Univariate Alarm Research Articles

Unbiased estimation based multivariate alarm design considering temporal and multimodal process characteristics

In alarm systems, conventional univariate alarm methods often result in frequent false and missing alarms, calling for an urgent need to introduce multivariate information. For the multivariate alarm design, although each variable’s detection sensitivity is improved with the assistance of other variables, it is also susceptible to anomalies of other variables. Therefore, it is a challenge to properly introduce multivariate information, especially for complex processes subject to operating condition changes. This paper proposes a multivariate alarm framework that complements temporal and multimodal process characteristics to better alarm for variables by unbiased estimation. The temporal and multimodal characteristics are explored by prediction-oriented network structure and reconstruction-oriented network structure, respectively. To make them properly integrated, pattern labels that reveal the modes change are designed and used as the bridge between the temporal and the multimodal parts. On the one hand, the consideration of two types of characteristics allows perceiving temporal-related and modal-related faults, promoting sensitive alarm performance. On the other hand, unifying the two types of networks can eliminate the estimation bias of normal variables, making them unsusceptible to anomalies of other variables and promoting accurate alarm performance. Experiments on the coal mill prove the effectiveness of the proposed method regarding false alarm rate and missing alarm rate.

Performance analysis of up/down counters in alarm design

Alarm systems, as a layer of protection, play a critical role in process safety. Modern alarm systems, however, pose a challenge in the process industry due to alarm floods, missed alarms, and a large number of false, nuisance, and chattering alarms. In this paper, we propose up/down counter configurations as a technique to improve the quality of univariate alarm systems. Mathematical models are developed based on the Markov chain to study the performance of up/down counters and analytical expressions for performance indices (the false alarm rate, missed alarm rate, and expected detection delay) are derived. The results are confirmed using Monte Carlo simulation and the utility of the proposed technique is demonstrated through examples. This proposed up/down counter is specifically compared to delay-timers as a common and well-studied alarm configuration technique.

Evidence fusion-based alarm system design considering coarse and fine changes of process variable

In view of the coarse and fine changes of process variable in industrial systems, this paper introduces a univariate alarm design method based on dynamic evidence fusion. Firstly, in order to describe the coarse statistical characteristics of historical sample data, the multi-transition data segmentation based on memory and forgetting strategies and the referential evidential matrix (REM) construction are presented. Secondly, the real-time sample of process variable is transformed into alarm evidence by matching with REM, and then such multiple pieces of alarm evidence continuously acquired in time are fused by evidence reasoning (ER) rule with the interval-valued fusion weights and reliabilities of alarm evidence, so as to accurately adapt the fine change of process variable. Finally, numerical experiment and motor rotor alarm experiment are implemented to validate that the proposed method has better performances than traditional alarm design methods.

PDF-Based Technique for Univariate Alarm Systems and Its Application to Mixture Distributions

PDF-Based Technique for Univariate Alarm Systems and Its Application to Mixture Distributions

Data classification and performance evaluation for the most commonly-used univariate alarm systems

Alarm systems are critically important for safe and efficient operations of industrial plants, but many industrial alarm systems are suffering from too many nuisance alarms. This paper proposes a method to classify normal and abnormal data segments and evaluate performance indices for the most commonly used univariate alarm systems. The proposed method consists of three steps. First, piece-wise linear representations are exploited in separating historical data samples of an analog process variable configured with alarms into data segments with same qualitative trends. Second, data segments are classified into normal, abnormal and unclassified conditions via a mean hypothesis test; a required assumption is that data segments in normal and abnormal conditions have different mean values being distinguishable from alarm thresholds. Third, based on the normal and abnormal data, performance indices of univariate alarm systems are calculated, including two newly formulated ones as the false alarm duration ratio and the missed alarm duration ratio. The effectiveness of the proposed method is illustrated by numerical and industrial examples.

Multivariate statistical approaches for an early detection of foaming in a refinery SCOT unit

Abstract The formation of foam in amine units is an issue that plant operators and field personnel are confronted with on a regular basis. The inability to take proper actions in due time may result in plant downtime and increased emissions. Steep rises in differential pressure indicate foam formation, and are monitored manually in practice. Antifoaming agent is added in order to reduce foaming, but this is usually carried out under time pressure. Hence, plant operating authorities have expressed a strong interest in a data-driven solution capable of providing an early warning against foaming. The classical univariate alarm associated with differential pressure can be ineffective for foaming detection due to high misdetection rates and its lateness of detection. Modern univariate approaches based on pattern recognition techniques may not be suitable either for an early detection, as no universally distinctive features of differential pressure are observed prior to foaming in the present study. In this contribution, the multivariate statistical process monitoring approach based on principal component analysis (PCA) is applied to the early detection of foaming in a continuously operated Shell Claus Off-gas Treating (SCOT) unit of a major refinery in Germany. The results are extended to facilitate fully automated and adaptive modeling based on exponentially weighted recursive principal component analysis (EWRPCA).

On the use of penalty approach for design and analysis of univariate alarm systems

Alarm systems indicate abnormal conditions of the underlined plant equipment enabling operators to take corrective actions, and bring the equipment back to its normal condition. This paper presents a new approach for designing a generalized delay timer based on the penalty scenario and Markov chain schemes. The penalty approach is an extension for the well-known “n-sample on/off delay timer” approach in designing alarm systems. Three performance indices named, False Alarm Rate (FAR), Missed Alarm Rate (MAR) and Average Alarm Delay (AAD) are derived for the proposed penalty approach using Markov theory. Also, a new index named “Mean Time to Alarm (MTTA)” is introduced to analyze the underline alarm system, and to compute AAD. Finally, the effectiveness of the proposed method is investigated and compared with that of the other methods through a case study.

