Remote Bridge Monitoring Using Infrasound

Infrasound sensors are commonly deployed in networked, multisensor arrays for localization and characterization of signals; thus, multiple arrays are required for optimal data. For this experiment, three arrays were deployed with sites identified through considerations for space availability, security, wind protection, topography, distance from source to receiver, and prevailing meteorological conditions (Simpson et al., Forthcoming). Arrays were located at the Sutter County Airport, the Marysville City Cemetery, and Beale AFB at distances of 3.0, 2.6, and 24 km, respectively. Each of these array locations was separated from the bridge of interest by urban terrain, industry, and/or significant topographical features, such as levee systems. For analysis, these arrays were denoted as FRA for the Sutter County Airport array, FRB for the Beale AFB array, and FRC for Marysville City Cemetery array. Fig. 1 presents the location of each array in relation to the bridge as well as a layout of each array.

Data collection took place over 7 days from February 26 through March 4, 2014, with data sampled at 1,000 Hz, and the gain was set at unity for all infrasound sensors. Each of the five-element infrasound arrays consisted of five infrasound sensors with porous hoses attached to the four inlet pipes as passive wind filters, two digitizers, and associated cables, connectors, batteries, and solar panels. Some specifications for the equipment used are included here, but additional information can be found in the work by Simpson et al. (Forthcoming). The Inter-Mountain Laboratories (IML, Sheridan, Wyoming) Model ST infrasound sensors utilized in this experiment had a nominal signal frequency band of 2–30 Hz, but content could extend below 2 Hz with roll-off beginning at approximately 1 Hz. Data from these components were streamed to REF TEK (Plano, Texas) 130S-01 digitizers with an input preamplifier and digital antialias filters and a high-precision external Global Positioning System (GPS) receiver/clock for time synchronization between each digitizer through the course of the experiment. Arrays were deployed in a 60 × 60-m cross pattern with infrasound sensors at the cardinal points and one infrasound sensor at the center. The cross configuration gives omnidirectional detection capabilities while simplifying data processing.

Understanding atmospheric conditions is imperative to monitoring infrasound, because the atmosphere is a time-varying propagation medium with temperature and wind effects dominating propagation characteristics. During this experiment, radiosondes were launched to capture the atmospheric conditions. The radiosondes sampled the atmosphere once per second from balloon release at ground level to balloon bursting height at maximum altitude (balloon dependent). This deployment used 350-g balloons reaching an average altitude of approximately 25 km and a travel time of approximately 2 h. Data obtained by the radiosondes included temperature, pressure, humidity, wind direction and speed, altitude, and GPS coordinates. To minimize interruption to flight operations and training at Beale AFB, launches were limited to every 6 h, starting at 0600 on March 1 and continuing through March 2, 2014. Radiosonde launches occurred from the Sutter County Airport.

Two REF TEK 130-SMA strong motion accelerographs were also deployed at Br 18–0009 (one on the sidewalk and one directly below it at the base of a pier, with both pointing north) on March 3, 2014, with the intention of gathering data for 24 h to try to capture the natural frequencies of Br 18–0009. The intent was to use these frequency data as validation for the collected infrasound data. Unfortunately, the sensors were vandalized while deployed, rendering the data unusable. A second experiment was conducted to try to capture these on-structure data as well as collect additional infrasound data in the area. Contact with Caltrans indicated that conditions at the structure had not changed from the first experiment to the second, making comparison of infrasound data from the first experiment to the on-structure instrumentation results in the second experiment valid.

