About Us(CRI) | ||||||||||||||||||

Estimations are free. For more information, please send a mail -->here<-- |
||||||||||||||||||

Power Spectral Density computation (Spectral Analysis) Page 3 of PD001A/B User Guide Table of Content Appendix 1; About numerical error generated by computers. (Case 2) When the signal itself is not a sinusoid but the composition of sinusoids of various frequencies In this case our signal will be decomposed into several different frequency components of sinusoids (please see equation (3)) and we need to pick up those components. If we do not know which frequency components to pick up but know the shape of our signal, we might take following steps. First create test data that contains only the signal. Next compute PSD of that test data. Finally inspect resultant PSD to decide which frequency components to pick up. (Case 3) When the signal is sinusoid but the amplitude of the signal varies in time (Case 3-1) Variation of amplitude itself is sinusoidal (Example 1) Amplitude of sinusoidal oscillation of frequency f0 varies sinusoidally. Frequency of amplitude variation is f1 and f1 is smaller than f0. Both f0 and f1 are constants. Figure 2-11(a) shows an example of time series of this case. S (bottom), our test data here, is S0 (middle) multiplied by S1 (top). The positions of mountaintops of S agree those of S0 (in phase) between 0.0 and 0.5 but the positions of mountaintops of S agree the positions of valley bottoms of S0 (out of phase) between 0.5 and 1.0. This is because value of S1 is positive between 0.0 and 0.5 but it becomes negative between 0.5 and 1.0. Whatever the phase relation is, we still think there is a variation of frequency f0 everywhere. However, PSD computation does not see it that way if time series is long enough. The reason of this is as follows. B in equation (4) is a constant and we set it to 1.0 in Figure 2-11. P0 and P1 are phase (please see Figure 1-2 for phase) and we set both of them zero in Figure 2-11 for the sake of simplicity.What equation (4) shows is that our signal is actually equal to a composite of two equal constant amplitude, B/2, sinusoids, SL and SH but of different frequencies. The important point here is that the amplitudes of both SL and SH are constant and that is why we see peaks of these sinusoids when we make a graph of PSD. Am in equation (3) must be constants as we described before. SL (top) and SH(middle) in Figure 2-11(b) are these two components in the second line of equation (4). Adding these two together produces S (bottom) and this S is the same as S in Figure 2-11(a). Although SL and SH in Figure 2-11(b) look the same except that they are out of phase, if we count the number of mountains or valleys we can see that frequency of these two are different. Thus, correct amplitude spectrum and PSD show there are two peaks at f2=f0-f1 and f3=f0+f1 and the amplitudes of those are identical. We also like to mention that the first line of this equation (blue underline) shows that it does not make any difference if we swap f0 and f1 as long as we also swap P0 and P1 at the same time. How long should data be to see this thing happen? We will show actual computational results next, but first let us explain it verbally. Assuming we have the data shown in Figure 2-11, if we increase the data length little by little, initially we will only see a single peak at f0. Amplitudes at frequencies other than f0 start increasing as we add more data, but none of them becomes outstanding and this will continue until our data length reaches to a half of the period of S1 (horizontal position is 0.5). When data length is 0.5, S becomes the same as S0 multiplied by a window function called a cosine window. When our data length is very short, frequency resolution may be too large to separate f0+f1 from f0-f1. We will start seeing twin-peaks instead of a single peak as data length becomes closer to one period of S1 (1.0 in Figure 2-11). When data length is equal to one period of S1 we will get twin-peaks and amplitudes of them are 0.5. As we add more data, we will see twin peaks most of the time but may see a single peak occasionally. In general we will see only twin-peaks if data length is more than a few times longer than the period of S1. For the mathematical explanation of this behavior, please see Appendix 3. As of the accuracy of amplitude estimation, the conclusion is that we will get best result when frequencies of both peaks match the frequency of PSD. Figure 2-12(a) shows time series of test data S in this case. You may have noticed that S in Figure 2-12(a) looks different from S in Figure 2-11 but this is because we set both P0 and P1 (phase; equation (4)) non-zero values this time. The unit of horizontal axis below Figure 2-12(d) is the period of S1. Actual data starts from 0 but the computational result using data shorter than one period long is quite confusing and thus we omit that part in this graph as before. Now let us call the