq. (8) yielding independent analytical functions Zk(t), k = 1, . . . ,32. Under assumption that the variations of cycle intervals and morphology of the corresponding force signals do not correlate, the Zk(t) shapes can be assigned in a random order to synthetic walking cycles. Even when the same Zk(t) is assigned to different cycles, the resulting signal generated by the coupled system of Eq. (9) will be slightly different. This is because there are two more random parameters x0i , and I0w;i , which inuence morphology of the signals. Higher

V. Racic, J.M.W. Brownjohn / Advanced Engineering Informatics 25 (2011) 259275


-3

269

x 10 6

a)

-6 6

b)

-6 2 c)

-2 2 d)

Acceleration/ body weight [1/kg]

-2 1.5 e)

-1.5 1.5 f)

-1.5 0.8

g)

-0.6 0.8 h)

-0.6 0 5 10 15 20 Time [s] 25 30 35 40

Fig. 15. Comparison of responses of (a and b) 1.81 Hz oscillator, (c and d) 3.62 Hz oscillator, (e and f) 5.43 Hz oscillator and (g and h) 18.1 Hz oscillator to real (blue) and synthetic (black) force time series of walking at mean pacing ratefw = 1.81 Hz. (For interpretation of the references to colour in this gure legend, the reader is referred to the web version of this article.)

270

V. Racic, J.M.W. Brownjohn / Advanced Engineering Informatics 25 (2011) 259275

1.5 1.4 1.3 1.2 1.1 1 0.9 0.8 0.7 0.6 1.5 1.4 1.3 1.2 1.1 1 0.9 0.8 0.7 0.6 0

a)

Force/ body weight [-]

b)

Force/ body weight [-]

10

15

20 Time [s]

25

30

35

40

Fig. 16. (a) Measured and (b) synthetic walking time series which accounts for inter-cycle variations of force amplitudes.

0.25

a)

Fourier amplitudes/ body weight [-]

0.20 0.15 0.10 0.05 0

0.25 b)

Fourier amplitudes/ body weight [-]

0.20 0.15 0.10 0.05 0 0 2 4 6 8 10 Frequency [Hz] 12 14 16 18

Fig. 17. Discrete Fourier amplitude spectra of the signals shown in Fig. 16.

shown in the histogram in Fig. 19. Each bar in the histogram is approximately 0.1 Hz wide. For example, all force records with the pacing rate in the range 1.581.68 Hz are gathered into one cluster. Such ne resolution is a key feature for development of a stochastic model outlined in the next section. Note that walking at different speeds does not necessarily yield different pacing rates. This is because people unconsciously control their step length to walk at their most comfortable pacing rate [45]. In the present study, the cluster for the rates between 1.88 and 1.98 Hz is the most numerous comprising 101 different walking force records. This might be an indicator that this is the range of the most comfortable pacing rates for the majority of participants in this study.

Each force history within a cluster was processed using the concept described in Section 3. Multiple sets of information (parameters of the Gaussian ts Zk(t) for and the ASD S0s f , autoregression coefcients q1 and q0 and the parameters of the Gaussian random noise) were extracted and stored within the cluster as MATLAB structure les [46]. This is an important aspect of the stochastic approach adapted in this paper, as it will be demonstrated in the next section. The key challenge now is to show that the force amplitudes are independent of body weight, so that weight normalised synthetic forces of kind shown in Figs. 13b and 16b can be scaled by the weight of any person drawn by chance from the worlds population. By doing this, randomisation of the modelling parameters will

V. Racic, J.M.W. Brownjohn / Advanced Engineering Informatics 25 (2011) 259275


-3

271

x 10 6

a)

-6

b)

Acceleration/ body weight [1/kg]

-2 1.5

c)

-1.5

0.8

d)

-0.6 0 5 10 15 20 Time [s] 25 30 35 40

Fig. 18. Dynamic responses of (a) 1.81 Hz oscillator, (b) 3.62 Hz oscillator, (c) 5.43 Hz oscillator and (d) 18.1 Hz oscillator to synthetic force time series shown in Fig. 16b (mean pacing rate is fw = 1.81 Hz).

