20.1 Motivation

In recent years, there has been a huge revolution in automation systems. The principle behind this change is due to the rise of distributed decision-making paradigms. Modern control system sensors have evolved from simple transducers of physical quantities to expert systems, able to assess the measures carried out and take the appropriate decisions in complete autonomy. The widespread adoption of this kind of sensor, widely known by the name of “smart sensor,” has allowed an increase in the flexibility of the plants, and at the same time, a simplification of the cabling of the entire infrastructure (Smith and Bowen, 1995).

The adoption of these intelligent sensors has led to important benefits on the overall performance of plants. However, for maintenance purposes, the new available architectures are not yet effective, because the use of smart sensors can lead to the definition of completely autonomous plant areas and functions. A typical smart sensor has enough built-in computational power and I/O peripherals to control a set of actuators without the need to be interconnected. In this case, the data needed for maintenance operations may not be available to the whole maintenance system.

Another fact that has led to the delocalization is the evolution of fieldbus: in recent years, there has been a change in the traditional bus architecture due to the increasing adoption of internet protocol (IP)-based fieldbus such as EtherCAT or PROFINET (Felser, 2005; Ferrari et al., 2006). These standards preserve the legacy of the Ethernet protocol, including the star topology and the division into autonomous subnetworks. This involves a further breakdown of the information flow, which must be carefully prepared in advance in order to assure that all the needed information is provided to the maintenance management system.

This trend is further enhanced by the new standard for “wireless fieldbus.” In this system, known by the name of wireless sensor network (WSN; Spencer et al., 2004; Lewis, 2004), the communication features are delegated to autonomous devices that are often configured to transmit only relevant values (e.g., an alarm when a threshold is passed). This is done with the aim of preserving the battery life, which is the most critical component of a WSN system. Moreover, due to the unreliability of the ether as a transmission medium, with these systems the availability of data cannot be properly guaranteed. This uncertainty plays a great relevance in maintenance systems, mainly for diagnostic purposes.

For complex plants, a distributed approach helps improve the performance by reducing the computational time, especially in cases using predictive algorithms on several devices, which can have a significant impact on the overall performance of the system. In particular, it is possible to exploit the computational power of each device, even smart sensors, in order to decrease the time needed to estimate the health of the entire plant. Moreover, in complex plants, serious difficulties are often encountered in completing the full circle of condition-based maintenance (CBM). The gathered data are generally huge in their amount, because they come from assets dispersed over a large geographical area. The data may need to be integrated to provide useful information. Finally, the availability of an expert for converting data into useful information for maintenance is needed (Campos, 2009). Good experts are rare; therefore, even if a condition monitoring program is in operation, failures still occur, defeating the very purpose for which the investment was made (Prakash, 2006; Rao et al., 2003). This particular difficulty has been felt for a long time, and researchers have tried to overcome the problem through the application of artificial intelligence (AI) techniques, such as expert systems, artificial neural networks, fuzzy logic, etc. (Campos, 2009).

In order to develop a distributed logic in a maintenance system, a multi-agent system (MAS) approach can be considered. The agent technology has evolved from AI, more specifically from distributed artificial intelligence (DAI). Therefore, MAS is considered a subdiscipline of AI (Sycara, 1998). With MAS, it is possible to tackle the dualism between the local and the global system that characterizes a decentralized maintenance management system. The local capabilities can be easily implemented using local agents that operate only on a single machine without the need to perform continuous interactions with other parts of the system. Conversely, the capabilities of the overall system can be implemented using a set of agents that migrate between different machines, or by using a set of messages exchanged between different agents operating on the machines.

The functionalities of the maintenance system, which can be locally implemented, are mainly those strictly related to the single machine. This set can include the field data acquisition, diagnostics, and prognostic capabilities. The functionality that requires an interaction between different agents is related to human machine interfaces, maintenance scheduling, alarm reporting, or advanced prognostic systems that work on groups of devices or on the entire plant. During the development of this part of the system, special attention must be given to the choice of the way the different agents interact with each other. This is the key designing part of a MAS and greatly affects the performance of the system.

