Summary

The project shall develop a systematic framework for sensor-fault detection, isolation, and accommodation by forcing a paradigm shift towards the development and the integration of signal processing and machine learning methodologies into novel hybrid-analytics solutions. Building upon ground-breaking concepts from graph signal processing, deep learning and transfer learning, SIGNIFY shall design and test tailored strategies from a Bayesian perspective to be used as tools for sensor validation when importing data from physical assets into digital systems. Designing optimization strategies exploiting real-time real-world data from sensors is one main value from the digital transformation. Unfortunately, sensors are prone to failures and injection of corrupted data into digital twins generates erroneous planning. When operating in closed loop, erroneous planning may lead to consequences ranging from performance degradation to lack of safety and risk of danger. The need for a validation tool before injecting sensor data into the digital twin is urgent in safety-critical applications.

Among relevant areas, Industry 4.0 focuses on development of safety-critical systems, where the high level of accuracy is needed when validating sensor data. In these systems it is hard to predict a malfunction by looking at the data without prior knowledge of the underlying phenomenon. Results from SIGNIFY will be general enough to apply to a large variety of scientific/application domains, however during the project 2 uses cases within the Industry 4.0 framework will be considered:

  • (UC1) Flow Assurance for CO2 Transport Operations;

  • (UC2) Low-Temperature CO2 Liquefaction and Phase Separation for Carbon Capture.

The facilities selected will allow the integration of physical models to benchmark sensor data and fit the Bayesian approach employed in SIGNIFY to combine signal processing and machine learning techniques. Performance improvement will be assessed in terms of validation accuracy.

Objectives

  • (O1) Establish a Bayesian framework for model-based sensor-fault detection, isolation, and accommodation (SFDIA) exploiting data spatial/temporal structures captured through linear/nonlinear estimation/detection techniques, while incorporating risk analysis at design stage;

  • (O2) Establish a Bayesian framework for data-driven SFDIA exploiting spatial/temporal structures captured through deep network architectures, while incorporating risk analysis at design stage;

  • (O3) Combine the solutions from the two frameworks above and develop hybrid-analytics solutions to SFDIA exploiting graph signal processing and transfer learning;

  • (O4) Validate the proposed solutions within the relevant use cases from Industry 4.0 exploiting available synthetic data and real-world measurements;

  • (O5) Integrate the framework developed with simplified process models as a tool for facilitating fault detection.

Work Packages

  • (WP1) Model-based approach to SFDIA: a signal processing perspective;

  • (WP2) Data-driven approach to SFDIA: a machine learning perspective;

  • (WP3) Performance assessment on synthetic data;

  • (WP4) Performance assessment on real-world measurements;

  • (WP5) Hybrid-analytics solutions.