Reliable State Estimation in a Truck-Semitrailer Combination using an Artificial Neural Network-Aided Extended Kalman Filter

TL;DR

This work tackles reliable state estimation for a truck-semitrailer system under varying payloads, a classic out-of-distribution challenge for learning-based methods. It introduces a Hybrid Extended Kalman Filter (H-EKF) that fuses an Artificial Neural Network (NARX) with a model-based EKF to estimate the articulation angle $\theta$, lateral tire forces $F_{y_{2j}}$, and the truck steering angle $\delta_1$ using only standard semitrailer sensors. Through extensive real-world experiments across in-distribution and out-of-distribution loading states, the H-EKF demonstrates improved accuracy over purely model-based EKF and ANN baselines, particularly for dynamic tire forces, while maintaining reliable generalization when ANN confidence is low. By incorporating a knn-based, multidimensional confidence measure to modulate ANN influence, the approach supports safer ADAS and autonomous trucking under payload variations with potential applicability beyond this specific system.

Abstract

Advanced driver assistance systems are critically dependent on reliable and accurate information regarding a vehicles' driving state. For estimation of unknown quantities, model-based and learning-based methods exist, but both suffer from individual limitations. On the one hand, model-based estimation performance is often limited by the models' accuracy. On the other hand, learning-based estimators usually do not perform well in "unknown" conditions (bad generalization), which is particularly critical for semitrailers as their payload changes significantly in operation. To the best of the authors' knowledge, this work is the first to analyze the capability of state-of-the-art estimators for semitrailers to generalize across "unknown" loading states. Moreover, a novel hybrid Extended Kalman Filter (H-EKF) that takes advantage of accurate Artificial Neural Network (ANN) estimates while preserving reliable generalization capability is presented. It estimates the articulation angle between truck and semitrailer, lateral tire forces and the truck steering angle utilizing sensor data of a standard semitrailer only. An experimental comparison based on a full-scale truck-semitrailer combination indicates the superiority of the H-EKF compared to a state-of-the-art extended Kalman filter and an ANN estimator.

Figures (3)

• Figure 1: Hybrid estimation scheme for reliable combination of model-based and learning-based estimators, using the confidence $\tau_k$ in the learning-based method. The confidence $\tau_k$ in the learning-based estimate $\hat{\boldsymbol{y}}_{\mathrm{ANN}_k}$ is determined by the similarity between the current operation point and the training data set. $z^{-1}$ denotes a delay by one time instant.
• Figure 2: ann training inputs $\{\boldsymbol{u}_{\mathrm{ANN}_j}^{\mathrm{tr}} \}_{j=1}^{N_{\mathrm{tr}}}$ (blue) and evaluation inputs (red, orange) from raw data set (left) and excerpt of standardized data set (right) around new evaluation input $\bar{\boldsymbol{u}}_{\mathrm{ANN}_k}$ (orange) with knn-search for $K = 25$. The evaluation data of loading states "full load", "partial load 1", and "no load" is in-distribution regarding the training data. The evaluation data of loading state "partial load 2" is out-of-distribution regarding the training data.
• Figure 3: Measurements and estimates of articulation angle $\theta$ (top) and lateral tire force $F_{y_{21}}$ (middle) with ann estimation errors and current confidences (bottom). Excerpts from evaluation data of loading state "full load" (left) and "partial load 2" (right). The ann performs well if the evaluation data is close to the training data (i. e., the confidence $\tau$ is high) and shows large errors if the confidence is close to $0$. The hekf incorporates ann estimates and improves estimation accuracy reliably by taking the confidence into account.