Enhanced fuzzy-filtered neural networks for material fatigue prognosis

Although fuzzy-filtered neural networks (FFNN) have been used in pattern classification because of their unique characteristics in feature extraction, they usually have poor performance in forecasting applications due to their structure complexities especially in their consequent reasoning part. In this paper, an enhanced FFNN, EFFNN, is proposed for time series forecasting and material fatigue prognosis. A novel neural network scheme is developed to facilitate computation implementation. A new conjugate technique is proposed to improve training efficiency. The effectiveness of the developed EFFNN scheme and the related training technique is demonstrated by a series of simulation tests. The EFFNN is also implemented for material fatigue prognosis. Test results show that the developed EFFNN predictor is an effective forecasting tool; it can capture system dynamics effectively and track system characteristics accurately.

Based on the developed fuzzy-filtered neural network predictor, the input information is firstly processed by the use of appropriate fuzzy filters to unsure that each feature only triggers the corresponding rule without firing other unrelated rules. An intensity function is suggested to formulate rule firing strength to quantify the influence of the rule to the overall output. A novel training technique is proposed to optimize the shape and location of the fuzzy channels and link weights so as to further improve the operation convergence and forecasting accuracy.Figure optionsDownload as PowerPoint slideHighlights
► An enhanced fuzzy filtered neural network technique is proposed for time series forecasting.
► A new conjugate technique is proposed for system training.
► The developed predictor can capture system's dynamic characteristics quickly and accurately.
► The new predictor can be used effectively for material fatigue property prognosis.