CSMIP STRONG-MOTION DATA PROCESSING

Transcription

1 CSMIP STRONG-MOTION DATA PROCESSING A.F. SHAKAL, M.J. HUANG, AND V.M. GRAIZER California Strong Motion Instrumentation Program California Geological Survey, California Department of Conservation 801 K Street, MS 13-35, Sacramento, California tshakal@conservation.ca.gov ABSTRACT Digitization and processing of accelerograms at CSMIP began in During the 1980s and early 1990s, a film scanning system patterned after that developed at USC by Trifunac and Lee was used to digitize analog film accelerograms. Since the mid 1990s, film accelerograms have been digitized using a PC-based flat-bed scanner system developed at CSMIP. Post-digitization processing of analog records is based on the methods developed during the Caltech Bluebook project and evolved at USC. Processing of digital accelerograms from modern recorders, which avoid film digitization problems, much is less complex. CSMIP uses a frequency-domain processing approach that follows the basic elements of the Caltech/USC methods. INTRODUCTION Digitization and processing of analog film accelerograms at the California Strong Motion Instrumentation Program (CSMIP) is based on the Caltech Bluebook project of the early 1970s, described by Hudson and others in reports by the Earthquake Engineering Research Laboratory [1]. The computer programs are based on those developed under that project, described by Trifunac and Lee [2]. Although advances and improvements have been made, the core reflects this origin. ANALOG FILM DIGITIZATION CSMIP began digitizing and processing film records in a joint effort with the USGS in the late 1970s. This joint effort was followed by a stand-alone processing project at CSMIP, begun in 1982, in which accelerogram films were scanned using a system patterned after the semiautomatic digitization system introduced by Trifunac and Lee [3] in 1979 at the University of Southern California. This system replaced the cursor-and-crosshairs of the Caltech Bluebook manual digitizing project with an optical-density sensor which moved, under the control of a computer, across a film image as it rotated on a cylinder, yielding an x-y grid of optical density values. Curve-following software applied to the grid of optical densities ultimately yield acceleration-time pairs like that obtained in manual digitization. The system was further studied in terms of noise and longer period response and certain improvements made to lower the system noise levels by Shakal and Ragsdale [4]. In the mid 1990s, as desktop computers became more widespread and powerful and highresolution scanners also became available, a desktop scanning system that is almost fully automated, except for the most difficult records, was developed by Cao and others [5]. The early desktop scanners were not adequate because the step size was too large and the trace was

2 approximated as a black-and-white image rather than the gray-scale imaging that more smoothly matches the accelerogram image. The PC-based scanner, after a period of parallel digitizing and testing, replaced the large mini-computer based optical scanning system. This desktop system, which has a lower noise level than the original large system, is used for digitizing film accelerograms at CSMIP, though there are increasingly fewer films to digitize as the network is converted to digital instruments. ACCELEROGRAM PROCESSING The Caltech project showed that displacements computed for the longest periods were heavily influenced by noise intrinsic to the film digitization and processing, and that this noise could be handled in a uniform way, as first suggested in 1977 by Trifunac [6]. In general, the log of the velocity response spectrum for records from intermediate-magnitude events increases from a low value at short periods until it reaches a maximum at intermediate periods, beyond which it decreases once again at longer periods. In contrast, the noise spectrum due to digitizing and processing a straight-line record increases in the long period range, while the data spectrum is decreasing. As a result, the spectrum of a digitized accelerogram plotted against period, before filtering, often has a characteristic shape, shown schematically in Figure 1. A noise spectrum is present for both analog and digital records, but it is generally higher for optically-digitized analog records than records from digital recorders. Figure 1: Typical spectrum from a digitized earthquake record (solid line) compared with the digitization noise spectrum. An objective in strong-motion data processing is to determine the useful data bandwidth where the signal adequately exceeds the noise. The principle of uniform processing as introduced by Trifunac [6] utilized the characteristic form of the response spectrum as the basis for selecting the filter period. Trifunac and Lee [7] used this approach to refilter the Caltech records, selecting the long-period filter corner based on where the accelerogram spectrum intersected the averaged spectrum for a set of straight-line accelerograms. This approach coincides with the signal-to-noise ratio (SNR) approach used in signal processing. In CSMIP processing a SNR of 2 to 3 is used to obtain an initial estimate of the long period filter corner. Specifically, a period is chosen to the left (toward shorter periods) of the minimum which occurs where the signal and noise spectra intersect; a period is then selected at which the spectrum has increased to 2-3 times that minimum.

