The Methods Of Anomaly Detection And Fault Diagnosis With Online Adaptive Learning Under Small Samples Based On Immune System

The main reasons why the existing intelligent fault diagnosis methods cannot bewidely used are as follows. First, fault samples are insufficient. Second, statedetection and fault diagnosis are separate from each other. Last, training and testingprocess are independent of each other. It is of great significance to research on themethod of anomaly detection and fault diagnosis with online adaptive learningreferring to biological immune mechanism, which is adaptable to the equipment, andhas a lower dependence on the number of fault samples.To improve the detector coverage and eliminate redundancy detectors oftraditional real-valued negative selection algorithm, the Boundary-Fixed NegativeSelection Algorithm (FB-NSA), Fine Boundary-Fixed Negative Selection Algorithm(FFB-NSA), Interface Detector with Boundary Samples (I-detector), and InterfaceDetector with Reduction Boundary Samples (RI-detector) are presented. Thedetection performances of these methods have been analyzed by simulationexperiments use152-dimensional datasets and Iris dataset, after further discussionthe algorithms of these methods. These methods show a better detection performance,compared to other commonly used anomaly detection methods in most cases. Andwhen the detection rate is similar, the number of detectors (boundary samples)decreases successively.After the detection performances of FB-NSA and I-detector are fully analyzed,the Boundary-Fixed Negative Selection Algorithm with Online Adaptive Learningunder Small Samples (OALFB-NSA) and Interface Detector with Online AdaptiveLearning under Small Samples (OALI-detector) are presented, and the reasons forover fitting and under fitting of OALFB-NSA and OALI-detector are also analyzed.The simulation experiments show that OALFB-NSA and OALI-detector are betterthan the traditional RNSA. The reasons why OALI-detector is better than OALFB-NSA are discussed. OALI-detector based on vaccination strategy isproposed. During the testing stage, negative vaccine can overcome over fitting toimprove the detection rate; positive vaccine can overcome under fitting to reducefalse alarm rate.Based on introducing the term of abnormal degree and abnormal level andcombining with the characteristics of OALI-detector, the adaptive hyper-ringdetector (AHr-detector) is presented. And the AHr-detector is an anomaly detectionand fault diagnosis method with online adaptive learning under small samples. Whenfault OALI-detector is built by limited fault samples in Nonself space, and the termof membership between classes is introduced, categorizing the known type faultsamples and clustering the unknown type fault samples are realized.When applied to the diagnosis of rolling bearing, it shows a higher diagnosticaccuracy, compared to other fault diagnosis methods under various conditions.AHr-detector has a broad application prospects, because it realizes integrationbetween anomaly detection and fault diagnosis, has online adaptive learning abilityunder small samples, has better diagnosis capability for both known faults andunknown faults, and adds the known fault samples in the existing model at any time.