In critical environments, it is necessary to keep machines in a reliable and working condition. To ensure this, the traditional approach was to perform reactive maintenance often combined with heavy amounts of redundancy or preventive maintenance, which in return risks the execution of unnecessary maintenance work including the replacement of completely healthy parts. In contrast, predictive maintenance systems monitor and/or model the current condition of machines and predict necessary action just before a part actually fails.
This has therefore a wide range of use cases in multiple different industries. To achieve exact predictions, a reliable monitoring of the current state is essential. The Signal Processing and Machine Learning group of the institute performs research towards solutions to estimate exact machine conditions and alleviate maintenance effort for modern industrial or automotive applications.
Automatic acoustic event detection extracts information about the occurrence of specific time-limited patterns, so-called events, from an acoustic signal. A specific event can often be associated with a specific sound source and its corresponding relevance. Consequently, the acoustic event detection can be used to determine the underlying causes for active sound sources. Our expertise on this field started with detecting dangerous situations around cars [1] and lately adopted state of the art methods using deep learning [2]. Our current best acoustic event detection system is able to achieve top-1 performance on well-known tasks and datasets [3].
In an automotive environment, a large variety of different sounds makes up the “normal” operating noises. As experienced drivers or service technicians can hear if there is a technical problem only from the sound of the vehicle, we deploy acoustic event detection to obtain an automated system for acoustic diagnostics. Traditional approaches using pattern matching algorithms, correlations, or frequency analysis were not able to solve this task in a satisfactory manner. Our new approach applies AI-based acoustic event detection to the vehicle diagnostics task [4]. It provides reliable and precise information of acoustically indicated problems in a vehicle. On this basis, more exact measurements of the vehicle condition will incrementally contribute towards predictive maintenance in the vehicular life cycle.
[1] P. Transfeld, S. Receveur, and T. Fingscheidt, “An Acoustic Event Detection Framework and Evaluation Metric for Surveillance in Cars,” in Proc. of Interspeech, Dresden, Germany, Sept. 2015, pp. 2927–2931
[2] J. Baumann, T. Lohrenz, A. Roy, and T. Fingscheidt, “Beyond the DCASE 2017 Challenge on Rare Sound Event Detection: A Proposal for a More Realistic Training and Test Framework,” in Proc. of ICASSP, Barcelona, Spain, May 2020, pp. 611–615
[3] J. Baumann, P. Meyer, T. Lohrenz, A. Roy, M. Papendieck and T. Fingscheidt, “A New DCASE 2017 Rare Sound Event Detection Benchmark Under Equal Training Data: CRNN With Multi-Width Kernels,” in Proc. of ICASSP, Toronto, Canada, Jun. 2021, pp. 865–869
[4] T. Fingscheidt, J. Baumann, M. Papendieck and A. Roy, “AI-Based, Automated Acoustic Diagnostics in Vehicles,” ATZ worldwide, vol. 07-08/2021, pp. 16-21, Jul. 2021