3 A practical aspect remains which can have a significant influence - the choice of one filter corner for each component, versus one corner for each record (so that all components are filtered the same). The latter approach (used by CSMIP) means that some components (e.g., horizontal) are filtered more severely because the filter is set to control noise in the component with the lowest amplitudes (often the vertical). Using the same filter band for all components means that multidimensional features of the data like particle motions and torsional motion can be analyzed by users without special procedures. Steps in Strong Motion Processing The post-digitization processing procedures and steps described below have their origin in the procedures developed in the Caltech EERL [1] project and modified by Trifunac and Lee [7] and others. Several basic steps can be defined (described more thoroughly in Shakal and others [8]). 1. Baseline Correction The raw digitized data points are interpolated to obtain equalinterval sampling, if necessary (e.g., 200 points/sec, for a 100 Hz folding frequency), and converted to acceleration units using the sensitivity constant of the accelerometer. At least a first-order base-line operation is performed, to make the data zero-mean. More involved baseline correction may also be performed on particular records, in some cases. In general an equivalent to complex baseline correction is performed via long-period filtering (Step 4), which has well-defined properties in the frequency domain, largely independent of the record length. Simple base-line correction using a constant or linear trend, where appropriate, is most effective in projects handling a large number of records. The results of this step are usually denoted as Volume (or Phase) 1, and released a rawdata product. 2. Instrument Correction The base-line adjusted data is corrected for instrument response using a simple finite-difference operator. The sensitivity constant is applied in Step 1. In frequency domain processing, discussed further below, the finite-difference process can be simply replaced by dividing the spectrum by the instrument response spectrum. 3. High Frequency Filtering After instrument correction, a filter is generally applied to remove high frequency noise. In the Caltech and USC processing, an Ormsby filter with a corner frequency at 23 Hz and a termination frequency at 25 Hz was applied. For modern digital records, CSMIP uses a Butterworth filter with a corner frequency near 80% of the final sampling rate and a 3 rd or 4 th order decay. After filtering, the data are decimated to the final sample rate at which the acceleration data will be distributed to the user community (50 points/second in the Caltech/USC data; CSMIP distributes data at 100 points/second today for data recorded digitally). In the case of time-domain processing, the instrument correction is performed prior to this decimation, rather than after, to improve the high-frequency accuracy (Shakal and Ragsdale [4]). 4. Initial Integration and Long Period Filtering An initial long period filter is applied to the instrument-corrected acceleration data with a cut-off corner near 15 seconds period (longer periods are filtered out), the maximum period traditionally used in Caltech/USC

4 processing. (Recent data has shown that a longer-period cut-off may be appropriate for digital records from large earthquakes.) Velocity and displacement are obtained by numerically integrating the acceleration and filtered using the same low-frequency filter. 5. Computation of Maximum-Bandwidth Response Spectra The pseudo-velocity (PSV) response spectra are calculated in the time domain for the full set of 91 periods defined in the Caltech project (0.04 seconds to 15 seconds) using the method of Nigam and Jennings [9]. The spectra are computed for damping values of 0, 2, 5, 10, and 20% of critical, and plotted for 0.04 to 15 seconds (the full bandwidth) for use in comparative analyses to select the best filter. 6. Long-Period Filter Selection A suite of long-period filters is applied to the data obtained in Step 3 using corner periods near the long-period minimum of the spectrum obtained in Step 5. The long-period intersection of the maximum-bandwidth spectrum and the average noise spectrum determined for the system is used to indicate the long-period limit of useful information in the record, as discussed above. The final value of the filter corner is selected after studying the resulting suite of displacement plots, comparing them to one another, to displacement plots computed from noise records, and to records obtained for nearby stations, if available. Choosing a filter period at which the signal-tonoise ratio (SNR) is 2 to 3 or greater usually gives a result with low noise, though this may be more conservative that some desire. Selection of the optimal long period filter remains the most difficult part of strong-motion processing. The effect of earthquake magnitude is to raise the response spectrum at long periods, so that the crossing with the noise spectrum may not occur in the usual strongmotion processing band. In CSMIP processing, the selection is made in a team approach, in which several individuals (e.g., the authors), propose the best corner and the reasons for it, based on the suite of displacement plots, spectra, and records from any nearby stations; the consensus that comes out of the effort determines the final filter corner used. 7. Final Output Preparations The acceleration, velocity and displacement time histories obtained using the final filter are plotted for presentation in reports and saved as files for distribution. The response spectra may be plotted with tripartite logarithmic scaling or the linear scaling most commonly used in building codes. Illustrative Example To illustrate the process described above, a copy of a raw accelerogram is shown in Figure 2a. This analog record contains three acceleration traces corresponding to the three component directions, two fixed reference traces, a half-second time mark pattern at the bottom and a time mark trace containing the IRIG-encoded time at the top. This record was digitized and scaled to produce the Vol. 1 (Phase 1) data plotted in Figure 2b. Steps 2 through 4 were then performed, and the full-bandwidth spectrum of Step 5 was computed; for the first component, the velocity spectrum of the full-bandwidth output is compared with the noise spectrum of the digitization system in Figure 2c. As part of Step 6, a suite of displacement time histories, computed using a range of long-period filter corners, are compared in Figure 2d to guide the selection of the