The second experiment took place November 13–18, 2015, and focused primarily on capturing the fundamental modes of Br 18–0009 utilizing on-structure instrumentation. In addition, two of the three array locations from the first experiment were reoccupied. Initial data-analysis efforts yielded no detection of the low-energy signals at FRB during the initial times of interest (temperature inversions) identified in the collected radiosonde data; therefore, the decision was made not to instrument this location for the second experiment. The arrays at the Sutter County Airport (FRA) location and the Marysville City Cemetery (FRC) location utilized the same instrumentation as described for the first experiment. A single sensor was placed at the bridge during this experiment as well to try to collect a more direct infrasound measurement from the bridge. The Hyperion (Tupelo, Mississippi) IFS-3000 series infrasound sensor located at the bridge utilized a dome wind filter rather than porous hoses and has a response within 3-dB variation from 0.01 to 100 Hz. Data collection took place November 13–18, 2015. In addition to infrasound sensor arrays, accelerometers were placed on the main spans of the bridge from November 16–18 to validate frequencies detected through infrasonic monitoring.

Twelve Gulf Coast Data Concepts (Waveland, Mississippi) multifunction extended life accelerometer data logger –×2 units (operated in high-gain mode), such as the one presented in Fig. 2, were used for continuous collection of bridge vibration data in this experiment. These devices combine a battery-powered digitizer and a Class C microelectromechanical system (MEMS)-based accelerometer. The accelerometer is a three-axis device with a range of ±2 g, whereas the data recorder can digitize up to 512 S/s on each channel with 16 bits of resolution. Each digitizer had 32 GB of storage capability, giving ample storage for the 46-h time frame during which the units were deployed. For a detailed description of this sensor’s performance and its comparison with similar devices, the reader is referred to work by Evans et al. (2014). Due to a significant time-stamping error detected with these units at their highest sample rate (512 S/s), the data in this experiment were collected at a sample rate of approximately 255 S/s. The data logger was configured to sequentially record files in 15-min blocks in its high-resolution time-stamped data-collection mode. The time-stamp information was accurate to 0.1 ms and was used during postprocessing to resample the data to a uniform rate. Individual data-logger clocks were synchronized, and cross-correlation-based delay estimation was used to perform final time alignment of the data.

These sensors were deployed on Br 18–0009 in three locations at Span 21 and in three similar locations at Span 22. The use of switchable magnets on the sensor-mounting plates, as seen in Fig. 2, allowed for rapid and slip-free attachment of the sensors to the steel bridge beams. This sensor-mounting approach avoided issues relating to epoxies or mechanical clamping systems and worked well in terms of rapid deployment and recovery from the snooper truck. The approximate locations of the sensors are presented in Fig. 2. Each sensor collected 46 h of data on each axis. Meteorological data were also collected during the course of the experiment. Launches were constrained by the training schedule for Beale AFB with launches occurring at dawn and dusk on November 14 and 15 from the Sutter County Airport for a total of four launches. Analysis of the data allowed for identification of times most favorable for detection of signals from the bridge of interest.

Analysis of infrasound data requires an understanding of the prevailing meteorological conditions of the area at the time of the experiment, because these conditions will directly impact infrasound propagation. Considerations include temperature profile, wind speed and direction, and understanding of the layers of the atmosphere. The atmosphere is composed of four main layers: the troposphere (0–12 km), stratosphere (12–50 km), mesosphere (50–80 km), and thermosphere (80–320 km) (Evers and Haak 2009; Hedlin et al. 2012). A signal propagates from the source and spherically radiates upward through the atmosphere until a point is reached at which the effective sound speed, based on temperature, wind direction, and wind speed, is higher than at the origin of the signal. At this point, the signal turns and is refracted back down to the receiver. The phase velocity of the signal as it moves across the array is indicative of the atmospheric layer at which the signal turned. A higher phase velocity across the array indicates a higher turning point of the signal. Higher turning points are typically associated with long propagation paths.

The close proximity of the arrays to Br 18–0009 is considered local propagation, roughly defined as less than 50 km (McKenna et al. 2012); therefore, atmospheric effects are limited to the troposphere (<10 km). Analysis of the radiosonde data showed no significant winds, thereby reducing the sound speed [Eq. (1)] (McKenna et al. 2008; McKisic 1997) to adiabatic sound speed (without wind vector component)where T = absolute temperature (K); and $\mathbf{n}·\mathbf{v}$ = component of wind speed in the direction of propagation.

