Damage detection and reliability assessment using analytically based artificial intelligence.

摘要

In the pursuit of further improvement of reliability and safety of dynamic systems, we have developed more effective and accurate methods for damage detection and reliability assessment by using analytical and analytically based artificial intelligence techniques. Damage detection problems are formulated as inverse eigenvalue problems. An exact functional relationship between system eigenvalues and damage parameters are developed and combined with partial eigenvector method or system perturbation method to obtain an unique and exact inverse solution. The damage detection problem is simplified by decomposing the problem into two stages--Isolation and Identification. All small or large, single or multiple damages can be detected precisely.; Based on the observation of the analytical knowledge and the effectiveness of decomposition, we then design an analytically based artificial neural network in modularized architecture for damage detection. The proposed analytically based neural network, due to the simple design, can eliminate intensive training and provide greater performance.; A very effective way to use time-domain data for real-time system health monitoring is also developed. This includes a general method for constructing simplified equivalent dynamic model and an innovative hybrid neural network architecture, which consists of a recurrent network for system identification and a multilayer percetron network for damage parameters identification. Simulation examples show that the proposed method can isolate faulty elements rapidly.; The final part of this research deals with reliability assessment with fuzzy information. Fuzzy-set theory is extended and applied to the reliability problems. An unified approach is developed to treat different types of variables including random, fuzzy, random-fuzzy hybrid, and random with fuzzy information in reliability analysis. Neural networks are proposed to construct the fuzzy membership functions. This fuzzy-neural-based approach offers a way to incorporate engineer judgments into reliability analysis, and opens a way to computerize and to integrate with other AI techniques for reliability analysis.