5 optimal filter. The final "corrected" acceleration, velocity and displacement (Phase 2 data), and response spectra (Phase 3 data) of Step 7 are shown in Figures 2e and 2f, respectively. Note that the final filter corners are shown in both plots. Figure 2a: The analog record obtained from the Obregon Park station in Los Angeles during the 1987 Whittier Earthquake. Figure 2b: Acceleration records (Phase 1 data) after the raw record in Figure 2a was digitized and scaled.

6 Figure 2c: Spectrum of the first component of the record in Figure 2a (using a wide filter bandwidth) is compared with the noise-level spectrum (PSV, 20% damping) for the CSMIP digitization system (Shakal and Ragsdale [4]) to help establish the long-period filter corner. Figure 2d: Displacement plots for the first component of the record in Figure 2a processed with different long-period filters. The plots are compared with each other, and the spectra are compared with the noise spectra in Figure 2c, to select the final long-period filter. In this case, the filter with a 3-second corner period was selected (Huang et al. [10]).

7 Figure 2e: Final instrument and baseline-corrected acceleration, velocity and displacement (Phase 2 data) for the first component of the record in Figure 2c. The filter bandwidth used in the processing is indicated on the plot. Figure 2f: Final response spectra (Phase 3 data) for the record in Figure 2c, first component. The spectra are plotted only for periods within the filter bandwidth determined in the processing. Usable Data Bandwidth Processed data has a bandwidth, extending from the period of the high frequency (short-period) filter to the corner of the long period filter, within which the final spectrum corresponds to the

8 original record. Outside of this band, filters have removed as much as possible of the information because of noise contamination; this band may be called the Usable Data Bandwidth (Figure 3). The UDB defines the range within which the results can be used in modeling of Figure 3: Definition of the usable data bandwidth (UDB) for a processed strong-motion record. The UDB gives the frequency or period range within which the data can be used for seismological and earthquake engineering applications. structural response, etc. Comparisons of modeling results to data outside this period band are not valid, of course, and will likely be misleading. Note that the spectra of the final result in Step 7 should only be plotted to the period of the long-period filter - beyond that the spectra are smoothly decreasing, but this is because of the asymptotic nature of the response spectrum at long periods and does not reflect the shape of the data spectrum. Time Domain vs Frequency Domain Some operations in strong motion processing may be most effectively performed in the frequency domain. The early Caltech processing, before the FFT was widely available, was all performed in the time domain. The Ormsby filter was selected as a relatively stable timedomain, recursive filter. Today the FFT is common, and filters like the Butterworth are easily applied in the frequency domain, and integration to obtain velocity and displacement is as simple as dividing the spectrum by iω. Tapering the data segment (avoiding discontinuities) is particularly important in frequencydomain processing. In CSMIP processing the ends of the acceleration time series are tapered with a raised cosine bell, or Tukey interim window (e.g., Bergland, [11]), once the mean has been removed, to prevent spurious side-lobe leakage in the spectrum. If there is an adequate preevent data segment, this does not affect the data at the beginning of the motion. Of course, in frequency domain processing the time-history length to be transformed to the frequency domain must have a power-of-two number of points. A zero-filled section is included at the end of adequate length (25% of the record length or more) to prevent problems with periodicity. Some operations, like complex baseline corrections, are most effectively done in the time domain. For example, if there is an a priori reason that certain time functions should be used to

