Adaptive acoustic recognition system based on generative adversarial networks reinforcement learning

The purpose of this study is to develop an adaptive acoustic diagnostic system capable of detecting rare critically important malfunctions of industrial equipment based on audio signal analysis. The combination of the modified architecture of the generative adversarial network WaveGAN, reinforcement learning algorithms Deep Q-Network multi-agent analysis with dynamic adaptation is used as a methodological basis. This integration allows not only to generate physically reliable synthetic signals to compensate for class imbalance, but also to adapt the behavior of the system depending on the characteristics of the acoustic environment the type of equipment. Experiments conducted using the MIMII dataset demonstrated high accuracy (up to 96 %) classification completeness on various types of equipment. The results obtained indicate the high stability of the system to external noise its ability to timely detect pulse transient defects. The scientific novelty of the work lies in the synthesis of a generative approach a multi-agent architecture with contextual adaptation to production conditions, which provides a comprehensive interpretable analysis of audio signals. The practical significance is due to the possibility of introducing the developed system into intelligent platforms for monitoring the technical condition of equipment in real-world production.

Authors: N. A. Verzun, M. O. Kolbanev, A. R. Salieva, R. M. Mukhtarbekova

Direction: Informatics, Computer Technologies And Control

Keywords: acoustic recognition, generative-adversarial networks, reinforcement learning, multi-agent systems, predictive maintenance, WaveGAN, industrial equipment, intelligent monitoring

---

View full article