Intelligent Sensor Based Bayesian Neural Network for Combined Parameters and States Estimation of a Brushed DC Motor

Abstract

Abstract-The objective of this paper is to develop an

Estimation of speed, armature temperature, and resistance in brushed DC machines using a CFNN based on BFGS BP

“…To the authors' knowledge, very few publications deal with the simultaneous estimation of speed and armature temperature of DC machines [25], especially when performed by intelligent techniques [29].…”

“…A very interesting approach was proposed in [25], applying and experimentally validating a transient EKF to estimate the speed and armature temperature in a BDC motor. However, the EKF has some limitations, in particular: (i) if the system is incorrectly modeled, the filter may quickly diverge; (ii) the EKF assumes Gaussian noise [26][27][28]; (iii) if the initial state estimate values are incorrect, the filter may also diverge; (iv) the EKF can be difficult to stabilize due to the sensitivity of the covariance matrices [27,29].…”

“…To the authors' knowledge, very few publications deal with the simultaneous estimation of speed and armature temperature of DC machines [25], especially when performed by intelligent techniques [29]. Artificial neural networks (ANNs) have demonstrated their ability in a wide variety of applications such as process control [30], identification [31], diagnostics [32], pattern recognition [33], robot vision [34], flight scheduling [35], finance and economics [36], and medical diagnosis [37].…”

“…In this paper, while referring to our previous study [29], in which an estimator based on a multilayer perceptron with Levenberg-Marquardt BP was developed in order to avoid the limitations of the standard ANN, a solution based on a cascade-forward neural network (CFNN) and Bayesian regulation BP (BRBP) is proposed. A highly accurate BRBP-based ANN was proposed in [38,39] but it requires an extremely long convergence time and is in fact known to be among the slowest algorithms to converge.…”

“…A highly accurate BRBP-based ANN was proposed in [38,39] but it requires an extremely long convergence time and is in fact known to be among the slowest algorithms to converge. Based on the approach already presented in [29], the purpose of this paper is to propose a novel approach using a learning algorithm that is a compromise between speed and accuracy. The BFGS can respond to these two constraints [39].…”