AI-based Performance Issue Detection for Cyber-Physical System Software Updates
Aitor Garciandia Alustiza
- DIRECTORS: Aitor Agirre and Goiuria Sagardui
- UNIVERSITY: Mondragon Unibertsitatea
Cyber-Physical Systems (CPSs) are systems that integrate computation and communication with the physical world. Software systems that are embedded in CPSs usually have a large life-cycle and are continuously evolving in order to incorporate new requirements, bug fixes, and to deal with hardware obsolescence. Despite the fact that the rise of new technologies provides an opportunity to improve CPS life-cycle management methods, there are still many complex issues that need to be addressed throughout its entire life cycle, from development to maintenance. In this sense, the increasing expansion of IoT and CPSs has highlighted the need for additional mechanisms related to the deployment and monitoring of these systems in operation, due to the challenge that represents the diversity of environments where these systems are deployed. The heterogeneity of the environments, along with the high configurability of CPSs, make testing the these systems under every possible condition impossible, increasing the possibility of errors appearing once the software is deployed in operation. This is especially concerning with performance errors, which are particularly difficult to detect in the lab.
In this context, this Ph.D. study proposes a mechanism to detect performance errors in Cyber-Physical Systems. The aim is to investigate different AI techniques to build a performance oracle, which predicts the expected performance of the system in operation and detects performance errors by comparing the predicted value with the actual monitored performance metrics. This prediction is obtained by training a performance model with data from the execution of previous versions of the software. The performance oracle is then encapsulated as a microservice, so that it can be seamlessly deployed along with other services to detect performance errors in operation. In order to make the performance oracle useful in multiple, heterogeneous environments, different training strategies have been investigated.
The method was evaluated by applying it to an industrial use case provided by Orona, using its dispatching traffic algorithm for elevator systems. Results show that the used AI techniques can be used to predict CPS performance and detect errors, even in installations where data for training was not available.