Michel Hell

Participatory Learning in Power Transformers Thermal Modeling

In this paper, we introduce a new approach based on the participatory learning paradigm to train a class of hybrid neurofuzzy networks whose aim is to model the thermal behavior of power transformers. The participatory learning paradigm is a training procedure that tends to emulate the human learning mechanism. An acceptance mechanism determines which observation is used for learning based upon their compatibility with the current beliefs. The proposed model is compared with actual data obtained from an experimental power transformer equipped with fiber-optic probes. Comparisons with alternative approaches suggested in the literature are included to show the effectiveness of participatory learning to model the thermal behavior of power transformers.

Recurrent Neurofuzzy Network in Thermal Modeling of Power Transformers

This work suggests recurrent neurofuzzy networks as a means to model the thermal condition of power transformers. Experimental results with actual data reported in the literature show that neurofuzzy modeling requires less computational effort, and is more robust and efficient than multilayer feedforward networks, a radial basis function network, and classic deterministic modeling approaches

Uninetworks in time series forecasting

This paper presents an approach for time series forecasting using a new class of fuzzy neural networks called uninetworks. Uninetworks are constructed using a recent generalization of the classic and and or logic neurons. These generalized logic neurons, called unineurons, provide a mechanism to implement general nonlinear processing and introduce important characteristics of biological neurons such as neuronal AND synaptic plasticity. Unineurons achieve synaptic and neuronal plasticity modifying their internal parameters in response to external changes. Thus, unineurons may individually vary from an and neuron to an or neuron (and vice-versa), depending upon the necessity of the modeling task. Besides, the proposed neural fuzzy networks are able to extract knowledge from input/output data and to encode it explicitly in the form of if-then rules. Therefore, linguistic models are obtained in a form suitable for human understanding. Experimental results show that the models proposed here are more general and perform best in terms of accuracy and computational costs when compared against alternative approaches suggested in the literature.

Hybrid neurofuzzy computing with nullneurons

In this paper we address a new type of elementary neurofuzzy unit called nullneuron. A nullneuron is a generalization of and/or neurons based on the concept of nullnorm, a category of fuzzy sets operators that generalizes triangular norms and conorms. The nullneuron model is parametrized by an element u, called the absorbing element. If the absorbing element u = 0, then the nullneuron becomes a and neuron and if u = 1, then the nullneuron becomes a dual or neuron. Also, we introduce a new learning scheme for hybrid neurofuzzy networks based on nullneurons using reinforcement learning. This learning scheme adjusts the weights associated with the individual inputs of the nullneurons, and learns the role of the nullneuron in the network (and or or) by individually adjusting the parameter u of each nullneuron. Nullneuron-based neural networks and the associated learning scheme is more general than similar neurofuzzy networks because they allow different triangular norms in the same network structure. Experimental results show that nullneuron-based networks provide accurate results with low computational effort.

Real-Time Model-Based Fault Detection and Diagnosis for Alternators and Induction Motors

This paper describes a real-time model-based fault detection and diagnosis software. The electric machines diagnosis system (EMDS) covers field winding shorted-turns fault in alternators and stator windings shorted-turns fault in induction motors. The EMDS has a modular architecture. The modules include: acquisition and data treatment; well-known parameters estimation algorithms, such as recursive least squares (RLS) and extended Kalman filter (EKF); dynamic models for faults simulation; faults detection and identification tools, such as M.L.P. and S.O.M. neural networks and fuzzy C-means (FCM) technique. The modules working together detect possible faulty conditions of various machines working in parallel through routing. A fast, safe and efficient data manipulation requires a great DataBase managing system (DBMS) performance. In our experiment, the EMDS real-time operation demonstrated that the proposed system could efficiently and effectively detect abnormal conditions resulting in lower-cost maintenance for the company.

