Data Acquisition and Network Enhancements

Introduction

Stations from nearly all networks operated by the BSL transmit data continuously to UC Berkeley for analysis and archive. As part of our operations, we have developed procedures for data acquisition and quality control in order to ensure a complete and high quality data collection. In this chapter, we describe the data acquisition systems and quality control procedures. We also document recent efforts in instrument testing and a collaborative experiment in early warning.

Data acquisition

Central-site data acquisition for the BDSN and NHFN is performed by two computer systems located at the BSL (Figure 8.1). These acquisition systems are also used for the Parkfield-Hollister electromagnetic array and for the BARD network. A third system used primarily as data exchange system with the USNSN receives a secondary feed from CMB, MOD, SAO, and WDC from the the NSN VSAT, and processes and transmits event data to the USNSN from HOPS, CMB, SAO, WDC, and YBH. Data acquisition for the HRSN follows a more complicated path, as described in Chapter 4.

Figure 8.1: Data flow from the BDSN, NHFN, HRSN, and BARD network into the BSL central processing facility.

Data acquisition and communication with the Quanterra data loggers depends both on the software on the recording systems and at the central site.

MultiSHEAR

In late 1998, Quanterra provided the first release of MultiSHEAR, an enhanced version of its data acquisition software that was year 2000 compliant, and updated components of the OS/9 operating system to address the year 2000 problem. MultiSHEAR contained a number of enchancements, especially in the area of multi-site data collection, and introduced a totally new configuration procedure. The BSL worked with Quanterra during 1999 to enhance the configuration procedures to address the diverse needs of the BDSN and HFN station configurations. During November and December 1999, all of the BSL Quanterra data loggers were updated to MultiSHEAR with the corresponding OS/9 modifications, which addressed the year 2000 problems.

The two significant features of MultiSHEAR that affected the BSL were the correction of a systematic timing error of the decimated channels from SHEAR and UltraSHEAR software in the Quanterra data logger, and the addition of multi-site data collection. The subject of timing corrections is addressed in Chapter 10.

Multi-site data acquisition

Prior to the release of MultiSHEAR, a Quanterra data logger could acquire, store, and telemeter data from only a single source such as its own A/D system. However, MultiSHEAR adds the new capability to acquire data from remote digital sources such as other Quanterra systems. The BSL will use this feature to create "hub" systems, which will acquire, store, and transmit data from several remote data loggers as well as the hub's own data. The San Francisco/Oakland Bay Bridge network will consist of two Quanterra 4120 hubs (Figures 8.2 and 8.3), each of which will acquire data from its own digitizers as well data from three remote diskless Quanterra Q730 systems. Each hub will provide local storage for all four sites, and will transmit the real-time the data from all four sites to the BSL over a 512Kb spread spectrum radio. The BSL developed the initial MultiSHEAR hub configuration procedure, and worked with Quanterra to refine and test the hub configurations.

Figure 8.2: Planned data flow from the East Bay Bridge seismic monitoring hub.

Figure 8.3: Planned data flow from the West Bay Bridge seismic monitoring hub.

Comserv

The BSL uses the comserv program for central data acquisition, which was developed by Quanterra. The comserv program receives data from a remote Quanterra data logger, and redistributes the data to one or more comserv client programs. The comserv clients used by REDI include datalog, which writes the data to disk files for archival purposes, cdafill, which writes the data to the shared memory region for REDI analysis, and other programs such as the seismic alarm process, the DAC480 system, and the feed for the Momento Mori Web page (Figure 8.4).

Figure 8.4: Dataflow in the REDI processing environment, showing waveform data coming in from the Quanterra data loggers (Q) into comserv. From comserv, data are logged to disk (via datalog), distributed to other computers (mserv), fed into the CDA for REDI processing, and spooled into a tracering for export.

The two computers that perform data acquisition also serve as REDI processing systems. In order to facilitate REDI processing, each system maintains a shared memory region that contains the most recent 30 minutes of data for each channel used by the REDI analysis system. All REDI analysis routines first attempt to use data in the shared memory region, and will only revert to retrieving data from disk files if the requested data is unavailable in the shared memory region.

Most BDSN and NHFN stations transmit data to only one or the other of the two REDI systems. In earlier system configurations, each station would transmit data to one of the two systems, which would write the data to local disk files, copy the data in its own shared memory region, and transmit the data via a socket to the other system's shared memory region. Each REDI system's shared memory region contained data from all stations, but each computer's filesystem contained data from only one half of the network. The REDI systems use the Network File System (NFS) to access remote files that reside on other computers. If a REDI analysis program required data that was not in the shared memory region, it would attempt to retrieve the data from the disk files from both REDI computers. If one of the two REDI computers was unavailable, this could cause the REDI processing to hang waiting for access to the other computer's files.

During 2000, we revised our data acquisition procedure to use two programs developed at Caltech. The comserv client program cs2m receives data from a comserv and multicasts the data over a private ethernet. The program mcast, a modified version of Quanterra's comserv program, receives the multicast data from cs2m, and provides a comserv-like interface to local comserv clients. This allows each REDI system to have a comserv server for every BDSN station. We added additional disk space to both of the REDI computers so that we use datalog to write a copy of data to each computer's filesystem. We reconfigured the REDI computers to only retrieve data from their own filesystems, thereby preventing a loss of one REDI computer from possibly hanging the other REDI computer's processing.

