Which of the following methodology is concerned with noticing weak signals

The concept of weak signals was first proposed by Ansoff & McDonnell for strategic planning and management, as “imprecise early indications about impending, impactful events” (Ansoff and McDonnell, 1990). Based on the definition, researchers emphasized that weak signals have crucial meanings for the future and provide a threat or opportunity for a business, but they are not mature at the time when they appear (Coffman, 1997, Holopainen and Toivonen, 2012). The concept has been applied in multiple domains for anticipating the future, such as technology foresight (Tabatabaei, 2011), defense (Koivisto et al., 2016), and natural disasters (Shelly et al., 2007). Since the definition can be different depending on contexts, the study summarized a few definitions and characteristics of weak signals in the domain of safety.

Vaughan (1997) defined weak signals as ambiguous information not showing clear threats to safety. The definition was expanded and specified later and referred a weak signal as a technical anomaly that “has no clear and direct connection to a potential danger, or that only occurs once and does not seem likely to occur again” (Vaughan, 2002, Vaughan, 2004). Similarly, Weick et al. (1999) defined weak signals as technical anomalies that are observed during operations. However, instead of only treating weak signals as technical issues, Guillaume (2011) pointed out weak signals could be re-occurring technical failures or deficiencies at upper management levels but they are ambiguous for anticipating future events. Additionally, Hollnagel (2004) proposed a definition of weak signals and revealed insights into why weak signals are ambiguous. Within a complex system, multiple functions are involved ranging from technological functions, human functions, to organization functions. Each function has its own performance variabilities. A weak signal can be a performance variability of one function while relatively the rest are noise. A weak signal is ambiguous since it does not cause detectable effects until it combines with noise and its constitution to a hazard is only amplified by the combination (Hollnagel, 2004).

“Weak signal” is not common terminology in the domain of safety. There are multiple terminologies resembling a similar spirit, such as precursors, early warning signals/signs and leading indicators. The terminologies bring more confusions to understand weak signals. For example, precursors have various definitions (Carroll, 2004, Körvers, 2004, Kunreuther et al., 2004, Saleh et al., 2013), and the one defined by Körvers (2004) is the same as the technical weak signal that was defined by Guillaume (2011). “Early warning signal” is treated as an interchangeable terminology of “weak signal”, which is an early warning of a precursor (Luyk, 2011). Additionally, a leading indicator is another form of an early warning to evaluate overall safety or risk in a system but is not interchangeable with a weak signal (Øien et al., 2011). To utilize leading indicators directly as signals for incident prediction is not feasible since the correlation between most leading indicators and event realization is unknown, vague, and unpredictable (Körvers, 2004, Luyk, 2011, Øien et al., 2011).

Even though weak signals are defined differently, the characteristics of weak signals that are addressed in literature are consistent. First, weak signals emerge for a long time before incidents and indicate potential risks of impactful events. They can be too early to be precise, but provide opportunities to respond (Guillaume, 2011, Luyk, 2011). Second, it is difficult to recognize weak signals since they cannot be interpreted in isolation and the impact of a weak signal is not noticeable until it combines with other signals (Brizon and Wybo, 2009, Hollnagel, 2004, Luyk, 2011). Besides, due to existence of noise, weak signals only can be interpreted after appropriate filtering and processing (Brizon and Wybo, 2009, Guillaume, 2011). Without tools to help people interpret weak signals, the interpretation process becomes subjective to individuals’ experience, knowledge, and willingness. When too much information is recorded in a system, weak signals are rarely picked up and connected (Guillaume, 2011, Vaughan, 1997, Vaughan, 2002). A few studies (Brizon and Wybo, 2009, Guillaume, 2011, Körvers, 2004, Luyk, 2011) were conducted aiming to help industries improve abilities to identify weak signals, but the studies mainly focused on improving organizational management qualitatively, instead of developing techniques to identify weak signals and predict incidents proactively.

