#### Event Title

#### Location

Brunswick, ME

#### Start Date

1-1-1992 12:00 AM

#### Description

Neural networks are finding increasing use as an adaptive signal classifier in many engineering applications. Artificial neural networks have been used for classifying NDE signals such as ultrasonic and eddy current signals. These networks consist of densely interconnected units with variable interconnection weights. The networks can be categorized according to their architecture and learning algorithm. The class of neural networks most commonly used is the multilayer perceptron network. The basic structure of this network consists of one input layer, one output layer, and one or more hidden layers. In order to properly classify input signals, the neural network must first be trained. The network is trained by presenting known input patterns to the input nodes and the corresponding desired output patterns to the output nodes. The network computes the output corresponding to each input and the error between the desired output and the actual output is used to drive an iterative algorithm for updating the weights in a manner that minimizes the error.

#### Book Title

Review of Progress in Quantitative Nondestructive Evaluation

#### Volume

11A

#### Chapter

Chapter 3: Interpretive Signal Processing and Image Reconstruction

#### Section

Neural Networks

#### Pages

685-691

#### DOI

10.1007/978-1-4615-3344-3_88

#### Copyright Owner

Springer-Verlag US

#### Copyright Date

January 1992

#### Language

en

#### File Format

application/pdf

New Training Algorithm for Neural Networks

Brunswick, ME

Neural networks are finding increasing use as an adaptive signal classifier in many engineering applications. Artificial neural networks have been used for classifying NDE signals such as ultrasonic and eddy current signals. These networks consist of densely interconnected units with variable interconnection weights. The networks can be categorized according to their architecture and learning algorithm. The class of neural networks most commonly used is the multilayer perceptron network. The basic structure of this network consists of one input layer, one output layer, and one or more hidden layers. In order to properly classify input signals, the neural network must first be trained. The network is trained by presenting known input patterns to the input nodes and the corresponding desired output patterns to the output nodes. The network computes the output corresponding to each input and the error between the desired output and the actual output is used to drive an iterative algorithm for updating the weights in a manner that minimizes the error.