We extended the multicasting approach to handle data received from other networks such as the NCSN and UNR. These data are received by Earthworm data exchange programs, and are then converted to MiniSEED and multicast in the same manner as the BDSN data. We use mserv on both REDI computers to receive the multicast data, and handle it in an identical fashion to the BDSN MiniSEED data.

Background Noise

BDSN data are routinely monitored for state-of-health. An automated analysis is computed weekly to characterize the seismic noise level recorded by each broadband seismometer. The estimation of the Power Spectral Density (PSD) of the ground motion recorded at a seismic station, provides an objective measure of background seismic noise characteristics over a wide range of frequencies. When used routinely, the PSD algorithm also provides an objective measure of seasonal and secular variation in the noise characteristics and aids in the early diagnoses of instrumental problems. A PSD estimation algorithm was developed in the early 1990's at the BSL for characterizing the background seismic noise and as a tool for quality control. As presently implemented, the algorithm sends the results via email to the engineering and some research staff members and it also generates a bargraph output which compares all the BDSN broadband stations by components. A summary of the results for 2000-2001 is displayed in Figure 2.3.

We have expanded our use of the weekly PSD results to monitor trends in the noise level at each station. In addition to the weekly bar graph, additional figures showing the analysis for the current year are produced. These cummulative PSD plots are generated for each station and show the noise level in 5 frequency bands for the broadband channels. These cummulative plots make it easier to spot certain problems, such as failure of a sensor. In addition to the station-based plots, a summary plot for each channel is produced, comparing all stations. These figures are presented as part of a noise analysis of the BDSN on the WWW at http://seismo.berkeley.edu/bdsn/psd/ and are part of the recently revamped BDSN Web pages at http://seismo.berkeley.edu/bdsn/.

Over the past year and a half, an exportable version of the PSD algorithm has been developed, with IRIS funding and for general use by IRIS and the seismological community. In order to facilitate portability and ease of use, the algorithm has been redesigned to acquire all requisite station transfer function information and seismic time series data from a SEED data volume. The algorithm has also evolved during the past year as feedback from users has led to improvements in the code. In particular, the algorithms which decode the transfer function information from the sensor information in the SEED volume and which smooth the PSD estimates at both ends of the frequency spectrum have been improved.

PSD algorithm

Description

In order to facilitate portability, the PSD algorithm is designed to acquire the requisite station, sensor, and time series data from a SEED data volume (Halbert et al., 1993). The "rdseed" algorithm developed by IRIS personnel is invoked to extract the station data, the sensor transfer function data, and to dump the time series data records in SAC binary format (Tapley and Tull, 1992). The resulting SAC binary format files are processed to calculate the background noise PSD.

Table 8.1: Abridged example of PSD algorithm results.

The algorithm uses a statistical approach to robustly estimate the background noise PSD. The PSD estimates are reported in dB relative to 1 (m/s$^2$)$^2$/Hz. The input time series is parsed into eight (possibly overlapping) time series and each of the resulting time series are appropriately windowed prior to calculating their PSD estimates. For short time series, less than 1.5 hours in length, the time series are detrended and sine tapered while for longer time series the dominant semi-diurnal gravitational tide signal is also removed to avoid biasing the long-period PSD estimates. The PSD estimates are smoothed and reported at twenty logarithmically spaced intervals per decade in period.

Owing to the statistical nature of the PSD algorithm, it is required that the time series to be processed contain at least 65,536 (2$^{16}$) contiguous samples. Shorter time series are not processed and a warning is issued. The PSD algorithm can process data with a wide variety of sampling rates (from $<$0.01 sps to $>$500 sps). A typical usage with broadband data is for the time series to contain one day of continuous LH (1 sps) data (86,400 samples). Since the sensor transfer function representation in the SEED data volume for a typical inertial seismometer does not include the static component of the response, the background noise PSD estimates for periods longer than approximately an hour will be biased high and hence they will be unreliable.

Figure 8.5: Background noise PSD estimates for the BDSN station YBH, sited in an abandoned hard rock mine in the Klamath National Forest, 430 km N of Berkeley. PSD estimates for eight one-day intervals, picked at random over the past three months, are plotted as solid lines. The low and high background noise models (Peterson, 1993) (dashed lines) are given for comparison. YBH exhibits consistently low background noise PSD levels at low frequencies and it is one of the quietest BDSN broadband stations.

Source Code

The original BSL in-house version of the PSD algorithm was written in Fortran 77. At the request of IRIS, the f2c translator was used to convert the Fortran source code to C source files. The source code was cleaned up and the necessary subroutines, not included in the f2c libraries, were added to expedite the conversion process. The only f2c option used in the conversion was "-c" which embeds the original Fortran 77 source code as comments in the C source files.

