A scale-adaptive positive selection algorithm based on B-cell immune mechanisms for anomaly detection
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 and the local restriction factor .
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.
Section snippets
A deficiency of current real-valued immune algorithms
In the anomaly detection algorithms based on AIS, the process that detectors detect normal or abnormal samples is compared to the process that immune cells recognize self or non-self antigens. Therefore, in immune algorithms, there are some mapping relationships between common concepts and immune concepts. The main mapping relationships in this paper are shown in Table 1. At the same time, normal and abnormal samples are also called self and non-self samples, respectively. As shown in Fig. 1,
Scale-adaptive positive selection algorithm (SA-PSA)
Based on the concept of the SAB-cells, we propose the scale-adaptive positive selection algorithm (SA-PSA), which is described in detail in this section. The flow chart of SA-PSA is shown in Fig. 3, which contains three phases: the preprocessing phase, the training phase, and the testing phase.
Experiments and discussion
We select the cell scale factor , the cell fusion factor and the local restriction factor as the main control parameters of SA-PSA. In addition, we select the detection rate DR, the false alarm rate FA and the number of memory SAB-cells n as the main performance indicators of SA-PSA. The detection rate DR and the false alarm rate FA are defined as Eq. (23) (Ji and Dasgupta, 2004, Stibor et al., 2005). Here TP, FN, FP, and TN are the counts of true positive, false
Application in bearing anomaly detection
In order to test the performance of SA-PSA in more complicated engineering applications, we perform a bearing anomaly detection experiment using Normal Baseline Data and 48k Drive End Bearing Fault Data, which are from Case Western Reserve University Bearing Data (Loparo, 2020). We perform wavelet decomposition on the raw time series data with “db16” wavelet for eight levels and use the high-frequency wavelet coefficients energy of each level as the signal feature (Li et al., 2017). In this
Conclusions
In this study, we proposed a new type of detector named scale-adaptive B-cells (SAB-cells) to realize a novel setting method of the detection radius of self detectors, in which the concentration characteristic of self samples is considered. Based on the concept of SAB-cells, we proposed a new anomaly detection algorithm named scale-adaptive positive selection algorithm (SA-PSA), which mainly simulates the B-cell immune mechanisms of clonal variation and network suppression. In SA-PSA, the
CRediT authorship contribution statement
Hongli Zhang: Conceptualization, Supervision, Funding acquisition. Zhongyuan Ren: Methodology, Software, Writing - original draft. Shaojie Xin: Investigation. Shulin Liu: Writing - review & editing. Chao Lan: Methodology, Software. Xin Sun: Investigation.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgments
This work was supported by the National Natural Science Foundation of China (Grants No. 61603238 and 11802168).
References (39)
- et al.
A condition-based maintenance policy for stochastically deteriorating systems
Reliab. Eng. Syst. Saf.
(2002) - et al.
Principles and practical application of the receiver-operating characteristic analysis for diagnostic tests
Prevent. Vet. Med.
(2000) - et al.
A novel fault diagnosis method based on optimal relevance vector machine
Neurocomputing
(2017) - et al.
A review on machinery diagnostics and prognostics implementing condition-based maintenance
Mech. Syst. Signal Process.
(2006) - et al.
A self-adaptive negative selection algorithm used for anomaly detection
Prog. Nat. Sci.
(2009) - et al.
Relevance vector machine for tool wear prediction
Mech. Syst. Signal Process.
(2019) - et al.
A boundary-fixed negative selection algorithm with online adaptive learning under small samples for anomaly detection
Eng. Appl. Artif. Intell.
(2016) - et al.
Multi-fault diagnosis for rotating machinery based on orthogonal supervised linear local tangent space alignment and least square support vector machine
Neurocomputing
(2015) - et al.
Prognosis of machine health condition using neuro-fuzzy systems
Mech. Syst. Signal Process.
(2004) - et al.
1128. Fault diagnosis of rolling bearing based on relevance vector machine and kernel principal component analysis
J. Vibroengineering
(2014)