Design of an Artificial Immune System for fault detection: A Negative
Expert Systems with Applications 37 (2010) 5507–5513
Contents lists available at ScienceDirect
Expert Systems with Applications journal homepage: www.elsevier.com/locate/eswa
Design of an Artificial Immune System for fault detection: A Negative Selection Approach C.A. Laurentys *, G. Ronacher, R.M. Palhares, W.M. Caminhas Federal University of Minas Gerais, Av. Antonio Carlos 6627, Belo Horizonte – MG, Brazil
a r t i c l e
i n f o
Keywords: Fault detection Artificial Immune System Negative selection Computational intelligence Decision support Dynamic systems
a b s t r a c t This paper presents a methodology that designs a fault detection Artificial Immune System (AIS) based on immune theory. The fault detection is a challenging problem due to increasing complexity of processes and agility necessary to avoid malfunction or accidents. The key fault detection challenge is determining the difference between normal and potential harmful activities. A promising solution is emerging in the form of AIS. The Self Nonself theory inspired an immune-based fault detection approach. This article proposes the AIS Multi-Operational Algorithm based on the Negative Selection Algorithm. The proposed algorithm is used to a DC motor fault model benchmark to compare its relative performance to others fault detection algorithms. The results show that the strategy developed is promising for incipient and abrupt fault detection. Ó 2010 Elsevier Ltd. All rights reserved.
1. Introduction The fault detection is a major problem in the area of engineering processes. It is one of the vital components for the Abnormal Event Management (AEM) which has attracted attention. The AEM deals with: detection, diagnosis and correction of abnormal conditions in real-time at processes operation. Nowadays plant operators perform operations with complex decision-making as: detecting abnormalities, identifying the fundamental cause, predicting consequences of failures beyond the planning and implementation of corrective actions (Fangping, 2003) In this context recently, it has been an advance in the area of 67 Artificial Immune System (De Almeida, Palhares, & Caminhas, 2010). The AEM is becoming increasingly challenging due the size and complexity of procedures and the broad scope of its activities, encompassing a variety of factors such as degradation of the process, measurements inadequate, incomplete and not reliable and its interrelation with the human action (Venkatasubramanian, Rengaswamy, & Yin, 2003). Regarding the complexity, the industries of manufacturing process and suffer pressure to increase the quality of products and environmental standards each time more restrictive. To meet the growing standard of quality, industrial processes added a set of observed variables. For example, there are references in the literature of cases up to 1500 variables to be observed by the second (Bailey, 1984).
* Corresponding author. E-mail address: [email protected] (C.A. Laurentys). 0957-4174/$ - see front matter Ó 2010 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2010.02.004
A large number of variables that interact dynamically during a process give high complexity of industrial systems that despite highly automated, they still are dependent on human performance in various aspects. Analyzing the context is not surprising that people responsible for AEM take often incorrect decisions. Industrial statistics shows that 70% accidents are caused by human mistakes (Kavuri & Venkatasubramanian, 1992). It is necessary a tool to help the monitoring of deviations from procedures, improving the handling of security problems and qualities intrinsic to the process. The challenge today is the automation of the AEM using computer systems (De Almeida, Bomfim, Menezes, & Caminhas, 2010). These systems aim to allow automatic AEM. In this context, recently, it has been an advance in the area of Artificial Immune System. Increasingly, there are groups and national and international events related to the subject. One area that recently has gain attention is the Immune Engineering. By Immune Engineering understand a formal structure for the development of AIS (Dasgupta & González, 2003). The AIS have emerged from attempts to model and apply the principles in the development of new immunological computational tools (de Castro Silva, 2001; de Castro Silva & Timmis, 2002a, 2002). Among many applications of SIA can be cited: recognition of patterns, clustering (de Castro Silva & Timmis, 2002b), learning (de Castro Silva & Zuben, 2000), detection of failures, anomalies (de Castro Silva & Zuben, 2002a), etc. The relevance of detection of anomalies and failures can be found in several works on related topics. One of the first algorithms was proposed by Forrest et al. and adapted to detect anomalies in time series (Dasgupta & Krishnakumar, 2004).
5508
C.A. Laurentys et al. / Expert Systems with Applications 37 (2010) 5507–5513
Immune-based techniques are gaining popularity in a wide area of applications showing that has emerged as a new branch of Artificial Intelligence. The main characteristics of AIS are the powerful information processing capability, pattern recognition, learning, memory and immune distributive nature provide rich metaphors for its artificial (computational) counterpart. In this context, Artificial Immune Systems (AIS) (Dasgupta & Forrest, 1996) are defined as a new computational paradigm based on metaphors of the biological immune systems. This paper proposes, implements and validates an AIS to automate the monitoring and fault detection phases of AEM in order to create a decision-making tool to support operator actions in a dynamic system avoiding malfunction or accidents. This paper will describe two existing algorithms based on negative selection and propose a new algorithm, called Multi-Operational (MO) Algorithm. The paper presents a performance comparison between the two algorithms and MO Algorithm into a pattern classification of the DC model motor. Considerations are made on the standardization of data and suggestions for future work are mentioned. The key contribution is an AIS inspired on the Self Nonself immune theory, specifically using a Negative Selection Algorithm, to automate the AEM fault detection. This article is organized as following: Negative Selection Overview section describe the found literature of Negative Selection Algorithm that inspired its use for the proposed AIS, Methodology section describes the Multi-Operational Algorithm fault detection inspired detailing the implementation of the AIS, Process validation brief description section defines the process that was applied for validation of the proposed AIS, Results and Discussion section presents the algorithm validation database and discuss key results achieved, Conclusion section point out the major benefits of using this approach. It is important to stress that the proposed methodology was applied to a fault detection model to compare its relative performance to others algorithms. The results show that the strategy developed is promising for incipient and abrupt fault detection in dynamic systems. 2. Negative Selection Overview 2.1. A brief immune description of Self Nonself For over 50 years immunologists have been based their thoughts, experiments, and clinical treatments on the idea that the immune system functions by making a distinction between self (related to belonging molecules in the organism) and nonself (related to foreign molecules in the organism) (de Castro, 2006; Forrest, Perelson, Allen, & Cherukuri, 1994). AIS take their inspiration from the operation of the human immune system, which is capable of recognizing virtually any pathogenic agent. The Self Nonself theory proposes that the recognizing is done by distinguishing the body’s own cells and molecules (self) from foreign ones (nonself). This is a classic immune theory to understand the immune system as a system that distinguishes self from nonself, where self is considered the body and nonself all pathogens. The natural immune system is a complex adaptive system which efficiently employs several mechanisms for recognizing novel patterns representing pathogens. The key role of the immune system is to recognize all cells (or molecules) within the body as self and categorize foreign cells as nonself to generate specific response.
