Fault diagnosis for a MSF using neural networks

This work outlines the development of a fault diagnostic system for a multi-stage flash (MSF) desalination plant using artificial neural networks (ANNs). This diagnostic system processes the plant data to determine whether the process state is normal or not. In the last case, the diagnostic system determines the cause of the abnormal process state. The diagnostic system has an ANN for each potential fault. Every ANN processes the plant data looking for symptoms of their respective faults. At a given time, the result reported by an ANN is an index between 0 and 1. This number represents the certainty about the corresponding fault is affecting the plant. The higher is the value, the higher is the certainty of the affirmation. The structure of each ANN is simpler than those reported in the bibliography; however, the performance is better. These results are obtained due to a careful selection of the diagnostic system output and the use of a special training method. That training method calculates an appropriate value for the output of each ANN instead of setting it at 0 or 1 only. The new value of the output does not depend on the fault that causes the inputs but it does only on the degree of matching between the observed evolution and the expected one for the fault corresponding to each ANN. Finally, a dynamic simulator was used to evaluate the performance of the diagnostic system.

A diagnostic system based on ANNs was introduced. The structure of each ANN is simpler than those presented in previous works; however, t ta td tf to X Xn — Fault — Fault maximum magnitude — Number of sensors supervised by the diagnostic system — Sampling time — Activating time — Developing time — Final time of simulation — Detecting time — Vector of quantitative values of process variables — Vector or array of normal values of process variables Greek α δX δX0 ∆t ∆X — Attenuating factor — Vector of normalized deviations of process variables — Vector of expected normalized deviations — Sampling interval — Vector of quantitative deviations of process variables ∆Xn — Vector of normal bands of process variables — Vector of certainties of dynamic states µD µF — Vector of certainties of faults — Vector of certainties of measured variµX ables