On-line fault detection and diagnosis of cyber-physical systems

In recent years we have seen a rise in the complexity of the architecture of physical systems, as these architectures become more complex also the correct functioning of such systems becomes more complicated. Furthermore, it is possible that an accurate model of the system is not provided therefore, even with complex algorithms to verify it’s proper functioning becomes complicated. The goal of the thesis is to provide a data-driven model where the model of the cyber-physical system (in this case a self driving car) is developed through a black-box model, an LSTM neural network. This network will learn the behavior of the system by providing the states at the next time step and will use these outputs to guarantee safety protocols. First, data from a simulator was obtained which was used to train a neural network to generate the control of the vehicle (steering angle and acceleration), so that a proper environment for a self driving vehicle was available. Second, once the control was developed it was necessary to verify the behavior of the vehicle. This was achieved by using a particle filter which verifies the probability distribution of the states at the next time steps and eventually these states were compared with the safety properties imposed on the vehicle. Once this was achieved a proper model of the system was developed, which in turn was used to verify it’s correctness.