One of the best solutions for performing this task is to use a hybrid approach, each of which is optimized for a specific task. As an example, for data provider agents (DPAs) the client-server paradigm is a good choice. With this approach, each DPA acts as a server listening for any direct request from other client agents. A more complex approach must be conceived, for example, to implement a scheduler for maintenance intervention: for this task, several different approaches can be adopted. One of the possibilities is the implementation of a centralized maintenance planning system, with a global agent, that collects all the requests, performs scheduling, and activates the maintenance intervention. Conversely, a distributed approach can adopt a market-based paradigm where the scheduling of the maintenance operation is governed by an auction-based mechanism (Adhau et al., 2012). As a result, the choice of the right cost parameter is mandatory, because it affects the overall performance of the system. An unwise choice could lead to a relevant production breakdown where a critical machine could be waiting for repair because the maintenance staff has been assigned to work on other equipment.

In summary, a multi-agent approach theoretically allows the system to easily implement a distributed approach, with complex relationships, but only with the definition of some simple rules. This would result in an extreme flexibility and adaptability of the system that can also tackle important changes in the plant under control.

20.2 An Overview of the Applications

This section offers an overview of some examples of applications of a MAS-based methodology for maintenance purposes. However, keep in mind that this is a cutting-edge approach. Thus, most of these studies are still in a prototypical phase or are related to a very specific application.

20.2.1 IMS-TEMIIS

IMS-TEMIIS has been developed as a demonstration prototype of an integrated platform for remote and advance maintenance strategies (Iung, 2003). This platform simulates a typical chemical plant with several pumps and valves controlled by a heterogeneous set of control systems on several different fieldbus (Figure 20.1).

Inside this platform, a MAS-based diagnostic system has been implemented. This system is composed of a limited set of agents, as shown in Figure 20.2, which are loosely intercoupled.

In this application, all the main functionalities, such as diagnostics and monitoring, are implemented by specific agents that autonomously perform all the necessary operations. When an agent detects a failure, a more complex and cooperative decision process is activated, through the use of communicator and synchronizer agents, in order to share the knowledge between diagnostic agents (DAs) and isolate the degradated components of the system.

This auxiliary agent also aims to ensure a reliable and affordable communication between different machines. The exchange of messages required for this process are managed by a custom middleware, named kernel, that abstracts the maintenance system from the hardware and allows the integration of different devices and communication protocols.

20.2.4 Maintenance Schedule for a Bus Fleet

This implementation relates to the maintenance management of a bus fleet (Zhou et al., 2004). In this application, only maintenance scheduling operations have been implemented. Fault isolation and identification is not considered, and consequently the system does not carry out any prognostic functionality. The architecture of this implementation is shown in Figure 20.3.

This approach is centralized because there is only a single agent responsible for the management of the request for maintenance operation. This agent, linked with a bus status tracking system and a bus service planning system (a couple of external systems responsible for the health assessment of the buses and the planning activities of the vehicles) receives all the maintenance orders, ranks them according to their priority, and retains a global knowledge about the status of the overall system. It is linked to several mediator agents, each of them responsible for a specific maintenance activity. Each mediator agent keeps a list of bay agents, which are capable of the maintenance type that the mediator agent manages and selects for performing the task. Each bay agent is part of the system that manages and performs the physical maintenance operation. In order to keep track of the status of the various bay and maintenance orders, a database agent is responsible for storing the actual state of the overall system in order to help mediator agents choose the best bay to perform the task. The use of database agents allows individual agents to be aware of the overall status of the system in order to find the optimum solution. This choice was made in order to minimize the impact of maintenance operations on the functionality of the transport system, and also to maximize the uptime of the maintenance bay.

20.3 Application Details: AI2MS, a MAS Based on a Biomimetic Approach

As shown in the previous paragraph, MAS can be very useful in implementing an advanced maintenance system in those scenarios where the standard centralized maintenance system is not easily applicable.