On the other hand, in process control of continuous processes much work has been done on what is called Fault Detection & Diagnosis (FDD) to provide early warnings of process deviations. A few recent studies of FDD are by (Cheng et al., 2021, Fazai et al., 2019, Nhat et al., 2020), and much more studies are summarized in several review papers by (Alauddin et al., 2018, Arunthavanathan et al., 2021, Nor et al., 2019, Venkatasubramanian et al., 2003a, Venkatasubramanian et al., 2003b, Venkatasubramanian et al., 2003c). FDD aims to automatically recognize deviations of monitored process variables by filtering noise disturbances to provide early warnings through safety instrumentation system for abnormal situation management. Such early warnings are necessary to alert deviations that have already occur during operation periods, however, available time for interventions can be limited resulting in uncontrollable consequences. For the perspective of preventive safety, this work is beyond the scope of FDD to identify weak signals such as organization performance, equipment maintenance and human factors. Identifying such weak signals is critical for monitoring safety performances of organizations based on risks that are resulted from their interactions, and guiding organizations to make appropriate corrective actions for eliminating or mitigating degradations in safety management.

Since weak signals are defined precisely in the domain of safety, a formal definition of weak signals needs to be proposed. Among the existing definitions that were mentioned in the previous section, the definition proposed by Hollnagel (2004) provides the most clarity through emphasizing multiple sources of weak signals and indicating that a weak signal alone cannot cause impactful events. However, a weak signal and noise were defined as relative terms depending on focuses. The difference between weak signals and noise needs clarifications. Additionally, Hollnagel (2004) states a detectable hazard is caused by the interaction between a weak signal and its noise. According to the Cambridge dictionary, a signal is something that “gives a message or a warning”. It is more sensible by treating anything representing a sign of a hazard as a signal, instead of noise. In view of the definition by Hollnagel (2004) and the understandings of weak signals in other literature, this study proposes new definitions of weak signals and noise:

Weak signals are performance variabilities of technological, organizational, or human functions whose interactions combine clues or signs giving rise to early prediction of a future unexpected event/incident.

Noise is performance variabilities of the functions, which have no or negligible impacts on a future event and do not provide information about the future event.

Compared to the definition by Hollnagel (2004), the proposed definition differentiates signals and noise depending on whether they contribute to predicting future events. Additionally, the definition of weak signals addresses three aspects. First, weak signals are not restricted to be at technological level, and it could be performance variabilities of human or organization functions. Second, weak signals exist as combinations. An individual weak signal does not cause noticeable impacts until it interacts with others. Otherwise, it should be treated as a strong signal. Lastly, weak signals are early predictors. Even though “early” is a characteristic that has been well recognized in existing literature, the scope of weak signals in terms of how early is early was rarely clarified. Only Luyk (2011) explicitly defined weak signals need to be early enough to predict incident precursors, which are precisely defined as “a chain of adverse events flowing an initiating off-nominal event and that can lead to an accident” (Saleh et al., 2013). According to the concepts of weak signals and precursors, the study further precisely defined the characteristic “early” of weak signals. Since precursors are adverse events leading to an incident, in order to be early enough to indicate the precursors, weak signals are restricted to be conditions that could warn for adverse events. For example, relief valve that fails is a precursor event, but unqualified maintenance of the relief valve is a condition which can be a weak signal of the failure.

The life cycle from identifying to acting on weak signals consists of three stages (Brizon and Wybo, 2009, Holopainen and Toivonen, 2012).

First, weak signals need to be observed in the system. The stage requires an organization to collect data of weak signals so they can be monitored for further evaluations.

The second stage is evaluating the relevance of weak signals. At this stage, weak signals need to be interpreted based on knowledge and context to recognize their indications of future events. In most cases, one who interprets the weak signals is not the one who has power to act. Therefore, the relevance needs to be evaluated so that one could transmit the existence of weak signals to decision-makers when one thinks the relevance is strong enough or it meets any criterion in organizational standards.

Once the existence of weak signals is transmitted to decision-makers, decision-makers could determine whether to act on the weak signals and to prioritize actions based on the significance of weak signals.

