In the area of real-time production monitoring technique, Guo et al. proposed an RFID-based smart decision support system architecture for production monitoring and scheduling in a distributed manufacturing environment. The RFID, cloud technology, and intelligent techniques are used in the architecture [9]. Chai et al. presented a cooperative manufacturing execution systems; it uses real-time information to realize lean production [10]. By applying the concept of cloud computing into manufacturing field, Wang and Xu provided an interoperable production mode [11]. Wang et al. developed a framework, that is, Wise-ShopFloor (web-based-integrated sensor-driven e-ShopFloor), for real-time monitoring and remote control of networked CNC machines [12]. Chen et al. proposed an RFID-based enterprise application integration framework for real-time management of dynamic manufacturing processes [13]. Chongwatpol and Sharda proposed a creative information visibility-based scheduling (VBS) rule that utilized information generated from the real-time traceability systems for tracking work in processes (WIPs), parts and components, and raw materials to adjust production schedules [14]. Arkan and Van proposed an RFID-based real-time location system (RTLS) solution for obtaining multiitem WIP visibility within a company [15]. Bevilacqua et al. presented a case study for implementation of an RFID system in a furniture industry involved in the fashion sector; the case company is an Italian enterprise which is a leader in the furniture industry [16].

In the production KPIs analysis aspect, Georgiadis and Michaloudis merged discrete event simulation and continuous dynamics in a unified hybrid simulation environment. By coordinating the simulations run in real time, the results can be associated with information from the shop-floor frontline [17]. Zang et al. applied an event-processing method for management systems using RFID, where the framework, information structures, optimization manners, and algorithms are discussed in detail [18]. Considering the difference businesses in different companies, Huang et al. presented an agent-based workflow management method [19] and Zhang et al. proposed a smart objects management system for RFID-enabled real-time reconfigurable manufacturing [20]. Kumaraguru et al. proposed an approach to integrate real-time analytics with continuous performance management, which can form the base for further research on fulfilling likely interoperability issues and essential standardization efforts to support improvement of a system [21]. Blaha et al. designed a paradigm that enables real-time monitoring using cognitive models and its implementation was demonstrated with a fatigue-sensitive task [22]. Li proposed an aggregation method by overlapping decomposition approach to approximate the production rate for systems with rework loops; the systems were decomposed into four serial production lines, with machines modified to accommodate the interactions with other machines and buffers [23]. Colledani and Tolio designed a new approximate analytical method to estimate the quality and productivity performance measures of asynchronous production lines, thus the associations between the quality control and system production logistics performance can be captured [24]. Lao and Liu integrated data envelopment analysis (DEA) and geographic information systems (GIS) to evaluate the performance of bus lines within a public transit system, considering both the operations and operational environment [25]. Lin et al. proposed a three-phase method to measure the performance of a highly value-added footwear manufacturing system considering reworking actions, in which the system consists of multiple production lines [26]. Using RFID-enabled shop-floor production data, Zhong et al. proposed a data mining approach (DT-based model) to estimate realistic standard operation time and unknown dispatching rules [27].

To analysis production anomalies, two main steps need to be taken, that is, the anomaly extraction and anomaly cause diagnosis. In the area of anomaly extraction, Evans et al. presented a distinct experience-based decision support system, which uses factual information of historical decision to compute the confidence factors [35]. Zarei et al. presented an intelligent method based on artificial neural networks (ANNs) to identify bearing defects of induction motors [36]. Cabral presented a PN approach for online diagnosis of DESs modeled by finite state automata; the method is based on the construction of a Petri net diagnoser (PND) which is constructed in polynomial time and requires less memory than other methods proposed in the literature [37]. Cabasino et al. proposed a method for online diagnosis of DESs based on labeled PNs [38]. He et al. proposed a DT learning-based model for bivariate process mean shift monitoring and fault identification [39]. Maier et al. presented new model-based approaches for testing and diagnosis of automation systems; the models are not created manually but learned automatically by monitoring the plant behavior [40]. Similarly, Faltinski et al. presented a new model-based approach for the prediction of energy consumption in production plants in order to detect anomalies. Hybrid timed automaton models of the supervised production plant are generated and executed in parallel to the system so that the anomalies can be detected automatically [41]. Yin et al. considered the real-time implementation of fault-tolerant control systems with performance optimization [42].

Up to now, comprehensive research on the real-time production PAM has been done. However, less research effort is paid to the following issues.