One of these scenarios is the maintenance of oil transfer via pipelines or a waste water treatment facility. These systems are composed of a huge number of devices, often placed in inaccessible areas with a large distance between them, and often without the possibility of data connections. For these scenarios, it is mandatory that the maintenance system is distributed, and because the different devices inside the plants are activated very infrequently (typically once a day), the time needed to acquire enough data related to the behavior of a single device directly on the field is too high to enable the implementation of a sound traditional centralized system.

In this harsh scenario, an intelligent maintenance management system, named AI2MS (Artifical Immune Intelligent Management System) has been conceived by integrating a MAS-based architecture with the main features of an artificial immune system (AIS).

AISs are defined (Timmis et al., 2008) as “adaptive systems, inspired by theoretical immunology and observed immune functions, principles, and models, which are applied to problem solving.” Immune systems are a natural defense system against foreign harmful substances and microorganisms (such as viruses or bacteria) called pathogens. An immune system provides many levels of protection. A first natural barrier against invasion is the skin. After that, there is a physiological environment where the temperature and pH establishes hostile conditions for some pathogens. Then, there is the innate immune system, composed of specialized cells such as macrophages, which are able to identify and capture a limited set of microorganisms (Somayaji et al., 1997; Castro and Zuben, 1999).

An adaptive immune system is a more complex system, capable of identifying new threats, creating a response to them, and embodying this knowledge. An AIS tries to reproduce strategies of the adaptive immune system to acquire its features: distributability, adaptability, abnormality detection, and disposability (Somayaji et al., 1997).

An adaptive immune system is composed mainly of lymphocytes: the B and T cells. The recognition process is performed via a chemical affinity between antibodies and the molecular structures of the invaders, called antigens. Each B cell has one particular antibody, and through a mutation process, new kinds of antibodies can be generated (Castro and Zuben, 1999; Aickelin and Dasgupta, 2005; Dasgupta, 2006).

The clonal selection reproduces the B cell, which is capable of identifying an antigen. The B cells could be differentiated into plasma cells that accelerate the immune response, or into memory cells that remain longer in the organism and are responsible for the acquired immunity (learning processing). Lymphocytes produced by the mutation process can identify cells of the organism (“self”) as invader cells, and this full immune response can result in damage to the host organism. Negative selection is a mechanism employed to avoid this problem. It occurs within the thymus. T cells are exposed to “self” molecular structures, and those that react against it are eliminated. The remaining T cells act as suppressors for B cells, avoiding the recognition of a “self” structure as an invader (Castro and Zuben, 1999; Dasgupta and Forrest, 1999; Aickelin and Dasgupta, 2005; Dasgupta, 2006).

The immune network theory proposed by Jerne (1973) is based on a mechanism that performs the recognition task via a network of interconnected B and T cells (Mizessyn and Ishida, 1993). These cells both stimulate and suppress each other in certain ways that lead to the stabilization of the network. Hence, the recognition task is performed at the system level, not as an individual task (Aickelin and Dasgupta, 2005, p. 187; Dasgupta, 2006; Ishiguro et al., 1994).

The developments within AIS are based on these three immunological theories with different approaches. Clonal selection and immune networks are mainly used as learning and memory mechanisms, and the negative selection principle is applied for the generation of detectors that are capable of classifying changes in self (Timmis et al., 2008).

From the preceding description, it turns out that an AIS is composed of several systems that must relate to each other in order to provide a prognostic functionality. All these functions are implemented using agents, some of which operate only in a single machine (local agents), whereas others can migrate between more machines in order to preserve a global behavior of the system.

In Figure 20.4, an overview of the AI2MS architecture is provided.

The different types of agents that constitute the architecture can be grouped in more clusters: DPAs, DAs, PAs, and service agents (ServAs). The remainder of this section provides a detailed description of their main role within the AI2MS system.

