1. Context This thesis forms part of an extended study funded by FonCSI (Fondation pour une Culture de Sécurité Industrielle) about learning systems of major hazard companies. All French industrial sites running a risky activity – e.g. petrochemicals, steel making plants – must put a Safety Management System in place, according to the Seveso II EC Directive . Amongst other things, it requires implementing a Feedback of Experience process (Retour d’Expériences – Rex - in French). The objective of the Rex is to report and treat any critical events in order to prevent and avoid accident recurrences; it is composed of the following stages: detect, analyze, learn and share the lessons drawn from accident analyses. Despite the use of Rex, many accidents still occur. Academics and industrials note that there are limitations to the Rex process (Gilbert & Bourdeaux 1999, Dien & Dechy 2007); it seems to be mainly reactive, and only to identify technical and procedural causes of accidents, which consequently limit the opportunities for learning – at the organizational level – for the sites. Researchers put the emphasis on means to promote and facilitate proactive safety management. Weak signals, defined as accident precursors, were perceived as a promising issue to meet such an objective. However, bodies of literature on weak signals (e.g. Caron-Fasan 2001, Brindejonc & Llory 2005 and Pariès & Bieder 2006) reveal the difficulty to recognize and treat them. They show that weak signals are fragmented and complex data which require a process of sense making to be carried out by the organization. We hypothesize that organizational factors block the treatment of these signals and maintain conditions in a favorable state, leaving these signals to persist in a weak state. 2. Objectives and methodology The objective of this PhD research is to explore the research literature on Rex and weak signals, explore the theoretical underpinning and understand the functioning of Rex system and its strengths and weaknesses in practice. It aims at defining weak signals and identifying blocking factors in order to address theoretical and practical propositions to respond to this challenging issue. We conducted two in-depth case studies using the approach of action-research to provide a rich picture of how Rex in practice functions: the first in a petrochemical plant, EDIA, and the second in a steel plant, ACIO. We carried out a sociological diagnosis at both sites in order to explore their maturity concerning weak signals management. From the cases we induce and hypothesize important filters and barriers to information transfer, comprehension and learning. 3. Results Accident analyses and a study of normal functioning revealed that the EDIA and ACIO sites identified much data relating to safety, which were processed by the formal tools implemented and used on site, but discarded other data. We categorized the latter data in the following way: - Latent and repetitive technical failures - Work coordination weaknesses - Underestimation of risk. We defined these data as weak signals because, with hindsight (during accident analysis), they revealed obvious and relevant information for prevention. Although identified as important data linked to the event, revealing organizational dysfunctions for safety, they were not perceived in advance as threats to safety and not included in the formal procedures of accident analysis. Neither site took the opportunity to learn from them; that is to say to deepen the analysis of the events (on the human and organizational dimensions) and to revise their own methods of investigation. The case studies led us to identify the roots of weak signals at two levels. - First level; safety management activities: For each of the three categories of weak signals, a normal activity was defined as the mark or root of that specific category. Thus, the root of technical weak signals was found in the system of management of technical failures, the root of the “work coordination” weak signals in the relative inadequacy of the bureaucratic organization for the coordination of contractors’ work, and the root of the “risk perception” weak signals in limits to the proactive and reactive analyses of risks. Within each of these three activities, we observed a process of information filtering which tended to distinguish between two types of data. On the one hand, we identified known or simple data, that is to say, data for which the detection and analysis tools and organizational mechanisms for managing and monitoring were predefined. These data get a simplified treatment, that is to say that the analysis and actions put in place to remedy the problems identified tend to be limited to technical and/or procedural aspects, with only marginal treatment of the organizational factor. On the other hand, more complex, unexpected data or weak signals, unlike the former, cannot enter, or even are discarded from the information management system, because they do not fit within the classifications or predefined treatment categories. The fragmentation of safety management and the process of filtering information were defined as two factors blocking the recognition and treatment of the three categories of weak signals. -Second level: organizational structure We showed that fragmentation and information filtering, blocking the detection and treatment of weak signals were, in turn, rooted in the organizational structure. Organizational fragmentation was shown to result in a distinction or de-synchronisation between the staff and line (operational) organization of work, particularly in relation to maintenance. Where the resources made available to operators to perform their work (procedures, competence, information, technical and/or organizational barriers) are incomplete or inadequate, this encourages them to violate or bypass the rules and create informal practices that suit the pressures of work and the constraints of business. These practices, which may threaten safety, nevertheless keep the system functioning and may, in some circumstances, be tolerated and normalized by management. The centralization and complexity of prescribed rules, leading to necessary adjustments, were shown to be a fertile seedbed for these informal practices, and to be implicated in the persistent weakness of the signals related to coordination and underestimation of risk. By analyzing the Rex systems in this way, we rejected the theoretical formulation of weak signals as inherently weak and reformulated them as signals made or kept weak by the way they are handled by the learning systems of the company and their ability to detect, code, transmit and handle signals. Finally, we addressed some recommendations. By comparing the EDIA and ACIO safety managements with the Delft model (Hale, 2003), we showed that both sites had gaps in operation of some specified principles of good practice. First of all, both sites did not manage to recognize and identify all the risks and to operationalize them in accident scenarios which could include the factors which potentially are weak signals. Then, there seems to be a weakness in the resources provided to ensure a continual assessment and reappraisal of the efficiency of the whole range of safety barriers necessary to maintain a good level of safety. Their safety management systems do not appear to be efficient in recognizing and treating such weak signals of accidents. Finally, both safety management systems proved to be not effective at organizational double-loop learning. The case studies showed that the EDIA and ACIO sites do not learn effectively from weak signals, because they still discard useful and important signals from their formal tools of accident analysis, whereas we believe that these signals have a huge potential for learning, for accident causes and for understanding and revision of the Rex process.