Given the local propagation distance expected with the source-receiver spacings, it can be assumed that the wavefront propagated as a horizontal plane, with a phase velocity observed at the array approximately equal to the observed sound speed near the ground. The adiabatic sound speeds were calculated for data collected by each balloon launch with a maximum value of approximately 344 m/s. This value was later used in the frequency-wave number (F-K) analysis.

InfraTool is an infrasound analysis tool within MatSeis and uses frequency-slowness processing to devolve data collected across an array into correlation, back azimuth, and slowness (Hart 2004). Two analysis schemes were devised to further analyze the output for determination of the presence of a signal, the inverse slope method and the Hough Transform, both of which were discussed by Hart (2004). InfraTool was originally designed to identify impulsive sources within uncorrelated noise but has proven to be a useful tool for preliminary processing to identify times of coherent continuous wave packetized signals and identify back azimuths. This allowed for rapid assessment of large quantities of data to determine times of interest for further processing.

For each of the arrays, InfraTool was used to process each hour of the 7-day data-collection period. InfraTool uses a Butterworth filter, with the number of poles and passbands established by the user. A Type I Chebyshev filter (investigated within Matlab) was considered for this analysis; however, too much ripple was introduced in the specified passband. The Butterworth filter is effectively a subset of the Chebyshev filter with zero allowed ripple. Butterworth filters were utilized here, because they give as flat a frequency response as possible in the passband of interest. In this experiment, a three-pole filter was specified to achieve fast roll-off while minimizing signal ringing. Each hour was processed twice, once with a passband from 0.5 to 10 Hz and again with a passband from 0.5 to 4 Hz. These bands were both well within the range defined as infrasound, with 0.5–10 Hz being a broader sweep and 0.5–4 Hz corresponding to a range comparable with frequencies seen for bridge modes in the literature. The window specified was a 10-s window with 50% overlap. Times of high correlation with a back azimuth in the direction of the bridge from each array for each 1-hour block were noted as possible detections and, thus, times of interest for further analysis. The March 1, 2014, 0900–1000 UTC time block from the FRC array contained high-correlation values (>0.8) with back azimuths consistently pointing at the source of interest and phase velocity approximately the measured sound speed.

Following the identification of the time of interest (March 1, 2014, 0900–1000 UTC) for FRC through InfraTool and analysis of meteorological data, infrasound data were manually processed with Geotool, a software package for visualization and characterization of seismic and acoustic data (Coyne and Henson 1995). This processing involved filtering the data, handpicking coherent signals, completing a Fourier analysis to identify modes and confirm that the modes are in the infrasound passband, and an F-K analysis to confirm the back azimuth from which the signal originated (McKenna et al. 2009c, 2012; McComas et al. 2016).

First, the time series was filtered with a causal three-pole 1- to 10-Hz bandpass Butterworth filter based on a literature review that indicated the first several modes of long-span bridges are often under 5 Hz (Cunha et al. 2013; Conte et al. 2008; Brownjohn et al. 2010; Ren et al. 2005; Caicedo et al. 2001; Pietrzko et al. 1996; Hsieh et al. 2006; Morassi and Tonon 2008; Briaud et al. 2011; Foti and Sabia 2011; Jianxin et al. 2013; Chen et al. 2014; Ko et al. 2010; Lee et al. 2012) as well as a previous experiment using infrasound for bridge detection (McKenna et al. 2009a, c). The data were then manually processed to identify coherent data signals, such as those seen in the works by Donn et al. (1974), McKenna et al. (2009a, c), and McComas et al. (2016), which were defined as a signal present on at least three of the five infrasound sensors [Fig. 3(a)]. Examples of the bridge signal packets are highlighted with gray boxes.