110 100 90

Number of occurrences [-]

80 70 60 50 40 30 20 10 0 1 1.5 2 Pacing rate [Hz] 2.5 3

and fast pacing rates. The scattered patterns for all three clusters indicate no correlation between the two parameters, thus they can be treated as independent variables. Furthermore, body weight can be modelled as a random parameter using a probability density function, as suggested by Hermanussen et al. [47] for German, Austrian and Norwegian citizens.

4.2. Procedure for generating synthetic forces The stochastic approach adopted in this paper rests on the assumption that a set of the modelling parameters selected randomly and equally likely from a cluster can be used to synthesise walking time series at any pacing rate which falls within the clusters narrow frequency range. This means that the modelling parameters extracted from real walking at particular rate can be used to generate synthetic walking forces at rates which are close enough. The ow chart in Fig. 21 illustrates the complete process (algorithm) of creating synthetic GRF signals. For a specied pacing rate and duration of walking, the algorithm rst estimates the total number of walking cycles N. Then, from a frequency cluster corresponding to the pacing rate, it selects by chance a set of multiple modelling parameters. At this point, the algorithm splits into two actions. First, it creates intervals T 0i using the ASD S0s f (Section 3.2) and then calculates normalised impulses I0w;i using the regressive

Fig. 19. Histogram of pacing rates yielding 20 clusters.

be extended to the maximum. The evidence to support this hypothesis is given in Fig. 20, where DLF of the fundamental harmonic is plotted versus the corresponding body weight for all signals in clusters 1.481.58 Hz, 1.881.98 Hz and 2.282.38 Hz. These clusters are selected as representatives of slow, moderate

272
0.14 a) 0.12 0.10

V. Racic, J.M.W. Brownjohn / Advanced Engineering Informatics 25 (2011) 259275


0.40 b) 0.35 0.30 0.40 0.25 0.08 0.20 0.06 0.04 0.02 400 500 600 700 800 900 1000 1100 1200 1300 0.15 0.10 0.05 400 500 600 700 800 900 1000 1100 1200 1300 0.35 0.30 0.20 0.15 0.10 0.05 450 500 550 600 650 700 750 800 850 900 950 0.55 0.50 0.45 c)

DLF [-]

Body weight [N]

Body weight [N]

Body weight [N]

Fig. 20. DLFs versus body weight for pacing rates in the range (a) 1.481.58 Hz, (b) 1.881.98 Hz and (c) 2.282.38 Hz.

input data: pacing rate and duration of walking number of cycles N

from the corresponding cluster select a set of the force parameters by chance random number seed ASD of variations of cycle intervals S (f)

random number seed

phases [-

shapes Zk(t) generation of synthetic walking cycles Ti generation of the corresponding angular frequencies i pdf of body weight random number seed body weight coupled system of equations weight normalised impulses Iw,i

weight normalised GRFs synthetic GRF signal


Fig. 21. Algorithm for generating synthetic time series.

model given by Eq. (6). The algorithm then assigns randomly and equally likely one of the functions Zk(t) to each of the N walking cycles. The next step integrates everything generated so far to run the dynamic model (Section 3.5) and therefore to generate weight normalised walking forces of a kind shown in Fig. 16b. Finally, these become equivalent walking force time histories when their amplitudes are additionally multiplied by random body weight. Fig. 22 shows examples of the signals generated when the model is run twice in succession for the pacing rate 2 Hz lasting 30 s. A visual comparison of the two signals provides convincing evidence that the model can account for the inter-subject variability in the shapes and force amplitudes. On the other hand, the ability of the model to generate different degrees of intra-subject variability for different persons becomes more obvious from comparison between the corresponding frequency spectra (Fig. 23). The broader

spread of energy around dominant harmonics (i.e. integer multiples of 2 Hz) and existence of sub-harmonics (i.e. integer multiples of 1 Hz) in Fig. 23a relative to b indicates that the rst virtual pedestrian varies their walking more.

5. Conclusions A data-driven mathematical concept has been developed for modelling dynamic loads generated by pedestrians walking. The ability to replicate temporal and spectral features of real walking loads, and therefore simulate reliably dynamic response of structures carrying pedestrians, gives this model a denite advantage over the existing walking models which can be found in the current UK and worldwide design guidelines for pedestrian structures, such as footbridges and oors. Therefore, the model presented in

V. Racic, J.M.W. Brownjohn / Advanced Engineering Informatics 25 (2011) 259275

273

a) 1100 1000 900 800 700 600 500 400 300 b) 1400 1200

Force [N]

Force [N]

1000 800 600 400 0 5 10 15 Time [s] 20 25 30

Fig. 22. Examples of synthetic time series for walking rate 2 Hz and duration 30 s.

