While applying EKF to the system, observation error matrix (π ), plant covariance matrix (ππ) , initial state estimate πΜ+ and initial covariance matrix (π0) needs to be introduced.
The EKF variables are often chosen arbitrarily and the systematic technique for choosing EKF variable is a topic of active research (Schneider et.al. 2013). Since the performance of the filter is largely influenced by the user defined input EKF parameters, the tuning of the covariance matricesβ parameters is very important step as it may result in instability of the system, even if everything else is correct (Sanjuro et.al. 2017). The values of the plant covariance matrix tell filter the difference between the filter model and the model with measured value. Similarly, the observation error matrixβs value defines the error in the measurements obtained from sensors. The effect of these two covariance matrices ππ and π on the system output can be defined in simple term as higher the values the more difference in the filter model and model with measured data and more error in the sensor measurements, respectively. The covariance matrix ππ was assumed to have constant values in this study because of the lack of availability of the actual data and measured data is based on simulation. It is assumed that π is not time dependent and is generally provided by the manufacturer of the measuring device. Moreover, the filter performance is also dependent of the accuracy of the initial states, the initial covariance matrix π0 and it defines the confidence in the initial values used. The initial values and π0 selection is very critical because sometimes the filter performance is hardly effected by ππ and π but highly effected by the initial values and π0 (Miller et.al. 2008).
3.4.1 Plant covariance matrix (ππ)
The most difficult matrix to define in EKF ππ, refers to the inaccuracies expected in the modeling errors for example in the state equations. The ππ can be set to zero if the system is accurate perfectly, however there is always a chance to have some modelling errors, measurement errors, discretization error etc. (Rhudy, 2015). The ππ is generally introduced by the hit and trial method, firstly a nonzero ππβs value is fed through the equation and upon the study of its effect in the output, it can be further tuned (Laamari, et.al. 2014). For example, Kang et.al. found that when the plant covariance was less than 1e-16, the estimated values for displacement was very close to the actual value (Kang et.al. 2020). The ππ in this study was based on its effect on the result, the extremely small value of ππas of Kang et.al.
had no significant improvement in filter performance, hence the ππ value was chosen to be 1e-1.
3.4.2 Observation error matrix (π )
The observation error covariance matrix represents the intensity of the divergence of obtained data from the actual one from sensors (Liu et.al. 2019). It is assumed that π value is not time dependent and, in most cases, provided by the sensor manufacturer. However, the observation error matrix can be derived as,
π = ππ ππππ. (π2) (22)
where ππ β₯1, π is the standard deviation of the output from the sensor, depending on the confidence on the deviceβs output ππ can be raised even beyond 1. In an experiment conducted by Miller et.al. where they used EKF to identify bearing coefficient, it was found that π value when set between (0.2-2) Β·10-6 m, could give acceptable result (Miller et.al.
2008). The value for this study was hence chosen to be 1Β·10-3 m.
3.4.3 Initial covariance matrix (π0)
The initial state estimation πΜ+ is required for the implementation of EKF, however the true value of πΜ+ is seldom known. Hence, a covariance matrix π0 needs to be introduced which defines the uncertainty of the initial state. Unlike observation error matrix, the range of initial state uncertainty cannot be measured physically. Therefore, it is very important to set the initial covariance matrix. If the value of π0 is too small in the meantime πΜ+ has high divergence compared to the measured one, the resulting Kalman gain is also small and the filter relies on the numerical model heavily and conversely filter might diverge. There is no standard rule to set the initial covariance matrix, as it is done by setting a certain value and analysing its effect on the outcome. The initial covariance in this study was set to be 1Β·10-4, meaning there is relatively high confidence on the initial values.
3.4.4 State transition matrix (PHI)
State transition matrix defines the state of system from initial time to measured time, if initial time is supposed to be π‘0 and reference time is π‘ the state transition matrix will relate the state of a system from π‘0 to π‘. In this study state transition matrix is diagonal matrix of 1.
4 RESULT
In this section the results obtained while varying the two chosen parameters i.e., unbalance and bearing coefficient for the measured, model and EKF is presented. First the result obtained from varying unbalance mass and its effect in the displacement at each of the observed node in measured (true) and model (bad model) was considered and then the EKFβs correction was introduced. Same procedure was followed but with bearing coefficient as the varying parameters. The result obtained by varying the two parameters, one at a time for measured and modelled model and EKFβs correction was found as follow.