Fig. 1. Diagram evidencing the context of the Self Nonself theory. The immune system based on Self Nonself theory (SNS) or based on infections-nonself (INS) and the danger theory that the immune response is triggered by alarms/dangerous signals (danger) (Deaton et al., 1997).
The Self Nonself theory suggests that the trigger to an immune response is foreignness of the cells (or molecules). During the years different theories have been proposed to explain the behavior of the immune response based on other triggers. Other theories (Deaton et al., 1997) are summarized by Fig. 1. 2.2. Negative Selection Algorithms Based on the Self Nonself theory some algorithms were presented in order to detect faults or anomalous events. Two of these algorithms (constant radius and variant radius) were implemented and analyzed in order to provide context for the MO Algorithm and are briefly presented in the next topics. 2.2.1. Constant Radius Algorithms The detection algorithm is a constant radius of relatively simple model used for the first contact with the Negative Selection Algorithm (NSA). The key points of NSA are described by the fluxogram in the Fig. 2. Illustrative results of Constant radius NSA is depicted at Fig. 3. This algorithm depends on a high number of detectors to achieve coverage, and a satisfactory coverage is not guaranteed. Analyzing the Fig. 3, the coverage obtained with an increased number of detectors has a high overlap of the detectors. This factor is undesirable because it increases the computational cost and makes more difficult the fault definition in order to specialize some future detectors for certain failures. Added to these factors the algorithm is not able to generate boundary between self and nonself, that is, hardly any generator detector contains is tangent to the self set (S). 2.2.2. Variable Radius Algorithms The V-detector or Variable Radius Algorithm NSA was initially proposed in Matzinger (2002) shows an evolution in the previous
C.A. Laurentys et al. / Expert Systems with Applications 37 (2010) 5507–5513
5509
Fig. 2. Constant radius algorithm NSA fluxogram.
Fig. 3. Results examples (a) m = 200 detectors and (b) m = 500 detectors. The blue circles define the self set (S) as a circular crown region and the pink ones define the detectors set (D).
algorithm. The Variable Radius Algorithm key feature is the fact that the detectors do not have constant radius. Moreover, it has a strategy that only add detectors which are not detected by other detectors. Fig. 4 shows the basic workflow of this algorithm. The Variable Radius Algorithm solves some drawbacks of the algorithm of constant radius, because the detectors will be tangent to the self set (S). However, this algorithm still has limitations: the high overlap of the detectors, in spite of having a mechanism to prevent such unwanted overlap, Non guaranteed coverage of the nonself set, since the calculation of coverage is given by the naive estimation described in de Castro Silva (2001). This estimation has been described as not being of high reliability.
3. Methodology The MO Algorithm was developed with the aim to promote better coverage of nonself space with greater efficiency, meaning in this context: lowest number of detectors with the largest radius possible. The detectors should be arranged in the most appropriate manner to minimize the overlap between them. For a better understanding will be first describe the operational tools used by MO Algorithm. The large number of operational tools is the responsible for the quality of results, and is the motivator of the name. 3.1. MO Algorithm operational tools This section will describe the tools used in the MO Algorithm.
The Fig. 5 illustrates the results of the algorithm for two different levels of coverage. The two algorithms studied have a set of characteristics that were inspiration to develop a new algorithm that has traces of genetic algorithms and presents a development in relation to previous results. The key aspects of the MO Algorithm will be described at the methodology section.
3.1.1. Tangent detector radius calculation In the MO Algorithm each detector has its radius (R) defined by the same procedure of the V-detector algorithm as defined by Eq. (1).
rd ¼ dist rs
ð1Þ
5510
C.A. Laurentys et al. / Expert Systems with Applications 37 (2010) 5507–5513
Define: Create new detector x
Self Set (S), Self radius (rs), Number of detectors (m).
No x is detected by other detector?
Yes
Yes
Minimum Coverage achieved?
No Final D
Generate x radius Tangent to S
Back to create a new detector
Add x to D.
Yes
Yes
No
x radius > rs?
Has been more than t attempts to find x radius > rs?
Yes
No
|D|<m?
Back to create a new detector.
No Final D
Fig. 4. Fluxogram of the V-detector algorithm.
Fig. 5. Examples of V-detector results (a) 99% estimated coverage and (b) 99.9% estimated coverage. The blue circles define the self set (S) and de pink ones defines the detectors set (D). (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
where rd is the detector radius, dist the distance between the generated detector and the nearest self space (S) point and rs is the self radius of the nearest self space. Applying this procedure the detectors will be tangent to the self space as illustrated by Fig. 6. 3.1.2. Detector moving Another operation available is the detector moving. By moving the detectors the overlap may reduce and achieve better coverage. The strategy to move the detectors is based on the Eq. (2).
C new ¼ C old þ a
C old C near kC old C near k
ð2Þ Fig. 6. Creating a new detector tangent to the nearest self space point.
