6 August 2019 in IEEE Transactions on Neural Networks and Learning Systems
Energy optimization for plug-in hybrid electric vehicles (PHEVs) is a challenging problem due to the system complexity and many physical and operational constraints in PHEVs. In this paper, we present a Q-learning-based in-vehicle learning system that is free of physical models and can robustly converge to an optimal energy control solution. The proposed machine learning algorithms combine neuro-dynamic programming (NDP) with future trip information to effectively estimate the expected future energy cost (expected cost-to-go) for a given vehicle state and control actions. The convergences of these learning algorithms were demonstrated on both fixed and randomly selected drive cycles. Based on the characteristics of these learning algorithms, we propose a two-stage deployment solution for PHEV power management applications. Furthermore, we introduce a new initialization strategy, which combines the optimal learning with a properly selected penalty function. This initialization scheme can reduce the learning convergence time by 70%, which is a significant improvement for in-vehicle implementation efficiency. Finally, we develop a neural network (NN) for predicting battery state-of-charge (SoC), rendering the proposed power management controller completely free of physical models.
17 October 2019 in IEEE Transactions on Industrial Electronics
11 December 2020 in IEEE Transactions on Intelligent Vehicles
A pre-crash detection and warning system in a host vehicle needs to accurately determine the position of each remote vehicle in its vicinity and the context of the driving environment. ADAS (Advanced driver-assistance systems) have extensively used camera radar and LIDAR for automatic detection of vehicles, pedestrians and other road users and their behaviors. However, these vehicle-resident sensors have short operation ranges and require objects to be within the line-of-sight. V2V communication has emerged to be a promising technology to augment vehicle-resident sensors with extended ability of an overall vehicle safety system by addressing a broader range of crash scenarios with improved warning timing. In this paper we present an intelligent system, Geo+NN, developed using the synergy of neural network and geometric modeling. We extract the key geometric features using an analytic geometric model and use them as input to a neural network that is trained on real-world V2V signals to detect and predict remote vehicles’ positions. Geo+NN system is evaluated on V2V communication data recorded during real-world driving trips by vehicles installed with DSRC devices. Experimental results show that Geo+NN has the capabilities of effectively detecting and predicting remote vehicles within the context of 8 different positions.
25 September 2019 in IEEE Transactions on Vehicular Technology
This paper proposes a dual level reinforcement estimator (DLRE) to estimate the electric mining haul trucks' (EMHTs') total mass and road slope. To design the estimator, the longitudinal dynamics model of EMHT is built and the model's parameters are identified by a recursive least square (RLS) with double forgetting factors. This DLRE consists of an initial level estimation and an extended level estimation. In the initial estimation, the total mass and the road slope are preliminarily estimated. In the extended estimation, to prevent the estimated mass error from leading to great errors in the road slope, the road slope is first estimated and then the total mass is estimated based on the estimated road slope. The DLRE overcomes the defects of the traditional recursive-strategy-based estimators whose cumulative estimation errors often lead to poor estimation performance or even lead to estimation divergence. The DLRE performance is validated with an EMHT of 930E, and the experimental results show that the total mass and the road slope are estimated with higher accuracy using the DLRE.
19 December 2019 in IEEE Transactions on Intelligent Vehicles
Drivers often engage in secondary in-vehicle activities that can be functional and/or to relieve monotony. Drivers believe they can safely do so when their perceived workload is low. However, driving requires concurrent execution of cognitive, physical, perceptual and motor tasks. Over allocation of a driver's attention to secondary tasks may impair the driver's control of the vehicle and attention to the surrounding traffic. Accurate assessment of a driver's workload is an important, but challenging, research topic with many applications in intelligent vehicle systems related to driving safety and enhance driving experience. In this paper, we present our research on driver workload detection based on driver's physiological, vehicle signals as well as traffic contexts such as congestion level and traffic events. We obtained a collection of data from real driving scenarios. Twenty participants were recruited, whose workload was scaled to "low" as a base level, and "elevated" as a higher level. We developed two convolutional neural networks for multivariate temporal series (MTS-CNN). Extensive experiments were conducted in two scopes: within driver and cross-driver. The within-driver scope concentrates on experiments using data from a single driver, while in the cross-driver scope, transfer learning is leveraged and discussed. The experimental results demonstrate that one of the proposed models, i.e., MTS-CNN2, which combines features captured by the convolutional layers at all levels, is capable of learning well from the combined temporal physiological and vehicle signals and obtains the best performance.
1 June 2007 in 2007 IEEE Power Engineering Society General Meeting
This paper presents research in model based fault diagnostics for the power electronics inverter based induction motor drives. A normal model and various faulted models of the inverter-motor combination were developed, and voltages and current signals were generated from those models to train an artificial neural network for fault diagnosis. Instead of simple open-loop circuits, our research focuses on closed loop circuits. Our simulation experiments show that this model-based fault diagnostic approach is effective in detecting single switch open-circuit faults as well as post-short-circuit conditions occurring in power electronics inverter based electrical drives.