Overview

The seismic array in the Factor Building generates data at a ferocious rate. For the most part, however, this network and its incredible data collection resources remain largely untouched until some kind of significant seismic event takes place (a large earthquake in the L.A. area, for example). At that point researchers examine how the earthquake’s motion translates through the various floors of the building, providing insight into the structural characteristics of Factor.

Our goal with this proposal was to make use of smaller-scale, mini-seismic events that take place in the building everyday. In short, we were hoping to capitalize on routine, natural experiments to identify vibrations patterns that can be monitored over the long term and possibly used to detect degradation or other changes in the structure. Hence the title “Measures of convenience.” From the beginning, our work focused on the elevators in Factor. Can we “see” the movement of the elevators in the data reported by the seismic array? If so, can we use these signals to formulate and learn statistical models related to the structural characteristics of the building?

In the last six months, we have made connections with a community of researchers working in the area of structural health monitoring (SHM). In its most direct implementation, SHM involves excitation with a (known) repeatable source of vibration. In inhabited buildings like Factor (which houses auditoriums, classrooms, and laboratories; faculty and administrative offices; and the Schools of Nursing and Medicine) this kind of experimentation is not feasible, and SHM researchers look to ambient vibrations.

It turns out that the movement of elevators provides an excellent, but previously unexamined, source of ambient vibrations; in tall buildings like Factor elevators move at sufficient speeds to generate observable vibration patterns. Technically, the vibration is a result of (at least) two sources: First, the wheels that guide the elevator’s counterweight, a large mass that rides in a track along the back wall of the elevator shaft, can develop “flat spots” that produce vibrations at high speeds; and second, the track itself can have irregularities that introduce extra vibration at specific locations. From an SHM perspective, the twist here is that while this excitation source is repeatable (and in fact repeats many times over the course of a day on its own) its spectral characteristics are not known and will change slowly over time.

In September, Robert Nigbor (UCLA Civil Engineering) joined our group. Bob and two of his students provide much needed expertise in SHM and have helped design, instrument and implement a series of small experiments in Factor. The involvement of Steve Osterday from UCLA Facilities Management has been critical; he has provided us with access to the building, approvals to install sensing hardware (accelerometers) and advice/supervision during deployments.

Approach

Our goal is to assess the visible (through Factor’s seismic array) impact that the elevators have on the building and to create a more formal characterization that can be used for SHM. During the current funding period (which began in late September) our approach has been iterative and experimental: First, we introduced a known excitation near the main bank of elevators and examined the output from the seismic array; next, we conducted simple sweeps of the building, running the elevators from basement to top floor, during quiet times (late at night and early in the morning) and again examined the output from the seismic array; and finally, we installed an accelerometer on the counterweight itself and characterized the excitation associated with elevator movement. At each stage in this process, we evaluated what we had learned from the data, consulted with Facilities/civil engineers to decide how to either improve data collection and to see if our general approach is consistent with their practice.

Figure: The top plot shows the spectrogram of the counterweight sensor data. The middle and bottom plot illustrate the spectrogram of the seismic sensors’ low frequency and high frequency vibration on each floor, respectively. In each case the elevator is run from the basement to the top floor and back again. This experimental design induces the “on-off” pattern in the top plot and the angled response in the bottom plot (as the counterweight moves past each floor).

Systems/Experiments

First, we introduced a known vibration source. An APS Dynamics Model Electroseis 400 was used to generate frequencies from 1 to 10 Hz in steps on 1Hz and then two final runs at 20 Hz and 40 Hz. This design was repeated in two directions, north-south and east-west, and at floors 4 and 8. We then examined the response from the seismic array. Although the energy associated with the response to the lower frequencies was inconsistent and floor-dependent (with the 4th floor having the strongest response), the higher frequencies were clearly visible on both floors. With these encouraging results, we then performed a series of experiments involving simple movements of the elevator. These took place late at night when regular circulation in the building was the lightest. Here we noticed that the elevator needed to be run quickly to generate any kind of response, so that moving between two adjacent floors was not sufficient. The velocity of the elevator dictates its vibration profile.

Finally, in cooperation with Steve Osterday from UCLA Facilities Management, we installed a tri-axial accelerometer and a Q330 data logger with 200 sampling rate on one of Factor’s elevator counterweights. This created a feed of data that recorded the elevator’s state. We again conducted a series of experiments late at night and now have a clean data set of elevator vibrations and building responses.

 

Accomplishments

When we lined counterweight and the seismic data in time, we saw clear responses from data in both low and high frequency range, in contrast to times when the elevators were not moving. Moreover, time-frequency analysis (figure) indicates that the low frequency data induced by the counterweight is somewhat location-independent, indicating that the Factor building is responding globally to the elevator. This falls into the typical structural health monitoring category. On the other hand, a fast-moving elevator also induces high frequency vibration on each floor, indicating the elevator can be used as a local excitation source. As of now, we are concentrating on the high frequency region and trying to represent the dynamic response of a building in terms of wave propagation. All these results strongly indicate that elevator-induced vibrations are statistically separable from other ambient vibration components. An abstract outlines our studies has been accepted for the 4th World Conference on Structural Control and Monitoring that will be held in San Diego this July and we continue on the development of a methodology to use the elevators as a repeatable source for structural health monitoring.

 

Future Directions

First, we anticipate our expanded conference paper being delivered to the relevant structural health monitoring journals. In addition, we anticipate reports being delivered to the relevant sensor network conferences as well as to more statistical venues. From a statistical point of view, our goal is to formulate a formal statistical model of these data. We will initially consider segmenting the counterweight data into locally stationary parts since manual segmentation of the data is time consuming and sometimes inaccurate. Automatic detection of locally stationary parts of the data is desirable. From here, we can fold in seismic sensor data and eventually examine the coherence between the two sets of data through a multivariate model that is fit to all the available data from the building. The feasibility of these approaches is all predicated on the patterns observed in the actual data. For CENS, we believe that this broad paradigm of data collection is generalizable and can provide a template for other CENS deployments. The archived data can be used to fuel CENS visualization and modeling efforts. Finally, our study will also be folded into the statistics Ph.D. curriculum through connections with Statistics 202a, our introductory programming and Statistics 260, our data collection course. It will also be part of an elective statistics course planned for 2007 on modeling issues related to sensor networks.

People

Andrew Baek (UCLA Statistics)
Mark Hansen (UCLA Statistics)
Steve Kang (UCLA Civil Engineering)
Robert Nigbor (UCLA Civil Engineering)
Steve Osterday (UCLA Facilities Management)
Salih Tileylioglu (UCLA Civil Engineering)