5511
C.A. Laurentys et al. / Expert Systems with Applications 37 (2010) 5507–5513
where C represents the center of:
4. Process validation description
new subscript indicates the new coordinates of the detector, old subscript refers to the prior coordinates of the detectors, near subscript indicates the nearest entity from the detector, a is an exponential factor directly related to the numbers of moving attempts (t). The Fig. 7 illustrates this operation.
3.1.3. Overlap measurement In the MO Algorithm each detector must go through several steps before being considered a mature detector to be added to the set of detectors (D). The index r is responsible for assessing the maturation of each detector, measuring its overlap with respect to other detectors. This index is defined by Eq. (3):
r¼
rd rref þ dist 2 rd
ð3Þ
where rd is the detector radius, rref is the radius of the reference detector and dist is the distance of the two detectors. It must be stress that the index r proposed to assess the maturity of the detector increases with rd and with the variable dist. Unitary value of r indicates a complete mature detector, meaning that its tangent to all other selected detectors. 3.1.4. Exponential radius decay Even using the operation describe it may be difficult to allocate some detectors. In these cases, exponential radius decay is used. 3.2. MO Algorithm The MO Algorithm creates random detectors and attempts to allocate them a pre defined set of operators and using principles of NSA in the nonself space. Fig. 8 describes the MO Algorithm. After creating a new random detector center and calculate its radius according to the tangent detector tool, it is overlapping is estimated by the overlap measurement tool. If it is above a preestablished limit, the detector is moved a few times (t at top). If this action is not sufficient to reduce the overlap, the radius of the detector suffers successive decays, and finally as a last resort, increases the allowed overlapping in order to try to allocate the detector. As shown by the other algorithms Fig. 9 illustrates the MO Algorithm using different conditions of stop. This process of creation and maturation of the detector continues until it reaches the number maximum of detector or a preset number of attempts to allocate detectors has been reached. In a general way, the last alternative has shown as most effective way of convergence, however a discussion of the convergence of algorithms based on AIS will be discussed further.
Near Center
Old Center
New Center
Fig. 7. Moving detectors to reduce overlapping.
The MO Algorithm was applied in a model of self-formulation of a failure to drive a motor of continuous current (de Castro Silva & Timmis, 2002a). Considering that ia, ifd, e, xr are the measured variables and using the equation of state notation, the model of continuous direct current (DC) motor is represented by:
2 3 L 3 2 3 kaa : Lraa kaa : Lafda :x3 0 x1 x_ 1 6 7 r 6_ 7 6 0 6 7 0 7 kaf : Lfd : x 7 4 x2 5 ¼ 6 4 25 fd 4 5 Lafd x x_ 3 Bm 3 :x2 0 J J m 2m 3 2 3 2 3 1 # 0 " 0 y1 cca :v a 6 La 1 7 kaa :k 6 7 6 7 6 7 þ 4 0 5:½T L 4 y2 5 þ 4 0 L 5: :v fd kaf :k ccfd fd 1 y3 0 0 2 f 32 3 2 3 2 3 k ia 0 0 y1 x1 ia 6 7 6 7 6 7 6 7 f ¼6 0 7 4 0 ki 54 x2 5; onde4 y2 5 ¼ 4 ifd 5eki ¼ 1 ki 2
fd
0
0
f k xr
x3
y3
xr
where the parameters of the engine are:
ra: resistance of the armature circuit; Rfd: resistance of the field circuit; La: inductance of the armature circuit; Lfd: inductance of the field circuit; Lafd: mutual inductance armor / field; TL: mechanical load; Bm: coefficient of viscous friction; Jm: moment of inertia of the engine. The coefficients applied to the fault model are:
kaa 2 f0; 1g, where (0) indicates the disconnection of the armor; K afd 2 f0; 1g, where (0) indicates the disconnection of the field; kcca 2 f0; 1g, where (1) indicates short-circuit the armor; kccfd 2 f0; 1g, where (1) indicates short-circuit of the field; f kia 2 f0; 1g, where (1) indicates fault sensor in the armor current; f kifd 2 f0; 1g, where (1) indicates fault sensor of the field current; f kxr 2 f0; 1g, where (1) indicates fault in the velocity sensor.
The model was simulated for 3 s, starting with the failure to act to half that time. It was defined that the MO Algorithm had a self space with three variables: the current circuit of armor, it speed of rotation and current of the field circuit.
5. Results and discussions The results for each fault are presented in Table 1 and the detection and the average time spent to detect the fault for a 100 sets of detectors generated by MO Algorithm. During the study it was noticed that a crucial factor for the proper functioning of the algorithms of AIS is the correct normalization of real data. It is recommended a definition of the self radius as a function of uncertainty and tolerances of the data collection system. Incorrect normalization prevents detection of a few failures, because the difference between data with and without faults can be very subtle and the algorithm will not be able to detect it. This was one of the problems encountered, which did not allow the faults 5 and 7 to be fully detected. This was caused by poor normalization and not by mistake of the proposed algorithm by itself.
5512
C.A. Laurentys et al. / Expert Systems with Applications 37 (2010) 5507–5513
Initialize: Self Set (S), Self radius (rs), # of detectors (m) Min radius (r_min).
Create new detector x with Tangent Radius Calculation
x Radius < minimum radius
Detector Assessment
Attempts < t
Detector Overlap Measurement
Yes
x overlapping lower than threshold ?
No
Attempts of Detector Moving < t ?
Yes
No
Attempts of Radius Exponential Decay < t or x Radius < r_min ?
No
Increase overlapping threshold for x
Yes Repeat Until |D| > m Or Attempts of allocation < t
Add x to D
Fig. 8. Description of the MO Algorithm focusing on the operational tools.
