Development and research of a reinforcement learning method for acoustic diagnostics of industrial equipment
Gespeichert in:
| Titel: | Development and research of a reinforcement learning method for acoustic diagnostics of industrial equipment |
|---|---|
| Autoren: | N. A. Verzun, M. O. Kolbanev, A. R. Salieva |
| Quelle: | Научно-технический вестник информационных технологий, механики и оптики, Vol 25, Iss 5, Pp 961-970 (2025) |
| Verlagsinformationen: | ITMO University, 2025. |
| Publikationsjahr: | 2025 |
| Bestand: | LCC:Information technology |
| Schlagwörter: | акустическая диагностика, промышленное оборудование, обучение с подкреплением, классификация состояний, rl-агент, спектральный анализ, Information technology, T58.5-58.64 |
| Beschreibung: | The actual problem of acoustic diagnostics of autonomously operating industrial equipment is investigated. An overview of existing approaches to acoustic diagnostics, including methods based on convolutional neural networks and learning algorithms with a teacher, is provided. Their limitations have been identified, such as the need to use large amounts of labeled data for training, poor adaptation to changing conditions, and the lack of a real-time decision-making mechanism. A new approach to acoustic diagnostics based on reinforcement learning methods is proposed, characterized by adaptability, high resistance to noise and the possibility of continuous learning in a dynamic environment. The proposed method for determining the state of equipment operability uses an approach based on the study of acoustic signals emitted by operating equipment. The method includes building a neural network, selecting audio recordings from open audio file libraries, and training the network using a reinforcement learning algorithm. The process of acoustic diagnostics of the state of serviceability/ malfunction of industrial equipment involves four stages: real-time recording of acoustic data of working equipment, extraction of signs of equipment condition, training with reinforcement of a neural network and making a decision on the serviceability / malfunction of the equipment. Based on tagged WAV audio files from open databases, an experiment was conducted to identify various states of the equipment: normal condition, initial stage of the defect, critical malfunction. The results showed classification accuracy from 89.7 % to 98.5 % and average response time from 0.5 to 0.7 seconds with low computing load (on average 36.5 % CPU and 509 MB RAM). Unlike the wellknown acoustic diagnostic systems based on teacher-learning algorithms for neural and convolutional neural networks on pre-marked datasets containing acoustic signals emitted by running equipment, the proposed approach implements the decomposition of the initial acoustic signals into spectral components. Each of these components is analyzed and provided with signs reflecting the state of serviceability or malfunction of the equipment. This approach allows you to: use reinforcement learning algorithms for strategic decision-making; reduce model training time by pre-selecting significant features; improve diagnostic accuracy; reduce computational load and hardware resource requirements. The developed algorithm can be used for continuous monitoring of equipment condition and predictive maintenance in autonomously functioning industrial systems. Its use will allow reliable and timely detection and classification of industrial equipment malfunctions. It is possible to refine the algorithm to meet the requirements for integration with the IoT infrastructure, increase resistance to external noise, and implement more advanced RL algorithms such as PPO. |
| Publikationsart: | article |
| Dateibeschreibung: | electronic resource |
| Sprache: | English Russian |
| ISSN: | 2226-1494 2500-0373 |
| Relation: | https://ntv.elpub.ru/jour/article/view/525; https://doaj.org/toc/2226-1494; https://doaj.org/toc/2500-0373 |
| DOI: | 10.17586/2226-1494-2025-25-5-961-970 |
| Zugangs-URL: | https://doaj.org/article/4a0c52714b2640a7a36f7f918f2d59fd |
| Dokumentencode: | edsdoj.4a0c52714b2640a7a36f7f918f2d59fd |
| Datenbank: | Directory of Open Access Journals |
Nájsť tento článok vo Web of Science | Abstract: | The actual problem of acoustic diagnostics of autonomously operating industrial equipment is investigated. An overview of existing approaches to acoustic diagnostics, including methods based on convolutional neural networks and learning algorithms with a teacher, is provided. Their limitations have been identified, such as the need to use large amounts of labeled data for training, poor adaptation to changing conditions, and the lack of a real-time decision-making mechanism. A new approach to acoustic diagnostics based on reinforcement learning methods is proposed, characterized by adaptability, high resistance to noise and the possibility of continuous learning in a dynamic environment. The proposed method for determining the state of equipment operability uses an approach based on the study of acoustic signals emitted by operating equipment. The method includes building a neural network, selecting audio recordings from open audio file libraries, and training the network using a reinforcement learning algorithm. The process of acoustic diagnostics of the state of serviceability/ malfunction of industrial equipment involves four stages: real-time recording of acoustic data of working equipment, extraction of signs of equipment condition, training with reinforcement of a neural network and making a decision on the serviceability / malfunction of the equipment. Based on tagged WAV audio files from open databases, an experiment was conducted to identify various states of the equipment: normal condition, initial stage of the defect, critical malfunction. The results showed classification accuracy from 89.7 % to 98.5 % and average response time from 0.5 to 0.7 seconds with low computing load (on average 36.5 % CPU and 509 MB RAM). Unlike the wellknown acoustic diagnostic systems based on teacher-learning algorithms for neural and convolutional neural networks on pre-marked datasets containing acoustic signals emitted by running equipment, the proposed approach implements the decomposition of the initial acoustic signals into spectral components. Each of these components is analyzed and provided with signs reflecting the state of serviceability or malfunction of the equipment. This approach allows you to: use reinforcement learning algorithms for strategic decision-making; reduce model training time by pre-selecting significant features; improve diagnostic accuracy; reduce computational load and hardware resource requirements. The developed algorithm can be used for continuous monitoring of equipment condition and predictive maintenance in autonomously functioning industrial systems. Its use will allow reliable and timely detection and classification of industrial equipment malfunctions. It is possible to refine the algorithm to meet the requirements for integration with the IoT infrastructure, increase resistance to external noise, and implement more advanced RL algorithms such as PPO. |
|---|---|
| ISSN: | 22261494 25000373 |
| DOI: | 10.17586/2226-1494-2025-25-5-961-970 |