Pattern Classification by an Incremental Learning Fuzzy Neural Network
Abstract
To detect and identify defects in machine condition health monitoring, classical neural classifiers, such as Multilayer Perceptron (MLP) neural networks, are proposed to supervise the monitored system. A drawback of classical neural classifiers, off-line and iterative learning algorithms, is a long training time. In addition, they are often stuck at local minima, unable to achieve the optimum solution. Furthennore, in an operating mode, it is possible that new faults are developing while a monitored system is running. These new classes of defects need to be instantly detected and distinguished from those that have been trained to the classifier. Those classical neural classifiers need to be retrained by both old and new patterns in order to learn new patterns without forgetting the learned patterns. Conventional classifiers cannot detect and learn the new fault types on-line real-time. Using incremental learning algorithms in the monitoring system it is possible to detect those new defects of machine conditions with the system operating while maintaining oLd knowledge. Inspired by the promising properties of an incremental learning algorithm named Fuzzy ARTMAP Neural Network, a new algorithm suitable for pattern classification based on fuzzy neural networks called an Incremental Learning Fuzzy Neuron Network (ILFN) is developed. The ILFN uses Gaussian neurons to represent the distributions of the input space, while the fuzzy ARTMAP neural network uses hyperboxes. The ILFN employs a hybrid supervised and unsupervised learning scheme to generate its prototypes. The network is a self-organized classifier with the capability of adaptive learning of new information without forgetting old knowledge. The classifier can detect new classes of patterns and update its parameters while in an operating mode. Moreover, it is an on-line (real-time) and fast learning algorithm without knowing a priori information. In addition, it has the capability to make soft (fuzzy) and hard (crisp) decisions, and.it is able to classify both linear separable and nonlinear separable problems. To prove the concept, simulations have been performed with the vibration data known as the Westland Data Set. This data set was obtained from the Internet at http://wisdom.ar1.psu.edulWestland/ collected from U.S. Navy CH-46E helicopters maintained by Applied Research Laboratory (ARL) at Penn State University. Using a simple Fast Fourier Transform (FFT) technique for the feature extraction part, the network, capable of one-pass, on-line, and incremental learning performed quite well. Training by various torque levels, the network achieved 100% correct prediction for the same torque level of testing data. Furthermore, the classification performance of the network has been tested using other benchmark data, such as the Fisher's Iris data, the two-spiral problem, and a vowel data set. Comparison studies among other well-known classifiers were preformed. The ILFN was found competitive with or even superior to many classifiers.
Collections
- OSU Theses [15752]