Survey

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Transcript

B.5.5. Comparing Methods to Denoise COP Signals The source for simulating naturalistic reference signals was the signals obtained by a force platform (Dinascan 600M, IBV, Valencia, Spain) at a sampling rate of 180 Hz, from a population of fourteen subjects, who performed sixteen trials each of manual lifting and lowering tasks in an experiment described in detail in Â§ 3.5. In total 224 were obtained. The first forty harmonics were used to approximate the COP reference signals, ensuring that the reconstructed signals are practically bandlimited to 40 f0 (i.e., Fourier transform is zero outside). Original signals were demeaned before be transformed into Fourier series. Then, the reconstructed signals were decimated by a factor of 6 using an FIR filter of order 30 to match the noise sample rate of 30 Hz. The duration of the new signals is of 30 sec, with resolution f o = 1 30 , and cut-off frequency = 1.33 Hz. The experimental obtained noise added to the reconstructed signals was generated by measuring different static loads placed at the geometrical center of the top plate of the force platform at sampling rate of 30 Hz for a period of 5 sec. In total 120 noise signals were generated, demeaned, and stored. To check whether the experimental obtained âzero-meanâ noise can be characterized as Gaussian white noise (GWN), we examine independence (Box-Pierce and Ljung-Box tests) and normality (Anderson-Darling and Shapiro-Wilks tests) of the time-series. Six noise signals was then pooled randomly, combined in random order, and added to the reconstructed reference signals to simulate COP registrations (Fig. B.9). In total 26880 signals were generated, superimposed with additive non-stationary noise. Three different denoise procedures were compared: low-pass filtering with a fourth-order zero-phase-shift Butterworth filter with cut-off frequency found following the residual analysis procedure (Winter, 2009), by quintic splines using the generalized crossed-validation natural splines smoothing algorithm (GCVSPL) (Woltring, 1986), and by quintic splines according to the âTrue Predicted Mean-squared Errorâ of Woltring (1986) following the proposed uncertainty analysis. B.5.6. B.5.6.1. Results and Discussion Noise Test The results showed that for Fz = constant the noise of the COP signal can be modelled as additive, zero-mean âwhite noiseâ. Different sampling rates influence the COP noise. This is obvious for the COP signals that were registered with different (low - very high) sampling rates (Fig. B.10). For example, the variances of the raw COP data for different sampling rates are COP M L RAW -30Hz = 0.54 mm2 , COP M L RAW -230Hz = 0.59 mm2 , COP M L RAW -500Hz = 0.61 mm2 , and COPAP RAW -30Hz = 1.10 mm2 , COPAP RAW -230Hz = 1.20 mm2 , COPAP RAW -500Hz = 1.30 mm2 . However, for a narrower frequency interval the assumption that the sampling rate do not influence the COP noise can be considered as correct. Notwithstanding, the noise elimination, was higher after oversampling spread the power over higher frequencies. The variance of the COP signals after low-pass filtering is COPAP BT W -30Hz = 0.24 mm2 , COPAP BT W -230Hz = 0.05 mm2 , COPAP BT W -500Hz = 0.03 mm2 and COP M L BT W -30Hz = 0.15 mm2 , COP M L BT W -230Hz = 0.03 mm2 , COP M L BT W -500Hz = 0.02 mm2 (Fig. B.11). Other studies have also been shown that cut-off frequency and sampling rate influence stabilometric parameters (Schmid et al., 2002; Scoppa et al., 2013). However, our 164