A scale-adaptive positive selection algorithm based on B-cell immune mechanisms for anomaly detection

Abstract

In current anomaly detection immune algorithms, the methods for setting the detection radius of detectors fail to take into account the concentration characteristic of self samples, which weaken their application effect. In response to this deficiency, we proposed a new type of detector named the scale-adaptive B-cells (SAB-cells) detector, and a novel algorithm named scale-adaptive positive selection algorithm (SA-PSA). This algorithm is mainly based on the B-cell immune mechanisms of clonal variation and network suppression. In SA-PSA, the detection radius of SAB-cells can be adaptively adjusted by clonal variation, and the number of redundant SAB-cells can be effectively compressed by fusion variation, so as to eventually obtain efficient detectors. Based on the Iris data set, firstly, we analyzed the effects of three main control parameters on SA-PSA; secondly, we compared SA-PSA with other mainstream anomaly detection immune algorithms by three performance indicators; thirdly, we performed the analysis of receiver operating characteristic (ROC) curve and verified the effectiveness of SA-PSA. At last, we also applied SA-PSA to bearing anomaly detection and further verified its effectiveness in more complicated engineering applications.

Introduction

With the increasing level of complexity and automation in modern machinery, operators and system maintenance personnel need more effective and efficient techniques to monitor the operation status of machines for the realization of condition-based maintenance (CBM) (Grall et al., 2002, Jardine et al., 2006). In CBM, the maintenance of a machine is usually performed based on an assessment or prediction of its health status instead of service time, and extended machine life, reduced downtime, and enhanced operational safety are achieved. At present, as a means of fault prevention, machine health detection (or anomaly detection) has been studied extensively, and a series of research results have been obtained. These results play a positive role in many important fields, including fault diagnosis (Kong et al., 2019, Xiao et al., 2019, Yan et al., 2014), signal processing (Wu et al., 2019, Zabihi et al., 2016), dynamics analysis (He et al., 2019b, Wei and Liu, 2019, Zhang et al., 2018), and reliability analysis (He et al., 2019a, Peng et al., 2016). Most of anomaly detection methods, such as support vector machine (SVM) (Chen et al., 2013, Su et al., 2015), relevance vector machine (RVM) (He et al., 2017, Wang et al., 2014), envelope spectrum analysis (Jiang et al., 2013) and artificial neural network (ANN) (Ali et al., 2015, Wang et al., 2004), establish the classification model of anomaly detection based on the training of a large number of fault samples. However, most of the time, a machine will stop working immediately once it fails, which makes the fault samples rare. Thus anomaly detection methods which depend on many fault samples are difficult to be applied to this case.

The immune system which is a highly dynamic system can recognize and classify many different and unseen antigen molecules at any given time, and therefore has been a rich inspiration source for anomaly detection and pattern classification (Hocine et al., 2017). One of the earliest engineering applications of the artificial immune system (AIS) was the negative selection algorithms (NSAs). The first NSA was proposed by Forrest et al. (1994), which was inspired by the mechanism of T-cell maturation in the thymus. As a one-class classification algorithm, NSA attracted widespread interest in anomaly detection because it can be trained only with normal samples (or self samples) and without abnormal samples (or non-self samples). With the in-depth study of NSA, a variety of modified versions of this algorithm were proposed (Ji and Dasgupta, 2004, Jinquan et al., 2009, Stibor et al., 2005).

In initial NSAs, self and non-self samples were represented with binary encoding. However, the number of detectors produced by binary NSAs is exponential with the size of problems analyzed, so these algorithms can hardly process many applications described in real-valued space (González et al., 2003, Ji and Dasgupta, 2007). Thus González et al. (2002) proposed the first real-valued NSA called real-valued negative selection (RNS). Based on this, Ji and Dasgupta (2004) described an enhanced real-valued NSA called V-detector. Since then, many other real-valued NSAs had been designed and applied in engineering such as aircraft fault diagnosis (Dasgupta et al., 2004).

In addition, many researchers combined NSA with other methods. Gómez et al. (2003) further extended the real-valued NSA into fuzzy classification and used a membership function to describe the degree of a sample belonging to self or non-self. González and Dasgupta (2003) combined the negative selection with other traditional classification algorithms to find a boundary between normal and abnormal classes. Esponda et al. (2004) combined the negative selection with the positive selection and proposed a formal framework to achieve the appropriate number of detectors.

The positive selection mentioned here is another immune calculation method, which simulates the specific recognition mechanism of human B-cells: for a specific antigen, the corresponding original B-cells are able to produce many mature B-cells through clonal variation; these mature B-cells have a high affinity to the specific antigen and can maintain the memory of it. Based on the immune network dynamics and the clonal selection mechanism of B-cells, Timmis (2000) put forward an artificial immune network model called a resource limited artificial immune system (RLAIS). De Casto and Von Zuben (2000) proposed another artificial immune network model called aiNet, which simulated the mechanisms of clonal variation and network suppression in specific immunity. Inspired by RLAIS, Watkins (2001) combined the concept of artificial recognition balls to propose a successful immune classifier called artificial immune recognition system (AIRS). AIRS simulated clonal variation, network suppression, limited resources, and other immune mechanisms. Although these B-cell immune mechanisms mentioned here have been used in clustering and classification, little attention has been paid to the anomaly detection.

In this study, we proposed a new type of detector named scale-adaptive B-cells (SAB-cells) and a novel anomaly detection algorithm named scale-adaptive positive selection algorithm (SA-PSA). In SA-PSA, the concentration characteristic of self samples is considered into the detection radius of SAB-cells. The antigenic closeness centrality is used as the action scale of SAB-cells so that their detection radius can be adaptively adjusted to effectively divide the state space into self space and non-self space. The number of redundant mature SAB-cells can also be effectively reduced by the process of fusion variation (see Section 3.2.3), which is controlled by the two parameters of the cell fusion factor ${\lambda }_f$ and the local restriction factor ${\lambda }_{pos}$.

The remaining sections of the paper are structured as follows. The concept of SAB-cells and the algorithm of SA-PSA are presented in Sections 2 Scale-adaptive B-cells, 3 Scale-adaptive positive selection algorithm (SA-PSA), respectively. The experiments and discussion are presented in Section 4. The engineering application of SA-PSA is presented in Section 5. In Section 6, conclusions are provided.