Author: Andrea Gaal
Learning by ear – Self-learning monitoring of compressors
They are used to operate heat pumps and jackhammers, fill car tires with air, clean surfaces, or for saliva suction in dentistry – the areas of application for compressors are manifold. A compressor compresses gas using mechanical force, thus increasing the pressure. The resulting energy is released when the pressure is discharged and can be used to drive tools and machines. This makes compressors an indispensable part of industrial production.
Interval-based maintenance – that can be done better
Maintenance work on compressors is often carried out at fixed intervals and process parameters are checked on a random basis. This is expensive and inefficient. In the CompWatch project, funded by the German Federal Ministry for Education and Research, the involved project partners develop methods for event-oriented maintenance intervals. This means that only when an error is imminent, a service is initiated, and the system is adjusted or repaired. The requirement for this is that possible errors are detected at an early stage. In this way, failures can be predicted and, ideally, also avoided.
Debugging by ear
One possibility of error detection are sounds: If a system sounds atypical, you know something is not running correctly. So if you permanently attach sensors, you can record acoustic signals and use the signatures contained in the data to infer the state of the compressor. In order to evaluate this data automatically using artificial intelligence (AI) and machine learning (ML) methods, all errors that can occur must be known – this is the only way to recognize them later.
A look into the crystal ball
The CompWatch project partners want to overcome this approach and determine and predict atypical conditions (anomalies) without such prior knowledge. Only then a seamless integration into new environments and running systems is possible.
The development of this new approach is based on the results of extensive experiments. In these, various sensors were used, which were tested for suitability when applied to different problems. Mounted at different positions of the compressor, the sensors had to detect anomalies and normal conditions.
Specifically, four structure-borne sound sensors were placed on the surface of compressor components to record the noise emitted, a microphone in the compressor to record airborne sound, and an acceleration sensor on the engine housing to record vibrations. This made it possible to detect leaks, dirty oil filters (simulated by a cover) and dirty air filters.
Demonstration of the possibilities
This is where the so-called neural networks come into play, which detect the difference between normal condition and anomaly based on the sensor signal. For this purpose, a software demonstrator was set up in which the “plan” is described how the signals recorded by the sensors are evaluated. Measurement data from different sensors can be read into this demonstrator. From this, a classifier can be trained for each sensor. Based on these initially recorded signals of the normal state, the system learns how to evaluate the recorded data and forwards the information to the plant operator. If anomalies are indicated, the operator can service or repair the compressor.
Big plans
Currently, the project partners are ensuring that the procedure works for a wide range of errors. In parallel, the implementation in marketable hardware is taking place.
What has been successfully demonstrated in the example of the compressor can also be used for other plant components in the future. Self-learning systems enable event-oriented maintenance and thus contribute to cost-efficient and effective condition monitoring as well as plant safety.
About the project
Partners with different competences are working together in the CompWatch project. SONOTEC GmbH (sensors), Petko GmbH (expert knowledge of compressor operation and maintenance) as well as Fraunhofer IKTS (AI-based algorhythms for error detection) are involved.
The project is funded by the Federal Ministry of Education and Research (BMBF) (funding reference: 02K18K012) and supervised by Projektträger Karlsruhe (PTKA).
Further information
- IKTS group »Machine Learning and Data Analysis«
- IKTS group »Cognitive Material Diagnostics«
- Industrial solution »Acoustic diagnostics – Defect detection | Signal evaluation | Quality assurance«
Stay informed: You are welcome to subscribe to our newsletter, read our other blog articles and follow us on LinkedIn, Instagram and YouTube.