Access Official Technical PDF

Share Summary:LinkedInEmail

Multi-sensor data integration is a familiar challenge in industrial automation: measurements arrive from different devices, at different sampling rates and in different formats, yet must be brought together for analysis. The Global Earthquake Forecast System (GEFS), a research initiative funded by the SCOR Foundation from 2015 to 2018, tackled a similar problem across satellite and ground-based observations. Its work offers a useful case study in data infrastructure, anomaly detection and the discipline needed to evaluate forecasting systems.

GEFS aimed to investigate whether multiple physical signals could help identify conditions associated with large earthquakes. The project did not establish a universally reliable prediction method. Instead, it built tools and datasets for testing a complex hypothesis, while confronting practical issues that also matter to industrial data programs: heterogeneous sources, scale, signal quality and validation.

Multi-sensor data integration starts with diverse sources

The project collected measurements spanning several domains. Satellite data included thermal infrared imagery, GPS measurements used to derive total electron content, and signals from the DEMETER and SWARM missions. Ground observations included magnetic fields, geoelectric measurements, air-ion counts, ozone and carbon monoxide. Earthquake catalogs supplied a reference for comparing observed signals with seismic events.

These sources differed in format, coverage, resolution and collection method. Some data were image files; others were compressed binary measurements or time-series records. The team reported assembling a substantial archive, including GPS data dating back to 1995 and satellite observations covering multiple years. Its planned direction was to convert raw inputs into a common time-series format so analytical modules could work across sources without being tied to a particular instrument or file type.

That is a central design principle for multi-sensor data integration: preserve the source data, but create a consistent representation for downstream processing. A shared structure makes it easier to compare signals and reuse analytics, while retaining information about each measurement’s origin and characteristics.

Building infrastructure for large, heterogeneous datasets

GEFS combined local and cloud storage with high-performance computing resources. The report describes satellite and ground data stored in many millions of files, often in compressed formats. Analysis required decompressing some of those files, which increased storage and processing demands. Time-series measurements were also migrated into a database designed to handle temporal data, with support for spatial information.

The project’s experience shows why data architecture must account for the full pipeline, not just storage capacity. Collection, transfer, decompression, normalization and analysis all affect whether data can be used efficiently. The team encountered transfer-software problems and changes to computing infrastructure, illustrating how platform dependencies can delay analytical work even when the data itself is available.

For automation teams, the transferable lesson is to plan for operational realities: data may be distributed across systems, encoded in proprietary formats or dependent on specific infrastructure. A common data model and clear access arrangements can reduce friction as new sources are added. GEFS also developed a collaboration platform intended to support shared data access, analysis and discussion among researchers.

Time-series anomaly detection without a predefined signal

A major analytical challenge was that the form of a potential precursor was not known in advance. GEFS therefore explored non-parametric methods that could identify unusual behavior without assuming a specific signal shape. One approach divided measurements into frequency bands, calculated statistics over time segments and compared segments with a reference noise period. Segments that differed sufficiently from the noise were marked as candidate anomalies.

Testing on synthetic signals showed that the method could detect several kinds of injected anomalies, but it also missed some signals and produced a false detection associated with the filter’s initial response. The report describes practical responses, including shifting the segmentation to test alternative boundaries and excluding an initial interval affected by filter behavior.

These limitations matter in any sensor analytics application. A detection algorithm should be assessed not only on whether it finds known events, but also on missed detections, false alarms and sensitivity to preprocessing choices. If the signal’s expected characteristics are uncertain, testing across multiple scenarios is particularly important.

Validation separates detection from forecasting claims

To examine detection performance with known events, the team recorded magnetic-field changes near a Zurich tramway station. Tram arrivals and departures served as target events, while nighttime measurements provided a background-noise period. The data were divided into training and validation sets: parameters were tuned using the training data, then evaluated on separate data not used in that tuning.

The project assessed results with Molchan diagrams, which compare the share of time covered by alarms with the share of target events missed. This approach makes the trade-off between alarm duration and missed events visible. More broadly, the train-and-validate structure helps distinguish a method that fits one dataset from one that may retain skill on independent observations.

Results across earthquake-related datasets were not uniform. The report describes promising findings from multi-station geoelectric measurements in Taiwan, including improved scores when observations from multiple stations were combined. By contrast, analysis of DEMETER satellite very-low-frequency data did not show the expected pre-earthquake behavior. The researchers therefore identified further work as necessary rather than treating all signal types as established precursors.

Practical implications for sensor analytics

The GEFS experience underscores several principles for technical teams integrating distributed measurements:

  • Standardize access: map varied raw formats into a consistent analytical representation while retaining source context.
  • Account for processing effects: filtering, segmentation and data gaps can influence anomaly results.
  • Separate tuning from evaluation: use independent validation data to reduce the risk of overstating performance.
  • Compare across signals: combining sensors may add useful information, but each data stream still requires its own quality and performance assessment.
  • Communicate uncertainty: an anomaly is an indicator for investigation, not proof of a future event.

For industrial automation, these are relevant foundations for condition monitoring and other sensor-driven decision-support workflows, without implying that earthquake forecasting methods transfer directly to plant operations. The practical value lies in the engineering discipline: integrate data deliberately, test algorithms against independent evidence and make limitations clear. GEFS illustrates both the potential of multi-sensor analysis and the work required before complex measurements can support dependable decisions.

Access Official Technical PDF

About The Author