20.3.1 Data Provider Agents

DPAs are involved in the provision of data from the field. There are two different types of agents: sensor agents (SAs) and sensor diagnostic agents (SDAs) (Figure 20.5). SAs are local agents that are located inside the machine and are responsible for the provisioning of field data to other agents. Each agent of this type handles a single sensor, so in a typical application there are many instances of these agents that operate at the same time, providing information to the other agents or to another system. SDAs are local agents responsible for the evaluation of the data provided by SAs. The main task of a SDA is to check the correct operation of a sensor and, in case of degradation, fix the problem. The SDA also assesses the fidelity of the information provided. If it is difficult to fix the problem generated by a sensor in a short time, these data could still be used with the necessary safety margins.

20.3.2 Diagnostic Agents

DAs are the core of the AI2MS. This group contains all the agents responsible for the prediction of the failure of a single machine or of the entire system (Figure 20.6).

20.3.2.1 Fault detection agents

Fault detection agents (FDA) are agents responsible for detecting a specific failure mode. These agents are the equivalent of the lymphocytes T-helper and B-memory cells of a biological immune system and, like these, are very specialized. In a typical application, inside each machine there is a large number of agents of this typology. Each of them is able to detect, with the most appropriate technique, a specific failure mode. To perform this task, each agent needs to interact with the various SAs to acquire the current state of the system and, by combining this information with the history of the system (stored inside the agent), it is able to detect a failure in the system.

According to the AIS approach, in order to implement a more robust system this kind of agent is generated by using clonal selection techniques. This implies that for each failure mode a set of agents is generated with slight changes in the detection parameters. After a training phase, only the most efficient one will be kept active inside the system.

20.3.2.2 New fault detection agents

The clonal selection methodology used in FDAs enables the detection of well-known failure modes. However, whenever a new failure mode occurs, their role is not effective. For this reason, in each device of the plant a specific agent is considered to detect an unknown malfunctioning. This agent, namely the new fault detection agents (NFDAs), is based on a negative selection methodology where a set of good state signatures of the device are used to train the agent, making it able to detect unknown failure modes of the system. This agent is analogous to the biological innate immune system. Given the non-specificity of this kind of agent, in each machine only one NFDA operates. As will be explained afterward, the implementation of this functionality is the core of the evolution of the overall AI2MS.

20.3.2.3 Cooperative detection agents

The combined use of FDAs and NFDAs makes the system able to detect almost all the failures occurring inside a single machine. However, often a failure can result in the abnormal behavior of several machines or, especially for this type of plant, a failure can occur in a part of the system not under control. Let’s consider, for example, a leakage in a pipeline. This kind of failure cannot be detected by local agents because each device inside the plant works correctly, so in order to detect this kind of malfunctioning, a cooperative detection agent (CDA) was conceived. This agent is a network agent that is capable of migrating between different machines and acquiring the necessary information to detect this kind of failure.

20.3.3 Prognostic Agents

PAs group the agents responsible to estimate the health assessment of a single machine or of the entire plant (Figure 20.7). They can play different roles.

20.3.3.1 Device health assessment agent

A device health assessment agent (DHAA) is a local agent that runs in a single copy in each machine to estimate its remaining useful life. This agent estimates the residual health of the system using advanced soft computing techniques, such as neural networks (DePold and Gass, 1999) or advanced pattern recognition (Lee et al., 2006) based on the data provided by the SAs and failure detection agents.

20.3.3.2 Plant health assessment agent

A plant health assessment agent (PHAA) is similar to a DHAA, acting globally through the network to estimate the health conditions of the entire plant. This agent migrates between different machines to acquire the needed data and also keeps track of the topology of the plant in order to identify redundancies and bottlenecks and thus perform a more accurate estimation.

20.3.4 Service Agents

ServAs regroup all the accessory functionalities that are not strictly related to maintenance operations. This group is responsible for the evolution of the entire system. Like the biological immune system, the AI2MS needs to evolve in order to be able to detect new kinds of failures, improving the overall reliability of the system and consequently increasing the performance of the overall system.

