Fault Detection Based On The Negative Selection Algorithm

With the growing complexity of modern industrial processes, fault diagnosis methodologies have attracted great attention for safe and effective operation during recent years. The Artificial Immune System (AIS) has provided new inspiration and ideas to the research of fault diagnosis. In this thesis, we study the Negative Selection Algorithm (NSA)-based anomaly detection technique with applications in fault diagnosis.In the NSA, the key problem is the trade-off between the number of detectors and the coverage of non-self space by detectors, that is, how to find the smallest number of detectors to cover the largest nonself space. Two novel NSA for fault detection has been developed here.The Particle Swarm Optimization (PSO) is first used to optimize the randomly generated detectors in the NSA. In our method, for a given number of detectors, the coverage of the nonself space is maximized, while the coverage of the self samples can be minimized. Simulations are performed using both sinusoidal time series (artificial datasets) and vibration signals of motor bearings with broken balls (real-world datasets). Experimental results show that the proposed NSA has remarkable advantages over the NSA optimized by the Simulated Annealing Algorithm in the nonself space coverage. The fault detection rate has been considerably increased as well.Moreover, another NSA based on incremental rectangle detectors, which can occupy more coverage of the non-self space using fewer detectors, is proposed. In this scheme, the D∞distance is employed to measure the self and nonself space coverage. The sizes of the detectors are increased exponentially in each dimension until they overlap with the self samples. The generation strategy of detectors ensures that every detector is extended to its maximum size in each dimension. The number and size of these detectors are optimized by eliminating the redundancy among the existing detectors. Consequently, the detection efficiency of individual detector is increased. The simulation of sinusoidal time series demonstrates that this new NSA performs much better than the above PSO-optimized NSA. Some datasets of different geometry shapes, such as shape comb, cross, intersection, pentagram, ring, stripe, triangle, are also applied to examine the proposed NSA.