Application of BP Network with Changing Initial Weights to CSTR Fault Diagnosis

Neural networks have been widely used in kinds of research fields. In this paper, faults of CSTR will be detected and diagnosed using an improved BP algorithm. Due to remarkable influence of initial weights on networks' training speed, great attention is paid to selection of initial weights. A compositional method is used to modify initial weights in order to avoid the low convergence and system paralysis caused by the randomicity of initial weights. The fault diagnosis of CSTR model is simulated to indicate higher performance of the improved algorithm compared with BP networks.