Stochastic Analysis of Controller Area Network Message Latencies with Observable Operator Models

We describe a novel machine learning approach for a stochastic analysis of controller area networks (CAN). Using CAN data of a vehicle we train Observable Operator Models (OOM) to learn patterns in the data. We present three approaches to construct such models from the data and investigate one in more detail. These models can then be used to extract probability distributions regarding message latencies and probabilities for deadline misses. The advantages of such an approach in contrast to theoretical analysis is that we make few to no assumptions about the CAN. Thus, the approach can generally be applied to period, data-driven and even mixed activation models. OOMs also model dependencies and are hence capable of predicting repeated and dependent occurrences of events such as repeated deadline misses. They are also intrinsically suitable as an online controller to observe events in real time and predict events based on the current state. The initial experiments that we carried out are explorative and limited in scope, but motivate further research. The major shortcoming of the approach in general is that it requires sufficient training data from which to learn the patterns of the CAN and an analytic construction of OOMs without training data is currently infeasible.