bin, frequency coverage of which includes f0-f1, signal bin1 and the bin, frequency coverage of which includes f0+f1, signal bin2. Figure 2-12(b) shows amplitudes of signal bin 1 (blue; lower frequency) and signal bin 2 (red; higher frequency). We did not detrend data and did not apply any window function for these computations. Vertical blue and red dashed lines show where frequencies of the signal bins agree with their respective actual signal frequencies (f0-f1 and f0+f1). As we mentioned before we cannot choose arbitrary frequencies for PSD computation. In these particular computations frequency of signal bin 1 always matches f0-f1 when frequency of signal bin 2 matches f0+f1. Thus, all of the red dashed lines are drawn over blue dashed lines. The horizontal solid black line shows signal peak amplitudes, 0.5. This graph shows that adding more data would not improve accuracy of amplitude estimation greatly if we already have more than a several period of S1 as in the case of Figure 2-4(b). We are well aware that this graph is too complicated to see the details and we will show another graph next for closer inspection. Figure 2-12(c) shows the phases of the signal bins. The unit of this graph is 1 pi radian (Thus, the maximum is +1.0 and the minimum is -1.0). Horizontal blue and red solid lines show the phase of actual signals. We show another graph for closer inspection next. Figure 2-12(d) shows frequency bandwidth of bins. We omit the results when we used a Tukey and a Hanning window to save a space. Figure 2-13 shows expanded view of amplitude and phase of signal bins for the data length between 7.8 and 8.8. Here we show the results when we used a Tukey and a Hanning window as well. These graphs show that the amplitudes of signal bins vary quite rapidly unlike those shown in Figure 2-4(b) but this is because the periods of twin-peaks are much shorter (frequencies are higher) than the period of S1. If we use an average period of twin-peaks, which is equal to the period of S0, as a unit for horizontal axis instead of period of S1 and expand horizontal axis accordingly, we will see slow variations similar to those shown in Figure 2-4(b). This graph shows that it is very important to pay attention to periods of twin-peaks to get the best amplitude and phase estimates. The vertical short solid blue line A indicates the data length where frequencies of both signal bin 1 and signal bin 2 agree with their respective signal frequencies. This data length is also the data length when the end point of data can be connected smoothly to the start point of the data and an application of a window function is not necessary as we will explain in (4-2-1). The amplitudes and phases of both signal bins are very close to actual signal amplitudes and phases. Figure 2-14(a) shows amplitude spectra when data length is A. This graph shows that the amplitude spectra have sharp twin-peaks. This is clearer in Figure 2-15(a), which shows expanded view of amplitude spectra around twin-peaks. When we use a Tukey window, amplitude spectrum has a relatively large value at frequencies other than signal frequencies. When we use a Hanning window, peaks become wider (3 points) due to the implicit FDS. These features are pretty much the same as those shown in Figure 2-5 and Figure 2-8 except that now we have twin-peaks instead of a single peak. Figure 2-16(a) is basically the same graph as Figure 2-15(a) except that we used a linear scaling so that it will be easier to read amplitude of peaks. This graph shows that the amplitude at frequencies other than signal frequencies is actually much smaller than the signal bin amplitudes despite the impression one may get from Figure 2-15. The vertical short solid blue line C in Figure 2-13 indicates the data length when only the frequency of signal bin 2 agree with signal frequency. When data length is C, Figure 2-15(c) and Figure 2-16(c) show that the higher frequency peak (signal bin 2) is sharper and the amplitude of it is closer to the signal amplitude, while the lower frequency peak is wider and its amplitude is not close to the signal amplitude. The vertical short solid blue line B in Figure 2-13 is at the middle point between A and C and the frequency of signal bin 1 is relatively close to the actual signal frequency. Figure 2-16(b) shows that the amplitude of signal bin 1 is relatively close to the actual signal amplitude when data length is B. The vertical short solid blue line D is at the middle point between C and another point where only the frequency of signal bin 2 agrees with actual signal frequency. When data length is D, the frequency differences between signal bins and their respective signals are almost the same. Figure 2-16(d) shows that the both peaks are wider. The effect of FDS is the same as before (Figure 2-10). Therefore we only show the graph similar to Figure 2-10(a) below. Again, FDS is not that helpful. Finally, if we subtract f0 from frequencies of all the bins (allowing negative frequencies) and draw a graph of PSD using these modified frequencies, we will get two-sided PSD of S1. The appearance of this new PSD is the same as that of PSD without frequency modification except for horizontal axis labels (frequencies). (Example 2) Amplitude of sinusoidal oscillation of frequency f0 is summation of a constant and a sinusoidal variation of frequency f1. The signal of this case is expressed as Here, all of the coefficients in equation (5) (B, C, f0, f1, P0 and P1) are constant. The right hand side in the upper line of equation (5) (blue underline) is equal to the right hand side of equation (4) plus a sinusoidal variation of frequency f0. Thus, amplitude spectrum and PSD show peaks at three frequencies in this case. The peak at f0 is located right in the middle of two other peaks at f2=f0-f1(or f1-f0 if f0 is smaller than f1) and f3=f0+f1 and its amplitude is C. If B/2, the amplitude of other two peaks, is much smaller than C, we might just ignore amplitude variation and treat this case as a variant of Case 1. If B/2 is not negligible we have to pick up amplitudes at all three frequencies. This case is more like an AM radio signal. Amplitude Modulation (AM) radio wave is a constant frequency sinusoidal wave (carrier wave), amplitude of which varies with (modulated by) music or voices. If equation (5) represents an AM radio signal and we tune to f0, we will hear a single monotonous tone, frequency of which is f1. We cannot distinguish tone at f0+f1 from f0-f1 because they are equally distanced from f0. (Case3-2) Amplitude of sinusoidal oscillation of frequency f0 varies but that variation does not resemble to a sinusoid. In this case variation of amplitude of a signal, D(t), itself might be expressed as summation of sinusoids of various frequencies. Then, this case is just an extension of a previous case and (Case4) When the signal is periodic but the signal does not look like a sinusoid (2-1-3-4) Obtain some idea how strong the variations within a specific frequency range is. (2-1-3-5) Compute digitally filtered time series data Figure 2-18(a) shows the results when we applied high pass filter (HPF), band pass filter (BPF) and low pass filter (LPF) to the data shown in Figure 1-1(a) as an example. (HPF is a filter that removes low frequency components. LPF is a filter that removes high frequency components. BPF is a filter that removes both high and low frequency components.) The shortest period (upper frequency limit of HPF) we can get is 2 hours and this is determined by sampling interval as we described before. The longest period, which is determined by data length, is 1250 days and that is the lower frequency limit of LPF. We also draw original time series data at the bottom (red) for comparison. This is basically the same as Figure 1-1(a). Low pass filtered data reproduces smoothed shape of original data well in this case. Since the summation of the frequency ranges of HPF, BPF and LPF contains entire frequency range, summation of these three filtered time series is supposed to be equal to original time series data. Figure 2-18(b) shows the difference between original time series data and the summation of these three filtered time series data. This graph shows small computational error we described before. The maximum value of original time series data exceeds 100 and the standard deviation is about 31.9. Therefore, these errors are much smaller (order of 1E-12) than original data and negligible for majority of applications but this is another example that the result of computation like this does always contain some small computational errors. As an extension of above, we might be able to use information of amplitudes and phases to construct signal from noise-contaminated data. We start our demonstration when noise is relatively small. Our first test data is similar to the case 3-2 in (2-1-3-3) and Here D0=0.8, D1=0.4, D2=0.3, f0=100, f1=4, f2=10 and all of these are constants. P0, P1 and P2 are also non-zero constants. We omit unit of any of these constants here because units are not important for our demonstration. We generated "Noise" using a random number generator. The standard deviation of generated noise is about 1.0 and this means strength of noise is about the same as that of the signal we want to detect. The signal part of this data includes variations at 5 frequencies. Namely, Let us defineV0, V1, V2, V3 and V4 as With these variables, equation (7) can be simply expressed as Figure 2-21 shows the expanded view of variations of amplitudes and phases of signal bins near A, B and C. Although A is the data length when frequencies of all of the signal bins match their respective signal frequencies, Figures 2-21(a) and (c) show that the amplitudes of these signal bins are close but do not match to actual signal amplitudes unlike the case shown in Figure 2-13. This is due to the noise we added. B is the data length when none of frequencies of signal bins