Fig. 9. Examples of results with the MO Algorithm with (a) convergence for the maximum number of detectors and (b) for attempts to limit the allocation of detectors. The blue cells define the self set and the pink ones are created by the MO Algorithm. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
Table 1 Detection rate and detection time for each fault using the MO Algorithm (de Castro Silva et al., 2000). Fault ID
Fault description
Detection rate (%)
Average time to detect (s)
1 2 3 4 5 6 7
Disconnection of the armor Armor short-circuit Disconnection of the field Field short-circuit Armor current sensor fault Field current sensor fault Velocity sensor fault
100 100 100 100 75 100 30
0.0060 0.0137 0.0040 0.0645 0.0040 0.0040 0.0040
In addition to mixed use of detectors in the region and rectangular shapes, circular and ellipsoid appears as interesting alternative to be explored to alleviate the problems of normalization. The ellipses may be interpreted as the difference in the uncertainties of the sensors of two different data. Another alternative is to create layers of detectors: sets of detectors for each system with a fault being reported only if there is equal response in the layers of detectors. The computational cost is an important factor to consider, both in training and in the process of the detection algorithms. The MO Algorithm has a high computational cost in training on the algorithm of constant radius and the V-detector, however its
C.A. Laurentys et al. / Expert Systems with Applications 37 (2010) 5507–5513
computational cost is less than other algorithms in the literature (Caminhas, 1997). The cost in training is not a risk factor for AIS, as the training is done only once for each system in the initial phase. 6. Conclusion This article proposed, implemented and validated an algorithm called Multi-Operational Algorithm. It was carried out a comparison on the operation of two other pre-existing algorithms in the literature. The MO Algorithm was applied in a pattern of failures in DC motor model. The results show that the strategy developed is promising for fault detection although some faults were not detected because of the data normalization. Such considerations have noted the computational cost relatively favorable for the detection, although the training can be improved. Acknowledgment This work has been supported by the Brazilian agencies CNPq and FAPEMIG. References Bailey, S. J. (1984). From desktop to plant floor, a CRT is the control operators window on the process. Control Engineering, 31(6), 86–90. Caminhas, W. M. (1997). Estratégias de detecção e diagnóstico de falhas em sistemas dinâmicos. Tese de Doutorado, Faculdade de Engenharia Elétrica e de Computação. UNICAMP. Dasgupta, D., & Forrest, S. (1996). Novelty detection in time series data using ideas from immunology. In proceedings of the international conference on intelligent systems. Dasgupta, D., & Krishnakumar, K. (2004). Negative selection algorithm for aircraft fault detection. In Artificial immune systems. Berlin/Heidelberg: Springer. Dasgupta, D., & González, F. (2003). Artificial immune system research in the last five years. IEEE Press.
5513
De Almeida, C. A., Bomfim, C.H., Menezes, B., & Caminhas, W. M. (2010). Design of a pipeline leakage detection using expert system: A novel approach. Applied Soft Computing Journal, doi:10.1016/j.asoc.2010.02.2005. De Almeida, C. A., Palhares, R. M., & Caminhas, W. M. (2010). Design of an artificial immune system based on danger model for fault detection. Expert System with Applications, doi:10.1016/j.eswa .2009.12.079. de Castro, L. N., (2006). Fundamentals of natural computing: Basic concepts, algorithms, and applications. Chapman and Hall/CRC Computer and Information Sciences. de Castro Silva, L. N. (2001). Engenharia Imunológica: Desenvolvimento e Aplicação de Ferramentas Computacionais Inspiradas em Sistemas Imunológicos Artificiais [Tese de Doutorado]. Campinas: DCA/FEEC/Unicamp. de Castro Silva, L. N., & Timmis, J. (2002a). An artificial immune network for multimodal function optimization. In Proceedings of IEEE congress on evolutionary computation (CEC’02) (Vol. 1, pp. 669–674). Hawaii: IEEE. de Castro Silva, L. N., & Zuben, F. J. V. (2000). An evolutionary immune network for data clustering. In Proceedings of the IEEE SBRN (Brazilian Symposium on Artificial Neural Networks) (pp. 84–89). Rio de Janeiro: SBRN. de Castro Silva, L. N., & Timmis, J. (2002b). Artificial immune systems: A novel paradigm to pattern recognition. In J. M. Corchado, L. Alonso, & C. Fyfe (Eds.), Artificial neural networks in pattern recognition (pp. 67–84). Paisley (UK): University of Paisley. de Castro Silva, L. N., & Zuben, F. J. V. (2002a). Learning and optimization using the clonal selection principle. IEEE Transactions on Evolutionary Computation, Special Issue on Artificial Immune Systems, IEEE, 6, 239–251. de Castro, L., & Timmis, J. (2002). Artificial immune systems: a new computational approach. London, UK: Springer-Verlag. Deaton, R., Garzon, M., Rose, J. A., Murphy, R. C., Stevens, S. E., Jr., & Franceschetti, D. R. (1997). DNA based artificial immune system for self–nonself discrimination. In Proceedings of the IEEE international conference on systems, Man, and Cybernetics, Orlando, Florida. Fangping, M. (2003). Multivariate statistical process monitoring and its integration with HAZOP analysis for abnormal event management. Master’s thesis, Purdue University. Forrest, S., Perelson, A. S., Allen, L., & Cherukuri, R. (1994). Self-nonself discrimination in a computer. In Proceedings of the IEEE symposium on research in security and privacy. IEEE Press. Kavuri, S., & Venkatasubramanian, P. (1992). Combining pattern classification and assumption-based techniques for process fault diagnosis. Computers and Chemical Engineering, 16(4), 299–312. Matzinger, P. (2002). The danger model: a renewed sense of self. Science Magazine, 296. Venkatasubramanian, V., Rengaswamy, R., & Yin, K. (2003). A review of fault detection and diagnosis. Computer and Chemical Engineering(27), 293–311.