The goal of this module is to construct a smart manufacturing environment based on IoT technologies, thus the information bridge between physical manufacturing systems and upper-level information systems can be established. The IoT technologies, such as RFID, wireless, etc., are applied to the traditional manufacturing things based on the information requirements. Besides, the optimal algorithms are used to decide the configuration pattern, so that a smart manufacturing environment can be established in an optimal manner. At last, the amounts of manufacturing data can be timely obtained by upper-level applications. More details can be found in Chapter 3.

This module aims to model the dynamic behavior of the manufacturing system and process the large number of raw manufacturing data captured to CrEs of KPPA. As shown in the middle of Fig. 7.1, multilevel event model are used to analyze real-time PP. Four levels of events are established in this model, namely primitive events (PEs), basic events (BEs), complex events (CEs), and CrEs. The extracting processes are fulfilled by using HTCPN technologies, which will obtain the up-level events by continuously reading and integrating the primitive data. More details will be presented in Sections 7.4 and 7.5.

Evaluating the real-time key PP condition is an important step in improving production efficiency. After the CrEs are obtained, the PPs are going to be evaluated to find the anomalies. First, the quantitative information of PP are obtained from the lower-level modules, and the information of the related factors are also called. Second, each kind of key PP are assessed to find the anomalies, respectively. The classic decision tree classifier, namely, C4.5, is used for the construction of anomaly extraction rules due to its ability to analyze continuous attributes. At last, the corresponding anomaly cause analysis rules are called to acquire the reasons as soon as an anomaly is extracted. Since most of the captured information of manufacturing resources has fuzzy attributes, the Fuzzy Interactive Dichotomizer 3 (Fuzzy-ID3) algorithm is used to obtain anomalies cause diagnosis rules. More details are given in Section 7.6.

HTCPN is a graphical modeling language with a well-defined semantics, which has been proved to be a suitable tool to model dynamic manufacturing system. HTCPN, as a normal extension of the classic PN, have three main advantages for production modeling: (1) HTCPN can decrease the dimension of traditional PN in a great extent; (2) HTCPN can model deterministic firing durations, which helps to obtain exact performance estimates; and (3) HTCPN can use a unified net to model similar components/subsystems, and the nets can be involved repeatedly when modeling a complex system.

To model the multilevel event, mainly five steps are considered to build an HTCPN.

As shown in Fig. 7.3, the connection between the established HTCPN and manufacturing resources is fulfilled by a middleware. On one hand, the manufacturing resources are monitored by the multiple sensors, and the real-time information are captured and uploaded to the middleware on time. On the other hand, the marking of the HTCPN alters based on the dynamic information acquired from the middleware. Thus, the tokens in the HTCPNs can be connected with the manufacturing resources. Once the status of a manufacturing resource alters, the color of the token will be altered immediately.

After the HTCPN model is established and the tokens are linked with the manufacturing resources, the model can change its markings according to the real-time condition. Thus, the performance of the PN can reflect the real-time PP. In order to obtain the PP easily, three actions are used to analyze the manufacturing processes.

In order to describe a new PP analysis problem simply and directly, the newly obtained information is represented as a new case (C).

A new production anomaly analysis event is evaluated based on anomaly analysis rules. It is significant to feedback the exception analysis, so that experiences are learned by knowledge mined from historical cases. Thereby, the possibility of the success of classification for new anomalies can be increased.

To evaluate production anomaly, multiple decision variables are involved. Although performance attributes are always with crisp values, anomaly attributes cover both crisp and fuzzy values owing to the ambiguity or uncertainty during information capturing stage. Variables with crisp values are the attributes that are calculated according to a traditional set, that is, the membership of any value for the attributes is either 1 or 0. Variable with a fuzzy value defines their value with the membership that varied from 0 to 1, which are always determined by the membership function.

Mainly three steps of rule induction are needed in DT learning. Step 1: create a large DT from historical cases according to feature selection approaches; Step 2: prune the branches and nodes that have little statistical influence on the tree; and Step 3: improve the understandability of obtained tree. Since the DT technologies have been widely discussed, we give the feature selection approaches briefly.

After a real-time PP case is acquired, the rules correlated to the information can be used to assess the PP condition. If one anomaly event is detected, the anomaly label (e.g., Label L1) will be affixed to it, thus the following anomaly cause diagnosis process can easily identify the anomaly.