The PSD code distribution along with examples of its usage are available via the IRIS DMC home page (http://www.iris.washington.edu).

Sample Output

Besides the information that is printed to the screen when PSD is executing, it outputs a station coordinate file ``instr.sdv.coord'', a sensor transfer function file ``instr.sdv.resp'', and a file with a SEED naming convention prefix and a ``.psd'' suffix for each data stream, that meets the minimum requirements of the PSD algorithm, extracted from the SEED data volume.

An abridged sample of the PSD output file ``YBH.BK.LHZ.XX.D.2000.178.0000.psd'' is given in Table 8.1, where the first 17 lines contain header information and the 18$^{th}$ through the last line give the PSD estimates. For each PSD estimate, the first column gives the period in seconds and the remaining columns give the lower bound (LB), minimum (min), median (med), maximum (max), and upper bound (UB) of the background noise acceleration PSD estimate. The minimum and maximum columns refer to the estimates which have the minimum and maximum sums over all periods (the time series corresponding to each is given in the 13$^{th}$ header line) while the lower bound (LB) and upper bound (UB) columns give the lowest and highest estimate for each period. The PSD estimate in the ``PSD min'' column is the most robust estimate of the background noise PSD.

A plot of the PSD minimum estimate versus period, observed at BDSN station YBH, is given in Figure 8.5.

Sensor Testing Facility

Introduction

The BSL has set up an instrumentation test facility in the Byerly Seismographic Vault (BKS) in order to systematically determine and compare the characteristics of up to eight sensors at a time. The test equipment consists of an eight channel Quanterra Q4120 high-resolution data logger and a custom interconnect panel that provides isolated power and preamplification (when required) in order to facilitate the interconnection and routing of signals from the sensors to the data logger with shielded signal lines. Upon acquisition of the 100 samples-per-second (sps) data from the instruments under test, PSD analysis and spectral phase coherency analysis are used to characterize and compare the performance of each sensor. Tilt tests and seismic signals with a sufficient signal level above the background seismic noise are also used to verify the absolute calibration of the sensors. Also, a simple shake table was constructed from a vertical seismometer and used to determine the linearity of a seismic sensor.

Figure 8.6: 8-channel Quanterra Q4120 data logger (lower left) that records the data from the instruments undergoing testing in the Byerly Vault.

Background and Motivation

BSL personnel have tested numerous sensors during the past several years and each test has been ad hoc in its implementation and execution. In order to expedite the setup and testing of the instruments for both BSL in-house use as well as for other groups, we have dedicated an eight-channel Quanterra Q4120 data logger (see Figure 8.6) and constructed cabling and a patch panel (see Figure 8.7) to enable the simultaneous testing of up to eight sensors with high (24-bit integer) resolution sampling at 100 samples per second (for a usable band width of 0-32 Hz). We housed the test equipment in Byerly Seismographic Vault, located in an old mining drift in the Claremont shales and cherts formation in Strawberry Canyon behind the Botanical Garden, because the site is seismically quieter than the BSL facilities on the UC Berkeley campus and because it has all the necessary infrastructure for a test facility, including a set of Streckeisen STS-1 broadband seismometers and a Kinemetrics FBA-23 strong motion accelerometer that are routinely recorded by the BDSN station BKS sited in the same vault and which provide reference seismic recordings of the ground motion.

Figure 8.7: Interconnecting patch panel (top center) which routes the signals from the sensors under test (lower left and STS-2 reference sensor center right) to the Q4120 data logger.

Sensor Testing Methodology

Sensor testing has three aspects: 1) data acquisition; 2) sensor calibration; and, 3) sensor performance.

The acquisition of high resolution digital data is absolutely essential for the testing of modern sensors. The Quanterra Q4120 data logger, which has 24-bit integer resolution (144 dB dynamic range) with a LSB of 2.38 $\mu$V and a 40 V P-P signal handling capability, is ideally suited for this task. Data acquisition has two modes, active and passive. In the active acquisition mode, a signal is induced by appropriate means as described below. In the passive acquisition mode, the sensors signals are recorded for extended periods of time, typically from a day or so to a couple of weeks or more, to acquire samples of both the background noise level as well as various natural seismic, gravitational tide, and atmospheric signals.

The calibration methodology employed is sensor dependent. In the case of broadband seismometers, induced tilting, dynamic shaking (see shake table section), dynamic driving of a calibration coil, the gravitational tide signal, and comparison with ground motions inferred from known sensors are used as appropriate. In the case of strong motion accelerometers, induced tilting, dynamic shaking, and comparison with ground motions inferred from known sensors are used as appropriate. In the case of barometric pressure sensors, static elevation changes and comparison with the atmospheric pressure inferred from known sensors are used as appropriate.

