TimeTector: A Twin-Branch Approach for Unsupervised Anomaly Detection in Livestock Sensor Noisy Data (TT-TBAD).

Area of Science:

• Machine Learning
• Data Science
• Sensor Data Analysis

Background:

• Unsupervised anomaly detection in multivariate time series sensor data is challenging due to complex spatial-temporal correlations and noisy data.
• Applications span livestock farming and agriculture (LF&A), Internet of Things (IoT), and human activity recognition (HAR).
• Existing methods struggle to establish standard patterns and accurately distinguish between normal, abnormal, and noisy data.

Purpose of the Study:

• To develop advanced machine learning methods for anomaly detection in multi-sensor time series data.
• To propose a novel approach that accurately identifies normal, abnormal, and noisy patterns, minimizing misinterpretation of mixed noisy data.
• To improve the robustness and accuracy of anomaly detection models in complex real-world scenarios.

Main Methods:

• Proposed a novel "TimeTector-Twin-Branch Shared LSTM Autoencoder" incorporating Multi-Head Attention mechanisms.
• Implemented a Twin-Branch method for simultaneous multi-task learning, including data reconstruction and prediction error.
• Evaluated the model on a custom dataset against several benchmark anomaly detection models.

Main Results:

• The proposed TimeTector model demonstrated lower reconstruction errors (MSE, MAE, RMSE) compared to baseline models.
• Achieved higher accuracy scores (precision, recall, F1) in anomaly detection than existing benchmark models.
• Successfully minimized the risk of misinterpreting models when dealing with mixed noisy data during training.

Conclusions:

• The TimeTector-Twin-Branch Shared LSTM Autoencoder significantly outperforms existing anomaly detection models.
• The novel approach effectively handles noisy multivariate time series data, improving detection accuracy.
• This research offers a robust solution for unsupervised anomaly detection in diverse sensor data applications.