20.3.4.1 Update agents

The update agents (UAs) are the agents responsible for the sharing of the knowledge between the different machines of the plants (Figure 20.8). There are two different kinds of UA. The update training agents (UTAs) share the raw data used for the training of the DA and PA; the failure mode update agents (FMUA) are responsible for the migration of the failure detection agents between different machines of the same type. All these agents are network agents that migrate between the machines, acquire new information, validate it, and share only that which is useful.

20.3.4.2 Evolution agents

Evolution agents (EAs) are the core of the entire AI2MS system. The role of these agents is the management of the evolution process that leads to the definition of new FDAs (Figure 20.9). This is a global agent which performs a comparison between the results of the NFDAs and CDAs of each machine of the plant and, in cooperation with maintenance personnel, evaluates whether the NFDAs have really acknowledged the fault mode and promoted it in FDAs. This is similar to the evolution of T-lymphocytes into the thymus gland. When a new FDA has been generated, UAs share the new detector with the other machines.

20.4 Benefits and Assessment

In order to validate its effectiveness, the AI2MS was implemented in a lab environment in order to analyze the performance of the core of the system. In particular, the system was applied on a specific test bench that is able to simulate a limited set of typical failure modes that can occur in an oil transfer system.

This system is based on an electric valve actuator that is able to reproduce two different kinds of failures: gear damage (by replacing the good gears with a degraded one) and a problem with the actuator (by increasing the resistive torque).

The overall system was developed using JADE (Bellifemine et al., 1999), a Java middleware specific to the implementation of distributed multi-agent architectures.

One of the most important benefits of JADE is that the developed applications are compliant with the FIPA (Bellifemine et al., 1999) standard. The use of this standard allows the system to easily interact with other components of the control system of the plant based on a multi-agent methodology. This feature is very important for the integration of the diagnostic system with the other components of the plants. In order to further strengthen it, the interaction of different agents is based on an acknowledged ontology in order to standardize the different messages between agents. With the use of the ontology and the support of a yellow page service for agent discovery, the integration with other systems can be easily reached with limited effort. An overview of the agent communication is shown in Figure 20.10.

The test bench (Figure 20.11) is equipped with vibrating sensors, positioned in the bearing of the motor shaft, that are managed by a SA to provide information to a set of DAs. This group is composed of 50 agents, and created using a clonal selection technique. Starting from the detector of three failure modes, the size of the group is chosen using a trial-and-error approach in order to reach a trade-off between computational power and the precision of the results.

Each DA uses a wavelet of packet energy to analyze the vibration data in order to extract a performance signature and estimate the health of the device.

Table 20.1 presents the results of the first test, which is instrumental in the evaluation of the detection capabilities of the system. For performing this experiment, we acquired 150 signatures of the system in three health conditions: normal condition, worn gear, and damaged gear (50 for each state). A set of 20 cycles of each operational condition was randomly chosen to train the AI2MS, and the remaining 90 signals (30 of each type) were used to test the system.

The AI2MS was composed by two different types of detection agents: Fault 1, responsible for the detection of the wear of the gear, and Fault 2, responsible for the detection of the damaged gear. There is also a NFDA, Fault 3, which is responsible for detecting an unknown failure mode (simulated by a random increase in resistive torque).

The proposed diagnosing algorithm has produced 1.1% false positives, 3.3% false negatives, and 11% inconclusive results. The performance of the system is, at the current state, lower with respect to a typical solution (Laurentys et al., 2010). However, it is expected that, with the implementation of the Service Agents, the performance will increase to a value comparable with other solutions. This is because the data set used for the test is very small, comparable to a month of operations on a single device. With the increase in the data set provided by the UTAs, the performance will likely improve.

Further steps in the testing activities will be:

• Implementing a feedback mechanism to the EAs, thus providing learning capabilities to the system.

• Modeling and implementing the collaborative DA to validate the collaborative approach proposed.

• Analysis of the intensity and quality of the message exchange, to evaluate the requirements of network support and the possibility of integration with the plant control network.

20.5 Discussion

The use of AIS methodology leads to remarkable benefits in terms of flexibility, adaptability, and performance optimization when used to implement a maintenance management system.

