HTP Graphics

LEAD RESEARCHER: LUIGI PANGIONE In recent years I have decided to radically change my career and move from working in controls of fusion plasma to robotics. Joining RACE allowed me to be part of the RAIN Hub. My role in RAIN relates to condition monitoring of a glovebox system and is, as my daughter pointed out once, to be "the doctor of the robots". That is the person who takes care of the health status of the robots before they get seriously ill! CONDITION MONITORING SYSTEM (CMS) SUMMARY // The Condition Monitoring System (CMS) work has two main purposes. First is to inform the operator of a possible upcoming fault. This will give the operator the ability to stop operations while the robot is still functioning and leave it in a safe state. Second, once the CMS recognises an abnormal situation, it should be able to assign it to an expected problem to help fault diagnosis. The problem of identifying an abnormal situation is more generally referred to as "anomaly detection". This applies in different areas such as fraud detection of bank transactions, intrusion of a network or sensor failure, just to name a few. Moreover, CMS can facilitate a shift in maintenance methods from being a planned process executed at frequent and regular periods of time, to a predicted process, executed when it is actually needed. This will allow the maintenance process to be optimised to extend equipment up-time and possibly reduce costs. "Unfortunately" faults are, by their nature, rare events. Moreover, manufacturers are very restrained in releasing data showing faults. This means that there is little chance of having "interesting" fault data to characterise CMS scenarios. UNIQUENESS // Historically CMS were based on alarms set on predefined thresholds derived by experience or on displacement of the behaviour from an ideal and known model. In recent years the use of Machine Learning (ML) and Neural Networks (NN) has gained ground with different methodology and topology of networks. We are investigating the latest network topologies and we are proposing different solutions with the intent of comparing them. In particular, we are investigating the use of Dirichlet Process Gaussian Mixture Model (DPGMM) and Variational Auto Encoder (VAE) to model the correct (baseline) behaviour of the robots. Our work is focussed on designing a metric to define an anomaly, which can be expressed in a simple probabilistic form. For example, we want to be able to say that the current sample has got 5% chance of being a fault. 60

• Cover