Contents lists available at ScienceDirect
Expert Systems with Applications journal homepage: www.elsevier.com/locate/eswa
Design of an Artificial Immune System for fault detection: A Negative Selection Approach C.A. Laurentys *, G. Ronacher, R.M. Palhares, W.M. Caminhas Federal University of Minas Gerais, Av. Antonio Carlos 6627, Belo Horizonte – MG, Brazil
a r t i c l e
i n f o
Keywords: Fault detection Artificial Immune System Negative selection Computational intelligence Decision support Dynamic systems
a b s t r a c t This paper presents a methodology that designs a fault detection Artificial Immune System (AIS) based on immune theory. The fault detection is a challenging problem due to increasing complexity of processes and agility necessary to avoid malfunction or accidents. The key fault detection challenge is determining the difference between normal and potential harmful activities. A promising solution is emerging in the form of AIS. The Self Nonself theory inspired an immune-based fault detection approach. This article proposes the AIS Multi-Operational Algorithm based on the Negative Selection Algorithm. The proposed algorithm is used to a DC motor fault model benchmark to compare its relative performance to others fault detection algorithms. The results show that the strategy developed is promising for incipient and abrupt fault detection. Ó 2010 Elsevier Ltd. All rights reserved.
1. Introduction The fault detection is a major problem in the area of engineering processes. It is one of the vital components for the Abnormal Event Management (AEM) which has attracted attention. The AEM deals with: detection, diagnosis and correction of abnormal conditions in real-time at processes operation. Nowadays plant operators perform operations with complex decision-making as: detecting abnormalities, identifying the fundamental cause, predicting consequences of failures beyond the planning and implementation of corrective actions (Fangping, 2003) In this context recently, it has been an advance in the area of 67 Artificial Immune System (De Almeida, Palhares, & Caminhas, 2010). The AEM is becoming increasingly challenging due the size and complexity of procedures and the broad scope of its activities, encompassing a variety of factors such as degradation of the process, measurements inadequate, incomplete and not reliable and its interrelation with the human action (Venkatasubramanian, Rengaswamy, & Yin, 2003). Regarding the complexity, the industries of manufacturing process and suffer pressure to increase the quality of products and environmental standards each time more restrictive. To meet the growing standard of quality, industrial processes added a set of observed variables. For example, there are references in the literature of cases up to 1500 variables to be observed by the second (Bailey, 1984).
* Corresponding author. E-mail address: [email protected] (C.A. Laurentys). 0957-4174/$ - see front matter Ó 2010 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2010.02.004
A large number of variables that interact dynamically during a process give high complexity of industrial systems that despite highly automated, they still are dependent on human performance in various aspects. Analyzing the context is not surprising that people responsible for AEM take often incorrect decisions. Industrial statistics shows that 70% accidents are caused by human mistakes (Kavuri & Venkatasubramanian, 1992). It is necessary a tool to help the monitoring of deviations from procedures, improving the handling of security problems and qualities intrinsic to the process. The challenge today is the automation of the AEM using computer systems (De Almeida, Bomfim, Menezes, & Caminhas, 2010). These systems aim to allow automatic AEM. In this context, recently, it has been an advance in the area of Artificial Immune System. Increasingly, there are groups and national and international events related to the subject. One area that recently has gain attention is the Immune Engineering. By Immune Engineering understand a formal structure for the development of AIS (Dasgupta & González, 2003). The AIS have emerged from attempts to model and apply the principles in the development of new immunological computational tools (de Castro Silva, 2001; de Castro Silva & Timmis, 2002a, 2002). Among many applications of SIA can be cited: recognition of patterns, clustering (de Castro Silva & Timmis, 2002b), learning (de Castro Silva & Zuben, 2000), detection of failures, anomalies (de Castro Silva & Zuben, 2002a), etc. The relevance of detection of anomalies and failures can be found in several works on related topics. One of the first algorithms was proposed by Forrest et al. and adapted to detect anomalies in time series (Dasgupta & Krishnakumar, 2004).
5508
C.A. Laurentys et al. / Expert Systems with Applications 37 (2010) 5507–5513
Immune-based techniques are gaining popularity in a wide area of applications showing that has emerged as a new branch of Artificial Intelligence. The main characteristics of AIS are the powerful information processing capability, pattern recognition, learning, memory and immune distributive nature provide rich metaphors for its artificial (computational) counterpart. In this context, Artificial Immune Systems (AIS) (Dasgupta & Forrest, 1996) are defined as a new computational paradigm based on metaphors of the biological immune systems. This paper proposes, implements and validates an AIS to automate the monitoring and fault detection phases of AEM in order to create a decision-making tool to support operator actions in a dynamic system avoiding malfunction or accidents. This paper will describe two existing algorithms based on negative selection and propose a new algorithm, called Multi-Operational (MO) Algorithm. The paper presents a performance comparison between the two algorithms and MO Algorithm into a pattern classification of the DC model motor. Considerations are made on the standardization of data and suggestions for future work are mentioned. The key contribution is an AIS inspired on the Self Nonself immune theory, specifically using a Negative Selection Algorithm, to automate the AEM fault detection. This article is organized as following: Negative Selection Overview section describe the found literature of Negative Selection Algorithm that inspired its use for the proposed AIS, Methodology section describes the Multi-Operational Algorithm fault detection inspired detailing the implementation of the AIS, Process validation brief description section defines the process that was applied for validation of the proposed AIS, Results and Discussion section presents the algorithm validation database and discuss key results achieved, Conclusion section point out the major benefits of using this approach. It is important to stress that the proposed methodology was applied to a fault detection model to compare its relative performance to others algorithms. The results show that the strategy developed is promising for incipient and abrupt fault detection in dynamic systems. 2. Negative Selection Overview 2.1. A brief immune description of Self Nonself For over 50 years immunologists have been based their thoughts, experiments, and clinical treatments on the idea that the immune system functions by making a distinction between self (related to belonging molecules in the organism) and nonself (related to foreign molecules in the organism) (de Castro, 2006; Forrest, Perelson, Allen, & Cherukuri, 1994). AIS take their inspiration from the operation of the human immune system, which is capable of recognizing virtually any pathogenic agent. The Self Nonself theory proposes that the recognizing is done by distinguishing the body’s own cells and molecules (self) from foreign ones (nonself). This is a classic immune theory to understand the immune system as a system that distinguishes self from nonself, where self is considered the body and nonself all pathogens. The natural immune system is a complex adaptive system which efficiently employs several mechanisms for recognizing novel patterns representing pathogens. The key role of the immune system is to recognize all cells (or molecules) within the body as self and categorize foreign cells as nonself to generate specific response.
