Real-value Negative Selection Algorithm: Application And Optimization

Artificial immune system which is based on the operation mechanism ofbiological immune system have accumulated a large number of successful applicationcases, such as a computer virus defense software. The current artificial immunesystem is constantly deepened and improved. It has become the hotspot in the field ofartificial intelligence after neural networks’ come into being.Negative selection algorithm is an important branch of artificial immune theory,which has applied to pattern recognition, anomaly detection, and virus protection andother projects. Just using little normal quantitative information, The algorithm candistinguish abnormal state from normal state, negative selection algorithm hascapability to distributed detection, memory and self-learning.The nut of Negative selection algorithm is detector’s generation, This thesislearned Real-Valued Detector Generation Algorithm based on the Partition-TestProess and designed a multi-classification model based on the latest information ofthis field, real value encoding in this model, and euclidean distance measure is theaffinity of the subject sample and the detector, the model were tested by iris data setsand oil dissolved gas data set in power transformer, the latter data set were collectedfrom the literature. In terms of transformer fault diagnosis, kernel principalcomponent analysis technique applied to extract the amount of sample characteristics.this thesis introduced least squares support vector machine and BP neural network todo comparison, results show that the model based on real negative selection has10percentage points higher fault detection rate.The current Real-Valued Detector Generation Algorithm was found existsautoimmune reaction when the input sample has an sparse distribution, means thatdetector will generated in self-distribution space, which will lead an misdiagnosis.furthermore, when the given sample has a high Dimensions, detector’s generate willconsume too much time, contrary to these problems, KPCA technology wasemployed, which is an powerful tool in compress feature field. And we specially designed two rules to detecting the presence of pathogenic detector, respectivelyusing irregular distributed self-data，annulus distributed self-data and nearly circledistributed self-data to have a simulation and research on MATLAB. Results showthat this method can effectively eliminate90%of the pathogenic detector, which laysa theoretical foundation for negative selection algorithm’s application in patternrecognition area.