Unsupervised Pattern Discovery in Automotive Time Series: Pattern-based Construction of Representative Driving Cycles

Description

In the last decade unsupervised pattern discovery in time series, i.e. the problem of finding recurrent similar subsequences in long multivariate time series without the need of querying subsequences, has earned more and more attention in research and industry. Pattern discovery was already successfully applied to various areas like seismology, medicine, robotics or music. Until now an application to automotive time series has not been investigated. This dissertation fills this desideratum by studying the special characteristics of vehicle sensor logs and proposing an appropriate approach for pattern discovery. To prove the benefit of pattern discovery methods in automotive applications, the algorithm is applied to construct representative driving cycles. Danksagung Zusammenfassung Abstract Contents Abbreviations Symbols List of Figures List of Tables 1 Introduction 1.1 Motivation 1.2 Structure of the Thesis 2 Related Work 2.1 Pattern Discovery in Time Series 2.1.1 Terminology 2.1.2 Matrix Profile 2.1.3 Discretization-based Approaches 2.1.4 Other Approaches 2.2 Representative Cycles 2.2.1 Possible Applications 2.2.2 Representativeness 2.2.3 Existing Construction Methods 2.3 The Idea of Pattern-based Cycles 3 Development of Pattern Discovery Algorithms for Automotive Time Series 3.1 Suitability of Approaches 3.1.1 Special Requirements in Automotive Time Series 3.1.2 Discussion of Suitability 3.2 Framework Introduction 3.3 Preprocessing—Reduction of Complexity 3.3.1 Problem of Symbolic Toggling 3.3.2 Problem of Multidimensionality 3.3.3 Problem of Time Warping 3.4 Creating a Dictionary 3.4.1 Adapted Sequitur Compression Algorithm 3.4.2 Pattern Enumeration Algorithm 3.5 Postprocessing—Selection of Relevant Patterns 3.5.1 Filtering of Irrelevant Patterns 3.5.2 Selection of Representative Patterns 4 Pattern-based Representative Cycles 4.1 Pattern-based Statistics 4.1.1 Characteristic Parameters 4.1.2 Use Case Differentiation 4.1.3 Pattern-based Distance Measure 4.1.4 Influence of Pattern Discovery 4.2 Identification of Full Trip RDC 4.2.1 Trips and Pattern Sets 4.2.2 Selection of RDC 4.3 Construction of RDC 4.3.1 Selection of Absolute Pattern Frequency 4.3.2 Solving a Directed Graph Problem 4.3.3 Design of Transitions between Patterns 5 Evaluation 5.1 Experimental Evaluation of Unsupervised Pattern Discovery 5.1.1 Synthetic Randomized Data Set 5.1.2 Real-Life Data Set 5.1.3 Dimensionality 5.1.4 Postprocessing 5.2 Representativeness of Driving Cycles 5.2.1 Basics 5.2.2 Identification of Full Trip RDC 5.2.3 Construction of RDC 5.2.4 Evenly Distributed Pattern Frequency RDCs 6 Conclusion A References