Feasibility of ANFIS towards multiclass event classification in PFBR considering dimensionality reduction using PCA. (January 2017)

• Record Type: Journal Article
• Authors: Prusty, Manas Ranjan; Jayanthi, T.; Chakraborty, Jaideep; Velusamy, K.
• Is Part Of: Annals of nuclear energy. Volume 99(2017:Jan.)
• Page Start: 311
• Page End: 320
• Publication Date: 2017-01
• Subjects: Prototype fast breeder reactor -- Adaptive neuro fuzzy inference system -- Principal component analysis -- Dimensionality reduction -- Factor analysis; Nuclear energy -- Periodicals; Nuclear engineering -- Periodicals; 621.4805
• DOI: 10.1016/j.anucene.2016.09.015
• Languages: English
• ISSNs: 0306-4549

Abstract

Highlights: A dimensionality reduction using principal component analysis (PCA) is achievable in nuclear power plant. PCA reduces the complexity of any classifier used for event classification. Adaptive neuro fuzzy inference system (ANFIS) works in acceptable range for such classification. PCA along with ANFIS improves the overall performance of ANFIS for event classification. Multiple ANFIS gives better performance over single ANFIS when the number of classes is more. Abstract: Nuclear power plant (NPP) such as Prototype Fast Breeder Reactor (PFBR), is a paradigm of complex engineering which is safety critical in nature. It sends copious plant signals to the main control room. An amalgamation of plenty of such plant signals escorts an operator to make appropriate decision during catastrophic circumstances. This decision must be quick and unambiguous in order to overcome any adverse situation. The concept of dimensionality reduction aids the purpose as it reduces the number of inputs to any system or model or classifier which ultimately makes the decision process faster. The fundamental aspect which should be taken paramount care is that there should not be any loss of information due to dimensionality reduction, as in a NPP, safety is the foremost goal. One of the most prevalently used dimensionality reduction algorithms is principal component analysis (PCA). This is achieved by dumping the principal components which has less variability. In this paper, the feasibility of dimensionality reduction is studied for classification of some of the events in PFBR. Two cases are considered in this paper out of which in first one, the event is divided into two sub events such as primary and secondary part of the event based on the importance of classification. The event data is fed to two separate classifiers which classify both the parts of the event separately. Finally, the event is classified as the concatenation of the outputs of both the classifiers. Unlike the first case, where two classifiers identify the event, here, in second case, a single classifier does the event classification. The classifier used is the adaptive neuro fuzzy inference system (ANFIS). This classifier has the advantage of both neural networks and fuzzy system where the neural network concept is used to tune the fuzzy membership function. PCA is used for dimensionality reduction and scree test for factor analysis. The performance of the PCA-ANFIS classifier is measured by calculating the area under the receiver operating characteristics curve (AUC) which is one of the most popularly used performance measures for any classifier. A comparative study is done on the AUC of all the PCA-k-ANFIS classifiers for both the cases mentioned above. Here, k represents the number of principal components considered as input data to the ANFIS classifier. A study on the significance of PCA along with ANFIS is also carried out by comparing it with an ANFIS classifier without using PCA.