Performance of a sensor is primarily characterized by determining the usable dynamic range, the linearity, and the noise characteristics. The usable dynamic range is the difference, usually expressed in dB, between the sensor saturation signal level and the sensor or background noise level. The usable dynamic range is usually frequency dependent. Characteization of the linearity and dynamic range of a broadband seismic sensor is problematic because seismic signals are a wide bandwidth transient phenomena (see the linearity and dynamic range section below). The linearity is characterized by determining the third-order intercept point. The usable dynamic range is quantified by the difference between the sensor saturation (full scale) signal level and the background signal or sensor noise floor and it is usually expressed in dB. PSD estimates of the background seismic (or sensor self noise), plotted as a function of frequency, are used to concisely quantify the performance of a broadband sensor (and also to quantify differences in the noise levels observed in various seismic vaults housing broadband seismic sensors). Another approach for characterizing the bandwidth and performance of a particular type of sensor is to simultaneously record two identical sensors placed on the same pier and calculate the phase coherency of their outputs as a function of frequency.

All of the above approaches to sensor testing have been employed in our analysis during the past year.

Figure 8.8: Simple vertical shake table, constructed using a Johnson-Matheson Model 6840 short-period seismometer, with a PMD sensor mounted on top of the shake table for testing.

Sensor Testing Algorithms

A number of computer algorithms have been developed over the past few years to aid in the testing and characterization of seismic and other types of sensors. As well as routinely viewing the recorded waveforms in the time domain, always very useful and highly recommended for detecting and identifying various types of problems, we also look at the output of several algorithms which do complex spectral domain analysis. The most useful of these algorithms has been the background noise PSD estimation algorithm described above. In addition, a few ad hoc algorithms have been also developed for investigation of noise problems encountered in the deployment and operation of the Hayward and Parkfield borehole networks. These algorithms were tailored to characterize, in part, specific spectral noise peaks generated by power line or spread spectrum radio interference.

Figure 8.9: Third-order intercept point plot for PMD sensor, measured at 4 Hz.

PMD Sensor Test

Starting in spring 2000, we acquired a pair of PMD Scientific, Inc. broadband three-component sensors for testing. Our primary interest was to compare the waveforms from the PMD sensors and the co-sited STS-1 sensors at BKS and to characterize the linearity and the useful dynamic range of the PMD sensors. The PMD sensors use an unique Molecular Electronic Transfer (MET) transducer (see Abramovich et al., 1997 for a detailed description) to detect the velocity of a highly conductive electrolytic fluid. It is noteworthy that, prior to the introduction of the PMD MET transducer, all seismic sensors with velocity transduction invariably used a coil and a magnet.

Owing to the lack of an electronic feedback mechanism in the PMD instrument, as is used in all modern high-performance broadband seismic sensors to improve their stability and linearity, the question arose as to its linearity. To answer the linearity question, we developed a simple shake table, as described below, to actively shake the PMD sensor and determine its response. The generation of PMD sensors tested by the BSL do not employ any feedback mechanism. We are scheduled to receive a pair of the latest generation PMD MET broadband seismometers, which use a magnetohydrodynamic feedback system to lower the self noise and to improve stability, for testing within the next couple on months.

The first part of the testing was to characterize the noise floor of the two PMD sensors and their response to seismic events. The calibration is checked by comparing microseismic ground motions inferred from the PMD sensors and the co-sited STS-1's. We found that the PMD self noise was consistent with the published specifications. The second part of the test was to determine the spectral phase coherency between the two Z-component PMD's and the Z-component STS-1. We found that the phase coherency was, as expected, higher over a wider bandwidth between the two PMD sensors than between either PMD sensor and the STS-1 owing to the higher frequency passband of the PMD sensors.

Shake Table

The problem of how to shake the seismic sensors in order to test the linearity was solved by constructing a vertical component shake table from a Johnson-Matheson (J-M) Model 6840 short-period vertical seismometer as shown in Figure 8.8. The large and stiff suspension of the J-M was ideally suited for modification to a simple shake table. A platform to hold the sensor and extra springs to support the additional weight we added to the J-M seismometer so that vertical shaking could be induced by actively driving the J-M signal coil at relatively low current levels. The natural frequency of the resulting shake table was $\approx$1.6 Hz. We found that the shake table could be driven over a useful range of $\approx$50 dB, from $\approx 3.4x10^{-6}$m rms (limited by the background noise PSD level of $\approx$-136 dB at BKS) to $\approx 10^{-3}$m rms (limited by the shake table suspension travel), and over a 0.2-5 Hz frequency range by driving the signal coil sinusoidally with a WaveTek function generator. Vertical and horizontal Wilcoxon 731A accelerometers are attached to the housing of the sensor under test to directly measure the amplitude and spectral characteristics of the induced shaking. The table and sensor were arranged so that their center of mass was on the center line of the J-M suspension axis in order to minimize cross axis coupling.