Nullneurons-Based Hybrid Neurofuzzy Network

In this paper we introduce design and learning schemes for hybrid neurofuzzy networks based on nullneurons. A nullneuron is a logic neuron that performs an operation psi parameterized by u (absorbing element). The nullneuron becomes a AND neuron if u = 0 and a dual OR neuron if u = 1. The operator psi is a composition of nullnorms. Based on input-output data, the learning procedure proposed here adjusts not only the weights associated with the individual inputs of the nullneurons, but also the type of the nullneuron in the network (AND or OR) learning the value of parameter u. Adjustment of u is done individually and after learning each nullneuron can be either a AND neuron or a OR neuron, independently of the state of the remaining nullneurons. Consequently, the neurofuzzy network presented in this paper is more general than alternative approaches discussed in the literature because it embeds a set of if-then rules that uses different connectives in their antecedents. Experimental results are included to show that the neurofuzzy network proposed provides accurate models after short period of learning time.

Participatory learning in the neurofuzzy short-term load forecasting

This paper presents a new approach for short-term load forecasting using the participatory learning paradigm. Participatory learning paradigm is a new training procedure that follows the human learning mechanism adopting an acceptance mechanism to determine which observation is used based upon its compatibility with the current beliefs. Here, participatory learning is used to train a class of hybrid neuro-fuzzy network to forecast 24-h daily energy consumption series of an electrical operation unit located at the Southeast region of Brazil. Experimental results show that the neurofuzzy approach with participatory learning requires less computational effort, is more robust, and more efficient than alternative neural methods. The approach is particularly efficient when training data reflects anomalous load conditions or contains spurious measurements. Comparisons with alternative approaches suggested in the literature are also included to show the effectiveness of participatory learning.

New Neurofuzzy Training Procedure Based on Participatory Learning Paradigm

In this paper we introduce a new approach to train neurofuzzy networks using the participatory learning concept. The participatory learning paradigm tends to emulate the human learning mechanism where an acceptance mechanism determines which observation is used for learning based upon their compatibility with the current beliefs. The performance of the proposed learning scheme is illustrated by an example involving a nonlinear system modeling problem: the thermal modeling of power transformers. Comparisons with other methods reported in the literature and between two dual network structures are also included. The experimental results show the effectiveness of participatory learning in neurofuzzy networks training.

Neurons and Neural Fuzzy Networks Based on Nullnorms

This paper suggests a new type of elementary unit for neural fuzzy networks based on the concept of nullnorm. A nullnorm is a category of fuzzy set-oriented operators that generalizes triangular norms and conorms. The new unit, called nullneuron, is a generalization of and or logic-based neurons parametrized by an element u, called the absorbing element. If the absorbing element u = 0, then the nullneuron becomes an and neuron and if u = 1, then the nullneuron becomes a dual or neuron. The paper also addresses two learning schemes for a class of hybrid neural fuzzy networks with nullneurons. The first scheme uses the gradient descent technique and the second reinforcement learning. Both learning schemes adjust not only the weights associated with the inputs of the nullneurons, but also the role of the nullneuron in the network (and or or) by individually adjusting the parameter u of each nullneuron. The neurofuzzy network presented here is more general than alternative approaches discussed in the literature because they allow different triangular norms in the same network structure. Experimental results show that nullneuron-based networks provide accurate results with low computational effort.

Recurrent neural approaches for power transformers thermal modeling

This paper introduces approaches for power transformer thermal modeling based on two conceptually different recurrent neural networks. The first is the Elman recurrent neural network model whereas the second is a recurrent neural fuzzy network constructed with fuzzy neurons based on triangular norms. These two models are used to model the thermal behavior of power transformers using data reported in literature. The paper details the neural modeling approaches and discusses their main capabilities and properties. Comparisons with the classic deterministic model and static neural modeling approaches are also reported. Computational experiments suggest that the recurrent neural fuzzy-based modeling approach outperforms the remaining models from both, computational processing speed and robustness point of view.