0.35

Fourier amplitudes/ body weight [-]

0.30 0.25 0.20 0.15 0.10 0.05 0

a)

0.25 b)

Fourier amplitudes/ body weight [-]

0.20 0.15 0.10 0.05 0 0 2 4 6 8 Frequency [Hz] 10 12 14

Fig. 23. Discrete Fourier amplitudes of the signals shown in Fig. 22.

this paper can serve as a framework for a more realistic dynamic performance assessment which can be adopted in the design practice. The modelling strategy takes a complex numerical approach, thus can be distributed to designers as a user-friendly software. For example, it can be a graphical user interface which will allow designers to change pacing rate, duration of force signals and walking path through direct manipulation of graphical elements such as windows, menus, check boxes and icons. Automated structural design would help reduce its costs and improve the efciency with which structural engineers can handle it [48]. The rst renement of the current design procedures would account for innate variability of individual walking excitation. To simplify dynamic analysis, this excitation is conventionally as-

sumed in guidelines to be perfectly periodic and therefore presentable as a series of identical footfalls replicated at precise intervals. However, this leads to signicant loss of information and introduction of inaccuracies in predicted dynamic response. As demonstrated in this paper, the proposed model accounts for variations of intervals, impulses and shapes of successive strides (walking cycles) to simulate the response better. The second renement addresses the articial division in the current design guidelines between low and high-frequency structures. This is due to the lack of a mathematical concept which can associate all frequency components of real walking excitation within a single model. Assumed perfect periodicity allows walking forces to be presentable as a Fourier series for low-frequency struc-