When a coherent packet was identified, then the characteristics of the signal were noted [e.g., duration, frequency content, back azimuth to source, and signal-to-noise ratio (SNR)]. Fourier analysis was completed for each element of the FRC array to identify the coherent frequency in the signal packet window [Fig. 3(b)]. If coherent frequencies were observed, then F-K analysis was completed to determine back azimuth to the source and SNR. Visualization of the aforementioned signal processing scheme is presented in Fig. 3, which highlights the time series and frequency analysis. Fig. 3 was created in Matlab, because Geotool does not provide an adequate export capability for generating figures. The coherent continuous wave signal packets were highlighted in the time series data through the use of a delay and sum beamformer to align the time series in the direction of 222°, aligning with the main bridge span [Fig. 3(a)] (Rost and Thomas 2002).

The Fourier analysis for Fig. 3 was completed using Welch’s method, which is a nonparametric method used to estimate the power in a signal at different frequencies. Implementation is typically achieved by dividing a time domain signal into smaller overlapping and windowed segments that are converted to the frequency domain using the Fourier transform and then averaged at each frequency (Welch 1967). This approach has the benefit of typically reducing noise in the spectral estimate with some loss in frequency resolution due to the shortened time windows. This method was utilized to determine the coherent frequency content of a short (09:02:43–09:02:59 UTC, March 1, 2014) set of continuous wave packets and 1-h window (0900–1000 UTC, March 1, 2014), with both processed using 16-s windows and 75% overlap. The 16-s window was selected because it was 10 times longer than the lowest frequency of interest. The short-duration Fourier analysis identified the prominent frequency content of those signal packets [Fig. 3(b)], with the hour-long analysis demonstrating the persistent nature of the signal [Fig. 3(c)].

F-K analysis is an array-processing method by which the complete slowness vector, composed of both the back azimuth and the horizontal slowness, can be determined, with slowness being defined as the inverse of the apparent velocity for the wave front of the signal as it moves across the array, and the back azimuth being defined as the angle of the wave front arriving at the array measured between north and the direction to the source in degrees (Rost and Thomas 2002). This analysis is only applicable for short time windows, because longer time frames may contain several different phases, making differentiation and localization of the back azimuth difficult to differentiate between phases (Rost and Thomas 2002). The packetized signals that were investigated here occurred in short bursts, so F-K analysis was applicable in this case. Within Geotool, users are allowed to set the maximum slowness to be considered. This parameter was set based on the observed adiabatic sound speed. It is important to note that adiabatic sound speed and phase velocity are two different parameters, but for suspected direct arrivals where the source-to-receiver distance is in the local propagation range, it is acceptable to assume that the two values are approximately equal.

Conversion of effective sound speed (in meters per second) to slowness (in seconds per degree) yielded a value of 310 s/degree. Inputting this parameter as the maximum slowness for the F-K analysis and hand tuning the results for an apparent velocity of 0.36 km/s, corresponding to the 361 m/s determined previously, yielded the results presented in Fig. 4.

The contours of the plot indicate relative power, with the back azimuth being associated with the highest relative power on the plot that matches the slowness for the expected return of the signal as it is directed back to the earth. The results indicated a back azimuth of 222° from FRC, which corresponded to Br 18–0009. Fig. 5 presents the range of back azimuths from the FRC array to Br 18–0009, which ranged from 209 to 223°. The back azimuth of 222° corresponded to one of the longest spans of the bridge, which was expected because the long spans were able to displace more air, generating the infrasonic waves we detected. The result of the F-K analysis and the coherent frequencies observed indicated the hypothesized source of Br 18–0009.

Reviewing the data comprehensively, it appeared that the periods of increased traffic, such as rush hour, transferred more energy into the higher harmonics, whereas the less busy periods were dominated more by the natural mode, near 2 Hz. This is thought to be a product of the traffic forcing function and the increased rate of the sources (i.e., the frequency of occurrence of traffic wheel loads). Analysis of a 46-h data-collection window, from approximately 1100 Pacific Standard Time (PST) on November 16 to 0930 PST on November 18, 2015, indicated that traffic was not as efficient at exciting the lowest modes of the structure.

