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

This dissertation investigates unsupervised pattern discovery in automotive time series, focusing on vehicle sensor logs. It proposes an approach for pattern discovery and demonstrates its effectiveness by constructing representative driving cycles.

Format: Paperback / softback
Length: 148 pages
Publication date: 24 March 2022
Publisher: Springer Fachmedien Wiesbaden

In recent years, the field of unsupervised pattern discovery in time series has garnered significant attention in both research and industry. This involves the identification of recurring similar subsequences within long multivariate time series without the need for explicit querying of subsequences. While pattern discovery techniques have already been successfully applied to diverse domains such as seismology, medicine, robotics, and music, their application to automotive time series has remained relatively unexplored. This dissertation aims to address this gap by examining the unique characteristics of vehicle sensor logs and proposing an effective approach for pattern discovery.

To demonstrate the potential benefits of pattern discovery methods in automotive applications, the proposed algorithm is applied to construct representative driving cycles. These driving cycles are designed to capture the diverse operational conditions and behaviors exhibited by vehicles. By analyzing these driving cycles, the algorithm identifies patterns that can be used to improve vehicle performance, diagnose faults, and optimize fuel consumption.

The dissertation begins by introducing the background and significance of unsupervised pattern discovery in time series. It then discusses the challenges associated with analyzing automotive time series, including the high dimensionality, noise, and temporal variability of sensor data. Next, the dissertation presents an overview of existing pattern discovery techniques and their limitations in automotive applications.

To address these challenges, the dissertation proposes a novel approach for pattern discovery in automotive time series. The approach is based on the concept of temporal convolutional networks (TCNs), which are highly effective in analyzing time series data with high dimensionality. The TCN is designed to capture the temporal dependencies between different sensor readings and identify recurring patterns within the data.

To evaluate the performance of the proposed approach, the algorithm is applied to construct representative driving cycles. The driving cycles are generated using real-world vehicle sensor data obtained from a fleet of test vehicles. The algorithm is then used to analyze the driving cycles and identify patterns that can be used to improve vehicle performance, diagnose faults, and optimize fuel consumption.

The results of the evaluation demonstrate the effectiveness of the proposed approach in identifying patterns that are relevant to automotive applications. The identified patterns can be used to improve vehicle performance by identifying inefficient driving behaviors, such as excessive braking or acceleration, and optimizing fuel consumption by identifying patterns that indicate optimal driving conditions. Additionally, the identified patterns can be used to diagnose faults by identifying abnormal sensor readings that may indicate a potential problem with the vehicle.

In conclusion, this dissertation presents a novel approach for pattern discovery in automotive time series. The approach is based on the concept of temporal convolutional networks and is designed to capture the temporal dependencies between different sensor readings and identify recurring patterns within the data. The evaluation demonstrates the effectiveness of the proposed approach in identifying patterns that are relevant to automotive applications and can be used to improve vehicle performance, diagnose faults, and optimize fuel consumption.

Weight: 233g
Dimension: 210 x 148 (mm)
ISBN-13: 9783658363352
Edition number: 1st ed. 2022