A multi-setpoint delay-timer alarming strategy for industrial alarm monitoring

Alarm systems play an important role in industry to notify operators of abnormal and fault situations. In real industrial plants, however, a majority of abnormal alarm phenomena, including false alarms and missed alarms, always interfere with operators' judgment. Mathematical analysis of traditional univariate alarm techniques indicates that the alarm setpoint, the dynamic alarm order, and the alarm algorithm are the three main elements of alarm annunciation and alarm clearance. Based on this, we proposed a multi-setpoint delay-timer alarming strategy which optimizes the false alarm rate (FAR), the missed alarm rate (MAR), and the averaged alarm delay (AAD) by increasing the number of univariate alarm setpoints. The proposed alarming strategy is able to achieve a balance between alarming accuracy and alarming sensitivity. Its effectiveness is demonstrated with both simulation and industrial case studies.

Analysis and design of time-deadbands for univariate alarm systems

Time-deadbands (or alarm latches) are popular alarm configuration methods used in industry to improve the alarm system performance. In this paper, time-deadband based configurations for the case of univariate alarm systems are analyzed. Mathematical models are developed based on Markov processes, and analytical expressions for performance indices (the false alarm rate, missed alarm rate, and expected detection delay) are derived. Systematic design procedures are also proposed, and the utility of the methods is illustrated through design examples.

Design of univariate alarm systems via rank order filters

Filtering is one of the techniques used in alarm system design to improve the performance of an alarm system. Due to the fact that the filtered data is no longer independent, computation of the performance indexes (false alarm rate (FAR), missed alarm rate (MAR) and expected detection delay (EDD)) is hard for filters. In this paper, rank order filters are applied in alarm system design. The output of rank order filters is restricted to one of the input samples, thus the probability density function (PDF) of the filtered data can be computed directly from the PDF of the raw data. This feature makes it possible to compute FAR and MAR for rank order filters directly. Further, a method to compute the expected detection delay is proposed for rank order filters despite the dependence of the filtered data. Simulation results shows that the order of rank order filters provide another degree-of-freedom in alarm system design besides the window size, which can be used to improve the alarm performance.

A heuristic method of variable selection based on principal component analysis and factor analysis for monitoring in a 300 kW MCFC power plant

In a commercialized 300 kW molten carbonate fuel cell (MCFC) power plant, a univariate alarm system that has only upper and lower limits is usually employed to identify abnormal conditions in the system. Even though univariate alarms have already been adopted for system monitoring, this simple monitoring system is limited for using in an extended monitoring system for fault diagnosis. Therefore, based on principal component analysis (PCA), a recursive variable grouping method for a multivariate monitoring system in a commercialized MCFC power plant is presented in this paper. In terms of development, since a principal component analysis model that contains all system variables cannot isolate a system fault, heuristic recursive variable selection method using factor analysis is presented here. To verify the performance of the fault detection, real plant operations data are used. Furthermore, comparison between type 1 and type 2 errors for four different variable groups demonstrates that the developed heuristic method works well when system faults occur. These monitoring techniques can reduce the number of false alarms occurring on site at MCFC power plant.

Performance Assessment and Design for Univariate Alarm Systems Based on FAR, MAR, and AAD

The performance of a univariate alarm system can be assessed in many cases by three indices, namely, the false alarm rate (FAR), missed alarm rate (MAR), and averaged alarm delay (AAD). First, this paper studies the definition and computation of the FAR, MAR, and AAD for the basic mechanism of alarm generation solely based on a trip point, and for the advanced mechanism of alarm generation by exploiting alarm on/off delays. Second, a systematic design of alarm systems is investigated based on the three performance indices and the tradeoffs among them. The computation of FAR, MAR, and AAD and the design of alarm systems require the probability density functions (PDFs) of the univariate process variable in the normal and abnormal conditions. Thus, a new method based on mean change detection is proposed to estimate the two PDFs. Numerical examples and an industrial case study are provided to validate the obtained theoretical results on the FAR, MAR and AAD, and to illustrate the proposed performance assessment and alarm system design procedures.

Application of Multivariate Statistics for Efficient Alarm Generation

Ever since the introduction of Distributed Control Systems (DCS), there has been little motivation for limiting the number of alarms that could be configured for process monitoring. As a result, plant operators are overwhelmed with many alarms during process upsets. Although a handful of these alarms are informative, many of them are a nuisance to the operator. However, all the alarms need to be acknowledged. Generally, check limits on univariate alarms are based on statistical quality control (three sigma limits, also known as Shewhart charts). While annunciating a univariate alarm on a particular variable, the information from other variables is often ignored. Modern day process plants have variables which are highly correlated. This correlation structure can be exploited in the efficient management of alarms. This work demonstrates the advantages of monitoring the PCA based T2 and Q statistic over individual process variables. Monitoring these higher level statistics will not only reduce the false alarm and missed alarm rates but also reduces the detection latency which is one of the main drawbacks of monitoring a filtered variable. Two simulation examples are shown to illustrate the utility of the proposed method.