透過聲紋事件偵測豬隻健康狀況

Abstract

本研究提出一個基於聲音的豬隻健康檢測系統，利用聲學事件來評估呼吸狀況，並實現畜牧業中的早期疾病檢測。我們透過放置在豬隻背部的聽診器捕捉呼吸聲音，並使用梅爾頻譜圖 (Mel spectrogram) 分析結合先進機器學習技術處理音頻數據。本研究解決了畜牧業聲學監測中的挑戰，包括嘈雜的音頻環境、呼吸資料的不足與類別不平衡，以及有限標註數據的問題。為克服這些限制，我們提出結合 k-means clustering 和 query attention 機制的 prototypical network 方法，並嘗試傳統機器學習模型，包括高斯混合模型 (GMM)、隱馬可夫模型 (HMM) 和梯度提升 (XGBoost, LightGBM) 等演算法作為對比模型。我們的實驗框架評估了多種編碼器架構，從傳統的特徵基礎方法到複雜的深度學習模型，包括卷積神經網路 (convolutional neural network, CNN)、Transformers 和四種預訓練聲學模型。本研究所提出的方法雖遜色於 baseline， 但其中 query attention 搭配 k-means 原型網路依舊展現優異性能，達到類似基準模型的效果，展現出模型成功區分不同的呼吸模式，實現非侵入性即時健康監測。

This study presents a sound-based health monitoring system for pigs using acoustic event detection to assess respiratory conditions and enable early disease detection in livestock farming. The system captures breathing sounds through stethoscopes placed on pigs' backs and processes the audio data using Mel spectrogram analysis combined with advanced machine learning techniques. We propose novel methodological approaches incorporating prototypical networks with k-means clustering and attention mechanisms, alongside traditional machine learning models including Gaussian mixture models, hidden Markov models, and gradient boosting algorithms. Our experimental framework evaluates multiple encoder architectures, ranging from traditional feature-based approaches to sophisticated deep learning models including Convolutional Neural Networks, Transformers, and pretrained audio models. The proposed prototypical network approaches demonstrate poorer performance than the deep learning baseline, but the k-means prototypical network with query attention shows a more stable result than the k-means prototypical network without the query attention mechanism when K is greater than 2, showing the effectiveness of how query attention is able to stabilize the k prototypes, while the linear encoder baseline approach achieves the best overall classification results. The system successfully distinguishes between different respiratory patterns, enabling non-invasive real-time health monitoring that reduces animal stress while improving disease detection efficiency.