A hybrid-type diagnostic system to give a final diagnostic result by integrating the results of sub-systems is one of promising ways to develop a flexible diagnostic system. This paper describes a hybrid-type diagnostic agent system for a fast breeder reactor “Monju”. This paper also proposes an integration technique of diagnostic results by sub-systems using what are called confidence values and trust values. Two simulations under the condition that a diagnostic sub-system diagnoses plant condition by a random process are conducted to examine the applicability of the proposed integration technique. The results suggest the applicability of the proposed technique and some applicable conditions for the diagnostic performances of sub-systems.

The objectives of the present study are to analyze plant transients caused by small abnormalities and to find plant parameters by which operators can recognize these small abnormalities. In order to evaluate the plant transient during an abnormal situation in the water system using the plant system code NETFLOW++, the turbine and feedwater (FW) systems should be analyzed with good precision. The code is validated using the measured data at “Monju”. Several abnormalities in the water system are candidates of the present study, e.g., FW control valve degradation, FW pump degradation, heat transfer degradation due to fouling on heat transfer tubes of the evaporator, loss-of-feedwater-heating, etc. All major components in the tertiary system are included in the calculation model such as the steam generators, the high-pressure turbine, the deaerator, the FW pump, the FW heaters, the FW control valves, the steam control valve, extraction lines and drainpipes. In case of a malfunction of a FW control valve resulting in low flow rate, a large temperature increase at the outlet of the evaporator is observed. On the other hand, a temperature decrease at the outlet of the evaporator occurs if heat transfer tubes in the evaporator have fouling. As a result of the calculations, it was determined that temperature at the outlet of the evaporator is a good indicator to detect abnormal situations.

In order to detect anomalies in rotating machines such as pumps at an early stage, we developed a system using wavelet transform. The pump diagnostic experiment equipment was designed taking into consideration the structure of the pump used for the water-steam system of the fast breeder reactor “Monju”. For improving detection capability, it is desirable to use a mother wavelet (MW) whose shape is similar to the anomaly signal that is required to be detected. We call the constructed MW on the basis of the real signal the real mother wavelet (RMW). The parasitic discrete wavelet transform (P-DWT) that has a large flexibility in design of the MW and a high processing speed was applied for detecting process signals. The vibration and sound signals were measured using the pump diagnostic experiment equipment when three types of anomalies (injection of an object, change of a balance of the impeller, and damage to the axis of the impeller) occur. Complex RMWs were constructed on the basis of the measured signals, and subsequently, parasitic filters were constructed. Signal detection was performed by calculating the fast wavelet instantaneous correlation using the parasitic filter. We evaluated three types of anomalies, and found that P-DWT is useful for detecting these anomalies. Furthermore, we developed a diagnostic agent using P-DWT as one of the diagnostic agents of our hybrid diagnostic agent system, which is intended to work together with the “Monju” distributed diagnostic agent system.

Accidents of industrial plants cause large loss on human, economic, social credibility. In recent, studies of diagnostic methods using techniques of machine learning are expected to detect early and correctly abnormality occurred in a plant. However, the general diagnostic machines are generated generally to require all process signals (hereafter, signals) for plant diagnosis. Thus if trouble occurs such as process sensor is broken, the diagnostic machine cannot diagnose or may decrease diagnostic performance.
Therefore, we propose an important process signal selection method to improve impact mitigation without reducing the diagnostic performance by reducing the adverse effect of noises on multi-agent diagnostic system. The advantage of our method is the general-purpose property that allows to be applied to various supervised machine learning and to set the various parameters to decide termination of search.
The experiment evaluation revealed that diagnostic machines generated by our method using SVM improved the impact mitigation and did not reduce performance about the diagnostic accuracy, the velocity of diagnosis, predictions of plant state near accident occurrence, in comparison with the basic diagnostic machine which diagnoses by using all signals.
This paper reports our proposed method and the results evaluated which our method was applied to the simulated abnormal of the fast-breeder reactor Monju.