9 compensate for a baseline shift that occurs at a certain point in an accelerogram, a time domain application is most effective. Accelerometer Offsets, Recording Site Tilts and Permanent Displacements Calculated long period displacements are very sensitive to the stability of the DC or zero level of the accelerometer. Small errors in the sensor zero-level can lead to significant errors in displacements and in long period spectrum levels. This problem has become more important for modern digital recorders, largely because more information is being extracted from the records due to their greater dynamic range. The ability of sensors to shift to a new zero value during or after strong (or even weak) motion is an important caution to consider in strong motion processing. This phenomenon had been noted, for example, by Iwan and others [12] in laboratory tests, and they proposed a numerical approach to manage this effect. This behavior was also observed in early small-amplitude field recordings, especially in early magnetic-tape digital accelerographs. It has also been observed in the records from other arrays and instruments types, such as those from the SMART array in Taiwan, and most recently the 1999 Taiwan earthquake (W.H.K. Lee, personal communication, 1999). It was also observed, with significant level changes, in records from the 1994 Northridge and other earthquakes. Correction for a changing zero level remains problematic, because it requires an assumption of the nature (time function) of the change. If the change can be approximated by a simple step, it is easy to introduce a correction in processing. If a change can be approximated as a constant during an interval of unknown duration, followed by a new constant value, the approach suggested by Iwan and others [12] can be used. However, if the change occurs over a period of time as a ramp, curve, oscillatory function, or a series of steps (as slippage in the sensor might cause), the appropriate correction function is uncertain. A zero-level change due to actual tilting of the instrument site may be hard to distinguish from a sensor zero-shift problem. The 1994 Northridge record analyzed by Shakal and others [13] from the Pacoima Dam Upper Left Abutment shows a clear zero-level shift. The time of the shift is the same on the two horizontal components, and there is little expression on the vertical component, important diagnostic features of real tilting. Also, the tilt amount obtained from analysis of the record (approximately 3 degrees) and its azimuth closely matches actual postearthquake measurements made on the concrete pad on which the instrument was mounted. In general, zero-level offsets or tilts remain an important difficulty in strong motion processing and the recovery of the longest period motions and permanent displacement. Sensor zero shift significantly affects the ability to determine permanent displacement without introducing assumed corrections. However, permanent displacement is generally not important in earthquake engineering unless a structure straddles a fault. In the period band of most importance in engineering the zero-shift effects are not as important as in seismological source modeling. Shakal and others [13] found little effects at periods less than 10 seconds, and Boore [14] found no important effects below 20

10 seconds in the examples he studied. In general, in CSMIP processing a manual process similar to the Iwan approach is used in the infrequent cases where sensor zero shifts occur. AUTOMATIC PRELIMINARY PROCESSING Advances in recording technology have opened the possibility of automated preliminary processing. Current steps in this direction show promise for the future (e.g., Figure 4). Figure 4: Automatic preliminary processing results - Three components of acceleration, velocity and displacement at Santa Barbara - Goleta station from a M4.4 earthquake that occurred on May 9, 2004 at 1:57 AM. Automated processing will never approach the quality level that human analysis can yield, but the rapid results obtained are very useful, though preliminary. To get the best possible results, the preliminary processing must be followed up by professional analysis of noise levels and possible filter settings, just as has been performed for years in conventional strong motion processing. Developments in recent years, such as communication with recorders at remote sites via landline, cell phone or radio, has opened new possibilities in rapid use of strong motion recordings. Field instruments can initiate communication with a central site upon recording an event, and transmit

