An optimal neural network for diagnosing multiple faults in chemical processes

Detection of incipient faults caused by gradual deterioration of the variables in chemical processes is considered. The authors focus on the issue of optimizing a set of learning data to obtain efficient diagnosis of multiple faults of different degrees in chemical processes. This problem may be considered as one of recognition of large patterns. By using the ability of neural networks to perform generalization, all patterns can be recognized by a small network, which has been trained by optimal learning data of single faults. A set of optimal learning data was found by which an optimal network can be easily trained to perform generalization. By using the optimal network trained by the optimal learning data, a diagnostic system for incipient faults can be constructed.<>