Fig. 1. Diagram evidencing the context of the Self Nonself theory. The immune system based on Self Nonself theory (SNS) or based on infections-nonself (INS) and the danger theory that the immune response is triggered by alarms/dangerous signals (danger) (Deaton et al., 1997).
The Self Nonself theory suggests that the trigger to an immune response is foreignness of the cells (or molecules). During the years different theories have been proposed to explain the behavior of the immune response based on other triggers. Other theories (Deaton et al., 1997) are summarized by Fig. 1. 2.2. Negative Selection Algorithms Based on the Self Nonself theory some algorithms were presented in order to detect faults or anomalous events. Two of these algorithms (constant radius and variant radius) were implemented and analyzed in order to provide context for the MO Algorithm and are briefly presented in the next topics. 2.2.1. Constant Radius Algorithms The detection algorithm is a constant radius of relatively simple model used for the first contact with the Negative Selection Algorithm (NSA). The key points of NSA are described by the fluxogram in the Fig. 2. Illustrative results of Constant radius NSA is depicted at Fig. 3. This algorithm depends on a high number of detectors to achieve coverage, and a satisfactory coverage is not guaranteed. Analyzing the Fig. 3, the coverage obtained with an increased number of detectors has a high overlap of the detectors. This factor is undesirable because it increases the computational cost and makes more difficult the fault definition in order to specialize some future detectors for certain failures. Added to these factors the algorithm is not able to generate boundary between self and nonself, that is, hardly any generator detector contains is tangent to the self set (S). 2.2.2. Variable Radius Algorithms The V-detector or Variable Radius Algorithm NSA was initially proposed in Matzinger (2002) shows an evolution in the previous
C.A. Laurentys et al. / Expert Systems with Applications 37 (2010) 5507–5513
5509
Fig. 2. Constant radius algorithm NSA fluxogram.
Fig. 3. Results examples (a) m = 200 detectors and (b) m = 500 detectors. The blue circles define the self set (S) as a circular crown region and the pink ones define the detectors set (D).
algorithm. The Variable Radius Algorithm key feature is the fact that the detectors do not have constant radius. Moreover, it has a strategy that only add detectors which are not detected by other detectors. Fig. 4 shows the basic workflow of this algorithm. The Variable Radius Algorithm solves some drawbacks of the algorithm of constant radius, because the detectors will be tangent to the self set (S). However, this algorithm still has limitations: the high overlap of the detectors, in spite of having a mechanism to prevent such unwanted overlap, Non guaranteed coverage of the nonself set, since the calculation of coverage is given by the naive estimation described in de Castro Silva (2001). This estimation has been described as not being of high reliability.
3. Methodology The MO Algorithm was developed with the aim to promote better coverage of nonself space with greater efficiency, meaning in this context: lowest number of detectors with the largest radius possible. The detectors should be arranged in the most appropriate manner to minimize the overlap between them. For a better understanding will be first describe the operational tools used by MO Algorithm. The large number of operational tools is the responsible for the quality of results, and is the motivator of the name. 3.1. MO Algorithm operational tools This section will describe the tools used in the MO Algorithm.
The Fig. 5 illustrates the results of the algorithm for two different levels of coverage. The two algorithms studied have a set of characteristics that were inspiration to develop a new algorithm that has traces of genetic algorithms and presents a development in relation to previous results. The key aspects of the MO Algorithm will be described at the methodology section.
3.1.1. Tangent detector radius calculation In the MO Algorithm each detector has its radius (R) defined by the same procedure of the V-detector algorithm as defined by Eq. (1).
rd ¼ dist rs
ð1Þ
5510
C.A. Laurentys et al. / Expert Systems with Applications 37 (2010) 5507–5513
Define: Create new detector x
Self Set (S), Self radius (rs), Number of detectors (m).
No x is detected by other detector?
Yes
Yes
Minimum Coverage achieved?
No Final D
Generate x radius Tangent to S
Back to create a new detector
Add x to D.
Yes
Yes
No
x radius > rs?
Has been more than t attempts to find x radius > rs?
Yes
No
|D|<m?
Back to create a new detector.
No Final D
Fig. 4. Fluxogram of the V-detector algorithm.
Fig. 5. Examples of V-detector results (a) 99% estimated coverage and (b) 99.9% estimated coverage. The blue circles define the self set (S) and de pink ones defines the detectors set (D). (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
where rd is the detector radius, dist the distance between the generated detector and the nearest self space (S) point and rs is the self radius of the nearest self space. Applying this procedure the detectors will be tangent to the self space as illustrated by Fig. 6. 3.1.2. Detector moving Another operation available is the detector moving. By moving the detectors the overlap may reduce and achieve better coverage. The strategy to move the detectors is based on the Eq. (2).
C new ¼ C old þ a
C old C near kC old C near k
ð2Þ Fig. 6. Creating a new detector tangent to the nearest self space point.
