Anomaly Detection for Sensor Signals Utilizing Deep Learning Autoencoder-Based Neural Networks.

Area of Science:

• Sensor technology
• Artificial intelligence
• Signal processing

Background:

• Anomaly detection is critical in sensor applications to prevent high-risk decisions.
• Deep learning effectively handles imbalanced datasets common in anomaly detection.
• Diverse and unknown anomaly features necessitate robust detection methods.

Purpose of the Study:

• To develop autoencoder-based prediction models for automatic anomaly detection in electrochemical aptasensor data.
• To evaluate different autoencoder architectures (vanilla, ULSTM, BLSTM) and their integration for anomaly detection.
• To assess the impact of dataset characteristics, such as signal length and amount of normal data, on model performance.

Main Methods:

• Utilized a semi-supervised learning approach with normal data for training deep learning neural networks.
• Developed autoencoder-based prediction models incorporating vanilla, unidirectional LSTM (ULSTM), and bidirectional LSTM (BLSTM) architectures.
• Employed Kernel Density Estimation (KDE) for anomaly threshold determination.
• Integrated results from multiple autoencoder networks for enhanced decision-making.

Main Results:

• Vanilla and integrated autoencoder models demonstrated comparable performance in anomaly detection.
• LSTM-based autoencoder models exhibited lower accuracy compared to vanilla and integrated approaches.
• An integrated model combining ULSTM and vanilla autoencoders achieved approximately 80% accuracy on data with lengthier signals.
• Model accuracy was positively correlated with the availability of sufficient normal data for training.

Conclusions:

• Proposed vanilla and integrated autoencoder models effectively detect abnormal data in electrochemical aptasensors.
• Sufficient normal data is crucial for successful training and reliable anomaly detection performance.
• The developed models offer an automated solution for identifying anomalies in sensor signals.