F. Yuan

Enhanced Multi-Stream Kalman Filter Training for Recurrent Networks

We present a framework for the training of time-lagged recurrent networks that has been used for a wide variety of both abstract problems and practical applications. Our method is based on rigorous computation of dynamic derivatives, using various forms of backpropagation through time (BPTT), a second-order weight update scheme that uses the extended Kaiman filter, and data delivery mechanics designed for sequential weight updates with broad coverage of the available data. We extend our previous discussions of this framework by discussing various alternative forms of BPTT. In addition, we consider explicitly the issue of dealing with and optimizing network initial states. We discuss the initial state problem from the standpoint of making time-series predictions.

Fuzzy logic anti-lock brake system for a limited range coefficient of friction surface

Describes the preliminary research on and implementation of a fuzzy logic controller to control wheel slip for an antilock brake system. The dynamics of braking systems are highly nonlinear and time-variant. Simulation was used to derive an initial rule base, which was then tested on an experimental brake system. The rules were further refined by analysis of the data acquired from vehicle braking maneuvers on a surface with a high coefficient of friction. The robustness of the fuzzy logic slip regulator was further tested by varying operating conditions and external environmental variables.<<ETX>>

Neural network modeling and control of an anti-lock brake system

The authors have previously described neural-network-based methods for modeling automotive systems and training near-optimal controllers. These methods are based on the premise that the physical system can be sufficiently instrumented during network training so that accurate evaluation of the effect of control actions is possible. In certain systems, such a automotive anti-lock braking (ABS), it may be costly to obtain the detailed data that would be required to exploit the full capabilities of neural methods. The present paper reports an initial simulation-based study to determine the performance potential of controllers designed with these methods. Such studies will help determine whether the cost of carrying out neural training methods on actual systems is justified.<<ETX>>

Neural control systems trained by dynamic gradient methods for automotive applications

The use of dynamic gradient-based training of neural controllers for automotive systems is illustrated. The authors use a recurrent structure that embeds an identification network and a neural controller and that properly treats both short- and long-term effects of controller weight changes. This results in an approximately optimal control strategy. Feedforward and hybrid feedforward-feedback neural controllers trained by dynamic backpropagation and a dynamic decoupled extended Kalman filter (DDEKF) are investigated. A quarter-car active suspension model is considered in both linear and nonlinear forms, and representative results are presented. Methods using higher-order information, e.g., DDEKF are very effective in comparison to methods based exclusively upon gradient descent, e.g., dynamic backpropagation (DBP). The use of a recurrent structure for obtaining derivatives for controller training is illustrated.<<ETX>>

Controlling cyclic combustion variations in lean-fueled spark-ignition engines

This paper describes the reduction of cyclic combustion variations in spark-ignited engines, especially under idle conditions in which the air-fuel mixture is lean of stoichiometry. Under such conditions, the combination of residual cylinder gas and parametric variations (such as variations in fuel preparation) gives rise to significant combustion instabilities that may lead to customerperceived engine roughness and transient emissions spikes. Such combustion instabilities may preclude operation at air-fuel ratios that would otherwise be advantageous for fuel economy and emissions. This approach exploits the recognition that a component of the observed combustion instability results from a noisedriven, nonlinear deterministic mechanism that can be actively stabilized by small feedback control actions which result in little if any additional use of fuel. Application of this approach on a test vehicle using crankshaft acceleration as a measure of torque and fuel pulse width modification as a control shows as much as 30% reduction in rms variation near the lean limit.

Vision Guided Robots For Automated Assembly

A vision guided robot for assembly is defined to be a robot/vision system that acquires robotic destination poses (location and orientation) by visual means so that the robots end-effector can be positioned at the desired poses. In this paper, the robot/vision system consists of a stereo-pair of CCD array cameras mounted to the end-effector of a six-axis revolute robot arm. From a systems point of view, accuracy issues of the vision system, the robot, and the manufacturing requirements are considered for the development of automated calibration methodologies for local and global work volumes of the robot/vision system. Resulting accuracy of local calibration on the order of 1.5 mm is sufficient for many automotive assembly applications. Multiple component assembly and robotic fastening has been demonstrated with the developed vision guided robot.

Training a hybrid neural-fuzzy system

It is shown that hybrid neural-fuzzy systems can be described almost as concisely as conventional layered neural networks and can be subjected to the same methods for training. Combining elements of neural and fuzzy systems in this way offers clear benefits whenever the training a neural network can be improved by incorporation of prior knowledge or where a fuzzy system requires careful tuning. The examples suggest that the inclusion of fuzzy elements in a neural network framework may, for certain applications, increase representational power with fewer parameters than would be required by merely increasing the number of conventional nodes and layers.<<ETX>>

Neural network control of a four-wheel ABS model

This paper presents simulation studies of neural network controller training carried out on a four-wheel anti-lock brake system (ABS) model. The indirect controller training method utilized here requires prior training of an identification network, which is then used in a sensitivity circuit that embeds the identification network and carries out a form of real-time recurrent learning. The training of both identification and controller networks makes use of a decoupled extended Kalman filter (DEKF) update scheme. To the extent that the model represents a real system, the resulting controller may be applicable to an actual vehicle. More likely, in our estimation, is that such studies will be useful in determining appropriate identification and controller architectures for on-vehicle training and in determining how to carry out such training.

Strategies and issues in applications of neural networks

The authors survey some of the fundamental aspects of neural networks that have been found crucial to their application to practical problems in diagnostics, modeling, and control. The analysis is a loosely connected collection of remarks on difficulties that have been encountered and the approach to dealing with them. Promising approaches now being explored and suggestions for future work are outlined. The issues raised concern the following: neural nets and engineering; training by higher order methods; sparse data and generalization; local representation networks; prestructured networks; scaling nodes; context switching; recurrent networks; neural controller development; and fusion of neural nets and fuzzy logic.<<ETX>>

Vision guided robots for automated assembly

A vision guided robot system consisting of a stereo-pair of CCD array cameras mounted on the end-effector of a six-axis revolute robot arm is presented. The accuracy of the vision system, the robot, and the manufacturing requirements is considered from a systems point of view. The development of automated calibration methodologies for local and global work volumes of the robot/vision system is discussed. The accuracy of local calibration which is on the order of 1.5 mm is sufficient for many automotive assembly applications. The assembly strategy is described, and multiple-component assembly and robotic fastening with the vision guided robot is reported.<<ETX>>