The overview of different maintenance needs, solved with a multi-agent implementation shown in this chapter, clearly underlines the adaptability of this methodology to solve all the tasks required for a modern intelligent maintenance system, particularly regarding highly critical plants with rigid constraints.

The practical implications arising from the adoption of these technologies are quite evident: the opportunity to implement CBM approaches in complex plants with a large number of delocalized assets allows the prediction of unexpected failures which, in turn, impacts both economical performance and safety.

For example, the specific mode of operation of the artificial immune algorithms allows for an easier implementation, by using agents, and thus enabling the creation of a maintenance system able to react to different known or unknown failure modes.

In fact, this methodology allows the use of a combination of different failure identification techniques that make the system more robust with respect to a classical system.

In particular, clonal selection techniques increase the robustness in comparison to a suboptimal tuning of detection algorithms, due to the large number of detector agents—slightly different from each other—that are running together, while the negative selection methodology allows good detection of new kinds of failure modes.

The use of multi-agent based solutions also allows the diagnostic system to be very robust with respect to internal failures. Due to the combination of autonomy and distributability, innate in this methodology, a failure in a part of the diagnostic system will not affect other components that operate in a normal way.

Finally, another strong benefit of MAS is related to the scalability of the system. Unlike traditional approaches, the efforts needed to implement the maintenance system does not depend on the complexity of the plants. In a MAS, most of the effort is dedicated to designing the functionality and relations between the different types of agents. The size of the plants usually affects only the number of agents used for diagnostic purposes. This means that this methodology has an excellent scalability. However, the difficulty in the system design makes the adoption of these systems less suitable in the case of simple plants where the preliminary design work could be much higher than the expected achievable benefits.

The design flexibility of such applications also leads to some important disadvantages that need to be taken into account when designing a MAS. The main problem with this kind of approach is the lack of predictability of the behavior of the overall system (Zuccolotto et al., 2013). This is due to the many different variables and conditions, which is very hard to estimate in advance. Another disadvantage of this approach is the increase in the computational power required, which results in the need for more expensive devices despite the possiblity of easily splitting the computational load into multiple devices (Carabelea et al., 2003).

Finally, another relevant criticality in the use of MASs is the inability to ensure the achievement of the maximum performance (reach the global optimum solution) in absolute terms. With a centralized approach, all data are accessible, so it is quite easy to check if the solution provided by the system reaches a global or local optimum. With a distributed approach, this is more complicated (Caridi and Cavalieri, 2004; Dorer and Calisti, 2005).

20.6 Conclusion

The adoption of MASs to implement a computerized maintenance management system is definitely a good solution whenever the system requires good adaptability or the system is so complicated that it is very hard to define a set of rules for the management of the system with a standard approach.

Compared to a standard centralized maintenance system, this type of approach can lead to substantial benefits such as:

• Distributed computing: It avoids the saturation of processing power during the analysis of large amounts of data provided by large plants.

• On-site computing: It avoids the requirement of a long-range communication system to centralize all the information.

• Social behavior: The particularity of the interaction between the different agents enables the implementation of the complex behavior that can be split into the actions and communication of several agents instead of being implemented in a monolithical algorithm. In this way, it is possible to simulate a social behavior; something that is not possible with a monolithical algorithm.

From a practical point of view, this application offers better reliability from diagnostics systems. This is particularly important in very complex and automated plants (i.e., process industry) in which the failure could lead not only to a great impact in terms of production loss, but also in terms of safety.

However, it is important to keep in mind the limits of the multi-agent approach, mainly regarding the difficulty in predicting the behavior of a maintenance system governed by a MAS. A wise design of the governance rule of the different agents, and of the communication between them, is necessary in order to develop a stable system that allows all the needed functionalities. Also, an exhaustive simulation approach must be used for the validation of the system in order to verify the proper functioning of the system.

As a matter of fact, in these kinds of systems the behavior of the overall system is mainly related not only to the behavior of single agents but also to the interaction between the different agents. The validation through simulation is the only possible way to have a better predictability of the proper functionality of the diagnostic system.