Two Steps Ahead of Potential Failures with Airborne Sound

When Peter Meyer opens the doors to the production hall, he stands in his realm. For outsiders, the clicking, rattling and hammering of the machines is simply loud. What Peter hears are steady, typical operating sounds from the various rotating mechanical components that mesh smoothly to keep the machine running. If something was wrong, it wouldn’t get past his trained ears.

The machines are part of his everyday life; Peter is responsible for their maintenance. It’s an important job, because if one of them breaks down, it can lead to costly downtime, supply bottlenecks, and lost revenue.

Sound as an Early Warning System

Relying on sound to determine if a machine is about to fail is not new. Experienced listeners like Peter have developed a sharp sense for unusual machine operating noises over the years. But the more machines he oversees, the harder it becomes to notice an impending failure in time. With advances in media and sensor technology, automatic acoustic monitoring has become a widely used approach to predictive maintenance.

Measuring structural vibrations is the most commonly used sound technique for predicting failures. If a machine is unbalanced, misaligned, or if mounting bolts are loose, vibration increases — an indicator that something is wrong. But in order to determine the right time to take the equipment out of service, operators must continuously monitor the malfunction and need great experience with similar problems.

The use of airborne sound for predictive maintenance has long been considered an unusual monitoring option. Airborne sound is sound that is transmitted through the air, like music or speech. This differs from vibration, where the sensor is mounted directly on the machine under test. Its susceptibility to background noise has made airborne sound impractical for monitoring machines in the past. Anyone who has ever stood in a production hall knows why hearing protection is an essential piece of equipment.

Improve Existing Forecasting Models with Airborne Sound

With advances in sensor technology and signal processing algorithms, airborne sound-based acoustic condition monitoring has become robust to background noises. This has allowed the development of systems that can replicate and enhance human hearing diagnostic capabilities by analyzing sound characteristics like time, frequency, amplitude, or velocity comparable to that of experienced human hearing — like Peter’s.

“Keeping assets running by predicting, simulating, and optimizing their health is a key priority for industrial companies,”  said Anton Kroeger, senior director of Natural Resources for Australia and New Zealand at SAP. “Using airborne sound for predictive maintenance could improve existing forecasting models and deliver additional value to our customers.”

To better understand the potential of airborne sound for predictive maintenance, the SAP Innovation Center Network joined forces with the Fraunhofer Institute for Digital Media Technology (IDMT). In a proof of concept, the SAP Innovation Center Network and Fraunhofer aim to explore how to predict and avoid planned and unplanned downtime of industrial equipment through airborne sound measurements. With their expertise in sound engineering and acoustic condition monitoring, Fraunhofer IDMT will install a set of acoustic monitoring microphones at a customer site to capture sound radiated by the machine under test.

“Airborne sound carries valuable information for acoustic condition monitoring to detect anomalies and classify the behavior of machinery in automated production processes,” said Judith Liebetrau, group lead for Industrial Media Applications at Fraunhofer IDMT. “Since airborne sound-based condition monitoring is a non-invasive approach, sensors can be easily installed on the existing assets.”

Big Impact with Small Adjustments

The measured acoustic data of the machines and their surroundings will be recorded locally and transmitted securely to Fraunhofer for analysis. Instead of processing these immense amounts of data centrally in the cloud, the SAP Innovation Center Network team will train machine learning models locally across multiple decentralized edge devices without exchanging any raw data. The learnings of these individual edge models are then combined and transferred to the cloud to train a common, more robust master model. Think of it as leveraging swarm intelligence. The master model shares this aggregated knowledge with the local edge models and is continuously updated with new data. This distributed learning approach is designed to improve performance, increase accuracy and scale to larger input data sizes.

“With Fraunhofer’s expertise, we are able to explore this new technology in action and potentially apply the technological advances as part of a standardized solution at enterprise scale,” said Kavitha Krishnan, product manager and head of the SAP Innovation Center Network in Bangalore. “The data and insights from this joint project will help SAP to better understand and evaluate the business impact for customers.”

If airborne sound proves to be a reliable indicator for machine operators such as Peter, companies can add another dimension to existing forecasting models to further increase machine uptime and productivity. In maintenance, even small adjustments can have a big impact: SAP’s industry experts estimate that a 1% increase in asset availability and the resulting higher production can generate $38 billion in incremental revenue for customers in asset-intensive industries.

Instead of relying on intuition and experience in understanding if a machine is working well or not, the use of airborne sound can potentially extend existing predictive maintenance models and help to be two steps ahead of potential failures.


Yaad Oren is head of the SAP Innovation Center Network.