Reshaping Bioacoustics Event Detection: Leveraging Few-Shot Learning (FSL) with Transductive Inference and Data Augmentation

Bioengineering, Journal Year: 2024, Volume and Issue: 11(7), P. 685 - 685

Published: July 5, 2024

Bioacoustic event detection is a demanding endeavor involving recognizing and classifying the sounds animals make in their natural habitats. Traditional supervised learning requires large amount of labeled data, which are hard to come by bioacoustics. This paper presents few-shot (FSL) method incorporating transductive inference data augmentation address issues too few events small volumes recordings. Here, iteratively alters class prototypes feature extractors seize essential patterns, whereas applies SpecAugment on Mel spectrogram features augment training data. The proposed approach evaluated using Detecting Classifying Acoustic Scenes Events (DCASE) 2022 2021 datasets. Extensive experimental results demonstrate that all components achieve significant F-score improvements 27% 10%, for DCASE-2022 DCASE-2021 datasets, respectively, compared recent advanced approaches. Moreover, our helpful FSL tasks because it effectively adapts from various animal species, recordings, durations.

Language: Английский

FedLSTM: A Federated Learning Framework for Sensor Fault Detection in Wireless Sensor Networks

Electronics, Journal Year: 2024, Volume and Issue: 13(24), P. 4907 - 4907

Published: Dec. 12, 2024

The rapid growth of Internet Things (IoT) devices has significantly increased reliance on sensor-generated data, which are essential to a wide range systems and services. Wireless sensor networks (WSNs), crucial this ecosystem, often deployed in diverse challenging environments, making them susceptible faults such as software bugs, communication breakdowns, hardware malfunctions. These issues can compromise data accuracy, stability, reliability, ultimately jeopardizing system security. While advanced fault detection methods WSNs leverage machine learning approach achieve high they typically rely centralized learning, face scalability privacy challenges, especially when transferring large volumes data. In our experimental setup, we employ decentralized using federated with long short-term memory (FedLSTM) for WSNs, thereby preserving client privacy. This study utilizes temperature enhanced synthetic simulate various common faults: bias, drift, spike, erratic, stuck, data-loss. We evaluate the performance FedLSTM against based precision, sensitivity, F1-score. Additionally, analyze impacts varying participation rates number local training epochs. comparative analysis established models like one-dimensional convolutional neural network multilayer perceptron demonstrate promising results maintaining while reducing overheads server load.

Language: Английский