Linearity and Dynamic Range

Characterization of the linearity and dynamic range of a broadband seismic sensor is problematic because seismic signals are a wide bandwidth transient phenomena while typical linearity tests utilize a narrow bandwidth continuous signal. One figure of merit that is indicative of the quality with which a seismic sensor can reproduce the ground motion signal is Total Harmonic Distortion (THD). The THD, in percent, is defined as:

$$THD = 100 \frac{\sqrt{ \sum\limits_{k=2}^N a_k^2 }}{a_1}$$

where: $a_1$ is the rms level of the fundamental drive signal and the $a_k$ are the rms levels of the harmonics that result from non-linear processes in the seismic sensor. Given the difficulties in reliably determining the sum of the $a_k$'s, we opted to measure instead the size of the third-order harmonic ($a_3$) as a function of the sinusoidal drive signal ($a_1$) and to determine the third-order intercept point. The third-order intercept point, i.e. the projected asymptotic point at which the fundamental and third harmonic signal levels are equal in amplitude, provides a useful figure of merit for the quality of a ground motion signal recorded by a seismic sensor. The methodology for determining the third-order intercept point is shown in Figure 8.9 where the fundamental and third harmonic output velocities are plotted as a function of the sinusoidal input velocity at a given frequency. Asymptotes are drawn through the fundamental and third harmonic slopes and their intersection is the third-order intercept point. Spectral analysis of the vertical and horizontal shake table motion, as recorded by a Wilcoxon 731A accelerometers mounted on the PMD sensor housing, shows that the non-linearity in the shake table suspension is $\approx$50 dB at the drive frequency and below the noise floor at the third harmonic frequency. Thus the PMD sensor and not the shake table is the source of the observed harmonic distortion.

Figure 8.10: View of the Druck PTX-1240 transducer (mounted inside the pipe and union joint) and the transducer electronics test board. The Druck transducer sensing port has been closed off with a length of 3/8" pipe to facilitate measurement of the sensor self noise.

Figure 8.11: Comparison of atmospheric pressure variation over a four day interval measured at two NDBC stations and by two sensors at the BKS vault. For comparison, the sea level NDBC buoy pressure data have been corrected to the 276 m BKS elevation using the standard formula (where z is the elevation above mean sea level in meters): $P(z) = P_o ( 1 - 0.02255 z )^{5.256}$.

Barographic Sensor Test

We have been testing a Druck PTX-1240 barometric sensor and various configurations of the associated hardware interface to determine the optimal configuration and the performance. The transducer, housing and support electronics are shown in Figure 8.10. The interface electronics is designed with a bias circuit to null the long term average atmospheric pressure and with a gain of 40 to achieve the desired sensitivity. The transducer housing is designed to be installed inside the YBH vault in a manner that provides good thermal coupling to the rock in order to achieve adequate thermal stability and the housing will be connected to a manifold that leads to the exterior of the vault where it will be connected to a noise reduction system of permeable hoses.

A sample recording of the barometric test system is shown in Figure 8.11. By comparing the long-period signals from the test system with the barometric pressure recorded hourly (with 10 Pa resolution) at nearby National Data Buoy Center (URL: http://www.ndbc.noaa.gov) buoys, at San Francisco ($\char93$46026; 53 km W) and at Half Moon Bay ($\char93$46012; 69 km SW), we empirically determined that the sensitivity of the Druck PTX-1240 pressure sensor (recorded by the Q4120 data logger at 100 samples per second (sps), scaled by a factor of 16, FIR LP filtered, and decimated to 1 sps) is (91$\pm$3) $\mu$Pa/count. When the Druck pressure transducer is installed at YBH and recorded at 1 sps, the anticipated sensitivity is (1.46$\pm$0.05) mPa/count.

The pressure inferred from the Motorola MPX-2010 silicon pressure transducer signal is shown as the dashed line in Figure 8.11. The compliance of this sensor is obviously attenuated at long periods and it is likely that the reference port seal on the Motorola transducer is leaking with a time constant of order 1 hour. This is consistent with the roll-off in the phase coherency between the two pressure transducers at BKS at frequencies below $\approx$0.1 mHz as shown in Figure 8.12.

The estimated pressure background noise PSD in the BKS Vault, as inferred from the Druck PTX-1230 pressure transducer recording, is shown in Figure 8.13. The noise floor of the Druck sensor system, as indicated by the dashed line is -44 dB when extrapolated to 1 Hz or about 20 dB higher than the International Monitoring System (IMS) sensor noise requirement. It appears that a large fraction of the self noise is due to the characteristics of the interface between the pressure transducer and the digitizer and we are exploring ways of reducing the noise floor.

Figure 8.12: Phase coherency between Druck PTX-1240 and Motorola MPX-2010 pressure transducers co-sited in the Byerly Seismographic Vault. The coherency fall off at low frequencies due to a problem with the Motorola sensor and at high frequencies due to differences in the pressure admittance between the two sensors owing predominantly to degradation of the signal-to-noise level in, and the placement of, the Motorola sensor within the sealed urethane foam box which surrounds the broadband seismic sensors.

Figure 8.13: Estimates of the atmospheric pressure noise PSD measured using the Druck PTX-1240 barometer inside the BKS Vault. The upper curve is measured with an open sensing port which is sensing the ambient pressure inside the vault. The lower curve is measured with the sensing port closed off with a short length of capped pipe as shown in Figure 8.10, in order to estimate the self-noise of the pressure transducer system. The corner at 0.20 Hz is due to restricted air flow in the vault and opening the internal vault doors raised the pressure noise PSD to the level indicated by the diamonds.

