Elsevier

Journal of Process Control

Volume 24, Issue 9, September 2014, Pages 1382-1393
Journal of Process Control

Nearest neighbors method for detecting transient disturbances in process and electromechanical systems

https://doi.org/10.1016/j.jprocont.2014.06.007Get rights and content

Highlights

  • The paper proposes a method to detect transient disturbances.

  • The method is robust to other features like oscillations, noise or changes in operation level.

  • It uses a nearest neighbors technique and assigns high anomaly indices for periods with transients.

  • Statistical analysis of the detection threshold, and recommended parameter values are included.

  • The paper also suggests a color map to visualize the detection results.

Abstract

Transient disturbances are increasingly relevant in process industries which rely on electromechanical equipment. Existing data-driven methods for detecting transient disturbances assume a distinct amplitude or time-frequency component. This paper proposes a detection method which is more generic and handles any short-term deviation of a measurement from its overall trend, regardless of whether the trend incorporates features such as oscillations, noise or changes in operation level. The method is based on a nearest neighbors technique and builds a vector of anomaly indices which are high for the period with the transient disturbance. The paper includes analyses of the statistical significance of the threshold proposed and of the sensitivity of the parameters, and it also suggests a color map to visualize the detection results. The method is demonstrated on experimental and industrial case studies.

Introduction

In the past decade, analysis of disturbances in chemical process systems, for example for detection and diagnosis, has shifted from being focused on single control loops [1] to taking a plant-wide approach [2], [3]. This extension made sense because chemical processes are highly interconnected, thus disturbances often propagate throughout the plant along mass and energy flows, and control signals.

Similarly, it would be meaningful to further extend disturbance analysis to the equipment and utilities which service the process. The reason is that these subsystems also interact with the process through energy and signal paths, and therefore disturbances can also propagate between them [4], [5].

The electrical utility and associated electromechanical equipment are increasingly important sources of disturbances to the process industry. Examples of electromechanical equipment in process plants are the electric motors used to drive pumps and compressors. The use of electrical energy to drive process machinery is increasingly common due to the greater energy-efficiency and easier maintenance of electromechanical equipment compared with traditional gas turbines [6]. Electromechanical equipment has its own fault modes and is susceptible to power quality disturbances, that is, deviations of the voltage or current in the power supply from their ideal behavior [7]. In addition, a recent report [8] indicates that power quality disturbances are becoming more frequent.

The susceptibility of the process to disturbances in the electrical utility and associated electromechanical equipment was evidenced in two recent episodes, in March 2013. A water pump failure in the first case [9] and a power cut in the second [10] disrupted two industrial gas processing plants and temporarily stopped the supply of natural gas to the UK.

The extension of process disturbance analysis to electromechanical measurements makes it essential to detect transient disturbances. The reason is that disturbances related to the electrical utility are mostly of a transient nature, caused by power imbalances in the grid which lead to momentary frequency and voltage instabilities [11]. A transient disturbance is defined in this paper as a sudden and short-lived deviation of a measurement from its previous and subsequent trend. After the transient, the measurement may return to its previous trend or follow a different trend. Examples of the former are voltage spikes and deviations caused by sensor faults [12]. Examples of the latter are the responses of the system to step changes, as in Fig. 1.

Transient disturbances are well known to practitioners in process industries but previous research on data-driven plant-wide analysis methods has tended to focus on persistent disturbances, which are repetitive and last for long time horizons [2], [3], [13]. Methods for persistent disturbances, however, are not appropriate for transient disturbances because they rely on the repetition of the abnormal dynamic episode [14], [15], [16], [17].

Typically, the data-driven methods for detecting transient disturbances described in academic literature rely on features such as time-domain amplitude or components in the time-frequency domain to distinguish transients from the normal trend of the measurement. Traditional methods of statistical process monitoring are examples of the former, using thresholds on the amplitude to flag unwanted operation levels. However, this approach is not appropriate if the dynamics of the system generating the measurement is oscillatory or cyclical in nature. Misra et al. [12] use wavelet decomposition to identify the transients in the time-frequency domain. The authors assume that only the parts of the signal affected by transient disturbances map to wavelet coefficients of high amplitude in the lower scales.

The detection method proposed in this paper is, however, more generic than these previous contributions in that it handles any short-term deviation from the overall measurement trend, irrespective of the frequency components or relative amplitude of that deviation. To do that, the basic idea is to consider the measurement as a time series and hence the transient disturbance as the unusual segment. This problem is framed as an anomaly detection problem and solved with nearest neighbors methods. Furthermore, the method proposed does not require the development of models of routine or healthy operation, in contrast to statistical process monitoring techniques.

The contributions of this paper are (i) a method to detect transient disturbances, (ii) statistical analysis of the threshold for detection, (iii) recommendations and analysis of parameter values, and (iv) a color map to visualize the results of the transient detection analysis in a compact way that suggests the propagation of the disturbance through the process.

