We should note that although the academia and the industry, as well as the dif-ferent industries and industrial sites, monitor (or propose to monitor) difdif-ferent key performance indicators for the measurement of alarm system performance, no single solution has been found as a universal measure of efficiency. In the present study, a systematic overview and categorization of alarm system performance met-rics were provided. Based on the aim and approach of the measurements, seven main categories were defined and discussed. A systematic difference in usage of
the term of alarm system performance was revealed between the academia and industry: the academia mainly measures the optimality of alarm thresholds, while the industry, in the absence of the information of correct alarms, mainly measures the overall alarm load and its highest contributors. Moreover, the main metrics monitored by the industry proved to show an extremely unbalanced picture to-wards the quantitative measurements instead of qualitative ones. I presented how the informativeness and actionability of alarms should be taken into consideration.
Finally, the analysis of the industrial hydrofluoric acid alkylation unit before and after the alarm rationalization process proved how extremely complex the ration-alization process is and called attention to the problems of unilateral measurement of alarm systems.
The categorization of alarm system performance metrics provides the basis for the 1.3 subthesis of my first thesis finding, as these metrics support the maintenance of a well-functioning alarm system, which is a prerequisite for the application of advanced data-driven alarm management solutions. My thesis findings are summerized in Section 8.
Decision trees for informative process alarm definition and alarm-based fault classification
Alarm messages in industrial processes are designed to draw attention to abnormal-ities that require timely assessment or intervention. However, in practice, alarms are arbitrarily and excessively defined by process operators resulting in numerous nuisance and chattering alarms that are simply a source of distraction. Numerous techniques are available for the retrospective filtering of alarm data, e.g., adding time delays and deadbands to existing alarm settings. As an alternative, in the present chapter, instead of filtering or modifying existing alarms, a method for the design of alarm messages being informative for fault detection is proposed which takes into consideration that the occurring alarm messages originally should be optimal for fault detection and identification. This methodology utilizes a ma-chine learning technique, the decision tree classifier, which provides linguistically well-interpretable models without the modification of the measured process vari-ables. Furthermore, an online application of the defined alarm messages for fault identification is presented using a sliding window-based data preprocessing ap-proach. The effectiveness of the proposed methodology is demonstrated in terms of the analysis of a well-known benchmark simulator of a vinyl-acetate production process, where the complexity of the simulator is considered to be sufficient for the testing of alarm systems.
47
Note to practitioners: Process-specific knowledge can be used to label histor-ical process data to normal operating and fault-specific periods. Alarm generation should be designed to be able to detect and isolate faulty states. Using decision trees, optimal "cuts" or alarm limits for the purpose of fault classification can be defined utilizing a labelled dataset. The results apply to a variety of indus-tries operating with online control systems, and especially timely in the chemical industry.
3.1 Introduction
Production costs, product quality and, most importantly, process safety are primar-ily affected by the efficiency of the monitoring and control of industrial processes.
Consequently, the development of advanced fault detection and diagnosis methods is the focus of many research studies. Numerous well-advanced model- and process history-based methods are available for the purpose of monitoring and fault detec-tion [59], however, still the simplest sensor-signal-based alarming technique is the most widespread and accepted state-of-the-art industrial practice due to its simpli-city. This technique is not just considered too simple for complex fault detection and diagnosis tasks (a single alarm should be associated with each malfunction), but the poor design and maintenance of alarm systems make the situation even worse.
Another reason for poor alarm configurations is the simpilicity of their deployment.
In distributed control systems (DCS) or supervisory control and data acquisition (SCADA) systems, the definition of alarm messages is determined purely compu-tationally and entails no significant costs at all. As a result, alarm thresholds are arbitrarily and excessively defined by process operators. These poorly configured alarm limits are the direct cause of the numerous process alarms. Redundant (uninformative alarms for a single abnormality [9]) and chattering alarms [47]
(chattering alarms are previously explained in Section 2.3.3) are quite common examples, but long-standing alarms are also well-known. In critical situations at plants, a flood of alarms overloads operators, significantly hindering or completely inhibiting the detection of root causes. The performance of alarm systems is de-graded by not just the poor parametrization of alarms, but generally by the many alarmed variables as well.
The primary indicators for the operators of (chemical) technologies from faulty regions of operation are the alarm variables. In addition to the indication of quality issues and malfunctions associated with distinct process variables, complex fault detection and isolation efforts should be supported by the co-occurrence of alarm messages. Alarms are raised when a certain process variable exceeds its associated limits, more specifically, a low or high alarm sounds when the variable falls below its lower or exceeds its upper specification limit (alarm threshold), respectively.
It should be noted that according to most industrial practices, multiple higher and lower thresholds can be defined as well - the respective alarm messages are referred to as, for example, high-high or low-low alarms, and although the present methodology can address this issue, it was decided to ignore these multiple upper and lower limits in most cases for the sake of simplicity. The purpose of these alarm messages, besides the most intuitive, univariate aim of drawing attention to the value of a single variable, is to indicate the presence of malfunctions (or their absence in the case of normal operating conditions). The philosophy of complex, multivariate and alarm-based fault diagnostic lies in the co-occurrence of multiple alarm messages, which form a characteristic fingerprint of certain malfunctions and indicate their presence. However, the analysis and interpretation of these alarm messages and, therefore, the way the operator extracts process-specific knowledge from them as well as their design approach are directly inter-dependent. Therefore, simply phrased, a two-way causality should describe the process of alarm design and the applied philosophy concerning fault identification: (1) the design of alarm messages determines how sounding alarms can be utilized for fault identification by the operators and, in turn, (2) the way the operators aim to utilize alarm messages for fault detection should determine the philosophy concerning the alarm design.
Therefore, the key research question of the present work is how to facilitate the work of the operators of chemical technologies by alarm messages that indicate the faults of the technology.