IRIS Instrumentation Test

The BSL instrumentation test facility was employed to characterize eight sensors for the IRIS Instrumentation Committee. The sensors, listed in Table 8.2, were tested concurrently with 100 Hz data acquired via the Q4120 data logger. The accelerometers under test, and of particular interest to IRIS, are shown in Figure 8.14. PSD analysis was used to compare and characterize the performance of each sensor. Tilt tests and seismic signals above the background seismic noise are also used to verify the absolute calibration of the sensors.

Figure 8.14: Picture of six accelerometers which were tested simultaneously. The Episensor is on top, the MEMS on the bottom, and from left to right center is the Wilcoxon X731B (large cylinder), the Endevco 86 (small cylinder), the Wilcoxon 9XL (aka IRIS SUMS) (small sensor with cable attached), and the RefTek 131-02 (rectangular). There are also three duplicate sensors present.

Table 8.2: Sensors tested for IRIS and transduction constants. Notes: 1- New MET sensor but without feedback circuitry; 2- STS-2 used as reference for comparison; 3- aka IRIS SUMS; 4- Includes 100x preamplification; 5- Includes 400x preamplification.

Manufacturer	Model	Serial Number	Sensitivity	Sensitivity
			(w/ preamp)	(sensor)
				(V/g)
Wilcoxon	X731B	P01	41.7 V/(m/s$^2$)	4.26
PMD	MET$^1$	none	1960 V/(m/s)	-
Streckeisen	STS-2$^2$	20022	1500 V/(m/s)	-
Endevco	86	AA01	74.4$^4$ V/(m/s$^2$)	7.29
MEMS	SF1500A	358	0.152 V/(m/s$^2$)	1.49
Kinemetrics	Episensor	812	7.59 V/(m/s$^2$)	74.4
Wilcoxon	9XL$^3$	P002	24.2$^4$ V/(m/s$^2$)	2.37
RefTek	131-02	251	97.1$^5$ V/(m/s$^2$)	2.38

An example of the raw data, which includes the P-wave from a recent major teleseism (M 7 06/03/2001 Kermedec Is., depth 191 km, distance 85 degrees), is shown in Figure 8.15. Note the wide variation in the signal characteristics which is due to differences in the sensor sensitivity, transduction, and noise characteristics. In this plot, and in subsequent plots, only the sensors owned by the BSL are identified. As part of this collaboration with the manufacturers, we have agreed not to identify the other sensors and so have labelled then Sensor A through Sensor D in Figures 8.15 - 8.18).

Figure 8.15: Sensor raw data. The signal in the middle of the plot is the P-wave from a major teleseism. Note that the raw output of two of the sensors, the MET and the STS-2, is proportional to velocity while the output of the others is proportional to acceleration.

A comparison of the absolute ground accelerations inferred from the signals recorded by each of the eight sensors is shown in Figure 8.16. Note the variation in the signal-to-noise ratio (SNR) and frequency characteristics of each sensor. As in Figure 8.15 the peak signal in the middle of the plot is the P-wave from the Kermedec teleseism.

Figure 8.16: Absolute ground acceleration inferred from each of the eight sensors. The calibrated complex response for each sensor has been deconvolved in the frequency domain to obtain the ground acceleration.

Figure 8.17 is a plot of data from Figure 8.16 which has been 0.3-2 Hz bandpass filtered to enhance the SNR of the teleseismic P-wave. The eight traces are plotted on an absolute scale and the P-P signal level is 335 $um/s^2$ (0.0335 gal) (-87 dB PSD in the 0.3-2 Hz band). Note that all sensors recorded the P-wave and that the sensors with the highest noise PSD levels (as shown in Figure 8.18 below exhibit a degraded SNR. The transduction constant for each of the sensors under test was empirically determined by comparison of their inferred ground acceleration against the ground acceleration inferred from the STS-2 recording and the results are given in Table 8.2.

Figure 8.17: 0.3-2 Hz bandpass filtered absolute ground acceleration inferred from each of the eight sensors (filtered version of Figure 8.16).

Figure 8.18 shows the background noise PSD for each of the eight sensors under test. Each plot shows PSD estimates from six 45-minute long time series taken hourly starting at 2001.152.0500 UT. The scales are all identical and, for reference, the background noise PSD for each sensor (except for the STS-2) includes the STS-2 noise PSD estimate shown as dashed lines. Note that four of the sensors (the Wilcoxon X731B, the PMD MET, the Streckeisen STS-2, and the Kinemetrics Episensor) track the background seismic noise level, observed at BKS, in the $\approx$0.07-7 second period band.

Figure 8.18: Background noise PSD estimates for each of the eight sensors being tested. The axes and scales are identical for all figures. The background noise PSD determined from the Streckeisen STS-2, shown in the top center figure, is duplicated in the other figures for comparison.

BDSN Instrumentation Tests

In addition to the above testing, we have tested several BDSN broadband seismometers and strong motion accelerometers. New, and recently repaired, sensors are routinely tested before deployment in the field. Also, we had a few BDSN sensors, both accelerometers and broadband sensors, which either become noisy or malfunctioned during the past year and they were tested at the test facility in the Byerly Seismographic Vault to identify and characterize the problem.