The paper is organized as follows. Section 2 presents the background and related work on the nearest neighbors concept. Section 3 explains the method proposed in detail, analyses the statistics of the threshold for detection, and presents the color map to visualize the results. These contributions are illustrated on a reference example, obtained from experimental work with a gas compression rig. Section 4 recommends appropriate values for the parameters of the method and analyses the sensitivity of the detection results to those values. Section 5 presents an industrial case study and uses it to test the methods. The paper ends with comments and conclusions.

Section snippets

Background

This section introduces the use of nearest neighbors to detect anomalous segments in a time series. This technique relies on a similarity measure, and on using the similarity assessment to define an anomaly index for each segment of the time series. Hence, this section also discusses similarity measures and anomaly index definitions.

Detection of transient disturbances

This section explains the method proposed to detect and characterize transient disturbances. A reference example is first presented and then used to illustrate the explanation.

Parameter settings and sensitivity

To generate the anomaly index vector, the following parameters have to be selected:

  • Embedding granularity τ.

  • Embedding dimension m.

  • Embedding step δ.

  • Number of nearest neighbors k.

The objective of this section is to find adequate values for these parameters, given the dynamics of a particular system. The section starts by relating the physical meaning of the parameters with the dynamics of the system, and then recommends the best parameter values and analyses the sensitivity of the detection

Application to industrial case study

The test case study is part of an industrial process of gas processing, courtesy of ABB Oil, Gas and Petrochemicals, Oslo, Norway. Fig. 11 shows the selected part of the process, which includes a gas-condensate separation section, with a separator, filters and stabilizer, and a gas recompression section, with scrubbers and compressors. The speed of the compressors is used to adjust the pressure in the system, either at the outlet of the separator or at the outlet of the stabilizer.

Fig. 12a

Conclusion

This paper has demonstrated the use of a nearest neighbors technique in the detection of transient disturbances in measurements from chemical process systems and associated electromechanical equipment. The extension of process disturbance analysis to electromechanical measurements is justified by the increasing use of electrically-driven machinery in process industries and the rising number of power quality incidents.

The proposed method is based on the concept of anomaly detection, and was

Acknowledgments

The authors gratefully acknowledge the financial support from the Portuguese Foundation for Science and Technology (FCT) under Fellowship SFRH/BD/61384/2009 and the Marie Curie FP7-IAPP project “REAL-SMART – using real-time measurements for monitoring and management of power transmission dynamics for the Smart Grid”, Contract No: PIAP-GA-2009-251304.

Inês M. Cecílio would also like to thank P. Lipnicki, D. Lewandowski, and M. Wojcik of ABB Corporate Research Center, Kraków, Poland, for enabling

References (38)

  • M.H.J. Bollen

    Understanding Power Quality Problems: Voltage Sags and Interruptions

    (2000)
  • Statnett

    Systemdrifts-og markedsutviklingsplan 2012

    (2012)
  • S. Pfeifer

    Gas price spike underlines UK supply fears

    (October 2013)
  • S. Pfeifer

    Weather and Low Stocks Add to Gas Worries

    (October 2013)
  • H. Bevrani

    Robust Power System Frequency Control

    (2009)
  • N. Thornhill

    Finding the source of nonlinearity in a process with plant-wide oscillation

    IEEE Trans. Control Syst. Technol.

    (2005)
  • M. Bauer et al.

    Finding the direction of disturbance propagation in a chemical process using transfer entropy

    IEEE Trans. Control Syst. Technol.

    (2007)
  • V. Chandola et al.

    Anomaly detection: a survey

    ACM Comput. Surv.

    (2009)
  • Cited by (31)

    • From oscillatory to non-oscillatory disturbances: A comparative review of root cause analysis methods

      2022, Journal of Process Control
      Citation Excerpt :

      Fig. 6 shows the values of the anomaly indices for the embedded vectors computed on the measurements of the transient weld disturbance of Fig. 5. The dotted lines indicate the detection threshold computed according to [53]. Table 6 indicates the index of the first abnormal embedded vector for each variable affected by the transient weld disturbance, which indicates when the disturbance is detected in each variable.

    • The autonomous industrial plant – future of process engineering, operations and maintenance

      2020, Journal of Process Control
      Citation Excerpt :

      Therefore, it is a much better fit to be provided to human operator (hence, supporting his/her monitoring task) and is also more suitable as input for the other autonomy capabilities. Among the most relevant technologies are industrial applications of computer vision [15], vibration analysis [16,17], machine learning [18,19], and deep learning, in particular with convolutional and recurrent deep learning networks [20]. Understand: Understanding builds upon the information provided by perception capabilities.

    View all citing articles on Scopus
    View full text