Wyniki wyszukiwania - BazTech

Ograniczanie wyników

2 Diagnostyka

2 2025

Znaleziono wyników: 2

Liczba wyników na stronie

Wyniki wyszukiwania

Wyszukiwano:
w słowach kluczowych: voiceprint recognition

Sortuj według:

Ogranicz wyniki do:

Reinforcement-enhanced deep transfer fusion network for multi-domain transformer fault diagnosis via acoustic signatures

Mu Mingliang, Niu Jianfei, Guo Guangqian, Tian Maoxiang, Li Jinlin, Li Jingyi

Diagnostyka

2025

Vol. 26, No. 3

art. no. 2025303

During operation, power transformers generate continuous vibrational signatures bearing mechanical faultinduced impulses, which constitute the fundamental evidence for equipment condition assessment. To investigate acoustic-fingerprint characteristics under varying operating conditions, this paper proposes a Reinforced Integrated Deep Transfer Learning Network (REDTLN) for multi-fault-domain diagnostics. The methodology first constructs multiple specialized Deep Transfer Learning Networks (DTLNs) using novel kernel Maximum Mean Discrepancy (kMMD) variants, enabling source-specific adaptation to enrich transferable feature representations. Subsequently, a unified unsupervised ensemble framework integrates multi-metric divergence measures, employing a reinforcement-guided combinatorial search algorithm to discover optimal DTLN integration rules. This intelligent fusion mechanism significantly enhances multisource transfer capability, improving diagnostic accuracy and robustness in dynamic noise environments and complex operational scenarios. Experimental results confirm the model's efficacy in precisely identifying abnormal states while maintaining sustained >95% accuracy for representative faults under diverse acousticinterference conditions.

High-voltage circuit breaker fault voiceprint recognition based on prototype similar domain adaptive spectral morphological neural network

Wang Dong, Zhou Lubo, Xie Liyun, Liu Xipu, Dong Shiqi, Liu Junhua

Diagnostyka

2025

Vol. 26, No. 3

art. no. 2025301

High-voltage circuit breakers will emit continuous vibration signals during operation. The signals contain a large number of pulses and fluctuations caused by faults. They are the main data source for evaluating the functioning condition for high-tension breaker switches. To aim for examine its vocal features in vibration acoustic signals of high-tension breaker switches in different operating states, a prototype similarity domain adaptive spectral morphological neural network (PSDA-SMNN) was proposed. First, the vibration signal is denoised using spectral morphology variational mode decomposition combined with fast singular value decomposition method. Secondly, the labeled data of a certain operational state serves utilized for this information origin area, while this untagged information from different operational states serves utilized for this training objective area, and the prototype network distance similarity is used to align the feature distribution between the domains; then, meta-training is used to the domain network undergoes internal supervised training, and the network in the target domain undergoes external unsupervised training using the virtual label backpropagation algorithm. Through internal and external loop training, the difference in feature distribution between domains is reduced, and unlabeled faults of high-tension breaker switches in varying operational states are recognized. Test outcomes indicate which the suggested framework is able to precisely identify the malfunction operational condition for high-tension breaker switches and detect common issues for high-tension breaker switches under various interference settings, having a classification precision of approximately 95%.