11 the recorded data. This approach was introduced several years ago by the U.S. Bureau of Reclamation (Viksne et al. [15]) and CSMIP (Shakal et al. [16]), and also recently by the U.S. Geological Survey. This rapidity adds an important additional dimension to the value of strong motion records. For automated, preliminary processing, a process similar to that described above may be used. For all but large-earthquake records, a long-period filter corner near 3 or 4 seconds can be used as a reasonable first estimate. Because sensors have a high natural frequency, as noted above, an approximate instrument correction as simple as multiplication by a sensitivity factor can often be used with little penalty. Of course, this preliminary processing must be augmented by careful processing and filter selection by analysts at a later time to obtain optimal quality results. In the California Integrated Seismic Network, automated processing of records leads to automated Internet reports on shaking, like ShakeMaps, and to the Internet Quick Reports at the CISN Engineering Data Center, at which rapidly show preliminarily processed data for use in event response. SUMMARY Strong motion processing at CSMIP has advanced from the painstaking digitization of film accelerograms by scanning systems to include automated preliminary processing of digital accelerograms. The noise spectrum of modern accelerographs holds promise for high accuracy processing. However, sensor problems and limitations restrict the potential value of this precision in achieving accuracy in processed results without human oversight. Preliminary processing provides much of what is needed in an immediate post-event response application of the recorded data, however. REFERENCES 1. Earthquake Engineering Research Laboratory. Strong Motion Earthquake Accelerograms, Digitized and Plotted Data (Vols. I through III) EERL 70-20, 71-50, and 72-80, California Institute of Technology, Pasadena 2. Trifunac MD and Lee V. Routine Computer Processing of Strong Motion Accelerograms, Earthquake Engineering Research Laboratory, Report EERL 73-03, California Institute of Technology, Pasadena, Trifunac MD and Lee V. Automatic Digitization and Processing of Strong Motion Accelerograms, Report CE 79-15, Dept. Civil Engineering, Univ. Southern California, Los Angeles, Shakal AF and Ragsdale JT. Acceleration, Velocity and Displacement Noise Analysis of the CSMIP Accelerogram Digitization System, Proc. 8th World Conference Earthquake Engineering, Vol. 2, , 1984.

12 5. Cao TQ, Darragh R and Shakal AF. Digitization of Strong Motion Accelerograms Using a Personal Computer and Flatbed Scanner (Abs), Seismological Research Letters 1994; 65(1): Trifunac MD. Uniformly Processed Strong Motion Earthquake Ground Accelerations in the Western United States of America for the Period from 1933 to 1971: Pseudo Relative Spectra and Processing Noise, Report CE , Univ. Southern California, Los Angeles, Trifunac MD and Lee V. Uniformly Processed Strong Earthquake Ground Accelerations in the Western United States of America for the Period from 1933 to 1971: Corrected Acceleration, Velocity and Displacement Curves, Report CE 78-01, Dept. Civil Engineering, Univ. Southern California, Los Angeles, Shakal AF, Huang MJ and Graizer V. Strong-Motion Data Processing, Lee WHK, Kanamori H, Jennings PC and Kisslinger C, Editors. In: International Handbook of Earthquake and Engineering Seismology, B, Amsterdam, Academic Press 2003: Nigam NC and Jennings PC. Calculation of Response Spectra from Strong-Motion Earthquake Records, Bull. Seism. Soc. Amer. 1969; 59: Huang MJ, Cao TQ, Parke DL, Shakal AF. Processed Strong Motion Data from the Whittier, California Earthquake of 1 October 1987, Report OSMS 89-03, Cal. Div. Mines and Geology, Sacramento, Bergland GD. A Guided Tour of the Fast Fourier Transform, Rabiner LR and Rader CM, Editors. In: Digital Signal Processing, New York, IEEE Press Iwan WD, Moser MA and Peng CY. Some Observations on Strong Motion Earthquake Measurements Using a Digital Accelerograph, Bull. Seism. Soc. Amer. 1985; 75: Shakal AF, Cao TQ and Darragh R. Processing of the Upper Left Abutment Record from Pacoima Dam for the Northridge Earthquake, Report OSMS 94-13, Calif. Div. Mines and Geology, Sacramento, Boore DM. Effect of Baseline Corrections on Response Spectra for Several Recordings of the 1999 Chi-Chi, Taiwan, Earthquake, Bull. Seism. Soc. Amer. 2001; 91: Viksne AC, Wood C and Copeland D. Seismic Monitoring/Strong Motion Program and Notification System, In: Water Operation and Maintenance Bulletin, No. 171, , Bureau of Reclamation, Denver, Shakal AF, Petersen CD, Cramlet AB and Darragh RB. CSMIP Near-Real-Time Strong Motion Monitoring System: Rapid Data Recovery and Processing for Event Response, In: Proceedings SMIP95 Seminar on Seismological and Engineering Implications of Recent Strong- Motion Data, Huang, MJ, Ed., 1995, California Div. Mines and Geology, Sacramento.