Detecting anomalies in multivariate time series sensor data across various domains like livestock farming, agriculture, the Internet of Things (IoT), and human activity recognition (HAR) presents significant challenges. These tasks require advanced ma...
Detecting anomalies in multivariate time series sensor data across various domains like livestock farming, agriculture, the Internet of Things (IoT), and human activity recognition (HAR) presents significant challenges. These tasks require advanced machine learning techniques to handle data from multiple sensors and identify discrepancies effectively. Our research concentrates on crafting cutting-edge machine-learning solutions for this purpose, particularly in environments where sensor data often includes anomalies, complicating the identification of consistent patterns that account for both spatial and temporal relationships. We have developed an innovative method known as Dual detection and Prediction techniques "TimeTector-Twin-Branch Shared LSTM Autoencoder," which integrates multiple Multi-Head Attention mechanisms. This approach not only aids in accurately distinguishing between normal, abnormal, and noisy data but also reduces the likelihood of models drawing incorrect inferences from mixed noisy data during their training phase. Our Twin-Branch system enhances this capability by supporting simultaneous tasks like data reconstruction and prediction error analysis, which boosts efficiency in multi-task learning. We evaluated our model against several standard benchmarks in anomaly detection, utilizing our specific dataset. The outcomes indicate our model achieves lower error rates (MSE, MAE, RMSE) in reconstruction tasks and superior accuracy metrics (precision, recall, and F1 score) compared to baseline models. These results affirm that our approach significantly improves upon existing models in the field.