Informatics Publishing Limited

P. S. Dhabe ¹, A. S. Shingare ², M. L. Dhore ²

Affiliations
1 Vishwakarma Institute of Technology, Pune, Maharashtra, IN
2 Department of Computer Engineering, Vishwakarma Institute of Technology, Pune, Maharashtra, IN

Abstract

In this paper original solving set algorithm for detection of possible outliers is updated to have less iterations and thus there by less time. Original algorithm selects initial solving set randomly, but if we select this set carefully using standard deviation of each pattern with respect to each other. The proposed modification requires less time and iterations than the original one. Our experimentation says that this modification requires around half to two third of the patterns in the initial solving set having maximum standard deviation. We have compared original and updated algorithms using synthetic 2-dimensional data set, as described in section II, as well as a fault diagnosis data set from NASA. We observed that the time required to detect outliers for updated algorithm is less than the original one and it exhibit better outlier detection rate than the original one along with better cluster entropy. Better outlier detection rate, less time required and better cluster entropy are the key features of this modification that makes it suitable for outlier detection from large data sets.

Keywords

Data Mining, Distance-Based Outlier, Fault Diagnosis, Outlier Detection.

Full Text

     

P. S. Dhabe ¹, S. G. Lade ², Snehal Pingale ², Rachana Prakash ², M. L. Dhore ²

Affiliations
1 Vishwakarma Institute of Technology, Pune, Maharashtra, IN
2 Department of Computer Engineering, Vishwakarma Institute of Technology, Pune, Maharashtra, IN

Abstract

This paper describes, proposed modified K-NN (MKNN) classifier, which calculates group prototypes from several patterns belonging to the same class and uses these prototypes for the recognition of patterns. Number of prototypes created by MKNN classifier is dependant on the distance factor d. More prototypes are created for smaller value of d and vice versa. We have compared performance of original KNN and MKNN using a fault diagnosis databases. From the experimentation, one can conclude that performance of MKNN is better than original KNN, in terms of percentage recognition rate and recall time per pattern, classification and classification time. MKNN, thus has increased the scope of original KNN for its application to large data sets, which was not possible previously.

Keywords

KNN Classifier, Group Prototypes, Pattern Recognition, Document Classification, Fault Diagnosis.

Full Text

     

Suja S. Panicker ¹, P. S. Dhabe ², M. L. Dhore ³

Affiliations
1 Computer Engineering Department, MAEER's Maharashtra Institute of Technology, Pune, Maharashtra, IN
2 Vishwakarma Institute of Technology, Pune, Maharashtra, IN
3 Department of Computer Engineering, Vishwakarma Institute of Technology, Pune, Maharashtra, IN

Abstract

In this paper Fuzzy Min-Max Neural Network (FMN) classifier is used for Fault Diagnosis applications. It is a 3-layer architecture and uses a fuzzy membership function to reason about class label of a test pattern. We have collected two standard data sets-one from UCI repository and other from NASA, for experimentation purpose. Each data set is divided in two sets namely Training and Testing, using around half of the patterns. Above said Neural Network is trained using Training set and its performance is calculated using Test set. From the calculated performance it is found that the FMN performs well for both the data sets. By observing training, one can note that training time is more, but since training needs to be done only once it should not be treated as a serious handicap. Recall time per pattern is very small, thus the given neural network can be used for real time fault diagnostic purpose.

Keywords

Fault Diagnosis, Fuzzy Min Max Neural Network, NASA ADAPT Data, UCI Pump Data.

Full Text

     

P. S. Dhabe ¹, A. S. Natu ², A. S. Naval ², P. H. Parmar ², A. M. Padwal ², M. L. Dhore ²

Affiliations
1 Vishwakarma Institute of Technology, Pune, Maharashtra, IN
2 Department of Computer Engineering, Vishwakarma Institute of Technology, Pune, Maharashtra, IN

Abstract

In this paper fuzzy hyper line segment neural network (FHLSNN) is modified to represent the learned knowledge in a compact manner using a short hypothesis. This modification has increased percentage recognition rate and reduced recall time per pattern. We have also redefined the membership function so that it suit for the suggested modification. The performance of FHLSNN and modified fuzzy hyper line segment neural network (MFHLSNN) is tested and compared with the two splits of Heart disease database collected from UCI repository. The results shows that MFHLSNN algorithm is found superior to FHLSNN with respect to the recall time per pattern and space requirements for storing the hyper lines.

Keywords

Fuzzy Neural Networks, Heart Disease Detection, Medical Diagnosis, Pattern Recognition.

Full Text