Profile heavy duty vehicle usage based on CAN bus data mining

In this thesis work a real case problem concerning heavy duty vehicles’ usage patterns identification is addressed. Even if in the literature there are several cars usage patterns identification, the same kind of analysis is less frequently carried out on industrial, construction or off-roads vehicles. However, thanks to the wide spread of IoT devices and the firmly established cars connected mobility, the heavy-duty in-vehicle connectivity is growing in importance. To this purpose, multivariate analysis of multiple CAN signals techniques based on clustering and patterns discovery from time series data is presented. At first, ultra-fine, asynchronous and heterogeneous signals have been analysed: the relevant parameters to be monitored have been identified, the most appropriate level of aggregation of data has been suggested and series characterized by different sampling rates have been properly combined. Then, a multivariate time series segmentation strategy based on an application of the VALMOD algorithm has been proposed. Finally, different clustering and patterns discovery methods are presented, inspecting signals properties both in time and frequency domain. The results of the proposed procedures have been finally evaluated applying them to a real use case: three different usage patterns have been identified, respectively corresponding to idle, moving or regular working and higher workload. The results have been validated both by domain experts and by means of the additional information provided by MEA0183 messages data. In conclusion, an autoencoder-based deep learning for multivariate time series clustering is presented to inspect the presence of hidden features that may have not been considered before.