Benito Fernandez

Nonlinear dynamic system identification using artificial neural networks (ANNs)

A recurrent multilayer perceptron (MLP) network topology is used in the identification of nonlinear dynamic systems from only the input/output measurements. This effort is part of a research program devoted to developing real-time diagnostics and predictive control techniques for large-scale complex nonlinear dynamic systems. The identification is performed in the discrete-time domain, with the learning algorithm being a modified form of the back-propagation (BP) rule. The recurrent dynamic network (RDN) developed is used for the identification of a simple power plant boiler with known nonlinear behavior. Results indicate that the RDN can reproduce the nonlinear response of the boiler while keeping the number of nodes roughly equal to the relative order of the system. A number of issues are identified regarding the behavior of the RDN which are unresolved and require further research. Use of the recurrent MLP structure with a variety of different learning algorithms may prove useful in utilizing artificial neural networks for recognition, classification, and prediction of dynamic patterns

Control algorithms for fault-tolerant robots

When a fault-tolerant robot fails, its fault responsive system detects and identifies the failure. During the recovery process, reconfiguration of the system isolates the fault, and a new system model and a suitable controller attempt to completely compensate for the faulty condition without interrupting the robots operation. In this paper, the authors address the recovery process for fault-tolerant serial robots when they experience actuator failure. For this purpose, the authors consider three controllers based on PID feedback, sliding control, and parameter adaptation methods. It is shown that the sliding control implemented with a boundary layer reduces the system errors efficiently when the errors are large, and the controller behaves like an ordinary PID feedback as the errors get smaller. Additionally, when failures cause uncertainty in system parameters, inclusion of parameter identification capability in the controller design is suggested. Although the work is valid for a general robot, simulation results are presented on a four-axis robot.<<ETX>>

Recurrent multilayer perceptron for nonlinear system identification

A hybrid feedforward/feedback neural network, namely a recurrent multilayer perceptron, is used to identify nonlinear dynamic systems in an input/output sense. The feedforward portion of the network architecture provides the well-known curve-fitting character, while the local information feedback, through recurrency and crosstalk, permits the capture of the temporal aspects of the unknown system. A dynamic learning algorithm is used to train the proposed network in a supervised manner. The derived dynamic learning algorithm exhibits a computationally desirable characteristic: both network sweeps involved in the algorithm are performed forward, enhancing its parallel implementation. The capability of the recurrent multilayer perceptron network to identify nonlinear systems, using dynamic backpropagation learning, is demonstrated through a simple example. The simulation results are encouraging, though test of the identification method on a real-world system is still under investigation.<<ETX>>

Control of robot manipulator through robust sliding linearization

A robust control method which can overcome the disturbances and parameter uncertainties of a robot manipulator is proposed. This sliding-mode-based technique provides performance robustness that guarantees the system attractiveness to a certain linear surface if the so-called matching conditions and bounds on perturbations and modeling errors are satisfied. The performance and robustness characteristics can be separated into two cascaded loops, an internal one that guarantees the surface attraction or linearization robustness and an external one that designs a controller based on the linearized system. The technique is applied to a two-bar robot with a spring at the end effector as an unknown interaction with the environment and an external load disturbance.<<ETX>>

Accelerated learning in multilayer networks

An accelerated learning algorithm, adaptive backpropagation (ABP), is proposed for the supervised training of multilayer networks. The algorithm is based on the principle of forced dynamics for the error functional. The algorithm does not require use of any information from previous updates, while it requires knowledge of exactly the same error terms used in standard backpropagation. The numerical simulation results indicate that there are certain advantages to using ABP. The method is consistently about an order of magnitude faster than the standard backpropagation method, and also faster than such accelerated algorithms as quickprop. There is no added tuning parameter, other than the learning rate, to which ABP appears to be less sensitive. However, the drawbacks of jumpy behavior in the vicinity of the local minima and the inability to eventually reach the global minimum exist.<<ETX>>

ASDM-a novel neural network model based on sparse distributed memory

A novel artificial neural network model based on P. Kanervas (MIT Press, 1988) sparse distributed memory (SDM) is presented. The model possesses many major advantages over the original SDM. It is adaptive in the sense that the memory cells are called into service or released from service by a global mechanism. Since memory cells are utilized depending on the load on the memory, the actual number of memory cells needed to implement the ASDN (adaptive SDM) is significantly smaller than what is required for the original SDM. The storing and retrieval procedures are much simpler, and analysis of the best match problem can be carried out in a deterministic setting. All the advantages of the original SDM are retained while the main drawback of the original SDM, namely, the huge number of physical memory locations, is removed. The concept of time-varying intensity of memory is introduced, and customized metrics for determining distance between two data objects are allowed

Artificial neural networks based system identification and control of nuclear power plant components

Research on a novel neural network (NN)-based architecture for enhancing diagnostics and control of nuclear power plant components is described. The suggested diagnostician self-adapts, self-explores, incorporates and extends a standard rule-based expert system. The proposed architecture represents an improvement over conventional systems, since it incorporates knowledge acquired through the pattern recognition capabilities of NNs or through experts. The project is focusing on a U-tube steam generator as the representative component which is quite complex and amenable to analysis. The research approach and significant results are summarized.<<ETX>>

Improved CNC machining using NARMA neural systems

The NARMA (normal autoregressive moving average) neural network is based on the ARMA model used in time series analysis. NARMA networks can perform nonlinear analysis. It is shown that NARMA networks are capable of learning linear and nonlinear system dynamics. The networks are used to improve computer numerical control (CNC) machining precision. NARMA neural networks are used to learn the mapping between the actual workpiece dimensions and the commanded dimensions, including the effect of the change of stiffness of the workpiece during cutting. After training, the network then predicts the necessary corrective CNC commands which can reduce machining inaccuracies significantly. Computer simulations are performed showing that the machining precision is greatly improved compared to uncorrective cutting.<<ETX>>

An associative memory knowledge base for diagnostic and high level control functions

The authors present a novel associative memory (AM) knowledge base for real-time diagnostics and high level control functions for complex large-scale dynamic systems. The architecture is an adaptive architecture based on the sparse distributed memory (ASDM). The authors also introduce the concept of time-varying intensity of memory, and allow generalized metrics for determining distance between two data objects. With the generalized metric concept, one can attach significance to the bit locations of the patterns; bits in a word are now weighted in significance by their respective locations. A neural expert system application of the ASDM is discussed.<<ETX>>