A number of different approaches were pursued to visualize the bridge’s response as a function of traffic characteristics. When vehicles crossed the expansion joint, the bridge in the area of the accelerometer was excited. These were counted as vehicle detections. Acceleration levels related to these detections, which are generally proportional to the wheel load impacts, were binned based on several peak-amplitude levels. Fourier analyses using Welch’s method with a 6-s window and zero overlap were then completed for each detection, and their spectral responses were averaged to get an estimated response as a function of detected peak-acceleration levels. An example composite spectrum showing this change in energy as a function of traffic peak-acceleration levels is presented in Fig. 6. The purple trace is the averaged composite for traffic peak impacts exceeding 0.075 G, and the red trace is averaged impacts under 0.075 G. The blue trace is from averaged impacts over 0.04 G, and the yellow trace is due to averaged impacts below 0.04 G. From this visualization, it is clear that some traffic loads produced up to 50 dB more response than others.

Fig. 3(d) presents Welch’s power spectral density estimate for a representative accelerometer deployed on Bridge span 22 corresponding to 0900–1000 UTC on November 17, 2015. The parameters for Welch’s method aligned with the infrasound processing, 16-s windows with 75% overlap. To minimize errors due to variations in traffic type, location, and rates of excitation, long observations over the 46-h deployment were compared with shorter 1-h durations using Welch’s spectral estimation method.

For both spans, the strongest vibrations were observed at the sensors located closest to the bridge deck’s outer edge. This was expected because that location was only constrained on one side and could have more motion. For Span 21, the dominant low frequency was 2.27 Hz, whereas the shorter Span 22 had a dominant response at approximately 2.32 Hz. For comparison purposes, an infrasound sensor was also located on the ground at the bridge. Welch’s power spectral estimate was performed using the same parameters as described previously for the infrasound analysis; however, there were some potential issues with the data from this sensor. To limit spatial aliasing in processing, infrasound sensors should be deployed at a distance to allow for at least one wavelength to be completed before the signal reaches the sensor. Because of this, the lowest modes of the sensor at the bridge could not be properly resolved. Infrasound technology is actually more efficient at detection of the lowest modes when it is used at a distance from the structure of interest. In addition, the close proximity of the sensor at the bridge means there was a possibility that some of the frequencies recorded were due to seismic coupling.

In comparing the results obtained from the analysis of the accelerometers and the infrasound data at array FRC located 2.6 km from the bridge, presented in Fig. 3, there was good agreement in the frequencies, seen in the highlighted bands in Figs. 3(b–d). Comparison of the accelerometer data [Fig. 3(d)] and the data from the FRC array located 2.6 km from Br 18–0009 [Figs. 3(b and c)] indicates that the modes at approximately 0.9 and 1.6 Hz were faint in the accelerometer data. One hypothesis is that the higher modes excited by the traffic may have overwhelmed the lower modes on the accelerometers, as in Conte et al. (2008). Conte et al. (2008) performed a series of experiments on the Alfred Zampa Memorial Bridge in California prior to the bridge opening to traffic. The bridge was instrumented with a series of accelerometers, and a series of tests were completed including forced vibration tests, consisting mostly of traffic loads, and ambient vibration tests, consisting mostly of excitation due to wind. Test results indicated that the forced vibration tests were better able to excite the higher-frequency modes, making the lower modes more difficult to identify. The ambient vibration tests, utilizing mostly wind, were shown to be more informative for identification of the lowest modes (Conte et al. 2008).

This observation holds for the current data as well. The on-structure accelerometers were very good at picking up the local effects of rapid transient loadings (wheel loads) at their points of placement. Although these effects overwhelmed the lower-order modes at a given point, they did not contribute as much to the gross structural motion that moved large masses of air, creating infrasound signals. The infrasound data presented in Fig. 3 corresponded to approximately 0900 UTC. In local time, that corresponds to 0200 PST, which would be a time with less traffic. A potential benefit of infrasound for structural health monitoring is that it naturally filters out the higher-order transient signals while emphasizing lower-order primary vibration modes that are most closely tied to global structural changes.