Automated Industrial Sound Analysis

The experienced machine operator or production engineer often sees and hears whether a production process is running properly or whether a machine is working flawlessly. He is therefore an important part of quality assurance. With sensors and corresponding signal analysis, automated test methods are used to try to emulate these human abilities. In contrast to visual AI-based methods, the potential of acoustic quality assurance in combination with machine learning methods is not yet exploited. In order to increase this potential, we face various challenges that arise in the industrial environment:

  • Adaptive or easily adaptable test methods for flexible production with changing production and process parameters.
  • High quality of automated production leads to a low number of "not ok” parts. Accordingly, machine learning algorithms that can work with less or unbalanced data are necessary.
  • Reliable and robust detection even under noisy environmental conditions.


  • Development of application-specific classification algorithms
  • Consulting regarding applied acoustic sensor technology
  • Workshops on basics of AI-based signal analysis




Train own AI models, improve quality control



Research project


Acoustic inspection of weld seams of safety-critical components as part of quality assurance


Research project


Digitized material and data value chains


Research project


Accelerated Product Development


Research project


Sensor Edge Cloud for Federated Learning

Using prototypical demonstrators, Fraunhofer IDMT tests and demonstrates the function and effectiveness of novel methods for automatic acoustic event detection.

Automatic detection of compressed air leaks

Reliably and automatically detect audible leaks thanks to airborne sound analysis and machine learning methods.

Compressed air is an indispensable resource for many German industrial and trade companies for the operation of machines and plants. At the same time, it also represents a high cost factor on the electricity bill. On average, companies waste 30 percent of the energy generated due to the unnoticed escape of expensive compressed air. To detect such leaks, they rely, among other things, on ultrasound testing methods and on the mere hearing of trained personnel.

Fraunhofer IDMT has now tested in an experiment whether this "hearing" can be automated and reproduced using microphones in combination with machine learning methods in order to develop a reliable system for leakage detection. First results show that this is generally possible.

Automatic detection of different materials based on pling sounds

The principle demonstrator "Air-Hockey table" offers visitors to trade fairs and conferences a sporty variety in addition to a business talk.

Using an Air-Hockey table modified for research purposes, the researchers are working on novel approaches for acoustic quality assurance in an industrial context. They use pucks, which are made of different materials and cause different, but very characteristic "pling" noises as soon as they hit the board of the game table. During the game, these acoustic signals occur so frequently and irregularly that they can be used for analysis by means of machine learning methods to make a reliable statement about the material the pucks are made of.

This non-contact method, in which Fraunhofer IDMT combines its many years of expertise in the fields of acoustic measurement technology, signal processing and machine learning, can be used, for example, to detect material defects or in the in-line monitoring of welding processes. If acoustically perceptible defects are already detected during the process, it can be aborted and restarted promptly. Fraunhofer IDMT's acoustic testing method is also non-destructive and thus serves to reduce expensive test scrap.

Automatic detection of material defects based on rolling sounds

The marble track demonstrator illustrates the performance of the acoustic monitoring method in industrial applications. The detection of defects in surfaces was chosen as an application example.

Three differently coated balls roll in random order through the ball track system. The moving noises of the balls are picked up by small microphones and analyzed by machine learning algorithms. In this way, the balls are monitored in real time and identified by their specific rolling sound. The analysis result - the order of arrival - is immediately shown graphically on a display. With this principle demonstrator, the institute shows new acoustic methods for quality assurance - contactless, non-destructive and safely integrated.

Automated acoustic crack detection

Reliably detect hairline cracks in turned parts with airborne sound analysis and AI

As part of the quality inspection, the turned parts shown usually undergo a 100 percent visual final inspection. They are checked for finest hairline cracks, which are hardly visible, but indicate a quality defect. In an experiment, Fraunhofer IDMT investigated to what extent hairline cracks can be detected automatically by airborne sound analysis with AI. As a testing step in the final inspection, the impact of a turned part on a plate was recorded and assigned to the classes "ok" and "not ok" accordingly. The scientists investigated whether the sound spectra of the impact allow a distinction between good and bad parts, which parameters influence these sound spectra (e.g. drop height or material of the plate) and which microphone positions are best suited for the recordings as well as for later analysis. First results showed that it is generally possible to distinguish turned parts into two classes based on the impact sound. 

The data sets have been presented in several scientific publications at international conferences and are supposed to serve the scientific community as potential benchmarks for comparative experiments.

Industrial sound datasets