基于CNN-BiLSTM的可穿戴传感器数据犬类行为识别
打开文本图片集
中图分类号：TP183 文献标志码：A 文章编号：1673-3851（2026）05-0306-09
Reference Format：YANG Yi，KE Jun. CNN-BiLSTM-based dogactivity detection using wearable sensor data[J]. Journal of Zhejiang Sci-Tech University，2O26，55（3）：306-314.
NN-BiLSTM-based dog activity detection using wearable sensor data
YANG Yi，KE Jun （School of Mechanical Engineering， Zhejiang Sci-Tech University，Hangzhou 310ol8，China）
Abstract： A hybrid detection method based on data augmentation and CNN-BiLSTM was proposed to meet the dificulties in feature extraction from wearable sensor data，class imbalance，and the challenge of capturing long-term temporal dependencies in canine behavior recognition. This method used robust normalization（RN） and principal component analysis （PCA） to eliminate outliers and reduce feature dimensions，and mitigated class imbalance through a resampling strategy. On this basis，convolutional neural network （CNN） was used to extract local spatial features and suppress high-frequency noise， bidirectional long short-term memory networks （BiLSTM） were introduced to construct a bidirectional temporal dependency model. The experimental results showed that compared to the unidirectional CNNLSTM model， the CNN-BiLSTM model proposed in this paper achieved an improvement of 2.3% in accuracy and 2.1% in Fl score，with the Fl score for the complex behavior "playing" being improved by 26.0% . Compared with other mainstream behavior recognition algorithms， CNN-BiLSTM maintained a high recognition accuracy even when handling up to nine types of behavioral categories. It provided a reliable solution for dog behavior monitoring and recognition based on wearable devices.
Key words：dog activity detection；convolutional neural network；wearable sensors； deep learning; bidirectional long short-term memory network; robust normalization
0引言
犬类行为识别旨在基于可穿戴传感器采集的相关数据实现犬类行为的精准识别，该技术对保障犬类健康和建立和谐人犬互动关系方面有重要意义[」犬类行为识别可帮助主人及时发现其健康隐患，进而进行干预治疗；可拓展应用于生态学以及野生动物保护领域，为其他动物的行为模式分析、迁徙路线跟踪和濒危物种的定位保护等提供技术支撑[2-3]
犬类行为识别的现有主流方法为人工观察与视频观测，但这类方法效率低、主观性强且准确率低。（剩余11919字）