Chicago’s towering skyscrapers perched beside Lake Michigan make it hard to imagine that millions of years ago, there was no city, no land upon which to build. Instead, in Silurian time, more than 400 million years in the past, a shallow tropical sea spanned the modern-day state of Illinois, as well as other parts of the northeastern U.S.

In these warm waters, reefs built from now-extinct corals and other sea creatures reached hundreds of feet thick in places. Sea lilies, known as crinoids, waved their feathery fronds as bryozoans formed colonies that eventually left behind mesh-like fossils. Cephalopods jetted through the sea as trilobites crawled in the mud.

Eventually, the sea became too shallow to support the reefs and filled with muds and sands of tidal flats. The carbonates drowned in the sediment; the reefs died.

Today, these ancient limestones and dolomites, often filled with fossils, are used in the construction industry. Quarries in and around the Chicago metropolitan area produce aggregates, stone sand, and other building materials.

These quarries share something in common with semi trucks that rumble along the interstate and stadiums filled with cheering fans. They all produce measurable seismic signals. These signals of urban and industrial environments can pose challenges for earthquake detection, resulting in missed detections because the quake-created waves are buried in the human-made noise. Scientists refer to signal-to-noise ratio, and this is an instance where the ratio may be too low to catch a quake. The result: Events like quarry blasts or the scream of a train have been incorrectly identified as earthquakes or tremors.

In work published in Seismological Research Letters, a team of scientists from Northwestern University, led by doctoral student Ann Mariam Thomas, explores seismic data from a broadband seismometer (NW.HQIL) installed in one of Chicago’s industrial corridors. The station, installed shortly after a magnitude 3.2 earthquake that struck the region in 2013, and in place until 2019, was meant to detect aftershocks. Its location — near dense residential districts, quarrying operations, a flood-control reservoir, major highways, railroads, a sports stadium, and a commercial airport — generated thousands of events in mere days of data collection.

Manually exploring such a dataset — spanning years — would be time-consuming and challenging. The team presents a semiautomated approach for finding clusters of distinct seismic events. With their focus on signals that look like earthquakes, the results from this study can be used to improve future earthquake detection methods.

A first step

“Our guiding philosophy was to start with the simplest approach and to iteratively add complexity,” the authors write.

Initially, they began by applying an algorithm that identifies groups, or clusters, of anomalous events that have the same source. This algorithm — the k-means clustering algorithm — is an unsupervised learning algorithm that groups unlabeled data into k number of clusters, where the authors determined the optimal value of k.

However, the team found that first applying a misfit detector to the data “produced one of the most substantial improvements in our results.”

The workflow

The final workflow begins with a misfit detector that uses power spectral density, or PSD. PSD shows how a signal’s power varies with frequency. The PSD misfit detector searches for the difference, or misfit, between the PSD of a given time window versus the background noise. In other words, how different is the signal in a particular window of time compared to the noise? PSD misfit detectors can find curious events — even those with low signal-to-noise ratios.

Deciding what constitutes “background noise” for the PSD misfit detector can be tricky because noise changes with time. Consider our urban and industrial setting. Traffic tends to be greater at rush hour, producing more vibrations at those times of day.

To tackle this problem, the team defined noise according to time, with four six-hour time spans: night (12 a.m. to 6 a.m.), morning, (6 a.m. to 12 p.m.), afternoon (12 p.m. to 6 p.m.) and evening (6 p.m. to 12 a.m.). To capture the background, they randomly selected 50 days between July 2014 and December 2019, excluding days with poor data quality.

The second step — the k-means clustering algorithm — helped the team identify similar signals that likely had similar sources. But how many clusters, k, were there? A lower number of clusters might combine signals with different sources into a single group. A larger number may differentiate events from the same source.

The team used both quantitative and qualitative methods to determine the number of clusters, landing on a k-value of 12. This k-value yielded multiple clusters containing the same type of event. For example, several clusters exhibit similar waveforms, but the events occurred at different times, so they’re in different clusters.

The final k-means clustering model was trained on 21,293 anomalous events detected by the PSD misfit detector in three weeks of HQIL data spanning August 1 to 21 of 2017.

The authors applied the model to 654,419 ten-second windows detected by the PSD misfit detector in two years of HQIL data (June 2017 to June 2019). Once the data were clustered, the team sorted the clusters into different event types using features like signal duration.

Event types

At this stage, the authors chose to focus on event type rather than assigned cluster. They were especially concerned with signals that indicate operational change — for example, machinery turning on. These kinds of signals have commonly been misclassified as earthquakes.

Four event types emerged: long duration anthropogenic noise, transitional signals, high-amplitude short-duration blasts, and low-amplitude blast-like events.

The first event type, type A, features long duration anthropogenic noise that tended to occur during weekday daytime hours. The signal’s frequency is prominent at 11 Hertz — close to the frequency associated with eight-pole motors. Motor poles refer to the number of magnetic poles — north-south pairs of magnets — that modulate the motor’s speed. The more the poles, the slower the motor spins, the greater the torque. An eight-pole motor, then, is meant for heavy-duty applications, slow and steady. Eight-pole conveyor belts at the nearest quarry assisting with the extraction of Silurian reef rocks may explain this event type.

The second event type B, the transitional signal, is similar to type A, but of shorter duration. It likely suggests turning something on — perhaps the conveyer belts. These events “frequently and commonly get misclassified by EQTransformer,” says Thomas, referencing a neural network designed to detect earthquakes and pick seismic P- and S-wave arrival times developed at Stanford. “I wouldn’t describe [type B events] as earthquake-like, but they do have a statistical feature that [EQTransformer] associates with earthquake events.”

The third event type C, high-amplitude short-duration blasts, is dominated by high-frequency energy (greater than 30 Hertz). This signal tends to occur between 11 a.m. and 2 p.m. local time on weekdays. These may be surface blasts at the nearest quarry. The authors confirmed at least a few of these signals to be surface quarry blasts by corresponding with local quarry representatives and residents.

The fourth event type D, low-amplitude blast-like events, features signals with maximum amplitude about 1 to 3 orders of magnitude lower than type C. Duration is also longer. These events, Thomas says, “are more earthquake-like; they are lower-frequency events with an onset and decay similar to earthquakes.” These signals occur at a wide range of hours with a slight concentration from 12 p.m. to 2 p.m. local time. The authors propose that these signals source from quarry blasts greater than a kilometer from station HQIL — so, quarries other than the one nearby.

A labeled dataset

The team created a dataset of 1,262 human-produced events recorded by station HQIL that they labeled by event type. They carefully inspected each event to confirm that the event type matched the description, and to check for any hidden signals. They also manually picked the onset time for each event. The majority of detected events, they note, were related to the nearby quarry.

The study demonstrates how even a single seismic station can be a cost-effective tool for industrial facility monitoring. However, with a single station, reliable source locations are not possible. With additional sensors, the workflow could be adapted to identify and locate industrial operations. In the future, the team aims to install a small network of sensors in the same environment, apply the same workflow, and create a robust catalog of located events.

Such data may also help them solve the mystery of the 2013 magnitude 3.2 earthquake that struck the region. “Northern Illinois is not an area known for earthquakes,” says Thomas. “But, a few natural earthquakes of magnitude 2-3 are detected every decade.” For instance, notable faults like the Sandwich Fault Zone are linked to natural seismicity. However, the 2013 event could be human-induced. “We are currently investigating,” Thomas says, “but this study is ongoing.”