Runtime Monitoring of Cyber-Physical Systems Using Data-driven Models

In recent years we have seen a rise in the complexity of physical systems that humans build. As these systems become more complex also their correct functioning becomes more challenging. Furthermore, it is often difficult to obtain an accurate model of the system making formal verification of such systems even more complicated. The goal of the thesis is to investigate a frameworks whereby the model of the cyber-physical system (in this case a self-driving car) is developed through a black-box modeling approach, a long short term memory neural network. A recurrent neural network which stores information over arbitrary time intervals. This network will learn the behavior of the system from training data and can be subsequently used to guarantee safety of the system. This approach could be significant for many applications and in particular for autonomous systems which are currently a focus of intense development. In these systems, deep learning is often used to process data and make decisions on what the system should do. First, data from a simulator was used to train a neural network to generate the control of the vehicle (steering angle and acceleration).This module was used as an example of soft computing that is used in many autonomous cars. Once the control module was developed it was necessary to verify the safety of the vehicle. This was achieved by using the runtime monitoring framework. A particle filter which computes the probability distribution of the states at each time step (belief) is the integral part of the monitor. By computing the belief of the system combined with the desired safety property, we can make a decision on whether the operation of the vehicle is safe. Finally, once we obtained a monitor with good accuracies we show that it is possible to also use a data driven model of the vehicle to monitor safety.