5511
C.A. Laurentys et al. / Expert Systems with Applications 37 (2010) 5507–5513
where C represents the center of:
4. Process validation description
new subscript indicates the new coordinates of the detector, old subscript refers to the prior coordinates of the detectors, near subscript indicates the nearest entity from the detector, a is an exponential factor directly related to the numbers of moving attempts (t). The Fig. 7 illustrates this operation.
3.1.3. Overlap measurement In the MO Algorithm each detector must go through several steps before being considered a mature detector to be added to the set of detectors (D). The index r is responsible for assessing the maturation of each detector, measuring its overlap with respect to other detectors. This index is defined by Eq. (3):
r¼
rd rref þ dist 2 rd
ð3Þ
where rd is the detector radius, rref is the radius of the reference detector and dist is the distance of the two detectors. It must be stress that the index r proposed to assess the maturity of the detector increases with rd and with the variable dist. Unitary value of r indicates a complete mature detector, meaning that its tangent to all other selected detectors. 3.1.4. Exponential radius decay Even using the operation describe it may be difficult to allocate some detectors. In these cases, exponential radius decay is used. 3.2. MO Algorithm The MO Algorithm creates random detectors and attempts to allocate them a pre defined set of operators and using principles of NSA in the nonself space. Fig. 8 describes the MO Algorithm. After creating a new random detector center and calculate its radius according to the tangent detector tool, it is overlapping is estimated by the overlap measurement tool. If it is above a preestablished limit, the detector is moved a few times (t at top). If this action is not sufficient to reduce the overlap, the radius of the detector suffers successive decays, and finally as a last resort, increases the allowed overlapping in order to try to allocate the detector. As shown by the other algorithms Fig. 9 illustrates the MO Algorithm using different conditions of stop. This process of creation and maturation of the detector continues until it reaches the number maximum of detector or a preset number of attempts to allocate detectors has been reached. In a general way, the last alternative has shown as most effective way of convergence, however a discussion of the convergence of algorithms based on AIS will be discussed further.
Near Center
Old Center
New Center
Fig. 7. Moving detectors to reduce overlapping.
The MO Algorithm was applied in a model of self-formulation of a failure to drive a motor of continuous current (de Castro Silva & Timmis, 2002a). Considering that ia, ifd, e, xr are the measured variables and using the equation of state notation, the model of continuous direct current (DC) motor is represented by:
2 3 L 3 2 3 kaa : Lraa kaa : Lafda :x3 0 x1 x_ 1 6 7 r 6_ 7 6 0 6 7 0 7 kaf : Lfd : x 7 4 x2 5 ¼ 6 4 25 fd 4 5 Lafd x x_ 3 Bm 3 :x2 0 J J m 2m 3 2 3 2 3 1 # 0 " 0 y1 cca :v a 6 La 1 7 kaa :k 6 7 6 7 6 7 þ 4 0 5:½T L 4 y2 5 þ 4 0 L 5: :v fd kaf :k ccfd fd 1 y3 0 0 2 f 32 3 2 3 2 3 k ia 0 0 y1 x1 ia 6 7 6 7 6 7 6 7 f ¼6 0 7 4 0 ki 54 x2 5; onde4 y2 5 ¼ 4 ifd 5eki ¼ 1 ki 2
fd
0
0
f k xr
x3
y3
xr
where the parameters of the engine are:
ra: resistance of the armature circuit; Rfd: resistance of the field circuit; La: inductance of the armature circuit; Lfd: inductance of the field circuit; Lafd: mutual inductance armor / field; TL: mechanical load; Bm: coefficient of viscous friction; Jm: moment of inertia of the engine. The coefficients applied to the fault model are:
kaa 2 f0; 1g, where (0) indicates the disconnection of the armor; K afd 2 f0; 1g, where (0) indicates the disconnection of the field; kcca 2 f0; 1g, where (1) indicates short-circuit the armor; kccfd 2 f0; 1g, where (1) indicates short-circuit of the field; f kia 2 f0; 1g, where (1) indicates fault sensor in the armor current; f kifd 2 f0; 1g, where (1) indicates fault sensor of the field current; f kxr 2 f0; 1g, where (1) indicates fault in the velocity sensor.
The model was simulated for 3 s, starting with the failure to act to half that time. It was defined that the MO Algorithm had a self space with three variables: the current circuit of armor, it speed of rotation and current of the field circuit.
5. Results and discussions The results for each fault are presented in Table 1 and the detection and the average time spent to detect the fault for a 100 sets of detectors generated by MO Algorithm. During the study it was noticed that a crucial factor for the proper functioning of the algorithms of AIS is the correct normalization of real data. It is recommended a definition of the self radius as a function of uncertainty and tolerances of the data collection system. Incorrect normalization prevents detection of a few failures, because the difference between data with and without faults can be very subtle and the algorithm will not be able to detect it. This was one of the problems encountered, which did not allow the faults 5 and 7 to be fully detected. This was caused by poor normalization and not by mistake of the proposed algorithm by itself.
5512
C.A. Laurentys et al. / Expert Systems with Applications 37 (2010) 5507–5513
Initialize: Self Set (S), Self radius (rs), # of detectors (m) Min radius (r_min).
Create new detector x with Tangent Radius Calculation
x Radius < minimum radius
Detector Assessment
Attempts < t
Detector Overlap Measurement
Yes
x overlapping lower than threshold ?
No
Attempts of Detector Moving < t ?
Yes
No
Attempts of Radius Exponential Decay < t or x Radius < r_min ?
No
Increase overlapping threshold for x
Yes Repeat Until |D| > m Or Attempts of allocation < t
Add x to D
Fig. 8. Description of the MO Algorithm focusing on the operational tools.
