No abstract

No abstract

This paper presents a neural network approach to fault diagnosis of dynamic engineering systems based on the classification of surfaces in system output vector space. A simple second-order system is used to illustrate graphically the nature of the diagnosis problem and to develop theory. The approach is then applied to the diagnosis of a laboratory-based hydraulic actuator circuit. Results are presented for networks trained on both simulation and experimental data. An important achievement is the diagnosis of experimental faults using a network trained only on simulation data.

The paper provides further details of the automated failure modes and effects analysis (FMEA) program outlined in Part 1. Some of the more dijj'jcult development problems are discussed, and solutions are presented. The ,functionality of the program was tested through application to two experimental rigs, namely a closed-loop hydrostatic transmission with a dynamometer and a regenerative pump test rig. Non-destructive faults, such as abnormally low relief valve settings and excessive loads, were manually inserted into these rigs, und the measured Cfects were compared with the predictions from the program lo validale the software. The diference in complexity and contigurution evident in the two examples considered serves to highlight the generality of the approach. The ease ofreconj5gurability oj the software reflects the key aim of producing a program capable of analysing a wide range of hydraulic circuits.

No abstract