Abstract— In this paper, we introduce a procedure for diagnosis and treatment of faults in productive systems, i.e., a supervision strategy that considers not only the normal behavior of system’s components, but also abnormal (faulty) conditions of them. The present approach uses Bayesian networks for the diagnosis and decision-making purposes, and Petri net for the synthesis, modeling and control purposes. The integration of these techniques guarantees the specified functionality of the system. Special emphasis is laid on methodological issues and industrial systems, where a hierarchical structure can be adopted. It is presented, as a case study, the material entry system of a continuous pickling line process of a steel industry. I. INTRODUCTION ONCERNING the new paradigms of customer’s dependent manufacturing with high quality, lower costs, and minimal production times, faults are events that must be avoided since they compromise the global system’s performance. Unfortunately, even in systems with preventive maintenance programs, faults are inherent phenomena on productive systems and, by nature, difficult to predict. In this context, this work introduces a procedure for diagnosis and treatment of faults in discrete productive systems based on Bayesian networks and Petri net. We aim for the integration of these two techniques in order to improve systems’ reliability and also to reduce undesirable plant interruptions. Productive systems [1], specially from a supervision and control point of view, can be classified as a type of discrete event system (DES), that in the last years have been the focus of many researches. Their characteristics have motivated researches to develop new control techniques in order to supervisor their behavior and, among several proposals, the Petri net approach has been considered a very useful tool to model and specify control strategies (for e.g. [2]). Productive systems consist of a physical layer and an information system that operate accordingly with a predefined logic. The physical layer consists of a set of devices, operators, equipments, etc. that Manuscript received May 18, 2007. This work was supported in part by the Brazilian governmental agencies CNPq, CAPES and FAPESP Roy Andres Gomez Morales is with Escola Politecnica da Universidade de Sao Paulo, Sao Paulo, SP, Brazil (e-mail: roy.morales@poli.usp.br). Jose Isidro Garcia Melo is with Escola Politecnica da Universidade de Sao Paulo, Sao Paulo, SP, Brazil, on leave from University del Valle, Cali, Colombia (e-mail: joisigar@poli.usp.br). Paulo Eigi Miyagi is with Escola Politecnica da Universidade de Sao Paulo, Sao Paulo, SP, Brazil (e-mail: pemiyagi@usp.br). realize services and/or perform material transformation. The information system consists of computational resources responsible for the data manipulation and for the control and supervision tasks. The main purpose is to assure the correct behavior of the physical components. Faults are randomly events that difficult such task. Fault occurrences in productive systems generally imply lower quality of the final products, partial breakdown of production, off-limit time deliver, economical lost and sometimes human injuries. Because of these reasons, there is the need for solutions not limited to normal conditions of the processes, but also solutions considering the detection, diagnosis and treatment of faults. Many works have concerned with fault detection in DES [3]–[5]. In [3] and [4] for instance, a Petri net approach is introduced for fault detection purposes. In [5] a finite automata approach is presented for fault diagnosis in DES. Petri net and finite automata approaches confirm their effectiveness to model and analyze the processes, however, they suffer of exponentially state space growing referring to the complexity of the system. This complicates the fault detection. On the other hand, Bayesian networks have been proved as a suitable tool in modeling of detection and diagnosis of systems [6]–[8]. In [9] for example a comparison between rule-based expert systems and probabilistic methods is presented proving that, for diagnostic purposes, Bayesian networks offers a more suitable technique on dealing with uncertainties. This work aims to explore the advantages derived from previous studies. That is, in this work: i. the diagnoser is modeled in a different level of the control level and only the faults are considered for the representation, ii. a representative of cognitive human expert’s knowledge is considered for the synthesis of the diagnoser model, iii. both, qualitative and quantitative information of the systems are considered for the synthesis of the diagnoser model, iv. abductive, inductive and nonmonotonic reasoning are also considered to improve the diagnosability of the system. This paper is organized as follows. In Section II we present the general concepts of Petri net and Bayesian Diagnosis and Treatment of Faults in Productive Systems based on Bayesian Networks and Petri Net Roy A. Gómez Morales, Jose I. Garcia Melo, and Paulo E. Miyagi, Senior Member, IEEE C