Fig. 9. Examples of results with the MO Algorithm with (a) convergence for the maximum number of detectors and (b) for attempts to limit the allocation of detectors. The blue cells define the self set and the pink ones are created by the MO Algorithm. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
Table 1 Detection rate and detection time for each fault using the MO Algorithm (de Castro Silva et al., 2000). Fault ID
Fault description
Detection rate (%)
Average time to detect (s)
1 2 3 4 5 6 7
Disconnection of the armor Armor short-circuit Disconnection of the field Field short-circuit Armor current sensor fault Field current sensor fault Velocity sensor fault
100 100 100 100 75 100 30
0.0060 0.0137 0.0040 0.0645 0.0040 0.0040 0.0040
In addition to mixed use of detectors in the region and rectangular shapes, circular and ellipsoid appears as interesting alternative to be explored to alleviate the problems of normalization. The ellipses may be interpreted as the difference in the uncertainties of the sensors of two different data. Another alternative is to create layers of detectors: sets of detectors for each system with a fault being reported only if there is equal response in the layers of detectors. The computational cost is an important factor to consider, both in training and in the process of the detection algorithms. The MO Algorithm has a high computational cost in training on the algorithm of constant radius and the V-detector, however its
C.A. Laurentys et al. / Expert Systems with Applications 37 (2010) 5507–5513
computational cost is less than other algorithms in the literature (Caminhas, 1997). The cost in training is not a risk factor for AIS, as the training is done only once for each system in the initial phase. 6. Conclusion This article proposed, implemented and validated an algorithm called Multi-Operational Algorithm. It was carried out a comparison on the operation of two other pre-existing algorithms in the literature. The MO Algorithm was applied in a pattern of failures in DC motor model. The results show that the strategy developed is promising for fault detection although some faults were not detected because of the data normalization. Such considerations have noted the computational cost relatively favorable for the detection, although the training can be improved. Acknowledgment This work has been supported by the Brazilian agencies CNPq and FAPEMIG. References Bailey, S. J. (1984). From desktop to plant floor, a CRT is the control operators window on the process. Control Engineering, 31(6), 86–90. Caminhas, W. M. (1997). Estratégias de detecção e diagnóstico de falhas em sistemas dinâmicos. Tese de Doutorado, Faculdade de Engenharia Elétrica e de Computação. UNICAMP. Dasgupta, D., & Forrest, S. (1996). Novelty detection in time series data using ideas from immunology. In proceedings of the international conference on intelligent systems. Dasgupta, D., & Krishnakumar, K. (2004). Negative selection algorithm for aircraft fault detection. In Artificial immune systems. Berlin/Heidelberg: Springer. Dasgupta, D., & González, F. (2003). Artificial immune system research in the last five years. IEEE Press.
5513
De Almeida, C. A., Bomfim, C.H., Menezes, B., & Caminhas, W. M. (2010). Design of a pipeline leakage detection using expert system: A novel approach. Applied Soft Computing Journal, doi:10.1016/j.asoc.2010.02.2005. De Almeida, C. A., Palhares, R. M., & Caminhas, W. M. (2010). Design of an artificial immune system based on danger model for fault detection. Expert System with Applications, doi:10.1016/j.eswa .2009.12.079. de Castro, L. N., (2006). Fundamentals of natural computing: Basic concepts, algorithms, and applications. Chapman and Hall/CRC Computer and Information Sciences. de Castro Silva, L. N. (2001). Engenharia Imunológica: Desenvolvimento e Aplicação de Ferramentas Computacionais Inspiradas em Sistemas Imunológicos Artificiais [Tese de Doutorado]. Campinas: DCA/FEEC/Unicamp. de Castro Silva, L. N., & Timmis, J. (2002a). An artificial immune network for multimodal function optimization. In Proceedings of IEEE congress on evolutionary computation (CEC’02) (Vol. 1, pp. 669–674). Hawaii: IEEE. de Castro Silva, L. N., & Zuben, F. J. V. (2000). An evolutionary immune network for data clustering. In Proceedings of the IEEE SBRN (Brazilian Symposium on Artificial Neural Networks) (pp. 84–89). Rio de Janeiro: SBRN. de Castro Silva, L. N., & Timmis, J. (2002b). Artificial immune systems: A novel paradigm to pattern recognition. In J. M. Corchado, L. Alonso, & C. Fyfe (Eds.), Artificial neural networks in pattern recognition (pp. 67–84). Paisley (UK): University of Paisley. de Castro Silva, L. N., & Zuben, F. J. V. (2002a). Learning and optimization using the clonal selection principle. IEEE Transactions on Evolutionary Computation, Special Issue on Artificial Immune Systems, IEEE, 6, 239–251. de Castro, L., & Timmis, J. (2002). Artificial immune systems: a new computational approach. London, UK: Springer-Verlag. Deaton, R., Garzon, M., Rose, J. A., Murphy, R. C., Stevens, S. E., Jr., & Franceschetti, D. R. (1997). DNA based artificial immune system for self–nonself discrimination. In Proceedings of the IEEE international conference on systems, Man, and Cybernetics, Orlando, Florida. Fangping, M. (2003). Multivariate statistical process monitoring and its integration with HAZOP analysis for abnormal event management. Master’s thesis, Purdue University. Forrest, S., Perelson, A. S., Allen, L., & Cherukuri, R. (1994). Self-nonself discrimination in a computer. In Proceedings of the IEEE symposium on research in security and privacy. IEEE Press. Kavuri, S., & Venkatasubramanian, P. (1992). Combining pattern classification and assumption-based techniques for process fault diagnosis. Computers and Chemical Engineering, 16(4), 299–312. Matzinger, P. (2002). The danger model: a renewed sense of self. Science Magazine, 296. Venkatasubramanian, V., Rengaswamy, R., & Yin, K. (2003). A review of fault detection and diagnosis. Computer and Chemical Engineering(27), 293–311.
