Periodic Reporting for period 1 - FAULT-LEARNING (Online Class Imbalance Learning for Fault Diagnosis of Critical Infrastructure Systems)
Periodo di rendicontazione: 2019-10-01 al 2021-09-30
Modern society relies heavily on the availability and smooth operation of large-scale complex engineering systems such as electrical power systems, water distribution systems, telecommunication networks and transportation systems. These are called critical infrastructure (CI) systems and a major part of our everyday life depends on their reliable and secure operation. A CI system failure can occur due to faults (e.g. equipment faults, human error, software bugs), natural disasters or malicious attacks and it can have severe societal, health and economic consequences. As the demand for CI services is growing rapidly, these systems become larger, more complex and have evolved into heterogeneous and distributed infrastructures. This has created interdependencies between them, with fault effects propagating to interdependent infrastructures. In fact, when multiple faults occur simultaneously, a high impact cascading failure could occur. For example, in November 2006, a local fault in Germany's power grid cascaded through large areas of Europe resulting in 10 million people left without electricity in Germany, France, Austria, Italy, Belgium and Spain. Considering such severe consequences of undetected cascading failures due to faults, it is crucial that a fault diagnosis engine must make online, real-time decisions and continuously improve and adapt to the sequential arrival of data. The overall objectives of the fellowship are to design and develop novel online learning algorithms with adaptation capabilities, with an application focus on real-time monitoring of CI systems, to accurately detect a potential fault and effectively isolate its exact location.
Despite the recent technological advances in the areas of sensing devices, real-time computation and intelligent systems, the task of monitoring CIs is becoming more difficult mainly because of three critical challenges. The first challenge is class imbalance, which occurs when at least one data class (e.g. “fault”) is under-represented compared to others (e.g. “normal”). It is a difficult problem as it causes a traditional learning algorithm to be ineffective, because of the low prediction power for the minority class examples. The second challenge is concept drift which occurs when the underlying relationship between the input and output data is dynamic, for example, this can occur due to software or hardware (e.g. sensor) faults. Learning in such nonstationary environments is very difficult. The third challenge is the lack of ground truth information, which is used to train a predictive model. In other words, the model typically needs to know when a prediction is correct or wrong; this is termed supervised learning. However, supervision is not available during real-time monitoring. The work performed by the research team have addressed all three challenges and have proposed the following:
- A novel supervised learning method called Adaptive REBAlancing (AREBA) that is robust to class imbalance and effective under concept drift conditions.
- A novel semi-supervised learning method called ActiSiamese that uses a specific type of semi-supervised learning called active learning, and a special type of neural networks called Siamese networks. ActiSiamese not only it addresses imbalance and drift, but it requires considerably less human input / supervision.
- A novel unsupervised learning method called AE-STREAM that uses a special type of a neural network called AutoEncoder (AE). The proposed AE-STREAM addresses imbalance and drift without any supervision.
- The proposed methods have been tested under realistic operating conditions using physical CI Testbeds for road anomalies detection and water leakage detection.
- A novel supervised learning method called Adaptive REBAlancing (AREBA) that is robust to class imbalance and effective under concept drift conditions.
- A novel semi-supervised learning method called ActiSiamese that uses a specific type of semi-supervised learning called active learning, and a special type of neural networks called Siamese networks. ActiSiamese not only it addresses imbalance and drift, but it requires considerably less human input / supervision.
- A novel unsupervised learning method called AE-STREAM that uses a special type of a neural network called AutoEncoder (AE). The proposed AE-STREAM addresses imbalance and drift without any supervision.
- The proposed methods have been tested under realistic operating conditions using physical CI Testbeds for road anomalies detection and water leakage detection.
The project has advanced the state-of-the-art in the following ways. We have provided new insights into learning from nonstationary and imbalanced data, a largely unexplored area that focuses on the combined challenges of class imbalance and concept drift in online learning. We have investigated this from the perspective of online, active, unsupervised learning. The proposed methods have been evaluated in an extensive experimental study, involving various synthetic and real-world datasets with different imbalance levels and drift characteristics, that spans over a wide range of domain areas and applications. The proposed algorithms (AREBA, ActiSiamese, AE-STREAM) have been shown to significantly outperform strong baselines and state-of-the-art methods. The proposed methods also performed well under realistic operating conditions for road anomalies detection and water leakage detection.