match their respective signal frequencies and C is the data length when frequencies of signal bin 3 and signal bin 4 are relatively close to their respective signal frequencies. Figure 2-22 shows amplitude spectra of test data (blue) and of signals (red; S in equation (8)) for data length A, B and C. No window function was applied for these computations. These graphs show that the spectrum of noise is fairly flat across the entire frequency range. When data length is A, end point and start point of data without noise (red) can be connected smoothly and thus Figure 2-22(a) shows that the amplitudes at frequencies other than the signal frequencies are zero within computational accuracy as in the case shown in Figure 2-5(a). It must be stressed here that it will be very difficult to achieve complete frequency match when we have signals of multiple frequencies. We chose signal frequencies very carefully for this test to achieve this condition. When data length is not A, we have spectral leakages and that raises spectrum across entire frequency range as before. Figures 2-22(b) and 2-22(c) show that the effect of spectral leakage at frequencies far from signals is much smaller than the effect of noise and is not detectable when we add noise in this example, however. Figure 2-23 shows expanded view of amplitude spectra of signals (noiseless; corresponding to red line in Figure 2-22) around signal frequencies. For these graphs we used linear scaling and the plus (+) marks in these graphs show the frequency and the amplitude of bins. The height of short horizontal black solid lines (5 lines for each) indicates actual signal amplitude for each signal and the central position of these lines indicates actual signal frequencies. Please note that they do not necessarily match frequencies of signal bins. When data length is A, Figure 2-23(a) shows that amplitudes of signal bins are very close to actual signal amplitudes and all the peaks are single point sharp peaks. When data length is not A, Figures 2-23(b) and (c) show that amplitudes of signal bins deviate from actual signal amplitudes and some peaks become wider. Figure 2-24 shows how amplitude spectra change if we add noise of this magnitude. Although noise in this case is large enough to obscure signals in Figure 2-19, peaks in amplitude spectra are easily identifiable. Comparing Figure 2-24 with Figure 2-23, we can say that the difference of amplitude of signal bins between noiseless data and noise added data is relatively small in this case. There seems to be several strategies to reproduce signals from noise-contaminated data. The first one is picking up only the signal bins and then using equation (3). We call this Method 1. This method will underestimate signal amplitude when frequencies of signal bins do not match actual signal frequencies and spectral leakage occurs. The other strategy is picking up signal bins and two neighboring bins up and down (or left and right in Figure 2-23 and 2-24) and then using equation (3). In other words, pick up three bins for each signal. We call this Method 2. Figure 2-25 shows time series of actual total signal (S in equation (8) and Figure 2-19) at the top, reproduced total signal (S' for abbreviation) by Method 1 using amplitude spectrum of noiseless data for data length A, B and C and their difference from S (=S-S'). The data lengths where amplitude spectra are computed for each computation are shown as vertical black solid lines. The standard deviation of difference between S and S' when data length is A is about 2.2E-5 and the difference is too small to be seen in Figure 2-25(c). The differences between S and S' when data length is B and C are still small but large enough to be seen in Figures 2-25(e) and (g). These differences are symmetrical with the minimum in the middle of the section left of vertical black lines (not in the middle of graph). Differences tend to increase at the far right, which suggests extrapolation is a bad idea as usual. The point we would like to make here is that if frequency does not match, signal reproduction will not be perfect even if the data contains only the signal but nothing else. Figure 2-26 shows S' by the Method 2 and their difference from S. Figures 2-26(d) and (e) show that the differences between S and S' decrease slightly than those of Method 1. If we add more bins, S' will eventually become almost identical to S since we only have S but nothing else in this case. Graphs of Figure 2-27 are the same as those of Figure 2-25 except that these are the results of noise added data (Data in Figure 2-19(a)). The difference between S and S' becomes larger. If we use Method 2 instead of Method 1 to reproduce S from noise added data, Figure 2-28 shows that the differences slightly decrease as before. However, unlike the case of noiseless data, adding more bins will make difference between S and S' larger at certain point because we are using noise added data here. |
||||||||||||||||||