Unfortunately, in complex systems such as process plants, challenges exist throughout the entire life cycle. Given the rapid development of computing technologies in the past decades, digitized process plants have the capability to collect a tremendous amount of various types of data from process operations, control rooms, business and information systems (Pasman, 2020, Qin, 2014, Xu et al., 2015). Information to be collected depends on the purpose of uses and knowledge. It is questionable whether the collected data covers all the sources of weak signals that are relevant to a selected hazard. In order to ensure weak signals are observable in the first stage of the life cycle, a technique is necessary to guide organizations to decide what information to be collected. On the other hand, even though weak signals are observable, evaluating the relevance of weak signals based on knowledge and awareness of individuals in the second stage is still challenging. In process industries, operations involve interactions among technological, human, and organizational components (Leveson, 2004, Perrow, 1999, Rasmussen and Whetton, 1997). The interactions among different components arise high complexities since the interactions could aggregate into nonlinear and circular relationships leading to emerging failures (Cameron et al., 2017). The complexity could be intellectually unmanageable making weak signals hard to identify and interpret (Leveson, 2000). Without solutions to overcome the challenges for the first two stages, it cannot guarantee weak signals are identified and transmitted to decision-makers with an explanation of significance. Furthermore, it is also doubtful whether the responses to weak signals are effective without an appropriate interpretation. In this study, a framework was developed to identify weak signals by addressing the challenges throughout the life cycle, which can help industry management to prevent incidents proactively.

Learning from past anomalies and incidents plays a critical role in interpreting weak signals (Brizon and Wybo, 2009, Guillaume, 2011), but advanced techniques are necessary to extract the knowledge from massive historical data. Machine learning techniques have caught increasing attention in the past decades for item classification and pattern recognition. They learn patterns from existing data then make predictions on future events (Ge et al., 2017, Han et al., 2011, Xu et al., 2015). In process industry, machine learning techniques are widely applied using process data for process and quality monitoring, fault identification and as soft sensor (Qin, 2014). However, other valuable information to identify weak signals such as equipment failures, quality-related issues, and performance measurement (Haji‐Kazemi and Andersen, 2013, Körvers, 2004) was seldom utilized to predict process incidents. Instead, such information was commonly used for preventing occupational incidents. Table 1 summarizes the studies which applied classification algorithms to extract knowledge from historical data. Most studies in the table utilized only historical incident data for understanding causes and consequences of past incidents, rather than predicting occurrences of potential incidents. On the other hand, data of both safe operations and incidents was used by (Goh and Chua, 2013, Goh et al., 2018, Poh et al., 2018, Sarkar et al., 2018) and the studies showed promising results of predicting incident occurrences relying on potential weak signals such as safety management elements. Therefore, applications of machine learning techniques in process industries should consider a wider range of data beyond process data to recognize and respond to existing weak signals in plants.

As discussed above, “weak signal” was not precisely defined in the existing literature and techniques to observe, evaluate, and respond to weak signals were not developed for process industry. This study proposed a formal definition of weak signals in the domain of safety. After reviewing characteristics and challenges of identifying weak signals, we developed a framework providing techniques to overcome the challenges throughout the life cycle of weak signals.

The rest of the paper is organized as follows. Section 2 details the development process of the framework. Section 3 demonstrates applications of the framework by a case study of a hypothetical batch polymerization process. Performances of the machine learning techniques in the framework are discussed. Additionally, how the framework can be applied in operations is demonstrated by an example instance. Finally, conclusions and future directions of the topic are addressed in Section 4.

Application in a batch polymerization process

The industrial-scale Methyl Methacrylic (MMA) polymerization process studied in Yu et al. (2020) was adopted in the current study. The framework to identify weak signals leading to potential temperature excursions was demonstrated through the case study. The batch process involves a field operator (FO) and a control room operator (CRO). In an 8-hour shift, CRO enters the recipe of the batch on the control panel, authorizes the solvent feeding process, starts the agitator, and authorizes the

Conclusions and future work

Proactive incident prevention requires the identification of weak signals and appropriate responses to the weak signals. The study developed a framework based on FRAM and machine learning techniques, which provides explicit guidance for the entire life cycle of weak signals from identifying to responding. The case study of a hypothetical batch process showed great potentials for applying the framework in real operations. Given the information of the potential weak signals that were extracted

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgments

The research presented in this paper was supported by the Mary Kay O′Connor Process Safety Center at the Texas A&M University. The data generation and the development of machine learning models were conducted with the advanced computing resources provided by Texas A&M High Performance Research Computing (HPRC). The consulting support provided by the HPRC staff is gratefully acknowledged. Additionally, the research has been dedicated to the memory of Dr. M. Sam Mannan, who initiated the research