Configurable on-board vehicle data logging with Principal Component Analysis
Examensarbete för masterexamen
Computer systems and networks (MPCSN), MSc
Modern trucks are complex electro-mechanical systems with dozens of networked electronic control units (ECUs). The operation of these systems is monitored through diagnostics and logging. Stakeholders are interested in diverse sets of the collected information. During product development, diagnostic information helps to improve the product, while traffic safety research can benefit from, for example, accident statistics. Recording and storing large amounts of data for stakeholders to analyze later is often not possible in desired scales due to on-board hardware constraints. Similarly, analysis of these data volumes can take a considerable effort. The solution proposed in this thesis uses statistics to implement an on-board logging unit (LU) that has low resource usage, is easily changed, and returns valuable and compressed information. The system consists of a configurable on-board unit that uses a multivariate statistical method called Principal Component Analysis (PCA). The PCA technique can create a compressed abstraction of the raw data, which highlights patterns and relations. The LU can send this data to an off-board data facility, where further analysis is possible. On-board compression allows the system to respond to a growing demand of diagnostic data without overtaxing the on-board resources. Stakeholders can dynamically upload new configurations to the vehicle to enable quick and frequent changes, which ensures that the currently most important information is gathered. This thesis work investigates the solution through development and test of a prototype unit. The developed software runs in a simulated environment with log files from real vehicles. The method gives versatile analysis abilities for linearly related signals. In tests is the method shown to be able to distinguish between different driving scenarios. Tests regarding monitoring tire pressure and brake performance investigate the sensitivity of the PCA method. In the case of tire pressure monitoring a radius change of 9.5 mm is detectable, which shows ten times higher detection sensitivity compared to the unprocessed signals. In the case of brake monitoring is it possible to detect a difference between braking on flat and sloping ground with 5% inclination. The developed PCA implementation managed to capture large amount of variance in the compressed datasets, a compression to 15-30% of the original data contained 70-90% of the variance. Calibration is needed for the method to perform well, which could possibly be time consuming. Through the performed tests is the applicability of PCA in the vehicular setting shown to be good The possibility to make better decisions based on available information is beneficial in all aspects related to vehicles. The possibilities for lower fatality in traffic accidents and better resource usage of both fuel and vehicle parts, makes the search for better vehicle data logging solutions important both with regard to resource management and safety requirements.
Data- och informationsvetenskap , Computer and Information Science