An Algorithm for Finding Process Identification Intervals from Normal Operating Data

Abstract

Performing experiments for system identification is often a time-consuming task which may also interfere with the process operation. With memory prices going down and the possibility of cloud storage, years of data is more and more commonly stored (without compression) in a history database. In such stored data, there may already be intervals informative enough for system identification. Therefore, the goal of this project was to find an algorithm that searches and marks intervals suitable for process identification (rather than completely autonomous system identification). For each loop, four stored variables are required: setpoint, manipulated variable, measured process output and mode of the controller. The essential features of the method are the search for excitation of the input and output, followed by the estimation of a Laguerre model combined with a hypothesis test to check that there is a causal relationship between process input and output. The use of Laguerre models is crucial to handle processes with deadtime without explicit delay estimation. The method was tested on three years of data from about 200 control loops. It was able to find all intervals in which known identification experiments were performed as well as many other useful intervals in closed/open loop operation.

Identification of continuous-time models with slowly time-varying parameters

“…The switch at t s,iv should be carried out once convergence of the RLSSVF estimates is achieved. This can be assessed by checks on one (or more) of the following: (i) the variability or steady state condition of ∆θ(t k ) =θ(t k−1 )−θ(t k ), (ii) the size of the parameter covariance matrix P (t k ), which can be measured in terms of trace or determinant, (iii) the condition number of the parameter covariance matrix P (t k ), (iv) the prediction error ε(t k ), and (v) the output error ε y (t k ) defined in (33).…”

“…For instance, the variability of ∆θ(t k ) can be measured through its variance, as computed recursively with, for example, a simple exponential moving average of the kind used in an analogous manner by [33]. The steady state condition of the output error can be assessed using the approach proposed in [34].…”

Historical Data Segmentation for System Identification: A Comparison of Methods

“…The problem was explicitly introduced by the first time in (Peretzki et al, 2011), where an approach based on the Laguerre Model structure was used to find suitable intervals for system identification in closed-loop systems. A more detailed version of this work was published in (Bittencourt et al, 2015). A similar approach can be found in Huang, 2013a), but using an Autoregressive Moving Average with Exogenous Inputs (ARX) model structure.…”

“…The multivariable problem was introduced in (Patel, 2016), where an extension of the works in (Peretzki et al, 2011) and in (Bittencourt et al, 2015) is proposed to include Multiple Input Multiple Output (MIMO) systems in the analysis of the open-loop identification scenario. The multivariable problem for closed-loop systems was addressed in Kroll, 2017a).…”

“…Herein, two methods are compared based on the reviewed literature. More specifically, the works published in (Peretzki et al, 2011), Huang, 2013a), (Bittencourt et al, 2015), (Ribeiro;Aguirre, 2015) and Kroll, 2017a) are summarized in two methods, both using numerical conditioning as the basis to validate adequate time intervals to obtain a process model operating in closed-loop control. Moreover, the top-down change-point detection presented in (Wang et al, 2018) is proposed as an optional first step of the explained methods.…”