274

V. Racic, J.M.W. Brownjohn / Advanced Engineering Informatics 25 (2011) 259275 [11] S. Zivanovic, A. Pavic, P. Reynolds, Probability based prediction of multi mode vibration response to walking excitation, Engineering Structures 29 (6) (2007) 942954. [12] V. Racic, Experimental Measurements and Mathematical Modelling of Nearperiodic Human-induced Force Signals, PhD Thesis. University of Shefeld, UK, 2009. [13] J.H. Rainer, G. Pernica, D.E. Allen, Dynamic loading and response of footbridges, Canadian Journal of Civil Engineering 15 (1988) 6671. [14] S.C. Kerr, Human Induced Loading on Staircases. Ph.D. Thesis, Department of Mechanical Engineering, University of London, UK, 1998. [15] P. Young, Improved Floor Vibration Prediction Methodologies, ARUP Vibration Seminar, 4 October, London, 2001. [16] J.M.W. Brownjohn, A. Pavic, P. Omenzetter, A spectral density approach for modelling continuous vertical forces on pedestrian structures due to walking, Canadian Journal of Civil Engineering 31 (2004) 6577. [17] S. Zivanovic, A. Pavic, P. Reynolds, Human-structure dynamic interaction in footbridges, Bridge Engineering 158 (BE4) (2005) 165177. [18] S. Zivanovic, A. Pavic, Probabilistic modelling of walking excitation for building oors, Journal of Performance of Constructed Facilities 23 (3) (2009) 132143. [19] C.J. Middleton, Dynamic Performance of High Frequency Floors, PhD Thesis. University of Shefeld, UK, 2009. [20] S.E. Mouring, B.R. Ellingwood, Guidelines to minimize oor vibrations from building occupants, ASCE Journal of Structural Engineering 120 (1994) 507 526. [21] A. Belli, P. Bui, A. Berger, A. Geyssant, J.R. Lacour, A treadmill ergometer for three-dimensional ground reaction forces measurement during walking, Journal of Biomechanics 34 (1) (2001) 105112. [22] P.E. McSharry, G.D. Clifford, L. Tarassenko, L.A. Smith, A dynamical model for generating synthetic electrocardiogram signals, IEEE Transactions of Biomedical Engineering 50 (2003) 289294. [23] J.R. Holmes, W. Holmes, Speech Synthesis and Recognition, second ed., Taylor and Francis, London, 2001. [24] A. Ebrahimpour, A. Hamam, R.L. Sack, W.N. Patten, Measuring and modelling dynamic loads imposed by moving crowds, ASCE Journal of Structural Engineering 122 (1996) 14681474. [25] J.B. Dingwell, B.L. Davis, A rehabilitation treadmill with software for providing real-time gait analysis and visual feedback, Journal of Biomechanical Engineering 118 (2) (1996) 253255. [26] W.C. Scott, H.J. Yack, C.A. Tucker, H.Y. Lin, Comparison of vertical ground reaction forces during overground and treadmill walking, Medicine & Science in Sports & Exercise 30 (10) (1988) 15371542. [27] J.B. Dingwell, J.P. Cusumano, P.R. Cavanagh, D. Sternad, Local dynamic stability versus kinematic variability of continuous overground and treadmill walking, Journal of Biomechanical Engineering 123 (2001) 2732. [28] R.C. Nelson, C.J. Dillman, P. Lagasse, P. Bickett, Biomechanics of overground versus treadmill running, Medicine and Science in Sports 4 (4) (1972) 233240. [29] G. Paolini, U. Della Croce, P.O. Riley, F.K. Newton, D.C. Kerrigan, Testing of a triinstrumented-treadmill unit for kinetic analysis of locomotion tasks in static and dynamic loading conditions, Medical Engineering & Physics 29 (2007) 404411. [30] P.O. Riley, G. Paolini, U. Della Croce, K.W. Paylo, D.C. Kerrigan, A kinematic and kinetic comparison of overground and treadmill walking in healthy subjects, Gait and Posture 26 (2007) 1724. [31] M. van de Putte, N. Hagemeister, N. St-Onge, G. Parent, J.A. de Guise, Habituation to treadmill walking, Bio-Medical Materials and Engineering 16 (2006) 4352. [32] HEF Medical Development, User Manuals, HEF Groupe, Lion, 2009. [33] Kistler, Kistler User Manuals, 2009. Available from: www.kistler.com. [34] J. Rose, J.G. Gamble, Human Walking, second ed., Williams & Wilkins, Baltimore, 1994. [35] C. Sahnaci, M. Kasperski, Random loads induced by walking, in: Proc. Sixth European Conference on Structural Dynamics (EURODYN), Millpress, Rotterdam, 2005, pp. 441446. [36] S.M. Pandit, S.-M. Wu, Time Series and System Analysis with Applications, John Wiley & Sons, Inc., New York, 1983. [37] D.E. Newland, An Introduction to Random Vibrations, Spectral and Wavelet Analysis, third ed., Pearson Education Limited, Harlow, 1993. [38] D.M. Bates, D.G. Watts, Nonlinear Regression and its Applications, Wiley, New York, 1998. [39] C.M. Bishop, Pattern Recognition and Machine Learning, Springer, New York, 2006. [40] Z.M. Kovacs-Vajna, A ngerprint verication system based on triangular matching and dynamic time warping, IEEE Transactions of Pattern Analysis and Machine Intelligence 22 (11) (2000) 12661276. [41] G. Tomasi, F. van den Bergand, C. Andersson, Correlation optimized warping and dynamic time warping as preprocessing methods for chromatographic data, Journal of Chemometrics 18 (2004) 231241. [42] J. Aach, G.M. Church, Aligning gene expression time series with time warping algorithms, Bioinformatics 17 (2001) 495508. [43] I.G.D. Strachan, Novel probabilistic algorithms for dynamic monitoring of electrocardiogram waveforms, in: Proceedings of the Sixth International Conference on Condition Monitoring and Machinery Failure Prevention Technologies, Dublin, Ireland, 2325 June 2009, pp. 545555. [44] R.L. Pimentel, A. Pavic, P. Waldron, Evaluation of design requirements for footbridges excited by vertical dynamic forces from walking, Canadian Journal of Civil Engineering 28 (5) (2001) 769777.

