Behavior Forests: Real-Time Discovery of Dynamic Behavior for Data Selection

TL;DR

This work tackles the data explosion in ADS by introducing Behavior Forests for real-time discovery of Dynamic Behavior without pre-training. The approach combines preprocessing (discretization, hysteresis filtering, dimension and numerosity reduction) with a bottom-up Behavior Tree construction to analytically describe patterns and guide data selection. It demonstrates that 822 distinct dynamic behaviors can be identified in vehicle data, achieving a data reduction of 96.01% while preserving relevant patterns, with additional validation on synthetic and ECG datasets. The method enables efficient data collection and real-time triggering for ADS development, and offers a path toward multivariate extensions and sensor fusion for fleet-scale deployment.

Abstract

Automated Driving Systems (ADS) development relies on utilizing real-world vehicle data. The volume of data generated by modern vehicles presents transmission, storage, and computational challenges. Focusing on Dynamic Behavior (DB) offers a promising approach to distinguish relevant from irrelevant information for ADS functionalities, thereby reducing data. Time series pattern recognition is beneficial for this task as it can analyze the temporal context of vehicle driving behavior. However, existing state-of-the-art methods often lack the adaptability to identify variable-length patterns or provide analytical descriptions of discovered patterns. This contribution proposes a Behavior Forest framework for real-time data selection by constructing a Behavior Graph during vehicle operation, facilitating analytical descriptions without pre-training. The method demonstrates its performance using a synthetically generated and electrocardiogram data set. An automotive time series data set is used to evaluate the data reduction capabilities, in which this method discarded 96.01% of the incoming data stream, while relevant DB remain included.

Figures (7)

• Figure 1: Concept for recording or discarding vehicle data during operation using a Behavior Forest approach.
• Figure 2: Preprocessing of data. The raw multivariate time signals (left, ) are discretized with predefined breakpoints (middle left, ). Using unification, the multivariate time series is transformed into a univariate time series (middle right, ) and finally compressed with a logarithmic numerosity reduction method (right, ).
• Figure 3: Exemplary construction of a Behaviour Tree from a time sequence with the indicated dvb.
• Figure 4: Visualization of the two-dimensional synthetic time series in (a, top), in which signal one is orange and signal two is blue . An extract of the Behavior Forest (b) represents four distinct patterns as a colored leafs. These colored leafs are highlighted according to the pattern in the original time series (a, top), and in the preprocessed symbolic sequence (a, bottom) which is the Behavior Forest algorithm input.
• Figure 5: Subset of normal /and abnormal /ECG time signals (left). Visualizing the features of found patterns (mean, variance, skew, kurtosis, minimum, maximum, median, 25th, and 75th percentiles) in a two-dimensional plane using the t-SNE algorithm (right).

Theorems & Definitions (5)

• Definition 1: Time Series $\mathcal{T}$ morchen2006time
• Definition 2: Subsequence $\mathcal{T}_{i,\mathrm{l}}$ Noering2022
• Definition 3: Pattern $\mathcal{P}$ based on Li2012
• Definition 4: Dynamic Behavior (DB)
• Definition 5: Discretization