TN-GAN-Based Pet Behavior Prediction through Multiple-Dimension Time-Series Augmentation

Area of Science:

• Machine Learning
• Animal Behavior Analysis
• Wearable Technology

Background:

• Behavioral prediction models face challenges with performance deterioration and data bias.
• Accurate behavioral recognition and prediction are crucial for pet monitoring systems.

Purpose of the Study:

• To propose a novel approach for behavioral prediction using text-to-numeric generative adversarial network (TN-GAN)-based multidimensional time-series augmentation.
• To minimize data bias in behavioral prediction models.
• To enhance the accuracy of recognizing, predicting, and detecting abnormal behaviors in pets.

Main Methods:

• Utilized nine-axis sensor data (accelerometer, gyroscope, geomagnetic) from a wearable pet device (ODROID N2+).
• Applied data preprocessing techniques including outlier removal (IQR), z-score normalization, and cubic spline interpolation for missing values.
• Developed a hybrid model combining convolutional neural networks (CNNs) for feature extraction and long short-term memory (LSTM) for time-series analysis.

Main Results:

• The TN-GAN-based augmentation effectively minimized data bias in the behavioral prediction model.
• The hybrid CNN-LSTM model demonstrated proficiency in extracting features and reflecting time-series patterns for accurate behavioral prediction.
• Performance evaluation indices confirmed the model's effectiveness in actual vs. predicted value assessment.

Conclusions:

• The proposed TN-GAN-based multidimensional time-series augmentation is effective in mitigating data bias for behavioral prediction.
• The hybrid CNN-LSTM model offers a robust solution for pet behavior recognition and prediction.
• This research provides a foundation for advanced pet monitoring systems capable of detecting abnormal behaviors.