Experiment with the Japanese UrEDAS

Introduction

The established joint notification system in Northern California (i.e., REDI) provides accurate and reliable determination of earthquake parameters, but there is a time delay between the occurrence of an event and the determination of its size. In an emergency, this time delay prevents actions which could mitigate damage from strong ground shaking. The present configuration of the system could be improved with the capabilitiy of rapid size determination. In addition, it will be advantageous to have detection capabilities independent of the dense short-period network, especially in an emergency when communications may be disrupted.

In an effort to develop such capability with the BDSN, we have started an experiment collocating a set of UrEDAS (Urgent Earthquake Detection and Alarm System; see Nakamura, 1996), an integrated real-time earthquake warning system, at BKS. The UrEDAS detects an earthquake using only 3 sec of the P-wave recorded at a single station and has been used in Japan for over a decade to alert bullet trains (which travel at a peak speed of close to $\sim$200 km/hr) to strong shaking in progress. While the weakness of UrEDAS that it is less accurate in event parameter determination, its advantage is speed. Through this experiment we explore the development of a single-station capability of rapid earthquake detection and location/magnitude estimate using data from the BDSN stations.

Figure 8.19: UrEDAS sensors installed at BKS site. The monitor and processing system are located in the next room (not shown here).

Figure 8.20: Schematical illustration of the UrEDAS collocation experiment with an STS-1 broadband instrument and an FBA-23 strong motion seismometer with a Quantera data logger Q980 at BKS.

UrEDAS collocating experiment at BKS

In November 1999 the crew of System and Data Research (SDR) visited the BSL to discuss a possible joint experiment with UrEDAS. The seismologists at BSL agreed to test the UrEDAS performance by collocating it at one of the BDSN stations. We chose the Byerly Vault both for its low noise levels and relative ease of access. As described above, Byerly Vault contains a set of Strekeisen STS-1 seismometers and an FBA-23 strong motion package as part of the BDSN station BKS. This site has been occupied by the BSL since 1962 and the site specific characteristics in seismic waves are fairly known.

The initial system installation was completed with the event detection and notification in February 2001 and was upgraded to transmit waveform data to the BSL in July 2001. Figures 8.19 and 8.20 show a picture of the UrEDAS sensor and an illustration of the network configuration.

Figure 8.21: Comparison: a. azimuths and b. magnitudes determined by UrEDAS and events archived in the CNSS composite catalog.

In the UrEDAS system the event detection level of velocity is pre-set; the epicentral azimuth is estimated from the direction of the initial motion projected on the horizontal plane; and the preliminary estimate of the distance and magnitude is based on the frequency content and amplitudes of P-wave first motions ($\sim$3 sec). An alarm can be issued if a hazardous earthquake is detected by P-waves. After S-wave arrival the preliminary estimate is revised.

Since the installation, $\sim$100 events (1.3$\leq$M$\leq$8.1 according to the CNSS catalog) including teleseismic events were detected. The UrEDAS performance was evaluated by comparing the event parameters with those recorded in the CNSS composite catalog. The event detection performance was satisfactory, although UrEDAS is designed to detect primarily local events (R$\leq$200 km) and does not have ability to distinguish between teleseismic and local events at present.

The epicentral distance ($R$) is estimated using the relation $log R = a \cdot log A + b \cdot log T + c \mbox{ }$ where $A$ is the amplitude of the initial P-wave motion (in mkine), $T$ its prominent period, and $a, b,$ and $c$ are constant. The magnitude is estimated from the prominent period ($T$) of the initial P-wave motion using the relation $M=3.2 \cdot log T + 5.26$. We do not suppose that these relations apply universally but are testing them empirically. The azimuth determination shows systematic biases, most likely due to the nearby faults where the impedance can change by $\sim$40% (Fig. 8.21a). The erroneous location estimates can be also attributed to the propagation path effects through the faults, the near-site structural heterogeneity and/or noise level. If the azimuth estimate becomes reliable, the combined information from two stations could also make a reasonable estimate of an epicentral distance.

The estimated magnitudes of small local events in the epicentral distance between 20 and 200 km were within the range expected from other experiments (see the comparison of magnitudes between UrEDAS and the CNSS catalog in Figure 8.21b). Magnitudes of smaller events (M $<$ 2.5 as listed in the CNSS catalog) tend to be overestimated, and those of events at farther distances (R$>$200 km) are underestimated.

Figure 8.22 illustrates the UrEDAS location determination for a local event that occurred on 04/26/2001 in comparison with the UCB/USGS determination. UrEDAS was triggered at 04:52:33 (22 sec after the origin time), estimated the size and location using the 3-sec P-wave, and the version 1 information was received by the BSL computer at 04:52:39, i.e., $\sim$6 sec after the detection. The magnitude estimated by UrEDAS ($M$2.6) is very close to $M_{L}$2.7 by UCB/USGS. Version 2 parameters were estimated including S-wave in $\sim$16 sec. The revised epicentral distance 112 km is very similar to the UCB/USGS estimate, although the absolute location is shifted due to the effects of Hayward Fault in the estimation of azimuth.