tures prone to resonant vibration under pedestrian excitation, whereas the same loads are modelled as a series of equivalent impulses in the case the human walking causes a transient response to the successive heel strikes. However, problems typically occur for a structure that has strong responses in both the high and low frequency region. This paper takes a step forward by providing a uniform model which can account for all energy levels of walking excitation, thus can simulate reliably dynamic response for a structure having an arbitrary fundamental frequency. Contrary to conventional approach to their modelling, walking forces are not deterministic but are a random phenomenon. Hence, the third renement of the current design procedures would follow principles used in experimental and analytical treatment of other random forces dynamically exciting civil engineering structures, such as wind, waves or earthquakes. Such comprehensive statistical treatment of moving human-induced dynamic forces currently does not exist anywhere in the world. Therefore, the steps taken in this paper, such as gathering a large number of time-varying loading records of walking, establishing a viable database of them and using it to develop a new generation of walking models for stochastic-based dynamic calculations of real-life structures, present a timely opportunity to advance the whole eld of vibration serviceability assessment of structures carrying pedestrians, such as footbridges and long-span oors. The model is based on straight line horizontal walking data but ought to apply for walking paths that are mildly curved in the horizontal plane. For vertical curves (e.g. for signicantly cambered footbridges) a databases of forces for ascent and descent would be required. Finally, the mathematical framework presented can be extended further to stochastic walking loads due to groups and crowds. At the moment, individual forces can be summed with random phase lags as suggested elsewhere [49]. However, there are indications that this is not what is happening in reality [1] and more research into synchronisation between people walking is needed.

Acknowledgements The authors acknowledge the nancial support provided by the UK Engineering and Physical Sciences Research Council (EPSRC) for Grant reference EP/E018734/1 (Human Walking and Running Forces: Novel Experimental Characterisation and Application in Civil Engineering Dynamics).

References
[1] V. Racic, A. Pavic, J.M.W. Brownjohn, Experimental identication and analytical modelling of human walking forces: literature review, Journal of Sound and Vibration 326 (2009) 149. [2] C.J. Middleton, J.M.W. Brownjohn, Response of high frequency oors: a literature review, Engineering Structures 32 (2) (2009) 337352. [3] British Standards Institution (BSI), Steel, Concrete and Composite Bridges. Part 2: Specication for Loads; Appendix C: Vibration Serviceability Requirements for Foot and Cycle Track Bridges (BS 5400), British Standards Institution, London, 1978. [4] The Highway Engineering Division, Ontario Highway Bridge Design Code CAN/ CSA-S6-06, The Canadian Standards Association, Toronto, 1983. [5] T.A. Wyatt, Design Guide on the Vibration of Floors, The Steel Construction Institute, Construction Industry Research and Information Association, London, 1989. [6] D.E. Allen, T.M. Murray, Design criterion for vibrations due to walking, Engineering Journal AISC 30 (4) (1993) 117129. [7] A. Pavic, M.R. Willford, Appendix G: vibration serviceability of post-tensioned concrete oors, in: Post-Tensioned Concrete Floors Design Handbook, second ed., Concrete Society, Slough, 2005, pp. 99107. [8] A.L. Smith, S.J. Hicks, P.J. Devine, Design of oors for vibration: a new approach, SCI P354, The Steel Construction Institute, Askot, 2007. [9] B.R. Ellis, On the response of long-span oors to walking loads generated by individuals and crowds, The Structural Engineer 10 (78) (2000) 1725. [10] J.W. Smith, Vibration of Structures, Chapman and Hall, London, 1988.

V. Racic, J.M.W. Brownjohn / Advanced Engineering Informatics 25 (2011) 259275 [45] N. Sekiya, H. Nagasaki, H. Ito, T. Furuna, Optimal walking in terms of variability in step length, Journal of Orthopaedic and Sports Physical Therapy 26 (1997) 266272. [46] MathWorks, Matlab User Guides, 2010. Available from: www.mathworks.com. [47] M. Hermanussen, H. Danker-Hopfe, G.W. Weber, Body weight and the shape of the natural distribution of weight, in very large samples of German, Austrian

275

and Norwegian conscripts, International Journal of Obesity 25 (2001) 1550 1553. [48] I. Flood, Towards the next generation of articial neural networks for civil engineering, Advanced Engineering Informatics 22 (2008) 414. [49] Y. Matsumoto, T. Nishioka, H. Shiojiri, K. Matsuzaki, Dynamic design of footbridges, in: IABSE Proceedings No. P-17/78, 1978, pp. 115.

Você também pode gostar