Self-Attention-Based Deep Learning for Missing Sensor Data Imputation in Real-Time Probe Card Monitoring

Area of Science:

• Industrial sensor networks
• Data integrity in machine learning
• Signal processing

Background:

• Real-time sensor data is crucial for industrial monitoring, anomaly detection, and predictive maintenance.
• Missing data from sensor malfunctions challenges data integrity and subsequent analysis.
• Effective imputation methods are needed to address data gaps in industrial sensor networks.

Purpose of the Study:

• To apply and evaluate a Self-Attention-based Imputation for Time Series model for reconstructing corrupted sensor signals.
• To compare the performance of the self-attention model against traditional imputation methods and a Bidirectional Recurrent Imputation for Time Series model.
• To assess the accuracy and computational efficiency of the self-attention model for industrial monitoring applications.

Main Methods:

• Utilized a Self-Attention-based Imputation for Time Series deep learning model.
• Applied the model to reconstruct signals from industrial sensors (accelerometers and microphones).
• Evaluated performance using time- and frequency-domain metrics, comparing against traditional methods and a recurrent neural network model.

Main Results:

• The self-attention model demonstrated competitive or superior accuracy, with an average 66% improvement in Mean Absolute Error over traditional methods.
• Significant accuracy gains were observed, particularly in scenarios with extensive data loss (25%-88% improvement).
• The attention-based architecture trained over twenty times faster per epoch than the recurrent-based model.

Conclusions:

• The Self-Attention-based Imputation for Time Series model is a robust and pragmatic solution for ensuring data integrity in industrial monitoring systems.
• The model achieves high fidelity in reconstructed signals, even with substantial data loss.
• The balance of high performance and computational efficiency makes the self-attention framework suitable for demanding monitoring applications.