Figure 8.22: UrEDAS location in version 1 (Ur1) using 3 sec of P-wave, and version 2 (Ur2) including also S-wave are compared with the location determined by REDI. The estimated azimuth in version 1 is in the same quadrant as that by the UCB/USGS. While the absolute location is mislocated, the distance from Ur2 to BKS is very close to that estimated by the UCB/USGS. $M_{L}$ is 2.6 by UrEDAS and 2.7 by the UCB/USGS.

We have done some preliminary comparison of the waveform data recorded by the BKS broadband instrument with those recorded by the UrEDAS. Because of the complexities of seismic structure, nonlinearities involved in the propagation of the complex faults areas, this problem does not lend itself to easy analysis without systematic and more advanced analyses and calibrations. We focus on improving the algorithm to rapidly evaluate preliminary earthquake source parameters, i.e., magnitude and location.

Discussion

The magnitude and location estimation using just the first 3 sec of P-waves is not always reliable. It is not always easy to detect an event from P-wave alone as P-wave first motions are relatively weak to exceed a certain pre-set trigger level, or if the trigger level is lowered, then false alarms may be issued frequently. It is also possible that the station location happens to be near the nodal of the focal mechanism. Assuming that the primary goal is to determine the event location and size as rapidly as possible, the fastest approach will prove to be a hybrid approach where the remote stations determine the azimuth and ramp growth rate and associated uncertainties and the central site uses a fuzzy logic algorithm to determine the location and size of the event. The primary advantage of this hybrid method is that the ramp growth rate can be reliably determined before the S-wave arrives. In the limiting case, and with a sufficiently high station density, one could even go so far as to determine and report from the remote sites using only the broadband P-wave impulse, and associated azimuth and apparent angle of incidence and estimates of their resolution and have the central site coalesce the data into a viable and rapid event report.

The critical issue for a successful installation of a UrEDAS type system in the BDSN is the calibration of specific site effects at individual stations. A joint use of the single station detection system with the current northern California earthquake notification system would significantly increase the capability of real-time earthquake warning system.

Acknowledgements

Under Barbara Romanowicz's general supervision, and with Doug Neuhauser and Bob Uhrhammer as head gurus, Lind Gee, and Rick McKenzie are involved in the data acquisition and quality control of BDSN and NHFN data. Development of the exportable version of the PSD algorithm was funded by the IRIS Consortium and performed by Bob Uhrhammer. We thank the personnel at the IRIS DMC and PASSCAL Instrument Center for helpful feedback.

Development of the sensor test facility and analysis system was a collaborative effort of Bob Uhrhammer, Tom McEvilly, John Friday, and Bill Karavas. IRIS and DTRA provided, in part, funding and/or incentive to set up and operate the facility and we thank them for their support. We also thank Igor Abramovich of PMD Scientific, Inc. for providing the opportunity to test a pair of their unique MET broadband seismic sensors.

Fumiko Tajima led the collaboration with SDR on testing the UrEDAS at BKS and has worked with Bob Uhrhammer to evaluate the system. Doug Neuhauser, Bill Karavas, John Friday, and Dave Rapkin helped with installation and maintenance. We thank Yutaka Nakamura and his colleagues at SDR for providing us with the installation of UrEDAS system and information on the accumulated data by this system.

Bob Uhrhammer, Doug Neuhauser, Lind Gee and Fumiko Tajima contributed to the preparation of this chapter.

References

Abramovich, I. A., V. M. Agafonov, M. E. Cobern, and V. A. Kozlov, Improved Wide-Band Molecular Electronic Seismometer and Data Acquisition system, Poster Session S31B-18, AGU Fall Meeting, San Francisco, 1997.

Halbert, S. E., R. Buland, and C. R. Hutt, Standard for the Exchange of Earthquake Data (SEED), Version V2.0, February 25, 1988. United States Geological Survey, Albuquerque Seismological Laboratory, Building 10002, Kirtland Air Force Base East, Albuquerque, New Mexico 87115, 82 pp., 1996.

Nakamura, Y., Real-time information systems for seismic hazards mitigation UrEDAS, HERAS and PIC, Quarterly Report of Railway Technical Research Institute, 37,, 112-127, 1996.

Peterson, J. Observations and Modeling of Seismic Background Noise, U.S. Geological Survey Open File Report 93-322, 94 pp., 1993.

Tapley, W. C. and J. E. Tull, SAC - Seismic Analysis Code: Users Manual, Lawrence Livermore National Laboratory, Revision 4, 388 pp., March 20, 1992.

Uhrhammer, R. A. and T. V. McEvilly, IRIS Sensor Tests: an interim report, Poster Presentation, 2001 IRIS Workshop, Jackson Lake Lodge, Moran, WY, June 6-9, 2001.

Berkeley Seismological Laboratory, 202 McCone Hall, UC Berkeley, Berkeley, CA 94720-4760
Questions or comments? Send e-mail: www@seismo.berkeley.edu
© 2001, The Regents of the University of California.